Merge branch 'master' into DRTVWR-515-maint

74 files changed, 4379 insertions, 3633 deletions

diff --git a/indra/llcommon/CMakeLists.txt b/indra/llcommon/CMakeLists.txt
index af41b9e460..eeb315ead6 100644
--- a/indra/llcommon/CMakeLists.txt
+++ b/indra/llcommon/CMakeLists.txt
@@ -1,4 +1,3 @@
-
 # -*- cmake -*-
 
 project(llcommon)
@@ -44,7 +43,6 @@ set(llcommon_SOURCE_FILES
     llcleanup.cpp
     llcommon.cpp
     llcommonutils.cpp
-    llcoro_get_id.cpp
     llcoros.cpp
     llcrc.cpp
     llcriticaldamp.cpp
@@ -106,6 +104,7 @@ set(llcommon_SOURCE_FILES
     llstring.cpp
     llstringtable.cpp
     llsys.cpp
+    lltempredirect.cpp
     llthread.cpp
     llthreadlocalstorage.cpp
     llthreadsafequeue.cpp
@@ -146,7 +145,7 @@ set(llcommon_HEADER_FILES
     llcleanup.h
     llcommon.h
     llcommonutils.h
-    llcoro_get_id.h
+    llcond.h
     llcoros.h
     llcrc.h
     llcriticaldamp.h
@@ -186,9 +185,9 @@ set(llcommon_HEADER_FILES
     llkeythrottle.h
     llleap.h
     llleaplistener.h
-    lllistenerwrapper.h
     llliveappconfig.h
     lllivefile.h
+    llmainthreadtask.h
     llmd5.h
     llmemory.h
     llmemorystream.h
@@ -230,6 +229,7 @@ set(llcommon_HEADER_FILES
     llstaticstringtable.h
     llstatsaccumulator.h
     llsys.h
+    lltempredirect.h
     llthread.h
     llthreadlocalstorage.h
     llthreadsafequeue.h
@@ -247,6 +247,7 @@ set(llcommon_HEADER_FILES
     llwin32headers.h
     llwin32headerslean.h
     llworkerthread.h
+    lockstatic.h
     stdtypes.h
     stringize.h
     timer.h
@@ -291,7 +292,7 @@ target_link_libraries(
     ${JSONCPP_LIBRARIES}
     ${ZLIB_LIBRARIES}
     ${WINDOWS_LIBRARIES}
-    ${BOOST_COROUTINE_LIBRARY}
+    ${BOOST_FIBER_LIBRARY}
     ${BOOST_CONTEXT_LIBRARY}
     ${BOOST_PROGRAM_OPTIONS_LIBRARY}
     ${BOOST_REGEX_LIBRARY}
@@ -320,13 +321,14 @@ if (LL_TESTS)
       ${LLCOMMON_LIBRARIES} 
       ${WINDOWS_LIBRARIES} 
       ${GOOGLEMOCK_LIBRARIES} 
-      ${BOOST_COROUTINE_LIBRARY} 
+      ${BOOST_FIBER_LIBRARY} 
       ${BOOST_CONTEXT_LIBRARY} 
       ${BOOST_THREAD_LIBRARY} 
       ${BOOST_SYSTEM_LIBRARY})
   LL_ADD_INTEGRATION_TEST(commonmisc "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(bitpack "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llbase64 "" "${test_libs}")
+  LL_ADD_INTEGRATION_TEST(llcond "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(lldate "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(lldeadmantimer "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(lldependencies "" "${test_libs}")
@@ -338,6 +340,7 @@ if (LL_TESTS)
   LL_ADD_INTEGRATION_TEST(llheteromap "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llinstancetracker "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llleap "" "${test_libs}")
+  LL_ADD_INTEGRATION_TEST(llmainthreadtask "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llpounceable "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llprocess "" "${test_libs}")
   LL_ADD_INTEGRATION_TEST(llprocessor "" "${test_libs}")
diff --git a/indra/llcommon/StackWalker.cpp b/indra/llcommon/StackWalker.cpp
index c0d3104099..56defc6465 100644
--- a/indra/llcommon/StackWalker.cpp
+++ b/indra/llcommon/StackWalker.cpp
@@ -98,7 +98,10 @@
 // If VC7 and later, then use the shipped 'dbghelp.h'-file
 #pragma pack(push,8)
 #if _MSC_VER >= 1300
+#pragma warning (push)
+#pragma warning (disable:4091) // a microsoft header has warnings. Very nice.
 #include <dbghelp.h>
+#pragma warning (pop)
 #else
 // inline the important dbghelp.h-declarations...
 typedef enum {
@@ -422,7 +425,7 @@ public:
   LPSTR m_szSymPath;
 
 #pragma pack(push,8)
-typedef struct IMAGEHLP_MODULE64_V3 {
+struct IMAGEHLP_MODULE64_V3 {
     DWORD    SizeOfStruct;           // set to sizeof(IMAGEHLP_MODULE64)
     DWORD64  BaseOfImage;            // base load address of module
     DWORD    ImageSize;              // virtual size of the loaded module
@@ -450,7 +453,7 @@ typedef struct IMAGEHLP_MODULE64_V3 {
     BOOL     Publics;                // contains public symbols
 };
 
-typedef struct IMAGEHLP_MODULE64_V2 {
+struct IMAGEHLP_MODULE64_V2 {
     DWORD    SizeOfStruct;           // set to sizeof(IMAGEHLP_MODULE64)
     DWORD64  BaseOfImage;            // base load address of module
     DWORD    ImageSize;              // virtual size of the loaded module
@@ -657,7 +660,7 @@ private:
     pGMI = (tGMI) GetProcAddress( hPsapi, "GetModuleInformation" );
     if ( (pEPM == NULL) || (pGMFNE == NULL) || (pGMBN == NULL) || (pGMI == NULL) )
     {
-      // we couldnīt find all functions
+      // we couldn't find all functions
       FreeLibrary(hPsapi);
       return FALSE;
     }
diff --git a/indra/llcommon/StackWalker.h b/indra/llcommon/StackWalker.h
index 834f89c471..4634765d0b 100644
--- a/indra/llcommon/StackWalker.h
+++ b/indra/llcommon/StackWalker.h
@@ -148,7 +148,7 @@ protected:
     CHAR loadedImageName[STACKWALK_MAX_NAMELEN];
   } CallstackEntry;
 
-  typedef enum CallstackEntryType {firstEntry, nextEntry, lastEntry};
+  enum CallstackEntryType {firstEntry, nextEntry, lastEntry};
 
   virtual void OnSymInit(LPCSTR szSearchPath, DWORD symOptions, LPCSTR szUserName);
   virtual void OnLoadModule(LPCSTR img, LPCSTR mod, DWORD64 baseAddr, DWORD size, DWORD result, LPCSTR symType, LPCSTR pdbName, ULONGLONG fileVersion);
diff --git a/indra/llcommon/llapp.cpp b/indra/llcommon/llapp.cpp
index 421af3006e..3dab632aef 100644
--- a/indra/llcommon/llapp.cpp
+++ b/indra/llcommon/llapp.cpp
@@ -49,6 +49,8 @@
 #include "google_breakpad/exception_handler.h"
 #include "stringize.h"
 #include "llcleanup.h"
+#include "llevents.h"
+#include "llsdutil.h"
 
 //
 // Signal handling
@@ -561,10 +563,42 @@ void LLApp::runErrorHandler()
 	LLApp::setStopped();
 }
 
+namespace
+{
+
+static std::map<LLApp::EAppStatus, const char*> statusDesc
+{
+    { LLApp::APP_STATUS_RUNNING,  "running" },
+    { LLApp::APP_STATUS_QUITTING, "quitting" },
+    { LLApp::APP_STATUS_STOPPED,  "stopped" },
+    { LLApp::APP_STATUS_ERROR,    "error" }
+};
+
+} // anonymous namespace
+
 // static
 void LLApp::setStatus(EAppStatus status)
 {
-	sStatus = status;
+    sStatus = status;
+
+    // This can also happen very late in the application lifecycle -- don't
+    // resurrect a deleted LLSingleton
+    if (! LLEventPumps::wasDeleted())
+    {
+        // notify interested parties of status change
+        LLSD statsd;
+        auto found = statusDesc.find(status);
+        if (found != statusDesc.end())
+        {
+            statsd = found->second;
+        }
+        else
+        {
+            // unknown status? at least report value
+            statsd = LLSD::Integer(status);
+        }
+        LLEventPumps::instance().obtain("LLApp").post(llsd::map("status", statsd));
+    }
 }
 
 
diff --git a/indra/llcommon/llapr.h b/indra/llcommon/llapr.h
index da50dda103..3c07976f42 100644
--- a/indra/llcommon/llapr.h
+++ b/indra/llcommon/llapr.h
@@ -41,17 +41,7 @@
 
 #include "llstring.h"
 
-#if LL_WINDOWS
-#pragma warning (push)
-#pragma warning (disable:4265)
-#endif
-// warning C4265: 'std::_Pad' : class has virtual functions, but destructor is not virtual
-
-#include <mutex>
-
-#if LL_WINDOWS
-#pragma warning (pop)
-#endif
+#include "mutex.h"
 
 struct apr_dso_handle_t;
 /**
diff --git a/indra/llcommon/llcond.h b/indra/llcommon/llcond.h
new file mode 100644
index 0000000000..e31b67d893
--- /dev/null
+++ b/indra/llcommon/llcond.h
@@ -0,0 +1,405 @@
+/**
+ * @file   llcond.h
+ * @author Nat Goodspeed
+ * @date   2019-07-10
+ * @brief  LLCond is a wrapper around condition_variable to encapsulate the
+ *         obligatory condition_variable usage pattern. We also provide
+ *         simplified versions LLScalarCond, LLBoolCond and LLOneShotCond.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if ! defined(LL_LLCOND_H)
+#define LL_LLCOND_H
+
+#include "llunits.h"
+#include "llcoros.h"
+#include LLCOROS_MUTEX_HEADER
+#include "mutex.h"
+#include <chrono>
+
+/**
+ * LLCond encapsulates the pattern required to use a condition_variable. It
+ * bundles subject data, a mutex and a condition_variable: the three required
+ * data objects. It provides wait() methods analogous to condition_variable,
+ * but using the contained condition_variable and the contained mutex. It
+ * provides modify() methods accepting an invocable to safely modify the
+ * contained data and notify waiters. These methods implicitly perform the
+ * required locking.
+ *
+ * The generic LLCond template assumes that DATA might be a struct or class.
+ * For a scalar DATA type, consider LLScalarCond instead. For specifically
+ * bool, consider LLBoolCond.
+ *
+ * Use of LLCoros::ConditionVariable makes LLCond work between
+ * coroutines as well as between threads.
+ */
+template <typename DATA>
+class LLCond
+{
+public:
+    typedef DATA value_type;
+
+private:
+    // This is the DATA controlled by the condition_variable.
+    value_type mData;
+    // condition_variable must be used in conjunction with a mutex. Use
+    // LLCoros::Mutex instead of std::mutex because the latter blocks
+    // the entire calling thread, whereas the former blocks only the current
+    // coroutine within the calling thread. Yet LLCoros::Mutex is safe to
+    // use across threads as well: it subsumes std::mutex functionality.
+    LLCoros::Mutex mMutex;
+    // Use LLCoros::ConditionVariable for the same reason.
+    LLCoros::ConditionVariable mCond;
+
+public:
+    /// LLCond can be explicitly initialized with a specific value for mData if
+    /// desired.
+    LLCond(const value_type& init=value_type()):
+        mData(init)
+    {}
+
+    /// LLCond is move-only
+    LLCond(const LLCond&) = delete;
+    LLCond& operator=(const LLCond&) = delete;
+
+    /// get() returns a const reference to the stored DATA. The only way to
+    /// get a non-const reference -- to modify the stored DATA -- is via
+    /// update_one() or update_all().
+    const value_type& get() const { return mData; }
+
+    /**
+     * Pass update_one() an invocable accepting non-const (DATA&). The
+     * invocable will presumably modify the referenced DATA. update_one()
+     * will lock the mutex, call the invocable and then call notify_one() on
+     * the condition_variable.
+     *
+     * For scalar DATA, it's simpler to use LLScalarCond::set_one(). Use
+     * update_one() when DATA is a struct or class.
+     */
+    template <typename MODIFY>
+    void update_one(MODIFY modify)
+    {
+        { // scope of lock can/should end before notify_one()
+            LLCoros::LockType lk(mMutex);
+            modify(mData);
+        }
+        mCond.notify_one();
+    }
+
+    /**
+     * Pass update_all() an invocable accepting non-const (DATA&). The
+     * invocable will presumably modify the referenced DATA. update_all()
+     * will lock the mutex, call the invocable and then call notify_all() on
+     * the condition_variable.
+     *
+     * For scalar DATA, it's simpler to use LLScalarCond::set_all(). Use
+     * update_all() when DATA is a struct or class.
+     */
+    template <typename MODIFY>
+    void update_all(MODIFY modify)
+    {
+        { // scope of lock can/should end before notify_all()
+            LLCoros::LockType lk(mMutex);
+            modify(mData);
+        }
+        mCond.notify_all();
+    }
+
+    /**
+     * Pass wait() a predicate accepting (const DATA&), returning bool. The
+     * predicate returns true when the condition for which it is waiting has
+     * been satisfied, presumably determined by examining the referenced DATA.
+     * wait() locks the mutex and, until the predicate returns true, calls
+     * wait() on the condition_variable.
+     */
+    template <typename Pred>
+    void wait(Pred pred)
+    {
+        LLCoros::LockType lk(mMutex);
+        // We must iterate explicitly since the predicate accepted by
+        // condition_variable::wait() requires a different signature:
+        // condition_variable::wait() calls its predicate with no arguments.
+        // Fortunately, the loop is straightforward.
+        // We advise the caller to pass a predicate accepting (const DATA&).
+        // But what if they instead pass a predicate accepting non-const
+        // (DATA&)? Such a predicate could modify mData, which would be Bad.
+        // Forbid that.
+        while (! pred(const_cast<const value_type&>(mData)))
+        {
+            mCond.wait(lk);
+        }
+    }
+
+    /**
+     * Pass wait_for() a chrono::duration, indicating how long we're willing
+     * to wait, and a predicate accepting (const DATA&), returning bool. The
+     * predicate returns true when the condition for which it is waiting has
+     * been satisfied, presumably determined by examining the referenced DATA.
+     * wait_for() locks the mutex and, until the predicate returns true, calls
+     * wait_for() on the condition_variable. wait_for() returns false if
+     * condition_variable::wait_for() timed out without the predicate
+     * returning true.
+     */
+    template <typename Rep, typename Period, typename Pred>
+    bool wait_for(const std::chrono::duration<Rep, Period>& timeout_duration, Pred pred)
+    {
+        // Instead of replicating wait_until() logic, convert duration to
+        // time_point and just call wait_until().
+        // An implementation in which we repeatedly called
+        // condition_variable::wait_for() with our passed duration would be
+        // wrong! We'd keep pushing the timeout time farther and farther into
+        // the future. This way, we establish a definite timeout time and
+        // stick to it.
+        return wait_until(std::chrono::steady_clock::now() + timeout_duration, pred);
+    }
+
+    /**
+     * This wait_for() overload accepts F32Milliseconds as the duration. Any
+     * duration unit defined in llunits.h is implicitly convertible to
+     * F32Milliseconds. The semantics of this method are the same as the
+     * generic wait_for() method.
+     */
+    template <typename Pred>
+    bool wait_for(F32Milliseconds timeout_duration, Pred pred)
+    {
+        return wait_for(convert(timeout_duration), pred);
+    }
+
+protected:
+    // convert F32Milliseconds to a chrono::duration
+    auto convert(F32Milliseconds duration)
+    {
+        // std::chrono::milliseconds doesn't like to be constructed from a
+        // float (F32), rubbing our nose in the thought that
+        // std::chrono::duration::rep is probably integral. Therefore
+        // converting F32Milliseconds to std::chrono::milliseconds would lose
+        // precision. Use std::chrono::microseconds instead. Extract the F32
+        // milliseconds from F32Milliseconds, scale to microseconds, construct
+        // std::chrono::microseconds from that value.
+        return std::chrono::microseconds{ std::chrono::microseconds::rep(duration.value() * 1000) };
+    }
+
+private:
+    /**
+     * Pass wait_until() a chrono::time_point, indicating the time at which we
+     * should stop waiting, and a predicate accepting (const DATA&), returning
+     * bool. The predicate returns true when the condition for which it is
+     * waiting has been satisfied, presumably determined by examining the
+     * referenced DATA. wait_until() locks the mutex and, until the predicate
+     * returns true, calls wait_until() on the condition_variable.
+     * wait_until() returns false if condition_variable::wait_until() timed
+     * out without the predicate returning true.
+     *
+     * Originally this class and its subclasses published wait_until() methods
+     * corresponding to each wait_for() method. But that raised all sorts of
+     * fascinating questions about the time zone of the passed time_point:
+     * local time? server time? UTC? The bottom line is that for LLCond
+     * timeout purposes, we really shouldn't have to care -- timeout duration
+     * is all we need. This private method remains because it's the simplest
+     * way to support iteratively waiting across spurious wakeups while
+     * honoring a fixed timeout.
+     */
+    template <typename Clock, typename Duration, typename Pred>
+    bool wait_until(const std::chrono::time_point<Clock, Duration>& timeout_time, Pred pred)
+    {
+        LLCoros::LockType lk(mMutex);
+        // We advise the caller to pass a predicate accepting (const DATA&).
+        // But what if they instead pass a predicate accepting non-const
+        // (DATA&)? Such a predicate could modify mData, which would be Bad.
+        // Forbid that.
+        while (! pred(const_cast<const value_type&>(mData)))
+        {
+            if (LLCoros::cv_status::timeout == mCond.wait_until(lk, timeout_time))
+            {
+                // It's possible that wait_until() timed out AND the predicate
+                // became true more or less simultaneously. Even though
+                // wait_until() timed out, check the predicate one more time.
+                return pred(const_cast<const value_type&>(mData));
+            }
+        }
+        return true;
+    }
+};
+
+template <typename DATA>
+class LLScalarCond: public LLCond<DATA>
+{
+    using super = LLCond<DATA>;
+
+public:
+    using typename super::value_type;
+    using super::get;
+    using super::wait;
+    using super::wait_for;
+
+    /// LLScalarCond can be explicitly initialized with a specific value for
+    /// mData if desired.
+    LLScalarCond(const value_type& init=value_type()):
+        super(init)
+    {}
+
+    /// Pass set_one() a new value to which to update mData. set_one() will
+    /// lock the mutex, update mData and then call notify_one() on the
+    /// condition_variable.
+    void set_one(const value_type& value)
+    {
+        super::update_one([&value](value_type& data){ data = value; });
+    }
+
+    /// Pass set_all() a new value to which to update mData. set_all() will
+    /// lock the mutex, update mData and then call notify_all() on the
+    /// condition_variable.
+    void set_all(const value_type& value)
+    {
+        super::update_all([&value](value_type& data){ data = value; });
+    }
+
+    /**
+     * Pass wait_equal() a value for which to wait. wait_equal() locks the
+     * mutex and, until the stored DATA equals that value, calls wait() on the
+     * condition_variable.
+     */
+    void wait_equal(const value_type& value)
+    {
+        super::wait([&value](const value_type& data){ return (data == value); });
+    }
+
+    /**
+     * Pass wait_for_equal() a chrono::duration, indicating how long we're
+     * willing to wait, and a value for which to wait. wait_for_equal() locks
+     * the mutex and, until the stored DATA equals that value, calls
+     * wait_for() on the condition_variable. wait_for_equal() returns false if
+     * condition_variable::wait_for() timed out without the stored DATA being
+     * equal to the passed value.
+     */
+    template <typename Rep, typename Period>
+    bool wait_for_equal(const std::chrono::duration<Rep, Period>& timeout_duration,
+                        const value_type& value)
+    {
+        return super::wait_for(timeout_duration,
+                               [&value](const value_type& data){ return (data == value); });
+    }
+
+    /**
+     * This wait_for_equal() overload accepts F32Milliseconds as the duration.
+     * Any duration unit defined in llunits.h is implicitly convertible to
+     * F32Milliseconds. The semantics of this method are the same as the
+     * generic wait_for_equal() method.
+     */
+    bool wait_for_equal(F32Milliseconds timeout_duration, const value_type& value)
+    {
+        return wait_for_equal(super::convert(timeout_duration), value);
+    }
+
+    /**
+     * Pass wait_unequal() a value from which to move away. wait_unequal()
+     * locks the mutex and, until the stored DATA no longer equals that value,
+     * calls wait() on the condition_variable.
+     */
+    void wait_unequal(const value_type& value)
+    {
+        super::wait([&value](const value_type& data){ return (data != value); });
+    }
+
+    /**
+     * Pass wait_for_unequal() a chrono::duration, indicating how long we're
+     * willing to wait, and a value from which to move away.
+     * wait_for_unequal() locks the mutex and, until the stored DATA no longer
+     * equals that value, calls wait_for() on the condition_variable.
+     * wait_for_unequal() returns false if condition_variable::wait_for()
+     * timed out with the stored DATA still being equal to the passed value.
+     */
+    template <typename Rep, typename Period>
+    bool wait_for_unequal(const std::chrono::duration<Rep, Period>& timeout_duration,
+                          const value_type& value)
+    {
+        return super::wait_for(timeout_duration,
+                               [&value](const value_type& data){ return (data != value); });
+    }
+
+    /**
+     * This wait_for_unequal() overload accepts F32Milliseconds as the duration.
+     * Any duration unit defined in llunits.h is implicitly convertible to
+     * F32Milliseconds. The semantics of this method are the same as the
+     * generic wait_for_unequal() method.
+     */
+    bool wait_for_unequal(F32Milliseconds timeout_duration, const value_type& value)
+    {
+        return wait_for_unequal(super::convert(timeout_duration), value);
+    }
+
+protected:
+    using super::convert;
+};
+
+/// Using bool as LLScalarCond's DATA seems like a particularly useful case
+using LLBoolCond = LLScalarCond<bool>;
+
+/// LLOneShotCond -- init false, set (and wait for) true
+class LLOneShotCond: public LLBoolCond
+{
+    using super = LLBoolCond;
+
+public:
+    using typename super::value_type;
+    using super::get;
+    using super::wait;
+    using super::wait_for;
+    using super::wait_equal;
+    using super::wait_for_equal;
+    using super::wait_unequal;
+    using super::wait_for_unequal;
+
+    /// The bool stored in LLOneShotCond is initially false
+    LLOneShotCond(): super(false) {}
+
+    /// LLOneShotCond assumes that nullary set_one() means to set its bool true
+    void set_one(bool value=true)
+    {
+        super::set_one(value);
+    }
+
+    /// LLOneShotCond assumes that nullary set_all() means to set its bool true
+    void set_all(bool value=true)
+    {
+        super::set_all(value);
+    }
+
+    /**
+     * wait() locks the mutex and, until the stored bool is true, calls wait()
+     * on the condition_variable.
+     */
+    void wait()
+    {
+        super::wait_unequal(false);
+    }
+
+    /**
+     * Pass wait_for() a chrono::duration, indicating how long we're willing
+     * to wait. wait_for() locks the mutex and, until the stored bool is true,
+     * calls wait_for() on the condition_variable. wait_for() returns false if
+     * condition_variable::wait_for() timed out without the stored bool being
+     * true.
+     */
+    template <typename Rep, typename Period>
+    bool wait_for(const std::chrono::duration<Rep, Period>& timeout_duration)
+    {
+        return super::wait_for_unequal(timeout_duration, false);
+    }
+
+    /**
+     * This wait_for() overload accepts F32Milliseconds as the duration.
+     * Any duration unit defined in llunits.h is implicitly convertible to
+     * F32Milliseconds. The semantics of this method are the same as the
+     * generic wait_for() method.
+     */
+    bool wait_for(F32Milliseconds timeout_duration)
+    {
+        return wait_for(super::convert(timeout_duration));
+    }
+};
+
+#endif /* ! defined(LL_LLCOND_H) */
diff --git a/indra/llcommon/llcoro_get_id.cpp b/indra/llcommon/llcoro_get_id.cpp
deleted file mode 100644
index 24ed1fe0c9..0000000000
--- a/indra/llcommon/llcoro_get_id.cpp
+++ /dev/null
@@ -1,32 +0,0 @@
-/**
- * @file   llcoro_get_id.cpp
- * @author Nat Goodspeed
- * @date   2016-09-03
- * @brief  Implementation for llcoro_get_id.
- * 
- * $LicenseInfo:firstyear=2016&license=viewerlgpl$
- * Copyright (c) 2016, Linden Research, Inc.
- * $/LicenseInfo$
- */
-
-// Precompiled header
-#include "linden_common.h"
-// associated header
-#include "llcoro_get_id.h"
-// STL headers
-// std headers
-// external library headers
-// other Linden headers
-#include "llcoros.h"
-
-namespace llcoro
-{
-
-id get_id()
-{
-    // An instance of Current can convert to LLCoros::CoroData*, which can
-    // implicitly convert to void*, which is an llcoro::id.
-    return LLCoros::Current();
-}
-
-} // llcoro
diff --git a/indra/llcommon/llcoro_get_id.h b/indra/llcommon/llcoro_get_id.h
deleted file mode 100644
index 4c1dca6f19..0000000000
--- a/indra/llcommon/llcoro_get_id.h
+++ /dev/null
@@ -1,30 +0,0 @@
-/**
- * @file   llcoro_get_id.h
- * @author Nat Goodspeed
- * @date   2016-09-03
- * @brief  Supplement the functionality in llcoro.h.
- *
- *         This is broken out as a separate header file to resolve
- *         circularity: LLCoros isa LLSingleton, yet LLSingleton machinery
- *         requires llcoro::get_id().
- *
- *         Be very suspicious of anyone else #including this header.
- * 
- * $LicenseInfo:firstyear=2016&license=viewerlgpl$
- * Copyright (c) 2016, Linden Research, Inc.
- * $/LicenseInfo$
- */
-
-#if ! defined(LL_LLCORO_GET_ID_H)
-#define LL_LLCORO_GET_ID_H
-
-namespace llcoro
-{
-
-/// Get an opaque, distinct token for the running coroutine (or main).
-typedef void* id;
-id get_id();
-
-} // llcoro
-
-#endif /* ! defined(LL_LLCORO_GET_ID_H) */
diff --git a/indra/llcommon/llcoros.cpp b/indra/llcommon/llcoros.cpp
index cc775775bf..262929006d 100644
--- a/indra/llcommon/llcoros.cpp
+++ b/indra/llcommon/llcoros.cpp
@@ -26,15 +26,30 @@
  * $/LicenseInfo$
  */
 
+#include "llwin32headers.h"
+
 // Precompiled header
 #include "linden_common.h"
 // associated header
 #include "llcoros.h"
 // STL headers
 // std headers
+#include <atomic>
 // external library headers
 #include <boost/bind.hpp>
+#include <boost/fiber/fiber.hpp>
+#ifndef BOOST_DISABLE_ASSERTS
+#define UNDO_BOOST_DISABLE_ASSERTS
+// with Boost 1.65.1, needed for Mac with this specific header
+#define BOOST_DISABLE_ASSERTS
+#endif
+#include <boost/fiber/protected_fixedsize_stack.hpp>
+#ifdef UNDO_BOOST_DISABLE_ASSERTS
+#undef UNDO_BOOST_DISABLE_ASSERTS
+#undef BOOST_DISABLE_ASSERTS
+#endif
 // other Linden headers
+#include "llapp.h"
 #include "lltimer.h"
 #include "llevents.h"
 #include "llerror.h"
@@ -45,85 +60,43 @@
 #include <excpt.h>
 #endif
 
-namespace {
-void no_op() {}
-} // anonymous namespace
-
-// Do nothing, when we need nothing done. This is a static member of LLCoros
-// because CoroData is a private nested class.
-void LLCoros::no_cleanup(CoroData*) {}
-
-// CoroData for the currently-running coroutine. Use a thread_specific_ptr
-// because each thread potentially has its own distinct pool of coroutines.
-LLCoros::Current::Current()
+// static
+LLCoros::CoroData& LLCoros::get_CoroData(const std::string& caller)
 {
-    // Use a function-static instance so this thread_specific_ptr is
-    // instantiated on demand. Since we happen to know it's consumed by
-    // LLSingleton, this is likely to happen before the runtime has finished
-    // initializing module-static data. For the same reason, we can't package
-    // this pointer in an LLSingleton.
-
-    // This thread_specific_ptr does NOT own the CoroData object! That's owned
-    // by LLCoros::mCoros. It merely identifies it. For this reason we
-    // instantiate it with a no-op cleanup function.
-    static boost::thread_specific_ptr<LLCoros::CoroData> sCurrent(LLCoros::no_cleanup);
-
-    // If this is the first time we're accessing sCurrent for the running
-    // thread, its get() will be NULL. This could be a problem, in that
-    // llcoro::get_id() would return the same (NULL) token value for the "main
-    // coroutine" in every thread, whereas what we really want is a distinct
-    // value for every distinct stack in the process. So if get() is NULL,
-    // give it a heap CoroData: this ensures that llcoro::get_id() will return
-    // distinct values.
-    // This tactic is "leaky": sCurrent explicitly does not destroy any
-    // CoroData to which it points, and we do NOT enter these "main coroutine"
-    // CoroData instances in the LLCoros::mCoros map. They are dummy entries,
-    // and they will leak at process shutdown: one CoroData per thread.
-    if (! sCurrent.get())
+    CoroData* current{ nullptr };
+    // be careful about attempted accesses in the final throes of app shutdown
+    if (! wasDeleted())
     {
-        // It's tempting to provide a distinct name for each thread's "main
-        // coroutine." But as getName() has always returned the empty string
-        // to mean "not in a coroutine," empty string should suffice here --
-        // and truthfully the additional (thread-safe!) machinery to ensure
-        // uniqueness just doesn't feel worth the trouble.
-        // We use a no-op callable and a minimal stack size because, although
-        // CoroData's constructor in fact initializes its mCoro with a
-        // coroutine with that stack size, no one ever actually enters it by
-        // calling mCoro().
-        sCurrent.reset(new CoroData(0,  // no prev
-                                    "", // not a named coroutine
-                                    no_op,  // no-op callable
-                                    1024)); // stacksize moot
+        current = instance().mCurrent.get();
+    }
+    // For the main() coroutine, the one NOT explicitly launched by launch(),
+    // we never explicitly set mCurrent. Use a static CoroData instance with
+    // canonical values.
+    if (! current)
+    {
+        static std::atomic<int> which_thread(0);
+        // Use alternate CoroData constructor.
+        static thread_local CoroData sMain(which_thread++);
+        // We need not reset() the local_ptr to this instance; we'll simply
+        // find it again every time we discover that current is null.
+        current = &sMain;
     }
-
-    mCurrent = &sCurrent;
-}
-
-//static
-LLCoros::CoroData& LLCoros::get_CoroData(const std::string& caller)
-{
-    CoroData* current = Current();
-    // With the dummy CoroData set in LLCoros::Current::Current(), this
-    // pointer should never be NULL.
-    llassert_always(current);
     return *current;
 }
 
 //static
-LLCoros::coro::self& LLCoros::get_self()
+LLCoros::coro::id LLCoros::get_self()
 {
-    CoroData& current = get_CoroData("get_self()");
-    if (! current.mSelf)
-    {
-        LL_ERRS("LLCoros") << "Calling get_self() from non-coroutine context!" << LL_ENDL;
-    }
-    return *current.mSelf;
+    return boost::this_fiber::get_id();
 }
 
 //static
 void LLCoros::set_consuming(bool consuming)
 {
-    get_CoroData("set_consuming()").mConsuming = consuming;
+    CoroData& data(get_CoroData("set_consuming()"));
+    // DO NOT call this on the main() coroutine.
+    llassert_always(! data.mName.empty());
+    data.mConsuming = consuming;
 }
 
 //static
@@ -132,89 +105,59 @@ bool LLCoros::get_consuming()
     return get_CoroData("get_consuming()").mConsuming;
 }
 
-llcoro::Suspending::Suspending()
+// static
+void LLCoros::setStatus(const std::string& status)
 {
-    LLCoros::Current current;
-    // Remember currently-running coroutine: we're about to suspend it.
-    mSuspended = current;
-    // Revert Current to the value it had at the moment we last switched
-    // into this coroutine.
-    current.reset(mSuspended->mPrev);
+    get_CoroData("setStatus()").mStatus = status;
 }
 
-llcoro::Suspending::~Suspending()
+// static
+std::string LLCoros::getStatus()
 {
-    LLCoros::Current current;
-    // Okay, we're back, update our mPrev
-    mSuspended->mPrev = current;
-    // and reinstate our Current.
-    current.reset(mSuspended);
+    return get_CoroData("getStatus()").mStatus;
 }
 
 LLCoros::LLCoros():
     // MAINT-2724: default coroutine stack size too small on Windows.
     // Previously we used
     // boost::context::guarded_stack_allocator::default_stacksize();
-    // empirically this is 64KB on Windows and Linux. Try quadrupling.
+    // empirically this is insufficient.
 #if ADDRESS_SIZE == 64
-    mStackSize(512*1024)
+    mStackSize(512*1024),
 #else
-    mStackSize(256*1024)
+    mStackSize(256*1024),
 #endif
+    // mCurrent does NOT own the current CoroData instance -- it simply
+    // points to it. So initialize it with a no-op deleter.
+    mCurrent{ [](CoroData*){} }
 {
-    // Register our cleanup() method for "mainloop" ticks
-    LLEventPumps::instance().obtain("mainloop").listen(
-        "LLCoros", boost::bind(&LLCoros::cleanup, this, _1));
 }
 
-bool LLCoros::cleanup(const LLSD&)
+LLCoros::~LLCoros()
 {
-    static std::string previousName;
-    static int previousCount = 0;
-    // Walk the mCoros map, checking and removing completed coroutines.
-    for (CoroMap::iterator mi(mCoros.begin()), mend(mCoros.end()); mi != mend; )
+    printActiveCoroutines("at entry to ~LLCoros()");
+    // Other LLApp status-change listeners do things like close
+    // work queues and inject the Stop exception into pending
+    // promises, to force coroutines waiting on those things to
+    // notice and terminate. The only problem is that by the time
+    // LLApp sets "quitting" status, the main loop has stopped
+    // pumping the fiber scheduler with yield() calls. A waiting
+    // coroutine still might not wake up until after resources on
+    // which it depends have been freed. Pump it a few times
+    // ourselves. Of course, stop pumping as soon as the last of
+    // the coroutines has terminated.
+    for (size_t count = 0; count < 10 && CoroData::instanceCount() > 0; ++count)
     {
-        // Has this coroutine exited (normal return, exception, exit() call)
-        // since last tick?
-        if (mi->second->mCoro.exited())
-        {
-            if (previousName != mi->first)
-            { 
-                previousName = mi->first;
-                previousCount = 1;
-            }
-            else
-            {
-                ++previousCount;
-            }
-               
-            if ((previousCount < 5) || !(previousCount % 50))
-            {
-                if (previousCount < 5)
-                    LL_DEBUGS("LLCoros") << "LLCoros: cleaning up coroutine " << mi->first << LL_ENDL;
-                else
-                    LL_DEBUGS("LLCoros") << "LLCoros: cleaning up coroutine " << mi->first << "("<< previousCount << ")" << LL_ENDL;
-
-            }
-            // The erase() call will invalidate its passed iterator value --
-            // so increment mi FIRST -- but pass its original value to
-            // erase(). This is what postincrement is all about.
-            mCoros.erase(mi++);
-        }
-        else
-        {
-            // Still live, just skip this entry as if incrementing at the top
-            // of the loop as usual.
-            ++mi;
-        }
+        // don't use llcoro::suspend() because that module depends
+        // on this one
+        boost::this_fiber::yield();
     }
-    return false;
+    printActiveCoroutines("after pumping");
 }
 
 std::string LLCoros::generateDistinctName(const std::string& prefix) const
 {
-    static std::string previousName;
-    static int previousCount = 0;
+    static int unique = 0;
 
     // Allowing empty name would make getName()'s not-found return ambiguous.
     if (prefix.empty())
@@ -225,37 +168,15 @@ std::string LLCoros::generateDistinctName(const std::string& prefix) const
     // If the specified name isn't already in the map, just use that.
     std::string name(prefix);
 
-    // Find the lowest numeric suffix that doesn't collide with an existing
-    // entry. Start with 2 just to make it more intuitive for any interested
-    // parties: e.g. "joe", "joe2", "joe3"...
-    for (int i = 2; ; name = STRINGIZE(prefix << i++))
+    // Until we find an unused name, append a numeric suffix for uniqueness.
+    while (CoroData::getInstance(name))
     {
-        if (mCoros.find(name) == mCoros.end())
-        {
-            if (previousName != name)
-            {
-                previousName = name;
-                previousCount = 1;
-            }
-            else
-            {
-                ++previousCount;
-            }
-
-            if ((previousCount < 5) || !(previousCount % 50))
-            {
-                if (previousCount < 5)
-                    LL_DEBUGS("LLCoros") << "LLCoros: launching coroutine " << name << LL_ENDL;
-                else
-                    LL_DEBUGS("LLCoros") << "LLCoros: launching coroutine " << name << "(" << previousCount << ")" << LL_ENDL;
-
-            }
-
-            return name;
-        }
+        name = STRINGIZE(prefix << unique++);
     }
+    return name;
 }
 
+/*==========================================================================*|
 bool LLCoros::kill(const std::string& name)
 {
     CoroMap::iterator found = mCoros.find(name);
@@ -269,10 +190,19 @@ bool LLCoros::kill(const std::string& name)
     mCoros.erase(found);
     return true;
 }
+|*==========================================================================*/
 
-std::string LLCoros::getName() const
+//static
+std::string LLCoros::getName()
 {
-    return Current()->mName;
+    return get_CoroData("getName()").mName;
+}
+
+//static
+std::string LLCoros::logname()
+{
+    LLCoros::CoroData& data(get_CoroData("logname()"));
+    return data.mName.empty()? data.getKey() : data.mName;
 }
 
 void LLCoros::setStackSize(S32 stacksize)
@@ -281,25 +211,46 @@ void LLCoros::setStackSize(S32 stacksize)
     mStackSize = stacksize;
 }
 
-void LLCoros::printActiveCoroutines()
+void LLCoros::printActiveCoroutines(const std::string& when)
 {
-    LL_INFOS("LLCoros") << "Number of active coroutines: " << (S32)mCoros.size() << LL_ENDL;
-    if (mCoros.size() > 0)
+    LL_INFOS("LLCoros") << "Number of active coroutines " << when
+                        << ": " << CoroData::instanceCount() << LL_ENDL;
+    if (CoroData::instanceCount() > 0)
     {
         LL_INFOS("LLCoros") << "-------------- List of active coroutines ------------";
-        CoroMap::iterator iter;
-        CoroMap::iterator end = mCoros.end();
         F64 time = LLTimer::getTotalSeconds();
-        for (iter = mCoros.begin(); iter != end; iter++)
+        for (auto& cd : CoroData::instance_snapshot())
         {
-            F64 life_time = time - iter->second->mCreationTime;
-            LL_CONT << LL_NEWLINE << "Name: " << iter->first << " life: " << life_time;
+            F64 life_time = time - cd.mCreationTime;
+            LL_CONT << LL_NEWLINE
+                    << cd.getKey() << ' ' << cd.mStatus << " life: " << life_time;
         }
         LL_CONT << LL_ENDL;
         LL_INFOS("LLCoros") << "-----------------------------------------------------" << LL_ENDL;
     }
 }
 
+std::string LLCoros::launch(const std::string& prefix, const callable_t& callable)
+{
+    std::string name(generateDistinctName(prefix));
+    // 'dispatch' means: enter the new fiber immediately, returning here only
+    // when the fiber yields for whatever reason.
+    // std::allocator_arg is a flag to indicate that the following argument is
+    // a StackAllocator.
+    // protected_fixedsize_stack sets a guard page past the end of the new
+    // stack so that stack underflow will result in an access violation
+    // instead of weird, subtle, possibly undiagnosed memory stomps.
+    boost::fibers::fiber newCoro(boost::fibers::launch::dispatch,
+                                 std::allocator_arg,
+                                 boost::fibers::protected_fixedsize_stack(mStackSize),
+                                 [this, &name, &callable](){ toplevel(name, callable); });
+    // You have two choices with a fiber instance: you can join() it or you
+    // can detach() it. If you try to destroy the instance before doing
+    // either, the program silently terminates. We don't need this handle.
+    newCoro.detach();
+    return name;
+}
+
 #if LL_WINDOWS
 
 static const U32 STATUS_MSC_EXCEPTION = 0xE06D7363; // compiler specific
@@ -337,13 +288,16 @@ void LLCoros::winlevel(const callable_t& callable)
 
 #endif
 
-// Top-level wrapper around caller's coroutine callable. This function accepts
-// the coroutine library's implicit coro::self& parameter and saves it, but
-// does not pass it down to the caller's callable.
-void LLCoros::toplevel(coro::self& self, CoroData* data, const callable_t& callable)
+// Top-level wrapper around caller's coroutine callable.
+// Normally we like to pass strings and such by const reference -- but in this
+// case, we WANT to copy both the name and the callable to our local stack!
+void LLCoros::toplevel(std::string name, callable_t callable)
 {
-    // capture the 'self' param in CoroData
-    data->mSelf = &self;
+    // keep the CoroData on this top-level function's stack frame
+    CoroData corodata(name);
+    // set it as current
+    mCurrent.reset(&corodata);
+
     // run the code the caller actually wants in the coroutine
     try
     {
@@ -353,75 +307,69 @@ void LLCoros::toplevel(coro::self& self, CoroData* data, const callable_t& calla
         callable();
 #endif
     }
+    catch (const Stop& exc)
+    {
+        LL_INFOS("LLCoros") << "coroutine " << name << " terminating because "
+                            << exc.what() << LL_ENDL;
+    }
     catch (const LLContinueError&)
     {
         // Any uncaught exception derived from LLContinueError will be caught
         // here and logged. This coroutine will terminate but the rest of the
         // viewer will carry on.
-        LOG_UNHANDLED_EXCEPTION(STRINGIZE("coroutine " << data->mName));
+        LOG_UNHANDLED_EXCEPTION(STRINGIZE("coroutine " << name));
     }
     catch (...)
     {
         // Any OTHER kind of uncaught exception will cause the viewer to
         // crash, hopefully informatively.
-        CRASH_ON_UNHANDLED_EXCEPTION(STRINGIZE("coroutine " << data->mName));
+        CRASH_ON_UNHANDLED_EXCEPTION(STRINGIZE("coroutine " << name));
     }
-    // This cleanup isn't perfectly symmetrical with the way we initially set
-    // data->mPrev, but this is our last chance to reset Current.
-    Current().reset(data->mPrev);
 }
 
-/*****************************************************************************
-*   MUST BE LAST
-*****************************************************************************/
-// Turn off MSVC optimizations for just LLCoros::launch() -- see
-// DEV-32777. But MSVC doesn't support push/pop for optimization flags as it
-// does for warning suppression, and we really don't want to force
-// optimization ON for other code even in Debug or RelWithDebInfo builds.
-
-#if LL_MSVC
-// work around broken optimizations
-#pragma warning(disable: 4748)
-#pragma warning(disable: 4355) // 'this' used in initializer list: yes, intentionally
-#pragma optimize("", off)
-#endif // LL_MSVC
+//static
+void LLCoros::checkStop()
+{
+    if (wasDeleted())
+    {
+        LLTHROW(Shutdown("LLCoros was deleted"));
+    }
+    // do this AFTER the check above, because getName() depends on
+    // get_CoroData(), which depends on the local_ptr in our instance().
+    if (getName().empty())
+    {
+        // Our Stop exception and its subclasses are intended to stop loitering
+        // coroutines. Don't throw it from the main coroutine.
+        return;
+    }
+    if (LLApp::isStopped())
+    {
+        LLTHROW(Stopped("viewer is stopped"));
+    }
+    if (! LLApp::isRunning())
+    {
+        LLTHROW(Stopping("viewer is stopping"));
+    }
+}
 
-LLCoros::CoroData::CoroData(CoroData* prev, const std::string& name,
-                            const callable_t& callable, S32 stacksize):
-    mPrev(prev),
+LLCoros::CoroData::CoroData(const std::string& name):
+    LLInstanceTracker<CoroData, std::string>(name),
     mName(name),
-    // Wrap the caller's callable in our toplevel() function so we can manage
-    // Current appropriately at startup and shutdown of each coroutine.
-    mCoro(boost::bind(toplevel, _1, this, callable), stacksize),
     // don't consume events unless specifically directed
     mConsuming(false),
-    mSelf(0),
     mCreationTime(LLTimer::getTotalSeconds())
 {
 }
 
-std::string LLCoros::launch(const std::string& prefix, const callable_t& callable)
+LLCoros::CoroData::CoroData(int n):
+    // This constructor is used for the thread_local instance belonging to the
+    // default coroutine on each thread. We must give each one a different
+    // LLInstanceTracker key because LLInstanceTracker's map spans all
+    // threads, but we want the default coroutine on each thread to have the
+    // empty string as its visible name because some consumers test for that.
+    LLInstanceTracker<CoroData, std::string>("main" + stringize(n)),
+    mName(),
+    mConsuming(false),
+    mCreationTime(LLTimer::getTotalSeconds())
 {
-    std::string name(generateDistinctName(prefix));
-    Current current;
-    // pass the current value of Current as previous context
-    CoroData* newCoro = new(std::nothrow) CoroData(current, name, callable, mStackSize);
-    if (newCoro == NULL)
-    {
-        // Out of memory?
-        printActiveCoroutines();
-        LL_ERRS("LLCoros") << "Failed to start coroutine: " << name << " Stacksize: " << mStackSize << " Total coroutines: " << mCoros.size() << LL_ENDL;
-    }
-    // Store it in our pointer map
-    mCoros.insert(name, newCoro);
-    // also set it as current
-    current.reset(newCoro);
-    /* Run the coroutine until its first wait, then return here */
-    (newCoro->mCoro)(std::nothrow);
-    return name;
 }
-
-#if LL_MSVC
-// reenable optimizations
-#pragma optimize("", on)
-#endif // LL_MSVC
diff --git a/indra/llcommon/llcoros.h b/indra/llcommon/llcoros.h
index c551413811..38c2356c99 100644
--- a/indra/llcommon/llcoros.h
+++ b/indra/llcommon/llcoros.h
@@ -29,21 +29,26 @@
 #if ! defined(LL_LLCOROS_H)
 #define LL_LLCOROS_H
 
-#include <boost/dcoroutine/coroutine.hpp>
-#include <boost/dcoroutine/future.hpp>
+#include "llexception.h"
+#include <boost/fiber/fss.hpp>
+#include <boost/fiber/future/promise.hpp>
+#include <boost/fiber/future/future.hpp>
+#include "mutex.h"
 #include "llsingleton.h"
-#include <boost/ptr_container/ptr_map.hpp>
+#include "llinstancetracker.h"
 #include <boost/function.hpp>
-#include <boost/thread/tss.hpp>
-#include <boost/noncopyable.hpp>
 #include <string>
-#include <stdexcept>
-#include "llcoro_get_id.h"          // for friend declaration
 
-// forward-declare helper class
-namespace llcoro
-{
-class Suspending;
+// e.g. #include LLCOROS_MUTEX_HEADER
+#define LLCOROS_MUTEX_HEADER   <boost/fiber/mutex.hpp>
+#define LLCOROS_CONDVAR_HEADER <boost/fiber/condition_variable.hpp>
+
+namespace boost {
+    namespace fibers {
+        class mutex;
+        enum class cv_status;
+        class condition_variable;
+    }
 }
 
 /**
@@ -76,19 +81,21 @@ class Suspending;
  * name prefix; from your prefix it generates a distinct name, registers the
  * new coroutine and returns the actual name.
  *
- * The name can be used to kill off the coroutine prematurely, if needed. It
- * can also provide diagnostic info: we can look up the name of the
+ * The name
+ * can provide diagnostic info: we can look up the name of the
  * currently-running coroutine.
- *
- * Finally, the next frame ("mainloop" event) after the coroutine terminates,
- * LLCoros will notice its demise and destroy it.
  */
 class LL_COMMON_API LLCoros: public LLSingleton<LLCoros>
 {
     LLSINGLETON(LLCoros);
+    ~LLCoros();
 public:
-    /// Canonical boost::dcoroutines::coroutine signature we use
-    typedef boost::dcoroutines::coroutine<void()> coro;
+    /// The viewer's use of the term "coroutine" became deeply embedded before
+    /// the industry term "fiber" emerged to distinguish userland threads from
+    /// simpler, more transient kinds of coroutines. Semantically they've
+    /// always been fibers. But at this point in history, we're pretty much
+    /// stuck with the term "coroutine."
+    typedef boost::fibers::fiber coro;
     /// Canonical callable type
     typedef boost::function<void()> callable_t;
 
@@ -119,10 +126,10 @@ public:
      * DEV-32777 comments for an explanation.
      *
      * Pass a nullary callable. It works to directly pass a nullary free
-     * function (or static method); for all other cases use boost::bind(). Of
-     * course, for a non-static class method, the first parameter must be the
-     * class instance. Any other parameters should be passed via the bind()
-     * expression.
+     * function (or static method); for other cases use a lambda expression,
+     * std::bind() or boost::bind(). Of course, for a non-static class method,
+     * the first parameter must be the class instance. Any other parameters
+     * should be passed via the enclosing expression.
      *
      * launch() tweaks the suggested name so it won't collide with any
      * existing coroutine instance, creates the coroutine instance, registers
@@ -138,7 +145,7 @@ public:
      * one prematurely. Returns @c true if the specified name was found and
      * still running at the time.
      */
-    bool kill(const std::string& name);
+//  bool kill(const std::string& name);
 
     /**
      * From within a coroutine, look up the (tweaked) name string by which
@@ -146,16 +153,27 @@ public:
      * (e.g. if the coroutine was launched by hand rather than using
      * LLCoros::launch()).
      */
-    std::string getName() const;
+    static std::string getName();
 
-    /// for delayed initialization
+    /**
+     * This variation returns a name suitable for log messages: the explicit
+     * name for an explicitly-launched coroutine, or "mainN" for the default
+     * coroutine on a thread.
+     */
+    static std::string logname();
+
+    /**
+     * For delayed initialization. To be clear, this will only affect
+     * coroutines launched @em after this point. The underlying facility
+     * provides no way to alter the stack size of any running coroutine.
+     */
     void setStackSize(S32 stacksize);
 
-    /// for delayed initialization
-    void printActiveCoroutines();
+    /// diagnostic
+    void printActiveCoroutines(const std::string& when=std::string());
 
-    /// get the current coro::self& for those who really really care
-    static coro::self& get_self();
+    /// get the current coro::id for those who really really care
+    static coro::id get_self();
 
     /**
      * Most coroutines, most of the time, don't "consume" the events for which
@@ -180,6 +198,7 @@ public:
         {
             set_consuming(consuming);
         }
+        OverrideConsuming(const OverrideConsuming&) = delete;
         ~OverrideConsuming()
         {
             set_consuming(mPrevConsuming);
@@ -189,142 +208,124 @@ public:
         bool mPrevConsuming;
     };
 
+    /// set string coroutine status for diagnostic purposes
+    static void setStatus(const std::string& status);
+    static std::string getStatus();
+
+    /// RAII control of status
+    class TempStatus
+    {
+    public:
+        TempStatus(const std::string& status):
+            mOldStatus(getStatus())
+        {
+            setStatus(status);
+        }
+        TempStatus(const TempStatus&) = delete;
+        ~TempStatus()
+        {
+            setStatus(mOldStatus);
+        }
+
+    private:
+        std::string mOldStatus;
+    };
+
+    /// thrown by checkStop()
+    // It may sound ironic that Stop is derived from LLContinueError, but the
+    // point is that LLContinueError is the category of exception that should
+    // not immediately crash the viewer. Stop and its subclasses are to notify
+    // coroutines that the viewer intends to shut down. The expected response
+    // is to terminate the coroutine, rather than abort the viewer.
+    struct Stop: public LLContinueError
+    {
+        Stop(const std::string& what): LLContinueError(what) {}
+    };
+
+    /// early stages
+    struct Stopping: public Stop
+    {
+        Stopping(const std::string& what): Stop(what) {}
+    };
+
+    /// cleaning up
+    struct Stopped: public Stop
+    {
+        Stopped(const std::string& what): Stop(what) {}
+    };
+
+    /// cleaned up -- not much survives!
+    struct Shutdown: public Stop
+    {
+        Shutdown(const std::string& what): Stop(what) {}
+    };
+
+    /// Call this intermittently if there's a chance your coroutine might
+    /// continue running into application shutdown. Throws Stop if LLCoros has
+    /// been cleaned up.
+    static void checkStop();
+
     /**
-     * Please do NOT directly use boost::dcoroutines::future! It is essential
-     * to maintain the "current" coroutine at every context switch. This
-     * Future wraps the essential boost::dcoroutines::future functionality
-     * with that maintenance.
+     * Aliases for promise and future. An older underlying future implementation
+     * required us to wrap future; that's no longer needed. However -- if it's
+     * important to restore kill() functionality, we might need to provide a
+     * proxy, so continue using the aliases.
      */
     template <typename T>
-    class Future;
+    using Promise = boost::fibers::promise<T>;
+    template <typename T>
+    using Future = boost::fibers::future<T>;
+    template <typename T>
+    static Future<T> getFuture(Promise<T>& promise) { return promise.get_future(); }
+
+    // use mutex, lock, condition_variable suitable for coroutines
+    using Mutex = boost::fibers::mutex;
+    using LockType = std::unique_lock<Mutex>;
+    using cv_status = boost::fibers::cv_status;
+    using ConditionVariable = boost::fibers::condition_variable;
+
+    /// for data local to each running coroutine
+    template <typename T>
+    using local_ptr = boost::fibers::fiber_specific_ptr<T>;
 
 private:
-    friend class llcoro::Suspending;
-    friend llcoro::id llcoro::get_id();
     std::string generateDistinctName(const std::string& prefix) const;
-    bool cleanup(const LLSD&);
+    void toplevel(std::string name, callable_t callable);
     struct CoroData;
-    static void no_cleanup(CoroData*);
 #if LL_WINDOWS
     static void winlevel(const callable_t& callable);
 #endif
-    static void toplevel(coro::self& self, CoroData* data, const callable_t& callable);
     static CoroData& get_CoroData(const std::string& caller);
 
     S32 mStackSize;
 
     // coroutine-local storage, as it were: one per coro we track
-    struct CoroData
+    struct CoroData: public LLInstanceTracker<CoroData, std::string>
     {
-        CoroData(CoroData* prev, const std::string& name,
-                 const callable_t& callable, S32 stacksize);
-
-        // The boost::dcoroutines library supports asymmetric coroutines. Every
-        // time we context switch out of a coroutine, we pass control to the
-        // previously-active one (or to the non-coroutine stack owned by the
-        // thread). So our management of the "current" coroutine must be able to
-        // restore the previous value when we're about to switch away.
-        CoroData* mPrev;
+        CoroData(const std::string& name);
+        CoroData(int n);
+
         // tweaked name of the current coroutine
         const std::string mName;
-        // the actual coroutine instance
-        LLCoros::coro mCoro;
         // set_consuming() state
         bool mConsuming;
-        // When the dcoroutine library calls a top-level callable, it implicitly
-        // passes coro::self& as the first parameter. All our consumer code used
-        // to explicitly pass coro::self& down through all levels of call stack,
-        // because at the leaf level we need it for context-switching. But since
-        // coroutines are based on cooperative switching, we can cause the
-        // top-level entry point to stash a pointer to the currently-running
-        // coroutine, and manage it appropriately as we switch out and back in.
-        // That eliminates the need to pass it as an explicit parameter down
-        // through every level, which is unfortunately viral in nature. Finding it
-        // implicitly rather than explicitly allows minor maintenance in which a
-        // leaf-level function adds a new async I/O call that suspends the calling
-        // coroutine, WITHOUT having to propagate coro::self& through every
-        // function signature down to that point -- and of course through every
-        // other caller of every such function.
-        LLCoros::coro::self* mSelf;
+        // setStatus() state
+        std::string mStatus;
         F64 mCreationTime; // since epoch
     };
-    typedef boost::ptr_map<std::string, CoroData> CoroMap;
-    CoroMap mCoros;
 
-    // Identify the current coroutine's CoroData. Use a little helper class so
-    // a caller can either use a temporary instance, or instantiate a named
-    // variable and access it multiple times.
-    class Current
-    {
-    public:
-        Current();
-
-        operator LLCoros::CoroData*() { return get(); }
-        LLCoros::CoroData* operator->() { return get(); }
-        LLCoros::CoroData* get() { return mCurrent->get(); }
-        void reset(LLCoros::CoroData* ptr) { mCurrent->reset(ptr); }
-
-    private:
-        boost::thread_specific_ptr<LLCoros::CoroData>* mCurrent;
-    };
+    // Identify the current coroutine's CoroData. This local_ptr isn't static
+    // because it's a member of an LLSingleton, and we rely on it being
+    // cleaned up in proper dependency order.
+    local_ptr<CoroData> mCurrent;
 };
 
 namespace llcoro
 {
 
-/// Instantiate one of these in a block surrounding any leaf point when
-/// control literally switches away from this coroutine.
-class Suspending: boost::noncopyable
-{
-public:
-    Suspending();
-    ~Suspending();
-
-private:
-    LLCoros::CoroData* mSuspended;
-};
-
-} // namespace llcoro
-
-template <typename T>
-class LLCoros::Future
-{
-    typedef boost::dcoroutines::future<T> dfuture;
-
-public:
-    Future():
-        mFuture(get_self())
-    {}
-
-    typedef typename boost::dcoroutines::make_callback_result<dfuture>::type callback_t;
-
-    callback_t make_callback()
-    {
-        return boost::dcoroutines::make_callback(mFuture);
-    }
-
-#ifndef LL_LINUX
-    explicit
-#endif
-    operator bool() const
-    {
-        return bool(mFuture);
-    }
-
-    bool operator!() const
-    {
-        return ! mFuture;
-    }
+inline
+std::string logname() { return LLCoros::logname(); }
 
-    T get()
-    {
-        // instantiate Suspending to manage the "current" coroutine
-        llcoro::Suspending suspended;
-        return *mFuture;
-    }
-
-private:
-    dfuture mFuture;
-};
+} // llcoro
 
 #endif /* ! defined(LL_LLCOROS_H) */
diff --git a/indra/llcommon/llerror.cpp b/indra/llcommon/llerror.cpp
index b46f49ba34..411412c883 100644
--- a/indra/llcommon/llerror.cpp
+++ b/indra/llcommon/llerror.cpp
@@ -39,6 +39,9 @@
 #if !LL_WINDOWS
 # include <syslog.h>
 # include <unistd.h>
+# include <sys/stat.h>
+#else
+# include <io.h>
 #endif // !LL_WINDOWS
 #include <vector>
 #include "string.h"
@@ -53,6 +56,13 @@
 #include "llstl.h"
 #include "lltimer.h"
 
+// On Mac, got:
+// #error "Boost.Stacktrace requires `_Unwind_Backtrace` function. Define
+// `_GNU_SOURCE` macro or `BOOST_STACKTRACE_GNU_SOURCE_NOT_REQUIRED` if
+// _Unwind_Backtrace is available without `_GNU_SOURCE`."
+#define BOOST_STACKTRACE_GNU_SOURCE_NOT_REQUIRED
+#include <boost/stacktrace.hpp>
+
 namespace {
 #if LL_WINDOWS
 	void debugger_print(const std::string& s)
@@ -118,27 +128,28 @@ namespace {
 	class RecordToFile : public LLError::Recorder
 	{
 	public:
-		RecordToFile(const std::string& filename)
+		RecordToFile(const std::string& filename):
+			mName(filename)
 		{
 			mFile.open(filename.c_str(), std::ios_base::out | std::ios_base::app);
 			if (!mFile)
 			{
 				LL_INFOS() << "Error setting log file to " << filename << LL_ENDL;
 			}
-            else
-            {
-                if (!LLError::getAlwaysFlush())
-                {
-                    mFile.sync_with_stdio(false);
-                }
-            }
+			else
+			{
+				if (!LLError::getAlwaysFlush())
+				{
+					mFile.sync_with_stdio(false);
+				}
+			}
 		}
-		
+
 		~RecordToFile()
 		{
 			mFile.close();
 		}
-		
+
         virtual bool enabled() override
         {
 #ifdef LL_RELEASE_FOR_DOWNLOAD
@@ -148,11 +159,13 @@ namespace {
 #endif
         }
         
-		bool okay() { return mFile.good(); }
-		
-		virtual void recordMessage(LLError::ELevel level,
-									const std::string& message) override
-		{
+        bool okay() const { return mFile.good(); }
+
+        std::string getFilename() const { return mName; }
+
+        virtual void recordMessage(LLError::ELevel level,
+                                    const std::string& message) override
+        {
             if (LLError::getAlwaysFlush())
             {
                 mFile << message << std::endl;
@@ -161,9 +174,10 @@ namespace {
             {
                 mFile << message << "\n";
             }
-		}
-	
+        }
+
 	private:
+		const std::string mName;
 		llofstream mFile;
 	};
 	
@@ -171,7 +185,7 @@ namespace {
 	class RecordToStderr : public LLError::Recorder
 	{
 	public:
-		RecordToStderr(bool timestamp) : mUseANSI(ANSI_PROBE) 
+		RecordToStderr(bool timestamp) : mUseANSI(checkANSI()) 
 		{
             this->showMultiline(true);
 		}
@@ -193,14 +207,12 @@ namespace {
 
 		virtual void recordMessage(LLError::ELevel level,
 					   const std::string& message) override
-		{            
+		{
             static std::string s_ansi_error = createANSI("31"); // red
             static std::string s_ansi_warn  = createANSI("34"); // blue
             static std::string s_ansi_debug = createANSI("35"); // magenta
 
-            mUseANSI = (ANSI_PROBE == mUseANSI) ? (checkANSI() ? ANSI_YES : ANSI_NO) : mUseANSI;
-
-			if (ANSI_YES == mUseANSI)
+			if (mUseANSI)
 			{
                 writeANSI((level == LLError::LEVEL_ERROR) ? s_ansi_error :
                           (level == LLError::LEVEL_WARN)  ? s_ansi_warn :
@@ -213,12 +225,7 @@ namespace {
 		}
 	
 	private:
-		enum ANSIState 
-		{
-			ANSI_PROBE, 
-			ANSI_YES, 
-			ANSI_NO
-		}					mUseANSI;
+		bool mUseANSI;
 
         LL_FORCE_INLINE void writeANSI(const std::string& ansi_code, const std::string& message)
 		{
@@ -229,16 +236,13 @@ namespace {
 			fprintf(stderr, "%s%s%s\n%s", s_ansi_bold.c_str(), ansi_code.c_str(), message.c_str(), s_ansi_reset.c_str() );
 		}
 
-		bool checkANSI(void)
+		static bool checkANSI(void)
 		{
-#if LL_LINUX || LL_DARWIN
 			// Check whether it's okay to use ANSI; if stderr is
 			// a tty then we assume yes.  Can be turned off with
 			// the LL_NO_ANSI_COLOR env var.
 			return (0 != isatty(2)) &&
 				(NULL == getenv("LL_NO_ANSI_COLOR"));
-#endif // LL_LINUX
-            return FALSE; // works in a cygwin shell... ;)
 		}
 	};
 
@@ -308,28 +312,35 @@ namespace LLError
 	{
 #ifdef __GNUC__
 		// GCC: type_info::name() returns a mangled class name,st demangle
-        // passing nullptr, 0 forces allocation of a unique buffer we can free
-        // fixing MAINT-8724 on OSX 10.14
+		// passing nullptr, 0 forces allocation of a unique buffer we can free
+		// fixing MAINT-8724 on OSX 10.14
 		int status = -1;
 		char* name = abi::__cxa_demangle(mangled, nullptr, 0, &status);
-        std::string result(name ? name : mangled);
-        free(name);
-        return result;
-#elif LL_WINDOWS
-		// DevStudio: type_info::name() includes the text "class " at the start
+		std::string result(name ? name : mangled);
+		free(name);
+		return result;
 
-		static const std::string class_prefix = "class ";
+#elif LL_WINDOWS
+		// Visual Studio: type_info::name() includes the text "class " at the start
 		std::string name = mangled;
-		if (0 != name.compare(0, class_prefix.length(), class_prefix))
+		for (const auto& prefix : std::vector<std::string>{ "class ", "struct " })
 		{
-			LL_DEBUGS() << "Did not see '" << class_prefix << "' prefix on '"
-					   << name << "'" << LL_ENDL;
-			return name;
+			if (0 == name.compare(0, prefix.length(), prefix))
+			{
+				return name.substr(prefix.length());
+			}
 		}
+		// huh, that's odd, we should see one or the other prefix -- but don't
+		// try to log unless logging is already initialized
+		if (is_available())
+		{
+			// in Python, " or ".join(vector) -- but in C++, a PITB
+			LL_DEBUGS() << "Did not see 'class' or 'struct' prefix on '"
+				<< name << "'" << LL_ENDL;
+		}
+		return name;
 
-		return name.substr(class_prefix.length());
-
-#else
+#else  // neither GCC nor Visual Studio
 		return mangled;
 #endif
 	}
@@ -408,7 +419,7 @@ namespace
 				return false;
 			}
 
-			if (configuration.isUndefined() || !configuration.isMap() || configuration.emptyMap())
+			if (! configuration || !configuration.isMap())
 			{
 				LL_WARNS() << filename() << " missing, ill-formed, or simply undefined"
 							" content; not changing configuration"
@@ -490,14 +501,11 @@ namespace LLError
 		
 		LLError::FatalFunction              mCrashFunction;
 		LLError::TimeFunction               mTimeFunction;
-		
+
 		Recorders                           mRecorders;
-		RecorderPtr                         mFileRecorder;
-		RecorderPtr                         mFixedBufferRecorder;
-		std::string                         mFileRecorderFileName;
-		
-		int									mShouldLogCallCounter;
-		
+
+		int                                 mShouldLogCallCounter;
+
 	private:
 		SettingsConfig();
 	};
@@ -531,9 +539,6 @@ namespace LLError
 		mCrashFunction(NULL),
 		mTimeFunction(NULL),
 		mRecorders(),
-		mFileRecorder(),
-		mFixedBufferRecorder(),
-		mFileRecorderFileName(),
 		mShouldLogCallCounter(0)
 	{
 	}
@@ -656,22 +661,38 @@ namespace LLError
 
 namespace
 {
-	bool shouldLogToStderr()
-	{
+    bool shouldLogToStderr()
+    {
 #if LL_DARWIN
-		// On Mac OS X, stderr from apps launched from the Finder goes to the
-		// console log.  It's generally considered bad form to spam too much
-		// there.
-		
-		// If stdin is a tty, assume the user launched from the command line and
-		// therefore wants to see stderr.  Otherwise, assume we've been launched
-		// from the finder and shouldn't spam stderr.
-		return isatty(0);
+        // On Mac OS X, stderr from apps launched from the Finder goes to the
+        // console log.  It's generally considered bad form to spam too much
+        // there. That scenario can be detected by noticing that stderr is a
+        // character device (S_IFCHR).
+
+        // If stderr is a tty or a pipe, assume the user launched from the
+        // command line or debugger and therefore wants to see stderr.
+        if (isatty(STDERR_FILENO))
+            return true;
+        // not a tty, but might still be a pipe -- check
+        struct stat st;
+        if (fstat(STDERR_FILENO, &st) < 0)
+        {
+            // capture errno right away, before engaging any other operations
+            auto errno_save = errno;
+            // this gets called during log-system setup -- can't log yet!
+            std::cerr << "shouldLogToStderr: fstat(" << STDERR_FILENO << ") failed, errno "
+                      << errno_save << std::endl;
+            // if we can't tell, err on the safe side and don't write stderr
+            return false;
+        }
+
+        // fstat() worked: return true only if stderr is a pipe
+        return ((st.st_mode & S_IFMT) == S_IFIFO);
 #else
-		return true;
+        return true;
 #endif
-	}
-	
+    }
+
 	bool stderrLogWantsTime()
 	{
 #if LL_WINDOWS
@@ -685,20 +706,19 @@ namespace
 	void commonInit(const std::string& user_dir, const std::string& app_dir, bool log_to_stderr = true)
 	{
 		LLError::Settings::getInstance()->reset();
-		
+
 		LLError::setDefaultLevel(LLError::LEVEL_INFO);
-        LLError::setAlwaysFlush(true);
-        LLError::setEnabledLogTypesMask(0xFFFFFFFF);
+		LLError::setAlwaysFlush(true);
+		LLError::setEnabledLogTypesMask(0xFFFFFFFF);
 		LLError::setFatalFunction(LLError::crashAndLoop);
 		LLError::setTimeFunction(LLError::utcTime);
 
 		// log_to_stderr is only false in the unit and integration tests to keep builds quieter
 		if (log_to_stderr && shouldLogToStderr())
 		{
-			LLError::RecorderPtr recordToStdErr(new RecordToStderr(stderrLogWantsTime()));
-			LLError::addRecorder(recordToStdErr);
+			LLError::logToStderr();
 		}
-		
+
 #if LL_WINDOWS
 		LLError::RecorderPtr recordToWinDebug(new RecordToWinDebug());
 		LLError::addRecorder(recordToWinDebug);
@@ -996,49 +1016,110 @@ namespace LLError
 		s->mRecorders.erase(std::remove(s->mRecorders.begin(), s->mRecorders.end(), recorder),
 							s->mRecorders.end());
 	}
+
+    // Find an entry in SettingsConfig::mRecorders whose RecorderPtr points to
+    // a Recorder subclass of type RECORDER. Return, not a RecorderPtr (which
+    // points to the Recorder base class), but a shared_ptr<RECORDER> which
+    // specifically points to the concrete RECORDER subclass instance, along
+    // with a Recorders::iterator indicating the position of that entry in
+    // mRecorders. The shared_ptr might be empty (operator!() returns true) if
+    // there was no such RECORDER subclass instance in mRecorders.
+    template <typename RECORDER>
+    std::pair<boost::shared_ptr<RECORDER>, Recorders::iterator>
+    findRecorderPos()
+    {
+        SettingsConfigPtr s = Settings::instance().getSettingsConfig();
+        // Since we promise to return an iterator, use a classic iterator
+        // loop.
+        auto end{s->mRecorders.end()};
+        for (Recorders::iterator it{s->mRecorders.begin()}; it != end; ++it)
+        {
+            // *it is a RecorderPtr, a shared_ptr<Recorder>. Use a
+            // dynamic_pointer_cast to try to downcast to test if it's also a
+            // shared_ptr<RECORDER>.
+            auto ptr = boost::dynamic_pointer_cast<RECORDER>(*it);
+            if (ptr)
+            {
+                // found the entry we want
+                return { ptr, it };
+            }
+        }
+        // dropped out of the loop without finding any such entry -- instead
+        // of default-constructing Recorders::iterator (which might or might
+        // not be valid), return a value that is valid but not dereferenceable.
+        return { {}, end };
+    }
+
+    // Find an entry in SettingsConfig::mRecorders whose RecorderPtr points to
+    // a Recorder subclass of type RECORDER. Return, not a RecorderPtr (which
+    // points to the Recorder base class), but a shared_ptr<RECORDER> which
+    // specifically points to the concrete RECORDER subclass instance. The
+    // shared_ptr might be empty (operator!() returns true) if there was no
+    // such RECORDER subclass instance in mRecorders.
+    template <typename RECORDER>
+    boost::shared_ptr<RECORDER> findRecorder()
+    {
+        return findRecorderPos<RECORDER>().first;
+    }
+
+    // Remove an entry from SettingsConfig::mRecorders whose RecorderPtr
+    // points to a Recorder subclass of type RECORDER. Return true if there
+    // was one and we removed it, false if there wasn't one to start with.
+    template <typename RECORDER>
+    bool removeRecorder()
+    {
+        auto found = findRecorderPos<RECORDER>();
+        if (found.first)
+        {
+            SettingsConfigPtr s = Settings::instance().getSettingsConfig();
+            s->mRecorders.erase(found.second);
+        }
+        return bool(found.first);
+    }
 }
 
 namespace LLError
 {
 	void logToFile(const std::string& file_name)
 	{
-		SettingsConfigPtr s = Settings::getInstance()->getSettingsConfig();
+		// remove any previous Recorder filling this role
+		removeRecorder<RecordToFile>();
 
-		removeRecorder(s->mFileRecorder);
-		s->mFileRecorder.reset();
-		s->mFileRecorderFileName.clear();
-		
 		if (!file_name.empty())
 		{
-            RecorderPtr recordToFile(new RecordToFile(file_name));
-            if (boost::dynamic_pointer_cast<RecordToFile>(recordToFile)->okay())
-            {
-                s->mFileRecorderFileName = file_name;
-                s->mFileRecorder = recordToFile;
-                addRecorder(recordToFile);
-            }
+			boost::shared_ptr<RecordToFile> recordToFile(new RecordToFile(file_name));
+			if (recordToFile->okay())
+			{
+				addRecorder(recordToFile);
+			}
 		}
 	}
-	
-	void logToFixedBuffer(LLLineBuffer* fixedBuffer)
+
+	std::string logFileName()
 	{
-		SettingsConfigPtr s = Settings::getInstance()->getSettingsConfig();
+		auto found = findRecorder<RecordToFile>();
+		return found? found->getFilename() : std::string();
+	}
 
-		removeRecorder(s->mFixedBufferRecorder);
-		s->mFixedBufferRecorder.reset();
-		
-		if (fixedBuffer)
-		{
-            RecorderPtr recordToFixedBuffer(new RecordToFixedBuffer(fixedBuffer));
-            s->mFixedBufferRecorder = recordToFixedBuffer;
-            addRecorder(recordToFixedBuffer);
+    void logToStderr()
+    {
+        if (! findRecorder<RecordToStderr>())
+        {
+            RecorderPtr recordToStdErr(new RecordToStderr(stderrLogWantsTime()));
+            addRecorder(recordToStdErr);
         }
-	}
+    }
 
-	std::string logFileName()
+	void logToFixedBuffer(LLLineBuffer* fixedBuffer)
 	{
-		SettingsConfigPtr s = Settings::getInstance()->getSettingsConfig();
-		return s->mFileRecorderFileName;
+		// remove any previous Recorder filling this role
+		removeRecorder<RecordToFixedBuffer>();
+
+		if (fixedBuffer)
+		{
+			RecorderPtr recordToFixedBuffer(new RecordToFixedBuffer(fixedBuffer));
+			addRecorder(recordToFixedBuffer);
+		}
 	}
 }
 
@@ -1154,8 +1235,25 @@ namespace
 }
 
 namespace {
-	LLMutex gLogMutex;
-	LLMutex gCallStacksLogMutex;
+	// We need a couple different mutexes, but we want to use the same mechanism
+	// for both. Make getMutex() a template function with different instances
+	// for different MutexDiscriminator values.
+	enum MutexDiscriminator
+	{
+		LOG_MUTEX,
+		STACKS_MUTEX
+	};
+	// Some logging calls happen very early in processing -- so early that our
+	// module-static variables aren't yet initialized. getMutex() wraps a
+	// function-static LLMutex so that early calls can still have a valid
+	// LLMutex instance.
+	template <MutexDiscriminator MTX>
+	LLMutex* getMutex()
+	{
+		// guaranteed to be initialized the first time control reaches here
+		static LLMutex sMutex;
+		return &sMutex;
+	}
 
 	bool checkLevelMap(const LevelMap& map, const std::string& key,
 						LLError::ELevel& level)
@@ -1203,7 +1301,7 @@ namespace LLError
 
 	bool Log::shouldLog(CallSite& site)
 	{
-		LLMutexTrylock lock(&gLogMutex, 5);
+		LLMutexTrylock lock(getMutex<LOG_MUTEX>(), 5);
 		if (!lock.isLocked())
 		{
 			return false;
@@ -1254,7 +1352,7 @@ namespace LLError
 
 	std::ostringstream* Log::out()
 	{
-		LLMutexTrylock lock(&gLogMutex,5);
+		LLMutexTrylock lock(getMutex<LOG_MUTEX>(),5);
 		// If we hit a logging request very late during shutdown processing,
 		// when either of the relevant LLSingletons has already been deleted,
 		// DO NOT resurrect them.
@@ -1274,7 +1372,7 @@ namespace LLError
 
 	void Log::flush(std::ostringstream* out, char* message)
 	{
-		LLMutexTrylock lock(&gLogMutex,5);
+		LLMutexTrylock lock(getMutex<LOG_MUTEX>(),5);
 		if (!lock.isLocked())
 		{
 			return;
@@ -1314,7 +1412,7 @@ namespace LLError
 
 	void Log::flush(std::ostringstream* out, const CallSite& site)
 	{
-		LLMutexTrylock lock(&gLogMutex,5);
+		LLMutexTrylock lock(getMutex<LOG_MUTEX>(),5);
 		if (!lock.isLocked())
 		{
 			return;
@@ -1486,129 +1584,133 @@ namespace LLError
 	S32    LLCallStacks::sIndex  = 0 ;
 
 	//static
-   void LLCallStacks::allocateStackBuffer()
-   {
-	   if(sBuffer == NULL)
-	   {
-		   sBuffer = new char*[512] ;
-		   sBuffer[0] = new char[512 * 128] ;
-		   for(S32 i = 1 ; i < 512 ; i++)
-		   {
-			   sBuffer[i] = sBuffer[i-1] + 128 ;
-		   }
-		   sIndex = 0 ;
-	   }
-   }
-
-   void LLCallStacks::freeStackBuffer()
-   {
-	   if(sBuffer != NULL)
-	   {
-		   delete [] sBuffer[0] ;
-		   delete [] sBuffer ;
-		   sBuffer = NULL ;
-	   }
-   }
-
-   //static
-   void LLCallStacks::push(const char* function, const int line)
-   {
-       LLMutexTrylock lock(&gCallStacksLogMutex, 5);
-       if (!lock.isLocked())
-       {
-           return;
-       }
-
-	   if(sBuffer == NULL)
-	   {
-		   allocateStackBuffer();
-	   }
-
-	   if(sIndex > 511)
-	   {
-		   clear() ;
-	   }
-
-	   strcpy(sBuffer[sIndex], function) ;
-	   sprintf(sBuffer[sIndex] + strlen(function), " line: %d ", line) ;
-	   sIndex++ ;
-
-	   return ;
-   }
+    void LLCallStacks::allocateStackBuffer()
+    {
+        if(sBuffer == NULL)
+        {
+            sBuffer = new char*[512] ;
+            sBuffer[0] = new char[512 * 128] ;
+            for(S32 i = 1 ; i < 512 ; i++)
+            {
+                sBuffer[i] = sBuffer[i-1] + 128 ;
+            }
+            sIndex = 0 ;
+        }
+    }
 
-	//static
-   std::ostringstream* LLCallStacks::insert(const char* function, const int line)
-   {
-       std::ostringstream* _out = LLError::Log::out();
-	   *_out << function << " line " << line << " " ;
-             
-	   return _out ;
-   }
-
-   //static
-   void LLCallStacks::end(std::ostringstream* _out)
-   {
-       LLMutexTrylock lock(&gCallStacksLogMutex, 5);
-       if (!lock.isLocked())
-       {
-           return;
-       }
-
-	   if(sBuffer == NULL)
-	   {
-		   allocateStackBuffer();
-	   }
-
-	   if(sIndex > 511)
-	   {
-		   clear() ;
-	   }
-
-	   LLError::Log::flush(_out, sBuffer[sIndex++]) ;	   
-   }
-
-   //static
-   void LLCallStacks::print()
-   {
-       LLMutexTrylock lock(&gCallStacksLogMutex, 5);
-       if (!lock.isLocked())
-       {
-           return;
-       }
-
-       if(sIndex > 0)
-       {
-           LL_INFOS() << " ************* PRINT OUT LL CALL STACKS ************* " << LL_ENDL;
-           while(sIndex > 0)
-           {                  
-			   sIndex-- ;
-               LL_INFOS() << sBuffer[sIndex] << LL_ENDL;
-           }
-           LL_INFOS() << " *************** END OF LL CALL STACKS *************** " << LL_ENDL;
-       }
-
-	   if(sBuffer != NULL)
-	   {
-		   freeStackBuffer();
-	   }
-   }
-
-   //static
-   void LLCallStacks::clear()
-   {
-       sIndex = 0 ;
-   }
-
-   //static
-   void LLCallStacks::cleanup()
-   {
-	   freeStackBuffer();
-   }
+    void LLCallStacks::freeStackBuffer()
+    {
+        if(sBuffer != NULL)
+        {
+            delete [] sBuffer[0] ;
+            delete [] sBuffer ;
+            sBuffer = NULL ;
+        }
+    }
+
+    //static
+    void LLCallStacks::push(const char* function, const int line)
+    {
+        LLMutexTrylock lock(getMutex<STACKS_MUTEX>(), 5);
+        if (!lock.isLocked())
+        {
+            return;
+        }
+
+        if(sBuffer == NULL)
+        {
+            allocateStackBuffer();
+        }
+
+        if(sIndex > 511)
+        {
+            clear() ;
+        }
+
+        strcpy(sBuffer[sIndex], function) ;
+        sprintf(sBuffer[sIndex] + strlen(function), " line: %d ", line) ;
+        sIndex++ ;
+
+        return ;
+    }
+
+    //static
+    std::ostringstream* LLCallStacks::insert(const char* function, const int line)
+    {
+        std::ostringstream* _out = LLError::Log::out();
+        *_out << function << " line " << line << " " ;
+        return _out ;
+    }
+
+    //static
+    void LLCallStacks::end(std::ostringstream* _out)
+    {
+        LLMutexTrylock lock(getMutex<STACKS_MUTEX>(), 5);
+        if (!lock.isLocked())
+        {
+            return;
+        }
+
+        if(sBuffer == NULL)
+        {
+            allocateStackBuffer();
+        }
+
+        if(sIndex > 511)
+        {
+            clear() ;
+        }
+
+        LLError::Log::flush(_out, sBuffer[sIndex++]) ;
+    }
+
+    //static
+    void LLCallStacks::print()
+    {
+        LLMutexTrylock lock(getMutex<STACKS_MUTEX>(), 5);
+        if (!lock.isLocked())
+        {
+            return;
+        }
+
+        if(sIndex > 0)
+        {
+            LL_INFOS() << " ************* PRINT OUT LL CALL STACKS ************* " << LL_ENDL;
+            while(sIndex > 0)
+            {                  
+                sIndex-- ;
+                LL_INFOS() << sBuffer[sIndex] << LL_ENDL;
+            }
+            LL_INFOS() << " *************** END OF LL CALL STACKS *************** " << LL_ENDL;
+        }
+
+        if(sBuffer != NULL)
+        {
+            freeStackBuffer();
+        }
+    }
+
+    //static
+    void LLCallStacks::clear()
+    {
+        sIndex = 0 ;
+    }
+
+    //static
+    void LLCallStacks::cleanup()
+    {
+        freeStackBuffer();
+    }
+
+    std::ostream& operator<<(std::ostream& out, const LLStacktrace&)
+    {
+        return out << boost::stacktrace::stacktrace();
+    }
 }
 
 bool debugLoggingEnabled(const std::string& tag)
 {
-    LLMutexTrylock lock(&gLogMutex, 5);
+    LLMutexTrylock lock(getMutex<LOG_MUTEX>(), 5);
     if (!lock.isLocked())
     {
         return false;
diff --git a/indra/llcommon/llerror.h b/indra/llcommon/llerror.h
index 0a78229555..ffaa464d77 100644
--- a/indra/llcommon/llerror.h
+++ b/indra/llcommon/llerror.h
@@ -191,9 +191,9 @@ namespace LLError
 		The classes CallSite and Log are used by the logging macros below.
 		They are not intended for general use.
 	*/
-	
+
 	struct CallSite;
-	
+
 	class LL_COMMON_API Log
 	{
 	public:
@@ -202,8 +202,17 @@ namespace LLError
 		static void flush(std::ostringstream* out, char* message);
 		static void flush(std::ostringstream*, const CallSite&);
 		static std::string demangle(const char* mangled);
+		/// classname<TYPE>()
+		template <typename T>
+		static std::string classname()             { return demangle(typeid(T).name()); }
+		/// classname(some_pointer)
+		template <typename T>
+		static std::string classname(T* const ptr) { return ptr? demangle(typeid(*ptr).name()) : "nullptr"; }
+		/// classname(some_reference)
+		template <typename T>
+		static std::string classname(const T& obj) { return demangle(typeid(obj).name()); }
 	};
-	
+
 	struct LL_COMMON_API CallSite
 	{
 		// Represents a specific place in the code where a message is logged
@@ -262,30 +271,36 @@ namespace LLError
 	class LL_COMMON_API NoClassInfo { };
 		// used to indicate no class info known for logging
 
-   //LLCallStacks keeps track of call stacks and output the call stacks to log file
-   //when LLAppViewer::handleViewerCrash() is triggered.
-   //
-   //Note: to be simple, efficient and necessary to keep track of correct call stacks, 
-	//LLCallStacks is designed not to be thread-safe.
-   //so try not to use it in multiple parallel threads at same time.
-   //Used in a single thread at a time is fine.
-   class LL_COMMON_API LLCallStacks
-   {
-   private:
-       static char**  sBuffer ;
-	   static S32     sIndex ;
-
-	   static void allocateStackBuffer();
-	   static void freeStackBuffer();
-          
-   public:   
-	   static void push(const char* function, const int line) ;
-	   static std::ostringstream* insert(const char* function, const int line) ;
-       static void print() ;
-       static void clear() ;
-	   static void end(std::ostringstream* _out) ;
-	   static void cleanup();
-   }; 
+    //LLCallStacks keeps track of call stacks and output the call stacks to log file
+    //when LLAppViewer::handleViewerCrash() is triggered.
+    //
+    //Note: to be simple, efficient and necessary to keep track of correct call stacks, 
+    //LLCallStacks is designed not to be thread-safe.
+    //so try not to use it in multiple parallel threads at same time.
+    //Used in a single thread at a time is fine.
+    class LL_COMMON_API LLCallStacks
+    {
+    private:
+        static char**  sBuffer ;
+        static S32     sIndex ;
+
+        static void allocateStackBuffer();
+        static void freeStackBuffer();
+              
+    public:   
+        static void push(const char* function, const int line) ;
+        static std::ostringstream* insert(const char* function, const int line) ;
+        static void print() ;
+        static void clear() ;
+        static void end(std::ostringstream* _out) ;
+        static void cleanup();
+    };
+
+    // class which, when streamed, inserts the current stack trace
+    struct LLStacktrace
+    {
+        friend std::ostream& operator<<(std::ostream& out, const LLStacktrace&);
+    };
 }
 
 //this is cheaper than llcallstacks if no need to output other variables to call stacks. 
@@ -381,8 +396,13 @@ typedef LLError::NoClassInfo _LL_CLASS_TO_LOG;
 #define LL_WARNS(...)	lllog(LLError::LEVEL_WARN, false, ##__VA_ARGS__)
 #define LL_ERRS(...)	lllog(LLError::LEVEL_ERROR, false, ##__VA_ARGS__)
 // alternative to llassert_always that prints explanatory message
-#define LL_WARNS_IF(exp, ...)	if (exp) LL_WARNS(##__VA_ARGS__) << "(" #exp ")"
-#define LL_ERRS_IF(exp, ...)	if (exp) LL_ERRS(##__VA_ARGS__) << "(" #exp ")"
+// note ## token paste operator hack used above will only work in gcc following
+// a comma and is completely unnecessary in VS since the comma is automatically
+// suppressed
+// https://gcc.gnu.org/onlinedocs/cpp/Variadic-Macros.html
+// https://docs.microsoft.com/en-us/cpp/preprocessor/variadic-macros?view=vs-2015
+#define LL_WARNS_IF(exp, ...)	if (exp) LL_WARNS(__VA_ARGS__) << "(" #exp ")"
+#define LL_ERRS_IF(exp, ...)	if (exp) LL_ERRS(__VA_ARGS__) << "(" #exp ")"
 
 // Only print the log message once (good for warnings or infos that would otherwise
 // spam the log file over and over, such as tighter loops).
diff --git a/indra/llcommon/llerrorcontrol.h b/indra/llcommon/llerrorcontrol.h
index 276d22fc36..bfa2269025 100644
--- a/indra/llcommon/llerrorcontrol.h
+++ b/indra/llcommon/llerrorcontrol.h
@@ -183,6 +183,7 @@ namespace LLError
 		// each error message is passed to each recorder via recordMessage()
 
 	LL_COMMON_API void logToFile(const std::string& filename);
+	LL_COMMON_API void logToStderr();
 	LL_COMMON_API void logToFixedBuffer(LLLineBuffer*);
 		// Utilities to add recorders for logging to a file or a fixed buffer
 		// A second call to the same function will remove the logger added
diff --git a/indra/llcommon/lleventcoro.cpp b/indra/llcommon/lleventcoro.cpp
index 56367b8f54..995356dc52 100644
--- a/indra/llcommon/lleventcoro.cpp
+++ b/indra/llcommon/lleventcoro.cpp
@@ -31,17 +31,17 @@
 // associated header
 #include "lleventcoro.h"
 // STL headers
-#include <map>
+#include <chrono>
+#include <exception>
 // std headers
 // external library headers
+#include <boost/fiber/operations.hpp>
 // other Linden headers
 #include "llsdserialize.h"
+#include "llsdutil.h"
 #include "llerror.h"
 #include "llcoros.h"
-#include "llmake.h"
-#include "llexception.h"
-
-#include "lleventfilter.h"
+#include "stringize.h"
 
 namespace
 {
@@ -62,7 +62,7 @@ namespace
 std::string listenerNameForCoro()
 {
     // If this coroutine was launched by LLCoros::launch(), find that name.
-    std::string name(LLCoros::instance().getName());
+    std::string name(LLCoros::getName());
     if (! name.empty())
     {
         return name;
@@ -92,137 +92,173 @@ std::string listenerNameForCoro()
  * In the degenerate case in which @a path is an empty array, @a dest will
  * @em become @a value rather than @em containing it.
  */
-void storeToLLSDPath(LLSD& dest, const LLSD& rawPath, const LLSD& value)
+void storeToLLSDPath(LLSD& dest, const LLSD& path, const LLSD& value)
 {
-    if (rawPath.isUndefined())
+    if (path.isUndefined())
     {
         // no-op case
         return;
     }
 
-    // Arrange to treat rawPath uniformly as an array. If it's not already an
-    // array, store it as the only entry in one.
-    LLSD path;
-    if (rawPath.isArray())
-    {
-        path = rawPath;
-    }
-    else
-    {
-        path.append(rawPath);
-    }
-
-    // Need to indicate a current destination -- but that current destination
-    // needs to change as we step through the path array. Where normally we'd
-    // use an LLSD& to capture a subscripted LLSD lvalue, this time we must
-    // instead use a pointer -- since it must be reassigned.
-    LLSD* pdest = &dest;
-
-    // Now loop through that array
-    for (LLSD::Integer i = 0; i < path.size(); ++i)
-    {
-        if (path[i].isString())
-        {
-            // *pdest is an LLSD map
-            pdest = &((*pdest)[path[i].asString()]);
-        }
-        else if (path[i].isInteger())
-        {
-            // *pdest is an LLSD array
-            pdest = &((*pdest)[path[i].asInteger()]);
-        }
-        else
-        {
-            // What do we do with Real or Array or Map or ...?
-            // As it's a coder error -- not a user error -- rub the coder's
-            // face in it so it gets fixed.
-            LL_ERRS("lleventcoro") << "storeToLLSDPath(" << dest << ", " << rawPath << ", " << value
-                                   << "): path[" << i << "] bad type " << path[i].type() << LL_ENDL;
-        }
-    }
-
-    // Here *pdest is where we should store value.
-    *pdest = value;
+    // Drill down to where we should store 'value'.
+    llsd::drill(dest, path) = value;
 }
 
-/// For LLCoros::Future<LLSD>::make_callback(), the callback has a signature
-/// like void callback(LLSD), which isn't a valid LLEventPump listener: such
-/// listeners must return bool.
-template <typename LISTENER>
-class FutureListener
-{
-public:
-    // FutureListener is instantiated on the coroutine stack: the stack, in
-    // other words, that wants to suspend.
-    FutureListener(const LISTENER& listener):
-        mListener(listener),
-        // Capture the suspending coroutine's flag as a consuming or
-        // non-consuming listener.
-        mConsume(LLCoros::get_consuming())
-    {}
-
-    // operator()() is called on the main stack: the stack on which the
-    // expected event is fired.
-    bool operator()(const LLSD& event)
-    {
-        mListener(event);
-        // tell upstream LLEventPump whether listener consumed
-        return mConsume;
-    }
-
-protected:
-    LISTENER mListener;
-    bool mConsume;
-};
-
 } // anonymous
 
 void llcoro::suspend()
 {
-    // By viewer convention, we post an event on the "mainloop" LLEventPump
-    // each iteration of the main event-handling loop. So waiting for a single
-    // event on "mainloop" gives us a one-frame suspend.
-    suspendUntilEventOn("mainloop");
+    LLCoros::checkStop();
+    LLCoros::TempStatus st("waiting one tick");
+    boost::this_fiber::yield();
 }
 
 void llcoro::suspendUntilTimeout(float seconds)
 {
-    LLEventTimeout timeout;
-
-    timeout.eventAfter(seconds, LLSD());
-    llcoro::suspendUntilEventOn(timeout);
+    LLCoros::checkStop();
+    // We used to call boost::this_fiber::sleep_for(). But some coroutines
+    // (e.g. LLExperienceCache::idleCoro()) sit in a suspendUntilTimeout()
+    // loop, in which case a sleep_for() call risks sleeping through shutdown.
+    // So instead, listen for "LLApp" state-changing events -- which
+    // fortunately is handled for us by suspendUntilEventOnWithTimeout().
+    // Wait for an event on a bogus LLEventPump on which nobody ever posts
+    // events. Don't make it static because that would force instantiation of
+    // the LLEventPumps LLSingleton registry at static initialization time.
+    // DO allow tweaking the name for uniqueness, this definitely gets
+    // re-entered on multiple coroutines!
+    // We could use an LLUUID if it were important to actively prohibit anyone
+    // from ever posting on this LLEventPump.
+    LLEventStream bogus("xyzzy", true);
+    // Timeout is the NORMAL case for this call!
+    static LLSD timedout;
+    // Deliver, but ignore, timedout when (as usual) we did not receive any
+    // "LLApp" event. The point is that suspendUntilEventOnWithTimeout() will
+    // itself throw Stopping when "LLApp" starts broadcasting shutdown events.
+    suspendUntilEventOnWithTimeout(bogus, seconds, timedout);
 }
 
-LLSD llcoro::postAndSuspend(const LLSD& event, const LLEventPumpOrPumpName& requestPump,
-                 const LLEventPumpOrPumpName& replyPump, const LLSD& replyPumpNamePath)
+namespace
 {
-    // declare the future
-    LLCoros::Future<LLSD> future;
-    // make a callback that will assign a value to the future, and listen on
-    // the specified LLEventPump with that callback
-    std::string listenerName(listenerNameForCoro());
-    LLTempBoundListener connection(
-        replyPump.getPump().listen(listenerName,
-                                   llmake<FutureListener>(future.make_callback())));
+
+// returns a listener on replyPumpP, also on "mainloop" -- both should be
+// stored in LLTempBoundListeners on the caller's stack frame
+std::pair<LLBoundListener, LLBoundListener>
+postAndSuspendSetup(const std::string& callerName,
+                    const std::string& listenerName,
+                    LLCoros::Promise<LLSD>& promise,
+                    const LLSD& event,
+                    const LLEventPumpOrPumpName& requestPumpP,
+                    const LLEventPumpOrPumpName& replyPumpP,
+                    const LLSD& replyPumpNamePath)
+{
+    // Before we get any farther -- should we be stopping instead of
+    // suspending?
+    LLCoros::checkStop();
+    // Get the consuming attribute for THIS coroutine, the one that's about to
+    // suspend. Don't call get_consuming() in the lambda body: that would
+    // return the consuming attribute for some other coroutine, most likely
+    // the main routine.
+    bool consuming(LLCoros::get_consuming());
+    // listen on the specified LLEventPump with a lambda that will assign a
+    // value to the promise, thus fulfilling its future
+    llassert_always_msg(replyPumpP, ("replyPump required for " + callerName));
+    LLEventPump& replyPump(replyPumpP.getPump());
+    // The relative order of the two listen() calls below would only matter if
+    // "LLApp" were an LLEventMailDrop. But if we ever go there, we'd want to
+    // notice the pending LLApp status first.
+    LLBoundListener stopper(
+        LLEventPumps::instance().obtain("LLApp").listen(
+            listenerName,
+            [&promise, listenerName](const LLSD& status)
+            {
+                // anything except "running" should wake up the waiting
+                // coroutine
+                auto& statsd = status["status"];
+                if (statsd.asString() != "running")
+                {
+                    LL_DEBUGS("lleventcoro") << listenerName
+                                             << " spotted status " << statsd
+                                             << ", throwing Stopping" << LL_ENDL;
+                    try
+                    {
+                        promise.set_exception(
+                            std::make_exception_ptr(
+                                LLCoros::Stopping("status " + statsd.asString())));
+                    }
+                    catch (const boost::fibers::promise_already_satisfied&)
+                    {
+                        LL_WARNS("lleventcoro") << listenerName
+                                                << " couldn't throw Stopping "
+                                                   "because promise already set" << LL_ENDL;
+                    }
+                }
+                // do not consume -- every listener must see status
+                return false;
+            }));
+    LLBoundListener connection(
+        replyPump.listen(
+            listenerName,
+            [&promise, consuming, listenerName](const LLSD& result)
+            {
+                try
+                {
+                    promise.set_value(result);
+                    // We did manage to propagate the result value to the
+                    // (real) listener. If we're supposed to indicate that
+                    // we've consumed it, do so.
+                    return consuming;
+                }
+                catch(boost::fibers::promise_already_satisfied & ex)
+                {
+                    LL_DEBUGS("lleventcoro") << "promise already satisfied in '"
+                        << listenerName << "': "  << ex.what() << LL_ENDL;
+                    // We could not propagate the result value to the
+                    // listener.
+                    return false;
+                }
+            }));
+
     // skip the "post" part if requestPump is default-constructed
-    if (requestPump)
+    if (requestPumpP)
     {
+        LLEventPump& requestPump(requestPumpP.getPump());
         // If replyPumpNamePath is non-empty, store the replyPump name in the
         // request event.
         LLSD modevent(event);
-        storeToLLSDPath(modevent, replyPumpNamePath, replyPump.getPump().getName());
-        LL_DEBUGS("lleventcoro") << "postAndSuspend(): coroutine " << listenerName
-                                 << " posting to " << requestPump.getPump().getName()
+        storeToLLSDPath(modevent, replyPumpNamePath, replyPump.getName());
+        LL_DEBUGS("lleventcoro") << callerName << ": coroutine " << listenerName
+                                 << " posting to " << requestPump.getName()
                                  << LL_ENDL;
 
         // *NOTE:Mani - Removed because modevent could contain user's hashed passwd.
         //                         << ": " << modevent << LL_ENDL;
-        requestPump.getPump().post(modevent);
+        requestPump.post(modevent);
     }
-    LL_DEBUGS("lleventcoro") << "postAndSuspend(): coroutine " << listenerName
-                             << " about to wait on LLEventPump " << replyPump.getPump().getName()
+    LL_DEBUGS("lleventcoro") << callerName << ": coroutine " << listenerName
+                             << " about to wait on LLEventPump " << replyPump.getName()
                              << LL_ENDL;
+    return { connection, stopper };
+}
+
+} // anonymous
+
+LLSD llcoro::postAndSuspend(const LLSD& event, const LLEventPumpOrPumpName& requestPump,
+                 const LLEventPumpOrPumpName& replyPump, const LLSD& replyPumpNamePath)
+{
+    LLCoros::Promise<LLSD> promise;
+    std::string listenerName(listenerNameForCoro());
+
+    // Store both connections into LLTempBoundListeners so we implicitly
+    // disconnect on return from this function.
+    auto connections =
+        postAndSuspendSetup("postAndSuspend()", listenerName, promise,
+                            event, requestPump, replyPump, replyPumpNamePath);
+    LLTempBoundListener connection(connections.first), stopper(connections.second);
+
+    // declare the future
+    LLCoros::Future<LLSD> future = LLCoros::getFuture(promise);
     // calling get() on the future makes us wait for it
+    LLCoros::TempStatus st(STRINGIZE("waiting for " << replyPump.getPump().getName()));
     LLSD value(future.get());
     LL_DEBUGS("lleventcoro") << "postAndSuspend(): coroutine " << listenerName
                              << " resuming with " << value << LL_ENDL;
@@ -230,147 +266,52 @@ LLSD llcoro::postAndSuspend(const LLSD& event, const LLEventPumpOrPumpName& requ
     return value;
 }
 
-LLSD llcoro::suspendUntilEventOnWithTimeout(const LLEventPumpOrPumpName& suspendPumpOrName, 
-        F32 timeoutin, const LLSD &timeoutResult)
-{
-    /**
-     * The timeout pump is attached upstream of of the waiting pump and will 
-     * pass the timeout event through it.  We CAN NOT attach downstream since
-     * doing so will cause the suspendPump to fire any waiting events immediately 
-     * and they will be lost.  This becomes especially problematic with the 
-     * LLEventTimeout(pump) constructor which will also attempt to fire those
-     * events using the virtual listen_impl method in the not yet fully constructed
-     * timeoutPump.
-     */
-    LLEventTimeout timeoutPump;
-    LLEventPump &suspendPump = suspendPumpOrName.getPump();
-
-    LLTempBoundListener timeoutListener(timeoutPump.listen(suspendPump.getName(), 
-            boost::bind(&LLEventPump::post, &suspendPump, _1)));
-
-    timeoutPump.eventAfter(timeoutin, timeoutResult);
-    return llcoro::suspendUntilEventOn(suspendPump);
-}
-
-namespace
-{
-
-/**
- * This helper is specifically for postAndSuspend2(). We use a single future
- * object, but we want to listen on two pumps with it. Since we must still
- * adapt from the callable constructed by boost::dcoroutines::make_callback()
- * (void return) to provide an event listener (bool return), we've adapted
- * FutureListener for the purpose. The basic idea is that we construct a
- * distinct instance of FutureListener2 -- binding different instance data --
- * for each of the pumps. Then, when a pump delivers an LLSD value to either
- * FutureListener2, it can combine that LLSD with its discriminator to feed
- * the future object.
- *
- * DISCRIM is a template argument so we can use llmake() rather than
- * having to write our own argument-deducing helper function.
- */
-template <typename LISTENER, typename DISCRIM>
-class FutureListener2: public FutureListener<LISTENER>
+LLSD llcoro::postAndSuspendWithTimeout(const LLSD& event,
+                                       const LLEventPumpOrPumpName& requestPump,
+                                       const LLEventPumpOrPumpName& replyPump,
+                                       const LLSD& replyPumpNamePath,
+                                       F32 timeout, const LLSD& timeoutResult)
 {
-    typedef FutureListener<LISTENER> super;
-
-public:
-    // instantiated on coroutine stack: the stack about to suspend
-    FutureListener2(const LISTENER& listener, DISCRIM discriminator):
-        super(listener),
-        mDiscrim(discriminator)
-    {}
-
-    // called on main stack: the stack on which event is fired
-    bool operator()(const LLSD& event)
-    {
-        // our future object is defined to accept LLEventWithID
-        super::mListener(LLEventWithID(event, mDiscrim));
-        // tell LLEventPump whether or not event was consumed
-        return super::mConsume;
-    }
-
-private:
-    const DISCRIM mDiscrim;
-};
+    LLCoros::Promise<LLSD> promise;
+    std::string listenerName(listenerNameForCoro());
 
-} // anonymous
+    // Store both connections into LLTempBoundListeners so we implicitly
+    // disconnect on return from this function.
+    auto connections =
+        postAndSuspendSetup("postAndSuspendWithTimeout()", listenerName, promise,
+                            event, requestPump, replyPump, replyPumpNamePath);
+    LLTempBoundListener connection(connections.first), stopper(connections.second);
 
-namespace llcoro
-{
-
-LLEventWithID postAndSuspend2(const LLSD& event,
-                           const LLEventPumpOrPumpName& requestPump,
-                           const LLEventPumpOrPumpName& replyPump0,
-                           const LLEventPumpOrPumpName& replyPump1,
-                           const LLSD& replyPump0NamePath,
-                           const LLSD& replyPump1NamePath)
-{
     // declare the future
-    LLCoros::Future<LLEventWithID> future;
-    // either callback will assign a value to this future; listen on
-    // each specified LLEventPump with a callback
-    std::string name(listenerNameForCoro());
-    LLTempBoundListener connection0(
-        replyPump0.getPump().listen(
-            name + "a",
-            llmake<FutureListener2>(future.make_callback(), 0)));
-    LLTempBoundListener connection1(
-        replyPump1.getPump().listen(
-            name + "b",
-            llmake<FutureListener2>(future.make_callback(), 1)));
-    // skip the "post" part if requestPump is default-constructed
-    if (requestPump)
+    LLCoros::Future<LLSD> future = LLCoros::getFuture(promise);
+    // wait for specified timeout
+    boost::fibers::future_status status;
     {
-        // If either replyPumpNamePath is non-empty, store the corresponding
-        // replyPump name in the request event.
-        LLSD modevent(event);
-        storeToLLSDPath(modevent, replyPump0NamePath,
-                        replyPump0.getPump().getName());
-        storeToLLSDPath(modevent, replyPump1NamePath,
-                        replyPump1.getPump().getName());
-        LL_DEBUGS("lleventcoro") << "postAndSuspend2(): coroutine " << name
-                                 << " posting to " << requestPump.getPump().getName()
-                                 << ": " << modevent << LL_ENDL;
-        requestPump.getPump().post(modevent);
+        LLCoros::TempStatus st(STRINGIZE("waiting for " << replyPump.getPump().getName()
+                                         << " for " << timeout << "s"));
+        // The fact that we accept non-integer seconds means we should probably
+        // use granularity finer than one second. However, given the overhead of
+        // the rest of our processing, it seems silly to use granularity finer
+        // than a millisecond.
+        status = future.wait_for(std::chrono::milliseconds(long(timeout * 1000)));
     }
-    LL_DEBUGS("lleventcoro") << "postAndSuspend2(): coroutine " << name
-                             << " about to wait on LLEventPumps " << replyPump0.getPump().getName()
-                             << ", " << replyPump1.getPump().getName() << LL_ENDL;
-    // calling get() on the future makes us wait for it
-    LLEventWithID value(future.get());
-    LL_DEBUGS("lleventcoro") << "postAndSuspend(): coroutine " << name
-                             << " resuming with (" << value.first << ", " << value.second << ")"
-                             << LL_ENDL;
-    // returning should disconnect both connections
-    return value;
-}
-
-LLSD errorException(const LLEventWithID& result, const std::string& desc)
-{
-    // If the result arrived on the error pump (pump 1), instead of
-    // returning it, deliver it via exception.
-    if (result.second)
+    // if the future is NOT yet ready, return timeoutResult instead
+    if (status == boost::fibers::future_status::timeout)
     {
-        LLTHROW(LLErrorEvent(desc, result.first));
+        LL_DEBUGS("lleventcoro") << "postAndSuspendWithTimeout(): coroutine " << listenerName
+                                 << " timed out after " << timeout << " seconds,"
+                                 << " resuming with " << timeoutResult << LL_ENDL;
+        return timeoutResult;
     }
-    // That way, our caller knows a simple return must be from the reply
-    // pump (pump 0).
-    return result.first;
-}
-
-LLSD errorLog(const LLEventWithID& result, const std::string& desc)
-{
-    // If the result arrived on the error pump (pump 1), log it as a fatal
-    // error.
-    if (result.second)
+    else
     {
-        LL_ERRS("errorLog") << desc << ":" << std::endl;
-        LLSDSerialize::toPrettyXML(result.first, LL_CONT);
-        LL_CONT << LL_ENDL;
+        llassert_always(status == boost::fibers::future_status::ready);
+
+        // future is now ready, no more waiting
+        LLSD value(future.get());
+        LL_DEBUGS("lleventcoro") << "postAndSuspendWithTimeout(): coroutine " << listenerName
+                                 << " resuming with " << value << LL_ENDL;
+        // returning should disconnect the connection
+        return value;
     }
-    // A simple return must therefore be from the reply pump (pump 0).
-    return result.first;
 }
-
-} // namespace llcoro
diff --git a/indra/llcommon/lleventcoro.h b/indra/llcommon/lleventcoro.h
index 84827aab4a..c0fe8b094f 100644
--- a/indra/llcommon/lleventcoro.h
+++ b/indra/llcommon/lleventcoro.h
@@ -29,12 +29,8 @@
 #if ! defined(LL_LLEVENTCORO_H)
 #define LL_LLEVENTCORO_H
 
-#include <boost/optional.hpp>
 #include <string>
-#include <utility>                  // std::pair
 #include "llevents.h"
-#include "llerror.h"
-#include "llexception.h"
 
 /**
  * Like LLListenerOrPumpName, this is a class intended for parameter lists:
@@ -147,117 +143,29 @@ LLSD suspendUntilEventOn(const LLEventPumpOrPumpName& pump)
     return postAndSuspend(LLSD(), LLEventPumpOrPumpName(), pump);
 }
 
+/// Like postAndSuspend(), but if we wait longer than @a timeout seconds,
+/// stop waiting and return @a timeoutResult instead.
+LLSD postAndSuspendWithTimeout(const LLSD& event,
+                               const LLEventPumpOrPumpName& requestPump,
+                               const LLEventPumpOrPumpName& replyPump,
+                               const LLSD& replyPumpNamePath,
+                               F32 timeout, const LLSD& timeoutResult);
+
 /// Suspend the coroutine until an event is fired on the identified pump
 /// or the timeout duration has elapsed.  If the timeout duration 
 /// elapses the specified LLSD is returned.
-LLSD suspendUntilEventOnWithTimeout(const LLEventPumpOrPumpName& suspendPumpOrName, F32 timeoutin, const LLSD &timeoutResult);
-
-} // namespace llcoro
-
-/// return type for two-pump variant of suspendUntilEventOn()
-typedef std::pair<LLSD, int> LLEventWithID;
-
-namespace llcoro
-{
-
-/**
- * This function waits for a reply on either of two specified LLEventPumps.
- * Otherwise, it closely resembles postAndSuspend(); please see the documentation
- * for that function for detailed parameter info.
- *
- * While we could have implemented the single-pump variant in terms of this
- * one, there's enough added complexity here to make it worthwhile to give the
- * single-pump variant its own straightforward implementation. Conversely,
- * though we could use preprocessor logic to generate n-pump overloads up to
- * BOOST_COROUTINE_WAIT_MAX, we don't foresee a use case. This two-pump
- * overload exists because certain event APIs are defined in terms of a reply
- * LLEventPump and an error LLEventPump.
- *
- * The LLEventWithID return value provides not only the received event, but
- * the index of the pump on which it arrived (0 or 1).
- *
- * @note
- * I'd have preferred to overload the name postAndSuspend() for both signatures.
- * But consider the following ambiguous call:
- * @code
- * postAndSuspend(LLSD(), requestPump, replyPump, "someString");
- * @endcode
- * "someString" could be converted to either LLSD (@a replyPumpNamePath for
- * the single-pump function) or LLEventOrPumpName (@a replyPump1 for two-pump
- * function).
- *
- * It seems less burdensome to write postAndSuspend2() than to write either
- * LLSD("someString") or LLEventOrPumpName("someString").
- */
-LLEventWithID postAndSuspend2(const LLSD& event,
-                           const LLEventPumpOrPumpName& requestPump,
-                           const LLEventPumpOrPumpName& replyPump0,
-                           const LLEventPumpOrPumpName& replyPump1,
-                           const LLSD& replyPump0NamePath=LLSD(),
-                           const LLSD& replyPump1NamePath=LLSD());
-
-/**
- * Wait for the next event on either of two specified LLEventPumps.
- */
 inline
-LLEventWithID
-suspendUntilEventOn(const LLEventPumpOrPumpName& pump0, const LLEventPumpOrPumpName& pump1)
+LLSD suspendUntilEventOnWithTimeout(const LLEventPumpOrPumpName& suspendPumpOrName,
+                                    F32 timeoutin, const LLSD &timeoutResult)
 {
-    // This is now a convenience wrapper for postAndSuspend2().
-    return postAndSuspend2(LLSD(), LLEventPumpOrPumpName(), pump0, pump1);
+    return postAndSuspendWithTimeout(LLSD(),                  // event
+                                     LLEventPumpOrPumpName(), // requestPump
+                                     suspendPumpOrName,       // replyPump
+                                     LLSD(),                  // replyPumpNamePath
+                                     timeoutin,
+                                     timeoutResult);
 }
 
-/**
- * Helper for the two-pump variant of suspendUntilEventOn(), e.g.:
- *
- * @code
- * LLSD reply = errorException(suspendUntilEventOn(replyPump, errorPump),
- *                             "error response from login.cgi");
- * @endcode
- *
- * Examines an LLEventWithID, assuming that the second pump (pump 1) is
- * listening for an error indication. If the incoming data arrived on pump 1,
- * throw an LLErrorEvent exception. If the incoming data arrived on pump 0,
- * just return it. Since a normal return can only be from pump 0, we no longer
- * need the LLEventWithID's discriminator int; we can just return the LLSD.
- *
- * @note I'm not worried about introducing the (fairly generic) name
- * errorException() into global namespace, because how many other overloads of
- * the same name are going to accept an LLEventWithID parameter?
- */
-LLSD errorException(const LLEventWithID& result, const std::string& desc);
-
-} // namespace llcoro
-
-/**
- * Exception thrown by errorException(). We don't call this LLEventError
- * because it's not an error in event processing: rather, this exception
- * announces an event that bears error information (for some other API).
- */
-class LL_COMMON_API LLErrorEvent: public LLException
-{
-public:
-    LLErrorEvent(const std::string& what, const LLSD& data):
-        LLException(what),
-        mData(data)
-    {}
-    virtual ~LLErrorEvent() throw() {}
-
-    LLSD getData() const { return mData; }
-
-private:
-    LLSD mData;
-};
-
-namespace llcoro
-{
-
-/**
- * Like errorException(), save that this trips a fatal error using LL_ERRS
- * rather than throwing an exception.
- */
-LL_COMMON_API LLSD errorLog(const LLEventWithID& result, const std::string& desc);
-
 } // namespace llcoro
 
 /**
@@ -304,84 +212,4 @@ private:
     LLEventStream mPump;
 };
 
-/**
- * Other event APIs require the names of two different LLEventPumps: one for
- * success response, the other for error response. Extend LLCoroEventPump
- * for the two-pump use case.
- */
-class LL_COMMON_API LLCoroEventPumps
-{
-public:
-    LLCoroEventPumps(const std::string& name="coro",
-                     const std::string& suff0="Reply",
-                     const std::string& suff1="Error"):
-        mPump0(name + suff0, true),   // allow tweaking the pump instance name
-        mPump1(name + suff1, true)
-    {}
-    /// request pump 0's name
-    std::string getName0() const { return mPump0.getName(); }
-    /// request pump 1's name
-    std::string getName1() const { return mPump1.getName(); }
-    /// request both names
-    std::pair<std::string, std::string> getNames() const
-    {
-        return std::pair<std::string, std::string>(mPump0.getName(), mPump1.getName());
-    }
-
-    /// request pump 0
-    LLEventPump& getPump0() { return mPump0; }
-    /// request pump 1
-    LLEventPump& getPump1() { return mPump1; }
-
-    /// suspendUntilEventOn(either of our two LLEventPumps)
-    LLEventWithID suspend()
-    {
-        return llcoro::suspendUntilEventOn(mPump0, mPump1);
-    }
-
-    /// errorException(suspend())
-    LLSD suspendWithException()
-    {
-        return llcoro::errorException(suspend(), std::string("Error event on ") + getName1());
-    }
-
-    /// errorLog(suspend())
-    LLSD suspendWithLog()
-    {
-        return llcoro::errorLog(suspend(), std::string("Error event on ") + getName1());
-    }
-
-    LLEventWithID postAndSuspend(const LLSD& event,
-                              const LLEventPumpOrPumpName& requestPump,
-                              const LLSD& replyPump0NamePath=LLSD(),
-                              const LLSD& replyPump1NamePath=LLSD())
-    {
-        return llcoro::postAndSuspend2(event, requestPump, mPump0, mPump1,
-                                    replyPump0NamePath, replyPump1NamePath);
-    }
-
-    LLSD postAndSuspendWithException(const LLSD& event,
-                                  const LLEventPumpOrPumpName& requestPump,
-                                  const LLSD& replyPump0NamePath=LLSD(),
-                                  const LLSD& replyPump1NamePath=LLSD())
-    {
-        return llcoro::errorException(postAndSuspend(event, requestPump,
-                                                  replyPump0NamePath, replyPump1NamePath),
-                                      std::string("Error event on ") + getName1());
-    }
-
-    LLSD postAndSuspendWithLog(const LLSD& event,
-                            const LLEventPumpOrPumpName& requestPump,
-                            const LLSD& replyPump0NamePath=LLSD(),
-                            const LLSD& replyPump1NamePath=LLSD())
-    {
-        return llcoro::errorLog(postAndSuspend(event, requestPump,
-                                            replyPump0NamePath, replyPump1NamePath),
-                                std::string("Error event on ") + getName1());
-    }
-
-private:
-    LLEventStream mPump0, mPump1;
-};
-
 #endif /* ! defined(LL_LLEVENTCORO_H) */
diff --git a/indra/llcommon/lleventfilter.cpp b/indra/llcommon/lleventfilter.cpp
index 9fb18dc67d..4cded7f88e 100644
--- a/indra/llcommon/lleventfilter.cpp
+++ b/indra/llcommon/lleventfilter.cpp
@@ -37,6 +37,9 @@
 // other Linden headers
 #include "llerror.h"                // LL_ERRS
 #include "llsdutil.h"               // llsd_matches()
+#include "stringize.h"
+#include "lleventtimer.h"
+#include "lldate.h"
 
 /*****************************************************************************
 *   LLEventFilter
@@ -182,6 +185,27 @@ bool LLEventTimeout::countdownElapsed() const
     return mTimer.hasExpired();
 }
 
+LLEventTimer* LLEventTimeout::post_every(F32 period, const std::string& pump, const LLSD& data)
+{
+    return LLEventTimer::run_every(
+        period,
+        [pump, data](){ LLEventPumps::instance().obtain(pump).post(data); });
+}
+
+LLEventTimer* LLEventTimeout::post_at(const LLDate& time, const std::string& pump, const LLSD& data)
+{
+    return LLEventTimer::run_at(
+        time,
+        [pump, data](){ LLEventPumps::instance().obtain(pump).post(data); });
+}
+
+LLEventTimer* LLEventTimeout::post_after(F32 interval, const std::string& pump, const LLSD& data)
+{
+    return LLEventTimer::run_after(
+        interval,
+        [pump, data](){ LLEventPumps::instance().obtain(pump).post(data); });
+}
+
 /*****************************************************************************
 *   LLEventBatch
 *****************************************************************************/
@@ -409,3 +433,61 @@ void LLEventBatchThrottle::setSize(std::size_t size)
         flush();
     }
 }
+
+/*****************************************************************************
+*   LLEventLogProxy
+*****************************************************************************/
+LLEventLogProxy::LLEventLogProxy(LLEventPump& source, const std::string& name, bool tweak):
+    // note: we are NOT using the constructor that implicitly connects!
+    LLEventFilter(name, tweak),
+    // instead we simply capture a reference to the subject LLEventPump
+    mPump(source)
+{
+}
+
+bool LLEventLogProxy::post(const LLSD& event) /* override */
+{
+    auto counter = mCounter++;
+    auto eventplus = event;
+    if (eventplus.type() == LLSD::TypeMap)
+    {
+        eventplus["_cnt"] = counter;
+    }
+    std::string hdr{STRINGIZE(getName() << ": post " << counter)};
+    LL_INFOS("LogProxy") << hdr << ": " << event << LL_ENDL;
+    bool result = mPump.post(eventplus);
+    LL_INFOS("LogProxy") << hdr << " => " << result << LL_ENDL;
+    return result;
+}
+
+LLBoundListener LLEventLogProxy::listen_impl(const std::string& name,
+                                             const LLEventListener& target,
+                                             const NameList& after,
+                                             const NameList& before)
+{
+    LL_DEBUGS("LogProxy") << "LLEventLogProxy('" << getName() << "').listen('"
+                          << name << "')" << LL_ENDL;
+    return mPump.listen(name,
+                        [this, name, target](const LLSD& event)->bool
+                        { return listener(name, target, event); },
+                        after,
+                        before);
+}
+
+bool LLEventLogProxy::listener(const std::string& name,
+                               const LLEventListener& target,
+                               const LLSD& event) const
+{
+    auto eventminus = event;
+    std::string counter{"**"};
+    if (eventminus.has("_cnt"))
+    {
+        counter = stringize(eventminus["_cnt"].asInteger());
+        eventminus.erase("_cnt");
+    }
+    std::string hdr{STRINGIZE(getName() << " to " << name << " " << counter)};
+    LL_INFOS("LogProxy") << hdr << ": " << eventminus << LL_ENDL;
+    bool result = target(eventminus);
+    LL_INFOS("LogProxy") << hdr << " => " << result << LL_ENDL;
+    return result;
+}
diff --git a/indra/llcommon/lleventfilter.h b/indra/llcommon/lleventfilter.h
index ff8fc9bc7f..48c2570732 100644
--- a/indra/llcommon/lleventfilter.h
+++ b/indra/llcommon/lleventfilter.h
@@ -32,8 +32,12 @@
 #include "llevents.h"
 #include "stdtypes.h"
 #include "lltimer.h"
+#include "llsdutil.h"
 #include <boost/function.hpp>
 
+class LLEventTimer;
+class LLDate;
+
 /**
  * Generic base class
  */
@@ -210,6 +214,19 @@ public:
     LLEventTimeout();
     LLEventTimeout(LLEventPump& source);
 
+    /// using LLEventTimeout as namespace for free functions
+    /// Post event to specified LLEventPump every period seconds. Delete
+    /// returned LLEventTimer* to cancel.
+    static LLEventTimer* post_every(F32 period, const std::string& pump, const LLSD& data);
+    /// Post event to specified LLEventPump at specified future time. Call
+    /// LLEventTimer::getInstance(returned pointer) to check whether it's still
+    /// pending; if so, delete the pointer to cancel.
+    static LLEventTimer* post_at(const LLDate& time, const std::string& pump, const LLSD& data);
+    /// Post event to specified LLEventPump after specified interval. Call
+    /// LLEventTimer::getInstance(returned pointer) to check whether it's still
+    /// pending; if so, delete the pointer to cancel.
+    static LLEventTimer* post_after(F32 interval, const std::string& pump, const LLSD& data);
+
 protected:
     virtual void setCountdown(F32 seconds);
     virtual bool countdownElapsed() const;
@@ -376,4 +393,149 @@ private:
     std::size_t mBatchSize;
 };
 
+/**
+ * LLStoreListener self-registers on the LLEventPump of interest, and
+ * unregisters on destruction. As long as it exists, a particular element is
+ * extracted from every event that comes through the upstream LLEventPump and
+ * stored into the target variable.
+ *
+ * This is implemented as a subclass of LLEventFilter, though strictly
+ * speaking it isn't really a "filter" at all: it never passes incoming events
+ * to its own listeners, if any.
+ *
+ * TBD: A variant based on output iterators that stores and then increments
+ * the iterator. Useful with boost::coroutine2!
+ */
+template <typename T>
+class LLStoreListener: public LLEventFilter
+{
+public:
+    // pass target and optional path to element
+    LLStoreListener(T& target, const LLSD& path=LLSD(), bool consume=false):
+        LLEventFilter("store"),
+        mTarget(target),
+        mPath(path),
+        mConsume(consume)
+    {}
+    // construct and connect
+    LLStoreListener(LLEventPump& source, T& target, const LLSD& path=LLSD(), bool consume=false):
+        LLEventFilter(source, "store"),
+        mTarget(target),
+        mPath(path),
+        mConsume(consume)
+    {}
+
+    // Calling post() with an LLSD event extracts the element indicated by
+    // path, then stores it to mTarget.
+    virtual bool post(const LLSD& event)
+    {
+        // Extract the element specified by 'mPath' from 'event'. To perform a
+        // generic type-appropriate store through mTarget, construct an
+        // LLSDParam<T> and store that, thus engaging LLSDParam's custom
+        // conversions.
+        mTarget = LLSDParam<T>(llsd::drill(event, mPath));
+        return mConsume;
+    }
+
+private:
+    T& mTarget;
+    const LLSD mPath;
+    const bool mConsume;
+};
+
+/*****************************************************************************
+*   LLEventLogProxy
+*****************************************************************************/
+/**
+ * LLEventLogProxy is a little different than the other LLEventFilter
+ * subclasses declared in this header file, in that it completely wraps the
+ * passed LLEventPump (both input and output) instead of simply processing its
+ * output. Of course, if someone directly posts to the wrapped LLEventPump by
+ * looking up its string name in LLEventPumps, LLEventLogProxy can't intercept
+ * that post() call. But as long as consuming code is willing to access the
+ * LLEventLogProxy instance instead of the wrapped LLEventPump, all event data
+ * both post()ed and received is logged.
+ *
+ * The proxy role means that LLEventLogProxy intercepts more of LLEventPump's
+ * API than a typical LLEventFilter subclass.
+ */
+class LLEventLogProxy: public LLEventFilter
+{
+    typedef LLEventFilter super;
+public:
+    /**
+     * Construct LLEventLogProxy, wrapping the specified LLEventPump.
+     * Unlike a typical LLEventFilter subclass, the name parameter is @emph
+     * not optional because typically you want LLEventLogProxy to completely
+     * replace the wrapped LLEventPump. So you give the subject LLEventPump
+     * some other name and give the LLEventLogProxy the name that would have
+     * been used for the subject LLEventPump.
+     */
+    LLEventLogProxy(LLEventPump& source, const std::string& name, bool tweak=false);
+
+    /// register a new listener
+    LLBoundListener listen_impl(const std::string& name, const LLEventListener& target,
+                                const NameList& after, const NameList& before);
+
+    /// Post an event to all listeners
+    virtual bool post(const LLSD& event) /* override */;
+
+private:
+    /// This method intercepts each call to any target listener. We pass it
+    /// the listener name and the caller's intended target listener plus the
+    /// posted LLSD event.
+    bool listener(const std::string& name,
+                  const LLEventListener& target,
+                  const LLSD& event) const;
+
+    LLEventPump& mPump;
+    LLSD::Integer mCounter{0};
+};
+
+/**
+ * LLEventPumpHolder<T> is a helper for LLEventLogProxyFor<T>. It simply
+ * stores an instance of T, presumably a subclass of LLEventPump. We derive
+ * LLEventLogProxyFor<T> from LLEventPumpHolder<T>, ensuring that
+ * LLEventPumpHolder's contained mWrappedPump is fully constructed before
+ * passing it to LLEventLogProxyFor's LLEventLogProxy base class constructor.
+ * But since LLEventPumpHolder<T> presents none of the LLEventPump API,
+ * LLEventLogProxyFor<T> inherits its methods unambiguously from
+ * LLEventLogProxy.
+ */
+template <class T>
+class LLEventPumpHolder
+{
+protected:
+    LLEventPumpHolder(const std::string& name, bool tweak=false):
+        mWrappedPump(name, tweak)
+    {}
+    T mWrappedPump;
+};
+
+/**
+ * LLEventLogProxyFor<T> is a wrapper around any of the LLEventPump subclasses.
+ * Instantiating an LLEventLogProxy<T> instantiates an internal T. Otherwise
+ * it behaves like LLEventLogProxy.
+ */
+template <class T>
+class LLEventLogProxyFor: private LLEventPumpHolder<T>, public LLEventLogProxy
+{
+    // We derive privately from LLEventPumpHolder because it's an
+    // implementation detail of LLEventLogProxyFor. The only reason it's a
+    // base class at all is to guarantee that it's constructed first so we can
+    // pass it to our LLEventLogProxy base class constructor.
+    typedef LLEventPumpHolder<T> holder;
+    typedef LLEventLogProxy super;
+
+public:
+    LLEventLogProxyFor(const std::string& name, bool tweak=false):
+        // our wrapped LLEventPump subclass instance gets a name suffix
+        // because that's not the LLEventPump we want consumers to obtain when
+        // they ask LLEventPumps for this name
+        holder(name + "-", tweak),
+        // it's our LLEventLogProxy that gets the passed name
+        super(holder::mWrappedPump, name, tweak)
+    {}
+};
+
 #endif /* ! defined(LL_LLEVENTFILTER_H) */
diff --git a/indra/llcommon/llevents.cpp b/indra/llcommon/llevents.cpp
index eedd8c92b5..64fb985951 100644
--- a/indra/llcommon/llevents.cpp
+++ b/indra/llcommon/llevents.cpp
@@ -45,6 +45,7 @@
 #include <cctype>
 // external library headers
 #include <boost/range/iterator_range.hpp>
+#include <boost/make_shared.hpp>
 #if LL_WINDOWS
 #pragma warning (push)
 #pragma warning (disable : 4701) // compiler thinks might use uninitialized var, but no
@@ -63,51 +64,23 @@
 #endif
 
 /*****************************************************************************
-*   queue_names: specify LLEventPump names that should be instantiated as
-*   LLEventQueue
-*****************************************************************************/
-/**
- * At present, we recognize particular requested LLEventPump names as needing
- * LLEventQueues. Later on we'll migrate this information to an external
- * configuration file.
- */
-const char* queue_names[] =
-{
-    "placeholder - replace with first real name string"
-};
-
-/*****************************************************************************
-*   If there's a "mainloop" pump, listen on that to flush all LLEventQueues
+*   LLEventPumps
 *****************************************************************************/
-struct RegisterFlush : public LLEventTrackable
-{
-    RegisterFlush():
-        pumps(LLEventPumps::instance())
+LLEventPumps::LLEventPumps():
+    mFactories
     {
-        pumps.obtain("mainloop").listen("flushLLEventQueues", boost::bind(&RegisterFlush::flush, this, _1));
-    }
-    bool flush(const LLSD&)
+        { "LLEventStream",   [](const std::string& name, bool tweak)
+                             { return new LLEventStream(name, tweak); } },
+        { "LLEventMailDrop", [](const std::string& name, bool tweak)
+                             { return new LLEventMailDrop(name, tweak); } }
+    },
+    mTypes
     {
-        pumps.flush();
-        return false;
+        // LLEventStream is the default for obtain(), so even if somebody DOES
+        // call obtain("placeholder"), this sample entry won't break anything.
+        { "placeholder", "LLEventStream" }
     }
-    ~RegisterFlush()
-    {
-        // LLEventTrackable handles stopListening for us.
-    }
-    LLEventPumps& pumps;
-};
-static RegisterFlush registerFlush;
-
-/*****************************************************************************
-*   LLEventPumps
-*****************************************************************************/
-LLEventPumps::LLEventPumps():
-    // Until we migrate this information to an external config file,
-    // initialize mQueueNames from the static queue_names array.
-    mQueueNames(boost::begin(queue_names), boost::end(queue_names))
-{
-}
+{}
 
 LLEventPump& LLEventPumps::obtain(const std::string& name)
 {
@@ -118,14 +91,31 @@ LLEventPump& LLEventPumps::obtain(const std::string& name)
         // name.
         return *found->second;
     }
-    // Here we must instantiate an LLEventPump subclass. 
-    LLEventPump* newInstance;
-    // Should this name be an LLEventQueue?
-    PumpNames::const_iterator nfound = mQueueNames.find(name);
-    if (nfound != mQueueNames.end())
-        newInstance = new LLEventQueue(name);
-    else
-        newInstance = new LLEventStream(name);
+
+    // Here we must instantiate an LLEventPump subclass. Is there a
+    // preregistered class name override for this specific instance name?
+    auto nfound = mTypes.find(name);
+    std::string type;
+    if (nfound != mTypes.end())
+    {
+        type = nfound->second;
+    }
+    // pass tweak=false: we already know there's no existing instance with
+    // this name
+    return make(name, false, type);
+}
+
+LLEventPump& LLEventPumps::make(const std::string& name, bool tweak,
+                                const std::string& type)
+{
+    // find the relevant factory for this (or default) type
+    auto found = mFactories.find(type.empty()? "LLEventStream" : type);
+    if (found == mFactories.end())
+    {
+        // Passing an unrecognized type name is a no-no
+        LLTHROW(BadType(type));
+    }
+    auto newInstance = (found->second)(name, tweak);
     // LLEventPump's constructor implicitly registers each new instance in
     // mPumpMap. But remember that we instantiated it (in mOurPumps) so we'll
     // delete it later.
@@ -143,14 +133,23 @@ bool LLEventPumps::post(const std::string&name, const LLSD&message)
     return (*found).second->post(message);
 }
 
-
 void LLEventPumps::flush()
 {
     // Flush every known LLEventPump instance. Leave it up to each instance to
     // decide what to do with the flush() call.
-    for (PumpMap::iterator pmi = mPumpMap.begin(), pmend = mPumpMap.end(); pmi != pmend; ++pmi)
+    for (PumpMap::value_type& pair : mPumpMap)
+    {
+        pair.second->flush();
+    }
+}
+
+void LLEventPumps::clear()
+{
+    // Clear every known LLEventPump instance. Leave it up to each instance to
+    // decide what to do with the clear() call.
+    for (PumpMap::value_type& pair : mPumpMap)
     {
-        pmi->second->flush();
+        pair.second->clear();
     }
 }
 
@@ -158,9 +157,9 @@ void LLEventPumps::reset()
 {
     // Reset every known LLEventPump instance. Leave it up to each instance to
     // decide what to do with the reset() call.
-    for (PumpMap::iterator pmi = mPumpMap.begin(), pmend = mPumpMap.end(); pmi != pmend; ++pmi)
+    for (PumpMap::value_type& pair : mPumpMap)
     {
-        pmi->second->reset();
+        pair.second->reset();
     }
 }
 
@@ -267,6 +266,9 @@ LLEventPumps::~LLEventPumps()
     {
         delete *mOurPumps.begin();
     }
+    // Reset every remaining registered LLEventPump subclass instance: those
+    // we DIDN'T instantiate using either make() or obtain().
+    reset();
 }
 
 /*****************************************************************************
@@ -283,7 +285,7 @@ LLEventPump::LLEventPump(const std::string& name, bool tweak):
     // Register every new instance with LLEventPumps
     mRegistry(LLEventPumps::instance().getHandle()),
     mName(mRegistry.get()->registerNew(*this, name, tweak)),
-    mSignal(new LLStandardSignal()),
+    mSignal(boost::make_shared<LLStandardSignal>()),
     mEnabled(true)
 {}
 
@@ -311,6 +313,14 @@ std::string LLEventPump::inventName(const std::string& pfx)
     return STRINGIZE(pfx << suffix++);
 }
 
+void LLEventPump::clear()
+{
+    // Destroy the original LLStandardSignal instance, replacing it with a
+    // whole new one.
+    mSignal = boost::make_shared<LLStandardSignal>();
+    mConnections.clear();
+}
+
 void LLEventPump::reset()
 {
     mSignal.reset();
@@ -553,7 +563,7 @@ bool LLEventMailDrop::post(const LLSD& event)
         // be posted to any future listeners when they attach.
         mEventHistory.push_back(event);
     }
-    
+
     return posted;
 }
 
@@ -583,46 +593,9 @@ LLBoundListener LLEventMailDrop::listen_impl(const std::string& name,
     return LLEventStream::listen_impl(name, listener, after, before);
 }
 
-
-/*****************************************************************************
-*   LLEventQueue
-*****************************************************************************/
-bool LLEventQueue::post(const LLSD& event)
-{
-    if (mEnabled)
-    {
-        // Defer sending this event by queueing it until flush()
-        mEventQueue.push_back(event);
-    }
-    // Unconditionally return false. We won't know until flush() whether a
-    // listener claims to have handled the event -- meanwhile, don't block
-    // other listeners.
-    return false;
-}
-
-void LLEventQueue::flush()
+void LLEventMailDrop::discard()
 {
-	if(!mSignal) return;
-		
-    // Consider the case when a given listener on this LLEventQueue posts yet
-    // another event on the same queue. If we loop over mEventQueue directly,
-    // we'll end up processing all those events during the same flush() call
-    // -- rather like an EventStream. Instead, copy mEventQueue and clear it,
-    // so that any new events posted to this LLEventQueue during flush() will
-    // be processed in the *next* flush() call.
-    EventQueue queue(mEventQueue);
-    mEventQueue.clear();
-    // NOTE NOTE NOTE: Any new access to member data beyond this point should
-    // cause us to move our LLStandardSignal object to a pimpl class along
-    // with said member data. Then the local shared_ptr will preserve both.
-
-    // DEV-43463: capture a local copy of mSignal. See LLEventStream::post()
-    // for detailed comments.
-    boost::shared_ptr<LLStandardSignal> signal(mSignal);
-    for ( ; ! queue.empty(); queue.pop_front())
-    {
-        (*signal)(queue.front());
-    }
+    mEventHistory.clear();
 }
 
 /*****************************************************************************
diff --git a/indra/llcommon/llevents.h b/indra/llcommon/llevents.h
index 62d97007ac..e380c108f4 100644
--- a/indra/llcommon/llevents.h
+++ b/indra/llcommon/llevents.h
@@ -37,6 +37,7 @@
 #include <set>
 #include <vector>
 #include <deque>
+#include <functional>
 #if LL_WINDOWS
 	#pragma warning (push)
 	#pragma warning (disable : 4263) // boost::signals2::expired_slot::what() has const mismatch
@@ -55,7 +56,6 @@
 #include <boost/visit_each.hpp>
 #include <boost/ref.hpp>            // reference_wrapper
 #include <boost/type_traits/is_pointer.hpp>
-#include <boost/function.hpp>
 #include <boost/static_assert.hpp>
 #include "llsd.h"
 #include "llsingleton.h"
@@ -211,8 +211,7 @@ public:
     /// exception if you try to call when empty
     struct Empty: public LLException
     {
-        Empty(const std::string& what):
-            LLException(std::string("LLListenerOrPumpName::Empty: ") + what) {}
+        Empty(const std::string& what): LLException("LLListenerOrPumpName::Empty: " + what) {}
     };
 
 private:
@@ -247,6 +246,30 @@ public:
      */
     LLEventPump& obtain(const std::string& name);
 
+    /// exception potentially thrown by make()
+    struct BadType: public LLException
+    {
+        BadType(const std::string& what): LLException("BadType: " + what) {}
+    };
+
+    /**
+     * Create an LLEventPump with suggested name (optionally of specified
+     * LLEventPump subclass type). As with obtain(), LLEventPumps owns the new
+     * instance.
+     *
+     * As with LLEventPump's constructor, make() could throw
+     * LLEventPump::DupPumpName unless you pass tweak=true.
+     *
+     * As with a hand-constructed LLEventPump subclass, if you pass
+     * tweak=true, the tweaked name can be obtained by LLEventPump::getName().
+     *
+     * Pass empty type to get the default LLEventStream.
+     *
+     * If you pass an unrecognized type string, make() throws BadType.
+     */
+    LLEventPump& make(const std::string& name, bool tweak=false,
+                      const std::string& type=std::string());
+
     /**
      * Find the named LLEventPump instance. If it exists post the message to it.
      * If the pump does not exist, do nothing.
@@ -264,6 +287,11 @@ public:
     void flush();
 
     /**
+     * Disconnect listeners from all known LLEventPump instances
+     */
+    void clear();
+
+    /**
      * Reset all known LLEventPump instances
      * workaround for DEV-35406 crash on shutdown
      */
@@ -298,44 +326,22 @@ testable:
     // destroyed.
     typedef std::set<LLEventPump*> PumpSet;
     PumpSet mOurPumps;
-    // LLEventPump names that should be instantiated as LLEventQueue rather
-    // than as LLEventStream
-    typedef std::set<std::string> PumpNames;
-    PumpNames mQueueNames;
+    // for make(), map string type name to LLEventPump subclass factory function
+    typedef std::map<std::string, std::function<LLEventPump*(const std::string&, bool)>> PumpFactories;
+    // Data used by make().
+    // One might think mFactories and mTypes could reasonably be static. So
+    // they could -- if not for the fact that make() or obtain() might be
+    // called before this module's static variables have been initialized.
+    // This is why we use singletons in the first place.
+    PumpFactories mFactories;
+
+    // for obtain(), map desired string instance name to string type when
+    // obtain() must create the instance
+    typedef std::map<std::string, std::string> InstanceTypes;
+    InstanceTypes mTypes;
 };
 
 /*****************************************************************************
-*   details
-*****************************************************************************/
-namespace LLEventDetail
-{
-    /// Any callable capable of connecting an LLEventListener to an
-    /// LLStandardSignal to produce an LLBoundListener can be mapped to this
-    /// signature.
-    typedef boost::function<LLBoundListener(const LLEventListener&)> ConnectFunc;
-
-    /// overload of visit_and_connect() when we have a string identifier available
-    template <typename LISTENER>
-    LLBoundListener visit_and_connect(const std::string& name,
-                                      const LISTENER& listener,
-                                      const ConnectFunc& connect_func);
-    /**
-     * Utility template function to use Visitor appropriately
-     *
-     * @param listener Callable to connect, typically a boost::bind()
-     * expression. This will be visited by Visitor using boost::visit_each().
-     * @param connect_func Callable that will connect() @a listener to an
-     * LLStandardSignal, returning LLBoundListener.
-     */
-    template <typename LISTENER>
-    LLBoundListener visit_and_connect(const LISTENER& listener,
-                                      const ConnectFunc& connect_func)
-    {
-        return visit_and_connect("", listener, connect_func);
-    }
-} // namespace LLEventDetail
-
-/*****************************************************************************
 *   LLEventTrackable
 *****************************************************************************/
 /**
@@ -369,11 +375,6 @@ namespace LLEventDetail
  *     instance, it attempts to dereference the <tt>Foo*</tt> pointer that was
  *     <tt>delete</tt>d but not zeroed.)
  *   - Undefined behavior results.
- * If you suspect you may encounter any such scenario, you're better off
- * managing the lifespan of your object with <tt>boost::shared_ptr</tt>.
- * Passing <tt>LLEventPump::listen()</tt> a <tt>boost::bind()</tt> expression
- * involving a <tt>boost::weak_ptr<Foo></tt> is recognized specially, engaging
- * thread-safe Boost.Signals2 machinery.
  */
 typedef boost::signals2::trackable LLEventTrackable;
 
@@ -382,7 +383,7 @@ typedef boost::signals2::trackable LLEventTrackable;
 *****************************************************************************/
 /**
  * LLEventPump is the base class interface through which we access the
- * concrete subclasses LLEventStream and LLEventQueue.
+ * concrete subclasses such as LLEventStream.
  *
  * @NOTE
  * LLEventPump derives from LLEventTrackable so that when you "chain"
@@ -403,8 +404,7 @@ public:
      */
     struct DupPumpName: public LLException
     {
-        DupPumpName(const std::string& what):
-            LLException(std::string("DupPumpName: ") + what) {}
+        DupPumpName(const std::string& what): LLException("DupPumpName: " + what) {}
     };
 
     /**
@@ -440,9 +440,7 @@ public:
      */
     struct DupListenerName: public ListenError
     {
-        DupListenerName(const std::string& what):
-            ListenError(std::string("DupListenerName: ") + what)
-        {}
+        DupListenerName(const std::string& what): ListenError("DupListenerName: " + what) {}
     };
     /**
      * exception thrown by listen(). The order dependencies specified for your
@@ -454,7 +452,7 @@ public:
      */
     struct Cycle: public ListenError
     {
-        Cycle(const std::string& what): ListenError(std::string("Cycle: ") + what) {}
+        Cycle(const std::string& what): ListenError("Cycle: " + what) {}
     };
     /**
      * exception thrown by listen(). This one means that your new listener
@@ -475,7 +473,7 @@ public:
      */
     struct OrderChange: public ListenError
     {
-        OrderChange(const std::string& what): ListenError(std::string("OrderChange: ") + what) {}
+        OrderChange(const std::string& what): ListenError("OrderChange: " + what) {}
     };
 
     /// used by listen()
@@ -512,44 +510,13 @@ public:
      * the result be assigned to a LLTempBoundListener or the listener is 
      * manually disconnected when no longer needed since there will be no
      * way to later find and disconnect this listener manually.
-     *
-     * If (as is typical) you pass a <tt>boost::bind()</tt> expression as @a
-     * listener, listen() will inspect the components of that expression. If a
-     * bound object matches any of several cases, the connection will
-     * automatically be disconnected when that object is destroyed.
-     *
-     * * You bind a <tt>boost::weak_ptr</tt>.
-     * * Binding a <tt>boost::shared_ptr</tt> that way would ensure that the
-     *   referenced object would @em never be destroyed, since the @c
-     *   shared_ptr stored in the LLEventPump would remain an outstanding
-     *   reference. Use the weaken() function to convert your @c shared_ptr to
-     *   @c weak_ptr. Because this is easy to forget, binding a @c shared_ptr
-     *   will produce a compile error (@c BOOST_STATIC_ASSERT failure).
-     * * You bind a simple pointer or reference to an object derived from
-     *   <tt>boost::enable_shared_from_this</tt>. (UNDER CONSTRUCTION)
-     * * You bind a simple pointer or reference to an object derived from
-     *   LLEventTrackable. Unlike the cases described above, though, this is
-     *   vulnerable to a couple of cross-thread race conditions, as described
-     *   in the LLEventTrackable documentation.
      */
-    template <typename LISTENER>
-    LLBoundListener listen(const std::string& name, const LISTENER& listener,
+    LLBoundListener listen(const std::string& name,
+                           const LLEventListener& listener,
                            const NameList& after=NameList(),
                            const NameList& before=NameList())
     {
-        // Examine listener, using our listen_impl() method to make the
-        // actual connection.
-        // This is why listen() is a template. Conversion from boost::bind()
-        // to LLEventListener performs type erasure, so it's important to look
-        // at the boost::bind object itself before that happens.
-        return LLEventDetail::visit_and_connect(name,
-                                                listener,
-                                                boost::bind(&LLEventPump::listen_invoke,
-                                                            this,
-                                                            name,
-                                                            _1,
-                                                            after,
-                                                            before));
+        return listen_impl(name, listener, after, before);
     }
 
     /// Get the LLBoundListener associated with the passed name (dummy
@@ -587,19 +554,12 @@ public:
 
 private:
     friend class LLEventPumps;
-
+    virtual void clear();
     virtual void reset();
 
 
 
 private:
-    LLBoundListener listen_invoke(const std::string& name, const LLEventListener& listener,
-        const NameList& after,
-        const NameList& before)
-    {
-        return this->listen_impl(name, listener, after, before);
-    }
-
     // must precede mName; see LLEventPump::LLEventPump()
     LLHandle<LLEventPumps> mRegistry;
 
@@ -663,11 +623,10 @@ public:
  * event *must* eventually reach a listener that will consume it, else the
  * queue will grow to arbitrary length.
  * 
- * @NOTE: When using an LLEventMailDrop (or LLEventQueue) with a LLEventTimeout or
+ * @NOTE: When using an LLEventMailDrop with an LLEventTimeout or
  * LLEventFilter attaching the filter downstream, using Timeout's constructor will
  * cause the MailDrop to discharge any of its stored events. The timeout should 
  * instead be connected upstream using its listen() method.  
- * See llcoro::suspendUntilEventOnWithTimeout() for an example.
  */
 class LL_COMMON_API LLEventMailDrop : public LLEventStream
 {
@@ -679,7 +638,8 @@ public:
     virtual bool post(const LLSD& event) override;
     
     /// Remove any history stored in the mail drop.
-    virtual void flush() override { mEventHistory.clear(); LLEventStream::flush(); };
+    void discard();
+
 protected:
     virtual LLBoundListener listen_impl(const std::string& name, const LLEventListener&,
                                         const NameList& after,
@@ -691,30 +651,6 @@ private:
 };
 
 /*****************************************************************************
-*   LLEventQueue
-*****************************************************************************/
-/**
- * LLEventQueue is a LLEventPump whose post() method defers calling registered
- * listeners until flush() is called.
- */
-class LL_COMMON_API LLEventQueue: public LLEventPump
-{
-public:
-    LLEventQueue(const std::string& name, bool tweak=false): LLEventPump(name, tweak) {}
-    virtual ~LLEventQueue() {}
-
-    /// Post an event to all listeners
-    virtual bool post(const LLSD& event);
-
-    /// flush queued events
-    virtual void flush();
-
-private:
-    typedef std::deque<LLSD> EventQueue;
-    EventQueue mEventQueue;
-};
-
-/*****************************************************************************
 *   LLReqID
 *****************************************************************************/
 /**
@@ -809,329 +745,6 @@ private:
 LL_COMMON_API bool sendReply(const LLSD& reply, const LLSD& request,
                              const std::string& replyKey="reply");
 
-/**
- * Base class for LLListenerWrapper. See visit_and_connect() and llwrap(). We
- * provide virtual @c accept_xxx() methods, customization points allowing a
- * subclass access to certain data visible at LLEventPump::listen() time.
- * Example subclass usage:
- *
- * @code
- * myEventPump.listen("somename",
- *                    llwrap<MyListenerWrapper>(boost::bind(&MyClass::method, instance, _1)));
- * @endcode
- *
- * Because of the anticipated usage (note the anonymous temporary
- * MyListenerWrapper instance in the example above), the @c accept_xxx()
- * methods must be @c const.
- */
-class LL_COMMON_API LLListenerWrapperBase
-{
-public:
-    /// New instance. The accept_xxx() machinery makes it important to use
-    /// shared_ptrs for our data. Many copies of this object are made before
-    /// the instance that actually ends up in the signal, yet accept_xxx()
-    /// will later be called on the @em original instance. All copies of the
-    /// same original instance must share the same data.
-    LLListenerWrapperBase():
-        mName(new std::string),
-        mConnection(new LLBoundListener)
-    {
-    }
-
-    /// Copy constructor. Copy shared_ptrs to original instance data.
-    LLListenerWrapperBase(const LLListenerWrapperBase& that):
-        mName(that.mName),
-        mConnection(that.mConnection)
-    {
-    }
-    virtual ~LLListenerWrapperBase() {}
-
-    /// Ask LLEventPump::listen() for the listener name
-    virtual void accept_name(const std::string& name) const
-    {
-        *mName = name;
-    }
-
-    /// Ask LLEventPump::listen() for the new connection
-    virtual void accept_connection(const LLBoundListener& connection) const
-    {
-        *mConnection = connection;
-    }
-
-protected:
-    /// Listener name.
-    boost::shared_ptr<std::string> mName;
-    /// Connection.
-    boost::shared_ptr<LLBoundListener> mConnection;
-};
-
-/*****************************************************************************
-*   Underpinnings
-*****************************************************************************/
-/**
- * We originally provided a suite of overloaded
- * LLEventTrackable::listenTo(LLEventPump&, ...) methods that would call
- * LLEventPump::listen(...) and then pass the returned LLBoundListener to
- * LLEventTrackable::track(). This was workable but error-prone: the coder
- * must remember to call listenTo() rather than the more straightforward
- * listen() method.
- *
- * Now we publish only the single canonical listen() method, so there's a
- * uniform mechanism. Having a single way to do this is good, in that there's
- * no question in the coder's mind which of several alternatives to choose.
- *
- * To support automatic connection management, we use boost::visit_each
- * (http://www.boost.org/doc/libs/1_37_0/doc/html/boost/visit_each.html) to
- * inspect each argument of a boost::bind expression. (Although the visit_each
- * mechanism was first introduced with the original Boost.Signals library, it
- * was only later documented.)
- *
- * Cases:
- * * At least one of the function's arguments is a boost::weak_ptr<T>. Pass
- *   the corresponding shared_ptr to slot_type::track(). Ideally that would be
- *   the object whose method we want to call, but in fact we do the same for
- *   any weak_ptr we might find among the bound arguments. If we're passing
- *   our bound method a weak_ptr to some object, wouldn't the destruction of
- *   that object invalidate the call? So we disconnect automatically when any
- *   such object is destroyed. This is the mechanism preferred by boost::
- *   signals2.
- * * One of the functions's arguments is a boost::shared_ptr<T>. This produces
- *   a compile error: the bound copy of the shared_ptr stored in the
- *   boost_bind object stored in the signal object would make the referenced
- *   T object immortal. We provide a weaken() function. Pass
- *   weaken(your_shared_ptr) instead. (We can inspect, but not modify, the
- *   boost::bind object. Otherwise we'd replace the shared_ptr with weak_ptr
- *   implicitly and just proceed.)
- * * One of the function's arguments is a plain pointer/reference to an object
- *   derived from boost::enable_shared_from_this. We assume that this object
- *   is managed using boost::shared_ptr, so we implicitly extract a shared_ptr
- *   and track that. (UNDER CONSTRUCTION)
- * * One of the function's arguments is derived from LLEventTrackable. Pass
- *   the LLBoundListener to its LLEventTrackable::track(). This is vulnerable
- *   to a couple different race conditions, as described in LLEventTrackable
- *   documentation. (NOTE: Now that LLEventTrackable is a typedef for
- *   boost::signals2::trackable, the Signals2 library handles this itself, so
- *   our visitor needs no special logic for this case.)
- * * Any other argument type is irrelevant to automatic connection management.
- */
-
-namespace LLEventDetail
-{
-    template <typename F>
-    const F& unwrap(const F& f) { return f; }
-
-    template <typename F>
-    const F& unwrap(const boost::reference_wrapper<F>& f) { return f.get(); }
-
-    // Most of the following is lifted from the Boost.Signals use of
-    // visit_each.
-    template<bool Cond> struct truth {};
-
-    /**
-     * boost::visit_each() Visitor, used on a template argument <tt>const F&
-     * f</tt> as follows (see visit_and_connect()):
-     * @code
-     * LLEventListener listener(f);
-     * Visitor visitor(listener); // bind listener so it can track() shared_ptrs
-     * using boost::visit_each;   // allow unqualified visit_each() call for ADL
-     * visit_each(visitor, unwrap(f));
-     * @endcode
-     */
-    class Visitor
-    {
-    public:
-        /**
-         * Visitor binds a reference to LLEventListener so we can track() any
-         * shared_ptrs we find in the argument list.
-         */
-        Visitor(LLEventListener& listener):
-            mListener(listener)
-        {
-        }
-
-        /**
-         * boost::visit_each() calls this method for each component of a
-         * boost::bind() expression.
-         */
-        template <typename T>
-        void operator()(const T& t) const
-        {
-            decode(t, 0);
-        }
-
-    private:
-        // decode() decides between a reference wrapper and anything else
-        // boost::ref() variant
-        template<typename T>
-        void decode(const boost::reference_wrapper<T>& t, int) const
-        {
-//          add_if_trackable(t.get_pointer());
-        }
-
-        // decode() anything else
-        template<typename T>
-        void decode(const T& t, long) const
-        {
-            typedef truth<(boost::is_pointer<T>::value)> is_a_pointer;
-            maybe_get_pointer(t, is_a_pointer());
-        }
-
-        // maybe_get_pointer() decides between a pointer and a non-pointer
-        // plain pointer variant
-        template<typename T>
-        void maybe_get_pointer(const T& t, truth<true>) const
-        {
-//          add_if_trackable(t);
-        }
-
-        // shared_ptr variant
-        template<typename T>
-        void maybe_get_pointer(const boost::shared_ptr<T>& t, truth<false>) const
-        {
-            // If we have a shared_ptr to this object, it doesn't matter
-            // whether the object is derived from LLEventTrackable, so no
-            // further analysis of T is needed.
-//          mListener.track(t);
-
-            // Make this case illegal. Passing a bound shared_ptr to
-            // slot_type::track() is useless, since the bound shared_ptr will
-            // keep the object alive anyway! Force the coder to cast to weak_ptr.
-
-            // Trivial as it is, make the BOOST_STATIC_ASSERT() condition
-            // dependent on template param so the macro is only evaluated if
-            // this method is in fact instantiated, as described here:
-            // http://www.boost.org/doc/libs/1_34_1/doc/html/boost_staticassert.html
-
-            // ATTENTION: Don't bind a shared_ptr<anything> using
-            // LLEventPump::listen(boost::bind()). Doing so captures a copy of
-            // the shared_ptr, making the referenced object effectively
-            // immortal. Use the weaken() function, e.g.:
-            // somepump.listen(boost::bind(...weaken(my_shared_ptr)...));
-            // This lets us automatically disconnect when the referenced
-            // object is destroyed.
-            BOOST_STATIC_ASSERT(sizeof(T) == 0);
-        }
-
-        // weak_ptr variant
-        template<typename T>
-        void maybe_get_pointer(const boost::weak_ptr<T>& t, truth<false>) const
-        {
-            // If we have a weak_ptr to this object, it doesn't matter
-            // whether the object is derived from LLEventTrackable, so no
-            // further analysis of T is needed.
-            mListener.track(t);
-//          std::cout << "Found weak_ptr<" << typeid(T).name() << ">!\n";
-        }
-
-#if 0
-        // reference to anything derived from boost::enable_shared_from_this
-        template <typename T>
-        inline void maybe_get_pointer(const boost::enable_shared_from_this<T>& ct,
-                                      truth<false>) const
-        {
-            // Use the slot_type::track(shared_ptr) mechanism. Cast away
-            // const-ness because (in our code base anyway) it's unusual
-            // to find shared_ptr<const T>.
-            boost::enable_shared_from_this<T>&
-                t(const_cast<boost::enable_shared_from_this<T>&>(ct));
-            std::cout << "Capturing shared_from_this()" << std::endl;
-            boost::shared_ptr<T> sp(t.shared_from_this());
-/*==========================================================================*|
-            std::cout << "Capturing weak_ptr" << std::endl;
-            boost::weak_ptr<T> wp(sp);
-|*==========================================================================*/
-            std::cout << "Tracking shared__ptr" << std::endl;
-            mListener.track(sp);
-        }
-#endif
-
-        // non-pointer variant
-        template<typename T>
-        void maybe_get_pointer(const T& t, truth<false>) const
-        {
-            // Take the address of this object, because the object itself may be
-            // trackable
-//          add_if_trackable(boost::addressof(t));
-        }
-
-/*==========================================================================*|
-        // add_if_trackable() adds LLEventTrackable objects to mTrackables
-        inline void add_if_trackable(const LLEventTrackable* t) const
-        {
-            if (t)
-            {
-            }
-        }
-
-        // pointer to anything not an LLEventTrackable subclass
-        inline void add_if_trackable(const void*) const
-        {
-        }
-
-        // pointer to free function
-        // The following construct uses the preprocessor to generate
-        // add_if_trackable() overloads accepting pointer-to-function taking
-        // 0, 1, ..., LLEVENTS_LISTENER_ARITY parameters of arbitrary type.
-#define BOOST_PP_LOCAL_MACRO(n)                                     \
-        template <typename R                                        \
-                  BOOST_PP_COMMA_IF(n)                              \
-                  BOOST_PP_ENUM_PARAMS(n, typename T)>              \
-        inline void                                                 \
-        add_if_trackable(R (*)(BOOST_PP_ENUM_PARAMS(n, T))) const   \
-        {                                                           \
-        }
-#define BOOST_PP_LOCAL_LIMITS (0, LLEVENTS_LISTENER_ARITY)
-#include BOOST_PP_LOCAL_ITERATE()
-#undef  BOOST_PP_LOCAL_MACRO
-#undef  BOOST_PP_LOCAL_LIMITS
-|*==========================================================================*/
-
-        /// Bind a reference to the LLEventListener to call its track() method.
-        LLEventListener& mListener;
-    };
-
-    /**
-     * Utility template function to use Visitor appropriately
-     *
-     * @param raw_listener Callable to connect, typically a boost::bind()
-     * expression. This will be visited by Visitor using boost::visit_each().
-     * @param connect_funct Callable that will connect() @a raw_listener to an
-     * LLStandardSignal, returning LLBoundListener.
-     */
-    template <typename LISTENER>
-    LLBoundListener visit_and_connect(const std::string& name,
-                                      const LISTENER& raw_listener,
-                                      const ConnectFunc& connect_func)
-    {
-        // Capture the listener
-        LLEventListener listener(raw_listener);
-        // Define our Visitor, binding the listener so we can call
-        // listener.track() if we discover any shared_ptr<Foo>.
-        LLEventDetail::Visitor visitor(listener);
-        // Allow unqualified visit_each() call for ADL
-        using boost::visit_each;
-        // Visit each component of a boost::bind() expression. Pass
-        // 'raw_listener', our template argument, rather than 'listener' from
-        // which type details have been erased. unwrap() comes from
-        // Boost.Signals, in case we were passed a boost::ref().
-        visit_each(visitor, LLEventDetail::unwrap(raw_listener));
-        // Make the connection using passed function.
-        LLBoundListener connection(connect_func(listener));
-        // If the LISTENER is an LLListenerWrapperBase subclass, pass it the
-        // desired information. It's important that we pass the raw_listener
-        // so the compiler can make decisions based on its original type.
-        const LLListenerWrapperBase* lwb =
-            ll_template_cast<const LLListenerWrapperBase*>(&raw_listener);
-        if (lwb)
-        {
-            lwb->accept_name(name);
-            lwb->accept_connection(connection);
-        }
-        // In any case, show new connection to caller.
-        return connection;
-    }
-} // namespace LLEventDetail
-
 // Somewhat to my surprise, passing boost::bind(...boost::weak_ptr<T>...) to
 // listen() fails in Boost code trying to instantiate LLEventListener (i.e.
 // LLStandardSignal::slot_type) because the boost::get_pointer() utility function isn't
@@ -1142,12 +755,4 @@ namespace boost
     T* get_pointer(const weak_ptr<T>& ptr) { return shared_ptr<T>(ptr).get(); }
 }
 
-/// Since we forbid use of listen(boost::bind(...shared_ptr<T>...)), provide an
-/// easy way to cast to the corresponding weak_ptr.
-template <typename T>
-boost::weak_ptr<T> weaken(const boost::shared_ptr<T>& ptr)
-{
-    return boost::weak_ptr<T>(ptr);
-}
-
 #endif /* ! defined(LL_LLEVENTS_H) */
diff --git a/indra/llcommon/lleventtimer.cpp b/indra/llcommon/lleventtimer.cpp
index 0d96e03da4..f575a7b6bf 100644
--- a/indra/llcommon/lleventtimer.cpp
+++ b/indra/llcommon/lleventtimer.cpp
@@ -57,29 +57,17 @@ LLEventTimer::~LLEventTimer()
 //static
 void LLEventTimer::updateClass() 
 {
-	std::list<LLEventTimer*> completed_timers;
-	for (instance_iter iter = beginInstances(); iter != endInstances(); ) 
+	for (auto& timer : instance_snapshot())
 	{
-		LLEventTimer& timer = *iter++;
 		F32 et = timer.mEventTimer.getElapsedTimeF32();
 		if (timer.mEventTimer.getStarted() && et > timer.mPeriod) {
 			timer.mEventTimer.reset();
 			if ( timer.tick() )
 			{
-				completed_timers.push_back( &timer );
+				delete &timer;
 			}
 		}
 	}
-
-	if ( completed_timers.size() > 0 )
-	{
-		for (std::list<LLEventTimer*>::iterator completed_iter = completed_timers.begin(); 
-			 completed_iter != completed_timers.end(); 
-			 completed_iter++ ) 
-		{
-			delete *completed_iter;
-		}
-	}
 }
 
 
diff --git a/indra/llcommon/lleventtimer.h b/indra/llcommon/lleventtimer.h
index dc918121e1..dbbfe0c6e6 100644
--- a/indra/llcommon/lleventtimer.h
+++ b/indra/llcommon/lleventtimer.h
@@ -40,16 +40,83 @@ public:
 	LLEventTimer(F32 period);	// period is the amount of time between each call to tick() in seconds
 	LLEventTimer(const LLDate& time);
 	virtual ~LLEventTimer();
-	
+
 	//function to be called at the supplied frequency
 	// Normally return FALSE; TRUE will delete the timer after the function returns.
 	virtual BOOL tick() = 0;
 
 	static void updateClass();
 
+	/// Schedule recurring calls to generic callable every period seconds.
+	/// Returns a pointer; if you delete it, cancels the recurring calls.
+	template <typename CALLABLE>
+	static LLEventTimer* run_every(F32 period, const CALLABLE& callable);
+
+	/// Schedule a future call to generic callable. Returns a pointer.
+	/// CAUTION: The object referenced by that pointer WILL BE DELETED once
+	/// the callback has been called! LLEventTimer::getInstance(pointer) (NOT
+	/// pointer->getInstance(pointer)!) can be used to test whether the
+	/// pointer is still valid. If it is, deleting it will cancel the
+	/// callback.
+	template <typename CALLABLE>
+	static LLEventTimer* run_at(const LLDate& time, const CALLABLE& callable);
+
+	/// Like run_at(), but after a time delta rather than at a timestamp.
+	/// Same CAUTION.
+	template <typename CALLABLE>
+	static LLEventTimer* run_after(F32 interval, const CALLABLE& callable);
+
 protected:
 	LLTimer mEventTimer;
 	F32 mPeriod;
+
+private:
+	template <typename CALLABLE>
+	class Generic;
+};
+
+template <typename CALLABLE>
+class LLEventTimer::Generic: public LLEventTimer
+{
+public:
+    // making TIME generic allows engaging either LLEventTimer constructor
+    template <typename TIME>
+    Generic(const TIME& time, bool once, const CALLABLE& callable):
+        LLEventTimer(time),
+        mOnce(once),
+        mCallable(callable)
+    {}
+    BOOL tick() override
+    {
+        mCallable();
+        // true tells updateClass() to delete this instance
+        return mOnce;
+    }
+
+private:
+    bool mOnce;
+    CALLABLE mCallable;
 };
 
+template <typename CALLABLE>
+LLEventTimer* LLEventTimer::run_every(F32 period, const CALLABLE& callable)
+{
+    // return false to schedule recurring calls
+    return new Generic<CALLABLE>(period, false, callable);
+}
+
+template <typename CALLABLE>
+LLEventTimer* LLEventTimer::run_at(const LLDate& time, const CALLABLE& callable)
+{
+    // return true for one-shot callback
+    return new Generic<CALLABLE>(time, true, callable);
+}
+
+template <typename CALLABLE>
+LLEventTimer* LLEventTimer::run_after(F32 interval, const CALLABLE& callable)
+{
+    // one-shot callback after specified interval
+    return new Generic<CALLABLE>(interval, true, callable);
+}
+
 #endif //LL_EVENTTIMER_H
diff --git a/indra/llcommon/llexception.cpp b/indra/llcommon/llexception.cpp
index b32ec2c9c9..5ce8958687 100644
--- a/indra/llcommon/llexception.cpp
+++ b/indra/llcommon/llexception.cpp
@@ -18,10 +18,28 @@
 #include <typeinfo>
 // external library headers
 #include <boost/exception/diagnostic_information.hpp>
+#include <boost/exception/error_info.hpp>
+// On Mac, got:
+// #error "Boost.Stacktrace requires `_Unwind_Backtrace` function. Define
+// `_GNU_SOURCE` macro or `BOOST_STACKTRACE_GNU_SOURCE_NOT_REQUIRED` if
+// _Unwind_Backtrace is available without `_GNU_SOURCE`."
+#define BOOST_STACKTRACE_GNU_SOURCE_NOT_REQUIRED
+#if LL_WINDOWS
+// On Windows, header-only implementation causes macro collisions -- use
+// prebuilt library
+#define BOOST_STACKTRACE_LINK
+#endif // LL_WINDOWS
+#include <boost/stacktrace.hpp>
 // other Linden headers
 #include "llerror.h"
 #include "llerrorcontrol.h"
 
+// used to attach and extract stacktrace information to/from boost::exception,
+// see https://www.boost.org/doc/libs/release/doc/html/stacktrace/getting_started.html#stacktrace.getting_started.exceptions_with_stacktrace
+// apparently the struct passed as the first template param needs no definition?
+typedef boost::error_info<struct errinfo_stacktrace_, boost::stacktrace::stacktrace>
+        errinfo_stacktrace;
+
 namespace {
 // used by crash_on_unhandled_exception_() and log_unhandled_exception_()
 void log_unhandled_exception_(LLError::ELevel level,
@@ -53,3 +71,17 @@ void log_unhandled_exception_(const char* file, int line, const char* pretty_fun
     // routinely, but we DO expect to return from this function.
     log_unhandled_exception_(LLError::LEVEL_WARN, file, line, pretty_function, context);
 }
+
+void annotate_exception_(boost::exception& exc)
+{
+    // https://www.boost.org/doc/libs/release/libs/exception/doc/tutorial_transporting_data.html
+    // "Adding of Arbitrary Data to Active Exception Objects"
+    // Given a boost::exception&, we can add boost::error_info items to it
+    // without knowing its leaf type.
+    // The stacktrace constructor that lets us skip a level -- and why would
+    // we always include annotate_exception_()? -- also requires a max depth.
+    // For the nullary constructor, the stacktrace class declaration itself
+    // passes static_cast<std::size_t>(-1), but that's kind of dubious.
+    // Anyway, which of us is really going to examine more than 100 frames?
+    exc << errinfo_stacktrace(boost::stacktrace::stacktrace(1, 100));
+}
diff --git a/indra/llcommon/llexception.h b/indra/llcommon/llexception.h
index dfcb7c192f..422dd8810a 100644
--- a/indra/llcommon/llexception.h
+++ b/indra/llcommon/llexception.h
@@ -67,9 +67,29 @@ struct LLContinueError: public LLException
  * enriches the exception's diagnostic_information() with the source file,
  * line and containing function of the LLTHROW() macro.
  */
-// Currently we implement that using BOOST_THROW_EXCEPTION(). Wrap it in
-// LLTHROW() in case we ever want to revisit that implementation decision.
-#define LLTHROW(x) BOOST_THROW_EXCEPTION(x)
+#define LLTHROW(x)                                                      \
+do {                                                                    \
+    /* Capture the exception object 'x' by value. (Exceptions must */   \
+    /* be copyable.) It might seem simpler to use                  */   \
+    /* BOOST_THROW_EXCEPTION(annotate_exception_(x)) instead of    */   \
+    /* three separate statements, but:                             */   \
+    /* - We want to throw 'x' with its original type, not just a   */   \
+    /*   reference to boost::exception.                            */   \
+    /* - To return x's original type, annotate_exception_() would  */   \
+    /*   have to be a template function.                           */   \
+    /* - We want annotate_exception_() to be opaque.               */   \
+    /* We also might consider embedding BOOST_THROW_EXCEPTION() in */   \
+    /* our helper function, but we want the filename and line info */   \
+    /* embedded by BOOST_THROW_EXCEPTION() to be the throw point   */   \
+    /* rather than always indicating the same line in              */   \
+    /* llexception.cpp.                                            */   \
+    auto exc{x};                                                        \
+    annotate_exception_(exc);                                           \
+    BOOST_THROW_EXCEPTION(exc);                                         \
+    /* Use the classic 'do { ... } while (0)' macro trick to wrap  */   \
+    /* our multiple statements.                                    */   \
+} while (0)
+void annotate_exception_(boost::exception& exc);
 
 /// Call this macro from a catch (...) clause
 #define CRASH_ON_UNHANDLED_EXCEPTION(CONTEXT) \
diff --git a/indra/llcommon/llfasttimer.cpp b/indra/llcommon/llfasttimer.cpp
index 3d28cd15b0..08ea668964 100644
--- a/indra/llcommon/llfasttimer.cpp
+++ b/indra/llcommon/llfasttimer.cpp
@@ -193,27 +193,26 @@ TimeBlockTreeNode& BlockTimerStatHandle::getTreeNode() const
 
 void BlockTimer::bootstrapTimerTree()
 {
-	for (BlockTimerStatHandle::instance_tracker_t::instance_iter it = BlockTimerStatHandle::instance_tracker_t::beginInstances(), end_it = BlockTimerStatHandle::instance_tracker_t::endInstances(); 
-		it != end_it; 
-		++it)
+	for (auto& base : BlockTimerStatHandle::instance_snapshot())
 	{
-		BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(*it);
+		// because of indirect derivation from LLInstanceTracker, have to downcast
+		BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(base);
 		if (&timer == &BlockTimer::getRootTimeBlock()) continue;
 
 		// bootstrap tree construction by attaching to last timer to be on stack
 		// when this timer was called
 		if (timer.getParent() == &BlockTimer::getRootTimeBlock())
-{
+		{
 			TimeBlockAccumulator& accumulator = timer.getCurrentAccumulator();
 
 			if (accumulator.mLastCaller)
-	{
+			{
 				timer.setParent(accumulator.mLastCaller);
 				accumulator.mParent = accumulator.mLastCaller;
-		}
+			}
 			// no need to push up tree on first use, flag can be set spuriously
 			accumulator.mMoveUpTree = false;
-	}
+		}
 	}
 }
 
@@ -306,12 +305,10 @@ void BlockTimer::processTimes()
 	updateTimes();
 
 	// reset for next frame
-	for (BlockTimerStatHandle::instance_tracker_t::instance_iter it = BlockTimerStatHandle::instance_tracker_t::beginInstances(),
-			end_it = BlockTimerStatHandle::instance_tracker_t::endInstances();
-		it != end_it;
-		++it)
+	for (auto& base : BlockTimerStatHandle::instance_snapshot())
 	{
-		BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(*it);
+		// because of indirect derivation from LLInstanceTracker, have to downcast
+		BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(base);
 		TimeBlockAccumulator& accumulator = timer.getCurrentAccumulator();
 
 		accumulator.mLastCaller = NULL;
@@ -362,12 +359,10 @@ void BlockTimer::logStats()
 		LLSD sd;
 
 		{
-			for (BlockTimerStatHandle::instance_tracker_t::instance_iter it = BlockTimerStatHandle::instance_tracker_t::beginInstances(), 
-				end_it = BlockTimerStatHandle::instance_tracker_t::endInstances(); 
-				it != end_it; 
-			++it)
+			for (auto& base : BlockTimerStatHandle::instance_snapshot())
 			{
-				BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(*it);
+				// because of indirect derivation from LLInstanceTracker, have to downcast
+				BlockTimerStatHandle& timer = static_cast<BlockTimerStatHandle&>(base);
 				LLTrace::PeriodicRecording& frame_recording = LLTrace::get_frame_recording();
 				sd[timer.getName()]["Time"] = (LLSD::Real) (frame_recording.getLastRecording().getSum(timer).value());	
 				sd[timer.getName()]["Calls"] = (LLSD::Integer) (frame_recording.getLastRecording().getSum(timer.callCount()));
diff --git a/indra/llcommon/llfasttimer.h b/indra/llcommon/llfasttimer.h
index d463fc9d65..5628a05b00 100644
--- a/indra/llcommon/llfasttimer.h
+++ b/indra/llcommon/llfasttimer.h
@@ -31,6 +31,10 @@
 #include "lltrace.h"
 #include "lltreeiterators.h"
 
+#if LL_WINDOWS
+#include <intrin.h>
+#endif
+
 #define LL_FAST_TIMER_ON 1
 #define LL_FASTTIMER_USE_RDTSC 1
 
@@ -85,6 +89,8 @@ public:
 	//	return __rdtsc();
 	//}
 
+	
+
 	// shift off lower 8 bits for lower resolution but longer term timing
 	// on 1Ghz machine, a 32-bit word will hold ~1000 seconds of timing
 #if LL_FASTTIMER_USE_RDTSC
diff --git a/indra/llcommon/llfile.h b/indra/llcommon/llfile.h
index 398938b729..9de095b45d 100644
--- a/indra/llcommon/llfile.h
+++ b/indra/llcommon/llfile.h
@@ -86,6 +86,69 @@ public:
 	static  const char * tmpdir();
 };
 
+/// RAII class
+class LLUniqueFile
+{
+public:
+    // empty
+    LLUniqueFile(): mFileHandle(nullptr) {}
+    // wrap (e.g.) result of LLFile::fopen()
+    LLUniqueFile(LLFILE* f): mFileHandle(f) {}
+    // no copy
+    LLUniqueFile(const LLUniqueFile&) = delete;
+    // move construction
+    LLUniqueFile(LLUniqueFile&& other)
+    {
+        mFileHandle = other.mFileHandle;
+        other.mFileHandle = nullptr;
+    }
+    // The point of LLUniqueFile is to close on destruction.
+    ~LLUniqueFile()
+    {
+        close();
+    }
+
+    // simple assignment
+    LLUniqueFile& operator=(LLFILE* f)
+    {
+        close();
+        mFileHandle = f;
+        return *this;
+    }
+    // copy assignment deleted
+    LLUniqueFile& operator=(const LLUniqueFile&) = delete;
+    // move assignment
+    LLUniqueFile& operator=(LLUniqueFile&& other)
+    {
+        close();
+        std::swap(mFileHandle, other.mFileHandle);
+        return *this;
+    }
+
+    // explicit close operation
+    void close()
+    {
+        if (mFileHandle)
+        {
+            // in case close() throws, set mFileHandle null FIRST
+            LLFILE* h{nullptr};
+            std::swap(h, mFileHandle);
+            LLFile::close(h);
+        }
+    }
+
+    // detect whether the wrapped LLFILE is open or not
+    explicit operator bool() const { return bool(mFileHandle); }
+    bool operator!() { return ! mFileHandle; }
+
+    // LLUniqueFile should be usable for any operation that accepts LLFILE*
+    // (or FILE* for that matter)
+    operator LLFILE*() const { return mFileHandle; }
+
+private:
+    LLFILE* mFileHandle;
+};
+
 #if LL_WINDOWS
 /**
  *  @brief  Controlling input for files.
diff --git a/indra/llcommon/llinstancetracker.cpp b/indra/llcommon/llinstancetracker.cpp
index 3f990f4869..e7193b70b5 100644
--- a/indra/llcommon/llinstancetracker.cpp
+++ b/indra/llcommon/llinstancetracker.cpp
@@ -27,25 +27,15 @@
 #include "linden_common.h"
 // associated header
 #include "llinstancetracker.h"
-#include "llapr.h"
-
+#include "llerror.h"
 // STL headers
 // std headers
 // external library headers
 // other Linden headers
 
-void LLInstanceTrackerBase::StaticBase::incrementDepth()
-{
-	++sIterationNestDepth;
-}
-
-void LLInstanceTrackerBase::StaticBase::decrementDepth()
-{
-	llassert(sIterationNestDepth);
-	--sIterationNestDepth;
-}
-
-U32 LLInstanceTrackerBase::StaticBase::getDepth()
+void LLInstanceTrackerPrivate::logerrs(const char* cls, const std::string& arg1,
+                                       const std::string& arg2, const std::string& arg3)
 {
-	return sIterationNestDepth;
+    LL_ERRS("LLInstanceTracker") << LLError::Log::demangle(cls)
+                                 << arg1 << arg2 << arg3 << LL_ENDL;
 }
diff --git a/indra/llcommon/llinstancetracker.h b/indra/llcommon/llinstancetracker.h
index 363d0bcbd5..402333cca7 100644
--- a/indra/llcommon/llinstancetracker.h
+++ b/indra/llcommon/llinstancetracker.h
@@ -28,354 +28,432 @@
 #ifndef LL_LLINSTANCETRACKER_H
 #define LL_LLINSTANCETRACKER_H
 
-#include <atomic>
 #include <map>
+#include <set>
+#include <vector>
 #include <typeinfo>
+#include <memory>
+#include <type_traits>
+
+#include "mutex.h"
 
-#include "llstringtable.h"
 #include <boost/iterator/transform_iterator.hpp>
 #include <boost/iterator/indirect_iterator.hpp>
+#include <boost/iterator/filter_iterator.hpp>
 
-// As of 2017-05-06, as far as nat knows, only clang supports __has_feature().
-// Unfortunately VS2013's preprocessor shortcut logic doesn't prevent it from
-// producing (fatal) warnings for defined(__clang__) && __has_feature(...).
-// Have to work around that.
-#if ! defined(__clang__)
-#define __has_feature(x) 0
-#endif // __clang__
-
-#if defined(LL_TEST_llinstancetracker) && __has_feature(cxx_noexcept)
-// ~LLInstanceTracker() performs llassert_always() validation. That's fine in
-// production code, since the llassert_always() is implemented as an LL_ERRS
-// message, which will crash-with-message. In our integration test executable,
-// though, this llassert_always() throws an exception instead so we can test
-// error conditions and continue running the test. However -- as of C++11,
-// destructors are implicitly noexcept(true). Unless we mark
-// ~LLInstanceTracker() noexcept(false), the test executable crashes even on
-// the ATTEMPT to throw.
-#define LLINSTANCETRACKER_DTOR_NOEXCEPT noexcept(false)
-#else
-// If we're building for production, or in fact building *any other* test, or
-// we're using a compiler that doesn't support __has_feature(), or we're not
-// compiling with a C++ version that supports noexcept -- don't specify it.
-#define LLINSTANCETRACKER_DTOR_NOEXCEPT
-#endif
+#include "lockstatic.h"
+#include "stringize.h"
 
-/**
- * Base class manages "class-static" data that must actually have singleton
- * semantics: one instance per process, rather than one instance per module as
- * sometimes happens with data simply declared static.
- */
-class LL_COMMON_API LLInstanceTrackerBase
+/*****************************************************************************
+*   StaticBase
+*****************************************************************************/
+namespace LLInstanceTrackerPrivate
 {
-protected:
-    /// It's not essential to derive your STATICDATA (for use with
-    /// getStatic()) from StaticBase; it's just that both known
-    /// implementations do.
     struct StaticBase
     {
-        StaticBase():
-            sIterationNestDepth(0)
-        {}
-
-		void incrementDepth();
-		void decrementDepth();
-		U32 getDepth();
-	private:
-#ifdef LL_WINDOWS
-		std::atomic_uint32_t sIterationNestDepth;
-#else
-		std::atomic_uint sIterationNestDepth;
-#endif
-	};
-};
+        // We need to be able to lock static data while manipulating it.
+        std::mutex mMutex;
+    };
 
-LL_COMMON_API void assert_main_thread();
+    void logerrs(const char* cls, const std::string&, const std::string&, const std::string&);
+} // namespace LLInstanceTrackerPrivate
 
+/*****************************************************************************
+*   LLInstanceTracker with key
+*****************************************************************************/
 enum EInstanceTrackerAllowKeyCollisions
 {
-	LLInstanceTrackerErrorOnCollision,
-	LLInstanceTrackerReplaceOnCollision
+    LLInstanceTrackerErrorOnCollision,
+    LLInstanceTrackerReplaceOnCollision
 };
 
 /// This mix-in class adds support for tracking all instances of the specified class parameter T
 /// The (optional) key associates a value of type KEY with a given instance of T, for quick lookup
 /// If KEY is not provided, then instances are stored in a simple set
 /// @NOTE: see explicit specialization below for default KEY==void case
-/// @NOTE: this class is not thread-safe unless used as read-only
-template<typename T, typename KEY = void, EInstanceTrackerAllowKeyCollisions KEY_COLLISION_BEHAVIOR = LLInstanceTrackerErrorOnCollision>
-class LLInstanceTracker : public LLInstanceTrackerBase
+template<typename T, typename KEY = void,
+         EInstanceTrackerAllowKeyCollisions KEY_COLLISION_BEHAVIOR = LLInstanceTrackerErrorOnCollision>
+class LLInstanceTracker
 {
-	typedef LLInstanceTracker<T, KEY> self_t;
-	typedef typename std::multimap<KEY, T*> InstanceMap;
-	struct StaticData: public StaticBase
-	{
-		InstanceMap sMap;
-	};
-	static StaticData& getStatic() { static StaticData sData; return sData;}
-	static InstanceMap& getMap_() { return getStatic().sMap; }
+    typedef std::map<KEY, std::shared_ptr<T>> InstanceMap;
+    struct StaticData: public LLInstanceTrackerPrivate::StaticBase
+    {
+        InstanceMap mMap;
+    };
+    typedef llthread::LockStatic<StaticData> LockStatic;
 
 public:
-	class instance_iter : public boost::iterator_facade<instance_iter, T, boost::forward_traversal_tag>
-	{
-	public:
-		typedef boost::iterator_facade<instance_iter, T, boost::forward_traversal_tag> super_t;
-		
-		instance_iter(const typename InstanceMap::iterator& it)
-		:	mIterator(it)
-		{
-			getStatic().incrementDepth();
-		}
-
-		~instance_iter()
-		{
-			getStatic().decrementDepth();
-		}
-
-
-	private:
-		friend class boost::iterator_core_access;
-
-		void increment() { mIterator++; }
-		bool equal(instance_iter const& other) const
-		{
-			return mIterator == other.mIterator;
-		}
-
-		T& dereference() const
-		{
-			return *(mIterator->second);
-		}
-
-		typename InstanceMap::iterator mIterator;
-	};
-
-	class key_iter : public boost::iterator_facade<key_iter, KEY, boost::forward_traversal_tag>
-	{
-	public:
-		typedef boost::iterator_facade<key_iter, KEY, boost::forward_traversal_tag> super_t;
-
-		key_iter(typename InstanceMap::iterator it)
-		:	mIterator(it)
-		{
-			getStatic().incrementDepth();
-		}
-
-		key_iter(const key_iter& other)
-		:	mIterator(other.mIterator)
-		{
-			getStatic().incrementDepth();
-		}
-
-		~key_iter()
-		{
-			getStatic().decrementDepth();
-		}
-
-
-	private:
-		friend class boost::iterator_core_access;
-
-		void increment() { mIterator++; }
-		bool equal(key_iter const& other) const
-		{
-			return mIterator == other.mIterator;
-		}
-
-		KEY& dereference() const
-		{
-			return const_cast<KEY&>(mIterator->first);
-		}
-
-		typename InstanceMap::iterator mIterator;
-	};
-
-	static T* getInstance(const KEY& k)
-	{
-		const InstanceMap& map(getMap_());
-		typename InstanceMap::const_iterator found = map.find(k);
-		return (found == map.end()) ? NULL : found->second;
-	}
-
-	static instance_iter beginInstances() 
-	{	
-		return instance_iter(getMap_().begin()); 
-	}
-
-	static instance_iter endInstances() 
-	{
-		return instance_iter(getMap_().end());
-	}
-
-	static S32 instanceCount() 
-	{ 
-		return getMap_().size(); 
-	}
-
-	static key_iter beginKeys()
-	{
-		return key_iter(getMap_().begin());
-	}
-	static key_iter endKeys()
-	{
-		return key_iter(getMap_().end());
-	}
+    // snapshot of std::pair<const KEY, std::shared_ptr<T>> pairs
+    class snapshot
+    {
+        // It's very important that what we store in this snapshot are
+        // weak_ptrs, NOT shared_ptrs. That's how we discover whether any
+        // instance has been deleted during the lifespan of a snapshot.
+        typedef std::vector<std::pair<const KEY, std::weak_ptr<T>>> VectorType;
+        // Dereferencing our iterator produces a std::shared_ptr for each
+        // instance that still exists. Since we store weak_ptrs, that involves
+        // two chained transformations:
+        // - a transform_iterator to lock the weak_ptr and return a shared_ptr
+        // - a filter_iterator to skip any shared_ptr that has become invalid.
+        // It is very important that we filter lazily, that is, during
+        // traversal. Any one of our stored weak_ptrs might expire during
+        // traversal.
+        typedef std::pair<const KEY, std::shared_ptr<T>> strong_pair;
+        // Note for future reference: nat has not yet had any luck (up to
+        // Boost 1.67) trying to use boost::transform_iterator with a hand-
+        // coded functor, only with actual functions. In my experience, an
+        // internal boost::result_of() operation fails, even with an explicit
+        // result_type typedef. But this works.
+        static strong_pair strengthen(typename VectorType::value_type& pair)
+        {
+            return { pair.first, pair.second.lock() };
+        }
+        static bool dead_skipper(const strong_pair& pair)
+        {
+            return bool(pair.second);
+        }
+
+    public:
+        snapshot():
+            // populate our vector with a snapshot of (locked!) InstanceMap
+            // note, this assigns pair<KEY, shared_ptr> to pair<KEY, weak_ptr>
+            mData(mLock->mMap.begin(), mLock->mMap.end())
+        {
+            // release the lock once we've populated mData
+            mLock.unlock();
+        }
+
+        // You can't make a transform_iterator (or anything else) that
+        // literally stores a C++ function (decltype(strengthen)) -- but you
+        // can make a transform_iterator based on a _function pointer._
+        typedef boost::transform_iterator<decltype(strengthen)*,
+                                          typename VectorType::iterator> strong_iterator;
+        typedef boost::filter_iterator<decltype(dead_skipper)*, strong_iterator> iterator;
+
+        iterator begin() { return make_iterator(mData.begin()); }
+        iterator end()   { return make_iterator(mData.end()); }
+
+    private:
+        iterator make_iterator(typename VectorType::iterator iter)
+        {
+            // transform_iterator only needs the base iterator and the transform.
+            // filter_iterator wants the predicate and both ends of the range.
+            return iterator(dead_skipper,
+                            strong_iterator(iter, strengthen),
+                            strong_iterator(mData.end(), strengthen));
+        }
+
+        // lock static data during construction
+#if ! LL_WINDOWS
+        LockStatic mLock;
+#else  // LL_WINDOWS
+        // We want to be able to use (e.g.) our instance_snapshot subclass as:
+        // for (auto& inst : T::instance_snapshot()) ...
+        // But when this snapshot base class directly contains LockStatic, as
+        // above, Visual Studio 2017 requires us to code instead:
+        // for (auto& inst : std::move(T::instance_snapshot())) ...
+        // nat thinks this should be unnecessary, as an anonymous class
+        // instance is already a temporary. It shouldn't need to be cast to
+        // rvalue reference (the role of std::move()). clang evidently agrees,
+        // as the short form works fine with Xcode on Mac.
+        // To support the succinct usage, instead of directly storing
+        // LockStatic, store std::shared_ptr<LockStatic>, which is copyable.
+        std::shared_ptr<LockStatic> mLockp{std::make_shared<LockStatic>()};
+        LockStatic& mLock{*mLockp};
+#endif // LL_WINDOWS
+        VectorType mData;
+    };
+
+    // iterate over this for references to each instance
+    class instance_snapshot: public snapshot
+    {
+    private:
+        static T& instance_getter(typename snapshot::iterator::reference pair)
+        {
+            return *pair.second;
+        }
+    public:
+        typedef boost::transform_iterator<decltype(instance_getter)*,
+                                          typename snapshot::iterator> iterator;
+        iterator begin() { return iterator(snapshot::begin(), instance_getter); }
+        iterator end()   { return iterator(snapshot::end(),   instance_getter); }
+
+        void deleteAll()
+        {
+            for (auto it(snapshot::begin()), end(snapshot::end()); it != end; ++it)
+            {
+                delete it->second.get();
+            }
+        }
+    };                   
+
+    // iterate over this for each key
+    class key_snapshot: public snapshot
+    {
+    private:
+        static KEY key_getter(typename snapshot::iterator::reference pair)
+        {
+            return pair.first;
+        }
+    public:
+        typedef boost::transform_iterator<decltype(key_getter)*,
+                                          typename snapshot::iterator> iterator;
+        iterator begin() { return iterator(snapshot::begin(), key_getter); }
+        iterator end()   { return iterator(snapshot::end(),   key_getter); }
+    };
+
+    static T* getInstance(const KEY& k)
+    {
+        LockStatic lock;
+        const InstanceMap& map(lock->mMap);
+        typename InstanceMap::const_iterator found = map.find(k);
+        return (found == map.end()) ? NULL : found->second.get();
+    }
+
+    static S32 instanceCount() 
+    { 
+        return LockStatic()->mMap.size(); 
+    }
 
 protected:
-	LLInstanceTracker(const KEY& key) 
-	{ 
-		// make sure static data outlives all instances
-		getStatic();
-		add_(key); 
-	}
-	virtual ~LLInstanceTracker() LLINSTANCETRACKER_DTOR_NOEXCEPT
-	{ 
-		// it's unsafe to delete instances of this type while all instances are being iterated over.
-		llassert_always(getStatic().getDepth() == 0);
-		remove_();
-	}
-	virtual void setKey(KEY key) { remove_(); add_(key); }
-	virtual const KEY& getKey() const { return mInstanceKey; }
+    LLInstanceTracker(const KEY& key) 
+    {
+        // We do not intend to manage the lifespan of this object with
+        // shared_ptr, so give it a no-op deleter. We store shared_ptrs in our
+        // InstanceMap specifically so snapshot can store weak_ptrs so we can
+        // detect deletions during traversals.
+        std::shared_ptr<T> ptr(static_cast<T*>(this), [](T*){});
+        LockStatic lock;
+        add_(lock, key, ptr);
+    }
+public:
+    virtual ~LLInstanceTracker()
+    {
+        LockStatic lock;
+        remove_(lock);
+    }
+protected:
+    virtual void setKey(KEY key)
+    {
+        LockStatic lock;
+        // Even though the shared_ptr we store in our map has a no-op deleter
+        // for T itself, letting the use count decrement to 0 will still
+        // delete the use-count object. Capture the shared_ptr we just removed
+        // and re-add it to the map with the new key.
+        auto ptr = remove_(lock);
+        add_(lock, key, ptr);
+    }
+public:
+    virtual const KEY& getKey() const { return mInstanceKey; }
 
 private:
-	LLInstanceTracker( const LLInstanceTracker& );
-	const LLInstanceTracker& operator=( const LLInstanceTracker& );
-
-	void add_(const KEY& key) 
-	{ 
-		mInstanceKey = key; 
-		InstanceMap& map = getMap_();
-		typename InstanceMap::iterator insertion_point_it = map.lower_bound(key);
-		if (insertion_point_it != map.end() 
-			&& insertion_point_it->first == key)
-		{ // found existing entry with that key
-			switch(KEY_COLLISION_BEHAVIOR)
-			{
-				case LLInstanceTrackerErrorOnCollision:
-				{
-					// use assert here instead of LL_ERRS(), otherwise the error will be ignored since this call is made during global object initialization
-					llassert_always_msg(false, "Instance with this same key already exists!");
-					break;
-				}
-				case LLInstanceTrackerReplaceOnCollision:
-				{
-					// replace pointer, but leave key (should have compared equal anyway)
-					insertion_point_it->second = static_cast<T*>(this);
-					break;
-				}
-				default:
-					break;
-			}
-		}
-		else
-		{ // new key
-			map.insert(insertion_point_it, std::make_pair(key, static_cast<T*>(this)));
-		}
-	}
-	void remove_()
-	{
-		InstanceMap& map = getMap_();
-		typename InstanceMap::iterator iter = map.find(mInstanceKey);
-		if (iter != map.end())
-		{
-			map.erase(iter);
-		}
-	}
+    LLInstanceTracker( const LLInstanceTracker& ) = delete;
+    LLInstanceTracker& operator=( const LLInstanceTracker& ) = delete;
+
+    // for logging
+    template <typename K>
+    static std::string report(K key) { return stringize(key); }
+    static std::string report(const std::string& key) { return "'" + key + "'"; }
+    static std::string report(const char* key) { return report(std::string(key)); }
+
+    // caller must instantiate LockStatic
+    void add_(LockStatic& lock, const KEY& key, const std::shared_ptr<T>& ptr) 
+    { 
+        mInstanceKey = key; 
+        InstanceMap& map = lock->mMap;
+        switch(KEY_COLLISION_BEHAVIOR)
+        {
+        case LLInstanceTrackerErrorOnCollision:
+        {
+            // map stores shared_ptr to self
+            auto pair = map.emplace(key, ptr);
+            if (! pair.second)
+            {
+                LLInstanceTrackerPrivate::logerrs(typeid(*this).name(), " instance with key ",
+                                                  report(key), " already exists!");
+            }
+            break;
+        }
+        case LLInstanceTrackerReplaceOnCollision:
+            map[key] = ptr;
+            break;
+        default:
+            break;
+        }
+    }
+    std::shared_ptr<T> remove_(LockStatic& lock)
+    {
+        InstanceMap& map = lock->mMap;
+        typename InstanceMap::iterator iter = map.find(mInstanceKey);
+        if (iter != map.end())
+        {
+            auto ret = iter->second;
+            map.erase(iter);
+            return ret;
+        }
+        return {};
+    }
 
 private:
-	KEY mInstanceKey;
+    KEY mInstanceKey;
 };
 
+/*****************************************************************************
+*   LLInstanceTracker without key
+*****************************************************************************/
+// TODO:
+// - For the case of omitted KEY template parameter, consider storing
+//   std::map<T*, std::shared_ptr<T>> instead of std::set<std::shared_ptr<T>>.
+//   That might let us share more of the implementation between KEY and
+//   non-KEY LLInstanceTracker subclasses.
+// - Even if not that, consider trying to unify the snapshot implementations.
+//   The trouble is that the 'iterator' published by each (and by their
+//   subclasses) must reflect the specific type of the callables that
+//   distinguish them. (Maybe make instance_snapshot() and key_snapshot()
+//   factory functions that pass lambdas to a factory function for the generic
+//   template class?)
+
 /// explicit specialization for default case where KEY is void
 /// use a simple std::set<T*>
 template<typename T, EInstanceTrackerAllowKeyCollisions KEY_COLLISION_BEHAVIOR>
-class LLInstanceTracker<T, void, KEY_COLLISION_BEHAVIOR> : public LLInstanceTrackerBase
+class LLInstanceTracker<T, void, KEY_COLLISION_BEHAVIOR>
 {
-	typedef LLInstanceTracker<T, void> self_t;
-	typedef typename std::set<T*> InstanceSet;
-	struct StaticData: public StaticBase
-	{
-		InstanceSet sSet;
-	};
-	static StaticData& getStatic() { static StaticData sData; return sData; }
-	static InstanceSet& getSet_() { return getStatic().sSet; }
+    typedef std::set<std::shared_ptr<T>> InstanceSet;
+    struct StaticData: public LLInstanceTrackerPrivate::StaticBase
+    {
+        InstanceSet mSet;
+    };
+    typedef llthread::LockStatic<StaticData> LockStatic;
 
 public:
+    /**
+     * Storing a dumb T* somewhere external is a bad idea, since
+     * LLInstanceTracker subclasses are explicitly destroyed rather than
+     * managed by smart pointers. It's legal to declare stack instances of an
+     * LLInstanceTracker subclass. But it's reasonable to store a
+     * std::weak_ptr<T>, which will become invalid when the T instance is
+     * destroyed.
+     */
+    std::weak_ptr<T> getWeak()
+    {
+        return mSelf;
+    }
+    
+    static S32 instanceCount() { return LockStatic()->mSet.size(); }
 
-	/**
-	 * Does a particular instance still exist? Of course, if you already have
-	 * a T* in hand, you need not call getInstance() to @em locate the
-	 * instance -- unlike the case where getInstance() accepts some kind of
-	 * key. Nonetheless this method is still useful to @em validate a
-	 * particular T*, since each instance's destructor removes itself from the
-	 * underlying set.
-	 */
-	static T* getInstance(T* k)
-	{
-		const InstanceSet& set(getSet_());
-		typename InstanceSet::const_iterator found = set.find(k);
-		return (found == set.end())? NULL : *found;
-	}
-	static S32 instanceCount() { return getSet_().size(); }
-
-	class instance_iter : public boost::iterator_facade<instance_iter, T, boost::forward_traversal_tag>
-	{
-	public:
-		instance_iter(const typename InstanceSet::iterator& it)
-		:	mIterator(it)
-		{
-			getStatic().incrementDepth();
-		}
-
-		instance_iter(const instance_iter& other)
-		:	mIterator(other.mIterator)
-		{
-			getStatic().incrementDepth();
-		}
-
-		~instance_iter()
-		{
-			getStatic().decrementDepth();
-		}
-
-	private:
-		friend class boost::iterator_core_access;
-
-		void increment() { mIterator++; }
-		bool equal(instance_iter const& other) const
-		{
-			return mIterator == other.mIterator;
-		}
-
-		T& dereference() const
-		{
-			return **mIterator;
-		}
-
-		typename InstanceSet::iterator mIterator;
-	};
-
-	static instance_iter beginInstances() {	return instance_iter(getSet_().begin()); }
-	static instance_iter endInstances() { return instance_iter(getSet_().end()); }
+    // snapshot of std::shared_ptr<T> pointers
+    class snapshot
+    {
+        // It's very important that what we store in this snapshot are
+        // weak_ptrs, NOT shared_ptrs. That's how we discover whether any
+        // instance has been deleted during the lifespan of a snapshot.
+        typedef std::vector<std::weak_ptr<T>> VectorType;
+        // Dereferencing our iterator produces a std::shared_ptr for each
+        // instance that still exists. Since we store weak_ptrs, that involves
+        // two chained transformations:
+        // - a transform_iterator to lock the weak_ptr and return a shared_ptr
+        // - a filter_iterator to skip any shared_ptr that has become invalid.
+        typedef std::shared_ptr<T> strong_ptr;
+        static strong_ptr strengthen(typename VectorType::value_type& ptr)
+        {
+            return ptr.lock();
+        }
+        static bool dead_skipper(const strong_ptr& ptr)
+        {
+            return bool(ptr);
+        }
+
+    public:
+        snapshot():
+            // populate our vector with a snapshot of (locked!) InstanceSet
+            // note, this assigns stored shared_ptrs to weak_ptrs for snapshot
+            mData(mLock->mSet.begin(), mLock->mSet.end())
+        {
+            // release the lock once we've populated mData
+            mLock.unlock();
+        }
+
+        typedef boost::transform_iterator<decltype(strengthen)*,
+                                          typename VectorType::iterator> strong_iterator;
+        typedef boost::filter_iterator<decltype(dead_skipper)*, strong_iterator> iterator;
+
+        iterator begin() { return make_iterator(mData.begin()); }
+        iterator end()   { return make_iterator(mData.end()); }
+
+    private:
+        iterator make_iterator(typename VectorType::iterator iter)
+        {
+            // transform_iterator only needs the base iterator and the transform.
+            // filter_iterator wants the predicate and both ends of the range.
+            return iterator(dead_skipper,
+                            strong_iterator(iter, strengthen),
+                            strong_iterator(mData.end(), strengthen));
+        }
+
+        // lock static data during construction
+#if ! LL_WINDOWS
+        LockStatic mLock;
+#else  // LL_WINDOWS
+        // We want to be able to use our instance_snapshot subclass as:
+        // for (auto& inst : T::instance_snapshot()) ...
+        // But when this snapshot base class directly contains LockStatic, as
+        // above, Visual Studio 2017 requires us to code instead:
+        // for (auto& inst : std::move(T::instance_snapshot())) ...
+        // nat thinks this should be unnecessary, as an anonymous class
+        // instance is already a temporary. It shouldn't need to be cast to
+        // rvalue reference (the role of std::move()). clang evidently agrees,
+        // as the short form works fine with Xcode on Mac.
+        // To support the succinct usage, instead of directly storing
+        // LockStatic, store std::shared_ptr<LockStatic>, which is copyable.
+        std::shared_ptr<LockStatic> mLockp{std::make_shared<LockStatic>()};
+        LockStatic& mLock{*mLockp};
+#endif // LL_WINDOWS
+        VectorType mData;
+    };
+
+    // iterate over this for references to each instance
+    struct instance_snapshot: public snapshot
+    {
+        typedef boost::indirect_iterator<typename snapshot::iterator> iterator;
+        iterator begin() { return iterator(snapshot::begin()); }
+        iterator end()   { return iterator(snapshot::end()); }
+
+        void deleteAll()
+        {
+            for (auto it(snapshot::begin()), end(snapshot::end()); it != end; ++it)
+            {
+                delete it->get();
+            }
+        }
+    };
 
 protected:
-	LLInstanceTracker()
-	{
-		// make sure static data outlives all instances
-		getStatic();
-		getSet_().insert(static_cast<T*>(this));
-	}
-	virtual ~LLInstanceTracker() LLINSTANCETRACKER_DTOR_NOEXCEPT
-	{
-		// it's unsafe to delete instances of this type while all instances are being iterated over.
-		llassert_always(getStatic().getDepth() == 0);
-		getSet_().erase(static_cast<T*>(this));
-	}
-
-	LLInstanceTracker(const LLInstanceTracker& other)
-	{
-		getSet_().insert(static_cast<T*>(this));
-	}
+    LLInstanceTracker()
+    {
+        // Since we do not intend for this shared_ptr to manage lifespan, give
+        // it a no-op deleter.
+        std::shared_ptr<T> ptr(static_cast<T*>(this), [](T*){});
+        // save corresponding weak_ptr for future reference
+        mSelf = ptr;
+        // Also store it in our class-static set to track this instance.
+        LockStatic()->mSet.emplace(ptr);
+    }
+public:
+    virtual ~LLInstanceTracker()
+    {
+        // convert weak_ptr to shared_ptr because that's what we store in our
+        // InstanceSet
+        LockStatic()->mSet.erase(mSelf.lock());
+    }
+protected:
+    LLInstanceTracker(const LLInstanceTracker& other):
+        LLInstanceTracker()
+    {}
+
+private:
+    // Storing a weak_ptr to self is a bit like deriving from
+    // std::enable_shared_from_this(), except more explicit.
+    std::weak_ptr<T> mSelf;
 };
 
 #endif
diff --git a/indra/llcommon/llleaplistener.cpp b/indra/llcommon/llleaplistener.cpp
index fa5730f112..3e6ce9092c 100644
--- a/indra/llcommon/llleaplistener.cpp
+++ b/indra/llcommon/llleaplistener.cpp
@@ -14,6 +14,8 @@
 // associated header
 #include "llleaplistener.h"
 // STL headers
+#include <map>
+#include <functional>
 // std headers
 // external library headers
 #include <boost/foreach.hpp>
@@ -60,16 +62,11 @@ LLLeapListener::LLLeapListener(const ConnectFunc& connect):
     LLSD need_name(LLSDMap("name", LLSD()));
     add("newpump",
         "Instantiate a new LLEventPump named like [\"name\"] and listen to it.\n"
-        "If [\"type\"] == \"LLEventQueue\", make LLEventQueue, else LLEventStream.\n"
+        "[\"type\"] == \"LLEventStream\", \"LLEventMailDrop\" et al.\n"
         "Events sent through new LLEventPump will be decorated with [\"pump\"]=name.\n"
         "Returns actual name in [\"name\"] (may be different if collision).",
         &LLLeapListener::newpump,
         need_name);
-    add("killpump",
-        "Delete LLEventPump [\"name\"] created by \"newpump\".\n"
-        "Returns [\"status\"] boolean indicating whether such a pump existed.",
-        &LLLeapListener::killpump,
-        need_name);
     LLSD need_source_listener(LLSDMap("source", LLSD())("listener", LLSD()));
     add("listen",
         "Listen to an existing LLEventPump named [\"source\"], with listener name\n"
@@ -124,40 +121,23 @@ void LLLeapListener::newpump(const LLSD& request)
     Response reply(LLSD(), request);
 
     std::string name = request["name"];
-    LLSD const & type = request["type"];
+    std::string type = request["type"];
 
-    LLEventPump * new_pump = NULL;
-    if (type.asString() == "LLEventQueue")
+    try
     {
-        new_pump = new LLEventQueue(name, true); // tweak name for uniqueness
+        // tweak name for uniqueness
+        LLEventPump& new_pump(LLEventPumps::instance().make(name, true, type));
+        name = new_pump.getName();
+        reply["name"] = name;
+
+        // Now listen on this new pump with our plugin listener
+        std::string myname("llleap");
+        saveListener(name, myname, mConnect(new_pump, myname));
     }
-    else
+    catch (const LLEventPumps::BadType& error)
     {
-        if (! (type.isUndefined() || type.asString() == "LLEventStream"))
-        {
-            reply.warn(STRINGIZE("unknown 'type' " << type << ", using LLEventStream"));
-        }
-        new_pump = new LLEventStream(name, true); // tweak name for uniqueness
+        reply.error(error.what());
     }
-
-    name = new_pump->getName();
-
-    mEventPumps.insert(name, new_pump);
-
-    // Now listen on this new pump with our plugin listener
-    std::string myname("llleap");
-    saveListener(name, myname, mConnect(*new_pump, myname));
-
-    reply["name"] = name;
-}
-
-void LLLeapListener::killpump(const LLSD& request)
-{
-    Response reply(LLSD(), request);
-
-    std::string name = request["name"];
-    // success == (nonzero number of entries were erased)
-    reply["status"] = bool(mEventPumps.erase(name));
 }
 
 void LLLeapListener::listen(const LLSD& request)
@@ -228,13 +208,11 @@ void LLLeapListener::getAPIs(const LLSD& request) const
 {
     Response reply(LLSD(), request);
 
-    for (LLEventAPI::instance_iter eai(LLEventAPI::beginInstances()),
-             eaend(LLEventAPI::endInstances());
-         eai != eaend; ++eai)
+    for (auto& ea : LLEventAPI::instance_snapshot())
     {
         LLSD info;
-        info["desc"] = eai->getDesc();
-        reply[eai->getName()] = info;
+        info["desc"] = ea.getDesc();
+        reply[ea.getName()] = info;
     }
 }
 
diff --git a/indra/llcommon/llleaplistener.h b/indra/llcommon/llleaplistener.h
index 2193d81b9e..0ca5893657 100644
--- a/indra/llcommon/llleaplistener.h
+++ b/indra/llcommon/llleaplistener.h
@@ -40,7 +40,6 @@ public:
 
 private:
     void newpump(const LLSD&);
-    void killpump(const LLSD&);
     void listen(const LLSD&);
     void stoplistening(const LLSD&);
     void ping(const LLSD&) const;
@@ -64,10 +63,6 @@ private:
     // and listener name.
     typedef std::map<std::pair<std::string, std::string>, LLBoundListener> ListenersMap;
     ListenersMap mListeners;
-    // Similar lifespan reasoning applies to LLEventPumps instantiated by
-    // newpump() operations.
-    typedef boost::ptr_map<std::string, LLEventPump> EventPumpsMap;
-    EventPumpsMap mEventPumps;
 };
 
 #endif /* ! defined(LL_LLLEAPLISTENER_H) */
diff --git a/indra/llcommon/lllistenerwrapper.h b/indra/llcommon/lllistenerwrapper.h
deleted file mode 100644
index 09d074abca..0000000000
--- a/indra/llcommon/lllistenerwrapper.h
+++ /dev/null
@@ -1,198 +0,0 @@
-/**
- * @file   lllistenerwrapper.h
- * @author Nat Goodspeed
- * @date   2009-11-30
- * @brief  Introduce LLListenerWrapper template
- * 
- * $LicenseInfo:firstyear=2009&license=viewerlgpl$
- * Second Life Viewer Source Code
- * Copyright (C) 2010, Linden Research, Inc.
- * 
- * This library is free software; you can redistribute it and/or
- * modify it under the terms of the GNU Lesser General Public
- * License as published by the Free Software Foundation;
- * version 2.1 of the License only.
- * 
- * This library is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
- * Lesser General Public License for more details.
- * 
- * You should have received a copy of the GNU Lesser General Public
- * License along with this library; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
- * 
- * Linden Research, Inc., 945 Battery Street, San Francisco, CA  94111  USA
- * $/LicenseInfo$
- */
-
-#if ! defined(LL_LLLISTENERWRAPPER_H)
-#define LL_LLLISTENERWRAPPER_H
-
-#include "llevents.h"               // LLListenerWrapperBase
-#include <boost/visit_each.hpp>
-
-/**
- * Template base class for coding wrappers for LLEventPump listeners.
- *
- * Derive your listener wrapper from LLListenerWrapper. You must use
- * LLLISTENER_WRAPPER_SUBCLASS() so your subclass will play nicely with
- * boost::visit_each (q.v.). That way boost::signals2 can still detect
- * derivation from LLEventTrackable, and so forth.
- */
-template <typename LISTENER>
-class LLListenerWrapper: public LLListenerWrapperBase
-{
-public:
-    /// Wrap an arbitrary listener object
-    LLListenerWrapper(const LISTENER& listener):
-        mListener(listener)
-    {}
-
-    /// call
-    virtual bool operator()(const LLSD& event)
-    {
-        return mListener(event);
-    }
-
-    /// Allow boost::visit_each() to peek at our mListener.
-    template <class V>
-    void accept_visitor(V& visitor) const
-    {
-        using boost::visit_each;
-        visit_each(visitor, mListener, 0);
-    }
-
-private:
-    LISTENER mListener;
-};
-
-/**
- * Specialize boost::visit_each() (leveraging ADL) to peek inside an
- * LLListenerWrapper<T> to traverse its LISTENER. We borrow the
- * accept_visitor() pattern from boost::bind(), avoiding the need to make
- * mListener public.
- */
-template <class V, typename T>
-void visit_each(V& visitor, const LLListenerWrapper<T>& wrapper, int)
-{
-    wrapper.accept_visitor(visitor);
-}
-
-/// use this (sigh!) for each subclass of LLListenerWrapper<T> you write
-#define LLLISTENER_WRAPPER_SUBCLASS(CLASS)                              \
-template <class V, typename T>                                          \
-void visit_each(V& visitor, const CLASS<T>& wrapper, int)               \
-{                                                                       \
-    visit_each(visitor, static_cast<const LLListenerWrapper<T>&>(wrapper), 0); \
-}                                                                       \
-                                                                        \
-/* Have to state this explicitly, rather than using LL_TEMPLATE_CONVERTIBLE, */ \
-/* because the source type is itself a template. */                     \
-template <typename T>                                                   \
-struct ll_template_cast_impl<const LLListenerWrapperBase*, const CLASS<T>*> \
-{                                                                       \
-    const LLListenerWrapperBase* operator()(const CLASS<T>* wrapper)    \
-    {                                                                   \
-        return wrapper;                                                 \
-    }                                                                   \
-}
-
-/**
- * Make an instance of a listener wrapper. Every wrapper class must be a
- * template accepting a listener object of arbitrary type. In particular, the
- * type of a boost::bind() expression is deliberately undocumented. So we
- * can't just write Wrapper<CorrectType>(boost::bind(...)). Instead we must
- * write llwrap<Wrapper>(boost::bind(...)).
- */
-template <template<typename> class WRAPPER, typename T>
-WRAPPER<T> llwrap(const T& listener)
-{
-    return WRAPPER<T>(listener);
-}
-
-/**
- * This LLListenerWrapper template subclass is used to report entry/exit to an
- * event listener, by changing this:
- * @code
- * someEventPump.listen("MyClass",
- *                      boost::bind(&MyClass::method, ptr, _1));
- * @endcode
- * to this:
- * @code
- * someEventPump.listen("MyClass",
- *                      llwrap<LLCoutListener>(
- *                      boost::bind(&MyClass::method, ptr, _1)));
- * @endcode
- */
-template <class LISTENER>
-class LLCoutListener: public LLListenerWrapper<LISTENER>
-{
-    typedef LLListenerWrapper<LISTENER> super;
-
-public:
-    /// Wrap an arbitrary listener object
-    LLCoutListener(const LISTENER& listener):
-        super(listener)
-    {}
-
-    /// call
-    virtual bool operator()(const LLSD& event)
-    {
-        std::cout << "Entering listener " << *super::mName << " with " << event << std::endl;
-        bool handled = super::operator()(event);
-        std::cout << "Leaving  listener " << *super::mName;
-        if (handled)
-        {
-            std::cout << " (handled)";
-        }
-        std::cout << std::endl;
-        return handled;
-    }
-};
-
-LLLISTENER_WRAPPER_SUBCLASS(LLCoutListener);
-
-/**
- * This LLListenerWrapper template subclass is used to log entry/exit to an
- * event listener, by changing this:
- * @code
- * someEventPump.listen("MyClass",
- *                      boost::bind(&MyClass::method, ptr, _1));
- * @endcode
- * to this:
- * @code
- * someEventPump.listen("MyClass",
- *                      llwrap<LLLogListener>(
- *                      boost::bind(&MyClass::method, ptr, _1)));
- * @endcode
- */
-template <class LISTENER>
-class LLLogListener: public LLListenerWrapper<LISTENER>
-{
-    typedef LLListenerWrapper<LISTENER> super;
-
-public:
-    /// Wrap an arbitrary listener object
-    LLLogListener(const LISTENER& listener):
-        super(listener)
-    {}
-
-    /// call
-    virtual bool operator()(const LLSD& event)
-    {
-        LL_DEBUGS("LLLogListener") << "Entering listener " << *super::mName << " with " << event << LL_ENDL;
-        bool handled = super::operator()(event);
-        LL_DEBUGS("LLLogListener") << "Leaving  listener " << *super::mName;
-        if (handled)
-        {
-            LL_CONT << " (handled)";
-        }
-        LL_CONT << LL_ENDL;
-        return handled;
-    }
-};
-
-LLLISTENER_WRAPPER_SUBCLASS(LLLogListener);
-
-#endif /* ! defined(LL_LLLISTENERWRAPPER_H) */
diff --git a/indra/llcommon/llmainthreadtask.cpp b/indra/llcommon/llmainthreadtask.cpp
new file mode 100644
index 0000000000..e0d70cacd8
--- /dev/null
+++ b/indra/llcommon/llmainthreadtask.cpp
@@ -0,0 +1,22 @@
+/**
+ * @file   llmainthreadtask.cpp
+ * @author Nat Goodspeed
+ * @date   2019-12-05
+ * @brief  Implementation for llmainthreadtask.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+// Precompiled header
+#include "linden_common.h"
+// associated header
+#include "llmainthreadtask.h"
+// STL headers
+// std headers
+// external library headers
+// other Linden headers
+
+// This file is required by our CMake integration-test machinery. It
+// contributes no code to the viewer executable.
diff --git a/indra/llcommon/llmainthreadtask.h b/indra/llcommon/llmainthreadtask.h
new file mode 100644
index 0000000000..d509b687c0
--- /dev/null
+++ b/indra/llcommon/llmainthreadtask.h
@@ -0,0 +1,99 @@
+/**
+ * @file   llmainthreadtask.h
+ * @author Nat Goodspeed
+ * @date   2019-12-04
+ * @brief  LLMainThreadTask dispatches work to the main thread. When invoked on
+ *         the main thread, it performs the work inline.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if ! defined(LL_LLMAINTHREADTASK_H)
+#define LL_LLMAINTHREADTASK_H
+
+#include "lleventtimer.h"
+#include "llthread.h"
+#include "llmake.h"
+#include <future>
+#include <type_traits>              // std::result_of
+
+/**
+ * LLMainThreadTask provides a way to perform some task specifically on the
+ * main thread, waiting for it to complete. A task consists of a C++ nullary
+ * invocable (i.e. any callable that requires no arguments) with arbitrary
+ * return type.
+ *
+ * Instead of instantiating LLMainThreadTask, pass your invocable to its
+ * static dispatch() method. dispatch() returns the result of calling your
+ * task. (Or, if your task throws an exception, dispatch() throws that
+ * exception. See std::packaged_task.)
+ *
+ * When you call dispatch() on the main thread (as determined by
+ * on_main_thread() in llthread.h), it simply calls your task and returns the
+ * result.
+ *
+ * When you call dispatch() on a secondary thread, it instantiates an
+ * LLEventTimer subclass scheduled immediately. Next time the main loop calls
+ * LLEventTimer::updateClass(), your task will be run, and LLMainThreadTask
+ * will fulfill a future with its result. Meanwhile the requesting thread
+ * blocks on that future. As soon as it is set, the requesting thread wakes up
+ * with the task result.
+ */
+class LLMainThreadTask
+{
+private:
+    // Don't instantiate this class -- use dispatch() instead.
+    LLMainThreadTask() {}
+
+public:
+    /// dispatch() is the only way to invoke this functionality.
+    template <typename CALLABLE>
+    static auto dispatch(CALLABLE&& callable) -> decltype(callable())
+    {
+        if (on_main_thread())
+        {
+            // we're already running on the main thread, perfect
+            return callable();
+        }
+        else
+        {
+            // It's essential to construct LLEventTimer subclass instances on
+            // the heap because, on completion, LLEventTimer deletes them.
+            // Once we enable C++17, we can use Class Template Argument
+            // Deduction. Until then, use llmake_heap().
+            auto* task = llmake_heap<Task>(std::forward<CALLABLE>(callable));
+            auto future = task->mTask.get_future();
+            // Now simply block on the future.
+            return future.get();
+        }
+    }
+
+private:
+    template <typename CALLABLE>
+    struct Task: public LLEventTimer
+    {
+        Task(CALLABLE&& callable):
+            // no wait time: call tick() next chance we get
+            LLEventTimer(0),
+            mTask(std::forward<CALLABLE>(callable))
+        {}
+        BOOL tick() override
+        {
+            // run the task on the main thread, will populate the future
+            // obtained by get_future()
+            mTask();
+            // tell LLEventTimer we're done (one shot)
+            return TRUE;
+        }
+        // Given arbitrary CALLABLE, which might be a lambda, how are we
+        // supposed to obtain its signature for std::packaged_task? It seems
+        // redundant to have to add an argument list to engage result_of, then
+        // add the argument list again to complete the signature. At least we
+        // only support a nullary CALLABLE.
+        std::packaged_task<typename std::result_of<CALLABLE()>::type()> mTask;
+    };
+};
+
+#endif /* ! defined(LL_LLMAINTHREADTASK_H) */
diff --git a/indra/llcommon/llmake.h b/indra/llcommon/llmake.h
index 08744f90fb..02463d97ea 100644
--- a/indra/llcommon/llmake.h
+++ b/indra/llcommon/llmake.h
@@ -12,10 +12,8 @@
  * 
  *         also relevant:
  *
- *         Template argument deduction for class templates
- *         http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0091r3.html
- *         was apparently adopted in June 2016? Unclear when compilers will
- *         portably support this, but there is hope.
+ *         Template argument deduction for class templates (C++17)
+ *         https://en.cppreference.com/w/cpp/language/class_template_argument_deduction
  *
  * $LicenseInfo:firstyear=2015&license=viewerlgpl$
  * Copyright (c) 2015, Linden Research, Inc.
@@ -25,37 +23,43 @@
 #if ! defined(LL_LLMAKE_H)
 #define LL_LLMAKE_H
 
-/*==========================================================================*|
-// When we allow ourselves to compile with C++11 features enabled, this form
-// should generically handle an arbitrary number of arguments.
-
+/**
+ * Usage: llmake<SomeTemplate>(args...)
+ *
+ * Deduces the types T... of 'args' and returns an instance of
+ * SomeTemplate<T...>(args...).
+ */
 template <template<typename...> class CLASS_TEMPLATE, typename... ARGS>
 CLASS_TEMPLATE<ARGS...> llmake(ARGS && ... args)
 {
     return CLASS_TEMPLATE<ARGS...>(std::forward<ARGS>(args)...);
 }
-|*==========================================================================*/
 
-// As of 2015-12-18, this is what we'll use instead. Add explicit overloads
-// for different numbers of template parameters as use cases arise.
+/// dumb pointer template just in case that's what's wanted
+template <typename T>
+using dumb_pointer = T*;
 
 /**
- * Usage: llmake<SomeTemplate>(arg)
+ * Same as llmake(), but returns a pointer to a new heap instance of
+ * SomeTemplate<T...>(args...) using the pointer of your choice.
  *
- * Deduces the type T of 'arg' and returns an instance of SomeTemplate<T>
- * initialized with 'arg'. Assumes a constructor accepting T (by value,
- * reference or whatever).
+ * @code
+ * auto* dumb  = llmake_heap<SomeTemplate>(args...);
+ * auto shared = llmake_heap<SomeTemplate, std::shared_ptr>(args...);
+ * auto unique = llmake_heap<SomeTemplate, std::unique_ptr>(args...);
+ * @endcode
  */
-template <template<typename> class CLASS_TEMPLATE, typename ARG1>
-CLASS_TEMPLATE<ARG1> llmake(const ARG1& arg1)
-{
-    return CLASS_TEMPLATE<ARG1>(arg1);
-}
-
-template <template<typename, typename> class CLASS_TEMPLATE, typename ARG1, typename ARG2>
-CLASS_TEMPLATE<ARG1, ARG2> llmake(const ARG1& arg1, const ARG2& arg2)
+// POINTER_TEMPLATE is characterized as template<typename...> rather than as
+// template<typename T> because (e.g.) std::unique_ptr has multiple template
+// arguments. Even though we only engage one, std::unique_ptr doesn't match a
+// template template parameter that itself takes only one template parameter.
+template <template<typename...> class CLASS_TEMPLATE,
+          template<typename...> class POINTER_TEMPLATE=dumb_pointer,
+          typename... ARGS>
+POINTER_TEMPLATE<CLASS_TEMPLATE<ARGS...>> llmake_heap(ARGS&&... args)
 {
-    return CLASS_TEMPLATE<ARG1, ARG2>(arg1, arg2);
+    return POINTER_TEMPLATE<CLASS_TEMPLATE<ARGS...>>(
+        new CLASS_TEMPLATE<ARGS...>(std::forward<ARGS>(args)...));
 }
 
 #endif /* ! defined(LL_LLMAKE_H) */
diff --git a/indra/llcommon/llmutex.cpp b/indra/llcommon/llmutex.cpp
index 75f43a4704..4d73c04d07 100644
--- a/indra/llcommon/llmutex.cpp
+++ b/indra/llcommon/llmutex.cpp
@@ -32,8 +32,7 @@
 //============================================================================
 
 LLMutex::LLMutex() :
- mCount(0),
- mLockingThread(NO_THREAD)
+ mCount(0)
 {
 }
 
@@ -55,7 +54,7 @@ void LLMutex::lock()
 	
 #if MUTEX_DEBUG
 	// Have to have the lock before we can access the debug info
-	U32 id = LLThread::currentID();
+	auto id = LLThread::currentID();
 	if (mIsLocked[id] != FALSE)
 		LL_ERRS() << "Already locked in Thread: " << id << LL_ENDL;
 	mIsLocked[id] = TRUE;
@@ -74,13 +73,13 @@ void LLMutex::unlock()
 	
 #if MUTEX_DEBUG
 	// Access the debug info while we have the lock
-	U32 id = LLThread::currentID();
+	auto id = LLThread::currentID();
 	if (mIsLocked[id] != TRUE)
 		LL_ERRS() << "Not locked in Thread: " << id << LL_ENDL;	
 	mIsLocked[id] = FALSE;
 #endif
 
-	mLockingThread = NO_THREAD;
+	mLockingThread = LLThread::id_t();
 	mMutex.unlock();
 }
 
@@ -102,7 +101,7 @@ bool LLMutex::isSelfLocked()
 	return mLockingThread == LLThread::currentID();
 }
 
-U32 LLMutex::lockingThread() const
+LLThread::id_t LLMutex::lockingThread() const
 {
 	return mLockingThread;
 }
@@ -122,7 +121,7 @@ bool LLMutex::trylock()
 	
 #if MUTEX_DEBUG
 	// Have to have the lock before we can access the debug info
-	U32 id = LLThread::currentID();
+	auto id = LLThread::currentID();
 	if (mIsLocked[id] != FALSE)
 		LL_ERRS() << "Already locked in Thread: " << id << LL_ENDL;
 	mIsLocked[id] = TRUE;
diff --git a/indra/llcommon/llmutex.h b/indra/llcommon/llmutex.h
index f841d7f950..838d7d34c0 100644
--- a/indra/llcommon/llmutex.h
+++ b/indra/llcommon/llmutex.h
@@ -28,20 +28,12 @@
 #define LL_LLMUTEX_H
 
 #include "stdtypes.h"
+#include "llthread.h"
 #include <boost/noncopyable.hpp>
 
-#if LL_WINDOWS
-#pragma warning (push)
-#pragma warning (disable:4265)
-#endif
-// 'std::_Pad' : class has virtual functions, but destructor is not virtual
-#include <mutex>
+#include "mutex.h"
 #include <condition_variable>
 
-#if LL_WINDOWS
-#pragma warning (pop)
-#endif
-
 //============================================================================
 
 #define MUTEX_DEBUG (LL_DEBUG || LL_RELEASE_WITH_DEBUG_INFO)
@@ -53,11 +45,6 @@
 class LL_COMMON_API LLMutex
 {
 public:
-	typedef enum
-	{
-		NO_THREAD = 0xFFFFFFFF
-	} e_locking_thread;
-
 	LLMutex();
 	virtual ~LLMutex();
 	
@@ -66,15 +53,15 @@ public:
 	void unlock();		// undefined behavior when called on mutex not being held
 	bool isLocked(); 	// non-blocking, but does do a lock/unlock so not free
 	bool isSelfLocked(); //return true if locked in a same thread
-	U32 lockingThread() const; //get ID of locking thread
-	
+	LLThread::id_t lockingThread() const; //get ID of locking thread
+
 protected:
 	std::mutex			mMutex;
 	mutable U32			mCount;
-	mutable U32			mLockingThread;
+	mutable LLThread::id_t	mLockingThread;
 	
 #if MUTEX_DEBUG
-	std::map<U32, BOOL> mIsLocked;
+	std::map<LLThread::id_t, BOOL> mIsLocked;
 #endif
 };
 
diff --git a/indra/llcommon/llpreprocessor.h b/indra/llcommon/llpreprocessor.h
index e8f9981437..bae402110a 100644
--- a/indra/llcommon/llpreprocessor.h
+++ b/indra/llcommon/llpreprocessor.h
@@ -232,4 +232,11 @@
 #define LL_COMPILE_TIME_MESSAGE(msg)
 #endif
 
+// __FUNCTION__ works on all the platforms we care about, but...
+#if LL_WINDOWS
+#define LL_PRETTY_FUNCTION __FUNCSIG__
+#else
+#define LL_PRETTY_FUNCTION __PRETTY_FUNCTION__
+#endif
+
 #endif	//	not LL_LINDEN_PREPROCESSOR_H
diff --git a/indra/llcommon/llprocess.cpp b/indra/llcommon/llprocess.cpp
index 1fa53f322b..23936f0526 100644
--- a/indra/llcommon/llprocess.cpp
+++ b/indra/llcommon/llprocess.cpp
@@ -994,9 +994,9 @@ void LLProcess::handle_status(int reason, int status)
 //	wi->rv = apr_proc_wait(wi->child, &wi->rc, &wi->why, APR_NOWAIT);
 	// It's just wrong to call apr_proc_wait() here. The only way APR knows to
 	// call us with APR_OC_REASON_DEATH is that it's already reaped this child
-	// process, so calling suspend() will only produce "huh?" from the OS. We
+	// process, so calling wait() will only produce "huh?" from the OS. We
 	// must rely on the status param passed in, which unfortunately comes
-	// straight from the OS suspend() call, which means we have to decode it by
+	// straight from the OS wait() call, which means we have to decode it by
 	// hand.
 	mStatus = interpret_status(status);
 	LL_INFOS("LLProcess") << getStatusString() << LL_ENDL;
diff --git a/indra/llcommon/llrefcount.h b/indra/llcommon/llrefcount.h
index fb0411d27b..7e4af6ea66 100644
--- a/indra/llcommon/llrefcount.h
+++ b/indra/llcommon/llrefcount.h
@@ -28,9 +28,10 @@
 
 #include <boost/noncopyable.hpp>
 #include <boost/intrusive_ptr.hpp>
-#include "llmutex.h"
 #include "llatomic.h"
 
+class LLMutex;
+
 //----------------------------------------------------------------------------
 // RefCount objects should generally only be accessed by way of LLPointer<>'s
 // see llthread.h for LLThreadSafeRefCount
diff --git a/indra/llcommon/llsdserialize.cpp b/indra/llcommon/llsdserialize.cpp
index 79934642ae..022a5d4659 100644
--- a/indra/llcommon/llsdserialize.cpp
+++ b/indra/llcommon/llsdserialize.cpp
@@ -66,7 +66,8 @@ const std::string LLSD_NOTATION_HEADER("llsd/notation");
  */
 
 // static
-void LLSDSerialize::serialize(const LLSD& sd, std::ostream& str, ELLSD_Serialize type, U32 options)
+void LLSDSerialize::serialize(const LLSD& sd, std::ostream& str, ELLSD_Serialize type,
+							  LLSDFormatter::EFormatterOptions options)
 {
 	LLPointer<LLSDFormatter> f = NULL;
 
@@ -174,10 +175,10 @@ bool LLSDSerialize::deserialize(LLSD& sd, std::istream& str, S32 max_bytes)
 	{
 		p = new LLSDXMLParser;
 	}
-    else if (header == LLSD_NOTATION_HEADER)
-    {
-        p = new LLSDNotationParser;
-    }
+	else if (header == LLSD_NOTATION_HEADER)
+	{
+		p = new LLSDNotationParser;
+	}
 	else
 	{
 		LL_WARNS() << "deserialize request for unknown ELLSD_Serialize" << LL_ENDL;
@@ -1234,9 +1235,11 @@ bool LLSDBinaryParser::parseString(
 /**
  * LLSDFormatter
  */
-LLSDFormatter::LLSDFormatter() :
-	mBoolAlpha(false)
+LLSDFormatter::LLSDFormatter(bool boolAlpha, const std::string& realFmt, EFormatterOptions options):
+    mOptions(options)
 {
+    boolalpha(boolAlpha);
+    realFormat(realFmt);
 }
 
 // virtual
@@ -1253,6 +1256,17 @@ void LLSDFormatter::realFormat(const std::string& format)
 	mRealFormat = format;
 }
 
+S32 LLSDFormatter::format(const LLSD& data, std::ostream& ostr) const
+{
+    // pass options captured by constructor
+    return format(data, ostr, mOptions);
+}
+
+S32 LLSDFormatter::format(const LLSD& data, std::ostream& ostr, EFormatterOptions options) const
+{
+    return format_impl(data, ostr, options, 0);
+}
+
 void LLSDFormatter::formatReal(LLSD::Real real, std::ostream& ostr) const
 {
 	std::string buffer = llformat(mRealFormat.c_str(), real);
@@ -1262,7 +1276,9 @@ void LLSDFormatter::formatReal(LLSD::Real real, std::ostream& ostr) const
 /**
  * LLSDNotationFormatter
  */
-LLSDNotationFormatter::LLSDNotationFormatter()
+LLSDNotationFormatter::LLSDNotationFormatter(bool boolAlpha, const std::string& realFormat,
+                                             EFormatterOptions options):
+    LLSDFormatter(boolAlpha, realFormat, options)
 {
 }
 
@@ -1278,14 +1294,8 @@ std::string LLSDNotationFormatter::escapeString(const std::string& in)
 	return ostr.str();
 }
 
-// virtual
-S32 LLSDNotationFormatter::format(const LLSD& data, std::ostream& ostr, U32 options) const
-{
-	S32 rv = format_impl(data, ostr, options, 0);
-	return rv;
-}
-
-S32 LLSDNotationFormatter::format_impl(const LLSD& data, std::ostream& ostr, U32 options, U32 level) const
+S32 LLSDNotationFormatter::format_impl(const LLSD& data, std::ostream& ostr,
+									   EFormatterOptions options, U32 level) const
 {
 	S32 format_count = 1;
 	std::string pre;
@@ -1406,21 +1416,33 @@ S32 LLSDNotationFormatter::format_impl(const LLSD& data, std::ostream& ostr, U32
 	{
 		// *FIX: memory inefficient.
 		const std::vector<U8>& buffer = data.asBinary();
-		ostr << "b(" << buffer.size() << ")\"";
-		if(buffer.size())
+		if (options & LLSDFormatter::OPTIONS_PRETTY_BINARY)
 		{
-			if (options & LLSDFormatter::OPTIONS_PRETTY_BINARY)
+			ostr << "b16\"";
+			if (! buffer.empty())
 			{
 				std::ios_base::fmtflags old_flags = ostr.flags();
 				ostr.setf( std::ios::hex, std::ios::basefield );
-				ostr << "0x";
+				// It shouldn't strictly matter whether the emitted hex digits
+				// are uppercase; LLSDNotationParser handles either; but as of
+				// 2020-05-13, Python's llbase.llsd requires uppercase hex.
+				ostr << std::uppercase;
+				auto oldfill(ostr.fill('0'));
+				auto oldwidth(ostr.width());
 				for (int i = 0; i < buffer.size(); i++)
 				{
-					ostr << (int) buffer[i];
+					// have to restate setw() before every conversion
+					ostr << std::setw(2) << (int) buffer[i];
 				}
+				ostr.width(oldwidth);
+				ostr.fill(oldfill);
 				ostr.flags(old_flags);
 			}
-			else
+		}
+		else                        // ! OPTIONS_PRETTY_BINARY
+		{
+			ostr << "b(" << buffer.size() << ")\"";
+			if (! buffer.empty())
 			{
 				ostr.write((const char*)&buffer[0], buffer.size());
 			}
@@ -1437,11 +1459,12 @@ S32 LLSDNotationFormatter::format_impl(const LLSD& data, std::ostream& ostr, U32
 	return format_count;
 }
 
-
 /**
  * LLSDBinaryFormatter
  */
-LLSDBinaryFormatter::LLSDBinaryFormatter()
+LLSDBinaryFormatter::LLSDBinaryFormatter(bool boolAlpha, const std::string& realFormat,
+                                         EFormatterOptions options):
+    LLSDFormatter(boolAlpha, realFormat, options)
 {
 }
 
@@ -1450,7 +1473,8 @@ LLSDBinaryFormatter::~LLSDBinaryFormatter()
 { }
 
 // virtual
-S32 LLSDBinaryFormatter::format(const LLSD& data, std::ostream& ostr, U32 options) const
+S32 LLSDBinaryFormatter::format_impl(const LLSD& data, std::ostream& ostr,
+									 EFormatterOptions options, U32 level) const
 {
 	S32 format_count = 1;
 	switch(data.type())
@@ -1466,7 +1490,7 @@ S32 LLSDBinaryFormatter::format(const LLSD& data, std::ostream& ostr, U32 option
 		{
 			ostr.put('k');
 			formatString((*iter).first, ostr);
-			format_count += format((*iter).second, ostr);
+			format_count += format_impl((*iter).second, ostr, options, level+1);
 		}
 		ostr.put('}');
 		break;
@@ -1481,7 +1505,7 @@ S32 LLSDBinaryFormatter::format(const LLSD& data, std::ostream& ostr, U32 option
 		LLSD::array_const_iterator end = data.endArray();
 		for(; iter != end; ++iter)
 		{
-			format_count += format(*iter, ostr);
+			format_count += format_impl(*iter, ostr, options, level+1);
 		}
 		ostr.put(']');
 		break;
diff --git a/indra/llcommon/llsdserialize.h b/indra/llcommon/llsdserialize.h
index fe0f4443ef..d6079fd9fa 100644
--- a/indra/llcommon/llsdserialize.h
+++ b/indra/llcommon/llsdserialize.h
@@ -435,7 +435,8 @@ public:
 	/** 
 	 * @brief Constructor
 	 */
-	LLSDFormatter();
+	LLSDFormatter(bool boolAlpha=false, const std::string& realFormat="",
+				  EFormatterOptions options=OPTIONS_PRETTY_BINARY);
 
 	/** 
 	 * @brief Set the boolean serialization format.
@@ -459,16 +460,38 @@ public:
 	void realFormat(const std::string& format);
 
 	/** 
-	 * @brief Call this method to format an LLSD to a stream.
+	 * @brief Call this method to format an LLSD to a stream with options as
+	 * set by the constructor.
+	 *
+	 * @param data The data to write.
+	 * @param ostr The destination stream for the data.
+	 * @return Returns The number of LLSD objects formatted out
+	 */
+	S32 format(const LLSD& data, std::ostream& ostr) const;
+
+	/** 
+	 * @brief Call this method to format an LLSD to a stream, passing options
+	 * explicitly.
 	 *
 	 * @param data The data to write.
 	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
+	 * @param options OPTIONS_NONE to emit LLSD::Binary as raw bytes
+	 * @return Returns The number of LLSD objects formatted out
 	 */
-	virtual S32 format(const LLSD& data, std::ostream& ostr, U32 options = LLSDFormatter::OPTIONS_NONE) const = 0;
+	virtual S32 format(const LLSD& data, std::ostream& ostr, EFormatterOptions options) const;
 
 protected:
 	/** 
+	 * @brief Implementation to format the data. This is called recursively.
+	 *
+	 * @param data The data to write.
+	 * @param ostr The destination stream for the data.
+	 * @return Returns The number of LLSD objects formatted out
+	 */
+	virtual S32 format_impl(const LLSD& data, std::ostream& ostr, EFormatterOptions options,
+							U32 level) const = 0;
+
+	/** 
 	 * @brief Helper method which appropriately obeys the real format.
 	 *
 	 * @param real The real value to format.
@@ -476,9 +499,9 @@ protected:
 	 */
 	void formatReal(LLSD::Real real, std::ostream& ostr) const;
 
-protected:
 	bool mBoolAlpha;
 	std::string mRealFormat;
+	EFormatterOptions mOptions;
 };
 
 
@@ -498,7 +521,8 @@ public:
 	/** 
 	 * @brief Constructor
 	 */
-	LLSDNotationFormatter();
+	LLSDNotationFormatter(bool boolAlpha=false, const std::string& realFormat="",
+						  EFormatterOptions options=OPTIONS_PRETTY_BINARY);
 
 	/** 
 	 * @brief Helper static method to return a notation escaped string
@@ -512,25 +536,16 @@ public:
 	 */
 	static std::string escapeString(const std::string& in);
 
-	/** 
-	 * @brief Call this method to format an LLSD to a stream.
-	 *
-	 * @param data The data to write.
-	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
-	 */
-	virtual S32 format(const LLSD& data, std::ostream& ostr, U32 options = LLSDFormatter::OPTIONS_NONE) const;
-
 protected:
-
 	/** 
 	 * @brief Implementation to format the data. This is called recursively.
 	 *
 	 * @param data The data to write.
 	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
+	 * @return Returns The number of LLSD objects formatted out
 	 */
-	S32 format_impl(const LLSD& data, std::ostream& ostr, U32 options, U32 level) const;
+	S32 format_impl(const LLSD& data, std::ostream& ostr, EFormatterOptions options,
+					U32 level) const override;
 };
 
 
@@ -550,7 +565,8 @@ public:
 	/** 
 	 * @brief Constructor
 	 */
-	LLSDXMLFormatter();
+	LLSDXMLFormatter(bool boolAlpha=false, const std::string& realFormat="",
+					 EFormatterOptions options=OPTIONS_PRETTY_BINARY);
 
 	/** 
 	 * @brief Helper static method to return an xml escaped string
@@ -565,20 +581,23 @@ public:
 	 *
 	 * @param data The data to write.
 	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
+	 * @return Returns The number of LLSD objects formatted out
 	 */
-	virtual S32 format(const LLSD& data, std::ostream& ostr, U32 options = LLSDFormatter::OPTIONS_NONE) const;
+	S32 format(const LLSD& data, std::ostream& ostr, EFormatterOptions options) const override;
 
-protected:
+	// also pull down base-class format() method that isn't overridden
+	using LLSDFormatter::format;
 
+protected:
 	/** 
 	 * @brief Implementation to format the data. This is called recursively.
 	 *
 	 * @param data The data to write.
 	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
+	 * @return Returns The number of LLSD objects formatted out
 	 */
-	S32 format_impl(const LLSD& data, std::ostream& ostr, U32 options, U32 level) const;
+	S32 format_impl(const LLSD& data, std::ostream& ostr, EFormatterOptions options,
+					U32 level) const override;
 };
 
 
@@ -618,18 +637,20 @@ public:
 	/** 
 	 * @brief Constructor
 	 */
-	LLSDBinaryFormatter();
+	LLSDBinaryFormatter(bool boolAlpha=false, const std::string& realFormat="",
+						EFormatterOptions options=OPTIONS_PRETTY_BINARY);
 
+protected:
 	/** 
-	 * @brief Call this method to format an LLSD to a stream.
+	 * @brief Implementation to format the data. This is called recursively.
 	 *
 	 * @param data The data to write.
 	 * @param ostr The destination stream for the data.
-	 * @return Returns The number of LLSD objects fomatted out
+	 * @return Returns The number of LLSD objects formatted out
 	 */
-	virtual S32 format(const LLSD& data, std::ostream& ostr, U32 options = LLSDFormatter::OPTIONS_NONE) const;
+	S32 format_impl(const LLSD& data, std::ostream& ostr, EFormatterOptions options,
+					U32 level) const override;
 
-protected:
 	/** 
 	 * @brief Helper method to serialize strings
 	 *
@@ -669,7 +690,8 @@ public:
 	/** 
 	 * @brief Constructor
 	 */
-	LLSDOStreamer(const LLSD& data, U32 options = LLSDFormatter::OPTIONS_NONE) :
+	LLSDOStreamer(const LLSD& data,
+				  LLSDFormatter::EFormatterOptions options=LLSDFormatter::OPTIONS_PRETTY_BINARY) :
 		mSD(data), mOptions(options) {}
 
 	/**
@@ -681,17 +703,17 @@ public:
 	 * @return Returns the stream passed in after streaming mSD.
 	 */
 	friend std::ostream& operator<<(
-		std::ostream& str,
-		const LLSDOStreamer<Formatter>& formatter)
+		std::ostream& out,
+		const LLSDOStreamer<Formatter>& streamer)
 	{
 		LLPointer<Formatter> f = new Formatter;
-		f->format(formatter.mSD, str, formatter.mOptions);
-		return str;
+		f->format(streamer.mSD, out, streamer.mOptions);
+		return out;
 	}
 
 protected:
 	LLSD mSD;
-	U32 mOptions;
+	LLSDFormatter::EFormatterOptions mOptions;
 };
 
 typedef LLSDOStreamer<LLSDNotationFormatter>	LLSDNotationStreamer;
@@ -724,7 +746,7 @@ public:
 	 * Generic in/outs
 	 */
 	static void serialize(const LLSD& sd, std::ostream& str, ELLSD_Serialize,
-		U32 options = LLSDFormatter::OPTIONS_NONE);
+						  LLSDFormatter::EFormatterOptions options=LLSDFormatter::OPTIONS_PRETTY_BINARY);
 
 	/**
 	 * @brief Examine a stream, and parse 1 sd object out based on contents.
@@ -752,9 +774,9 @@ public:
 	static S32 toPrettyBinaryNotation(const LLSD& sd, std::ostream& str)
 	{
 		LLPointer<LLSDNotationFormatter> f = new LLSDNotationFormatter;
-		return f->format(sd, str, 
-				LLSDFormatter::OPTIONS_PRETTY | 
-				LLSDFormatter::OPTIONS_PRETTY_BINARY);
+		return f->format(sd, str,
+						 LLSDFormatter::EFormatterOptions(LLSDFormatter::OPTIONS_PRETTY | 
+														  LLSDFormatter::OPTIONS_PRETTY_BINARY));
 	}
 	static S32 fromNotation(LLSD& sd, std::istream& str, S32 max_bytes)
 	{
diff --git a/indra/llcommon/llsdserialize_xml.cpp b/indra/llcommon/llsdserialize_xml.cpp
index 6d0fe862b9..0da824d694 100644
--- a/indra/llcommon/llsdserialize_xml.cpp
+++ b/indra/llcommon/llsdserialize_xml.cpp
@@ -45,7 +45,9 @@ extern "C"
 /**
  * LLSDXMLFormatter
  */
-LLSDXMLFormatter::LLSDXMLFormatter()
+LLSDXMLFormatter::LLSDXMLFormatter(bool boolAlpha, const std::string& realFormat,
+                                   EFormatterOptions options):
+    LLSDFormatter(boolAlpha, realFormat, options)
 {
 }
 
@@ -55,7 +57,8 @@ LLSDXMLFormatter::~LLSDXMLFormatter()
 }
 
 // virtual
-S32 LLSDXMLFormatter::format(const LLSD& data, std::ostream& ostr, U32 options) const
+S32 LLSDXMLFormatter::format(const LLSD& data, std::ostream& ostr,
+							 EFormatterOptions options) const
 {
 	std::streamsize old_precision = ostr.precision(25);
 
@@ -72,7 +75,8 @@ S32 LLSDXMLFormatter::format(const LLSD& data, std::ostream& ostr, U32 options)
 	return rv;
 }
 
-S32 LLSDXMLFormatter::format_impl(const LLSD& data, std::ostream& ostr, U32 options, U32 level) const
+S32 LLSDXMLFormatter::format_impl(const LLSD& data, std::ostream& ostr,
+								  EFormatterOptions options, U32 level) const
 {
 	S32 format_count = 1;
 	std::string pre;
diff --git a/indra/llcommon/llsdutil.cpp b/indra/llcommon/llsdutil.cpp
index 6a23c443a0..d44387cc55 100644
--- a/indra/llcommon/llsdutil.cpp
+++ b/indra/llcommon/llsdutil.cpp
@@ -856,6 +856,74 @@ bool llsd_equals(const LLSD& lhs, const LLSD& rhs, int bits)
     }
 }
 
+/*****************************************************************************
+*   llsd::drill()
+*****************************************************************************/
+namespace llsd
+{
+
+LLSD& drill(LLSD& blob, const LLSD& rawPath)
+{
+    // Treat rawPath uniformly as an array. If it's not already an array,
+    // store it as the only entry in one. (But let's say Undefined means an
+    // empty array.)
+    LLSD path;
+    if (rawPath.isArray() || rawPath.isUndefined())
+    {
+        path = rawPath;
+    }
+    else
+    {
+        path.append(rawPath);
+    }
+
+    // Need to indicate a current destination -- but that current destination
+    // must change as we step through the path array. Where normally we'd use
+    // an LLSD& to capture a subscripted LLSD lvalue, this time we must
+    // instead use a pointer -- since it must be reassigned.
+    // Start by pointing to the input blob exactly as is.
+    LLSD* located{&blob};
+
+    // Extract the element of interest by walking path. Use an explicit index
+    // so that, in case of a bogus type in path, we can identify the specific
+    // path entry that's bad.
+    for (LLSD::Integer i = 0; i < path.size(); ++i)
+    {
+        const LLSD& key{path[i]};
+        if (key.isString())
+        {
+            // a string path element is a map key
+            located = &((*located)[key.asString()]);
+        }
+        else if (key.isInteger())
+        {
+            // an integer path element is an array index
+            located = &((*located)[key.asInteger()]);
+        }
+        else
+        {
+            // What do we do with Real or Array or Map or ...?
+            // As it's a coder error -- not a user error -- rub the coder's
+            // face in it so it gets fixed.
+            LL_ERRS("llsdutil") << "drill(" << blob << ", " << rawPath
+                                << "): path[" << i << "] bad type "
+                                << sTypes.lookup(key.type()) << LL_ENDL;
+        }
+    }
+
+    // dereference the pointer to return a reference to the element we found
+    return *located;
+}
+
+LLSD drill(const LLSD& blob, const LLSD& path)
+{
+    // non-const drill() does exactly what we want. Temporarily cast away
+    // const-ness and use that.
+    return drill(const_cast<LLSD&>(blob), path);
+}
+
+} // namespace llsd
+
 // Construct a deep partial clone of of an LLSD object. primitive types share 
 // references, however maps, arrays and binary objects are duplicated. An optional
 // filter may be include to exclude/include keys in a map. 
diff --git a/indra/llcommon/llsdutil.h b/indra/llcommon/llsdutil.h
index 863be04c8a..84be95ba54 100644
--- a/indra/llcommon/llsdutil.h
+++ b/indra/llcommon/llsdutil.h
@@ -143,6 +143,16 @@ LL_COMMON_API std::string llsd_matches(const LLSD& prototype, const LLSD& data,
 /// equality rather than bitwise equality, pass @a bits as for
 /// is_approx_equal_fraction().
 LL_COMMON_API bool llsd_equals(const LLSD& lhs, const LLSD& rhs, int bits=-1);
+/// If you don't care about LLSD::Real equality
+inline bool operator==(const LLSD& lhs, const LLSD& rhs)
+{
+    return llsd_equals(lhs, rhs);
+}
+inline bool operator!=(const LLSD& lhs, const LLSD& rhs)
+{
+    // operator!=() should always be the negation of operator==()
+    return ! (lhs == rhs);
+}
 
 // Simple function to copy data out of input & output iterators if
 // there is no need for casting.
@@ -156,6 +166,31 @@ template<typename Input> LLSD llsd_copy_array(Input iter, Input end)
 	return dest;
 }
 
+namespace llsd
+{
+
+/**
+ * Drill down to locate an element in 'blob' according to 'path', where 'path'
+ * is one of the following:
+ *
+ * - LLSD::String: 'blob' is an LLSD::Map. Find the entry with key 'path'.
+ * - LLSD::Integer: 'blob' is an LLSD::Array. Find the entry with index 'path'.
+ * - Any other 'path' type will be interpreted as LLSD::Array, and 'blob' is a
+ *   nested structure. For each element of 'path':
+ *   - If it's an LLSD::Integer, select the entry with that index from an
+ *     LLSD::Array at that level.
+ *   - If it's an LLSD::String, select the entry with that key from an
+ *     LLSD::Map at that level.
+ *   - Anything else is an error.
+ *
+ * By implication, if path.isUndefined() or otherwise equivalent to an empty
+ * LLSD::Array, drill() returns 'blob' as is.
+ */
+LLSD  drill(const LLSD& blob, const LLSD& path);
+LLSD& drill(      LLSD& blob, const LLSD& path);
+
+}
+
 /*****************************************************************************
 *   LLSDArray
 *****************************************************************************/
@@ -225,6 +260,36 @@ private:
     LLSD _data;
 };
 
+namespace llsd
+{
+
+/**
+ * Construct an LLSD::Array inline, using modern C++ variadic arguments.
+ */
+
+// recursion tail
+inline
+void array_(LLSD&) {}
+
+// recursive call
+template <typename T0, typename... Ts>
+void array_(LLSD& data, T0&& v0, Ts&&... vs)
+{
+    data.append(std::forward<T0>(v0));
+    array_(data, std::forward<Ts>(vs)...);
+}
+
+// public interface
+template <typename... Ts>
+LLSD array(Ts&&... vs)
+{
+    LLSD data;
+    array_(data, std::forward<Ts>(vs)...);
+    return data;
+}
+
+} // namespace llsd
+
 /*****************************************************************************
 *   LLSDMap
 *****************************************************************************/
@@ -269,6 +334,36 @@ private:
     LLSD _data;
 };
 
+namespace llsd
+{
+
+/**
+ * Construct an LLSD::Map inline, using modern C++ variadic arguments.
+ */
+
+// recursion tail
+inline
+void map_(LLSD&) {}
+
+// recursive call
+template <typename T0, typename... Ts>
+void map_(LLSD& data, const LLSD::String& k0, T0&& v0, Ts&&... vs)
+{
+    data[k0] = v0;
+    map_(data, std::forward<Ts>(vs)...);
+}
+
+// public interface
+template <typename... Ts>
+LLSD map(Ts&&... vs)
+{
+    LLSD data;
+    map_(data, std::forward<Ts>(vs)...);
+    return data;
+}
+
+} // namespace llsd
+
 /*****************************************************************************
 *   LLSDParam
 *****************************************************************************/
@@ -452,6 +547,16 @@ LLSD llsd_clone(LLSD value, LLSD filter = LLSD());
 // the filter parameter.
 LLSD llsd_shallow(LLSD value, LLSD filter = LLSD());
 
+namespace llsd
+{
+
+// llsd namespace aliases
+inline
+LLSD clone  (LLSD value, LLSD filter=LLSD()) { return llsd_clone  (value, filter); }
+inline
+LLSD shallow(LLSD value, LLSD filter=LLSD()) { return llsd_shallow(value, filter); }
+
+} // namespace llsd
 
 // Specialization for generating a hash value from an LLSD block. 
 template <>
diff --git a/indra/llcommon/llsingleton.cpp b/indra/llcommon/llsingleton.cpp
index c45c144570..d3d25201b2 100644
--- a/indra/llcommon/llsingleton.cpp
+++ b/indra/llcommon/llsingleton.cpp
@@ -30,9 +30,9 @@
 #include "llerror.h"
 #include "llerrorcontrol.h"         // LLError::is_available()
 #include "lldependencies.h"
-#include "llcoro_get_id.h"
+#include "llexception.h"
+#include "llcoros.h"
 #include <boost/foreach.hpp>
-#include <boost/unordered_map.hpp>
 #include <algorithm>
 #include <iostream>                 // std::cerr in dire emergency
 #include <sstream>
@@ -42,8 +42,6 @@ namespace {
 void log(LLError::ELevel level,
          const char* p1, const char* p2, const char* p3, const char* p4);
 
-void logdebugs(const char* p1="", const char* p2="", const char* p3="", const char* p4="");
-
 bool oktolog();
 } // anonymous namespace
 
@@ -57,63 +55,131 @@ bool oktolog();
 class LLSingletonBase::MasterList:
     public LLSingleton<LLSingletonBase::MasterList>
 {
+private:
     LLSINGLETON_EMPTY_CTOR(MasterList);
 
-public:
-    // No need to make this private with accessors; nobody outside this source
-    // file can see it.
+    // Independently of the LLSingleton locks governing construction,
+    // destruction and other state changes of the MasterList instance itself,
+    // we must also defend each of the data structures owned by the
+    // MasterList.
+    // This must be a recursive_mutex because, while the lock is held for
+    // manipulating some data in the master list, we must also check whether
+    // it's safe to log -- which involves querying a different LLSingleton --
+    // which requires accessing the master list.
+    typedef std::recursive_mutex mutex_t;
+    typedef std::unique_lock<mutex_t> lock_t;
 
+    mutex_t mMutex;
+
+public:
+    // Instantiate this to both obtain a reference to MasterList::instance()
+    // and lock its mutex for the lifespan of this Lock instance.
+    class Lock
+    {
+    public:
+        Lock():
+            mMasterList(MasterList::instance()),
+            mLock(mMasterList.mMutex)
+        {}
+        Lock(const Lock&) = delete;
+        Lock& operator=(const Lock&) = delete;
+        MasterList& get() const { return mMasterList; }
+        operator MasterList&() const { return get(); }
+
+    protected:
+        MasterList& mMasterList;
+        MasterList::lock_t mLock;
+    };
+
+private:
     // This is the master list of all instantiated LLSingletons (save the
     // MasterList itself) in arbitrary order. You MUST call dep_sort() before
     // traversing this list.
-    LLSingletonBase::list_t mMaster;
+    list_t mMaster;
+
+public:
+    // Instantiate this to obtain a reference to MasterList::mMaster and to
+    // hold the MasterList lock for the lifespan of this LockedMaster
+    // instance.
+    struct LockedMaster: public Lock
+    {
+        list_t& get() const { return mMasterList.mMaster; }
+        operator list_t&() const { return get(); }
+    };
 
+private:
     // We need to maintain a stack of LLSingletons currently being
     // initialized, either in the constructor or in initSingleton(). However,
     // managing that as a stack depends on having a DISTINCT 'initializing'
     // stack for every C++ stack in the process! And we have a distinct C++
-    // stack for every running coroutine. It would be interesting and cool to
-    // implement a generic coroutine-local-storage mechanism and use that
-    // here. The trouble is that LLCoros is itself an LLSingleton, so
-    // depending on LLCoros functionality could dig us into infinite
-    // recursion. (Moreover, when we reimplement LLCoros on top of
-    // Boost.Fiber, that library already provides fiber_specific_ptr -- so
-    // it's not worth a great deal of time and energy implementing a generic
-    // equivalent on top of boost::dcoroutine, which is on its way out.)
-    // Instead, use a map of llcoro::id to select the appropriate
-    // coro-specific 'initializing' stack. llcoro::get_id() is carefully
-    // implemented to avoid requiring LLCoros.
-    typedef boost::unordered_map<llcoro::id, LLSingletonBase::list_t> InitializingMap;
-    InitializingMap mInitializing;
-
-    // non-static method, cf. LLSingletonBase::get_initializing()
+    // stack for every running coroutine. Therefore this stack must be based
+    // on a coroutine-local pointer.
+    // This local_ptr isn't static because it's a member of an LLSingleton.
+    LLCoros::local_ptr<list_t> mInitializing;
+
+public:
+    // Instantiate this to obtain a reference to the coroutine-specific
+    // initializing list and to hold the MasterList lock for the lifespan of
+    // this LockedInitializing instance.
+    struct LockedInitializing: public Lock
+    {
+    public:
+        LockedInitializing():
+            // only do the lookup once, cache the result
+            // note that the lock is already locked during this lookup
+            mList(&mMasterList.get_initializing_())
+        {}
+        list_t& get() const
+        {
+            if (! mList)
+            {
+                LLTHROW(LLException("Trying to use LockedInitializing "
+                                    "after cleanup_initializing()"));
+            }
+            return *mList;
+        }
+        operator list_t&() const { return get(); }
+        void log(const char* verb, const char* name);
+        void cleanup_initializing()
+        {
+            mMasterList.cleanup_initializing_();
+            mList = nullptr;
+        }
+
+    private:
+        // Store pointer since cleanup_initializing() must clear it.
+        list_t* mList;
+    };
+
+private:
     list_t& get_initializing_()
     {
-        // map::operator[] has find-or-create semantics, exactly what we need
-        // here. It returns a reference to the selected mapped_type instance.
-        return mInitializing[llcoro::get_id()];
+        LLSingletonBase::list_t* current = mInitializing.get();
+        if (! current)
+        {
+            // If the running coroutine doesn't already have an initializing
+            // stack, allocate a new one and save it for future reference.
+            current = new LLSingletonBase::list_t();
+            mInitializing.reset(current);
+        }
+        return *current;
     }
 
+    // By the time mInitializing is destroyed, its value for every coroutine
+    // except the running one must have been reset() to nullptr. So every time
+    // we pop the list to empty, reset() the running coroutine's local_ptr.
     void cleanup_initializing_()
     {
-        InitializingMap::iterator found = mInitializing.find(llcoro::get_id());
-        if (found != mInitializing.end())
-        {
-            mInitializing.erase(found);
-        }
+        mInitializing.reset(nullptr);
     }
 };
 
-//static
-LLSingletonBase::list_t& LLSingletonBase::get_master()
-{
-    return LLSingletonBase::MasterList::instance().mMaster;
-}
-
 void LLSingletonBase::add_master()
 {
     // As each new LLSingleton is constructed, add to the master list.
-    get_master().push_back(this);
+    // This temporary LockedMaster should suffice to hold the MasterList lock
+    // during the push_back() call.
+    MasterList::LockedMaster().get().push_back(this);
 }
 
 void LLSingletonBase::remove_master()
@@ -125,27 +191,32 @@ void LLSingletonBase::remove_master()
     // master list, and remove this item IF FOUND. We have few enough
     // LLSingletons, and they are so rarely destroyed (once per run), that the
     // cost of a linear search should not be an issue.
-    get_master().remove(this);
+    // This temporary LockedMaster should suffice to hold the MasterList lock
+    // during the remove() call.
+    MasterList::LockedMaster().get().remove(this);
 }
 
 //static
-LLSingletonBase::list_t& LLSingletonBase::get_initializing()
+LLSingletonBase::list_t::size_type LLSingletonBase::get_initializing_size()
 {
-    return LLSingletonBase::MasterList::instance().get_initializing_();
+    return MasterList::LockedInitializing().get().size();
 }
 
 LLSingletonBase::~LLSingletonBase() {}
 
 void LLSingletonBase::push_initializing(const char* name)
 {
+    MasterList::LockedInitializing locked_list;
     // log BEFORE pushing so logging singletons don't cry circularity
-    log_initializing("Pushing", name);
-    get_initializing().push_back(this);
+    locked_list.log("Pushing", name);
+    locked_list.get().push_back(this);
 }
 
 void LLSingletonBase::pop_initializing()
 {
-    list_t& list(get_initializing());
+    // Lock the MasterList for the duration of this call
+    MasterList::LockedInitializing locked_list;
+    list_t& list(locked_list.get());
 
     if (list.empty())
     {
@@ -165,7 +236,7 @@ void LLSingletonBase::pop_initializing()
     // entirely.
     if (list.empty())
     {
-        MasterList::instance().cleanup_initializing_();
+        locked_list.cleanup_initializing();
     }
 
     // Now validate the newly-popped LLSingleton.
@@ -177,7 +248,7 @@ void LLSingletonBase::pop_initializing()
     }
 
     // log AFTER popping so logging singletons don't cry circularity
-    log_initializing("Popping", typeid(*back).name());
+    locked_list.log("Popping", typeid(*back).name());
 }
 
 void LLSingletonBase::reset_initializing(list_t::size_type size)
@@ -191,7 +262,8 @@ void LLSingletonBase::reset_initializing(list_t::size_type size)
     // push_initializing() call in LLSingletonBase's constructor. So only
     // remove the stack top if in fact we've pushed something more than the
     // previous size.
-    list_t& list(get_initializing());
+    MasterList::LockedInitializing locked_list;
+    list_t& list(locked_list.get());
 
     while (list.size() > size)
     {
@@ -201,29 +273,32 @@ void LLSingletonBase::reset_initializing(list_t::size_type size)
     // as in pop_initializing()
     if (list.empty())
     {
-        MasterList::instance().cleanup_initializing_();
+        locked_list.cleanup_initializing();
     }
 }
 
-//static
-void LLSingletonBase::log_initializing(const char* verb, const char* name)
+void LLSingletonBase::MasterList::LockedInitializing::log(const char* verb, const char* name)
 {
     if (oktolog())
     {
         LL_DEBUGS("LLSingleton") << verb << ' ' << demangle(name) << ';';
-        list_t& list(get_initializing());
-        for (list_t::const_reverse_iterator ri(list.rbegin()), rend(list.rend());
-             ri != rend; ++ri)
+        if (mList)
         {
-            LLSingletonBase* sb(*ri);
-            LL_CONT << ' ' << classname(sb);
+            for (list_t::const_reverse_iterator ri(mList->rbegin()), rend(mList->rend());
+                 ri != rend; ++ri)
+            {
+                LLSingletonBase* sb(*ri);
+                LL_CONT << ' ' << classname(sb);
+            }
         }
         LL_ENDL;
     }
 }
 
-void LLSingletonBase::capture_dependency(list_t& initializing, EInitState initState)
+void LLSingletonBase::capture_dependency()
 {
+    MasterList::LockedInitializing locked_list;
+    list_t& initializing(locked_list.get());
     // Did this getInstance() call come from another LLSingleton, or from
     // vanilla application code? Note that although this is a nontrivial
     // method, the vast majority of its calls arrive here with initializing
@@ -252,21 +327,8 @@ void LLSingletonBase::capture_dependency(list_t& initializing, EInitState initSt
                 LLSingletonBase* foundp(*found);
                 out << classname(foundp) << " -> ";
             }
-            // We promise to capture dependencies from both the constructor
-            // and the initSingleton() method, so an LLSingleton's instance
-            // pointer is on the initializing list during both. Now that we've
-            // detected circularity, though, we must distinguish the two. If
-            // the recursive call is from the constructor, we CAN'T honor it:
-            // otherwise we'd be returning a pointer to a partially-
-            // constructed object! But from initSingleton() is okay: that
-            // method exists specifically to support circularity.
             // Decide which log helper to call.
-            if (initState == CONSTRUCTING)
-            {
-                logerrs("LLSingleton circularity in Constructor: ", out.str().c_str(),
-                    classname(this).c_str(), "");
-            }
-            else if (it_next == initializing.end())
+            if (it_next == initializing.end())
             {
                 // Points to self after construction, but during initialization.
                 // Singletons can initialize other classes that depend onto them,
@@ -309,12 +371,12 @@ LLSingletonBase::vec_t LLSingletonBase::dep_sort()
     // SingletonDeps through the life of the program, dynamically adding and
     // removing LLSingletons as they are created and destroyed, in practice
     // it's less messy to construct it on demand. The overhead of doing so
-    // should happen basically twice: once for cleanupAll(), once for
-    // deleteAll().
+    // should happen basically once: for deleteAll().
     typedef LLDependencies<LLSingletonBase*> SingletonDeps;
     SingletonDeps sdeps;
-    list_t& master(get_master());
-    BOOST_FOREACH(LLSingletonBase* sp, master)
+    // Lock while traversing the master list 
+    MasterList::LockedMaster master;
+    for (LLSingletonBase* sp : master.get())
     {
         // Build the SingletonDeps structure by adding, for each
         // LLSingletonBase* sp in the master list, sp itself. It has no
@@ -326,51 +388,32 @@ LLSingletonBase::vec_t LLSingletonBase::dep_sort()
                   SingletonDeps::KeyList(sp->mDepends.begin(), sp->mDepends.end()));
     }
     vec_t ret;
-    ret.reserve(master.size());
+    ret.reserve(master.get().size());
     // We should be able to effect this with a transform_iterator that
     // extracts just the first (key) element from each sorted_iterator, then
     // uses vec_t's range constructor... but frankly this is more
     // straightforward, as long as we remember the above reserve() call!
-    BOOST_FOREACH(SingletonDeps::sorted_iterator::value_type pair, sdeps.sort())
+    for (const SingletonDeps::sorted_iterator::value_type& pair : sdeps.sort())
     {
         ret.push_back(pair.first);
     }
     // The master list is not itself pushed onto the master list. Add it as
     // the very last entry -- it is the LLSingleton on which ALL others
     // depend! -- so our caller will process it.
-    ret.push_back(MasterList::getInstance());
+    ret.push_back(&master.Lock::get());
     return ret;
 }
 
-//static
-void LLSingletonBase::cleanupAll()
+void LLSingletonBase::cleanup_()
 {
-    // It's essential to traverse these in dependency order.
-    BOOST_FOREACH(LLSingletonBase* sp, dep_sort())
+    logdebugs("calling ", classname(this).c_str(), "::cleanupSingleton()");
+    try
     {
-        // Call cleanupSingleton() only if we haven't already done so for this
-        // instance.
-        if (! sp->mCleaned)
-        {
-            sp->mCleaned = true;
-
-            logdebugs("calling ",
-                      classname(sp).c_str(), "::cleanupSingleton()");
-            try
-            {
-                sp->cleanupSingleton();
-            }
-            catch (const std::exception& e)
-            {
-                logwarns("Exception in ", classname(sp).c_str(),
-                         "::cleanupSingleton(): ", e.what());
-            }
-            catch (...)
-            {
-                logwarns("Unknown exception in ", classname(sp).c_str(),
-                         "::cleanupSingleton()");
-            }
-        }
+        cleanupSingleton();
+    }
+    catch (...)
+    {
+        LOG_UNHANDLED_EXCEPTION(classname(this) + "::cleanupSingleton()");
     }
 }
 
@@ -441,10 +484,6 @@ void log(LLError::ELevel level,
     }
 }
 
-void logdebugs(const char* p1, const char* p2, const char* p3, const char* p4)
-{
-    log(LLError::LEVEL_DEBUG, p1, p2, p3, p4);
-}
 } // anonymous namespace        
 
 //static
@@ -454,6 +493,18 @@ void LLSingletonBase::logwarns(const char* p1, const char* p2, const char* p3, c
 }
 
 //static
+void LLSingletonBase::loginfos(const char* p1, const char* p2, const char* p3, const char* p4)
+{
+    log(LLError::LEVEL_INFO, p1, p2, p3, p4);
+}
+
+//static
+void LLSingletonBase::logdebugs(const char* p1, const char* p2, const char* p3, const char* p4)
+{
+    log(LLError::LEVEL_DEBUG, p1, p2, p3, p4);
+}
+
+//static
 void LLSingletonBase::logerrs(const char* p1, const char* p2, const char* p3, const char* p4)
 {
     log(LLError::LEVEL_ERROR, p1, p2, p3, p4);
diff --git a/indra/llcommon/llsingleton.h b/indra/llcommon/llsingleton.h
index 7def9b019c..30a5b21cf8 100644
--- a/indra/llcommon/llsingleton.h
+++ b/indra/llcommon/llsingleton.h
@@ -30,18 +30,10 @@
 #include <list>
 #include <vector>
 #include <typeinfo>
-
-#if LL_WINDOWS
-#pragma warning (push)
-#pragma warning (disable:4265)
-#endif
-// warning C4265: 'std::_Pad' : class has virtual functions, but destructor is not virtual
-
-#include <mutex>
-
-#if LL_WINDOWS
-#pragma warning (pop)
-#endif
+#include "mutex.h"
+#include "lockstatic.h"
+#include "llthread.h"               // on_main_thread()
+#include "llmainthreadtask.h"
 
 class LLSingletonBase: private boost::noncopyable
 {
@@ -51,15 +43,13 @@ public:
 private:
     // All existing LLSingleton instances are tracked in this master list.
     typedef std::list<LLSingletonBase*> list_t;
-    static list_t& get_master();
-    // This, on the other hand, is a stack whose top indicates the LLSingleton
-    // currently being initialized.
-    static list_t& get_initializing();
+    // Size of stack whose top indicates the LLSingleton currently being
+    // initialized.
+    static list_t::size_type get_initializing_size();
     // Produce a vector<LLSingletonBase*> of master list, in dependency order.
     typedef std::vector<LLSingletonBase*> vec_t;
     static vec_t dep_sort();
 
-    bool mCleaned;                  // cleanupSingleton() has been called
     // we directly depend on these other LLSingletons
     typedef boost::unordered_set<LLSingletonBase*> set_t;
     set_t mDepends;
@@ -68,8 +58,8 @@ protected:
     typedef enum e_init_state
     {
         UNINITIALIZED = 0,          // must be default-initialized state
+        QUEUED,                     // construction queued, not yet executing
         CONSTRUCTING,               // within DERIVED_TYPE constructor
-        CONSTRUCTED,                // finished DERIVED_TYPE constructor
         INITIALIZING,               // within DERIVED_TYPE::initSingleton()
         INITIALIZED,                // normal case
         DELETED                     // deleteSingleton() or deleteAll() called
@@ -115,21 +105,23 @@ protected:
     // Remove 'this' from the init stack in case of exception in the
     // LLSingleton subclass constructor.
     static void reset_initializing(list_t::size_type size);
-private:
-    // logging
-    static void log_initializing(const char* verb, const char* name);
 protected:
     // If a given call to B::getInstance() happens during either A::A() or
     // A::initSingleton(), record that A directly depends on B.
-    void capture_dependency(list_t& initializing, EInitState);
+    void capture_dependency();
 
-    // delegate LL_ERRS() logging to llsingleton.cpp
+    // delegate logging calls to llsingleton.cpp
     static void logerrs(const char* p1, const char* p2="",
                         const char* p3="", const char* p4="");
-    // delegate LL_WARNS() logging to llsingleton.cpp
     static void logwarns(const char* p1, const char* p2="",
                          const char* p3="", const char* p4="");
+    static void loginfos(const char* p1, const char* p2="",
+                         const char* p3="", const char* p4="");
+    static void logdebugs(const char* p1, const char* p2="",
+                          const char* p3="", const char* p4="");
     static std::string demangle(const char* mangled);
+    // these classname() declarations restate template functions declared in
+    // llerror.h because we avoid #including that here
     template <typename T>
     static std::string classname()       { return demangle(typeid(T).name()); }
     template <typename T>
@@ -139,6 +131,9 @@ protected:
     virtual void initSingleton() {}
     virtual void cleanupSingleton() {}
 
+    // internal wrapper around calls to cleanupSingleton()
+    void cleanup_();
+
     // deleteSingleton() isn't -- and shouldn't be -- a virtual method. It's a
     // class static. However, given only Foo*, deleteAll() does need to be
     // able to reach Foo::deleteSingleton(). Make LLSingleton (which declares
@@ -148,32 +143,15 @@ protected:
 
 public:
     /**
-     * Call this to call the cleanupSingleton() method for every LLSingleton
-     * constructed since the start of the last cleanupAll() call. (Any
-     * LLSingleton constructed DURING a cleanupAll() call won't be cleaned up
-     * until the next cleanupAll() call.) cleanupSingleton() neither deletes
-     * nor destroys its LLSingleton; therefore it's safe to include logic that
-     * might take significant realtime or even throw an exception.
-     *
-     * The most important property of cleanupAll() is that cleanupSingleton()
-     * methods are called in dependency order, leaf classes last. Thus, given
-     * two LLSingleton subclasses A and B, if A's dependency on B is properly
-     * expressed as a B::getInstance() or B::instance() call during either
-     * A::A() or A::initSingleton(), B will be cleaned up after A.
-     *
-     * If a cleanupSingleton() method throws an exception, the exception is
-     * logged, but cleanupAll() attempts to continue calling the rest of the
-     * cleanupSingleton() methods.
-     */
-    static void cleanupAll();
-    /**
-     * Call this to call the deleteSingleton() method for every LLSingleton
-     * constructed since the start of the last deleteAll() call. (Any
-     * LLSingleton constructed DURING a deleteAll() call won't be cleaned up
-     * until the next deleteAll() call.) deleteSingleton() deletes and
-     * destroys its LLSingleton. Any cleanup logic that might take significant
-     * realtime -- or throw an exception -- must not be placed in your
-     * LLSingleton's destructor, but rather in its cleanupSingleton() method.
+     * deleteAll() calls the cleanupSingleton() and deleteSingleton() methods
+     * for every LLSingleton constructed since the start of the last
+     * deleteAll() call. (Any LLSingleton constructed DURING a deleteAll()
+     * call won't be cleaned up until the next deleteAll() call.)
+     * deleteSingleton() deletes and destroys its LLSingleton. Any cleanup
+     * logic that might take significant realtime -- or throw an exception --
+     * must not be placed in your LLSingleton's destructor, but rather in its
+     * cleanupSingleton() method, which is called implicitly by
+     * deleteSingleton().
      *
      * The most important property of deleteAll() is that deleteSingleton()
      * methods are called in dependency order, leaf classes last. Thus, given
@@ -181,9 +159,9 @@ public:
      * expressed as a B::getInstance() or B::instance() call during either
      * A::A() or A::initSingleton(), B will be cleaned up after A.
      *
-     * If a deleteSingleton() method throws an exception, the exception is
-     * logged, but deleteAll() attempts to continue calling the rest of the
-     * deleteSingleton() methods.
+     * If a cleanupSingleton() or deleteSingleton() method throws an
+     * exception, the exception is logged, but deleteAll() attempts to
+     * continue calling the rest of the deleteSingleton() methods.
      */
     static void deleteAll();
 };
@@ -203,9 +181,16 @@ struct LLSingleton_manage_master
     {
         LLSingletonBase::reset_initializing(size);
     }
-    // For any LLSingleton subclass except the MasterList, obtain the init
-    // stack from the MasterList singleton instance.
-    LLSingletonBase::list_t& get_initializing() { return LLSingletonBase::get_initializing(); }
+    // For any LLSingleton subclass except the MasterList, obtain the size of
+    // the init stack from the MasterList singleton instance.
+    LLSingletonBase::list_t::size_type get_initializing_size()
+    {
+        return LLSingletonBase::get_initializing_size();
+    }
+    void capture_dependency(LLSingletonBase* sb)
+    {
+        sb->capture_dependency();
+    }
 };
 
 // But for the specific case of LLSingletonBase::MasterList, don't.
@@ -218,20 +203,14 @@ struct LLSingleton_manage_master<LLSingletonBase::MasterList>
     void pop_initializing (LLSingletonBase*) {}
     // since we never pushed, no need to clean up
     void reset_initializing(LLSingletonBase::list_t::size_type size) {}
-    LLSingletonBase::list_t& get_initializing()
-    {
-        // The MasterList shouldn't depend on any other LLSingletons. We'd
-        // get into trouble if we tried to recursively engage that machinery.
-        static LLSingletonBase::list_t sDummyList;
-        return sDummyList;
-    }
+    LLSingletonBase::list_t::size_type get_initializing_size() { return 0; }
+    void capture_dependency(LLSingletonBase*) {}
 };
 
 // Now we can implement LLSingletonBase's template constructor.
 template <typename DERIVED_TYPE>
 LLSingletonBase::LLSingletonBase(tag<DERIVED_TYPE>):
-    mCleaned(false),
-    mDeleteSingleton(NULL)
+    mDeleteSingleton(nullptr)
 {
     // This is the earliest possible point at which we can push this new
     // instance onto the init stack. LLSingleton::constructSingleton() can't
@@ -273,10 +252,19 @@ class LLParamSingleton;
  * leading back to yours, move the instance reference from your constructor to
  * your initSingleton() method.
  *
- * If you override LLSingleton<T>::cleanupSingleton(), your method will be
- * called if someone calls LLSingletonBase::cleanupAll(). The significant part
- * of this promise is that cleanupAll() will call individual
- * cleanupSingleton() methods in reverse dependency order.
+ * If you override LLSingleton<T>::cleanupSingleton(), your method will
+ * implicitly be called by LLSingleton<T>::deleteSingleton() just before the
+ * instance is destroyed. We introduce a special cleanupSingleton() method
+ * because cleanupSingleton() operations can involve nontrivial realtime, or
+ * throw an exception. A destructor should do neither!
+ *
+ * If your cleanupSingleton() method throws an exception, we log that
+ * exception but carry on.
+ *
+ * If at some point you call LLSingletonBase::deleteAll(), all remaining
+ * LLSingleton<T> instances will be destroyed in reverse dependency order. (Or
+ * call MySubclass::deleteSingleton() to specifically destroy the canonical
+ * MySubclass instance.)
  *
  * That is, consider LLSingleton subclasses C, B and A. A depends on B, which
  * in turn depends on C. These dependencies are expressed as calls to
@@ -284,33 +272,34 @@ class LLParamSingleton;
  * It shouldn't matter whether these calls appear in A::A() or
  * A::initSingleton(), likewise B::B() or B::initSingleton().
  *
- * We promise that if you later call LLSingletonBase::cleanupAll():
- * 1. A::cleanupSingleton() will be called before
- * 2. B::cleanupSingleton(), which will be called before
- * 3. C::cleanupSingleton().
+ * We promise that if you later call LLSingletonBase::deleteAll():
+ * 1. A::deleteSingleton() will be called before
+ * 2. B::deleteSingleton(), which will be called before
+ * 3. C::deleteSingleton().
  * Put differently, if your LLSingleton subclass constructor or
  * initSingleton() method explicitly depends on some other LLSingleton
  * subclass, you may continue to rely on that other subclass in your
  * cleanupSingleton() method.
- *
- * We introduce a special cleanupSingleton() method because cleanupSingleton()
- * operations can involve nontrivial realtime, or might throw an exception. A
- * destructor should do neither!
- *
- * If your cleanupSingleton() method throws an exception, we log that
- * exception but proceed with the remaining cleanupSingleton() calls.
- *
- * Similarly, if at some point you call LLSingletonBase::deleteAll(), all
- * remaining LLSingleton instances will be destroyed in dependency order. (Or
- * call MySubclass::deleteSingleton() to specifically destroy the canonical
- * MySubclass instance.)
- *
- * As currently written, LLSingleton is not thread-safe.
  */
 template <typename DERIVED_TYPE>
 class LLSingleton : public LLSingletonBase
 {
 private:
+    // LLSingleton<DERIVED_TYPE> must have a distinct instance of
+    // SingletonData for every distinct DERIVED_TYPE. It's tempting to
+    // consider hoisting SingletonData up into LLSingletonBase. Don't do it.
+    struct SingletonData
+    {
+        // Use a recursive_mutex in case of constructor circularity. With a
+        // non-recursive mutex, that would result in deadlock.
+        typedef std::recursive_mutex mutex_t;
+        mutex_t mMutex;             // LockStatic looks for mMutex
+
+        EInitState      mInitState{UNINITIALIZED};
+        DERIVED_TYPE*   mInstance{nullptr};
+    };
+    typedef llthread::LockStatic<SingletonData> LockStatic;
+
     // Allow LLParamSingleton subclass -- but NOT DERIVED_TYPE itself -- to
     // access our private members.
     friend class LLParamSingleton<DERIVED_TYPE>;
@@ -319,17 +308,17 @@ private:
     // purpose for its subclass LLParamSingleton is to support Singletons
     // requiring constructor arguments. constructSingleton() supports both use
     // cases.
+    // Accepting LockStatic& requires that the caller has already locked our
+    // static data before calling.
     template <typename... Args>
-    static void constructSingleton(Args&&... args)
+    static void constructSingleton(LockStatic& lk, Args&&... args)
     {
-        auto prev_size = LLSingleton_manage_master<DERIVED_TYPE>().get_initializing().size();
-        // getInstance() calls are from within constructor
-        sData.mInitState = CONSTRUCTING;
+        auto prev_size = LLSingleton_manage_master<DERIVED_TYPE>().get_initializing_size();
+        // Any getInstance() calls after this point are from within constructor
+        lk->mInitState = CONSTRUCTING;
         try
         {
-            sData.mInstance = new DERIVED_TYPE(std::forward<Args>(args)...);
-            // we have called constructor, have not yet called initSingleton()
-            sData.mInitState = CONSTRUCTED;
+            lk->mInstance = new DERIVED_TYPE(std::forward<Args>(args)...);
         }
         catch (const std::exception& err)
         {
@@ -343,62 +332,56 @@ private:
             // There isn't a separate EInitState value meaning "we attempted
             // to construct this LLSingleton subclass but could not," so use
             // DELETED. That seems slightly more appropriate than UNINITIALIZED.
-            sData.mInitState = DELETED;
+            lk->mInitState = DELETED;
             // propagate the exception
             throw;
         }
-    }
 
-    static void finishInitializing()
-    {
-        // getInstance() calls are from within initSingleton()
-        sData.mInitState = INITIALIZING;
+        // Any getInstance() calls after this point are from within initSingleton()
+        lk->mInitState = INITIALIZING;
         try
         {
             // initialize singleton after constructing it so that it can
             // reference other singletons which in turn depend on it, thus
             // breaking cyclic dependencies
-            sData.mInstance->initSingleton();
-            sData.mInitState = INITIALIZED;
+            lk->mInstance->initSingleton();
+            lk->mInitState = INITIALIZED;
 
             // pop this off stack of initializing singletons
-            pop_initializing();
+            pop_initializing(lk->mInstance);
         }
         catch (const std::exception& err)
         {
             // pop this off stack of initializing singletons here, too --
             // BEFORE logging, so log-machinery LLSingletons don't record a
             // dependency on DERIVED_TYPE!
-            pop_initializing();
+            pop_initializing(lk->mInstance);
             logwarns("Error in ", classname<DERIVED_TYPE>().c_str(),
                      "::initSingleton(): ", err.what());
-            // and get rid of the instance entirely
+            // Get rid of the instance entirely. This call depends on our
+            // recursive_mutex. We could have a deleteSingleton(LockStatic&)
+            // overload and pass lk, but we don't strictly need it.
             deleteSingleton();
             // propagate the exception
             throw;
         }
     }
 
-    static void pop_initializing()
+    static void pop_initializing(LLSingletonBase* sb)
     {
         // route through LLSingleton_manage_master so we Do The Right Thing
         // (namely, nothing) for MasterList
-        LLSingleton_manage_master<DERIVED_TYPE>().pop_initializing(sData.mInstance);
+        LLSingleton_manage_master<DERIVED_TYPE>().pop_initializing(sb);
     }
 
-    // Without this 'using' declaration, the static method we're declaring
-    // here would hide the base-class method we want it to call.
-    using LLSingletonBase::capture_dependency;
-    static void capture_dependency()
+    static void capture_dependency(LLSingletonBase* sb)
     {
         // By this point, if DERIVED_TYPE was pushed onto the initializing
         // stack, it has been popped off. So the top of that stack, if any, is
         // an LLSingleton that directly depends on DERIVED_TYPE. If
         // getInstance() was called by another LLSingleton, rather than from
         // vanilla application code, record the dependency.
-        sData.mInstance->capture_dependency(
-            LLSingleton_manage_master<DERIVED_TYPE>().get_initializing(),
-            sData.mInitState);
+        LLSingleton_manage_master<DERIVED_TYPE>().capture_dependency(sb);
     }
 
     // We know of no way to instruct the compiler that every subclass
@@ -411,20 +394,6 @@ private:
     // subclass body.
     virtual void you_must_use_LLSINGLETON_macro() = 0;
 
-    // The purpose of this struct is to engage the C++11 guarantee that static
-    // variables declared in function scope are initialized exactly once, even
-    // if multiple threads concurrently reach the same declaration.
-    // https://en.cppreference.com/w/cpp/language/storage_duration#Static_local_variables
-    // Since getInstance() declares a static instance of SingletonInitializer,
-    // only the first call to getInstance() calls constructSingleton().
-    struct SingletonInitializer
-    {
-        SingletonInitializer()
-        {
-            constructSingleton();
-        }
-    };
-
 protected:
     // Pass DERIVED_TYPE explicitly to LLSingletonBase's constructor because,
     // until our subclass constructor completes, *this isn't yet a
@@ -439,97 +408,176 @@ protected:
         LLSingleton_manage_master<DERIVED_TYPE>().add(this);
     }
 
-public:
+protected:
     virtual ~LLSingleton()
     {
-        // remove this instance from the master list
+        // This phase of cleanup is performed in the destructor rather than in
+        // deleteSingleton() to defend against manual deletion. When we moved
+        // cleanup to deleteSingleton(), we hit crashes due to dangling
+        // pointers in the MasterList.
+        LockStatic lk;
+        lk->mInstance  = nullptr;
+        lk->mInitState = DELETED;
+
+        // Remove this instance from the master list.
         LLSingleton_manage_master<DERIVED_TYPE>().remove(this);
-        sData.mInstance = NULL;
-        sData.mInitState = DELETED;
     }
 
+public:
     /**
-     * @brief Immediately delete the singleton.
+     * @brief Cleanup and destroy the singleton instance.
      *
-     * A subsequent call to LLProxy::getInstance() will construct a new
-     * instance of the class.
+     * deleteSingleton() calls this instance's cleanupSingleton() method and
+     * then destroys the instance.
      *
-     * Without an explicit call to LLSingletonBase::deleteAll(), LLSingletons
-     * are implicitly destroyed after main() has exited and the C++ runtime is
-     * cleaning up statically-constructed objects. Some classes derived from
-     * LLSingleton have objects that are part of a runtime system that is
-     * terminated before main() exits. Calling the destructor of those objects
-     * after the termination of their respective systems can cause crashes and
-     * other problems during termination of the project. Using this method to
-     * destroy the singleton early can prevent these crashes.
+     * A subsequent call to LLSingleton<T>::getInstance() will construct a new
+     * instance of the class.
      *
-     * An example where this is needed is for a LLSingleton that has an APR
-     * object as a member that makes APR calls on destruction. The APR system is
-     * shut down explicitly before main() exits. This causes a crash on exit.
-     * Using this method before the call to apr_terminate() and NOT calling
-     * getInstance() again will prevent the crash.
+     * Without an explicit call to LLSingletonBase::deleteAll(), or
+     * LLSingleton<T>::deleteSingleton(), LLSingleton instances are simply
+     * leaked. (Allowing implicit destruction at shutdown caused too many
+     * problems.)
      */
     static void deleteSingleton()
     {
-        delete sData.mInstance;
-        // SingletonData state handled by destructor, above
+        // Hold the lock while we call cleanupSingleton() and the destructor.
+        // Our destructor also instantiates LockStatic, requiring a recursive
+        // mutex.
+        LockStatic lk;
+        // of course, only cleanup and delete if there's something there
+        if (lk->mInstance)
+        {
+            lk->mInstance->cleanup_();
+            delete lk->mInstance;
+            // destructor clears mInstance (and mInitState)
+        }
     }
 
     static DERIVED_TYPE* getInstance()
     {
-        // call constructSingleton() only the first time we get here
-        static SingletonInitializer sInitializer;
-
-        switch (sData.mInitState)
-        {
-        case UNINITIALIZED:
-            // should never be uninitialized at this point
-            logerrs("Uninitialized singleton ",
-                    classname<DERIVED_TYPE>().c_str());
-            return NULL;
-
-        case CONSTRUCTING:
-            // here if DERIVED_TYPE's constructor (directly or indirectly)
-            // calls DERIVED_TYPE::getInstance()
-            logerrs("Tried to access singleton ",
-                    classname<DERIVED_TYPE>().c_str(),
-                    " from singleton constructor!");
-            return NULL;
-
-        case CONSTRUCTED:
-            // first time through: set to CONSTRUCTED by
-            // constructSingleton(), called by sInitializer's constructor;
-            // still have to call initSingleton()
-            finishInitializing();
-            break;
-
-        case INITIALIZING:
-            // here if DERIVED_TYPE::initSingleton() (directly or indirectly)
-            // calls DERIVED_TYPE::getInstance(): go ahead and allow it
-        case INITIALIZED:
-            // normal subsequent calls
-            break;
-
-        case DELETED:
-            // called after deleteSingleton()
-            logwarns("Trying to access deleted singleton ",
-                     classname<DERIVED_TYPE>().c_str(),
-                     " -- creating new instance");
-            // This recovery sequence is NOT thread-safe! We would need a
-            // recursive_mutex a la LLParamSingleton.
-            constructSingleton();
-            finishInitializing();
-            break;
-        }
-
-        // record the dependency, if any: check if we got here from another
-        // LLSingleton's constructor or initSingleton() method
-        capture_dependency();
-        return sData.mInstance;
+        // We know the viewer has LLSingleton dependency circularities. If you
+        // feel strongly motivated to eliminate them, cheers and good luck.
+        // (At that point we could consider a much simpler locking mechanism.)
+
+        // If A and B depend on each other, and thread T1 requests A at the
+        // same moment thread T2 requests B, you could get a sequence like this:
+        // - T1 locks A
+        // - T2 locks B
+        // - T1, having constructed A, calls A::initSingleton(), which calls
+        //   B::getInstance() and blocks on B's lock
+        // - T2, having constructed B, calls B::initSingleton(), which calls
+        //   A::getInstance() and blocks on A's lock
+        // In other words, classic deadlock.
+
+        // Avoid that by constructing and initializing every LLSingleton on
+        // the main thread. In that scenario:
+        // - T1 locks A
+        // - T2 locks B
+        // - T1 discovers A is UNINITIALIZED, so it queues a task for the main
+        //   thread, unlocks A and blocks on the std::future.
+        // - T2 discovers B is UNINITIALIZED, so it queues a task for the main
+        //   thread, unlocks B and blocks on the std::future.
+        // - The main thread executes T1's request for A. It locks A and
+        //   starts to construct it.
+        // - A::initSingleton() calls B::getInstance(). Fine: nobody's holding
+        //   B's lock.
+        // - The main thread locks B, constructs B, calls B::initSingleton(),
+        //   which calls A::getInstance(), which returns A.
+        // - B::getInstance() returns B to A::initSingleton(), unlocking B.
+        // - A::getInstance() returns A to the task wrapper, unlocking A.
+        // - The task wrapper passes A to T1 via the future. T1 resumes.
+        // - The main thread executes T2's request for B. Oh look, B already
+        //   exists. The task wrapper passes B to T2 via the future. T2
+        //   resumes.
+        // This still works even if one of T1 or T2 *is* the main thread.
+        // This still works even if thread T3 requests B at the same moment as
+        // T2. Finding B still UNINITIALIZED, T3 also queues a task for the
+        // main thread, unlocks B and blocks on a (distinct) std::future. By
+        // the time the main thread executes T3's request for B, B already
+        // exists, and is simply delivered via the future.
+
+        { // nested scope for 'lk'
+            // In case racing threads call getInstance() at the same moment,
+            // serialize the calls.
+            LockStatic lk;
+
+            switch (lk->mInitState)
+            {
+            case CONSTRUCTING:
+                // here if DERIVED_TYPE's constructor (directly or indirectly)
+                // calls DERIVED_TYPE::getInstance()
+                logerrs("Tried to access singleton ",
+                        classname<DERIVED_TYPE>().c_str(),
+                        " from singleton constructor!");
+                return nullptr;
+
+            case INITIALIZING:
+                // here if DERIVED_TYPE::initSingleton() (directly or indirectly)
+                // calls DERIVED_TYPE::getInstance(): go ahead and allow it
+            case INITIALIZED:
+                // normal subsequent calls
+                // record the dependency, if any: check if we got here from another
+                // LLSingleton's constructor or initSingleton() method
+                capture_dependency(lk->mInstance);
+                return lk->mInstance;
+
+            case DELETED:
+                // called after deleteSingleton()
+                logwarns("Trying to access deleted singleton ",
+                         classname<DERIVED_TYPE>().c_str(),
+                         " -- creating new instance");
+                // fall through
+            case UNINITIALIZED:
+            case QUEUED:
+                // QUEUED means some secondary thread has already requested an
+                // instance, but for present purposes that's semantically
+                // identical to UNINITIALIZED: either way, we must ourselves
+                // request an instance.
+                break;
+            }
+
+            // Here we need to construct a new instance.
+            if (on_main_thread())
+            {
+                // On the main thread, directly construct the instance while
+                // holding the lock.
+                constructSingleton(lk);
+                capture_dependency(lk->mInstance);
+                return lk->mInstance;
+            }
+
+            // Here we need to construct a new instance, but we're on a secondary
+            // thread.
+            lk->mInitState = QUEUED;
+        } // unlock 'lk'
+
+        // Per the comment block above, dispatch to the main thread.
+        loginfos(classname<DERIVED_TYPE>().c_str(),
+                 "::getInstance() dispatching to main thread");
+        auto instance = LLMainThreadTask::dispatch(
+            [](){
+                // VERY IMPORTANT to call getInstance() on the main thread,
+                // rather than going straight to constructSingleton()!
+                // During the time window before mInitState is INITIALIZED,
+                // multiple requests might be queued. It's essential that, as
+                // the main thread processes them, only the FIRST such request
+                // actually constructs the instance -- every subsequent one
+                // simply returns the existing instance.
+                loginfos(classname<DERIVED_TYPE>().c_str(),
+                         "::getInstance() on main thread");
+                return getInstance();
+            });
+        // record the dependency chain tracked on THIS thread, not the main
+        // thread (consider a getInstance() overload with a tag param that
+        // suppresses dep tracking when dispatched to the main thread)
+        capture_dependency(instance);
+        loginfos(classname<DERIVED_TYPE>().c_str(),
+                 "::getInstance() returning on requesting thread");
+        return instance;
     }
 
     // Reference version of getInstance()
-    // Preferred over getInstance() as it disallows checking for NULL
+    // Preferred over getInstance() as it disallows checking for nullptr
     static DERIVED_TYPE& instance()
     {
         return *getInstance();
@@ -539,7 +587,9 @@ public:
     // Use this to avoid accessing singletons before they can safely be constructed.
     static bool instanceExists()
     {
-        return sData.mInitState == INITIALIZED;
+        // defend any access to sData from racing threads
+        LockStatic lk;
+        return lk->mInitState == INITIALIZED;
     }
 
     // Has this singleton been deleted? This can be useful during shutdown
@@ -547,23 +597,12 @@ public:
     // cleaned up.
     static bool wasDeleted()
     {
-        return sData.mInitState == DELETED;
+        // defend any access to sData from racing threads
+        LockStatic lk;
+        return lk->mInitState == DELETED;
     }
-
-private:
-    struct SingletonData
-    {
-        // explicitly has a default constructor so that member variables are zero initialized in BSS
-        // and only changed by singleton logic, not constructor running during startup
-        EInitState      mInitState;
-        DERIVED_TYPE*   mInstance;
-    };
-    static SingletonData sData;
 };
 
-template<typename T>
-typename LLSingleton<T>::SingletonData LLSingleton<T>::sData;
-
 
 /**
  * LLParamSingleton<T> is like LLSingleton<T>, except in the following ways:
@@ -588,47 +627,86 @@ class LLParamSingleton : public LLSingleton<DERIVED_TYPE>
 {
 private:
     typedef LLSingleton<DERIVED_TYPE> super;
-    // Use a recursive_mutex in case of constructor circularity. With a
-    // non-recursive mutex, that would result in deadlock rather than the
-    // logerrs() call in getInstance().
-    typedef std::recursive_mutex mutex_t;
+    using typename super::LockStatic;
 
-public:
-    using super::deleteSingleton;
-    using super::instanceExists;
-    using super::wasDeleted;
-
-    // Passes arguments to DERIVED_TYPE's constructor and sets appropriate states
+    // Passes arguments to DERIVED_TYPE's constructor and sets appropriate
+    // states, returning a pointer to the new instance.
     template <typename... Args>
-    static void initParamSingleton(Args&&... args)
+    static DERIVED_TYPE* initParamSingleton_(Args&&... args)
     {
         // In case racing threads both call initParamSingleton() at the same
         // time, serialize them. One should initialize; the other should see
         // mInitState already set.
-        std::unique_lock<mutex_t> lk(getMutex());
+        LockStatic lk;
         // For organizational purposes this function shouldn't be called twice
-        if (super::sData.mInitState != super::UNINITIALIZED)
+        if (lk->mInitState != super::UNINITIALIZED)
         {
             super::logerrs("Tried to initialize singleton ",
                            super::template classname<DERIVED_TYPE>().c_str(),
                            " twice!");
+            return nullptr;
+        }
+        else if (on_main_thread())
+        {
+            // on the main thread, simply construct instance while holding lock
+            super::logdebugs(super::template classname<DERIVED_TYPE>().c_str(),
+                             "::initParamSingleton()");
+            super::constructSingleton(lk, std::forward<Args>(args)...);
+            return lk->mInstance;
         }
         else
         {
-            super::constructSingleton(std::forward<Args>(args)...);
-            super::finishInitializing();
+            // on secondary thread, dispatch to main thread --
+            // set state so we catch any other calls before the main thread
+            // picks up the task
+            lk->mInitState = super::QUEUED;
+            // very important to unlock here so main thread can actually process
+            lk.unlock();
+            super::loginfos(super::template classname<DERIVED_TYPE>().c_str(),
+                            "::initParamSingleton() dispatching to main thread");
+            // Normally it would be the height of folly to reference-bind
+            // 'args' into a lambda to be executed on some other thread! By
+            // the time that thread executed the lambda, the references would
+            // all be dangling, and Bad Things would result. But
+            // LLMainThreadTask::dispatch() promises to block until the passed
+            // task has completed. So in this case we know the references will
+            // remain valid until the lambda has run, so we dare to bind
+            // references.
+            auto instance = LLMainThreadTask::dispatch(
+                [&](){
+                    super::loginfos(super::template classname<DERIVED_TYPE>().c_str(),
+                                    "::initParamSingleton() on main thread");
+                    return initParamSingleton_(std::forward<Args>(args)...);
+                });
+            super::loginfos(super::template classname<DERIVED_TYPE>().c_str(),
+                            "::initParamSingleton() returning on requesting thread");
+            return instance;
         }
     }
 
+public:
+    using super::deleteSingleton;
+    using super::instanceExists;
+    using super::wasDeleted;
+
+    /// initParamSingleton() constructs the instance, returning a reference.
+    /// Pass whatever arguments are required to construct DERIVED_TYPE.
+    template <typename... Args>
+    static DERIVED_TYPE& initParamSingleton(Args&&... args)
+    {
+        return *initParamSingleton_(std::forward<Args>(args)...);
+    }
+
     static DERIVED_TYPE* getInstance()
     {
         // In case racing threads call getInstance() at the same moment as
         // initParamSingleton(), serialize the calls.
-        std::unique_lock<mutex_t> lk(getMutex());
+        LockStatic lk;
 
-        switch (super::sData.mInitState)
+        switch (lk->mInitState)
         {
         case super::UNINITIALIZED:
+        case super::QUEUED:
             super::logerrs("Uninitialized param singleton ",
                            super::template classname<DERIVED_TYPE>().c_str());
             break;
@@ -639,25 +717,13 @@ public:
                            " from singleton constructor!");
             break;
 
-        case super::CONSTRUCTED:
-            // Should never happen!? The CONSTRUCTED state is specifically to
-            // navigate through LLSingleton::SingletonInitializer getting
-            // constructed (once) before LLSingleton::getInstance()'s switch
-            // on mInitState. But our initParamSingleton() method calls
-            // constructSingleton() and then calls finishInitializing(), which
-            // immediately sets INITIALIZING. Why are we here?
-            super::logerrs("Param singleton ",
-                           super::template classname<DERIVED_TYPE>().c_str(),
-                           "::initSingleton() not yet called");
-            break;
-
         case super::INITIALIZING:
             // As with LLSingleton, explicitly permit circular calls from
             // within initSingleton()
         case super::INITIALIZED:
             // for any valid call, capture dependencies
-            super::capture_dependency();
-            return super::sData.mInstance;
+            super::capture_dependency(lk->mInstance);
+            return lk->mInstance;
 
         case super::DELETED:
             super::logerrs("Trying to access deleted param singleton ",
@@ -677,30 +743,6 @@ public:
     {
         return *getInstance();
     }
-
-private:
-    // sMutex must be a function-local static rather than a static member. One
-    // of the essential features of LLSingleton and friends is that they must
-    // support getInstance() even when the containing module's static
-    // variables have not yet been runtime-initialized. A mutex requires
-    // construction. A static class member might not yet have been
-    // constructed.
-    //
-    // We could store a dumb mutex_t*, notice when it's NULL and allocate a
-    // heap mutex -- but that's vulnerable to race conditions. And we can't
-    // defend the dumb pointer with another mutex.
-    //
-    // We could store a std::atomic<mutex_t*> -- but a default-constructed
-    // std::atomic<T> does not contain a valid T, even a default-constructed
-    // T! Which means std::atomic, too, requires runtime initialization.
-    //
-    // But a function-local static is guaranteed to be initialized exactly
-    // once, the first time control reaches that declaration.
-    static mutex_t& getMutex()
-    {
-        static mutex_t sMutex;
-        return sMutex;
-    }
 };
 
 /**
@@ -725,9 +767,9 @@ public:
     using super::instanceExists;
     using super::wasDeleted;
 
-    static void construct()
+    static DT* construct()
     {
-        super::initParamSingleton();
+        return super::initParamSingleton();
     }
 };
 
diff --git a/indra/llcommon/llstacktrace.cpp b/indra/llcommon/llstacktrace.cpp
index bbf0e1e141..80057bf0f2 100644
--- a/indra/llcommon/llstacktrace.cpp
+++ b/indra/llcommon/llstacktrace.cpp
@@ -33,7 +33,10 @@
 #include <sstream>
 
 #include "llwin32headerslean.h"
-#include "Dbghelp.h"
+#pragma warning (push)
+#pragma warning (disable:4091) // a microsoft header has warnings. Very nice.
+#include <dbghelp.h>
+#pragma warning (pop)
 
 typedef USHORT NTAPI RtlCaptureStackBackTrace_Function(
     IN ULONG frames_to_skip,
diff --git a/indra/llcommon/llstring.cpp b/indra/llcommon/llstring.cpp
index 0174c411b4..0290eea143 100644
--- a/indra/llcommon/llstring.cpp
+++ b/indra/llcommon/llstring.cpp
@@ -657,22 +657,6 @@ std::string utf8str_removeCRLF(const std::string& utf8str)
 }
 
 #if LL_WINDOWS
-// documentation moved to header. Phoenix 2007-11-27
-namespace snprintf_hack
-{
-	int snprintf(char *str, size_t size, const char *format, ...)
-	{
-		va_list args;
-		va_start(args, format);
-
-		int num_written = _vsnprintf(str, size, format, args); /* Flawfinder: ignore */
-		va_end(args);
-		
-		str[size-1] = '\0'; // always null terminate
-		return num_written;
-	}
-}
-
 std::string ll_convert_wide_to_string(const wchar_t* in)
 {
 	return ll_convert_wide_to_string(in, CP_UTF8);
diff --git a/indra/llcommon/llstring.h b/indra/llcommon/llstring.h
index b619a9e48c..6b1a1e0a03 100644
--- a/indra/llcommon/llstring.h
+++ b/indra/llcommon/llstring.h
@@ -707,32 +707,6 @@ LL_COMMON_API std::string utf8str_removeCRLF(const std::string& utf8str);
 //@{
 
 /**
- * @brief Implementation the expected snprintf interface.
- *
- * If the size of the passed in buffer is not large enough to hold the string,
- * two bad things happen:
- * 1. resulting formatted string is NOT null terminated
- * 2. Depending on the platform, the return value could be a) the required
- *    size of the buffer to copy the entire formatted string or b) -1.
- *    On Windows with VS.Net 2003, it returns -1 e.g. 
- *
- * safe_snprintf always adds a NULL terminator so that the caller does not
- * need to check for return value or need to add the NULL terminator.
- * It does not, however change the return value - to let the caller know
- * that the passed in buffer size was not large enough to hold the
- * formatted string.
- *
- */
-
-// Deal with the differeneces on Windows
-namespace snprintf_hack
-{
-	LL_COMMON_API int snprintf(char *str, size_t size, const char *format, ...);
-}
-
-using snprintf_hack::snprintf;
-
-/**
  * @brief Convert a wide string to std::string
  *
  * This replaces the unsafe W2A macro from ATL.
diff --git a/indra/llcommon/lltempredirect.cpp b/indra/llcommon/lltempredirect.cpp
new file mode 100644
index 0000000000..ec194c1d29
--- /dev/null
+++ b/indra/llcommon/lltempredirect.cpp
@@ -0,0 +1,138 @@
+/**
+ * @file   lltempredirect.cpp
+ * @author Nat Goodspeed
+ * @date   2019-10-31
+ * @brief  Implementation for lltempredirect.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+// Precompiled header
+#include "linden_common.h"
+// associated header
+#include "lltempredirect.h"
+// STL headers
+// std headers
+#if !LL_WINDOWS
+# include <unistd.h>
+#else
+# include <io.h>
+#endif // !LL_WINDOWS
+// external library headers
+// other Linden headers
+
+/*****************************************************************************
+*   llfd
+*****************************************************************************/
+// We could restate the implementation of each of llfd::close(), etc., but
+// this is way more succinct.
+#if LL_WINDOWS
+#define fhclose  _close
+#define fhdup    _dup
+#define fhdup2   _dup2
+#define fhfdopen _fdopen
+#define fhfileno _fileno
+#else
+#define fhclose  ::close
+#define fhdup    ::dup
+#define fhdup2   ::dup2
+#define fhfdopen ::fdopen
+#define fhfileno ::fileno
+#endif
+
+int llfd::close(int fd)
+{
+    return fhclose(fd);
+}
+
+int llfd::dup(int target)
+{
+    return fhdup(target);
+}
+
+int llfd::dup2(int target, int reference)
+{
+    return fhdup2(target, reference);
+}
+
+FILE* llfd::open(int fd, const char* mode)
+{
+    return fhfdopen(fd, mode);
+}
+
+int llfd::fileno(FILE* stream)
+{
+    return fhfileno(stream);
+}
+
+/*****************************************************************************
+*   LLTempRedirect
+*****************************************************************************/
+LLTempRedirect::LLTempRedirect():
+    mOrigTarget(-1),                // -1 is an invalid file descriptor
+    mReference(-1)
+{}
+
+LLTempRedirect::LLTempRedirect(FILE* target, FILE* reference):
+    LLTempRedirect((target?    fhfileno(target)    : -1),
+                   (reference? fhfileno(reference) : -1))
+{}
+
+LLTempRedirect::LLTempRedirect(int target, int reference):
+    // capture a duplicate file descriptor for the file originally targeted by
+    // 'reference'
+    mOrigTarget((reference >= 0)? fhdup(reference) : -1),
+    mReference(reference)
+{
+    if (target >= 0 && reference >= 0)
+    {
+        // As promised, force 'reference' to refer to 'target'. This first
+        // implicitly closes 'reference', which is why we first capture a
+        // duplicate so the original target file stays open.
+        fhdup2(target, reference);
+    }
+}
+
+LLTempRedirect::LLTempRedirect(LLTempRedirect&& other)
+{
+    mOrigTarget = other.mOrigTarget;
+    mReference  = other.mReference;
+    // other LLTempRedirect must be in moved-from state so its destructor
+    // won't repeat the same operations as ours!
+    other.mOrigTarget = -1;
+    other.mReference  = -1;
+}
+
+LLTempRedirect::~LLTempRedirect()
+{
+    reset();
+}
+
+void LLTempRedirect::reset()
+{
+    // If this instance was default-constructed (or constructed with an
+    // invalid file descriptor), skip the following.
+    if (mOrigTarget >= 0)
+    {
+        // Restore mReference to point to mOrigTarget. This implicitly closes
+        // the duplicate created by our constructor of its 'target' file
+        // descriptor.
+        fhdup2(mOrigTarget, mReference);
+        // mOrigTarget has served its purpose
+        fhclose(mOrigTarget);
+    }
+    // assign these because reset() is also responsible for a "moved from"
+    // instance
+    mOrigTarget = -1;
+    mReference  = -1;
+}
+
+LLTempRedirect& LLTempRedirect::operator=(LLTempRedirect&& other)
+{
+    reset();
+    std::swap(mOrigTarget, other.mOrigTarget);
+    std::swap(mReference,  other.mReference);
+    return *this;
+}
diff --git a/indra/llcommon/lltempredirect.h b/indra/llcommon/lltempredirect.h
new file mode 100644
index 0000000000..33e05dc06b
--- /dev/null
+++ b/indra/llcommon/lltempredirect.h
@@ -0,0 +1,91 @@
+/**
+ * @file   lltempredirect.h
+ * @author Nat Goodspeed
+ * @date   2019-10-31
+ * @brief  RAII low-level file-descriptor redirection
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if ! defined(LL_LLTEMPREDIRECT_H)
+#define LL_LLTEMPREDIRECT_H
+
+// Functions in this namespace are intended to insulate the caller from the
+// aggravating distinction between ::close() and Microsoft _close().
+namespace llfd
+{
+
+int close(int fd);
+int dup(int target);
+int dup2(int target, int reference);
+FILE* open(int fd, const char* mode);
+int fileno(FILE* stream);
+
+} // namespace llfd
+
+/**
+ * LLTempRedirect is an RAII class that performs file redirection on low-level
+ * file descriptors, expressed as ints. (Use llfd::fileno() to obtain the file
+ * descriptor from a classic-C FILE*. There is no portable way to obtain the
+ * file descriptor from a std::fstream.)
+ *
+ * Instantiate LLTempRedirect with a target file descriptor (e.g. for some
+ * open file) and a reference file descriptor (e.g. for stderr). From that
+ * point until the LLTempRedirect instance is destroyed, all OS-level writes
+ * to the reference file descriptor will be redirected to the target file.
+ *
+ * Because dup2() is used for redirection, the original passed target file
+ * descriptor remains open. If you want LLTempRedirect's destructor to close
+ * the target file, close() the target file descriptor after passing it to
+ * LLTempRedirect's constructor.
+ *
+ * LLTempRedirect's constructor saves the original target of the reference
+ * file descriptor. Its destructor restores the reference file descriptor to
+ * point once again to its original target.
+ */
+class LLTempRedirect
+{
+public:
+    LLTempRedirect();
+    /**
+     * For the lifespan of this LLTempRedirect instance, all writes to
+     * 'reference' will be redirected to 'target'. When this LLTempRedirect is
+     * destroyed, the original target for 'reference' will be restored.
+     *
+     * Pass 'target' as NULL if you simply want to save and restore
+     * 'reference' against possible redirection in the meantime.
+     */
+    LLTempRedirect(FILE* target, FILE* reference);
+    /**
+     * For the lifespan of this LLTempRedirect instance, all writes to
+     * 'reference' will be redirected to 'target'. When this LLTempRedirect is
+     * destroyed, the original target for 'reference' will be restored.
+     *
+     * Pass 'target' as -1 if you simply want to save and restore
+     * 'reference' against possible redirection in the meantime.
+     */
+    LLTempRedirect(int target,   int reference);
+    LLTempRedirect(const LLTempRedirect&) = delete;
+    LLTempRedirect(LLTempRedirect&& other);
+
+    ~LLTempRedirect();
+
+    LLTempRedirect& operator=(const LLTempRedirect&) = delete;
+    LLTempRedirect& operator=(LLTempRedirect&& other);
+
+    /// returns (duplicate file descriptor for) the original target of the
+    /// 'reference' file descriptor passed to our constructor
+    int getOriginalTarget() const { return mOrigTarget; }
+    /// returns the original 'reference' file descriptor passed to our
+    /// constructor
+    int getReference()      const { return mReference; }
+
+private:
+    void reset();
+
+    int mOrigTarget, mReference;
+};
+
+#endif /* ! defined(LL_LLTEMPREDIRECT_H) */
diff --git a/indra/llcommon/llthread.cpp b/indra/llcommon/llthread.cpp
index a4171729db..0b9dec969c 100644
--- a/indra/llcommon/llthread.cpp
+++ b/indra/llcommon/llthread.cpp
@@ -92,26 +92,39 @@ void set_thread_name( DWORD dwThreadID, const char* threadName)
 // }
 // 
 //----------------------------------------------------------------------------
+namespace
+{
 
-U32 LL_THREAD_LOCAL sThreadID = 0;
+    LLThread::id_t main_thread()
+    {
+        // Using a function-static variable to identify the main thread
+        // requires that control reach here from the main thread before it
+        // reaches here from any other thread. We simply trust that whichever
+        // thread gets here first is the main thread.
+        static LLThread::id_t s_thread_id = LLThread::currentID();
+        return s_thread_id;
+    }
 
-U32 LLThread::sIDIter = 0;
+} // anonymous namespace
 
+LL_COMMON_API bool on_main_thread()
+{
+    return (LLThread::currentID() == main_thread());
+}
 
 LL_COMMON_API void assert_main_thread()
 {
-    static U32 s_thread_id = LLThread::currentID();
-    if (LLThread::currentID() != s_thread_id)
+    auto curr = LLThread::currentID();
+    auto main = main_thread();
+    if (curr != main)
     {
-        LL_WARNS() << "Illegal execution from thread id " << (S32) LLThread::currentID()
-            << " outside main thread " << (S32) s_thread_id << LL_ENDL;
+        LL_WARNS() << "Illegal execution from thread id " << curr
+            << " outside main thread " << main << LL_ENDL;
     }
 }
 
-void LLThread::registerThreadID()
-{
-    sThreadID = ++sIDIter;
-}
+// this function has become moot
+void LLThread::registerThreadID() {}
 
 //
 // Handed to the APR thread creation function
@@ -122,11 +135,12 @@ void LLThread::threadRun()
     set_thread_name(-1, mName.c_str());
 #endif
 
+    // this is the first point at which we're actually running in the new thread
+    mID = currentID();
+
     // for now, hard code all LLThreads to report to single master thread recorder, which is known to be running on main thread
     mRecorder = new LLTrace::ThreadRecorder(*LLTrace::get_master_thread_recorder());
 
-    sThreadID = mID;
-
     // Run the user supplied function
     do 
     {
@@ -168,8 +182,6 @@ LLThread::LLThread(const std::string& name, apr_pool_t *poolp) :
     mStatus(STOPPED),
     mRecorder(NULL)
 {
-
-    mID = ++sIDIter;
     mRunCondition = new LLCondition();
     mDataLock = new LLMutex();
     mLocalAPRFilePoolp = NULL ;
@@ -347,9 +359,9 @@ void LLThread::setQuitting()
 }
 
 // static
-U32 LLThread::currentID()
+LLThread::id_t LLThread::currentID()
 {
-    return sThreadID;
+    return std::this_thread::get_id();
 }
 
 // static
@@ -376,6 +388,16 @@ void LLThread::wakeLocked()
     }
 }
 
+void LLThread::lockData()
+{
+    mDataLock->lock();
+}
+
+void LLThread::unlockData()
+{
+    mDataLock->unlock();
+}
+
 //============================================================================
 
 //----------------------------------------------------------------------------
diff --git a/indra/llcommon/llthread.h b/indra/llcommon/llthread.h
index 863c9051f3..5cd0731f6c 100644
--- a/indra/llcommon/llthread.h
+++ b/indra/llcommon/llthread.h
@@ -30,12 +30,9 @@
 #include "llapp.h"
 #include "llapr.h"
 #include "boost/intrusive_ptr.hpp"
-#include "llmutex.h"
 #include "llrefcount.h"
 #include <thread>
 
-LL_COMMON_API void assert_main_thread();
-
 namespace LLTrace
 {
     class ThreadRecorder;
@@ -45,7 +42,6 @@ class LL_COMMON_API LLThread
 {
 private:
     friend class LLMutex;
-    static U32 sIDIter;
 
 public:
     typedef enum e_thread_status
@@ -55,6 +51,7 @@ public:
         QUITTING= 2,    // Someone wants this thread to quit
         CRASHED = -1    // An uncaught exception was thrown by the thread
     } EThreadStatus;
+    typedef std::thread::id id_t;
 
     LLThread(const std::string& name, apr_pool_t *poolp = NULL);
     virtual ~LLThread(); // Warning!  You almost NEVER want to destroy a thread unless it's in the STOPPED state.
@@ -64,7 +61,7 @@ public:
     bool isStopped() const { return (STOPPED == mStatus) || (CRASHED == mStatus); }
     bool isCrashed() const { return (CRASHED == mStatus); } 
     
-    static U32 currentID(); // Return ID of current thread
+    static id_t currentID(); // Return ID of current thread
     static void yield(); // Static because it can be called by the main thread, which doesn't have an LLThread data structure.
     
 public:
@@ -88,7 +85,7 @@ public:
 
     LLVolatileAPRPool* getLocalAPRFilePool() { return mLocalAPRFilePoolp ; }
 
-    U32 getID() const { return mID; }
+    id_t getID() const { return mID; }
 
     // Called by threads *not* created via LLThread to register some
     // internal state used by LLMutex.  You must call this once early
@@ -109,7 +106,7 @@ protected:
 
     std::thread        *mThreadp;
     EThreadStatus       mStatus;
-    U32                 mID;
+    id_t                mID;
     LLTrace::ThreadRecorder* mRecorder;
 
     //a local apr_pool for APRFile operations in this thread. If it exists, LLAPRFile::sAPRFilePoolp should not be used.
@@ -126,8 +123,8 @@ protected:
     virtual bool runCondition(void);
 
     // Lock/Unlock Run Condition -- use around modification of any variable used in runCondition()
-    inline void lockData();
-    inline void unlockData();
+    void lockData();
+    void unlockData();
     
     // This is the predicate that decides whether the thread should sleep.  
     // It should only be called with mDataLock locked, since the virtual runCondition() function may need to access
@@ -142,17 +139,6 @@ protected:
 };
 
 
-void LLThread::lockData()
-{
-    mDataLock->lock();
-}
-
-void LLThread::unlockData()
-{
-    mDataLock->unlock();
-}
-
-
 //============================================================================
 
 // Simple responder for self destructing callbacks
@@ -168,5 +154,6 @@ public:
 //============================================================================
 
 extern LL_COMMON_API void assert_main_thread();
+extern LL_COMMON_API bool on_main_thread();
 
 #endif // LL_LLTHREAD_H
diff --git a/indra/llcommon/llthreadlocalstorage.cpp b/indra/llcommon/llthreadlocalstorage.cpp
index 8cef05caac..d8a063e8d5 100644
--- a/indra/llcommon/llthreadlocalstorage.cpp
+++ b/indra/llcommon/llthreadlocalstorage.cpp
@@ -93,11 +93,9 @@ void LLThreadLocalPointerBase::initAllThreadLocalStorage()
 {
 	if (!sInitialized)
 	{
-		for (LLInstanceTracker<LLThreadLocalPointerBase>::instance_iter it = beginInstances(), end_it = endInstances();
-			it != end_it;
-			++it)
+		for (auto& base : instance_snapshot())
 		{
-			(*it).initStorage();
+			base.initStorage();
 		}
 		sInitialized = true;
 	}
@@ -108,11 +106,9 @@ void LLThreadLocalPointerBase::destroyAllThreadLocalStorage()
 {
 	if (sInitialized)
 	{
-		//for (LLInstanceTracker<LLThreadLocalPointerBase>::instance_iter it = beginInstances(), end_it = endInstances();
-		//	it != end_it;
-		//	++it)
+		//for (auto& base : instance_snapshot())
 		//{
-		//	(*it).destroyStorage();
+		//	base.destroyStorage();
 		//}
 		sInitialized = false;
 	}
diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h
index b0bddac8e5..30dd507f73 100644
--- a/indra/llcommon/llthreadsafequeue.h
+++ b/indra/llcommon/llthreadsafequeue.h
@@ -30,18 +30,12 @@
 #include "llexception.h"
 #include <deque>
 #include <string>
-
-#if LL_WINDOWS
-#pragma warning (push)
-#pragma warning (disable:4265)
-#endif
-// 'std::_Pad' : class has virtual functions, but destructor is not virtual
-#include <mutex>
-#include <condition_variable>
-
-#if LL_WINDOWS
-#pragma warning (pop)
-#endif
+#include <chrono>
+#include "mutex.h"
+#include "llcoros.h"
+#include LLCOROS_MUTEX_HEADER
+#include <boost/fiber/timed_mutex.hpp>
+#include LLCOROS_CONDVAR_HEADER
 
 //
 // A general queue exception.
@@ -88,18 +82,28 @@ public:
 	// Add an element to the front of queue (will block if the queue has
 	// reached capacity).
 	//
-	// This call will raise an interrupt error if the queue is deleted while
+	// This call will raise an interrupt error if the queue is closed while
 	// the caller is blocked.
 	void pushFront(ElementT const & element);
 	
-	// Try to add an element to the front ofqueue without blocking. Returns
+	// Try to add an element to the front of queue without blocking. Returns
 	// true only if the element was actually added.
 	bool tryPushFront(ElementT const & element);
-	
+
+	// Try to add an element to the front of queue, blocking if full but with
+	// timeout. Returns true if the element was added.
+	// There are potentially two different timeouts involved: how long to try
+	// to lock the mutex, versus how long to wait for the queue to stop being
+	// full. Careful settings for each timeout might be orders of magnitude
+	// apart. However, this method conflates them.
+	template <typename Rep, typename Period>
+	bool tryPushFrontFor(const std::chrono::duration<Rep, Period>& timeout,
+						 ElementT const & element);
+
 	// Pop the element at the end of the queue (will block if the queue is
 	// empty).
 	//
-	// This call will raise an interrupt error if the queue is deleted while
+	// This call will raise an interrupt error if the queue is closed while
 	// the caller is blocked.
 	ElementT popBack(void);
 	
@@ -110,13 +114,29 @@ public:
 	// Returns the size of the queue.
 	size_t size();
 
+	// closes the queue:
+	// - every subsequent pushFront() call will throw LLThreadSafeQueueInterrupt
+	// - every subsequent tryPushFront() call will return false
+	// - popBack() calls will return normally until the queue is drained, then
+	//   every subsequent popBack() will throw LLThreadSafeQueueInterrupt
+	// - tryPopBack() calls will return normally until the queue is drained,
+	//   then every subsequent tryPopBack() call will return false
+	void close();
+
+	// detect closed state
+	bool isClosed();
+	// inverse of isClosed()
+	explicit operator bool();
+
 private:
 	std::deque< ElementT > mStorage;
 	U32 mCapacity;
+	bool mClosed;
 
-	std::mutex mLock;
-	std::condition_variable mCapacityCond;
-	std::condition_variable mEmptyCond;
+	boost::fibers::timed_mutex mLock;
+	typedef std::unique_lock<decltype(mLock)> lock_t;
+	boost::fibers::condition_variable_any mCapacityCond;
+	boost::fibers::condition_variable_any mEmptyCond;
 };
 
 // LLThreadSafeQueue
@@ -124,7 +144,8 @@ private:
 
 template<typename ElementT>
 LLThreadSafeQueue<ElementT>::LLThreadSafeQueue(U32 capacity) :
-mCapacity(capacity)
+    mCapacity(capacity),
+    mClosed(false)
 {
 }
 
@@ -132,13 +153,18 @@ mCapacity(capacity)
 template<typename ElementT>
 void LLThreadSafeQueue<ElementT>::pushFront(ElementT const & element)
 {
+    lock_t lock1(mLock);
     while (true)
     {
-        std::unique_lock<std::mutex> lock1(mLock);
+        if (mClosed)
+        {
+            LLTHROW(LLThreadSafeQueueInterrupt());
+        }
 
         if (mStorage.size() < mCapacity)
         {
             mStorage.push_front(element);
+            lock1.unlock();
             mEmptyCond.notify_one();
             return;
         }
@@ -149,17 +175,61 @@ void LLThreadSafeQueue<ElementT>::pushFront(ElementT const & element)
 }
 
 
+template <typename ElementT>
+template <typename Rep, typename Period>
+bool LLThreadSafeQueue<ElementT>::tryPushFrontFor(const std::chrono::duration<Rep, Period>& timeout,
+                                                  ElementT const & element)
+{
+    // Convert duration to time_point: passing the same timeout duration to
+    // each of multiple calls is wrong.
+    auto endpoint = std::chrono::steady_clock::now() + timeout;
+
+    lock_t lock1(mLock, std::defer_lock);
+    if (!lock1.try_lock_until(endpoint))
+        return false;
+
+    while (true)
+    {
+        if (mClosed)
+        {
+            return false;
+        }
+
+        if (mStorage.size() < mCapacity)
+        {
+            mStorage.push_front(element);
+            lock1.unlock();
+            mEmptyCond.notify_one();
+            return true;
+        }
+
+        // Storage Full. Wait for signal.
+        if (LLCoros::cv_status::timeout == mCapacityCond.wait_until(lock1, endpoint))
+        {
+            // timed out -- formally we might recheck both conditions above
+            return false;
+        }
+        // If we didn't time out, we were notified for some reason. Loop back
+        // to check.
+    }
+}
+
+
 template<typename ElementT>
 bool LLThreadSafeQueue<ElementT>::tryPushFront(ElementT const & element)
 {
-    std::unique_lock<std::mutex> lock1(mLock, std::defer_lock);
+    lock_t lock1(mLock, std::defer_lock);
     if (!lock1.try_lock())
         return false;
 
+    if (mClosed)
+        return false;
+
     if (mStorage.size() >= mCapacity)
         return false;
 
     mStorage.push_front(element);
+    lock1.unlock();
     mEmptyCond.notify_one();
     return true;
 }
@@ -168,18 +238,23 @@ bool LLThreadSafeQueue<ElementT>::tryPushFront(ElementT const & element)
 template<typename ElementT>
 ElementT LLThreadSafeQueue<ElementT>::popBack(void)
 {
+    lock_t lock1(mLock);
     while (true)
     {
-        std::unique_lock<std::mutex> lock1(mLock);
-
         if (!mStorage.empty())
         {
             ElementT value = mStorage.back();
             mStorage.pop_back();
+            lock1.unlock();
             mCapacityCond.notify_one();
             return value;
         }
 
+        if (mClosed)
+        {
+            LLTHROW(LLThreadSafeQueueInterrupt());
+        }
+
         // Storage empty. Wait for signal.
         mEmptyCond.wait(lock1);
     }
@@ -189,15 +264,18 @@ ElementT LLThreadSafeQueue<ElementT>::popBack(void)
 template<typename ElementT>
 bool LLThreadSafeQueue<ElementT>::tryPopBack(ElementT & element)
 {
-    std::unique_lock<std::mutex> lock1(mLock, std::defer_lock);
+    lock_t lock1(mLock, std::defer_lock);
     if (!lock1.try_lock())
         return false;
 
+    // no need to check mClosed: tryPopBack() behavior when the queue is
+    // closed is implemented by simple inability to push any new elements
     if (mStorage.empty())
         return false;
 
     element = mStorage.back();
     mStorage.pop_back();
+    lock1.unlock();
     mCapacityCond.notify_one();
     return true;
 }
@@ -206,8 +284,34 @@ bool LLThreadSafeQueue<ElementT>::tryPopBack(ElementT & element)
 template<typename ElementT>
 size_t LLThreadSafeQueue<ElementT>::size(void)
 {
-    std::lock_guard<std::mutex> lock(mLock);
+    lock_t lock(mLock);
     return mStorage.size();
 }
 
+template<typename ElementT>
+void LLThreadSafeQueue<ElementT>::close()
+{
+    lock_t lock(mLock);
+    mClosed = true;
+    lock.unlock();
+    // wake up any blocked popBack() calls
+    mEmptyCond.notify_all();
+    // wake up any blocked pushFront() calls
+    mCapacityCond.notify_all();
+}
+
+template<typename ElementT>
+bool LLThreadSafeQueue<ElementT>::isClosed()
+{
+    lock_t lock(mLock);
+    return mClosed;
+}
+
+template<typename ElementT>
+LLThreadSafeQueue<ElementT>::operator bool()
+{
+    lock_t lock(mLock);
+    return ! mClosed;
+}
+
 #endif
diff --git a/indra/llcommon/lltrace.h b/indra/llcommon/lltrace.h
index 79ff55b739..0d0cd6f581 100644
--- a/indra/llcommon/lltrace.h
+++ b/indra/llcommon/lltrace.h
@@ -57,7 +57,7 @@ class StatBase
 {
 public:
 	StatBase(const char* name, const char* description);
-	virtual ~StatBase() LLINSTANCETRACKER_DTOR_NOEXCEPT	{}
+	virtual ~StatBase()	{}
 	virtual const char* getUnitLabel() const;
 
 	const std::string& getName() const { return mName; }
diff --git a/indra/llcommon/lltraceaccumulators.cpp b/indra/llcommon/lltraceaccumulators.cpp
index 385d31edd7..b1c23c6fb7 100644
--- a/indra/llcommon/lltraceaccumulators.cpp
+++ b/indra/llcommon/lltraceaccumulators.cpp
@@ -291,8 +291,8 @@ void EventAccumulator::reset( const EventAccumulator* other )
 {
 	mNumSamples = 0;
 	mSum = 0;
-	mMin = NaN;
-	mMax = NaN;
+	mMin = F32(NaN);
+	mMax = F32(NaN);
 	mMean = NaN;
 	mSumOfSquares = 0;
 	mLastValue = other ? other->mLastValue : NaN;
diff --git a/indra/llcommon/lltraceaccumulators.h b/indra/llcommon/lltraceaccumulators.h
index 6f27b97dff..8eb5338a2a 100644
--- a/indra/llcommon/lltraceaccumulators.h
+++ b/indra/llcommon/lltraceaccumulators.h
@@ -242,8 +242,8 @@ namespace LLTrace
 
 		EventAccumulator()
 		:	mSum(0),
-			mMin(NaN),
-			mMax(NaN),
+			mMin(F32(NaN)),
+			mMax(F32(NaN)),
 			mMean(NaN),
 			mSumOfSquares(0),
 			mNumSamples(0),
@@ -313,8 +313,8 @@ namespace LLTrace
 
 		SampleAccumulator()
 		:	mSum(0),
-			mMin(NaN),
-			mMax(NaN),
+			mMin(F32(NaN)),
+			mMax(F32(NaN)),
 			mMean(NaN),
 			mSumOfSquares(0),
 			mLastSampleTimeStamp(0),
diff --git a/indra/llcommon/lltracethreadrecorder.cpp b/indra/llcommon/lltracethreadrecorder.cpp
index 181fc2f058..025dc57044 100644
--- a/indra/llcommon/lltracethreadrecorder.cpp
+++ b/indra/llcommon/lltracethreadrecorder.cpp
@@ -28,6 +28,7 @@
 #include "lltracethreadrecorder.h"
 #include "llfasttimer.h"
 #include "lltrace.h"
+#include "llstl.h"
 
 namespace LLTrace
 {
@@ -64,16 +65,15 @@ void ThreadRecorder::init()
 	activate(&mThreadRecordingBuffers);
 
 	// initialize time block parent pointers
-	for (BlockTimerStatHandle::instance_tracker_t::instance_iter it = BlockTimerStatHandle::instance_tracker_t::beginInstances(), end_it = BlockTimerStatHandle::instance_tracker_t::endInstances(); 
-		it != end_it; 
-		++it)
+	for (auto& base : BlockTimerStatHandle::instance_snapshot())
 	{
-		BlockTimerStatHandle& time_block = static_cast<BlockTimerStatHandle&>(*it);
-		TimeBlockTreeNode& tree_node = mTimeBlockTreeNodes[it->getIndex()];
+		// because of indirect derivation from LLInstanceTracker, have to downcast
+		BlockTimerStatHandle& time_block = static_cast<BlockTimerStatHandle&>(base);
+		TimeBlockTreeNode& tree_node = mTimeBlockTreeNodes[time_block.getIndex()];
 		tree_node.mBlock = &time_block;
 		tree_node.mParent = &root_time_block;
 
-		it->getCurrentAccumulator().mParent = &root_time_block;
+		time_block.getCurrentAccumulator().mParent = &root_time_block;
 	}
 
 	mRootTimer = new BlockTimer(root_time_block);
diff --git a/indra/llcommon/lluuid.cpp b/indra/llcommon/lluuid.cpp
index 8f33d789eb..b05630c6b5 100644
--- a/indra/llcommon/lluuid.cpp
+++ b/indra/llcommon/lluuid.cpp
@@ -43,6 +43,7 @@
 #include "llstring.h"
 #include "lltimer.h"
 #include "llthread.h"
+#include "llmutex.h"
 
 const LLUUID LLUUID::null;
 const LLTransactionID LLTransactionID::tnull;
@@ -738,7 +739,7 @@ void LLUUID::getCurrentTime(uuid_time_t *timestamp)
       getSystemTime(&time_last);
       uuids_this_tick = uuids_per_tick;
       init = TRUE;
-	  mMutex = new LLMutex();
+      mMutex = new LLMutex();
    }
 
    uuid_time_t time_now = {0,0};
diff --git a/indra/llcommon/llworkerthread.h b/indra/llcommon/llworkerthread.h
index b1a6f61360..0387e75c65 100644
--- a/indra/llcommon/llworkerthread.h
+++ b/indra/llcommon/llworkerthread.h
@@ -34,6 +34,7 @@
 
 #include "llqueuedthread.h"
 #include "llatomic.h"
+#include "llmutex.h"
 
 #define USE_FRAME_CALLBACK_MANAGER 0
 
diff --git a/indra/llcommon/lockstatic.h b/indra/llcommon/lockstatic.h
new file mode 100644
index 0000000000..96c53c6473
--- /dev/null
+++ b/indra/llcommon/lockstatic.h
@@ -0,0 +1,73 @@
+/**
+ * @file   lockstatic.h
+ * @author Nat Goodspeed
+ * @date   2019-12-03
+ * @brief  LockStatic class provides mutex-guarded access to the specified
+ *         static data.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if ! defined(LL_LOCKSTATIC_H)
+#define LL_LOCKSTATIC_H
+
+#include "mutex.h"                  // std::unique_lock
+
+namespace llthread
+{
+
+// Instantiate this template to obtain a pointer to the canonical static
+// instance of Static while holding a lock on that instance. Use of
+// Static::mMutex presumes that Static declares some suitable mMutex.
+template <typename Static>
+class LockStatic
+{
+    typedef std::unique_lock<decltype(Static::mMutex)> lock_t;
+public:
+    LockStatic():
+        mData(getStatic()),
+        mLock(mData->mMutex)
+    {}
+    Static* get() const { return mData; }
+    operator Static*() const { return get(); }
+    Static* operator->() const { return get(); }
+    // sometimes we must explicitly unlock...
+    void unlock()
+    {
+        // but once we do, access is no longer permitted
+        mData = nullptr;
+        mLock.unlock();
+    }
+protected:
+    Static* mData;
+    lock_t mLock;
+private:
+    Static* getStatic()
+    {
+        // Static::mMutex must be function-local static rather than class-
+        // static. Some of our consumers must function properly (therefore
+        // lock properly) even when the containing module's static variables
+        // have not yet been runtime-initialized. A mutex requires
+        // construction. A static class member might not yet have been
+        // constructed.
+        //
+        // We could store a dumb mutex_t*, notice when it's NULL and allocate a
+        // heap mutex -- but that's vulnerable to race conditions. And we can't
+        // defend the dumb pointer with another mutex.
+        //
+        // We could store a std::atomic<mutex_t*> -- but a default-constructed
+        // std::atomic<T> does not contain a valid T, even a default-constructed
+        // T! Which means std::atomic, too, requires runtime initialization.
+        //
+        // But a function-local static is guaranteed to be initialized exactly
+        // once: the first time control reaches that declaration.
+        static Static sData;
+        return &sData;
+    }
+};
+
+} // llthread namespace
+
+#endif /* ! defined(LL_LOCKSTATIC_H) */
diff --git a/indra/llcommon/mutex.h b/indra/llcommon/mutex.h
new file mode 100644
index 0000000000..90d0942270
--- /dev/null
+++ b/indra/llcommon/mutex.h
@@ -0,0 +1,22 @@
+/**
+ * @file   mutex.h
+ * @author Nat Goodspeed
+ * @date   2019-12-03
+ * @brief  Wrap <mutex> in odious boilerplate
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if LL_WINDOWS
+#pragma warning (push)
+#pragma warning (disable:4265)
+#endif
+// warning C4265: 'std::_Pad' : class has virtual functions, but destructor is not virtual
+
+#include <mutex>
+
+#if LL_WINDOWS
+#pragma warning (pop)
+#endif
diff --git a/indra/llcommon/tests/llcond_test.cpp b/indra/llcommon/tests/llcond_test.cpp
new file mode 100644
index 0000000000..478149eacf
--- /dev/null
+++ b/indra/llcommon/tests/llcond_test.cpp
@@ -0,0 +1,67 @@
+/**
+ * @file   llcond_test.cpp
+ * @author Nat Goodspeed
+ * @date   2019-07-18
+ * @brief  Test for llcond.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+// Precompiled header
+#include "linden_common.h"
+// associated header
+#include "llcond.h"
+// STL headers
+// std headers
+// external library headers
+// other Linden headers
+#include "../test/lltut.h"
+#include "llcoros.h"
+
+/*****************************************************************************
+*   TUT
+*****************************************************************************/
+namespace tut
+{
+    struct llcond_data
+    {
+        LLScalarCond<int> cond{0};
+    };
+    typedef test_group<llcond_data> llcond_group;
+    typedef llcond_group::object object;
+    llcond_group llcondgrp("llcond");
+
+    template<> template<>
+    void object::test<1>()
+    {
+        set_test_name("Immediate gratification");
+        cond.set_one(1);
+        ensure("wait_for_equal() failed", 
+               cond.wait_for_equal(F32Milliseconds(1), 1));
+        ensure("wait_for_unequal() should have failed",
+               ! cond.wait_for_unequal(F32Milliseconds(1), 1));
+    }
+
+    template<> template<>
+    void object::test<2>()
+    {
+        set_test_name("Simple two-coroutine test");
+        LLCoros::instance().launch(
+            "test<2>",
+            [this]()
+            {
+                // Lambda immediately entered -- control comes here first.
+                ensure_equals(cond.get(), 0);
+                cond.set_all(1);
+                cond.wait_equal(2);
+                ensure_equals(cond.get(), 2);
+                cond.set_all(3);
+            });
+        // Main coroutine is resumed only when the lambda waits.
+        ensure_equals(cond.get(), 1);
+        cond.set_all(2);
+        cond.wait_equal(3);
+    }
+} // namespace tut
diff --git a/indra/llcommon/tests/lleventcoro_test.cpp b/indra/llcommon/tests/lleventcoro_test.cpp
index fa02d2bb1a..032923a108 100644
--- a/indra/llcommon/tests/lleventcoro_test.cpp
+++ b/indra/llcommon/tests/lleventcoro_test.cpp
@@ -26,102 +26,33 @@
  * $/LicenseInfo$
  */
 
-/*****************************************************************************/
-//  test<1>() is cloned from a Boost.Coroutine example program whose copyright
-//  info is reproduced here:
-/*---------------------------------------------------------------------------*/
-//  Copyright (c) 2006, Giovanni P. Deretta
-//
-//  This code may be used under either of the following two licences:
-//
-//  Permission is hereby granted, free of charge, to any person obtaining a copy 
-//  of this software and associated documentation files (the "Software"), to deal 
-//  in the Software without restriction, including without limitation the rights 
-//  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 
-//  copies of the Software, and to permit persons to whom the Software is 
-//  furnished to do so, subject to the following conditions:
-//
-//  The above copyright notice and this permission notice shall be included in 
-//  all copies or substantial portions of the Software.
-//
-//  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 
-//  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 
-//  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 
-//  THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 
-//  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 
-//  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 
-//  THE SOFTWARE. OF SUCH DAMAGE.
-//
-//  Or:
-//
-//  Distributed under the Boost Software License, Version 1.0.
-//  (See accompanying file LICENSE_1_0.txt or copy at
-//  http://www.boost.org/LICENSE_1_0.txt)
-/*****************************************************************************/
-
 #define BOOST_RESULT_OF_USE_TR1 1
-// On some platforms, Boost.Coroutine must #define magic symbols before
-// #including platform-API headers. Naturally, that's ineffective unless the
-// Boost.Coroutine #include is the *first* #include of the platform header.
-// That means that client code must generally #include Boost.Coroutine headers
-// before anything else.
-#include <boost/dcoroutine/coroutine.hpp>
 #include <boost/bind.hpp>
 #include <boost/range.hpp>
 #include <boost/utility.hpp>
 #include <boost/shared_ptr.hpp>
+#include <boost/make_shared.hpp>
 
 #include "linden_common.h"
 
 #include <iostream>
 #include <string>
+#include <typeinfo>
 
 #include "../test/lltut.h"
+#include "../test/lltestapp.h"
 #include "llsd.h"
 #include "llsdutil.h"
 #include "llevents.h"
-#include "tests/wrapllerrs.h"
-#include "stringize.h"
 #include "llcoros.h"
+#include "lleventfilter.h"
 #include "lleventcoro.h"
 #include "../test/debug.h"
+#include "../test/sync.h"
 
 using namespace llcoro;
 
 /*****************************************************************************
-*   from the banana.cpp example program borrowed for test<1>()
-*****************************************************************************/
-namespace coroutines = boost::dcoroutines;
-using coroutines::coroutine;
-
-template<typename Iter>
-bool match(Iter first, Iter last, std::string match) {
-  std::string::iterator i = match.begin();
-  for(; (first != last) && (i != match.end()); ++i) {
-    if (*first != *i)
-      return false;
-    ++first;
-  }
-  return i == match.end();
-}
-
-template<typename BidirectionalIterator> 
-BidirectionalIterator 
-match_substring(BidirectionalIterator begin, 
-		BidirectionalIterator end, 
-		std::string xmatch,
-		BOOST_DEDUCED_TYPENAME coroutine<BidirectionalIterator(void)>::self& self) { 
-//BidirectionalIterator begin_ = begin;
-  for(; begin != end; ++begin) 
-    if(match(begin, end, xmatch)) {
-      self.yield(begin);
-    }
-  return end;
-} 
-
-typedef coroutine<std::string::iterator(void)> match_coroutine_type;
-
-/*****************************************************************************
 *   Test helpers
 *****************************************************************************/
 /// Simulate an event API whose response is immediate: sent on receipt of the
@@ -131,8 +62,9 @@ typedef coroutine<std::string::iterator(void)> match_coroutine_type;
 class ImmediateAPI
 {
 public:
-    ImmediateAPI():
-        mPump("immediate", true)
+    ImmediateAPI(Sync& sync):
+        mPump("immediate", true),
+        mSync(sync)
     {
         mPump.listen("API", boost::bind(&ImmediateAPI::operator(), this, _1));
     }
@@ -141,20 +73,18 @@ public:
 
     // Invoke this with an LLSD map containing:
     // ["value"]: Integer value. We will reply with ["value"] + 1.
-    // ["reply"]: Name of LLEventPump on which to send success response.
-    // ["error"]: Name of LLEventPump on which to send error response.
-    // ["fail"]: Presence of this key selects ["error"], else ["success"] as
-    // the name of the pump on which to send the response.
+    // ["reply"]: Name of LLEventPump on which to send response.
     bool operator()(const LLSD& event) const
     {
+        mSync.bump();
         LLSD::Integer value(event["value"]);
-        LLSD::String replyPumpName(event.has("fail")? "error" : "reply");
-        LLEventPumps::instance().obtain(event[replyPumpName]).post(value + 1);
+        LLEventPumps::instance().obtain(event["reply"]).post(value + 1);
         return false;
     }
 
 private:
     LLEventStream mPump;
+    Sync& mSync;
 };
 
 /*****************************************************************************
@@ -162,633 +92,247 @@ private:
 *****************************************************************************/
 namespace tut
 {
-    struct coroutine_data {};
-    typedef test_group<coroutine_data> coroutine_group;
+    struct test_data
+    {
+        Sync mSync;
+        ImmediateAPI immediateAPI{mSync};
+        std::string replyName, errorName, threw, stringdata;
+        LLSD result, errordata;
+        int which;
+        LLTestApp testApp;
+
+        void explicit_wait(boost::shared_ptr<LLCoros::Promise<std::string>>& cbp);
+        void waitForEventOn1();
+        void coroPump();
+        void postAndWait1();
+        void coroPumpPost();
+    };
+    typedef test_group<test_data> coroutine_group;
     typedef coroutine_group::object object;
     coroutine_group coroutinegrp("coroutine");
 
-    template<> template<>
-    void object::test<1>()
-    {
-        set_test_name("From banana.cpp example program in Boost.Coroutine distro");
-        std::string buffer = "banananana"; 
-        std::string match = "nana"; 
-        std::string::iterator begin = buffer.begin();
-        std::string::iterator end = buffer.end();
-
-#if defined(BOOST_CORO_POSIX_IMPL)
-//      std::cout << "Using Boost.Coroutine " << BOOST_CORO_POSIX_IMPL << '\n';
-#else
-//      std::cout << "Using non-Posix Boost.Coroutine implementation" << std::endl;
-#endif
-
-        typedef std::string::iterator signature(std::string::iterator, 
-                                                std::string::iterator, 
-                                                std::string,
-                                                match_coroutine_type::self&);
-
-        coroutine<std::string::iterator(void)> matcher
-            (boost::bind(static_cast<signature*>(match_substring), 
-                         begin, 
-                         end, 
-                         match, 
-                         _1)); 
-
-        std::string::iterator i = matcher();
-/*==========================================================================*|
-        while(matcher && i != buffer.end()) {
-            std::cout <<"Match at: "<< std::distance(buffer.begin(), i)<<'\n'; 
-            i = matcher();
-        }
-|*==========================================================================*/
-        size_t matches[] = { 2, 4, 6 };
-        for (size_t *mi(boost::begin(matches)), *mend(boost::end(matches));
-             mi != mend; ++mi, i = matcher())
-        {
-            ensure("more", matcher);
-            ensure("found", i != buffer.end());
-            ensure_equals("value", std::distance(buffer.begin(), i), *mi);
-        }
-        ensure("done", ! matcher);
-    }
-
-    // use static data so we can intersperse coroutine functions with the
-    // tests that engage them
-    ImmediateAPI immediateAPI;
-    std::string replyName, errorName, threw, stringdata;
-    LLSD result, errordata;
-    int which;
-
-    // reinit vars at the start of each test
-    void clear()
-    {
-        replyName.clear();
-        errorName.clear();
-        threw.clear();
-        stringdata.clear();
-        result = LLSD();
-        errordata = LLSD();
-        which = 0;
-    }
-
-    void explicit_wait(boost::shared_ptr<LLCoros::Future<std::string>::callback_t>& cbp)
+    void test_data::explicit_wait(boost::shared_ptr<LLCoros::Promise<std::string>>& cbp)
     {
         BEGIN
         {
+            mSync.bump();
             // The point of this test is to verify / illustrate suspending a
             // coroutine for something other than an LLEventPump. In other
             // words, this shows how to adapt to any async operation that
             // provides a callback-style notification (and prove that it
             // works).
 
-            LLCoros::Future<std::string> future;
-            // get the callback from that future
-            LLCoros::Future<std::string>::callback_t callback(future.make_callback());
-
             // Perhaps we would send a request to a remote server and arrange
-            // for 'callback' to be called on response. Of course that might
-            // involve an adapter object from the actual callback signature to
-            // the signature of 'callback' -- in this case, void(std::string).
-            // For test purposes, instead of handing 'callback' (or the
+            // for cbp->set_value() to be called on response.
+            // For test purposes, instead of handing 'callback' (or an
             // adapter) off to some I/O subsystem, we'll just pass it back to
             // our caller.
-            cbp.reset(new LLCoros::Future<std::string>::callback_t(callback));
+            cbp = boost::make_shared<LLCoros::Promise<std::string>>();
+            LLCoros::Future<std::string> future = LLCoros::getFuture(*cbp);
 
-            ensure("Not yet", ! future);
             // calling get() on the future causes us to suspend
             debug("about to suspend");
             stringdata = future.get();
-            ensure("Got it", bool(future));
+            mSync.bump();
+            ensure_equals("Got it", stringdata, "received");
         }
         END
     }
 
     template<> template<>
-    void object::test<2>()
+    void object::test<1>()
     {
-        clear();
         set_test_name("explicit_wait");
         DEBUG;
 
         // Construct the coroutine instance that will run explicit_wait.
-        boost::shared_ptr<LLCoros::Future<std::string>::callback_t> respond;
-        LLCoros::instance().launch("test<2>",
-                                   boost::bind(explicit_wait, boost::ref(respond)));
+        boost::shared_ptr<LLCoros::Promise<std::string>> respond;
+        LLCoros::instance().launch("test<1>",
+                                   [this, &respond](){ explicit_wait(respond); });
+        mSync.bump();
         // When the coroutine waits for the future, it returns here.
         debug("about to respond");
-        // Now we're the I/O subsystem delivering a result. This immediately
-        // transfers control back to the coroutine.
-        (*respond)("received");
+        // Now we're the I/O subsystem delivering a result. This should make
+        // the coroutine ready.
+        respond->set_value("received");
+        // but give it a chance to wake up
+        mSync.yield();
         // ensure the coroutine ran and woke up again with the intended result
         ensure_equals(stringdata, "received");
     }
 
-    void waitForEventOn1()
+    void test_data::waitForEventOn1()
     {
         BEGIN
         {
+            mSync.bump();
             result = suspendUntilEventOn("source");
+            mSync.bump();
         }
         END
     }
 
     template<> template<>
-    void object::test<3>()
+    void object::test<2>()
     {
-        clear();
         set_test_name("waitForEventOn1");
         DEBUG;
-        LLCoros::instance().launch("test<3>", waitForEventOn1);
+        LLCoros::instance().launch("test<2>", [this](){ waitForEventOn1(); });
+        mSync.bump();
         debug("about to send");
         LLEventPumps::instance().obtain("source").post("received");
+        // give waitForEventOn1() a chance to run
+        mSync.yield();
         debug("back from send");
         ensure_equals(result.asString(), "received");
     }
 
-    void waitForEventOn2()
-    {
-        BEGIN
-        {
-            LLEventWithID pair = suspendUntilEventOn("reply", "error");
-            result = pair.first;
-            which  = pair.second;
-            debug(STRINGIZE("result = " << result << ", which = " << which));
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<4>()
-    {
-        clear();
-        set_test_name("waitForEventOn2 reply");
-        {
-        DEBUG;
-        LLCoros::instance().launch("test<4>", waitForEventOn2);
-        debug("about to send");
-        LLEventPumps::instance().obtain("reply").post("received");
-        debug("back from send");
-        }
-        ensure_equals(result.asString(), "received");
-        ensure_equals("which pump", which, 0);
-    }
-
-    template<> template<>
-    void object::test<5>()
-    {
-        clear();
-        set_test_name("waitForEventOn2 error");
-        DEBUG;
-        LLCoros::instance().launch("test<5>", waitForEventOn2);
-        debug("about to send");
-        LLEventPumps::instance().obtain("error").post("badness");
-        debug("back from send");
-        ensure_equals(result.asString(), "badness");
-        ensure_equals("which pump", which, 1);
-    }
-
-    void coroPump()
+    void test_data::coroPump()
     {
         BEGIN
         {
+            mSync.bump();
             LLCoroEventPump waiter;
             replyName = waiter.getName();
             result = waiter.suspend();
+            mSync.bump();
         }
         END
     }
 
     template<> template<>
-    void object::test<6>()
+    void object::test<3>()
     {
-        clear();
         set_test_name("coroPump");
         DEBUG;
-        LLCoros::instance().launch("test<6>", coroPump);
+        LLCoros::instance().launch("test<3>", [this](){ coroPump(); });
+        mSync.bump();
         debug("about to send");
         LLEventPumps::instance().obtain(replyName).post("received");
+        // give coroPump() a chance to run
+        mSync.yield();
         debug("back from send");
         ensure_equals(result.asString(), "received");
     }
 
-    void coroPumps()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            replyName = waiter.getName0();
-            errorName = waiter.getName1();
-            LLEventWithID pair(waiter.suspend());
-            result = pair.first;
-            which  = pair.second;
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<7>()
-    {
-        clear();
-        set_test_name("coroPumps reply");
-        DEBUG;
-        LLCoros::instance().launch("test<7>", coroPumps);
-        debug("about to send");
-        LLEventPumps::instance().obtain(replyName).post("received");
-        debug("back from send");
-        ensure_equals(result.asString(), "received");
-        ensure_equals("which pump", which, 0);
-    }
-
-    template<> template<>
-    void object::test<8>()
-    {
-        clear();
-        set_test_name("coroPumps error");
-        DEBUG;
-        LLCoros::instance().launch("test<8>", coroPumps);
-        debug("about to send");
-        LLEventPumps::instance().obtain(errorName).post("badness");
-        debug("back from send");
-        ensure_equals(result.asString(), "badness");
-        ensure_equals("which pump", which, 1);
-    }
-
-    void coroPumpsNoEx()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            replyName = waiter.getName0();
-            errorName = waiter.getName1();
-            result = waiter.suspendWithException();
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<9>()
-    {
-        clear();
-        set_test_name("coroPumpsNoEx");
-        DEBUG;
-        LLCoros::instance().launch("test<9>", coroPumpsNoEx);
-        debug("about to send");
-        LLEventPumps::instance().obtain(replyName).post("received");
-        debug("back from send");
-        ensure_equals(result.asString(), "received");
-    }
-
-    void coroPumpsEx()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            replyName = waiter.getName0();
-            errorName = waiter.getName1();
-            try
-            {
-                result = waiter.suspendWithException();
-                debug("no exception");
-            }
-            catch (const LLErrorEvent& e)
-            {
-                debug(STRINGIZE("exception " << e.what()));
-                errordata = e.getData();
-            }
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<10>()
-    {
-        clear();
-        set_test_name("coroPumpsEx");
-        DEBUG;
-        LLCoros::instance().launch("test<10>", coroPumpsEx);
-        debug("about to send");
-        LLEventPumps::instance().obtain(errorName).post("badness");
-        debug("back from send");
-        ensure("no result", result.isUndefined());
-        ensure_equals("got error", errordata.asString(), "badness");
-    }
-
-    void coroPumpsNoLog()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            replyName = waiter.getName0();
-            errorName = waiter.getName1();
-            result = waiter.suspendWithLog();
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<11>()
-    {
-        clear();
-        set_test_name("coroPumpsNoLog");
-        DEBUG;
-        LLCoros::instance().launch("test<11>", coroPumpsNoLog);
-        debug("about to send");
-        LLEventPumps::instance().obtain(replyName).post("received");
-        debug("back from send");
-        ensure_equals(result.asString(), "received");
-    }
-
-    void coroPumpsLog()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            replyName = waiter.getName0();
-            errorName = waiter.getName1();
-            WrapLLErrs capture;
-            threw = capture.catch_llerrs([&waiter, &debug](){
-                    result = waiter.suspendWithLog();
-                    debug("no exception");
-                });
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<12>()
-    {
-        clear();
-        set_test_name("coroPumpsLog");
-        DEBUG;
-        LLCoros::instance().launch("test<12>", coroPumpsLog);
-        debug("about to send");
-        LLEventPumps::instance().obtain(errorName).post("badness");
-        debug("back from send");
-        ensure("no result", result.isUndefined());
-        ensure_contains("got error", threw, "badness");
-    }
-
-    void postAndWait1()
+    void test_data::postAndWait1()
     {
         BEGIN
         {
+            mSync.bump();
             result = postAndSuspend(LLSDMap("value", 17),       // request event
                                  immediateAPI.getPump(),     // requestPump
                                  "reply1",                   // replyPump
                                  "reply");                   // request["reply"] = name
+            mSync.bump();
         }
         END
     }
 
     template<> template<>
-    void object::test<13>()
+    void object::test<4>()
     {
-        clear();
         set_test_name("postAndWait1");
         DEBUG;
-        LLCoros::instance().launch("test<13>", postAndWait1);
+        LLCoros::instance().launch("test<4>", [this](){ postAndWait1(); });
         ensure_equals(result.asInteger(), 18);
     }
 
-    void postAndWait2()
-    {
-        BEGIN
-        {
-            LLEventWithID pair = ::postAndSuspend2(LLSDMap("value", 18),
-                                                immediateAPI.getPump(),
-                                                "reply2",
-                                                "error2",
-                                                "reply",
-                                                "error");
-            result = pair.first;
-            which  = pair.second;
-            debug(STRINGIZE("result = " << result << ", which = " << which));
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<14>()
-    {
-        clear();
-        set_test_name("postAndWait2");
-        DEBUG;
-        LLCoros::instance().launch("test<14>", postAndWait2);
-        ensure_equals(result.asInteger(), 19);
-        ensure_equals(which, 0);
-    }
-
-    void postAndWait2_1()
-    {
-        BEGIN
-        {
-            LLEventWithID pair = ::postAndSuspend2(LLSDMap("value", 18)("fail", LLSD()),
-                                                immediateAPI.getPump(),
-                                                "reply2",
-                                                "error2",
-                                                "reply",
-                                                "error");
-            result = pair.first;
-            which  = pair.second;
-            debug(STRINGIZE("result = " << result << ", which = " << which));
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<15>()
-    {
-        clear();
-        set_test_name("postAndWait2_1");
-        DEBUG;
-        LLCoros::instance().launch("test<15>", postAndWait2_1);
-        ensure_equals(result.asInteger(), 19);
-        ensure_equals(which, 1);
-    }
-
-    void coroPumpPost()
+    void test_data::coroPumpPost()
     {
         BEGIN
         {
+            mSync.bump();
             LLCoroEventPump waiter;
             result = waiter.postAndSuspend(LLSDMap("value", 17),
                                         immediateAPI.getPump(), "reply");
+            mSync.bump();
         }
         END
     }
 
     template<> template<>
-    void object::test<16>()
+    void object::test<5>()
     {
-        clear();
         set_test_name("coroPumpPost");
         DEBUG;
-        LLCoros::instance().launch("test<16>", coroPumpPost);
+        LLCoros::instance().launch("test<5>", [this](){ coroPumpPost(); });
         ensure_equals(result.asInteger(), 18);
     }
 
-    void coroPumpsPost()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            LLEventWithID pair(waiter.postAndSuspend(LLSDMap("value", 23),
-                                                  immediateAPI.getPump(), "reply", "error"));
-            result = pair.first;
-            which  = pair.second;
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<17>()
-    {
-        clear();
-        set_test_name("coroPumpsPost reply");
-        DEBUG;
-        LLCoros::instance().launch("test<17>", coroPumpsPost);
-        ensure_equals(result.asInteger(), 24);
-        ensure_equals("which pump", which, 0);
-    }
-
-    void coroPumpsPost_1()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            LLEventWithID pair(
-                waiter.postAndSuspend(LLSDMap("value", 23)("fail", LLSD()),
-                                   immediateAPI.getPump(), "reply", "error"));
-            result = pair.first;
-            which  = pair.second;
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<18>()
-    {
-        clear();
-        set_test_name("coroPumpsPost error");
-        DEBUG;
-        LLCoros::instance().launch("test<18>", coroPumpsPost_1);
-        ensure_equals(result.asInteger(), 24);
-        ensure_equals("which pump", which, 1);
-    }
-
-    void coroPumpsPostNoEx()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            result = waiter.postAndSuspendWithException(LLSDMap("value", 8),
-                                                     immediateAPI.getPump(), "reply", "error");
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<19>()
-    {
-        clear();
-        set_test_name("coroPumpsPostNoEx");
-        DEBUG;
-        LLCoros::instance().launch("test<19>", coroPumpsPostNoEx);
-        ensure_equals(result.asInteger(), 9);
-    }
-
-    void coroPumpsPostEx()
-    {
-        BEGIN
+    template <class PUMP>
+    void test()
+    {
+        PUMP pump(typeid(PUMP).name());
+        bool running{false};
+        LLSD data{LLSD::emptyArray()};
+        // start things off by posting once before even starting the listener
+        // coro
+        LL_DEBUGS() << "test() posting first" << LL_ENDL;
+        LLSD first{LLSDMap("desc", "first")("value", 0)};
+        bool consumed = pump.post(first);
+        ensure("should not have consumed first", ! consumed);
+        // now launch the coro
+        LL_DEBUGS() << "test() launching listener coro" << LL_ENDL;
+        running = true;
+        LLCoros::instance().launch(
+            "listener",
+            [&pump, &running, &data](){
+                // important for this test that we consume posted values
+                LLCoros::instance().set_consuming(true);
+                // should immediately retrieve 'first' without waiting
+                LL_DEBUGS() << "listener coro waiting for first" << LL_ENDL;
+                data.append(llcoro::suspendUntilEventOnWithTimeout(pump, 0.1, LLSD()));
+                // Don't use ensure() from within the coro -- ensure() failure
+                // throws tut::fail, which won't propagate out to the main
+                // test driver, which will result in an odd failure.
+                // Wait for 'second' because it's not already pending.
+                LL_DEBUGS() << "listener coro waiting for second" << LL_ENDL;
+                data.append(llcoro::suspendUntilEventOnWithTimeout(pump, 0.1, LLSD()));
+                // and wait for 'third', which should involve no further waiting
+                LL_DEBUGS() << "listener coro waiting for third" << LL_ENDL;
+                data.append(llcoro::suspendUntilEventOnWithTimeout(pump, 0.1, LLSD()));
+                LL_DEBUGS() << "listener coro done" << LL_ENDL;
+                running = false;
+            });
+        // back from coro at the point where it's waiting for 'second'
+        LL_DEBUGS() << "test() posting second" << LL_ENDL;
+        LLSD second{llsd::map("desc", "second", "value", 1)};
+        consumed = pump.post(second);
+        ensure("should have consumed second", consumed);
+        // This is a key point: even though we've post()ed the value for which
+        // the coroutine is waiting, it's actually still suspended until we
+        // pause for some other reason. The coroutine will only pick up one
+        // value at a time from our 'pump'. It's important to exercise the
+        // case when we post() two values before it picks up either.
+        LL_DEBUGS() << "test() posting third" << LL_ENDL;
+        LLSD third{llsd::map("desc", "third", "value", 2)};
+        consumed = pump.post(third);
+        ensure("should NOT yet have consumed third", ! consumed);
+        // now just wait for coro to finish -- which it eventually will, given
+        // that all its suspend calls have short timeouts.
+        while (running)
         {
-            LLCoroEventPumps waiter;
-            try
-            {
-                result = waiter.postAndSuspendWithException(
-                    LLSDMap("value", 9)("fail", LLSD()),
-                    immediateAPI.getPump(), "reply", "error");
-                debug("no exception");
-            }
-            catch (const LLErrorEvent& e)
-            {
-                debug(STRINGIZE("exception " << e.what()));
-                errordata = e.getData();
-            }
+            LL_DEBUGS() << "test() waiting for coro done" << LL_ENDL;
+            llcoro::suspendUntilTimeout(0.1);
         }
-        END
+        // okay, verify expected results
+        ensure_equals("should have received three values", data,
+                      llsd::array(first, second, third));
+        LL_DEBUGS() << "test() done" << LL_ENDL;
     }
 
     template<> template<>
-    void object::test<20>()
-    {
-        clear();
-        set_test_name("coroPumpsPostEx");
-        DEBUG;
-        LLCoros::instance().launch("test<20>", coroPumpsPostEx);
-        ensure("no result", result.isUndefined());
-        ensure_equals("got error", errordata.asInteger(), 10);
-    }
-
-    void coroPumpsPostNoLog()
-    {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            result = waiter.postAndSuspendWithLog(LLSDMap("value", 30),
-                                               immediateAPI.getPump(), "reply", "error");
-        }
-        END
-    }
-
-    template<> template<>
-    void object::test<21>()
-    {
-        clear();
-        set_test_name("coroPumpsPostNoLog");
-        DEBUG;
-        LLCoros::instance().launch("test<21>", coroPumpsPostNoLog);
-        ensure_equals(result.asInteger(), 31);
-    }
-
-    void coroPumpsPostLog()
+    void object::test<6>()
     {
-        BEGIN
-        {
-            LLCoroEventPumps waiter;
-            WrapLLErrs capture;
-            threw = capture.catch_llerrs(
-                [&waiter, &debug](){
-                    result = waiter.postAndSuspendWithLog(
-                        LLSDMap("value", 31)("fail", LLSD()),
-                        immediateAPI.getPump(), "reply", "error");
-                    debug("no exception");
-                });
-        }
-        END
+        set_test_name("LLEventMailDrop");
+        tut::test<LLEventMailDrop>();
     }
 
     template<> template<>
-    void object::test<22>()
+    void object::test<7>()
     {
-        clear();
-        set_test_name("coroPumpsPostLog");
-        DEBUG;
-        LLCoros::instance().launch("test<22>", coroPumpsPostLog);
-        ensure("no result", result.isUndefined());
-        ensure_contains("got error", threw, "32");
+        set_test_name("LLEventLogProxyFor<LLEventMailDrop>");
+        tut::test< LLEventLogProxyFor<LLEventMailDrop> >();
     }
 }
-
-/*==========================================================================*|
-#include <boost/context/guarded_stack_allocator.hpp>
-
-namespace tut
-{
-    template<> template<>
-    void object::test<23>()
-    {
-        set_test_name("stacksize");
-        std::cout << "default_stacksize: " << boost::context::guarded_stack_allocator::default_stacksize() << '\n';
-    }
-} // namespace tut
-|*==========================================================================*/
diff --git a/indra/llcommon/tests/lleventdispatcher_test.cpp b/indra/llcommon/tests/lleventdispatcher_test.cpp
index a181d5c941..9da1ecfd67 100644
--- a/indra/llcommon/tests/lleventdispatcher_test.cpp
+++ b/indra/llcommon/tests/lleventdispatcher_test.cpp
@@ -23,6 +23,7 @@
 #include "stringize.h"
 #include "tests/wrapllerrs.h"
 #include "../test/catch_and_store_what_in.h"
+#include "../test/debug.h"
 
 #include <map>
 #include <string>
@@ -46,15 +47,6 @@ using boost::lambda::var;
 using namespace llsd;
 
 /*****************************************************************************
-*   Output control
-*****************************************************************************/
-#ifdef DEBUG_ON
-using std::cout;
-#else
-static std::ostringstream cout;
-#endif
-
-/*****************************************************************************
 *   Example data, functions, classes
 *****************************************************************************/
 // We don't need a whole lot of different arbitrary-params methods, just (no |
@@ -155,13 +147,13 @@ struct Vars
     /*------------- no-args (non-const, const, static) methods -------------*/
     void method0()
     {
-        cout << "method0()\n";
+        debug()("method0()");
         i = 17;
     }
 
     void cmethod0() const
     {
-        cout << 'c';
+        debug()('c', NONL);
         const_cast<Vars*>(this)->method0();
     }
 
@@ -170,13 +162,13 @@ struct Vars
     /*------------ Callable (non-const, const, static) methods -------------*/
     void method1(const LLSD& obj)
     {
-        cout << "method1(" << obj << ")\n";
+        debug()("method1(", obj, ")");
         llsd = obj;
     }
 
     void cmethod1(const LLSD& obj) const
     {
-        cout << 'c';
+        debug()('c', NONL);
         const_cast<Vars*>(this)->method1(obj);
     }
 
@@ -196,12 +188,12 @@ struct Vars
         else
             vcp = std::string("'") + cp + "'";
 
-        cout << "methodna(" << b
-             << ", " << i
-             << ", " << f
-             << ", " << d
-             << ", " << vcp
-             << ")\n";
+        debug()("methodna(", b,
+              ", ", i,
+              ", ", f,
+              ", ", d,
+              ", ", vcp,
+              ")");
 
         this->b = b;
         this->i = i;
@@ -218,12 +210,12 @@ struct Vars
             vbin << std::hex << std::setfill('0') << std::setw(2) << unsigned(byte);
         }
 
-        cout << "methodnb(" << "'" << s << "'"
-             << ", " << uuid
-             << ", " << date
-             << ", '" << uri << "'"
-             << ", " << vbin.str()
-             << ")\n";
+        debug()("methodnb(", "'", s, "'",
+              ", ", uuid,
+              ", ", date,
+              ", '", uri, "'",
+              ", ", vbin.str(),
+              ")");
 
         this->s = s;
         this->uuid = uuid;
@@ -234,18 +226,30 @@ struct Vars
 
     void cmethodna(NPARAMSa) const
     {
-        cout << 'c';
+        debug()('c', NONL);
         const_cast<Vars*>(this)->methodna(NARGSa);
     }
 
     void cmethodnb(NPARAMSb) const
     {
-        cout << 'c';
+        debug()('c', NONL);
         const_cast<Vars*>(this)->methodnb(NARGSb);
     }
 
     static void smethodna(NPARAMSa);
     static void smethodnb(NPARAMSb);
+
+    static Debug& debug()
+    {
+        // Lazily initialize this Debug instance so it can notice if main()
+        // has forcibly set LOGTEST. If it were simply a static member, it
+        // would already have examined the environment variable by the time
+        // main() gets around to checking command-line switches. Since we have
+        // a global static Vars instance, the same would be true of a plain
+        // non-static member.
+        static Debug sDebug("Vars");
+        return sDebug;
+    }
 };
 /*------- Global Vars instance for free functions and static methods -------*/
 static Vars g;
@@ -253,25 +257,25 @@ static Vars g;
 /*------------ Static Vars method implementations reference 'g' ------------*/
 void Vars::smethod0()
 {
-    cout << "smethod0() -> ";
+    debug()("smethod0() -> ", NONL);
     g.method0();
 }
 
 void Vars::smethod1(const LLSD& obj)
 {
-    cout << "smethod1(" << obj << ") -> ";
+    debug()("smethod1(", obj, ") -> ", NONL);
     g.method1(obj);
 }
 
 void Vars::smethodna(NPARAMSa)
 {
-    cout << "smethodna(...) -> ";
+    debug()("smethodna(...) -> ", NONL);
     g.methodna(NARGSa);
 }
 
 void Vars::smethodnb(NPARAMSb)
 {
-    cout << "smethodnb(...) -> ";
+    debug()("smethodnb(...) -> ", NONL);
     g.methodnb(NARGSb);
 }
 
@@ -284,25 +288,25 @@ void clear()
 /*------------------- Free functions also reference 'g' --------------------*/
 void free0()
 {
-    cout << "free0() -> ";
+    g.debug()("free0() -> ", NONL);
     g.method0();
 }
 
 void free1(const LLSD& obj)
 {
-    cout << "free1(" << obj << ") -> ";
+    g.debug()("free1(", obj, ") -> ", NONL);
     g.method1(obj);
 }
 
 void freena(NPARAMSa)
 {
-    cout << "freena(...) -> ";
+    g.debug()("freena(...) -> ", NONL);
     g.methodna(NARGSa);
 }
 
 void freenb(NPARAMSb)
 {
-    cout << "freenb(...) -> ";
+    g.debug()("freenb(...) -> ", NONL);
     g.methodnb(NARGSb);
 }
 
@@ -313,6 +317,7 @@ namespace tut
 {
     struct lleventdispatcher_data
     {
+        Debug debug{"test"};
         WrapLLErrs redirect;
         Dispatcher work;
         Vars v;
@@ -431,12 +436,17 @@ namespace tut
             // Same for freenb() et al.
             params = LLSDMap("a", LLSDArray("b")("i")("f")("d")("cp"))
                             ("b", LLSDArray("s")("uuid")("date")("uri")("bin"));
-            cout << "params:\n" << params << "\nparams[\"a\"]:\n" << params["a"] << "\nparams[\"b\"]:\n" << params["b"] << std::endl;
+            debug("params:\n",
+                  params, "\n"
+                  "params[\"a\"]:\n",
+                  params["a"], "\n"
+                  "params[\"b\"]:\n",
+                  params["b"]);
             // default LLSD::Binary value   
             std::vector<U8> binary;
             for (size_t ix = 0, h = 0xaa; ix < 6; ++ix, h += 0x11)
             {
-                binary.push_back(h);
+                binary.push_back((U8)h);
             }
             // Full defaults arrays. We actually don't care what the LLUUID or
             // LLDate values are, as long as they're different from the
@@ -448,7 +458,8 @@ namespace tut
                                                    (LLDate::now())
                                                    (LLURI("http://www.ietf.org/rfc/rfc3986.txt"))
                                                    (binary));
-            cout << "dft_array_full:\n" << dft_array_full << std::endl;
+            debug("dft_array_full:\n",
+                  dft_array_full);
             // Partial defaults arrays.
             foreach(LLSD::String a, ab)
             {
@@ -457,7 +468,8 @@ namespace tut
                     llsd_copy_array(dft_array_full[a].beginArray() + partition,
                                     dft_array_full[a].endArray());
             }
-            cout << "dft_array_partial:\n" << dft_array_partial << std::endl;
+            debug("dft_array_partial:\n",
+                  dft_array_partial);
 
             foreach(LLSD::String a, ab)
             {
@@ -473,7 +485,10 @@ namespace tut
                     dft_map_partial[a][params[a][ix].asString()] = dft_array_full[a][ix];
                 }
             }
-            cout << "dft_map_full:\n" << dft_map_full << "\ndft_map_partial:\n" << dft_map_partial << '\n';
+            debug("dft_map_full:\n",
+                  dft_map_full, "\n"
+                  "dft_map_partial:\n",
+                  dft_map_partial);
 
             // (Free function | static method) with (no | arbitrary) params,
             // map style, no (empty array) defaults
@@ -918,7 +933,12 @@ namespace tut
                                                  params[a].endArray()),
                                  dft_array_partial[a]);
         }
-        cout << "allreq:\n" << allreq << "\nleftreq:\n" << leftreq << "\nrightdft:\n" << rightdft << std::endl;
+        debug("allreq:\n",
+              allreq, "\n"
+              "leftreq:\n",
+              leftreq, "\n"
+              "rightdft:\n",
+              rightdft);
 
         // Generate maps containing parameter names not provided by the
         // dft_map_partial maps.
@@ -930,7 +950,8 @@ namespace tut
                 skipreq[a].erase(me.first);
             }
         }
-        cout << "skipreq:\n" << skipreq << std::endl;
+        debug("skipreq:\n",
+              skipreq);
 
         LLSD groups(LLSDArray       // array of groups
 
@@ -975,7 +996,11 @@ namespace tut
             LLSD names(grp[0]);
             LLSD required(grp[1][0]);
             LLSD optional(grp[1][1]);
-            cout << "For " << names << ",\n" << "required:\n" << required << "\noptional:\n" << optional << std::endl;
+            debug("For ", names, ",\n",
+                  "required:\n",
+                  required, "\n"
+                  "optional:\n",
+                  optional);
 
             // Loop through 'names'
             foreach(LLSD nm, inArray(names))
@@ -1145,7 +1170,7 @@ namespace tut
         std::vector<U8> binary;
         for (size_t h(0x01), i(0); i < 5; h+= 0x22, ++i)
         {
-            binary.push_back(h);
+            binary.push_back((U8)h);
         }
         LLSD args(LLSDMap("a", LLSDArray(true)(17)(3.14)(123.456)("char*"))
                          ("b", LLSDArray("string")
@@ -1163,7 +1188,7 @@ namespace tut
         }
         // Adjust expect["a"]["cp"] for special Vars::cp treatment.
         expect["a"]["cp"] = std::string("'") + expect["a"]["cp"].asString() + "'";
-        cout << "expect: " << expect << '\n';
+        debug("expect: ", expect);
 
         // Use substantially the same logic for args and argsplus
         LLSD argsarrays(LLSDArray(args)(argsplus));
@@ -1218,7 +1243,8 @@ namespace tut
         {
             array_overfull[a].append("bogus");
         }
-        cout << "array_full: " << array_full << "\narray_overfull: " << array_overfull << std::endl;
+        debug("array_full: ", array_full, "\n"
+              "array_overfull: ", array_overfull);
         // We rather hope that LLDate::now() will generate a timestamp
         // distinct from the one it generated in the constructor, moments ago.
         ensure_not_equals("Timestamps too close",
@@ -1233,7 +1259,8 @@ namespace tut
             map_overfull[a] = map_full[a];
             map_overfull[a]["extra"] = "ignore";
         }
-        cout << "map_full: " << map_full << "\nmap_overfull: " << map_overfull << std::endl;
+        debug("map_full: ", map_full, "\n"
+              "map_overfull: ", map_overfull);
         LLSD expect(map_full);
         // Twiddle the const char* param.
         expect["a"]["cp"] = std::string("'") + expect["a"]["cp"].asString() + "'";
@@ -1248,7 +1275,7 @@ namespace tut
         // so won't bother returning it. Predict that behavior to match the
         // LLSD values.
         expect["a"].erase("b");
-        cout << "expect: " << expect << std::endl;
+        debug("expect: ", expect);
         // For this test, calling functions registered with different sets of
         // parameter defaults should make NO DIFFERENCE WHATSOEVER. Every call
         // should pass all params.
diff --git a/indra/llcommon/tests/lleventfilter_test.cpp b/indra/llcommon/tests/lleventfilter_test.cpp
index 1875013794..fa2cb03e95 100644
--- a/indra/llcommon/tests/lleventfilter_test.cpp
+++ b/indra/llcommon/tests/lleventfilter_test.cpp
@@ -36,9 +36,12 @@
 // other Linden headers
 #include "../test/lltut.h"
 #include "stringize.h"
+#include "llsdutil.h"
 #include "listener.h"
 #include "tests/wrapllerrs.h"
 
+#include <typeinfo>
+
 /*****************************************************************************
 *   Test classes
 *****************************************************************************/
@@ -401,6 +404,78 @@ namespace tut
         throttle.post(";17");
         ensure_equals("17", cat.result, "136;12;17"); // "17" delivered
     }
+
+    template<class PUMP>
+    void test()
+    {
+        PUMP pump(typeid(PUMP).name());
+        LLSD data{LLSD::emptyArray()};
+        bool consumed{true};
+        // listener that appends to 'data'
+        // but that also returns the current value of 'consumed'
+        // Instantiate this separately because we're going to listen()
+        // multiple times with the same lambda: LLEventMailDrop only replays
+        // queued events on a new listen() call.
+        auto lambda =
+            [&data, &consumed](const LLSD& event)->bool
+            {
+                data.append(event);
+                return consumed;
+            };
+        {
+            LLTempBoundListener conn = pump.listen("lambda", lambda);
+            pump.post("first");
+        }
+        // first post() should certainly be received by listener
+        ensure_equals("first", data, llsd::array("first"));
+        // the question is, since consumed was true, did it queue the value?
+        data = LLSD::emptyArray();
+        {
+            // if it queued the value, it would be delivered on subsequent
+            // listen() call
+            LLTempBoundListener conn = pump.listen("lambda", lambda);
+        }
+        ensure_equals("empty1", data, LLSD::emptyArray());
+        data = LLSD::emptyArray();
+        // now let's NOT consume the posted data
+        consumed = false;
+        {
+            LLTempBoundListener conn = pump.listen("lambda", lambda);
+            pump.post("second");
+            pump.post("third");
+        }
+        // the two events still arrive
+        ensure_equals("second,third1", data, llsd::array("second", "third"));
+        data = LLSD::emptyArray();
+        {
+            // when we reconnect, these should be delivered again
+            // but this time they should be consumed
+            consumed = true;
+            LLTempBoundListener conn = pump.listen("lambda", lambda);
+        }
+        // unconsumed events were delivered again
+        ensure_equals("second,third2", data, llsd::array("second", "third"));
+        data = LLSD::emptyArray();
+        {
+            // when we reconnect this time, no more unconsumed events
+            LLTempBoundListener conn = pump.listen("lambda", lambda);
+        }
+        ensure_equals("empty2", data, LLSD::emptyArray());
+    }
+
+    template<> template<>
+    void filter_object::test<6>()
+    {
+        set_test_name("LLEventMailDrop");
+        tut::test<LLEventMailDrop>();
+    }
+
+    template<> template<>
+    void filter_object::test<7>()
+    {
+        set_test_name("LLEventLogProxyFor<LLEventMailDrop>");
+        tut::test< LLEventLogProxyFor<LLEventMailDrop> >();
+    }
 } // namespace tut
 
 /*****************************************************************************
diff --git a/indra/llcommon/tests/llexception_test.cpp b/indra/llcommon/tests/llexception_test.cpp
index 6bee1943c2..8ddf636cd1 100644
--- a/indra/llcommon/tests/llexception_test.cpp
+++ b/indra/llcommon/tests/llexception_test.cpp
@@ -305,4 +305,19 @@ namespace tut
         std::cout << center("int", '=', margin) << std::endl;
         catch_several(throw_int, "throw_int");
     }
+
+    template<> template<>
+    void object::test<2>()
+    {
+        set_test_name("reporting exceptions");
+
+        try
+        {
+            LLTHROW(LLException("badness"));
+        }
+        catch (...)
+        {
+            LOG_UNHANDLED_EXCEPTION("llexception test<2>()");
+        }
+    }
 } // namespace tut
diff --git a/indra/llcommon/tests/llinstancetracker_test.cpp b/indra/llcommon/tests/llinstancetracker_test.cpp
index d94fc0c56d..9b89159625 100644
--- a/indra/llcommon/tests/llinstancetracker_test.cpp
+++ b/indra/llcommon/tests/llinstancetracker_test.cpp
@@ -41,7 +41,6 @@
 #include <boost/scoped_ptr.hpp>
 // other Linden headers
 #include "../test/lltut.h"
-#include "wrapllerrs.h"
 
 struct Badness: public std::runtime_error
 {
@@ -112,24 +111,22 @@ namespace tut
     void object::test<2>()
     {
         ensure_equals(Unkeyed::instanceCount(), 0);
-        Unkeyed* dangling = NULL;
+        std::weak_ptr<Unkeyed> dangling;
         {
             Unkeyed one;
             ensure_equals(Unkeyed::instanceCount(), 1);
-            Unkeyed* found = Unkeyed::getInstance(&one);
-            ensure_equals(found, &one);
+            std::weak_ptr<Unkeyed> found = one.getWeak();
+            ensure(! found.expired());
             {
                 boost::scoped_ptr<Unkeyed> two(new Unkeyed);
                 ensure_equals(Unkeyed::instanceCount(), 2);
-                Unkeyed* found = Unkeyed::getInstance(two.get());
-                ensure_equals(found, two.get());
             }
             ensure_equals(Unkeyed::instanceCount(), 1);
-            // store an unwise pointer to a temp Unkeyed instance
-            dangling = &one;
+            // store a weak pointer to a temp Unkeyed instance
+            dangling = found;
         } // make that instance vanish
         // check the now-invalid pointer to the destroyed instance
-        ensure("getInstance(T*) failed to track destruction", ! Unkeyed::getInstance(dangling));
+        ensure("weak_ptr<Unkeyed> failed to track destruction", dangling.expired());
         ensure_equals(Unkeyed::instanceCount(), 0);
     }
 
@@ -142,7 +139,8 @@ namespace tut
         // reimplement LLInstanceTracker using, say, a hash map instead of a
         // std::map. We DO insist that every key appear exactly once.
         typedef std::vector<std::string> StringVector;
-        StringVector keys(Keyed::beginKeys(), Keyed::endKeys());
+        auto snap = Keyed::key_snapshot();
+        StringVector keys(snap.begin(), snap.end());
         std::sort(keys.begin(), keys.end());
         StringVector::const_iterator ki(keys.begin());
         ensure_equals(*ki++, "one");
@@ -153,17 +151,15 @@ namespace tut
         ensure("didn't reach end", ki == keys.end());
 
         // Use a somewhat different approach to order independence with
-        // beginInstances(): explicitly capture the instances we know in a
+        // instance_snapshot(): explicitly capture the instances we know in a
         // set, and delete them as we iterate through.
         typedef std::set<Keyed*> InstanceSet;
         InstanceSet instances;
         instances.insert(&one);
         instances.insert(&two);
         instances.insert(&three);
-        for (Keyed::instance_iter ii(Keyed::beginInstances()), iend(Keyed::endInstances());
-             ii != iend; ++ii)
+        for (auto& ref : Keyed::instance_snapshot())
         {
-            Keyed& ref = *ii;
             ensure_equals("spurious instance", instances.erase(&ref), 1);
         }
         ensure_equals("unreported instance", instances.size(), 0);
@@ -180,11 +176,10 @@ namespace tut
         instances.insert(&two);
         instances.insert(&three);
 
-		for (Unkeyed::instance_iter ii(Unkeyed::beginInstances()), iend(Unkeyed::endInstances()); ii != iend; ++ii)
-		{
-			Unkeyed& ref = *ii;
-			ensure_equals("spurious instance", instances.erase(&ref), 1);
-		}
+        for (auto& ref : Unkeyed::instance_snapshot())
+        {
+            ensure_equals("spurious instance", instances.erase(&ref), 1);
+        }
 
         ensure_equals("unreported instance", instances.size(), 0);
     }
@@ -192,49 +187,49 @@ namespace tut
     template<> template<>
     void object::test<5>()
     {
-        set_test_name("delete Keyed with outstanding instance_iter");
-        std::string what;
-        Keyed* keyed = new Keyed("delete Keyed with outstanding instance_iter");
-        {
-            WrapLLErrs wrapper;
-            Keyed::instance_iter i(Keyed::beginInstances());
-            what = wrapper.catch_llerrs([&keyed](){
-                    delete keyed;
-                });
-        }
-        ensure(! what.empty());
+        std::string desc("delete Keyed with outstanding instance_snapshot");
+        set_test_name(desc);
+        Keyed* keyed = new Keyed(desc);
+        // capture a snapshot but do not yet traverse it
+        auto snapshot = Keyed::instance_snapshot();
+        // delete the one instance
+        delete keyed;
+        // traversing the snapshot should reflect the deletion
+        // avoid ensure_equals() because it requires the ability to stream the
+        // two values to std::ostream
+        ensure(snapshot.begin() == snapshot.end());
     }
 
     template<> template<>
     void object::test<6>()
     {
-        set_test_name("delete Keyed with outstanding key_iter");
-        std::string what;
-        Keyed* keyed = new Keyed("delete Keyed with outstanding key_it");
-        {
-            WrapLLErrs wrapper;
-            Keyed::key_iter i(Keyed::beginKeys());
-            what = wrapper.catch_llerrs([&keyed](){
-                    delete keyed;
-                });
-        }
-        ensure(! what.empty());
+        std::string desc("delete Keyed with outstanding key_snapshot");
+        set_test_name(desc);
+        Keyed* keyed = new Keyed(desc);
+        // capture a snapshot but do not yet traverse it
+        auto snapshot = Keyed::key_snapshot();
+        // delete the one instance
+        delete keyed;
+        // traversing the snapshot should reflect the deletion
+        // avoid ensure_equals() because it requires the ability to stream the
+        // two values to std::ostream
+        ensure(snapshot.begin() == snapshot.end());
     }
 
     template<> template<>
     void object::test<7>()
     {
-        set_test_name("delete Unkeyed with outstanding instance_iter");
+        set_test_name("delete Unkeyed with outstanding instance_snapshot");
         std::string what;
         Unkeyed* unkeyed = new Unkeyed;
-        {
-            WrapLLErrs wrapper;
-            Unkeyed::instance_iter i(Unkeyed::beginInstances());
-            what = wrapper.catch_llerrs([&unkeyed](){
-                    delete unkeyed;
-                });
-        }
-        ensure(! what.empty());
+        // capture a snapshot but do not yet traverse it
+        auto snapshot = Unkeyed::instance_snapshot();
+        // delete the one instance
+        delete unkeyed;
+        // traversing the snapshot should reflect the deletion
+        // avoid ensure_equals() because it requires the ability to stream the
+        // two values to std::ostream
+        ensure(snapshot.begin() == snapshot.end());
     }
 
     template<> template<>
@@ -246,11 +241,9 @@ namespace tut
         // We can't use the iterator-range InstanceSet constructor because
         // beginInstances() returns an iterator that dereferences to an
         // Unkeyed&, not an Unkeyed*.
-        for (Unkeyed::instance_iter uki(Unkeyed::beginInstances()),
-                                    ukend(Unkeyed::endInstances());
-             uki != ukend; ++uki)
+        for (auto& ref : Unkeyed::instance_snapshot())
         {
-            existing.insert(&*uki);
+            existing.insert(&ref);
         }
         try
         {
@@ -273,11 +266,9 @@ namespace tut
         // instances was also present in the original set. If that's not true,
         // it's because our new Unkeyed ended up in the updated set despite
         // its constructor exception.
-        for (Unkeyed::instance_iter uki(Unkeyed::beginInstances()),
-                                    ukend(Unkeyed::endInstances());
-             uki != ukend; ++uki)
+        for (auto& ref : Unkeyed::instance_snapshot())
         {
-            ensure("failed to remove instance", existing.find(&*uki) != existing.end());
+            ensure("failed to remove instance", existing.find(&ref) != existing.end());
         }
     }
 } // namespace tut
diff --git a/indra/llcommon/tests/llleap_test.cpp b/indra/llcommon/tests/llleap_test.cpp
index bf0a74d10d..9d71e327d8 100644
--- a/indra/llcommon/tests/llleap_test.cpp
+++ b/indra/llcommon/tests/llleap_test.cpp
@@ -49,24 +49,28 @@ const size_t BUFFERED_LENGTH = 1023*1024; // try wrangling just under a megabyte
 
 #endif
 
-void waitfor(const std::vector<LLLeap*>& instances, int timeout=60)
+// capture std::weak_ptrs to LLLeap instances so we can tell when they expire
+typedef std::vector<std::weak_ptr<LLLeap>> LLLeapVector;
+
+void waitfor(const LLLeapVector& instances, int timeout=60)
 {
     int i;
     for (i = 0; i < timeout; ++i)
     {
         // Every iteration, test whether any of the passed LLLeap instances
         // still exist (are still running).
-        std::vector<LLLeap*>::const_iterator vli(instances.begin()), vlend(instances.end());
-        for ( ; vli != vlend; ++vli)
+        bool found = false;
+        for (auto& ptr : instances)
         {
-            // getInstance() returns NULL if it's terminated/gone, non-NULL if
-            // it's still running
-            if (LLLeap::getInstance(*vli))
+            if (! ptr.expired())
+            {
+                found = true;
                 break;
+            }
         }
         // If we made it through all of 'instances' without finding one that's
         // still running, we're done.
-        if (vli == vlend)
+        if (! found)
         {
 /*==========================================================================*|
             std::cout << instances.size() << " LLLeap instances terminated in "
@@ -86,8 +90,8 @@ void waitfor(const std::vector<LLLeap*>& instances, int timeout=60)
 
 void waitfor(LLLeap* instance, int timeout=60)
 {
-    std::vector<LLLeap*> instances;
-    instances.push_back(instance);
+    LLLeapVector instances;
+    instances.push_back(instance->getWeak());
     waitfor(instances, timeout);
 }
 
@@ -218,11 +222,11 @@ namespace tut
         NamedTempFile script("py",
                              "import time\n"
                              "time.sleep(1)\n");
-        std::vector<LLLeap*> instances;
+        LLLeapVector instances;
         instances.push_back(LLLeap::create(get_test_name(),
-                                           sv(list_of(PYTHON)(script.getName()))));
+                                           sv(list_of(PYTHON)(script.getName())))->getWeak());
         instances.push_back(LLLeap::create(get_test_name(),
-                                           sv(list_of(PYTHON)(script.getName()))));
+                                           sv(list_of(PYTHON)(script.getName())))->getWeak());
         // In this case we're simply establishing that two LLLeap instances
         // can coexist without throwing exceptions or bombing in any other
         // way. Wait for them to terminate.
diff --git a/indra/llcommon/tests/llmainthreadtask_test.cpp b/indra/llcommon/tests/llmainthreadtask_test.cpp
new file mode 100644
index 0000000000..69b11ccafb
--- /dev/null
+++ b/indra/llcommon/tests/llmainthreadtask_test.cpp
@@ -0,0 +1,137 @@
+/**
+ * @file   llmainthreadtask_test.cpp
+ * @author Nat Goodspeed
+ * @date   2019-12-05
+ * @brief  Test for llmainthreadtask.
+ * 
+ * $LicenseInfo:firstyear=2019&license=viewerlgpl$
+ * Copyright (c) 2019, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+// Precompiled header
+#include "linden_common.h"
+// associated header
+#include "llmainthreadtask.h"
+// STL headers
+// std headers
+#include <atomic>
+// external library headers
+// other Linden headers
+#include "../test/lltut.h"
+#include "../test/sync.h"
+#include "llthread.h"               // on_main_thread()
+#include "lleventtimer.h"
+#include "lockstatic.h"
+
+/*****************************************************************************
+*   TUT
+*****************************************************************************/
+namespace tut
+{
+    struct llmainthreadtask_data
+    {
+        // 5-second timeout
+        Sync mSync{F32Milliseconds(5000.0f)};
+
+        llmainthreadtask_data()
+        {
+            // we're not testing the result; this is just to cache the
+            // initial thread as the main thread.
+            on_main_thread();
+        }
+    };
+    typedef test_group<llmainthreadtask_data> llmainthreadtask_group;
+    typedef llmainthreadtask_group::object object;
+    llmainthreadtask_group llmainthreadtaskgrp("llmainthreadtask");
+
+    template<> template<>
+    void object::test<1>()
+    {
+        set_test_name("inline");
+        bool ran = false;
+        bool result = LLMainThreadTask::dispatch(
+            [&ran]()->bool{
+                ran = true;
+                return true;
+            });
+        ensure("didn't run lambda", ran);
+        ensure("didn't return result", result);
+    }
+
+    struct StaticData
+    {
+        std::mutex mMutex;          // LockStatic looks for mMutex
+        bool ran{false};
+    };
+    typedef llthread::LockStatic<StaticData> LockStatic;
+
+    template<> template<>
+    void object::test<2>()
+    {
+        set_test_name("cross-thread");
+        skip("This test is prone to build-time hangs");
+        std::atomic_bool result(false);
+        // wrapping our thread lambda in a packaged_task will catch any
+        // exceptions it might throw and deliver them via future
+        std::packaged_task<void()> thread_work(
+            [this, &result](){
+                // unblock test<2>()'s yield_until(1)
+                mSync.set(1);
+                // dispatch work to main thread -- should block here
+                bool on_main(
+                    LLMainThreadTask::dispatch(
+                        []()->bool{
+                            // have to lock static mutex to set static data
+                            LockStatic()->ran = true;
+                            // indicate whether task was run on the main thread
+                            return on_main_thread();
+                        }));
+                // wait for test<2>() to unblock us again
+                mSync.yield_until(3);
+                result = on_main;
+            });
+        auto thread_result = thread_work.get_future();
+        std::thread thread;
+        try
+        {
+            // run thread_work
+            thread = std::thread(std::move(thread_work));
+            // wait for thread to set(1)
+            mSync.yield_until(1);
+            // try to acquire the lock, should block because thread has it
+            LockStatic lk;
+            // wake up when dispatch() unlocks the static mutex
+            ensure("shouldn't have run yet", !lk->ran);
+            ensure("shouldn't have returned yet", !result);
+            // unlock so the task can acquire the lock
+            lk.unlock();
+            // run the task -- should unblock thread, which will immediately block
+            // on mSync
+            LLEventTimer::updateClass();
+            // 'lk', having unlocked, can no longer be used to access; relock with
+            // a new LockStatic instance
+            ensure("should now have run", LockStatic()->ran);
+            ensure("returned too early", !result);
+            // okay, let thread perform the assignment
+            mSync.set(3);
+        }
+        catch (...)
+        {
+            // A test failure exception anywhere in the try block can cause
+            // the test program to terminate without explanation when
+            // ~thread() finds that 'thread' is still joinable. We could
+            // either join() or detach() it -- but since it might be blocked
+            // waiting for something from the main thread that now can never
+            // happen, it's safer to detach it.
+            thread.detach();
+            throw;
+        }
+        // 'thread' should be all done now
+        thread.join();
+        // deliver any exception thrown by thread_work
+        thread_result.get();
+        ensure("ran changed", LockStatic()->ran);
+        ensure("didn't run on main thread", result);
+    }
+} // namespace tut
diff --git a/indra/llcommon/tests/llprocess_test.cpp b/indra/llcommon/tests/llprocess_test.cpp
index 222d832084..f0eafa8201 100644
--- a/indra/llcommon/tests/llprocess_test.cpp
+++ b/indra/llcommon/tests/llprocess_test.cpp
@@ -493,14 +493,18 @@ namespace tut
         }
 //      std::cout << "child done: rv = " << rv << " (" << manager.strerror(rv) << "), why = " << why << ", rc = " << rc << '\n';
         aprchk_("apr_proc_wait(wi->child, &wi->rc, &wi->why, APR_NOWAIT)", wi.rv, APR_CHILD_DONE);
-        ensure_equals_(wi.why, APR_PROC_EXIT);
-        ensure_equals_(wi.rc, 0);
 
         // Beyond merely executing all the above successfully, verify that we
         // obtained expected output -- and that we duly got control while
         // waiting, proving the non-blocking nature of these pipes.
         try
         {
+            // Perform these ensure_equals_() within this try/catch so that if
+            // we don't get expected results, we'll dump whatever we did get
+            // to help diagnose.
+            ensure_equals_(wi.why, APR_PROC_EXIT);
+            ensure_equals_(wi.rc, 0);
+
             unsigned i = 0;
             ensure("blocking I/O on child pipe (0)", history[i].tries);
             ensure_equals_(history[i].which, "out");
diff --git a/indra/llcommon/tests/llsdserialize_test.cpp b/indra/llcommon/tests/llsdserialize_test.cpp
index 6ac974e659..642c1c3879 100644
--- a/indra/llcommon/tests/llsdserialize_test.cpp
+++ b/indra/llcommon/tests/llsdserialize_test.cpp
@@ -271,10 +271,10 @@ namespace tut
 		LLSD w;
 		mParser->reset();	// reset() call is needed since test code re-uses mParser
 		mParser->parse(stream, w, stream.str().size());
-		
+
 		try
 		{
-			ensure_equals(msg.c_str(), w, v);
+			ensure_equals(msg, w, v);
 		}
 		catch (...)
 		{
@@ -432,6 +432,7 @@ namespace tut
 		
 		const char source[] = "it must be a blue moon again";
 		std::vector<U8> data;
+		// note, includes terminating '\0'
 		copy(&source[0], &source[sizeof(source)], back_inserter(data));
 		
 		v = data;
@@ -468,28 +469,36 @@ namespace tut
 		checkRoundTrip(msg + " many nested maps", v);
 	}
 	
-	typedef tut::test_group<TestLLSDSerializeData> TestLLSDSerialzeGroup;
-	typedef TestLLSDSerialzeGroup::object TestLLSDSerializeObject;
-	TestLLSDSerialzeGroup gTestLLSDSerializeGroup("llsd serialization");
+	typedef tut::test_group<TestLLSDSerializeData> TestLLSDSerializeGroup;
+	typedef TestLLSDSerializeGroup::object TestLLSDSerializeObject;
+	TestLLSDSerializeGroup gTestLLSDSerializeGroup("llsd serialization");
 
 	template<> template<> 
 	void TestLLSDSerializeObject::test<1>()
 	{
-		mFormatter = new LLSDNotationFormatter();
+		mFormatter = new LLSDNotationFormatter(false, "", LLSDFormatter::OPTIONS_PRETTY_BINARY);
 		mParser = new LLSDNotationParser();
-		doRoundTripTests("notation serialization");
+		doRoundTripTests("pretty binary notation serialization");
 	}
-	
+
 	template<> template<> 
 	void TestLLSDSerializeObject::test<2>()
 	{
+		mFormatter = new LLSDNotationFormatter(false, "", LLSDFormatter::OPTIONS_NONE);
+		mParser = new LLSDNotationParser();
+		doRoundTripTests("raw binary notation serialization");
+	}
+
+	template<> template<> 
+	void TestLLSDSerializeObject::test<3>()
+	{
 		mFormatter = new LLSDXMLFormatter();
 		mParser = new LLSDXMLParser();
 		doRoundTripTests("xml serialization");
 	}
-	
+
 	template<> template<> 
-	void TestLLSDSerializeObject::test<3>()
+	void TestLLSDSerializeObject::test<4>()
 	{
 		mFormatter = new LLSDBinaryFormatter();
 		mParser = new LLSDBinaryParser();
diff --git a/indra/llcommon/tests/llsingleton_test.cpp b/indra/llcommon/tests/llsingleton_test.cpp
index 75ddff9d7d..15ffe68e67 100644
--- a/indra/llcommon/tests/llsingleton_test.cpp
+++ b/indra/llcommon/tests/llsingleton_test.cpp
@@ -143,8 +143,6 @@ namespace tut
                                                                         \
         (void)CLS::instance();                                          \
         ensure_equals(sLog, #CLS "i" #CLS);                             \
-        LLSingletonBase::cleanupAll();                                  \
-        ensure_equals(sLog, #CLS "i" #CLS "x" #CLS);                    \
         LLSingletonBase::deleteAll();                                   \
         ensure_equals(sLog, #CLS "i" #CLS "x" #CLS "~" #CLS);           \
     }                                                                   \
@@ -159,10 +157,8 @@ namespace tut
                                                                         \
         (void)CLS::instance();                                          \
         ensure_equals(sLog, #CLS #OTHER "i" #OTHER "i" #CLS);           \
-        LLSingletonBase::cleanupAll();                                  \
-        ensure_equals(sLog, #CLS #OTHER "i" #OTHER "i" #CLS "x" #CLS "x" #OTHER); \
         LLSingletonBase::deleteAll();                                   \
-        ensure_equals(sLog, #CLS #OTHER "i" #OTHER "i" #CLS "x" #CLS "x" #OTHER "~" #CLS "~" #OTHER); \
+        ensure_equals(sLog, #CLS #OTHER "i" #OTHER "i" #CLS "x" #CLS "~" #CLS "x" #OTHER "~" #OTHER); \
     }                                                                   \
                                                                         \
     template<> template<>                                               \
@@ -175,10 +171,8 @@ namespace tut
                                                                         \
         (void)CLS::instance();                                          \
         ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER);           \
-        LLSingletonBase::cleanupAll();                                  \
-        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "x" #OTHER); \
         LLSingletonBase::deleteAll();                                   \
-        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "x" #OTHER "~" #CLS "~" #OTHER); \
+        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "~" #CLS "x" #OTHER "~" #OTHER); \
     }                                                                   \
                                                                         \
     template<> template<>                                               \
@@ -191,10 +185,8 @@ namespace tut
                                                                         \
         (void)CLS::instance();                                          \
         ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER);           \
-        LLSingletonBase::cleanupAll();                                  \
-        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "x" #OTHER); \
         LLSingletonBase::deleteAll();                                   \
-        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "x" #OTHER "~" #CLS "~" #OTHER); \
+        ensure_equals(sLog, #CLS "i" #CLS #OTHER "i" #OTHER "x" #CLS "~" #CLS "x" #OTHER "~" #OTHER); \
     }
 
     TESTS(A, B, 4, 5, 6, 7)