summaryrefslogtreecommitdiff
path: root/indra/llcommon
diff options
context:
space:
mode:
Diffstat (limited to 'indra/llcommon')
-rw-r--r--indra/llcommon/CMakeLists.txt2
-rw-r--r--indra/llcommon/coro_scheduler.cpp164
-rw-r--r--indra/llcommon/coro_scheduler.h73
3 files changed, 239 insertions, 0 deletions
diff --git a/indra/llcommon/CMakeLists.txt b/indra/llcommon/CMakeLists.txt
index 20670d7ebe..221fc2bb3f 100644
--- a/indra/llcommon/CMakeLists.txt
+++ b/indra/llcommon/CMakeLists.txt
@@ -18,6 +18,7 @@ include(Tracy)
set(llcommon_SOURCE_FILES
apply.cpp
commoncontrol.cpp
+ coro_scheduler.cpp
hbxxh.cpp
indra_constants.cpp
lazyeventapi.cpp
@@ -125,6 +126,7 @@ set(llcommon_HEADER_FILES
chrono.h
classic_callback.h
commoncontrol.h
+ coro_scheduler.h
ctype_workaround.h
fix_macros.h
fsyspath.h
diff --git a/indra/llcommon/coro_scheduler.cpp b/indra/llcommon/coro_scheduler.cpp
new file mode 100644
index 0000000000..337162cbd5
--- /dev/null
+++ b/indra/llcommon/coro_scheduler.cpp
@@ -0,0 +1,164 @@
+/**
+ * @file coro_scheduler.cpp
+ * @author Nat Goodspeed
+ * @date 2024-08-05
+ * @brief Implementation for llcoro::scheduler.
+ *
+ * $LicenseInfo:firstyear=2024&license=viewerlgpl$
+ * Copyright (c) 2024, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+// Precompiled header
+#include "linden_common.h"
+// associated header
+#include "coro_scheduler.h"
+// STL headers
+// std headers
+#include <iomanip>
+// external library headers
+#include <boost/fiber/operations.hpp>
+// other Linden headers
+#include "llcallbacklist.h"
+#include "lldate.h"
+#include "llerror.h"
+
+namespace llcoro
+{
+
+const F32 scheduler::DEFAULT_TIMESLICE{ LL::Timers::DEFAULT_TIMESLICE };
+
+const std::string qname("General");
+
+scheduler::scheduler():
+ // Since use_scheduling_algorithm() must be called before any other
+ // Boost.Fibers operations, we can assume that the calling fiber is in
+ // fact the main fiber.
+ mMainID(boost::this_fiber::get_id()),
+ mStart(LLDate::now().secondsSinceEpoch()),
+ mQueue(LL::WorkQueue::getInstance(qname))
+{}
+
+void scheduler::awakened( boost::fibers::context* ctx) noexcept
+{
+ if (ctx->get_id() == mMainID)
+ {
+ // If the fiber that just came ready is the main fiber, record its
+ // pointer.
+ llassert(! mMainCtx);
+ mMainCtx = ctx;
+ }
+ // Delegate to round_robin::awakened() as usual, even for the main fiber.
+ // This way, as long as other fibers don't take too long, we can just let
+ // normal round_robin processing pass control to the main fiber.
+ super::awakened(ctx);
+}
+
+boost::fibers::context* scheduler::pick_next() noexcept
+{
+ // count calls to pick_next()
+ ++mSwitches;
+ // pick_next() is called when the previous fiber has suspended, and we
+ // need to pick another. Did the previous pick_next() call pick the main
+ // fiber? If so, it's the main fiber that just suspended.
+ auto now = LLDate::now().secondsSinceEpoch();
+ if (mMainRunning)
+ {
+ mMainRunning = false;
+ mMainLast = now;
+ }
+
+ boost::fibers::context* next;
+
+ // When the main fiber is ready, and it's been more than mTimeslice since
+ // the main fiber last ran, it's time to intervene.
+ F32 elapsed(now - mMainLast);
+ if (mMainCtx && elapsed > mTimeslice)
+ {
+ // We claim that the main fiber is not only stored in mMainCtx, but is
+ // also queued (somewhere) in our ready list.
+ llassert(mMainCtx->ready_is_linked());
+ // The usefulness of a doubly-linked list is that, given only a
+ // pointer to an item, we can unlink it.
+ mMainCtx->ready_unlink();
+ // Instead of delegating to round_robin::pick_next() to pop the head
+ // of the queue, override by returning mMainCtx.
+ next = mMainCtx;
+
+ /*------------------------- logging stuff --------------------------*/
+ // Unless this log tag is enabled, don't even bother posting.
+ LL_DEBUGS("LLCoros.scheduler");
+ // This feature is inherently hard to verify. The logging in the
+ // lambda below seems useful, but also seems like a lot of overhead
+ // for a coroutine context switch. Try posting the logging lambda to a
+ // ThreadPool to offload that overhead. However, if this is still
+ // taking an unreasonable amount of context-switch time, this whole
+ // passage could be skipped.
+
+ // Record this event for logging, but push it off to a thread pool to
+ // perform that work. Presumably std::weak_ptr::lock() is cheaper than
+ // WorkQueue::getInstance().
+ LL::WorkQueue::ptr_t queue{ mQueue.lock() };
+ // We probably started before the relevant WorkQueue was created.
+ if (! queue)
+ {
+ // Try again to locate the specified WorkQueue.
+ queue = LL::WorkQueue::getInstance(qname);
+ mQueue = queue;
+ }
+ // Both the lock() call and the getInstance() call might have failed.
+ if (queue)
+ {
+ // Bind values. Do NOT bind 'this' to avoid cross-thread access!
+ // It would be interesting to know from what queue position we
+ // unlinked the main fiber, out of how many in the ready list.
+ // Unfortunately round_robin::rqueue_ is private, not protected,
+ // so we have no access.
+ queue->post(
+ [switches=mSwitches, start=mStart, elapsed, now]
+ ()
+ {
+ U32 runtime(U32(now) - U32(start));
+ U32 minutes(runtime / 60u);
+ U32 seconds(runtime % 60u);
+ // use stringize to avoid lasting side effects to the
+ // logging ostream
+ LL_DEBUGS("LLCoros.scheduler")
+ << "At time "
+ << stringize(minutes, ":", std::setw(2), std::setfill('0'), seconds)
+ << " (" << switches << " switches), coroutines took "
+ << stringize(std::setprecision(4), elapsed)
+ << " sec, main coroutine jumped queue"
+ << LL_ENDL;
+ });
+ }
+ LL_ENDL;
+ /*----------------------- end logging stuff ------------------------*/
+ }
+ else
+ {
+ // Either the main fiber isn't yet ready, or it hasn't yet been
+ // mTimeslice seconds since the last time the main fiber ran. Business
+ // as usual.
+ next = super::pick_next();
+ }
+
+ // super::pick_next() could also have returned the main fiber, which is
+ // why this is a separate test instead of being folded into the override
+ // case above.
+ if (next && next->get_id() == mMainID)
+ {
+ // we're about to resume the main fiber: it's no longer "ready"
+ mMainCtx = nullptr;
+ // instead, it's "running"
+ mMainRunning = true;
+ }
+ return next;
+}
+
+void scheduler::use()
+{
+ boost::fibers::use_scheduling_algorithm<scheduler>();
+}
+
+} // namespace llcoro
diff --git a/indra/llcommon/coro_scheduler.h b/indra/llcommon/coro_scheduler.h
new file mode 100644
index 0000000000..a7572ccf4d
--- /dev/null
+++ b/indra/llcommon/coro_scheduler.h
@@ -0,0 +1,73 @@
+/**
+ * @file coro_scheduler.h
+ * @author Nat Goodspeed
+ * @date 2024-08-05
+ * @brief Custom scheduler for viewer's Boost.Fibers (aka coroutines)
+ *
+ * $LicenseInfo:firstyear=2024&license=viewerlgpl$
+ * Copyright (c) 2024, Linden Research, Inc.
+ * $/LicenseInfo$
+ */
+
+#if ! defined(LL_CORO_SCHEDULER_H)
+#define LL_CORO_SCHEDULER_H
+
+#include "workqueue.h"
+#include <boost/fiber/fiber.hpp>
+#include <boost/fiber/algo/round_robin.hpp>
+
+/**
+ * llcoro::scheduler is specifically intended for the viewer's main thread.
+ * Its role is to ensure that the main coroutine, responsible for UI
+ * operations and coordinating everything else, doesn't get starved by
+ * secondary coroutines -- however many of those there might be.
+ *
+ * The simple boost::fibers::algo::round_robin scheduler could result in
+ * arbitrary time lag between resumptions of the main coroutine. Of course
+ * every well-behaved viewer coroutine must be coded to yield before too much
+ * real time has elapsed, but sheer volume of secondary coroutines could still
+ * consume unreasonable real time before cycling back to the main coroutine.
+ */
+
+namespace llcoro
+{
+
+class scheduler: public boost::fibers::algo::round_robin
+{
+ using super = boost::fibers::algo::round_robin;
+public:
+ // If the main fiber is ready, and it's been at least this long since the
+ // main fiber last ran, jump the main fiber to the head of the queue.
+ static const F32 DEFAULT_TIMESLICE;
+
+ scheduler();
+ void awakened( boost::fibers::context*) noexcept override;
+ boost::fibers::context* pick_next() noexcept override;
+
+ static void use();
+
+private:
+ // This is the fiber::id of the main fiber. We use this to discover
+ // whether the fiber passed to awakened() is in fact the main fiber.
+ boost::fibers::fiber::id mMainID;
+ // This context* is nullptr until awakened() notices that the main fiber
+ // has become ready, at which point it contains the main fiber's context*.
+ boost::fibers::context* mMainCtx{};
+ // Set when pick_next() returns the main fiber.
+ bool mMainRunning{ false };
+ // If it's been at least this long since the last time the main fiber got
+ // control, jump it to the head of the queue.
+ F32 mTimeslice{ DEFAULT_TIMESLICE };
+ // Timestamp as of the last time we suspended the main fiber.
+ F32 mMainLast{ 0 };
+ // Timestamp of start time
+ F32 mStart{ 0 };
+ // count context switches
+ U64 mSwitches{ 0 };
+ // WorkQueue for deferred logging
+ LL::WorkQueue::weak_t mQueue;
+};
+
+} // namespace llcoro
+
+#endif /* ! defined(LL_CORO_SCHEDULER_H) */