Merge pull request #3583 'faster avatar loading' into release/2025.03

18 files changed, 1655 insertions, 908 deletions

diff --git a/indra/llappearance/llavatarappearance.cpp b/indra/llappearance/llavatarappearance.cpp
index 95d55c835f..3d66809ed6 100644
--- a/indra/llappearance/llavatarappearance.cpp
+++ b/indra/llappearance/llavatarappearance.cpp
@@ -799,7 +799,6 @@ void LLAvatarAppearance::buildCharacter()
     bool status = loadAvatar();
     stop_glerror();
 
-//  gPrintMessagesThisFrame = true;
     LL_DEBUGS() << "Avatar load took " << timer.getElapsedTimeF32() << " seconds." << LL_ENDL;
 
     if (!status)
diff --git a/indra/llmessage/llpacketbuffer.cpp b/indra/llmessage/llpacketbuffer.cpp
index dc5c7a73cb..0b04a560be 100644
--- a/indra/llmessage/llpacketbuffer.cpp
+++ b/indra/llmessage/llpacketbuffer.cpp
@@ -32,8 +32,6 @@
 #include "lltimer.h"
 #include "llhost.h"
 
-///////////////////////////////////////////////////////////
-
 LLPacketBuffer::LLPacketBuffer(const LLHost &host, const char *datap, const S32 size) : mHost(host)
 {
     mSize = 0;
@@ -41,7 +39,7 @@ LLPacketBuffer::LLPacketBuffer(const LLHost &host, const char *datap, const S32
 
     if (size > NET_BUFFER_SIZE)
     {
-        LL_ERRS() << "Sending packet > " << NET_BUFFER_SIZE << " of size " << size << LL_ENDL;
+        LL_ERRS() << "Constructing packet with size=" << size << " > " << NET_BUFFER_SIZE << LL_ENDL;
     }
     else
     {
@@ -51,7 +49,6 @@ LLPacketBuffer::LLPacketBuffer(const LLHost &host, const char *datap, const S32
             mSize = size;
         }
     }
-
 }
 
 LLPacketBuffer::LLPacketBuffer (S32 hSocket)
@@ -59,18 +56,29 @@ LLPacketBuffer::LLPacketBuffer (S32 hSocket)
     init(hSocket);
 }
 
-///////////////////////////////////////////////////////////
-
 LLPacketBuffer::~LLPacketBuffer ()
 {
 }
 
-///////////////////////////////////////////////////////////
-
-void LLPacketBuffer::init (S32 hSocket)
+void LLPacketBuffer::init(S32 hSocket)
 {
     mSize = receive_packet(hSocket, mData);
     mHost = ::get_sender();
     mReceivingIF = ::get_receiving_interface();
 }
 
+void LLPacketBuffer::init(const char* buffer, S32 data_size, const LLHost& host)
+{
+    if (data_size > NET_BUFFER_SIZE)
+    {
+        LL_ERRS() << "Initializing packet with size=" << data_size << " > " << NET_BUFFER_SIZE << LL_ENDL;
+    }
+    else
+    {
+        memcpy(mData, buffer, data_size);
+        mSize = data_size;
+        mHost = host;
+        mReceivingIF = ::get_receiving_interface();
+    }
+}
+
diff --git a/indra/llmessage/llpacketbuffer.h b/indra/llmessage/llpacketbuffer.h
index a2d2973fb0..ac4012d330 100644
--- a/indra/llmessage/llpacketbuffer.h
+++ b/indra/llmessage/llpacketbuffer.h
@@ -35,20 +35,22 @@ class LLPacketBuffer
 {
 public:
     LLPacketBuffer(const LLHost &host, const char *datap, const S32 size);
-    LLPacketBuffer(S32 hSocket);           // receive a packet
+    LLPacketBuffer(S32 hSocket);    // receive a packet
     ~LLPacketBuffer();
 
     S32         getSize() const                 { return mSize; }
     const char  *getData() const                { return mData; }
     LLHost      getHost() const                 { return mHost; }
     LLHost      getReceivingInterface() const   { return mReceivingIF; }
+
     void init(S32 hSocket);
+    void init(const char* buffer, S32 data_size, const LLHost& host);
 
 protected:
-    char    mData[NET_BUFFER_SIZE];        // packet data       /* Flawfinder : ignore */
-    S32     mSize;          // size of buffer in bytes
-    LLHost  mHost;         // source/dest IP and port
-    LLHost  mReceivingIF;         // source/dest IP and port
+    char    mData[NET_BUFFER_SIZE]; // packet data       /* Flawfinder : ignore */
+    S32     mSize;                  // size of buffer in bytes
+    LLHost  mHost;                  // source/dest IP and port
+    LLHost  mReceivingIF;           // source/dest IP and port
 };
 
 #endif
diff --git a/indra/llmessage/llpacketring.cpp b/indra/llmessage/llpacketring.cpp
index be838770a8..ae5a2168db 100644
--- a/indra/llmessage/llpacketring.cpp
+++ b/indra/llmessage/llpacketring.cpp
@@ -1,6 +1,6 @@
 /**
  * @file llpacketring.cpp
- * @brief implementation of LLPacketRing class for a packet.
+ * @brief implementation of LLPacketRing class.
  *
  * $LicenseInfo:firstyear=2001&license=viewerlgpl$
  * Second Life Viewer Source Code
@@ -43,329 +43,301 @@
 #include "message.h"
 #include "u64.h"
 
-///////////////////////////////////////////////////////////
-LLPacketRing::LLPacketRing () :
-    mUseInThrottle(false),
-    mUseOutThrottle(false),
-    mInThrottle(256000.f),
-    mOutThrottle(64000.f),
-    mActualBitsIn(0),
-    mActualBitsOut(0),
-    mMaxBufferLength(64000),
-    mInBufferLength(0),
-    mOutBufferLength(0),
-    mDropPercentage(0.0f),
-    mPacketsToDrop(0x0)
+constexpr S16 MAX_BUFFER_RING_SIZE = 1024;
+constexpr S16 DEFAULT_BUFFER_RING_SIZE = 256;
+
+LLPacketRing::LLPacketRing ()
+    : mPacketRing(DEFAULT_BUFFER_RING_SIZE, nullptr)
 {
+    LLHost invalid_host;
+    for (size_t i = 0; i < mPacketRing.size(); ++i)
+    {
+        mPacketRing[i] = new LLPacketBuffer(invalid_host, nullptr, 0);
+    }
 }
 
-///////////////////////////////////////////////////////////
 LLPacketRing::~LLPacketRing ()
 {
-    cleanup();
+    for (auto packet : mPacketRing)
+    {
+        delete packet;
+    }
+    mPacketRing.clear();
+    mNumBufferedPackets = 0;
+    mNumBufferedBytes = 0;
+    mHeadIndex = 0;
 }
 
-///////////////////////////////////////////////////////////
-void LLPacketRing::cleanup ()
+S32 LLPacketRing::receivePacket (S32 socket, char *datap)
 {
-    LLPacketBuffer *packetp;
+    bool drop = computeDrop();
+    return (mNumBufferedPackets > 0) ?
+        receiveOrDropBufferedPacket(datap, drop) :
+        receiveOrDropPacket(socket, datap, drop);
+}
 
-    while (!mReceiveQueue.empty())
+bool send_packet_helper(int socket, const char * datap, S32 data_size, LLHost host)
+{
+    if (!LLProxy::isSOCKSProxyEnabled())
     {
-        packetp = mReceiveQueue.front();
-        delete packetp;
-        mReceiveQueue.pop();
+        return send_packet(socket, datap, data_size, host.getAddress(), host.getPort());
     }
 
-    while (!mSendQueue.empty())
-    {
-        packetp = mSendQueue.front();
-        delete packetp;
-        mSendQueue.pop();
-    }
-}
+    char headered_send_buffer[NET_BUFFER_SIZE + SOCKS_HEADER_SIZE];
 
-///////////////////////////////////////////////////////////
-void LLPacketRing::dropPackets (U32 num_to_drop)
-{
-    mPacketsToDrop += num_to_drop;
-}
+    proxywrap_t *socks_header = static_cast<proxywrap_t*>(static_cast<void*>(&headered_send_buffer));
+    socks_header->rsv   = 0;
+    socks_header->addr  = host.getAddress();
+    socks_header->port  = htons(host.getPort());
+    socks_header->atype = ADDRESS_IPV4;
+    socks_header->frag  = 0;
 
-///////////////////////////////////////////////////////////
-void LLPacketRing::setDropPercentage (F32 percent_to_drop)
-{
-    mDropPercentage = percent_to_drop;
-}
+    memcpy(headered_send_buffer + SOCKS_HEADER_SIZE, datap, data_size);
 
-void LLPacketRing::setUseInThrottle(const bool use_throttle)
-{
-    mUseInThrottle = use_throttle;
+    return send_packet( socket,
+                        headered_send_buffer,
+                        data_size + SOCKS_HEADER_SIZE,
+                        LLProxy::getInstance()->getUDPProxy().getAddress(),
+                        LLProxy::getInstance()->getUDPProxy().getPort());
 }
 
-void LLPacketRing::setUseOutThrottle(const bool use_throttle)
+bool LLPacketRing::sendPacket(int socket, const char * datap, S32 data_size, LLHost host)
 {
-    mUseOutThrottle = use_throttle;
+    mActualBytesOut += data_size;
+    return send_packet_helper(socket, datap, data_size, host);
 }
 
-void LLPacketRing::setInBandwidth(const F32 bps)
+void LLPacketRing::dropPackets (U32 num_to_drop)
 {
-    mInThrottle.setRate(bps);
+    mPacketsToDrop += num_to_drop;
 }
 
-void LLPacketRing::setOutBandwidth(const F32 bps)
+void LLPacketRing::setDropPercentage (F32 percent_to_drop)
 {
-    mOutThrottle.setRate(bps);
+    mDropPercentage = percent_to_drop;
 }
-///////////////////////////////////////////////////////////
-S32 LLPacketRing::receiveFromRing (S32 socket, char *datap)
-{
 
-    if (mInThrottle.checkOverflow(0))
-    {
-        // We don't have enough bandwidth, don't give them a packet.
-        return 0;
-    }
-
-    LLPacketBuffer *packetp = NULL;
-    if (mReceiveQueue.empty())
+bool LLPacketRing::computeDrop()
+{
+    bool drop= (mDropPercentage > 0.0f && (ll_frand(100.f) < mDropPercentage));
+    if (drop)
     {
-        // No packets on the queue, don't give them any.
-        return 0;
+        ++mPacketsToDrop;
     }
-
-    S32 packet_size = 0;
-    packetp = mReceiveQueue.front();
-    mReceiveQueue.pop();
-    packet_size = packetp->getSize();
-    if (packetp->getData() != NULL)
+    if (mPacketsToDrop > 0)
     {
-        memcpy(datap, packetp->getData(), packet_size); /*Flawfinder: ignore*/
+        --mPacketsToDrop;
+        drop = true;
     }
-    // need to set sender IP/port!!
-    mLastSender = packetp->getHost();
-    mLastReceivingIF = packetp->getReceivingInterface();
-    delete packetp;
-
-    this->mInBufferLength -= packet_size;
-
-    // Adjust the throttle
-    mInThrottle.throttleOverflow(packet_size * 8.f);
-    return packet_size;
+    return drop;
 }
 
-///////////////////////////////////////////////////////////
-S32 LLPacketRing::receivePacket (S32 socket, char *datap)
+S32 LLPacketRing::receiveOrDropPacket(S32 socket, char *datap, bool drop)
 {
     S32 packet_size = 0;
 
-    // If using the throttle, simulate a limited size input buffer.
-    if (mUseInThrottle)
+    // pull straight from socket
+    if (LLProxy::isSOCKSProxyEnabled())
     {
-        bool done = false;
-
-        // push any current net packet (if any) onto delay ring
-        while (!done)
+        char buffer[NET_BUFFER_SIZE + SOCKS_HEADER_SIZE];   /* Flawfinder ignore */
+        packet_size = receive_packet(socket, buffer);
+        if (packet_size > 0)
         {
-            LLPacketBuffer *packetp;
-            packetp = new LLPacketBuffer(socket);
-
-            if (packetp->getSize())
-            {
-                mActualBitsIn += packetp->getSize() * 8;
-
-                // Fake packet loss
-                if (mDropPercentage && (ll_frand(100.f) < mDropPercentage))
-                {
-                    mPacketsToDrop++;
-                }
-
-                if (mPacketsToDrop)
-                {
-                    delete packetp;
-                    packetp = NULL;
-                    packet_size = 0;
-                    mPacketsToDrop--;
-                }
-            }
+            mActualBytesIn += packet_size;
+        }
 
-            // If we faked packet loss, then we don't have a packet
-            // to use for buffer overflow testing
-            if (packetp)
+        if (packet_size > SOCKS_HEADER_SIZE)
+        {
+            if (drop)
             {
-                if (mInBufferLength + packetp->getSize() > mMaxBufferLength)
-                {
-                    // Toss it.
-                    LL_WARNS() << "Throwing away packet, overflowing buffer" << LL_ENDL;
-                    delete packetp;
-                    packetp = NULL;
-                }
-                else if (packetp->getSize())
-                {
-                    mReceiveQueue.push(packetp);
-                    mInBufferLength += packetp->getSize();
-                }
-                else
-                {
-                    delete packetp;
-                    packetp = NULL;
-                    done = true;
-                }
+                packet_size = 0;
             }
             else
             {
-                // No packetp, keep going? - no packetp == faked packet loss
-            }
-        }
-
-        // Now, grab data off of the receive queue according to our
-        // throttled bandwidth settings.
-        packet_size = receiveFromRing(socket, datap);
-    }
-    else
-    {
-        // no delay, pull straight from net
-        if (LLProxy::isSOCKSProxyEnabled())
-        {
-            U8 buffer[NET_BUFFER_SIZE + SOCKS_HEADER_SIZE];
-            packet_size = receive_packet(socket, static_cast<char*>(static_cast<void*>(buffer)));
-
-            if (packet_size > SOCKS_HEADER_SIZE)
-            {
                 // *FIX We are assuming ATYP is 0x01 (IPv4), not 0x03 (hostname) or 0x04 (IPv6)
-                memcpy(datap, buffer + SOCKS_HEADER_SIZE, packet_size - SOCKS_HEADER_SIZE);
+                packet_size -= SOCKS_HEADER_SIZE; // The unwrapped packet size
+                memcpy(datap, buffer + SOCKS_HEADER_SIZE, packet_size);
                 proxywrap_t * header = static_cast<proxywrap_t*>(static_cast<void*>(buffer));
                 mLastSender.setAddress(header->addr);
                 mLastSender.setPort(ntohs(header->port));
-
-                packet_size -= SOCKS_HEADER_SIZE; // The unwrapped packet size
-            }
-            else
-            {
-                packet_size = 0;
+                mLastReceivingIF = ::get_receiving_interface();
             }
         }
         else
         {
-            packet_size = receive_packet(socket, datap);
-            mLastSender = ::get_sender();
+            packet_size = 0;
         }
-
-        mLastReceivingIF = ::get_receiving_interface();
-
-        if (packet_size)  // did we actually get a packet?
+    }
+    else
+    {
+        packet_size = receive_packet(socket, datap);
+        if (packet_size > 0)
         {
-            if (mDropPercentage && (ll_frand(100.f) < mDropPercentage))
+            mActualBytesIn += packet_size;
+            if (drop)
             {
-                mPacketsToDrop++;
+                packet_size = 0;
             }
-
-            if (mPacketsToDrop)
+            else
             {
-                packet_size = 0;
-                mPacketsToDrop--;
+                mLastSender = ::get_sender();
+                mLastReceivingIF = ::get_receiving_interface();
             }
         }
     }
-
     return packet_size;
 }
 
-bool LLPacketRing::sendPacket(int h_socket, char * send_buffer, S32 buf_size, LLHost host)
+S32 LLPacketRing::receiveOrDropBufferedPacket(char *datap, bool drop)
 {
-    bool status = true;
-    if (!mUseOutThrottle)
+    assert(mNumBufferedPackets > 0);
+    S32 packet_size = 0;
+
+    S16 ring_size = (S16)(mPacketRing.size());
+    S16 packet_index = (mHeadIndex + ring_size - mNumBufferedPackets) % ring_size;
+    LLPacketBuffer* packet = mPacketRing[packet_index];
+    packet_size = packet->getSize();
+    mLastSender = packet->getHost();
+    mLastReceivingIF = packet->getReceivingInterface();
+
+    --mNumBufferedPackets;
+    mNumBufferedBytes -= packet_size;
+    if (mNumBufferedPackets == 0)
     {
-        return sendPacketImpl(h_socket, send_buffer, buf_size, host );
+        assert(mNumBufferedBytes == 0);
+    }
+
+    if (!drop)
+    {
+        assert(packet_size > 0);
+        memcpy(datap, packet->getData(), packet_size);
     }
     else
     {
-        mActualBitsOut += buf_size * 8;
-        LLPacketBuffer *packetp = NULL;
-        // See if we've got enough throttle to send a packet.
-        while (!mOutThrottle.checkOverflow(0.f))
-        {
-            // While we have enough bandwidth, send a packet from the queue or the current packet
+        packet_size = 0;
+    }
+    return packet_size;
+}
+
+S32 LLPacketRing::bufferInboundPacket(S32 socket)
+{
+    if (mNumBufferedPackets == mPacketRing.size() && mNumBufferedPackets < MAX_BUFFER_RING_SIZE)
+    {
+        expandRing();
+    }
 
-            S32 packet_size = 0;
-            if (!mSendQueue.empty())
+    LLPacketBuffer* packet = mPacketRing[mHeadIndex];
+    S32 old_packet_size = packet->getSize();
+    S32 packet_size = 0;
+    if (LLProxy::isSOCKSProxyEnabled())
+    {
+        char buffer[NET_BUFFER_SIZE + SOCKS_HEADER_SIZE];   /* Flawfinder ignore */
+        packet_size = receive_packet(socket, buffer);
+        if (packet_size > 0)
+        {
+            mActualBytesIn += packet_size;
+            if (packet_size > SOCKS_HEADER_SIZE)
             {
-                // Send a packet off of the queue
-                LLPacketBuffer *packetp = mSendQueue.front();
-                mSendQueue.pop();
+                // *FIX We are assuming ATYP is 0x01 (IPv4), not 0x03 (hostname) or 0x04 (IPv6)
 
-                mOutBufferLength -= packetp->getSize();
-                packet_size = packetp->getSize();
+                proxywrap_t * header = static_cast<proxywrap_t*>(static_cast<void*>(buffer));
+                LLHost sender;
+                sender.setAddress(header->addr);
+                sender.setPort(ntohs(header->port));
 
-                status = sendPacketImpl(h_socket, packetp->getData(), packet_size, packetp->getHost());
+                packet_size -= SOCKS_HEADER_SIZE; // The unwrapped packet size
+                packet->init(buffer + SOCKS_HEADER_SIZE, packet_size, sender);
 
-                delete packetp;
-                // Update the throttle
-                mOutThrottle.throttleOverflow(packet_size * 8.f);
+                mHeadIndex = (mHeadIndex + 1) % (S16)(mPacketRing.size());
+                if (mNumBufferedPackets < MAX_BUFFER_RING_SIZE)
+                {
+                    ++mNumBufferedPackets;
+                    mNumBufferedBytes += packet_size;
+                }
+                else
+                {
+                    // we overwrote an older packet
+                    mNumBufferedBytes += packet_size - old_packet_size;
+                }
             }
             else
             {
-                // If the queue's empty, we can just send this packet right away.
-                status =  sendPacketImpl(h_socket, send_buffer, buf_size, host );
-                packet_size = buf_size;
-
-                // Update the throttle
-                mOutThrottle.throttleOverflow(packet_size * 8.f);
-
-                // This was the packet we're sending now, there are no other packets
-                // that we need to send
-                return status;
+                packet_size = 0;
             }
-
-        }
-
-        // We haven't sent the incoming packet, add it to the queue
-        if (mOutBufferLength + buf_size > mMaxBufferLength)
-        {
-            // Nuke this packet, we overflowed the buffer.
-            // Toss it.
-            LL_WARNS() << "Throwing away outbound packet, overflowing buffer" << LL_ENDL;
         }
-        else
+    }
+    else
+    {
+        packet->init(socket);
+        packet_size = packet->getSize();
+        if (packet_size > 0)
         {
-            static LLTimer queue_timer;
-            if ((mOutBufferLength > 4192) && queue_timer.getElapsedTimeF32() > 1.f)
+            mActualBytesIn += packet_size;
+
+            mHeadIndex = (mHeadIndex + 1) % (S16)(mPacketRing.size());
+            if (mNumBufferedPackets < MAX_BUFFER_RING_SIZE)
             {
-                // Add it to the queue
-                LL_INFOS() << "Outbound packet queue " << mOutBufferLength << " bytes" << LL_ENDL;
-                queue_timer.reset();
+                ++mNumBufferedPackets;
+                mNumBufferedBytes += packet_size;
+            }
+            else
+            {
+                // we overwrote an older packet
+                mNumBufferedBytes += packet_size - old_packet_size;
             }
-            packetp = new LLPacketBuffer(host, send_buffer, buf_size);
-
-            mOutBufferLength += packetp->getSize();
-            mSendQueue.push(packetp);
         }
     }
-
-    return status;
+    return packet_size;
 }
 
-bool LLPacketRing::sendPacketImpl(int h_socket, const char * send_buffer, S32 buf_size, LLHost host)
+S32 LLPacketRing::drainSocket(S32 socket)
 {
-
-    if (!LLProxy::isSOCKSProxyEnabled())
+    // drain into buffer
+    S32 packet_size = 1;
+    S32 num_loops = 0;
+    S32 old_num_packets = mNumBufferedPackets;
+    while (packet_size > 0)
     {
-        return send_packet(h_socket, send_buffer, buf_size, host.getAddress(), host.getPort());
+        packet_size = bufferInboundPacket(socket);
+        ++num_loops;
     }
+    S32 num_dropped_packets = (num_loops - 1 + old_num_packets) - mNumBufferedPackets;
+    if (num_dropped_packets > 0)
+    {
+        LL_WARNS("Messaging") << "dropped " << num_dropped_packets << " UDP packets" << LL_ENDL;
+    }
+    return (S32)(mNumBufferedPackets);
+}
 
-    char headered_send_buffer[NET_BUFFER_SIZE + SOCKS_HEADER_SIZE];
+bool LLPacketRing::expandRing()
+{
+    // compute larger size
+    constexpr S16 BUFFER_RING_EXPANSION = 256;
+    S16 old_size = (S16)(mPacketRing.size());
+    S16 new_size = llmin(old_size + BUFFER_RING_EXPANSION, MAX_BUFFER_RING_SIZE);
+    if (new_size == old_size)
+    {
+        // mPacketRing is already maxed out
+        return false;
+    }
 
-    proxywrap_t *socks_header = static_cast<proxywrap_t*>(static_cast<void*>(&headered_send_buffer));
-    socks_header->rsv   = 0;
-    socks_header->addr  = host.getAddress();
-    socks_header->port  = htons(host.getPort());
-    socks_header->atype = ADDRESS_IPV4;
-    socks_header->frag  = 0;
+    // make a larger ring and copy packet pointers
+    std::vector<LLPacketBuffer*> new_ring(new_size, nullptr);
+    for (S16 i = 0; i < old_size; ++i)
+    {
+        S16 j = (mHeadIndex + i) % old_size;
+        new_ring[i] = mPacketRing[j];
+    }
 
-    memcpy(headered_send_buffer + SOCKS_HEADER_SIZE, send_buffer, buf_size);
+    // allocate new packets for the remainder of new_ring
+    LLHost invalid_host;
+    for (S16 i = old_size; i < new_size; ++i)
+    {
+        new_ring[i] = new LLPacketBuffer(invalid_host, nullptr, 0);
+    }
 
-    return send_packet( h_socket,
-                        headered_send_buffer,
-                        buf_size + SOCKS_HEADER_SIZE,
-                        LLProxy::getInstance()->getUDPProxy().getAddress(),
-                        LLProxy::getInstance()->getUDPProxy().getPort());
+    // swap the rings and reset mHeadIndex
+    mPacketRing.swap(new_ring);
+    mHeadIndex = mNumBufferedPackets;
+    return true;
 }
diff --git a/indra/llmessage/llpacketring.h b/indra/llmessage/llpacketring.h
index f0e95f8524..0dff2c63b1 100644
--- a/indra/llmessage/llpacketring.h
+++ b/indra/llmessage/llpacketring.h
@@ -25,16 +25,14 @@
  * $/LicenseInfo$
  */
 
-#ifndef LL_LLPACKETRING_H
-#define LL_LLPACKETRING_H
+#pragma once
 
-#include <queue>
+#include <vector>
 
 #include "llhost.h"
 #include "llpacketbuffer.h"
-#include "llproxy.h"
 #include "llthrottle.h"
-#include "net.h"
+
 
 class LLPacketRing
 {
@@ -42,60 +40,65 @@ public:
     LLPacketRing();
     ~LLPacketRing();
 
-    void cleanup();
+    // receive one packet: either buffered or from the socket
+    S32  receivePacket (S32 socket, char *datap);
+
+    // send one packet
+    bool sendPacket(int h_socket, const char * send_buffer, S32 buf_size, LLHost host);
+
+    // drains packets from socket and returns final mNumBufferedPackets
+    S32 drainSocket(S32 socket);
 
     void dropPackets(U32);
     void setDropPercentage (F32 percent_to_drop);
-    void setUseInThrottle(const bool use_throttle);
-    void setUseOutThrottle(const bool use_throttle);
-    void setInBandwidth(const F32 bps);
-    void setOutBandwidth(const F32 bps);
-    S32  receivePacket (S32 socket, char *datap);
-    S32  receiveFromRing (S32 socket, char *datap);
 
-    bool sendPacket(int h_socket, char * send_buffer, S32 buf_size, LLHost host);
+    inline LLHost getLastSender() const;
+    inline LLHost getLastReceivingInterface() const;
 
-    inline LLHost getLastSender();
-    inline LLHost getLastReceivingInterface();
+    S32 getActualInBytes() const { return mActualBytesIn; }
+    S32 getActualOutBytes() const { return mActualBytesOut; }
+    S32 getAndResetActualInBits()   { S32 bits = mActualBytesIn * 8; mActualBytesIn = 0; return bits;}
+    S32 getAndResetActualOutBits()  { S32 bits = mActualBytesOut * 8; mActualBytesOut = 0; return bits;}
 
-    S32 getAndResetActualInBits()               { S32 bits = mActualBitsIn; mActualBitsIn = 0; return bits;}
-    S32 getAndResetActualOutBits()              { S32 bits = mActualBitsOut; mActualBitsOut = 0; return bits;}
+    S32 getNumBufferedPackets() const { return (S32)(mNumBufferedPackets); }
+    S32 getNumBufferedBytes() const { return mNumBufferedBytes; }
 protected:
-    bool mUseInThrottle;
-    bool mUseOutThrottle;
+    // returns 'true' if we should intentionally drop a packet
+    bool computeDrop();
 
-    // For simulating a lower-bandwidth connection - BPS
-    LLThrottle mInThrottle;
-    LLThrottle mOutThrottle;
+    // returns packet_size of received packet, zero or less if no packet found
+    S32 receiveOrDropPacket(S32 socket, char *datap, bool drop);
+    S32 receiveOrDropBufferedPacket(char *datap, bool drop);
 
-    S32 mActualBitsIn;
-    S32 mActualBitsOut;
-    S32 mMaxBufferLength;           // How much data can we queue up before dropping data.
-    S32 mInBufferLength;            // Current incoming buffer length
-    S32 mOutBufferLength;           // Current outgoing buffer length
+    // returns packet_size of packet buffered
+    S32 bufferInboundPacket(S32 socket);
 
-    F32 mDropPercentage;            // % of packets to drop
-    U32 mPacketsToDrop;             // drop next n packets
+    // returns 'true' if ring was expanded
+    bool expandRing();
 
-    std::queue<LLPacketBuffer *> mReceiveQueue;
-    std::queue<LLPacketBuffer *> mSendQueue;
+protected:
+    std::vector<LLPacketBuffer*> mPacketRing;
+    S16 mHeadIndex { 0 };
+    S16 mNumBufferedPackets { 0 };
+    S32 mNumBufferedBytes { 0 };
+
+    S32 mActualBytesIn { 0 };
+    S32 mActualBytesOut { 0 };
+    F32 mDropPercentage { 0.0f };   // % of inbound packets to drop
+    U32 mPacketsToDrop { 0 };       // drop next inbound n packets
 
+    // These are the sender and receiving_interface for the last packet delivered by receivePacket()
     LLHost mLastSender;
     LLHost mLastReceivingIF;
-
-private:
-    bool sendPacketImpl(int h_socket, const char * send_buffer, S32 buf_size, LLHost host);
 };
 
 
-inline LLHost LLPacketRing::getLastSender()
+inline LLHost LLPacketRing::getLastSender() const
 {
     return mLastSender;
 }
 
-inline LLHost LLPacketRing::getLastReceivingInterface()
+inline LLHost LLPacketRing::getLastReceivingInterface() const
 {
     return mLastReceivingIF;
 }
-
-#endif
diff --git a/indra/llmessage/message.cpp b/indra/llmessage/message.cpp
index cfa5178fc6..c130b7a6db 100644
--- a/indra/llmessage/message.cpp
+++ b/indra/llmessage/message.cpp
@@ -656,8 +656,7 @@ bool LLMessageSystem::checkMessages(LockMessageChecker&, S64 frame_count )
 
             // UseCircuitCode is allowed in even from an invalid circuit, so that
             // we can toss circuits around.
-            if(
-                valid_packet &&
+            else if (
                 !cdp &&
                 (mTemplateMessageReader->getMessageName() !=
                  _PREHASH_UseCircuitCode))
@@ -667,8 +666,7 @@ bool LLMessageSystem::checkMessages(LockMessageChecker&, S64 frame_count )
                 valid_packet = false;
             }
 
-            if(
-                valid_packet &&
+            if ( valid_packet &&
                 cdp &&
                 !cdp->getTrusted() &&
                 mTemplateMessageReader->isTrusted())
@@ -680,7 +678,7 @@ bool LLMessageSystem::checkMessages(LockMessageChecker&, S64 frame_count )
                 valid_packet = false;
             }
 
-            if( valid_packet )
+            if ( valid_packet )
             {
                 logValidMsg(cdp, host, recv_reliable, recv_resent, acks>0 );
                 valid_packet = mTemplateMessageReader->readMessage(buffer, host);
@@ -821,6 +819,11 @@ void LLMessageSystem::processAcks(LockMessageChecker&, F32 collect_time)
     }
 }
 
+S32 LLMessageSystem::drainUdpSocket()
+{
+    return mPacketRing.drainSocket(mSocket);
+}
+
 void LLMessageSystem::copyMessageReceivedToSend()
 {
     // NOTE: babbage: switch builder to match reader to avoid
diff --git a/indra/llmessage/message.h b/indra/llmessage/message.h
index b4b0d94021..1844d5e7cd 100644
--- a/indra/llmessage/message.h
+++ b/indra/llmessage/message.h
@@ -417,6 +417,9 @@ public:
     bool    checkMessages(LockMessageChecker&, S64 frame_count = 0 );
     void    processAcks(LockMessageChecker&, F32 collect_time = 0.f);
 
+    // returns total number of buffered packets after the drain
+    S32     drainUdpSocket();
+
     bool    isMessageFast(const char *msg);
     bool    isMessage(const char *msg)
     {
diff --git a/indra/llmessage/net.cpp b/indra/llmessage/net.cpp
index f153c938cf..2be5a9e5b6 100644
--- a/indra/llmessage/net.cpp
+++ b/indra/llmessage/net.cpp
@@ -76,14 +76,8 @@ static U32 gsnReceivingIFAddr = INVALID_HOST_IP_ADDRESS; // Address to which dat
 const char* LOOPBACK_ADDRESS_STRING = "127.0.0.1";
 const char* BROADCAST_ADDRESS_STRING = "255.255.255.255";
 
-#if LL_DARWIN
-    // macOS returns an error when trying to set these to 400000.  Smaller values succeed.
-    const int   SEND_BUFFER_SIZE    = 200000;
-    const int   RECEIVE_BUFFER_SIZE = 200000;
-#else // LL_DARWIN
-    const int   SEND_BUFFER_SIZE    = 400000;
-    const int   RECEIVE_BUFFER_SIZE = 400000;
-#endif // LL_DARWIN
+const int   SEND_BUFFER_SIZE    = 200000;
+const int   RECEIVE_BUFFER_SIZE = 800000;
 
 // universal functions (cross-platform)
 
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index 689f21d29f..a05466b476 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -349,10 +349,6 @@ std::string gLastVersionChannel;
 LLVector3 gWindVec(3.0, 3.0, 0.0);
 LLVector3 gRelativeWindVec(0.0, 0.0, 0.0);
 
-U32 gPacketsIn = 0;
-
-bool gPrintMessagesThisFrame = false;
-
 bool gRandomizeFramerate = false;
 bool gPeriodicSlowFrame = false;
 
@@ -361,6 +357,7 @@ bool gLLErrorActivated = false;
 bool gLogoutInProgress = false;
 
 bool gSimulateMemLeak = false;
+bool gDoDisconnect = false;
 
 // We don't want anyone, especially threads working on the graphics pipeline,
 // to have to block due to this WorkQueue being full.
@@ -374,7 +371,6 @@ const std::string MARKER_FILE_NAME("SecondLife.exec_marker");
 const std::string START_MARKER_FILE_NAME("SecondLife.start_marker");
 const std::string ERROR_MARKER_FILE_NAME("SecondLife.error_marker");
 const std::string LOGOUT_MARKER_FILE_NAME("SecondLife.logout_marker");
-static bool gDoDisconnect = false;
 static std::string gLaunchFileOnQuit;
 
 // Used on Win32 for other apps to identify our window (eg, win_setup)
@@ -1493,9 +1489,9 @@ bool LLAppViewer::doFrame()
 
         {
             LL_PROFILE_ZONE_NAMED_CATEGORY_APP("df pauseMainloopTimeout");
-        pingMainloopTimeout("Main:Sleep");
+            pingMainloopTimeout("Main:Sleep");
 
-        pauseMainloopTimeout();
+            pauseMainloopTimeout();
         }
 
         // Sleep and run background threads
@@ -4242,7 +4238,7 @@ U32 LLAppViewer::getTextureCacheVersion()
 U32 LLAppViewer::getDiskCacheVersion()
 {
     // Viewer disk cache version intorduced in Simple Cache Viewer, change if the cache format changes.
-    const U32 DISK_CACHE_VERSION = 1;
+    const U32 DISK_CACHE_VERSION = 2;
 
     return DISK_CACHE_VERSION ;
 }
@@ -4901,6 +4897,20 @@ void LLAppViewer::idle()
 
     if (gTeleportDisplay)
     {
+        if (gAgent.getTeleportState() == LLAgent::TELEPORT_ARRIVING)
+        {
+            // Teleported, but waiting for things to load, start processing surface data
+            {
+                LL_RECORD_BLOCK_TIME(FTM_NETWORK);
+                gVLManager.unpackData();
+            }
+            {
+                LL_RECORD_BLOCK_TIME(FTM_REGION_UPDATE);
+                const F32 max_region_update_time = .001f; // 1ms
+                LLWorld::getInstance()->updateRegions(max_region_update_time);
+            }
+        }
+
         return;
     }
 
@@ -5337,12 +5347,9 @@ void LLAppViewer::idleNameCache()
 // Handle messages, and all message related stuff
 //
 
-#define TIME_THROTTLE_MESSAGES
 
-#ifdef TIME_THROTTLE_MESSAGES
-#define CHECK_MESSAGES_DEFAULT_MAX_TIME .020f // 50 ms = 50 fps (just for messages!)
+constexpr F32 CHECK_MESSAGES_DEFAULT_MAX_TIME = 0.020f; // 50 ms = 50 fps (just for messages!)
 static F32 CheckMessagesMaxTime = CHECK_MESSAGES_DEFAULT_MAX_TIME;
-#endif
 
 static LLTrace::BlockTimerStatHandle FTM_IDLE_NETWORK("Idle Network");
 static LLTrace::BlockTimerStatHandle FTM_MESSAGE_ACKS("Message Acks");
@@ -5369,6 +5376,7 @@ void LLAppViewer::idleNetwork()
         F32 total_time = 0.0f;
 
         {
+            bool needs_drain = false;
             LockMessageChecker lmc(gMessageSystem);
             while (lmc.checkAllMessages(frame_count, gServicePump))
             {
@@ -5381,54 +5389,44 @@ void LLAppViewer::idleNetwork()
                 }
 
                 total_decoded++;
-                gPacketsIn++;
 
                 if (total_decoded > MESSAGE_MAX_PER_FRAME)
                 {
+                    needs_drain = true;
                     break;
                 }
 
-#ifdef TIME_THROTTLE_MESSAGES
                 // Prevent slow packets from completely destroying the frame rate.
                 // This usually happens due to clumps of avatars taking huge amount
                 // of network processing time (which needs to be fixed, but this is
                 // a good limit anyway).
                 total_time = check_message_timer.getElapsedTimeF32();
                 if (total_time >= CheckMessagesMaxTime)
+                {
+                    needs_drain = true;
                     break;
-#endif
+                }
+            }
+            if (needs_drain || gMessageSystem->mPacketRing.getNumBufferedPackets() > 0)
+            {
+                // Rather than allow packets to silently backup on the socket
+                // we drain them into our own buffer so we know how many exist.
+                S32 num_buffered_packets = gMessageSystem->drainUdpSocket();
+                if (num_buffered_packets > 0)
+                {
+                    // Increase CheckMessagesMaxTime so that we will eventually catch up
+                    CheckMessagesMaxTime *= 1.035f; // 3.5% ~= 2x in 20 frames, ~8x in 60 frames
+                }
+            }
+            else
+            {
+                // Reset CheckMessagesMaxTime to default value
+                CheckMessagesMaxTime = CHECK_MESSAGES_DEFAULT_MAX_TIME;
             }
 
             // Handle per-frame message system processing.
             lmc.processAcks(gSavedSettings.getF32("AckCollectTime"));
         }
-
-#ifdef TIME_THROTTLE_MESSAGES
-        if (total_time >= CheckMessagesMaxTime)
-        {
-            // Increase CheckMessagesMaxTime so that we will eventually catch up
-            CheckMessagesMaxTime *= 1.035f; // 3.5% ~= x2 in 20 frames, ~8x in 60 frames
-        }
-        else
-        {
-            // Reset CheckMessagesMaxTime to default value
-            CheckMessagesMaxTime = CHECK_MESSAGES_DEFAULT_MAX_TIME;
-        }
-#endif
-
-        // Decode enqueued messages...
-        S32 remaining_possible_decodes = MESSAGE_MAX_PER_FRAME - total_decoded;
-
-        if( remaining_possible_decodes <= 0 )
-        {
-            LL_INFOS() << "Maxed out number of messages per frame at " << MESSAGE_MAX_PER_FRAME << LL_ENDL;
-        }
-
-        if (gPrintMessagesThisFrame)
-        {
-            LL_INFOS() << "Decoded " << total_decoded << " msgs this frame!" << LL_ENDL;
-            gPrintMessagesThisFrame = false;
-        }
     }
     add(LLStatViewer::NUM_NEW_OBJECTS, gObjectList.mNumNewObjects);
 
diff --git a/indra/newview/llappviewer.h b/indra/newview/llappviewer.h
index 542379c828..10d71c9d42 100644
--- a/indra/newview/llappviewer.h
+++ b/indra/newview/llappviewer.h
@@ -413,11 +413,10 @@ extern std::string gLastVersionChannel;
 
 extern LLVector3 gWindVec;
 extern LLVector3 gRelativeWindVec;
-extern U32  gPacketsIn;
-extern bool gPrintMessagesThisFrame;
 
 extern bool gRandomizeFramerate;
 extern bool gPeriodicSlowFrame;
+extern bool gDoDisconnect;
 
 extern bool gSimulateMemLeak;
 
diff --git a/indra/newview/llmeshrepository.cpp b/indra/newview/llmeshrepository.cpp
index 93453f507c..e1fa84b4d8 100644
--- a/indra/newview/llmeshrepository.cpp
+++ b/indra/newview/llmeshrepository.cpp
@@ -255,12 +255,12 @@
 //     mSkinInfoQ               mMutex        rw.repo.mMutex, rw.main.mMutex [5] (was:  [0])
 //     mDecompositionRequests   mMutex        rw.repo.mMutex, ro.repo.none [5]
 //     mPhysicsShapeRequests    mMutex        rw.repo.mMutex, ro.repo.none [5]
-//     mDecompositionQ          mMutex        rw.repo.mMutex, rw.main.mMutex [5] (was:  [0])
+//     mDecompositionQ          mMutex        rw.repo.mLoadedMutex, rw.main.mLoadedMutex [5] (was:  [0])
 //     mHeaderReqQ              mMutex        ro.repo.none [5], rw.repo.mMutex, rw.any.mMutex
 //     mLODReqQ                 mMutex        ro.repo.none [5], rw.repo.mMutex, rw.any.mMutex
-//     mUnavailableQ            mMutex        rw.repo.none [0], ro.main.none [5], rw.main.mMutex
-//     mLoadedQ                 mMutex        rw.repo.mMutex, ro.main.none [5], rw.main.mMutex
-//     mPendingLOD              mMutex        rw.repo.mMutex, rw.any.mMutex
+//     mUnavailableQ            mMutex        rw.repo.none [0], ro.main.none [5], rw.main.mLoadedMutex
+//     mLoadedQ                 mMutex        rw.repo.mLoadedMutex, ro.main.none [5], rw.main.mLoadedMutex
+//     mPendingLOD              mMutex        rw.repo.mPendingMutex, rw.any.mPendingMutex
 //     mGetMeshCapability       mMutex        rw.main.mMutex, ro.repo.mMutex (was:  [0])
 //     mGetMesh2Capability      mMutex        rw.main.mMutex, ro.repo.mMutex (was:  [0])
 //     mGetMeshVersion          mMutex        rw.main.mMutex, ro.repo.mMutex
@@ -343,20 +343,22 @@ static LLFastTimer::DeclareTimer FTM_MESH_FETCH("Mesh Fetch");
 
 LLMeshRepository gMeshRepo;
 
-const S32 MESH_HEADER_SIZE = 4096;                      // Important:  assumption is that headers fit in this space
+constexpr U32 CACHE_PREAMBLE_VERSION = 1;
+constexpr S32 CACHE_PREAMBLE_SIZE = sizeof(U32) * 3; //version, header_size, flags
+constexpr S32 MESH_HEADER_SIZE = 4096;                      // Important:  assumption is that headers fit in this space
 
 
-const S32 REQUEST2_HIGH_WATER_MIN = 32;                 // Limits for GetMesh2 regions
-const S32 REQUEST2_HIGH_WATER_MAX = 100;
-const S32 REQUEST2_LOW_WATER_MIN = 16;
-const S32 REQUEST2_LOW_WATER_MAX = 50;
+constexpr S32 REQUEST2_HIGH_WATER_MIN = 32;                 // Limits for GetMesh2 regions
+constexpr S32 REQUEST2_HIGH_WATER_MAX = 100;
+constexpr S32 REQUEST2_LOW_WATER_MIN = 16;
+constexpr S32 REQUEST2_LOW_WATER_MAX = 50;
 
-const U32 LARGE_MESH_FETCH_THRESHOLD = 1U << 21;        // Size at which requests goes to narrow/slow queue
-const long SMALL_MESH_XFER_TIMEOUT = 120L;              // Seconds to complete xfer, small mesh downloads
-const long LARGE_MESH_XFER_TIMEOUT = 600L;              // Seconds to complete xfer, large downloads
+constexpr U32 LARGE_MESH_FETCH_THRESHOLD = 1U << 21;        // Size at which requests goes to narrow/slow queue
+constexpr long SMALL_MESH_XFER_TIMEOUT = 120L;              // Seconds to complete xfer, small mesh downloads
+constexpr long LARGE_MESH_XFER_TIMEOUT = 600L;              // Seconds to complete xfer, large downloads
 
-const U32 DOWNLOAD_RETRY_LIMIT = 8;
-const F32 DOWNLOAD_RETRY_DELAY = 0.5f; // seconds
+constexpr U32 DOWNLOAD_RETRY_LIMIT = 8;
+constexpr F32 DOWNLOAD_RETRY_DELAY = 0.5f; // seconds
 
 // Would normally like to retry on uploads as some
 // retryable failures would be recoverable.  Unfortunately,
@@ -366,7 +368,7 @@ const F32 DOWNLOAD_RETRY_DELAY = 0.5f; // seconds
 // cap which then produces a 404 on retry destroying some
 // (occasionally) useful error information.  We'll leave
 // upload retries to the user as in the past.  SH-4667.
-const long UPLOAD_RETRY_LIMIT = 0L;
+constexpr long UPLOAD_RETRY_LIMIT = 0L;
 
 // Maximum mesh version to support.  Three least significant digits are reserved for the minor version,
 // with major version changes indicating a format change that is not backwards compatible and should not
@@ -374,7 +376,7 @@ const long UPLOAD_RETRY_LIMIT = 0L;
 // present, the version is 0.001. A viewer that can parse version 0.001 can also parse versions up to 0.999,
 // but not 1.0 (integer 1000).
 // See wiki at https://wiki.secondlife.com/wiki/Mesh/Mesh_Asset_Format
-const S32 MAX_MESH_VERSION = 999;
+constexpr S32 MAX_MESH_VERSION = 999;
 
 U32 LLMeshRepository::sBytesReceived = 0;
 U32 LLMeshRepository::sMeshRequestCount = 0;
@@ -386,12 +388,12 @@ U32 LLMeshRepository::sLODProcessing = 0;
 U32 LLMeshRepository::sLODPending = 0;
 
 U32 LLMeshRepository::sCacheBytesRead = 0;
-U32 LLMeshRepository::sCacheBytesWritten = 0;
+std::atomic<U32> LLMeshRepository::sCacheBytesWritten = 0;
 U32 LLMeshRepository::sCacheBytesHeaders = 0;
 U32 LLMeshRepository::sCacheBytesSkins = 0;
 U32 LLMeshRepository::sCacheBytesDecomps = 0;
 U32 LLMeshRepository::sCacheReads = 0;
-U32 LLMeshRepository::sCacheWrites = 0;
+std::atomic<U32> LLMeshRepository::sCacheWrites = 0;
 U32 LLMeshRepository::sMaxLockHoldoffs = 0;
 
 LLDeadmanTimer LLMeshRepository::sQuiescentTimer(15.0, false);  // true -> gather cpu metrics
@@ -550,6 +552,7 @@ LLViewerFetchedTexture* LLMeshUploadThread::FindViewerTexture(const LLImportMate
 
 std::atomic<S32> LLMeshRepoThread::sActiveHeaderRequests = 0;
 std::atomic<S32> LLMeshRepoThread::sActiveLODRequests = 0;
+std::atomic<S32> LLMeshRepoThread::sActiveSkinRequests = 0;
 U32 LLMeshRepoThread::sMaxConcurrentRequests = 1;
 S32 LLMeshRepoThread::sRequestLowWater = REQUEST2_LOW_WATER_MIN;
 S32 LLMeshRepoThread::sRequestHighWater = REQUEST2_HIGH_WATER_MIN;
@@ -584,6 +587,7 @@ public:
         : LLCore::HttpHandler(),
           mMeshParams(),
           mProcessed(false),
+          mHasDataOwnership(true),
           mHttpHandle(LLCORE_HTTP_HANDLE_INVALID),
           mOffset(offset),
           mRequestedBytes(requested_bytes)
@@ -607,6 +611,9 @@ public:
     LLCore::HttpHandle mHttpHandle;
     U32 mOffset;
     U32 mRequestedBytes;
+
+protected:
+    bool mHasDataOwnership = true;
 };
 
 
@@ -659,6 +666,9 @@ public:
     virtual void processData(LLCore::BufferArray * body, S32 body_offset, U8 * data, S32 data_size);
     virtual void processFailure(LLCore::HttpStatus status);
 
+private:
+    void processLod(U8* data, S32 data_size);
+
 public:
     S32 mLOD;
 };
@@ -674,13 +684,17 @@ public:
     LLMeshSkinInfoHandler(const LLUUID& id, U32 offset, U32 requested_bytes)
         : LLMeshHandlerBase(offset, requested_bytes),
           mMeshID(id)
-    {}
+    {
+        LLMeshRepoThread::incActiveSkinRequests();
+    }
     virtual ~LLMeshSkinInfoHandler();
 
 protected:
     LLMeshSkinInfoHandler(const LLMeshSkinInfoHandler &);       // Not defined
     void operator=(const LLMeshSkinInfoHandler &);              // Not defined
 
+    void processSkin(U8* data, S32 data_size);
+
 public:
     virtual void processData(LLCore::BufferArray * body, S32 body_offset, U8 * data, S32 data_size);
     virtual void processFailure(LLCore::HttpStatus status);
@@ -823,6 +837,14 @@ void log_upload_error(LLCore::HttpStatus status, const LLSD& content,
     gMeshRepo.uploadError(args);
 }
 
+void write_preamble(LLFileSystem &file, S32 header_bytes, S32 flags)
+{
+    LLMeshRepository::sCacheBytesWritten += CACHE_PREAMBLE_SIZE;
+    file.write((U8*)&CACHE_PREAMBLE_VERSION, sizeof(U32));
+    file.write((U8*)&header_bytes, sizeof(U32));
+    file.write((U8*)&flags, sizeof(U32));
+}
+
 LLMeshRepoThread::LLMeshRepoThread()
 : LLThread("mesh repo"),
   mHttpRequest(NULL),
@@ -837,6 +859,9 @@ LLMeshRepoThread::LLMeshRepoThread()
 
     mMutex = new LLMutex();
     mHeaderMutex = new LLMutex();
+    mLoadedMutex = new LLMutex();
+    mPendingMutex = new LLMutex();
+    mSkinMapMutex = new LLMutex();
     mSignal = new LLCondition();
     mHttpRequest = new LLCore::HttpRequest;
     mHttpOptions = LLCore::HttpOptions::ptr_t(new LLCore::HttpOptions);
@@ -849,6 +874,11 @@ LLMeshRepoThread::LLMeshRepoThread()
     mHttpHeaders->append(HTTP_OUT_HEADER_ACCEPT, HTTP_CONTENT_VND_LL_MESH);
     mHttpPolicyClass = app_core_http.getPolicy(LLAppCoreHttp::AP_MESH2);
     mHttpLargePolicyClass = app_core_http.getPolicy(LLAppCoreHttp::AP_LARGE_MESH);
+
+    // Lod processing is expensive due to the number of requests
+    // and a need to do expensive cacheOptimize().
+    mMeshThreadPool.reset(new LL::ThreadPool("MeshLodProcessing", 2));
+    mMeshThreadPool->start();
 }
 
 
@@ -875,13 +905,21 @@ LLMeshRepoThread::~LLMeshRepoThread()
     }
 
     delete mHttpRequest;
-    mHttpRequest = NULL;
+    mHttpRequest = nullptr;
     delete mMutex;
-    mMutex = NULL;
+    mMutex = nullptr;
     delete mHeaderMutex;
-    mHeaderMutex = NULL;
+    mHeaderMutex = nullptr;
+    delete mLoadedMutex;
+    mLoadedMutex = nullptr;
+    delete mPendingMutex;
+    mPendingMutex = nullptr;
+    delete mSkinMapMutex;
+    mSkinMapMutex = nullptr;
     delete mSignal;
-    mSignal = NULL;
+    mSignal = nullptr;
+    delete[] mDiskCacheBuffer;
+    mDiskCacheBuffer = nullptr;
 }
 
 void LLMeshRepoThread::run()
@@ -908,6 +946,7 @@ void LLMeshRepoThread::run()
         // On the other hand, this may actually be an effective and efficient scheme...
 
         mSignal->wait();
+        LL_PROFILE_ZONE_NAMED("mesh_thread_loop")
 
         if (LLApp::isExiting())
         {
@@ -925,11 +964,55 @@ void LLMeshRepoThread::run()
         }
         sRequestWaterLevel = static_cast<S32>(mHttpRequestSet.size());            // Stats data update
 
-        // NOTE: order of queue processing intentionally favors LOD requests over header requests
+        // NOTE: order of queue processing intentionally favors LOD and Skin requests over header requests
         // Todo: we are processing mLODReqQ, mHeaderReqQ, mSkinRequests, mDecompositionRequests and mPhysicsShapeRequests
         // in relatively similar manners, remake code to simplify/unify the process,
         // like processRequests(&requestQ, fetchFunction); which does same thing for each element
 
+        if (mHttpRequestSet.size() < sRequestHighWater
+            && !mSkinRequests.empty())
+        {
+            if (!mSkinRequests.empty())
+            {
+                std::list<UUIDBasedRequest> incomplete;
+                while (!mSkinRequests.empty() && mHttpRequestSet.size() < sRequestHighWater)
+                {
+
+                    mMutex->lock();
+                    auto req = mSkinRequests.front();
+                    mSkinRequests.pop_front();
+                    mMutex->unlock();
+                    if (req.isDelayed())
+                    {
+                        incomplete.emplace_back(req);
+                    }
+                    else if (!fetchMeshSkinInfo(req.mId))
+                    {
+                        if (req.canRetry())
+                        {
+                            req.updateTime();
+                            incomplete.emplace_back(req);
+                        }
+                        else
+                        {
+                            LLMutexLock locker(mLoadedMutex);
+                            mSkinUnavailableQ.push_back(req);
+                            LL_DEBUGS() << "mSkinReqQ failed: " << req.mId << LL_ENDL;
+                        }
+                    }
+                }
+
+                if (!incomplete.empty())
+                {
+                    LLMutexLock locker(mMutex);
+                    for (const auto& req : incomplete)
+                    {
+                        mSkinRequests.push_back(req);
+                    }
+                }
+            }
+        }
+
         if (!mLODReqQ.empty() && mHttpRequestSet.size() < sRequestHighWater)
         {
             std::list<LODRequest> incomplete;
@@ -950,7 +1033,7 @@ void LLMeshRepoThread::run()
                     // failed to load before, wait a bit
                     incomplete.push_front(req);
                 }
-                else if (!fetchMeshLOD(req.mMeshParams, req.mLOD, req.canRetry()))
+                else if (!fetchMeshLOD(req.mMeshParams, req.mLOD))
                 {
                     if (req.canRetry())
                     {
@@ -961,7 +1044,7 @@ void LLMeshRepoThread::run()
                     else
                     {
                         // too many fails
-                        LLMutexLock lock(mMutex);
+                        LLMutexLock lock(mLoadedMutex);
                         mUnavailableQ.push_back(req);
                         LL_WARNS() << "Failed to load " << req.mMeshParams << " , skip" << LL_ENDL;
                     }
@@ -998,7 +1081,7 @@ void LLMeshRepoThread::run()
                     // failed to load before, wait a bit
                     incomplete.push_front(req);
                 }
-                else if (!fetchMeshHeader(req.mMeshParams, req.canRetry()))
+                else if (!fetchMeshHeader(req.mMeshParams))
                 {
                     if (req.canRetry())
                     {
@@ -1028,54 +1111,13 @@ void LLMeshRepoThread::run()
         // performing long-duration actions.
 
         if (mHttpRequestSet.size() < sRequestHighWater
-            && (!mSkinRequests.empty()
-            || !mDecompositionRequests.empty()
+            && (!mDecompositionRequests.empty()
             || !mPhysicsShapeRequests.empty()))
         {
             // Something to do probably, lock and double-check.  We don't want
             // to hold the lock long here.  That will stall main thread activities
             // so we bounce it.
 
-            if (!mSkinRequests.empty())
-            {
-                std::list<UUIDBasedRequest> incomplete;
-                while (!mSkinRequests.empty() && mHttpRequestSet.size() < sRequestHighWater)
-                {
-
-                    mMutex->lock();
-                    auto req = mSkinRequests.front();
-                    mSkinRequests.pop_front();
-                    mMutex->unlock();
-                    if (req.isDelayed())
-                    {
-                        incomplete.emplace_back(req);
-                    }
-                    else if (!fetchMeshSkinInfo(req.mId, req.canRetry()))
-                    {
-                        if (req.canRetry())
-                        {
-                            req.updateTime();
-                            incomplete.emplace_back(req);
-                        }
-                        else
-                        {
-                            LLMutexLock locker(mMutex);
-                            mSkinUnavailableQ.push_back(req);
-                            LL_DEBUGS() << "mSkinReqQ failed: " << req.mId << LL_ENDL;
-                        }
-                    }
-                }
-
-                if (!incomplete.empty())
-                {
-                    LLMutexLock locker(mMutex);
-                    for (const auto& req : incomplete)
-                    {
-                        mSkinRequests.push_back(req);
-                    }
-                }
-            }
-
             // holding lock, try next list
             // *TODO:  For UI/debug-oriented lists, we might drop the fine-
             // grained locking as there's a lowered expectation of smoothness
@@ -1103,7 +1145,7 @@ void LLMeshRepoThread::run()
                         }
                         else
                         {
-                            LL_DEBUGS() << "mDecompositionRequests failed: " << req.mId << LL_ENDL;
+                            LL_DEBUGS(LOG_MESH) << "mDecompositionRequests failed: " << req.mId << LL_ENDL;
                         }
                     }
                 }
@@ -1139,7 +1181,7 @@ void LLMeshRepoThread::run()
                         }
                         else
                         {
-                            LL_DEBUGS() << "mPhysicsShapeRequests failed: " << req.mId << LL_ENDL;
+                            LL_DEBUGS(LOG_MESH) << "mPhysicsShapeRequests failed: " << req.mId << LL_ENDL;
                         }
                     }
                 }
@@ -1195,40 +1237,91 @@ void LLMeshRepoThread::lockAndLoadMeshLOD(const LLVolumeParams& mesh_params, S32
     }
 }
 
-
 void LLMeshRepoThread::loadMeshLOD(const LLVolumeParams& mesh_params, S32 lod)
 { //could be called from any thread
     const LLUUID& mesh_id = mesh_params.getSculptID();
-    LLMutexLock lock(mMutex);
-    LLMutexLock header_lock(mHeaderMutex);
-    mesh_header_map::iterator iter = mMeshHeader.find(mesh_id);
-    if (iter != mMeshHeader.end())
+    loadMeshLOD(mesh_id, mesh_params, lod);
+}
+
+void LLMeshRepoThread::loadMeshLOD(const LLUUID& mesh_id, const LLVolumeParams& mesh_params, S32 lod)
+{
+    if (hasHeader(mesh_id))
     { //if we have the header, request LOD byte range
 
         LODRequest req(mesh_params, lod);
         {
+            LLMutexLock lock(mMutex);
             mLODReqQ.push(req);
             LLMeshRepository::sLODProcessing++;
         }
     }
     else
     {
+        LLMutexLock lock(mPendingMutex);
         HeaderRequest req(mesh_params);
         pending_lod_map::iterator pending = mPendingLOD.find(mesh_id);
 
         if (pending != mPendingLOD.end())
-        { //append this lod request to existing header request
-            pending->second.push_back(lod);
-            llassert(pending->second.size() <= LLModel::NUM_LODS);
+        {
+            //append this lod request to existing header request
+            if (lod < LLModel::NUM_LODS && lod >= 0)
+            {
+                pending->second[lod]++;
+            }
+            else
+            {
+                LL_WARNS(LOG_MESH) << "Invalid LOD request: " << lod << "for mesh" << mesh_id << LL_ENDL;
+            }
+            llassert_msg(lod < LLModel::NUM_LODS, "Requested lod is out of bounds");
         }
         else
-        { //if no header request is pending, fetch header
+        {
+            //if no header request is pending, fetch header
+            auto& array = mPendingLOD[mesh_id];
+            std::fill(array.begin(), array.end(), 0);
+            array[lod]++;
+
+            LLMutexLock lock(mMutex);
             mHeaderReqQ.push(req);
-            mPendingLOD[mesh_id].push_back(lod);
         }
     }
 }
 
+U8* LLMeshRepoThread::getDiskCacheBuffer(S32 size)
+{
+    if (mDiskCacheBufferSize < size)
+    {
+        const S32 MINIMUM_BUFFER_SIZE = 8192; // a minimum to avoid frequent early reallocations
+        size = llmax(MINIMUM_BUFFER_SIZE, size);
+        delete[] mDiskCacheBuffer;
+        try
+        {
+            mDiskCacheBuffer = new U8[size];
+        }
+        catch (std::bad_alloc&)
+        {
+            LL_WARNS(LOG_MESH) << "Failed to allocate memory for mesh thread's buffer, size: " << size << LL_ENDL;
+            mDiskCacheBuffer = NULL;
+
+            // Not sure what size is reasonable
+            // but if 30MB allocation failed, we definitely have issues
+            const S32 MAX_SIZE = 30 * 1024 * 1024; //30MB
+            if (size < MAX_SIZE)
+            {
+                LLAppViewer::instance()->outOfMemorySoftQuit();
+            } // else ignore failures for anomalously large data
+        }
+        mDiskCacheBufferSize = size;
+    }
+    else
+    {
+        // reusing old buffer, reset heading bytes to ensure
+        // old content won't be parsable if something fails.
+        memset(mDiskCacheBuffer, 0, 16);
+    }
+    return mDiskCacheBuffer;
+}
+
 // Mutex:  must be holding mMutex when called
 void LLMeshRepoThread::setGetMeshCap(const std::string & mesh_cap)
 {
@@ -1327,7 +1420,7 @@ LLCore::HttpHandle LLMeshRepoThread::getByteRange(const std::string & url,
 }
 
 
-bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
+bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id)
 {
     LL_PROFILE_ZONE_SCOPED;
     if (!mHeaderMutex)
@@ -1337,7 +1430,7 @@ bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
 
     mHeaderMutex->lock();
 
-    auto header_it = mMeshHeader.find(mesh_id);
+    mesh_header_map::const_iterator header_it = mMeshHeader.find(mesh_id);
     if (header_it == mMeshHeader.end())
     { //we have no header info for this mesh, do nothing
         mHeaderMutex->unlock();
@@ -1346,23 +1439,24 @@ bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
 
     ++LLMeshRepository::sMeshRequestCount;
     bool ret = true;
-    U32 header_size = header_it->second.first;
+    const LLMeshHeader& header = header_it->second;
+    U32 header_size = header.mHeaderSize;
 
     if (header_size > 0)
     {
-        const LLMeshHeader& header = header_it->second.second;
-
         S32 version = header.mVersion;
         S32 offset = header_size + header.mSkinOffset;
         S32 size = header.mSkinSize;
+        bool in_cache = header.mSkinInCache;
 
         mHeaderMutex->unlock();
 
         if (version <= MAX_MESH_VERSION && offset >= 0 && size > 0)
         {
             //check cache for mesh skin info
+            S32 disk_ofset = offset + CACHE_PREAMBLE_SIZE;
             LLFileSystem file(mesh_id, LLAssetType::AT_MESH);
-            if (file.getSize() >= offset + size)
+            if (in_cache && file.getSize() >= disk_ofset + size)
             {
                 U8* buffer = new(std::nothrow) U8[size];
                 if (!buffer)
@@ -1370,19 +1464,19 @@ bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
                     LL_WARNS(LOG_MESH) << "Failed to allocate memory for skin info, size: " << size << LL_ENDL;
 
                     // Not sure what size is reasonable for skin info,
-                    // but if 20MB allocation failed, we definetely have issues
+                    // but if 30MB allocation failed, we definitely have issues
                     const S32 MAX_SIZE = 30 * 1024 * 1024; //30MB
                     if (size < MAX_SIZE)
                     {
                         LLAppViewer::instance()->outOfMemorySoftQuit();
                     } // else ignore failures for anomalously large data
-                    LLMutexLock locker(mMutex);
+                    LLMutexLock locker(mLoadedMutex);
                     mSkinUnavailableQ.emplace_back(mesh_id);
                     return true;
                 }
                 LLMeshRepository::sCacheBytesRead += size;
                 ++LLMeshRepository::sCacheReads;
-                file.seek(offset);
+                file.seek(disk_ofset);
                 file.read(buffer, size);
 
                 //make sure buffer isn't all 0's by checking the first 1KB (reserved block but not written)
@@ -1393,14 +1487,67 @@ bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
                 }
 
                 if (!zero)
-                { //attempt to parse
-                    if (skinInfoReceived(mesh_id, buffer, size))
+                {
+                    //attempt to parse
+                    bool posted = mMeshThreadPool->getQueue().post(
+                        [mesh_id, buffer, size]
+                        ()
+                    {
+                        if (!gMeshRepo.mThread->skinInfoReceived(mesh_id, buffer, size))
+                        {
+                            // either header is faulty or something else overwrote the cache
+                            S32 header_size = 0;
+                            U32 header_flags = 0;
+                            {
+                                LL_DEBUGS(LOG_MESH) << "Mesh header for ID " << mesh_id << " cache mismatch." << LL_ENDL;
+
+                                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                                auto header_it = gMeshRepo.mThread->mMeshHeader.find(mesh_id);
+                                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                                {
+                                    LLMeshHeader& header = header_it->second;
+                                    // for safety just mark everything as missing
+                                    header.mSkinInCache = false;
+                                    header.mPhysicsConvexInCache = false;
+                                    header.mPhysicsMeshInCache = false;
+                                    for (S32 i = 0; i < LLModel::NUM_LODS; ++i)
+                                    {
+                                        header.mLodInCache[i] = false;
+                                    }
+                                    header_size = header.mHeaderSize;
+                                    header_flags = header.getFlags();
+                                }
+                            }
+
+                            if (header_size > 0)
+                            {
+                                LLFileSystem file(mesh_id, LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
+                                if (file.getMaxSize() >= CACHE_PREAMBLE_SIZE)
+                                {
+                                    write_preamble(file, header_size, header_flags);
+                                }
+                            }
+
+                            {
+                                LLMutexLock lock(gMeshRepo.mThread->mMutex);
+                                UUIDBasedRequest req(mesh_id);
+                                gMeshRepo.mThread->mSkinRequests.push_back(req);
+                            }
+                        }
+                        delete[] buffer;
+                    });
+                    if (posted)
+                    {
+                        // lambda owns buffer
+                        return true;
+                    }
+                    else if (skinInfoReceived(mesh_id, buffer, size))
                     {
                         delete[] buffer;
                         return true;
                     }
                 }
-
                 delete[] buffer;
             }
 
@@ -1420,26 +1567,21 @@ bool LLMeshRepoThread::fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry)
                                        << LL_ENDL;
                     ret = false;
                 }
-                else if(can_retry)
+                else
                 {
                     handler->mHttpHandle = handle;
                     mHttpRequestSet.insert(handler);
                 }
-                else
-                {
-                    LLMutexLock locker(mMutex);
-                    mSkinUnavailableQ.emplace_back(mesh_id);
-                }
             }
             else
             {
-                LLMutexLock locker(mMutex);
+                LLMutexLock locker(mLoadedMutex);
                 mSkinUnavailableQ.emplace_back(mesh_id);
             }
         }
         else
         {
-            LLMutexLock locker(mMutex);
+            LLMutexLock locker(mLoadedMutex);
             mSkinUnavailableQ.emplace_back(mesh_id);
         }
     }
@@ -1470,42 +1612,35 @@ bool LLMeshRepoThread::fetchMeshDecomposition(const LLUUID& mesh_id)
     }
 
     ++LLMeshRepository::sMeshRequestCount;
-    U32 header_size = header_it->second.first;
+    const auto& header = header_it->second;
+    U32 header_size = header.mHeaderSize;
     bool ret = true;
 
     if (header_size > 0)
     {
-        const auto& header = header_it->second.second;
         S32 version = header.mVersion;
         S32 offset = header_size + header.mPhysicsConvexOffset;
         S32 size = header.mPhysicsConvexSize;
+        bool in_cache = header.mPhysicsConvexInCache;
 
         mHeaderMutex->unlock();
 
         if (version <= MAX_MESH_VERSION && offset >= 0 && size > 0)
         {
-            //check cache for mesh skin info
+            // check cache for mesh decomposition
+            S32 disk_ofset = offset + CACHE_PREAMBLE_SIZE;
             LLFileSystem file(mesh_id, LLAssetType::AT_MESH);
-            if (file.getSize() >= offset+size)
+            if (in_cache && file.getSize() >= disk_ofset + size)
             {
-                U8* buffer = new(std::nothrow) U8[size];
+                U8* buffer = getDiskCacheBuffer(size);
                 if (!buffer)
                 {
-                    LL_WARNS(LOG_MESH) << "Failed to allocate memory for mesh decomposition, size: " << size << LL_ENDL;
-
-                    // Not sure what size is reasonable for decomposition
-                    // but if 20MB allocation failed, we definetely have issues
-                    const S32 MAX_SIZE = 30 * 1024 * 1024; //30MB
-                    if (size < MAX_SIZE)
-                    {
-                        LLAppViewer::instance()->outOfMemorySoftQuit();
-                    } // else ignore failures for anomalously large decompositiions
                     return true;
                 }
                 LLMeshRepository::sCacheBytesRead += size;
                 ++LLMeshRepository::sCacheReads;
 
-                file.seek(offset);
+                file.seek(disk_ofset);
                 file.read(buffer, size);
 
                 //make sure buffer isn't all 0's by checking the first 1KB (reserved block but not written)
@@ -1519,12 +1654,9 @@ bool LLMeshRepoThread::fetchMeshDecomposition(const LLUUID& mesh_id)
                 { //attempt to parse
                     if (decompositionReceived(mesh_id, buffer, size))
                     {
-                        delete[] buffer;
                         return true;
                     }
                 }
-
-                delete[] buffer;
             }
 
             //reading from cache failed for whatever reason, fetch from sim
@@ -1578,41 +1710,36 @@ bool LLMeshRepoThread::fetchMeshPhysicsShape(const LLUUID& mesh_id)
     }
 
     ++LLMeshRepository::sMeshRequestCount;
-    U32 header_size = header_it->second.first;
+    const auto& header = header_it->second;
+    U32 header_size = header.mHeaderSize;
     bool ret = true;
 
     if (header_size > 0)
     {
-        const auto& header = header_it->second.second;
         S32 version = header.mVersion;
         S32 offset = header_size + header.mPhysicsMeshOffset;
         S32 size = header.mPhysicsMeshSize;
+        bool in_cache = header.mPhysicsMeshInCache;
 
         mHeaderMutex->unlock();
 
+        // todo: check header.mHasPhysicsMesh
         if (version <= MAX_MESH_VERSION && offset >= 0 && size > 0)
         {
             //check cache for mesh physics shape info
+            S32 disk_ofset = offset + CACHE_PREAMBLE_SIZE;
             LLFileSystem file(mesh_id, LLAssetType::AT_MESH);
-            if (file.getSize() >= offset+size)
+            if (in_cache && file.getSize() >= disk_ofset +size)
             {
                 LLMeshRepository::sCacheBytesRead += size;
                 ++LLMeshRepository::sCacheReads;
-                file.seek(offset);
-                U8* buffer = new(std::nothrow) U8[size];
+
+                U8* buffer = getDiskCacheBuffer(size);
                 if (!buffer)
                 {
-                    LL_WARNS(LOG_MESH) << "Failed to allocate memory for mesh decomposition, size: " << size << LL_ENDL;
-
-                    // Not sure what size is reasonable for physcis
-                    // but if 20MB allocation failed, we definetely have issues
-                    const S32 MAX_SIZE = 30 * 1024 * 1024; //30MB
-                    if (size < MAX_SIZE)
-                    {
-                        LLAppViewer::instance()->outOfMemorySoftQuit();
-                    } // else ignore failures for anomalously large data
                     return true;
                 }
+                file.seek(disk_ofset);
                 file.read(buffer, size);
 
                 //make sure buffer isn't all 0's by checking the first 1KB (reserved block but not written)
@@ -1626,12 +1753,9 @@ bool LLMeshRepoThread::fetchMeshPhysicsShape(const LLUUID& mesh_id)
                 { //attempt to parse
                     if (physicsShapeReceived(mesh_id, buffer, size) == MESH_OK)
                     {
-                        delete[] buffer;
                         return true;
                     }
                 }
-
-                delete[] buffer;
             }
 
             //reading from cache failed for whatever reason, fetch from sim
@@ -1699,8 +1823,22 @@ void LLMeshRepoThread::decActiveHeaderRequests()
     --LLMeshRepoThread::sActiveHeaderRequests;
 }
 
+//static
+void LLMeshRepoThread::incActiveSkinRequests()
+{
+    LLMutexLock lock(gMeshRepo.mThread->mMutex);
+    ++LLMeshRepoThread::sActiveSkinRequests;
+}
+
+//static
+void LLMeshRepoThread::decActiveSkinRequests()
+{
+    LLMutexLock lock(gMeshRepo.mThread->mMutex);
+    --LLMeshRepoThread::sActiveSkinRequests;
+}
+
 //return false if failed to get header
-bool LLMeshRepoThread::fetchMeshHeader(const LLVolumeParams& mesh_params, bool can_retry)
+bool LLMeshRepoThread::fetchMeshHeader(const LLVolumeParams& mesh_params)
 {
     LL_PROFILE_ZONE_SCOPED;
     ++LLMeshRepository::sMeshRequestCount;
@@ -1714,17 +1852,34 @@ bool LLMeshRepoThread::fetchMeshHeader(const LLVolumeParams& mesh_params, bool c
         if (size > 0)
         {
             // *NOTE:  if the header size is ever more than 4KB, this will break
-            U8 buffer[MESH_HEADER_SIZE];
-            S32 bytes = llmin(size, MESH_HEADER_SIZE);
+            constexpr S32 DISK_MINIMAL_READ = 4096;
+            U8 buffer[DISK_MINIMAL_READ * 2];
+            S32 bytes = llmin(size, DISK_MINIMAL_READ);
             LLMeshRepository::sCacheBytesRead += bytes;
             ++LLMeshRepository::sCacheReads;
+
             file.read(buffer, bytes);
-            if (headerReceived(mesh_params, buffer, bytes) == MESH_OK)
+
+            U32 version = 0;
+            memcpy(&version, buffer, sizeof(U32));
+            if (version == CACHE_PREAMBLE_VERSION)
             {
-                LL_DEBUGS(LOG_MESH) << "Mesh/Cache: Mesh header for ID " << mesh_params.getSculptID() << " - was retrieved from the cache." << LL_ENDL;
+                S32 header_size = 0;
+                memcpy(&header_size, buffer + sizeof(U32), sizeof(S32));
+                if (header_size + CACHE_PREAMBLE_SIZE > DISK_MINIMAL_READ)
+                {
+                    bytes = llmin(size , DISK_MINIMAL_READ * 2);
+                    file.read(buffer + DISK_MINIMAL_READ, bytes - DISK_MINIMAL_READ);
+                }
+                U32 flags = 0;
+                memcpy(&flags, buffer + 2 * sizeof(U32), sizeof(U32));
+                if (headerReceived(mesh_params, buffer + CACHE_PREAMBLE_SIZE, bytes - CACHE_PREAMBLE_SIZE, flags) == MESH_OK)
+                {
+                    LL_DEBUGS(LOG_MESH) << "Mesh/Cache: Mesh header for ID " << mesh_params.getSculptID() << " - was retrieved from the cache." << LL_ENDL;
 
-                // Found mesh in cache
-                return true;
+                    // Found mesh in cache
+                    return true;
+                }
             }
         }
     }
@@ -1753,7 +1908,7 @@ bool LLMeshRepoThread::fetchMeshHeader(const LLVolumeParams& mesh_params, bool c
                                << LL_ENDL;
             retval = false;
         }
-        else if (can_retry)
+        else
         {
             handler->mHttpHandle = handle;
             mHttpRequestSet.insert(handler);
@@ -1764,7 +1919,7 @@ bool LLMeshRepoThread::fetchMeshHeader(const LLVolumeParams& mesh_params, bool c
 }
 
 //return false if failed to get mesh lod.
-bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod, bool can_retry)
+bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod)
 {
     LL_PROFILE_ZONE_SCOPED;
     if (!mHeaderMutex)
@@ -1784,41 +1939,43 @@ bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod,
     ++LLMeshRepository::sMeshRequestCount;
     bool retval = true;
 
-    U32 header_size = header_it->second.first;
+    const auto& header = header_it->second;
+    U32 header_size = header.mHeaderSize;
     if (header_size > 0)
     {
-        const auto& header = header_it->second.second;
         S32 version = header.mVersion;
         S32 offset = header_size + header.mLodOffset[lod];
         S32 size = header.mLodSize[lod];
+        bool in_cache = header.mLodInCache[lod];
         mHeaderMutex->unlock();
 
         if (version <= MAX_MESH_VERSION && offset >= 0 && size > 0)
         {
-
+            S32 disk_ofset = offset + CACHE_PREAMBLE_SIZE;
             //check cache for mesh asset
             LLFileSystem file(mesh_id, LLAssetType::AT_MESH);
-            if (file.getSize() >= offset+size)
+            if (in_cache && (file.getSize() >= disk_ofset + size))
             {
-                U8* buffer = new(std::nothrow) U8[size];
+                U8* buffer = new(std::nothrow) U8[size]; // todo, make buffer thread local and read in thread?
                 if (!buffer)
                 {
                     LL_WARNS(LOG_MESH) << "Can't allocate memory for mesh " << mesh_id << " LOD " << lod << ", size: " << size << LL_ENDL;
 
                     // Not sure what size is reasonable for a mesh,
-                    // but if 20MB allocation failed, we definetely have issues
+                    // but if 30MB allocation failed, we definitely have issues
                     const S32 MAX_SIZE = 30 * 1024 * 1024; //30MB
                     if (size < MAX_SIZE)
                     {
                         LLAppViewer::instance()->outOfMemorySoftQuit();
                     } // else ignore failures for anomalously large data
-                    LLMutexLock lock(mMutex);
+
+                    LLMutexLock lock(mLoadedMutex);
                     mUnavailableQ.push_back(LODRequest(mesh_params, lod));
                     return true;
                 }
                 LLMeshRepository::sCacheBytesRead += size;
                 ++LLMeshRepository::sCacheReads;
-                file.seek(offset);
+                file.seek(disk_ofset);
                 file.read(buffer, size);
 
                 //make sure buffer isn't all 0's by checking the first 1KB (reserved block but not written)
@@ -1829,15 +1986,76 @@ bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod,
                 }
 
                 if (!zero)
-                { //attempt to parse
-                    if (lodReceived(mesh_params, lod, buffer, size) == MESH_OK)
+                {
+                    //attempt to parse
+                    const LLVolumeParams params(mesh_params);
+                    bool posted = mMeshThreadPool->getQueue().post(
+                        [params, mesh_id, lod, buffer, size]
+                        ()
                     {
+                        if (gMeshRepo.mThread->lodReceived(params, lod, buffer, size) == MESH_OK)
+                        {
+                            LL_DEBUGS(LOG_MESH) << "Mesh/Cache: Mesh body for ID " << mesh_id << " - was retrieved from the cache." << LL_ENDL;
+                        }
+                        else
+                        {
+                            // either header is faulty or something else overwrote the cache
+                            S32 header_size = 0;
+                            U32 header_flags = 0;
+                            {
+                                LL_DEBUGS(LOG_MESH) << "Mesh header for ID " << mesh_id << " cache mismatch." << LL_ENDL;
+
+                                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                                auto header_it = gMeshRepo.mThread->mMeshHeader.find(mesh_id);
+                                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                                {
+                                    LLMeshHeader& header = header_it->second;
+                                    // for safety just mark everything as missing
+                                    header.mSkinInCache = false;
+                                    header.mPhysicsConvexInCache = false;
+                                    header.mPhysicsMeshInCache = false;
+                                    for (S32 i = 0; i < LLModel::NUM_LODS; ++i)
+                                    {
+                                        header.mLodInCache[i] = false;
+                                    }
+                                    header_size = header.mHeaderSize;
+                                    header_flags = header.getFlags();
+                                }
+                            }
+
+                            if (header_size > 0)
+                            {
+                                LLFileSystem file(mesh_id, LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
+                                if (file.getMaxSize() >= CACHE_PREAMBLE_SIZE)
+                                {
+                                    write_preamble(file, header_size, header_flags);
+                                }
+                            }
+
+                            {
+                                LLMutexLock lock(gMeshRepo.mThread->mMutex);
+                                LODRequest req(params, lod);
+                                gMeshRepo.mThread->mLODReqQ.push(req);
+                                LLMeshRepository::sLODProcessing++;
+                            }
+                        }
                         delete[] buffer;
+                    });
 
+                    if (posted)
+                    {
+                        // now lambda owns buffer
+                        return true;
+                    }
+                    else if (lodReceived(mesh_params, lod, buffer, size) == MESH_OK)
+                    {
+                        delete[] buffer;
                         LL_DEBUGS(LOG_MESH) << "Mesh/Cache: Mesh body for ID " << mesh_id << " - was retrieved from the cache." << LL_ENDL;
 
                         return true;
                     }
+
                 }
 
                 delete[] buffer;
@@ -1861,27 +2079,22 @@ bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod,
                                        << LL_ENDL;
                     retval = false;
                 }
-                else if (can_retry)
+                else
                 {
+                    // we already made a request, store the handle
                     handler->mHttpHandle = handle;
                     mHttpRequestSet.insert(handler);
-                    // *NOTE:  Allowing a re-request, not marking as unavailable.  Is that correct?
-                }
-                else
-                {
-                    LLMutexLock lock(mMutex);
-                    mUnavailableQ.push_back(LODRequest(mesh_params, lod));
                 }
             }
             else
             {
-                LLMutexLock lock(mMutex);
+                LLMutexLock lock(mLoadedMutex);
                 mUnavailableQ.push_back(LODRequest(mesh_params, lod));
             }
         }
         else
         {
-            LLMutexLock lock(mMutex);
+            LLMutexLock lock(mLoadedMutex);
             mUnavailableQ.push_back(LODRequest(mesh_params, lod));
         }
     }
@@ -1893,14 +2106,19 @@ bool LLMeshRepoThread::fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod,
     return retval;
 }
 
-EMeshProcessingResult LLMeshRepoThread::headerReceived(const LLVolumeParams& mesh_params, U8* data, S32 data_size)
+EMeshProcessingResult LLMeshRepoThread::headerReceived(const LLVolumeParams& mesh_params, U8* data, S32 data_size, U32 flags)
 {
+    LL_PROFILE_ZONE_SCOPED;
     const LLUUID mesh_id = mesh_params.getSculptID();
     LLSD header_data;
 
     LLMeshHeader header;
 
     llssize header_size = 0;
+    S32 skin_offset = -1;
+    S32 skin_size = -1;
+    S32 lod_offset[LLModel::NUM_LODS] = { -1 };
+    S32 lod_size[LLModel::NUM_LODS] = { -1 };
     if (data_size > 0)
     {
         llssize dsize = data_size;
@@ -1933,7 +2151,40 @@ EMeshProcessingResult LLMeshRepoThread::headerReceived(const LLVolumeParams& mes
         // make sure there is at least one lod, function returns -1 and marks as 404 otherwise
         else if (LLMeshRepository::getActualMeshLOD(header, 0) >= 0)
         {
-            header_size += stream.tellg();
+            header.mHeaderSize = (S32)stream.tellg();
+            header_size += header.mHeaderSize;
+            skin_offset = header.mSkinOffset;
+            skin_size = header.mSkinSize;
+
+            memcpy(lod_offset, header.mLodOffset, sizeof(lod_offset));
+            memcpy(lod_size, header.mLodSize, sizeof(lod_size));
+
+            if (flags != 0)
+            {
+                header.setFromFlags(flags);
+            }
+            else
+            {
+                if (header.mSkinSize > 0 && header_size + header.mSkinOffset + header.mSkinSize < data_size)
+                {
+                    header.mSkinInCache = true;
+                }
+                if (header.mPhysicsConvexSize > 0 && header_size + header.mPhysicsConvexOffset + header.mPhysicsConvexSize < data_size)
+                {
+                    header.mPhysicsConvexInCache = true;
+                }
+                if (header.mPhysicsMeshSize > 0 && header_size + header.mPhysicsMeshOffset + header.mPhysicsMeshSize < data_size)
+                {
+                    header.mPhysicsMeshInCache = true;
+                }
+                for (S32 i = 0; i < LLModel::NUM_LODS; ++i)
+                {
+                    if (lod_size[i] > 0 && header_size + lod_offset[i] + lod_size[i] < data_size)
+                    {
+                        header.mLodInCache[i] = true;
+                    }
+                }
+            }
         }
     }
     else
@@ -1947,35 +2198,84 @@ EMeshProcessingResult LLMeshRepoThread::headerReceived(const LLVolumeParams& mes
 
         {
             LLMutexLock lock(mHeaderMutex);
-            mMeshHeader[mesh_id] = { (U32)header_size, header };
+            mMeshHeader[mesh_id] = header;
             LLMeshRepository::sCacheBytesHeaders += (U32)header_size;
         }
 
         // immediately request SkinInfo since we'll need it before we can render any LoD if it is present
+        if (skin_offset >= 0 && skin_size > 0)
         {
-            LLMutexLock lock(gMeshRepo.mMeshMutex);
+            {
+                LLMutexLock lock(gMeshRepo.mMeshMutex);
 
-            if (gMeshRepo.mLoadingSkins.find(mesh_id) == gMeshRepo.mLoadingSkins.end())
+                if (gMeshRepo.mLoadingSkins.find(mesh_id) == gMeshRepo.mLoadingSkins.end())
+                {
+                    gMeshRepo.mLoadingSkins[mesh_id] = {}; // add an empty vector to indicate to main thread that we are loading skin info
+                }
+            }
+
+            S32 offset = (S32)header_size + skin_offset;
+            bool request_skin = true;
+            if (offset + skin_size < data_size)
+            {
+                request_skin = !skinInfoReceived(mesh_id, data + offset, skin_size);
+            }
+            if (request_skin)
             {
-                gMeshRepo.mLoadingSkins[mesh_id] = {}; // add an empty vector to indicate to main thread that we are loading skin info
+                mSkinRequests.push_back(UUIDBasedRequest(mesh_id));
             }
         }
 
-        fetchMeshSkinInfo(mesh_id);
-
-        LLMutexLock lock(mMutex); // make sure only one thread access mPendingLOD at the same time.
+        std::array<S32, LLModel::NUM_LODS> pending_lods;
+        bool has_pending_lods = false;
+        {
+            LLMutexLock lock(mPendingMutex); // make sure only one thread access mPendingLOD at the same time.
+            pending_lod_map::iterator iter = mPendingLOD.find(mesh_id);
+            if (iter != mPendingLOD.end())
+            {
+                pending_lods = iter->second;
+                mPendingLOD.erase(iter);
+                has_pending_lods = true;
+            }
+        }
 
         //check for pending requests
-        pending_lod_map::iterator iter = mPendingLOD.find(mesh_id);
-        if (iter != mPendingLOD.end())
+        if (has_pending_lods)
         {
-            for (U32 i = 0; i < iter->second.size(); ++i)
+            for (S32 i = 0; i < pending_lods.size(); ++i)
             {
-                LODRequest req(mesh_params, iter->second[i]);
-                mLODReqQ.push(req);
-                LLMeshRepository::sLODProcessing++;
+                if (pending_lods[i] > 1)
+                {
+                    // mLoadingMeshes should be protecting from dupplciates, but looks
+                    // like this is possible if object rezzes, unregisterMesh, then
+                    // rezzes again before first request completes.
+                    // mLoadingMeshes might need to change a bit to not rerequest if
+                    // mesh is already pending.
+                    //
+                    // Todo: Improve mLoadingMeshes and once done turn this into an assert.
+                    // Low priority since such situation should be relatively rare
+                    LL_INFOS(LOG_MESH) << "Multiple dupplicate requests for mesd ID:  " << mesh_id << " LOD: " << i
+                        << LL_ENDL;
+                }
+                if (pending_lods[i] > 0 && lod_size[i] > 0)
+                {
+                    // try to load from data we just received
+                    bool request_lod = true;
+                    S32 offset = (S32)header_size + lod_offset[i];
+                    if (offset + lod_size[i] <= data_size)
+                    {
+                        // initial request is 4096 bytes, it's big enough to fit this lod
+                        request_lod = lodReceived(mesh_params, i, data + offset, lod_size[i]) != MESH_OK;
+                    }
+                    if (request_lod)
+                    {
+                        LLMutexLock lock(mMutex);
+                        LODRequest req(mesh_params, i);
+                        mLODReqQ.push(req);
+                        LLMeshRepository::sLODProcessing++;
+                    }
+                }
             }
-            mPendingLOD.erase(iter);
         }
     }
 
@@ -1995,8 +2295,16 @@ EMeshProcessingResult LLMeshRepoThread::lodReceived(const LLVolumeParams& mesh_p
         if (volume->getNumFaces() > 0)
         {
             // if we have a valid SkinInfo, cache per-joint bounding boxes for this LOD
-            LLMeshSkinInfo* skin_info = mSkinMap[mesh_params.getSculptID()];
-            if (skin_info && isAgentAvatarValid())
+            LLPointer<LLMeshSkinInfo> skin_info = nullptr;
+            {
+                LLMutexLock lock(mSkinMapMutex);
+                skin_map::iterator iter = mSkinMap.find(mesh_params.getSculptID());
+                if (iter != mSkinMap.end())
+                {
+                    skin_info = iter->second;
+                }
+            }
+            if (skin_info.notNull() && isAgentAvatarValid())
             {
                 for (S32 i = 0; i < volume->getNumFaces(); ++i)
                 {
@@ -2008,7 +2316,7 @@ EMeshProcessingResult LLMeshRepoThread::lodReceived(const LLVolumeParams& mesh_p
 
             LoadedMesh mesh(volume, mesh_params, lod);
             {
-                LLMutexLock lock(mMutex);
+                LLMutexLock lock(mLoadedMutex);
                 mLoadedQ.push_back(mesh);
                 // LLPointer is not thread safe, since we added this pointer into
                 // threaded list, make sure counter gets decreased inside mutex lock
@@ -2026,6 +2334,7 @@ EMeshProcessingResult LLMeshRepoThread::lodReceived(const LLVolumeParams& mesh_p
 
 bool LLMeshRepoThread::skinInfoReceived(const LLUUID& mesh_id, U8* data, S32 data_size)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LLSD skin;
 
     if (data_size > 0)
@@ -2060,12 +2369,15 @@ bool LLMeshRepoThread::skinInfoReceived(const LLUUID& mesh_id, U8* data, S32 dat
 
         // copy the skin info for the background thread so we can use it
         // to calculate per-joint bounding boxes when volumes are loaded
-        mSkinMap[mesh_id] = new LLMeshSkinInfo(*info);
+        {
+            LLMutexLock lock(mSkinMapMutex);
+            mSkinMap[mesh_id] = new LLMeshSkinInfo(*info);
+        }
 
         {
             // Move the LLPointer in to the skin info queue to avoid reference
             // count modification after we leave the lock
-            LLMutexLock lock(mMutex);
+            LLMutexLock lock(mLoadedMutex);
             mSkinInfoQ.emplace_back(std::move(info));
         }
     }
@@ -2075,6 +2387,7 @@ bool LLMeshRepoThread::skinInfoReceived(const LLUUID& mesh_id, U8* data, S32 dat
 
 bool LLMeshRepoThread::decompositionReceived(const LLUUID& mesh_id, U8* data, S32 data_size)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LLSD decomp;
 
     if (data_size > 0)
@@ -2101,7 +2414,7 @@ bool LLMeshRepoThread::decompositionReceived(const LLUUID& mesh_id, U8* data, S3
         LLModel::Decomposition* d = new LLModel::Decomposition(decomp);
         d->mMeshID = mesh_id;
         {
-            LLMutexLock lock(mMutex);
+            LLMutexLock lock(mLoadedMutex);
             mDecompositionQ.push_back(d);
         }
     }
@@ -2111,6 +2424,7 @@ bool LLMeshRepoThread::decompositionReceived(const LLUUID& mesh_id, U8* data, S3
 
 EMeshProcessingResult LLMeshRepoThread::physicsShapeReceived(const LLUUID& mesh_id, U8* data, S32 data_size)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LLSD physics_shape;
 
     LLModel::Decomposition* d = new LLModel::Decomposition();
@@ -2150,7 +2464,7 @@ EMeshProcessingResult LLMeshRepoThread::physicsShapeReceived(const LLUUID& mesh_
     }
 
     {
-        LLMutexLock lock(mMutex);
+        LLMutexLock lock(mLoadedMutex);
         mDecompositionQ.push_back(d);
     }
     return MESH_OK;
@@ -2202,7 +2516,6 @@ LLMeshUploadThread::~LLMeshUploadThread()
     mHttpRequest = NULL;
     delete mMutex;
     mMutex = NULL;
-
 }
 
 LLMeshUploadThread::DecompRequest::DecompRequest(LLModel* mdl, LLModel* base_model, LLMeshUploadThread* thread)
@@ -2943,15 +3256,17 @@ void LLMeshRepoThread::notifyLoadedMeshes()
         return;
     }
 
+    LL_PROFILE_ZONE_SCOPED;
+
     if (!mLoadedQ.empty())
     {
         std::deque<LoadedMesh> loaded_queue;
 
-        mMutex->lock();
+        mLoadedMutex->lock();
         if (!mLoadedQ.empty())
         {
             loaded_queue.swap(mLoadedQ);
-            mMutex->unlock();
+            mLoadedMutex->unlock();
 
             update_metrics = true;
 
@@ -2960,40 +3275,47 @@ void LLMeshRepoThread::notifyLoadedMeshes()
             {
                 if (mesh.mVolume->getNumVolumeFaces() > 0)
                 {
-                    gMeshRepo.notifyMeshLoaded(mesh.mMeshParams, mesh.mVolume);
+                    gMeshRepo.notifyMeshLoaded(mesh.mMeshParams, mesh.mVolume, mesh.mLOD);
                 }
                 else
                 {
-                    gMeshRepo.notifyMeshUnavailable(mesh.mMeshParams,
-                        LLVolumeLODGroup::getVolumeDetailFromScale(mesh.mVolume->getDetail()));
+                    gMeshRepo.notifyMeshUnavailable(mesh.mMeshParams, mesh.mLOD, LLVolumeLODGroup::getVolumeDetailFromScale(mesh.mVolume->getDetail()));
                 }
             }
         }
+        else
+        {
+            mLoadedMutex->unlock();
+        }
     }
 
     if (!mUnavailableQ.empty())
     {
         std::deque<LODRequest> unavil_queue;
 
-        mMutex->lock();
+        mLoadedMutex->lock();
         if (!mUnavailableQ.empty())
         {
             unavil_queue.swap(mUnavailableQ);
-            mMutex->unlock();
+            mLoadedMutex->unlock();
 
             update_metrics = true;
 
             // Process the elements free of the lock
             for (const auto& req : unavil_queue)
             {
-                gMeshRepo.notifyMeshUnavailable(req.mMeshParams, req.mLOD);
+                gMeshRepo.notifyMeshUnavailable(req.mMeshParams, req.mLOD, req.mLOD);
             }
         }
+        else
+        {
+            mLoadedMutex->unlock();
+        }
     }
 
     if (!mSkinInfoQ.empty() || !mSkinUnavailableQ.empty() || ! mDecompositionQ.empty())
     {
-        if (mMutex->trylock())
+        if (mLoadedMutex->trylock())
         {
             std::deque<LLPointer<LLMeshSkinInfo>> skin_info_q;
             std::deque<UUIDBasedRequest> skin_info_unavail_q;
@@ -3014,7 +3336,7 @@ void LLMeshRepoThread::notifyLoadedMeshes()
                 decomp_q.swap(mDecompositionQ);
             }
 
-            mMutex->unlock();
+            mLoadedMutex->unlock();
 
             // Process the elements free of the lock
             while (! skin_info_q.empty())
@@ -3051,9 +3373,11 @@ S32 LLMeshRepoThread::getActualMeshLOD(const LLVolumeParams& mesh_params, S32 lo
 
     if (iter != mMeshHeader.end())
     {
-        auto& header = iter->second.second;
-
-        return LLMeshRepository::getActualMeshLOD(header, lod);
+        auto& header = iter->second;
+        if (header.mHeaderSize > 0)
+        {
+            return LLMeshRepository::getActualMeshLOD(header, lod);
+        }
     }
 
     return lod;
@@ -3220,7 +3544,10 @@ void LLMeshHandlerBase::onCompleted(LLCore::HttpHandle handle, LLCore::HttpRespo
 
         processData(body, body_offset, data, static_cast<S32>(data_size) - body_offset);
 
-        delete [] data;
+        if (mHasDataOwnership)
+        {
+            delete [] data;
+        }
     }
 
     // Release handler
@@ -3253,7 +3580,7 @@ void LLMeshHeaderHandler::processFailure(LLCore::HttpStatus status)
                        << LL_ENDL;
 
     // Can't get the header so none of the LODs will be available
-    LLMutexLock lock(gMeshRepo.mThread->mMutex);
+    LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
     for (int i(0); i < LLVolumeLODGroup::NUM_LODS; ++i)
     {
         gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, i));
@@ -3263,6 +3590,7 @@ void LLMeshHeaderHandler::processFailure(LLCore::HttpStatus status)
 void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* body_offset */,
                                       U8 * data, S32 data_size)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LLUUID mesh_id = mMeshParams.getSculptID();
     bool success = (!MESH_HEADER_PROCESS_FAILED)
         && ((data != NULL) == (data_size > 0)); // if we have data but no size or have size but no data, something is wrong;
@@ -3282,7 +3610,7 @@ void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* b
                            << LL_ENDL;
 
         // Can't get the header so none of the LODs will be available
-        LLMutexLock lock(gMeshRepo.mThread->mMutex);
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
         for (int i(0); i < LLVolumeLODGroup::NUM_LODS; ++i)
         {
             gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, i));
@@ -3298,8 +3626,8 @@ void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* b
         LLMeshRepoThread::mesh_header_map::iterator iter = gMeshRepo.mThread->mMeshHeader.find(mesh_id);
         if (iter != gMeshRepo.mThread->mMeshHeader.end())
         {
-            header_bytes = (S32)iter->second.first;
-            header = iter->second.second;
+            header = iter->second;
+            header_bytes = header.mHeaderSize;
         }
 
         if (header_bytes > 0
@@ -3324,7 +3652,7 @@ void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* b
             // LLSD is smart and can work like smart pointer, is not thread safe.
             gMeshRepo.mThread->mHeaderMutex->unlock();
 
-            S32 bytes = lod_bytes + header_bytes;
+            S32 bytes = lod_bytes + header_bytes + CACHE_PREAMBLE_SIZE;
 
 
             // It's possible for the remote asset to have more data than is needed for the local cache
@@ -3337,6 +3665,11 @@ void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* b
                 LLMeshRepository::sCacheBytesWritten += data_size;
                 ++LLMeshRepository::sCacheWrites;
 
+                // write preamble
+                U32 flags = header.getFlags();
+                write_preamble(file, header_bytes, flags);
+
+                // write header
                 file.write(data, data_size);
 
                 S32 remaining = bytes - file.tell();
@@ -3359,7 +3692,7 @@ void LLMeshHeaderHandler::processData(LLCore::BufferArray * /* body */, S32 /* b
             gMeshRepo.mThread->mHeaderMutex->unlock();
 
             // headerReceived() parsed header, but header's data is invalid so none of the LODs will be available
-            LLMutexLock lock(gMeshRepo.mThread->mMutex);
+            LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
             for (int i(0); i < LLVolumeLODGroup::NUM_LODS; ++i)
             {
                 gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, i));
@@ -3388,9 +3721,64 @@ void LLMeshLODHandler::processFailure(LLCore::HttpStatus status)
                        << " (" << status.toTerseString() << ").  Not retrying."
                        << LL_ENDL;
 
-    LLMutexLock lock(gMeshRepo.mThread->mMutex);
+    LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
     gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, mLOD));
 }
+void LLMeshLODHandler::processLod(U8* data, S32 data_size)
+{
+    EMeshProcessingResult result = gMeshRepo.mThread->lodReceived(mMeshParams, mLOD, data, data_size);
+    if (result == MESH_OK)
+    {
+        // good fetch from sim, write to cache
+        LLFileSystem file(mMeshParams.getSculptID(), LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
+
+        S32 offset = mOffset + CACHE_PREAMBLE_SIZE;
+        S32 size = mRequestedBytes;
+
+        if (file.getSize() >= offset + size)
+        {
+            S32 header_bytes = 0;
+            U32 flags = 0;
+            {
+                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                LLMeshRepoThread::mesh_header_map::iterator header_it = gMeshRepo.mThread->mMeshHeader.find(mMeshParams.getSculptID());
+                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                {
+                    LLMeshHeader& header = header_it->second;
+                    // update header
+                    if (!header.mLodInCache[mLOD])
+                    {
+                        header.mLodInCache[mLOD] = true;
+                        header_bytes = header.mHeaderSize;
+                        flags = header.getFlags();
+                    }
+                    // todo: handle else because we shouldn't have requested twice?
+                }
+            }
+            if (flags > 0)
+            {
+                write_preamble(file, header_bytes, flags);
+            }
+
+            file.seek(offset, 0);
+            file.write(data, size);
+            LLMeshRepository::sCacheBytesWritten += size;
+            ++LLMeshRepository::sCacheWrites;
+        }
+    }
+    else
+    {
+        LL_WARNS(LOG_MESH) << "Error during mesh LOD processing.  ID:  " << mMeshParams.getSculptID()
+            << ", Reason: " << result
+            << " LOD: " << mLOD
+            << " Data size: " << data_size
+            << " Not retrying."
+            << LL_ENDL;
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
+        gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, mLOD));
+    }
+}
 
 void LLMeshLODHandler::processData(LLCore::BufferArray * /* body */, S32 /* body_offset */,
                                    U8 * data, S32 data_size)
@@ -3399,33 +3787,26 @@ void LLMeshLODHandler::processData(LLCore::BufferArray * /* body */, S32 /* body
     if ((!MESH_LOD_PROCESS_FAILED)
         && ((data != NULL) == (data_size > 0))) // if we have data but no size or have size but no data, something is wrong
     {
-        EMeshProcessingResult result = gMeshRepo.mThread->lodReceived(mMeshParams, mLOD, data, data_size);
-        if (result == MESH_OK)
+        LLMeshHandlerBase::ptr_t shrd_handler = shared_from_this();
+        bool posted = gMeshRepo.mThread->mMeshThreadPool->getQueue().post(
+            [shrd_handler, data, data_size]
+            ()
         {
-            // good fetch from sim, write to cache
-            LLFileSystem file(mMeshParams.getSculptID(), LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
-
-            S32 offset = mOffset;
-            S32 size = mRequestedBytes;
+            LLMeshLODHandler* handler = (LLMeshLODHandler * )shrd_handler.get();
+            handler->processLod(data, data_size);
+            delete[] data;
+        });
 
-            if (file.getSize() >= offset+size)
-            {
-                file.seek(offset);
-                file.write(data, size);
-                LLMeshRepository::sCacheBytesWritten += size;
-                ++LLMeshRepository::sCacheWrites;
-            }
+        if (posted)
+        {
+            // ownership of data was passed to the lambda
+            mHasDataOwnership = false;
         }
         else
         {
-            LL_WARNS(LOG_MESH) << "Error during mesh LOD processing.  ID:  " << mMeshParams.getSculptID()
-                               << ", Reason: " << result
-                               << " LOD: " << mLOD
-                               << " Data size: " << data_size
-                               << " Not retrying."
-                               << LL_ENDL;
-            LLMutexLock lock(gMeshRepo.mThread->mMutex);
-            gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, mLOD));
+            // mesh thread dies later than event queue, so this is normal
+            LL_INFOS_ONCE(LOG_MESH) << "Failed to post work into mMeshThreadPool" << LL_ENDL;
+            processLod(data, data_size);
         }
     }
     else
@@ -3435,7 +3816,7 @@ void LLMeshLODHandler::processData(LLCore::BufferArray * /* body */, S32 /* body
                            << " LOD: " << mLOD
                            << " Data size: " << data_size
                            << LL_ENDL;
-        LLMutexLock lock(gMeshRepo.mThread->mMutex);
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
         gMeshRepo.mThread->mUnavailableQ.push_back(LLMeshRepoThread::LODRequest(mMeshParams, mLOD));
     }
 }
@@ -3446,6 +3827,7 @@ LLMeshSkinInfoHandler::~LLMeshSkinInfoHandler()
     {
         LL_WARNS(LOG_MESH) << "deleting unprocessed request handler (may be ok on exit)" << LL_ENDL;
     }
+    LLMeshRepoThread::decActiveSkinRequests();
 }
 
 void LLMeshSkinInfoHandler::processFailure(LLCore::HttpStatus status)
@@ -3454,38 +3836,98 @@ void LLMeshSkinInfoHandler::processFailure(LLCore::HttpStatus status)
                        << ", Reason:  " << status.toString()
                        << " (" << status.toTerseString() << ").  Not retrying."
                        << LL_ENDL;
-        LLMutexLock lock(gMeshRepo.mThread->mMutex);
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
         gMeshRepo.mThread->mSkinUnavailableQ.emplace_back(mMeshID);
 }
 
-void LLMeshSkinInfoHandler::processData(LLCore::BufferArray * /* body */, S32 /* body_offset */,
-                                        U8 * data, S32 data_size)
+void LLMeshSkinInfoHandler::processSkin(U8* data, S32 data_size)
 {
-    LL_PROFILE_ZONE_SCOPED;
-    if ((!MESH_SKIN_INFO_PROCESS_FAILED)
-        && ((data != NULL) == (data_size > 0)) // if we have data but no size or have size but no data, something is wrong
-        && gMeshRepo.mThread->skinInfoReceived(mMeshID, data, data_size))
+    if (gMeshRepo.mThread->skinInfoReceived(mMeshID, data, data_size))
     {
         // good fetch from sim, write to cache
         LLFileSystem file(mMeshID, LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
 
-        S32 offset = mOffset;
+        S32 offset = mOffset + CACHE_PREAMBLE_SIZE;
         S32 size = mRequestedBytes;
 
-        if (file.getSize() >= offset+size)
+        if (file.getSize() >= offset + size)
         {
             LLMeshRepository::sCacheBytesWritten += size;
             ++LLMeshRepository::sCacheWrites;
-            file.seek(offset);
+
+            S32 header_bytes = 0;
+            U32 flags = 0;
+            {
+                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                LLMeshRepoThread::mesh_header_map::iterator header_it = gMeshRepo.mThread->mMeshHeader.find(mMeshID);
+                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                {
+                    LLMeshHeader& header = header_it->second;
+                    // update header
+                    if (!header.mSkinInCache)
+                    {
+                        header.mSkinInCache = true;
+                        header_bytes = header.mHeaderSize;
+                        flags = header.getFlags();
+                    }
+                    // todo: handle else because we shouldn't have requested twice?
+                }
+            }
+            if (flags > 0)
+            {
+                write_preamble(file, header_bytes, flags);
+            }
+
+            file.seek(offset, 0);
             file.write(data, size);
         }
     }
     else
     {
         LL_WARNS(LOG_MESH) << "Error during mesh skin info processing.  ID:  " << mMeshID
+            << ", Unknown reason.  Not retrying."
+            << LL_ENDL;
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
+        gMeshRepo.mThread->mSkinUnavailableQ.emplace_back(mMeshID);
+    }
+}
+
+void LLMeshSkinInfoHandler::processData(LLCore::BufferArray * /* body */, S32 /* body_offset */,
+                                        U8 * data, S32 data_size)
+{
+    LL_PROFILE_ZONE_SCOPED;
+    if ((!MESH_SKIN_INFO_PROCESS_FAILED)
+        && ((data != NULL) == (data_size > 0))) // if we have data but no size or have size but no data, something is wrong
+    {
+        LLMeshHandlerBase::ptr_t shrd_handler = shared_from_this();
+        bool posted = gMeshRepo.mThread->mMeshThreadPool->getQueue().post(
+            [shrd_handler, data, data_size]
+            ()
+        {
+            LLMeshSkinInfoHandler* handler = (LLMeshSkinInfoHandler*)shrd_handler.get();
+            handler->processSkin(data, data_size);
+            delete[] data;
+        });
+
+        if (posted)
+        {
+            // ownership of data was passed to the lambda
+            mHasDataOwnership = false;
+        }
+        else
+        {
+            // mesh thread dies later than event queue, so this is normal
+            LL_INFOS_ONCE(LOG_MESH) << "Failed to post work into mMeshThreadPool" << LL_ENDL;
+            processSkin(data, data_size);
+        }
+    }
+    else
+    {
+        LL_WARNS(LOG_MESH) << "Error during mesh skin info processing.  ID:  " << mMeshID
                            << ", Unknown reason.  Not retrying."
                            << LL_ENDL;
-        LLMutexLock lock(gMeshRepo.mThread->mMutex);
+        LLMutexLock lock(gMeshRepo.mThread->mLoadedMutex);
         gMeshRepo.mThread->mSkinUnavailableQ.emplace_back(mMeshID);
     }
 }
@@ -3519,14 +3961,39 @@ void LLMeshDecompositionHandler::processData(LLCore::BufferArray * /* body */, S
         // good fetch from sim, write to cache
         LLFileSystem file(mMeshID, LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
 
-        S32 offset = mOffset;
+        S32 offset = mOffset + CACHE_PREAMBLE_SIZE;
         S32 size = mRequestedBytes;
 
         if (file.getSize() >= offset+size)
         {
             LLMeshRepository::sCacheBytesWritten += size;
             ++LLMeshRepository::sCacheWrites;
-            file.seek(offset);
+
+            S32 header_bytes = 0;
+            U32 flags = 0;
+            {
+                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                LLMeshRepoThread::mesh_header_map::iterator header_it = gMeshRepo.mThread->mMeshHeader.find(mMeshID);
+                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                {
+                    LLMeshHeader& header = header_it->second;
+                    // update header
+                    if (!header.mPhysicsConvexInCache)
+                    {
+                        header.mPhysicsConvexInCache = true;
+                        header_bytes = header.mHeaderSize;
+                        flags = header.getFlags();
+                    }
+                    // todo: handle else because we shouldn't have requested twice?
+                }
+            }
+            if (flags > 0)
+            {
+                write_preamble(file, header_bytes, flags);
+            }
+
+            file.seek(offset, 0);
             file.write(data, size);
         }
     }
@@ -3567,14 +4034,39 @@ void LLMeshPhysicsShapeHandler::processData(LLCore::BufferArray * /* body */, S3
         // good fetch from sim, write to cache for caching
         LLFileSystem file(mMeshID, LLAssetType::AT_MESH, LLFileSystem::READ_WRITE);
 
-        S32 offset = mOffset;
+        S32 offset = mOffset + CACHE_PREAMBLE_SIZE;
         S32 size = mRequestedBytes;
 
         if (file.getSize() >= offset+size)
         {
             LLMeshRepository::sCacheBytesWritten += size;
             ++LLMeshRepository::sCacheWrites;
-            file.seek(offset);
+
+            S32 header_bytes = 0;
+            U32 flags = 0;
+            {
+                LLMutexLock lock(gMeshRepo.mThread->mHeaderMutex);
+
+                LLMeshRepoThread::mesh_header_map::iterator header_it = gMeshRepo.mThread->mMeshHeader.find(mMeshID);
+                if (header_it != gMeshRepo.mThread->mMeshHeader.end())
+                {
+                    LLMeshHeader& header = header_it->second;
+                    // update header
+                    if (!header.mPhysicsMeshInCache)
+                    {
+                        header.mPhysicsMeshInCache = true;
+                        header_bytes = header.mHeaderSize;
+                        flags = header.getFlags();
+                    }
+                    // todo: handle else because we shouldn't have requested twice?
+                }
+            }
+            if (flags > 0)
+            {
+                write_preamble(file, header_bytes, flags);
+            }
+
+            file.seek(offset, 0);
             file.write(data, size);
         }
     }
@@ -3636,6 +4128,7 @@ void LLMeshRepository::shutdown()
     }
 
     mThread->mSignal->broadcast();
+    mThread->mMeshThreadPool->close();
 
     while (!mThread->isStopped())
     {
@@ -3710,24 +4203,24 @@ void LLMeshRepository::unregisterMesh(LLVOVolume* vobj)
     }
 }
 
-S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_params, S32 detail, S32 last_lod)
+S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_params, S32 new_lod, S32 last_lod)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_NETWORK; //LL_LL_RECORD_BLOCK_TIME(FTM_MESH_FETCH);
 
     // Manage time-to-load metrics for mesh download operations.
     metricsProgress(1);
 
-    if (detail < 0 || detail >= LLVolumeLODGroup::NUM_LODS)
+    if (new_lod < 0 || new_lod >= LLVolumeLODGroup::NUM_LODS)
     {
-        return detail;
+        return new_lod;
     }
 
     {
         LLMutexLock lock(mMeshMutex);
         //add volume to list of loading meshes
         const auto& mesh_id = mesh_params.getSculptID();
-        mesh_load_map::iterator iter = mLoadingMeshes[detail].find(mesh_id);
-        if (iter != mLoadingMeshes[detail].end())
+        mesh_load_map::iterator iter = mLoadingMeshes[new_lod].find(mesh_id);
+        if (iter != mLoadingMeshes[new_lod].end())
         { //request pending for this mesh, append volume id to list
             auto it = std::find(iter->second.begin(), iter->second.end(), vobj);
             if (it == iter->second.end()) {
@@ -3737,8 +4230,8 @@ S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_para
         else
         {
             //first request for this mesh
-            mLoadingMeshes[detail][mesh_id].push_back(vobj);
-            mPendingRequests.push_back(LLMeshRepoThread::LODRequest(mesh_params, detail));
+            mLoadingMeshes[new_lod][mesh_id].push_back(vobj);
+            mPendingRequests.emplace_back(new PendingRequestLOD(mesh_params, new_lod));
             LLMeshRepository::sLODPending++;
         }
     }
@@ -3767,7 +4260,7 @@ S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_para
             }
 
             //next, see what the next lowest LOD available might be
-            for (S32 i = detail-1; i >= 0; --i)
+            for (S32 i = new_lod -1; i >= 0; --i)
             {
                 LLVolume* lod = group->refLOD(i);
                 if (lod && lod->isMeshAssetLoaded() && lod->getNumVolumeFaces() > 0)
@@ -3780,7 +4273,7 @@ S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_para
             }
 
             //no lower LOD is a available, is a higher lod available?
-            for (S32 i = detail+1; i < LLVolumeLODGroup::NUM_LODS; ++i)
+            for (S32 i = new_lod+1; i < LLVolumeLODGroup::NUM_LODS; ++i)
             {
                 LLVolume* lod = group->refLOD(i);
                 if (lod && lod->isMeshAssetLoaded() && lod->getNumVolumeFaces() > 0)
@@ -3794,7 +4287,51 @@ S32 LLMeshRepository::loadMesh(LLVOVolume* vobj, const LLVolumeParams& mesh_para
         }
     }
 
-    return detail;
+    return new_lod;
+}
+
+F32 calculate_score(LLVOVolume* object)
+{
+    if (!object)
+    {
+        return -1.f;
+    }
+    LLDrawable* drawable = object->mDrawable;
+    if (!drawable)
+    {
+        return -1;
+    }
+    if (drawable->isState(LLDrawable::RIGGED) || object->isAttachment())
+    {
+        LLVOAvatar* avatar = object->getAvatar();
+        LLDrawable* av_drawable = avatar ? avatar->mDrawable : nullptr;
+        if (avatar && av_drawable)
+        {
+            // See LLVOVolume::calcLOD()
+            F32 radius;
+            if (avatar->isControlAvatar())
+            {
+                const LLVector3* box = avatar->getLastAnimExtents();
+                LLVector3 diag = box[1] - box[0];
+                radius = diag.magVec() * 0.5f;
+            }
+            else
+            {
+                // Volume in a rigged mesh attached to a regular avatar.
+                const LLVector3* box = avatar->getLastAnimExtents();
+                LLVector3 diag = box[1] - box[0];
+                radius = diag.magVec();
+
+                if (!avatar->isSelf() && !avatar->hasFirstFullAttachmentData())
+                {
+                    // slightly deprioritize avatars that are still receiving data
+                    radius *= 0.9f;
+                }
+            }
+            return radius / llmax(av_drawable->mDistanceWRTCamera, 1.f);
+        }
+    }
+    return drawable->getRadius() / llmax(drawable->mDistanceWRTCamera, 1.f);
 }
 
 void LLMeshRepository::notifyLoadedMeshes()
@@ -3918,6 +4455,7 @@ void LLMeshRepository::notifyLoadedMeshes()
             // erase from background thread
             mThread->mWorkQueue.post([=, this]()
                 {
+                    LLMutexLock(mThread->mSkinMapMutex);
                     mThread->mSkinMap.erase(id);
                 });
         }
@@ -3974,8 +4512,12 @@ void LLMeshRepository::notifyLoadedMeshes()
             mUploadErrorQ.pop();
         }
 
-        S32 active_count = LLMeshRepoThread::sActiveHeaderRequests + LLMeshRepoThread::sActiveLODRequests;
-        if (active_count < LLMeshRepoThread::sRequestLowWater)
+        // mPendingRequests go into queues, queues go into active http requests.
+        // Checking sRequestHighWater to keep queues at least somewhat populated
+        // for faster transition into http
+        S32 active_count = LLMeshRepoThread::sActiveHeaderRequests + LLMeshRepoThread::sActiveLODRequests + LLMeshRepoThread::sActiveSkinRequests;
+        active_count += (S32)(mThread->mLODReqQ.size() + mThread->mHeaderReqQ.size() + mThread->mSkinInfoQ.size());
+        if (active_count < LLMeshRepoThread::sRequestHighWater)
         {
             S32 push_count = LLMeshRepoThread::sRequestHighWater - active_count;
 
@@ -3996,50 +4538,69 @@ void LLMeshRepository::notifyLoadedMeshes()
                         F32 max_score = 0.f;
                         for (auto obj_iter = iter->second.begin(); obj_iter != iter->second.end(); ++obj_iter)
                         {
-                            LLVOVolume* object = *obj_iter;
-                            if (object)
+                            F32 cur_score = calculate_score(*obj_iter);
+                            if (cur_score >= 0.f)
                             {
-                                LLDrawable* drawable = object->mDrawable;
-                                if (drawable)
-                                {
-                                    F32 cur_score = drawable->getRadius()/llmax(drawable->mDistanceWRTCamera, 1.f);
-                                    max_score = llmax(max_score, cur_score);
-                                }
+                                max_score = llmax(max_score, cur_score);
                             }
                         }
 
                         score_map[iter->first] = max_score;
                     }
                 }
+                for (mesh_load_map::iterator iter = mLoadingSkins.begin(); iter != mLoadingSkins.end(); ++iter)
+                {
+                    F32 max_score = 0.f;
+                    for (auto obj_iter = iter->second.begin(); obj_iter != iter->second.end(); ++obj_iter)
+                    {
+                        F32 cur_score = calculate_score(*obj_iter);
+                        if (cur_score >= 0.f)
+                        {
+                            max_score = llmax(max_score, cur_score);
+                        }
+                    }
+
+                    score_map[iter->first] = max_score;
+                }
 
                 //set "score" for pending requests
-                for (std::vector<LLMeshRepoThread::LODRequest>::iterator iter = mPendingRequests.begin(); iter != mPendingRequests.end(); ++iter)
+                for (std::unique_ptr<PendingRequestBase>& req_p : mPendingRequests)
                 {
-                    iter->mScore = score_map[iter->mMeshParams.getSculptID()];
+                    req_p->setScore(score_map[req_p->getId()]);
                 }
 
                 //sort by "score"
                 std::partial_sort(mPendingRequests.begin(), mPendingRequests.begin() + push_count,
-                                  mPendingRequests.end(), LLMeshRepoThread::CompareScoreGreater());
+                                  mPendingRequests.end(), PendingRequestBase::CompareScoreGreater());
             }
-
             while (!mPendingRequests.empty() && push_count > 0)
             {
-                LLMeshRepoThread::LODRequest& request = mPendingRequests.front();
-                mThread->loadMeshLOD(request.mMeshParams, request.mLOD);
+                std::unique_ptr<PendingRequestBase>& req_p = mPendingRequests.front();
+                switch (req_p->getRequestType())
+                {
+                case MESH_REQUEST_LOD:
+                    {
+                        PendingRequestLOD* lod = (PendingRequestLOD*)req_p.get();
+                        mThread->loadMeshLOD(lod->mMeshParams, lod->mLOD);
+                        LLMeshRepository::sLODPending--;
+                        break;
+                    }
+                case MESH_REQUEST_SKIN:
+                    {
+                        PendingRequestUUID* skin = (PendingRequestUUID*)req_p.get();
+                        mThread->loadMeshSkinInfo(skin->getId());
+                        break;
+                    }
+
+                default:
+                    LL_ERRS() << "Unknown request type in LLMeshRepository::notifyLoadedMeshes" << LL_ENDL;
+                    break;
+                }
                 mPendingRequests.erase(mPendingRequests.begin());
-                LLMeshRepository::sLODPending--;
                 push_count--;
             }
         }
 
-        //send skin info requests
-        while (!mPendingSkinRequests.empty())
-        {
-            mThread->loadMeshSkinInfo(mPendingSkinRequests.front());
-            mPendingSkinRequests.pop();
-        }
-
         //send decomposition requests
         while (!mPendingDecompositionRequests.empty())
         {
@@ -4116,15 +4677,14 @@ void LLMeshRepository::notifyDecompositionReceived(LLModel::Decomposition* decom
     }
 }
 
-void LLMeshRepository::notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVolume* volume)
+void LLMeshRepository::notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVolume* volume, S32 lod)
 { //called from main thread
-    S32 detail = LLVolumeLODGroup::getVolumeDetailFromScale(volume->getDetail());
 
     //get list of objects waiting to be notified this mesh is loaded
     const auto& mesh_id = mesh_params.getSculptID();
-    mesh_load_map::iterator obj_iter = mLoadingMeshes[detail].find(mesh_id);
+    mesh_load_map::iterator obj_iter = mLoadingMeshes[lod].find(mesh_id);
 
-    if (volume && obj_iter != mLoadingMeshes[detail].end())
+    if (volume && obj_iter != mLoadingMeshes[lod].end())
     {
         //make sure target volume is still valid
         if (volume->getNumVolumeFaces() <= 0)
@@ -4134,6 +4694,7 @@ void LLMeshRepository::notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVol
         }
 
         { //update system volume
+            S32 detail = LLVolumeLODGroup::getVolumeDetailFromScale(volume->getDetail());
             LLVolume* sys_volume = LLPrimitive::getVolumeManager()->refVolume(mesh_params, detail);
             if (sys_volume)
             {
@@ -4157,22 +4718,22 @@ void LLMeshRepository::notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVol
             }
         }
 
-        mLoadingMeshes[detail].erase(obj_iter);
+        mLoadingMeshes[lod].erase(obj_iter);
 
         LLViewerStatsRecorder::instance().meshLoaded();
     }
 }
 
-void LLMeshRepository::notifyMeshUnavailable(const LLVolumeParams& mesh_params, S32 lod)
+void LLMeshRepository::notifyMeshUnavailable(const LLVolumeParams& mesh_params, S32 request_lod, S32 volume_lod)
 { //called from main thread
     //get list of objects waiting to be notified this mesh is loaded
     const auto& mesh_id = mesh_params.getSculptID();
-    mesh_load_map::iterator obj_iter = mLoadingMeshes[lod].find(mesh_id);
-    if (obj_iter != mLoadingMeshes[lod].end())
+    mesh_load_map::iterator obj_iter = mLoadingMeshes[request_lod].find(mesh_id);
+    if (obj_iter != mLoadingMeshes[request_lod].end())
     {
-        F32 detail = LLVolumeLODGroup::getVolumeScaleFromDetail(lod);
+        F32 detail = LLVolumeLODGroup::getVolumeScaleFromDetail(volume_lod);
 
-        LLVolume* sys_volume = LLPrimitive::getVolumeManager()->refVolume(mesh_params, lod);
+        LLVolume* sys_volume = LLPrimitive::getVolumeManager()->refVolume(mesh_params, volume_lod);
         if (sys_volume)
         {
             sys_volume->setMeshAssetUnavaliable(true);
@@ -4189,12 +4750,12 @@ void LLMeshRepository::notifyMeshUnavailable(const LLVolumeParams& mesh_params,
                     obj_volume->getDetail() == detail &&
                     obj_volume->getParams() == mesh_params)
                 { //should force volume to find most appropriate LOD
-                    vobj->setVolume(obj_volume->getParams(), lod);
+                    vobj->setVolume(obj_volume->getParams(), volume_lod);
                 }
             }
         }
 
-        mLoadingMeshes[lod].erase(obj_iter);
+        mLoadingMeshes[request_lod].erase(obj_iter);
     }
 }
 
@@ -4231,7 +4792,7 @@ const LLMeshSkinInfo* LLMeshRepository::getSkinInfo(const LLUUID& mesh_id, LLVOV
             {
                 //first request for this mesh
                 mLoadingSkins[mesh_id].push_back(requesting_obj);
-                mPendingSkinRequests.push(mesh_id);
+                mPendingRequests.emplace_back(new PendingRequestUUID(mesh_id, MESH_REQUEST_SKIN));
             }
         }
     }
@@ -4358,7 +4919,7 @@ bool LLMeshRepository::hasSkinInfo(const LLUUID& mesh_id)
     return false;
 }
 
-bool LLMeshRepository::hasHeader(const LLUUID& mesh_id)
+bool LLMeshRepository::hasHeader(const LLUUID& mesh_id) const
 {
     if (mesh_id.isNull())
     {
@@ -4368,13 +4929,13 @@ bool LLMeshRepository::hasHeader(const LLUUID& mesh_id)
     return mThread->hasHeader(mesh_id);
 }
 
-bool LLMeshRepoThread::hasPhysicsShapeInHeader(const LLUUID& mesh_id)
+bool LLMeshRepoThread::hasPhysicsShapeInHeader(const LLUUID& mesh_id) const
 {
     LLMutexLock lock(mHeaderMutex);
-    mesh_header_map::iterator iter = mMeshHeader.find(mesh_id);
-    if (iter != mMeshHeader.end() && iter->second.first > 0)
+    mesh_header_map::const_iterator iter = mMeshHeader.find(mesh_id);
+    if (iter != mMeshHeader.end() && iter->second.mHeaderSize > 0)
     {
-        LLMeshHeader &mesh = iter->second.second;
+        const LLMeshHeader &mesh = iter->second;
         if (mesh.mPhysicsMeshSize > 0)
         {
             return true;
@@ -4384,13 +4945,13 @@ bool LLMeshRepoThread::hasPhysicsShapeInHeader(const LLUUID& mesh_id)
     return false;
 }
 
-bool LLMeshRepoThread::hasSkinInfoInHeader(const LLUUID& mesh_id)
+bool LLMeshRepoThread::hasSkinInfoInHeader(const LLUUID& mesh_id) const
 {
     LLMutexLock lock(mHeaderMutex);
-    mesh_header_map::iterator iter = mMeshHeader.find(mesh_id);
-    if (iter != mMeshHeader.end() && iter->second.first > 0)
+    mesh_header_map::const_iterator iter = mMeshHeader.find(mesh_id);
+    if (iter != mMeshHeader.end() && iter->second.mHeaderSize > 0)
     {
-        LLMeshHeader& mesh = iter->second.second;
+        const LLMeshHeader& mesh = iter->second;
         if (mesh.mSkinOffset >= 0
             && mesh.mSkinSize > 0)
         {
@@ -4401,10 +4962,10 @@ bool LLMeshRepoThread::hasSkinInfoInHeader(const LLUUID& mesh_id)
     return false;
 }
 
-bool LLMeshRepoThread::hasHeader(const LLUUID& mesh_id)
+bool LLMeshRepoThread::hasHeader(const LLUUID& mesh_id) const
 {
     LLMutexLock lock(mHeaderMutex);
-    mesh_header_map::iterator iter = mMeshHeader.find(mesh_id);
+    mesh_header_map::const_iterator iter = mMeshHeader.find(mesh_id);
     return iter != mMeshHeader.end();
 }
 
@@ -4419,16 +4980,16 @@ void LLMeshRepository::uploadModel(std::vector<LLModelInstance>& data, LLVector3
     mUploadWaitList.push_back(thread);
 }
 
-S32 LLMeshRepository::getMeshSize(const LLUUID& mesh_id, S32 lod)
+S32 LLMeshRepository::getMeshSize(const LLUUID& mesh_id, S32 lod) const
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_VOLUME;
     if (mThread && mesh_id.notNull() && LLPrimitive::NO_LOD != lod)
     {
         LLMutexLock lock(mThread->mHeaderMutex);
-        LLMeshRepoThread::mesh_header_map::iterator iter = mThread->mMeshHeader.find(mesh_id);
-        if (iter != mThread->mMeshHeader.end() && iter->second.first > 0)
+        LLMeshRepoThread::mesh_header_map::const_iterator iter = mThread->mMeshHeader.find(mesh_id);
+        if (iter != mThread->mMeshHeader.end() && iter->second.mHeaderSize > 0)
         {
-            const LLMeshHeader& header = iter->second.second;
+            const LLMeshHeader& header = iter->second;
 
             if (header.m404)
             {
@@ -4532,9 +5093,9 @@ F32 LLMeshRepository::getStreamingCostLegacy(LLUUID mesh_id, F32 radius, S32* by
     {
         LLMutexLock lock(mThread->mHeaderMutex);
         LLMeshRepoThread::mesh_header_map::iterator iter = mThread->mMeshHeader.find(mesh_id);
-        if (iter != mThread->mMeshHeader.end() && iter->second.first > 0)
+        if (iter != mThread->mMeshHeader.end() && iter->second.mHeaderSize > 0)
         {
-            result  = getStreamingCostLegacy(iter->second.second, radius, bytes, bytes_visible, lod, unscaled_value);
+            result  = getStreamingCostLegacy(iter->second, radius, bytes, bytes_visible, lod, unscaled_value);
         }
     }
     if (result > 0.f)
@@ -4726,7 +5287,7 @@ bool LLMeshCostData::init(const LLMeshHeader& header)
 }
 
 
-S32 LLMeshCostData::getSizeByLOD(S32 lod)
+S32 LLMeshCostData::getSizeByLOD(S32 lod) const
 {
     if (llclamp(lod,0,3) != lod)
     {
@@ -4735,12 +5296,12 @@ S32 LLMeshCostData::getSizeByLOD(S32 lod)
     return mSizeByLOD[lod];
 }
 
-S32 LLMeshCostData::getSizeTotal()
+S32 LLMeshCostData::getSizeTotal() const
 {
     return mSizeByLOD[0] + mSizeByLOD[1] + mSizeByLOD[2] + mSizeByLOD[3];
 }
 
-F32 LLMeshCostData::getEstTrisByLOD(S32 lod)
+F32 LLMeshCostData::getEstTrisByLOD(S32 lod) const
 {
     if (llclamp(lod,0,3) != lod)
     {
@@ -4749,12 +5310,12 @@ F32 LLMeshCostData::getEstTrisByLOD(S32 lod)
     return mEstTrisByLOD[lod];
 }
 
-F32 LLMeshCostData::getEstTrisMax()
+F32 LLMeshCostData::getEstTrisMax() const
 {
     return llmax(mEstTrisByLOD[0], mEstTrisByLOD[1], mEstTrisByLOD[2], mEstTrisByLOD[3]);
 }
 
-F32 LLMeshCostData::getRadiusWeightedTris(F32 radius)
+F32 LLMeshCostData::getRadiusWeightedTris(F32 radius) const
 {
     F32 max_distance = 512.f;
 
@@ -4798,7 +5359,7 @@ F32 LLMeshCostData::getRadiusWeightedTris(F32 radius)
     return weighted_avg;
 }
 
-F32 LLMeshCostData::getEstTrisForStreamingCost()
+F32 LLMeshCostData::getEstTrisForStreamingCost() const
 {
     LL_DEBUGS("StreamingCost") << "tris_by_lod: "
                                << mEstTrisByLOD[0] << ", "
@@ -4808,7 +5369,7 @@ F32 LLMeshCostData::getEstTrisForStreamingCost()
 
     F32 charged_tris = mEstTrisByLOD[3];
     F32 allowed_tris = mEstTrisByLOD[3];
-    const F32 ENFORCE_FLOOR = 64.0f;
+    constexpr F32 ENFORCE_FLOOR = 64.0f;
     for (S32 i=2; i>=0; i--)
     {
         // How many tris can we have in this LOD without affecting land impact?
@@ -4825,13 +5386,13 @@ F32 LLMeshCostData::getEstTrisForStreamingCost()
     return charged_tris;
 }
 
-F32 LLMeshCostData::getRadiusBasedStreamingCost(F32 radius)
+F32 LLMeshCostData::getRadiusBasedStreamingCost(F32 radius) const
 {
     static LLCachedControl<U32> mesh_triangle_budget(gSavedSettings, "MeshTriangleBudget");
     return getRadiusWeightedTris(radius)/mesh_triangle_budget*15000.f;
 }
 
-F32 LLMeshCostData::getTriangleBasedStreamingCost()
+F32 LLMeshCostData::getTriangleBasedStreamingCost() const
 {
     F32 result = ANIMATED_OBJECT_COST_PER_KTRI * 0.001f * getEstTrisForStreamingCost();
     return result;
@@ -4846,9 +5407,9 @@ bool LLMeshRepository::getCostData(LLUUID mesh_id, LLMeshCostData& data)
     {
         LLMutexLock lock(mThread->mHeaderMutex);
         LLMeshRepoThread::mesh_header_map::iterator iter = mThread->mMeshHeader.find(mesh_id);
-        if (iter != mThread->mMeshHeader.end() && iter->second.first > 0)
+        if (iter != mThread->mMeshHeader.end() && iter->second.mHeaderSize > 0)
         {
-            LLMeshHeader& header = iter->second.second;
+            LLMeshHeader& header = iter->second;
 
             bool header_invalid = (header.m404
                                    || header.mLodSize[0] <= 0
diff --git a/indra/newview/llmeshrepository.h b/indra/newview/llmeshrepository.h
index 3d25a4dd78..0d9da32e27 100644
--- a/indra/newview/llmeshrepository.h
+++ b/indra/newview/llmeshrepository.h
@@ -63,6 +63,16 @@ typedef enum e_mesh_processing_result_enum
     MESH_UNKNOWN
 } EMeshProcessingResult;
 
+typedef enum e_mesh_request_type_enum
+{
+    MESH_REQUEST_HEADER,
+    MESH_REQUEST_LOD,
+    MESH_REQUEST_SKIN,
+    MESH_REQUEST_DECOMPOSITION,
+    MESH_REQUEST_PHYSICS,
+    MESH_REQUEST_UKNOWN
+} EMeshRequestType;
+
 class LLMeshUploadData
 {
 public:
@@ -183,7 +193,8 @@ public:
 class RequestStats
 {
 public:
-    RequestStats() : mRetries(0) {};
+
+    RequestStats() :mRetries(0) {};
 
     void updateTime();
     bool canRetry() const;
@@ -195,6 +206,67 @@ private:
     LLFrameTimer mTimer;
 };
 
+
+class PendingRequestBase
+{
+public:
+    struct CompareScoreGreater
+    {
+        bool operator()(const std::unique_ptr<PendingRequestBase>& lhs, const std::unique_ptr<PendingRequestBase>& rhs)
+        {
+            return lhs->mScore > rhs->mScore; // greatest = first
+        }
+    };
+
+    PendingRequestBase() : mScore(0.f) {};
+    virtual ~PendingRequestBase() {}
+
+    bool operator<(const PendingRequestBase& rhs) const
+    {
+        return mId < rhs.mId;
+    }
+
+    void setScore(F32 score) { mScore = score; }
+    F32 getScore() const { return mScore; }
+    LLUUID getId() const { return mId; }
+    virtual EMeshRequestType getRequestType() const = 0;
+
+protected:
+    F32 mScore;
+    LLUUID mId;
+};
+
+class PendingRequestLOD : public PendingRequestBase
+{
+public:
+    LLVolumeParams  mMeshParams;
+    S32 mLOD;
+
+    PendingRequestLOD(const LLVolumeParams& mesh_params, S32 lod)
+        : PendingRequestBase(), mMeshParams(mesh_params), mLOD(lod)
+    {
+        mId = mMeshParams.getSculptID();
+    }
+
+    EMeshRequestType getRequestType() const override { return MESH_REQUEST_LOD; }
+};
+
+class PendingRequestUUID : public PendingRequestBase
+{
+public:
+
+    PendingRequestUUID(const LLUUID& id, EMeshRequestType type)
+        : PendingRequestBase(), mRequestType(type)
+    {
+        mId = id;
+    }
+
+    EMeshRequestType getRequestType() const override { return mRequestType; }
+
+private:
+    EMeshRequestType mRequestType;
+};
+
 class LLMeshHeader
 {
 public:
@@ -235,19 +307,67 @@ public:
 
         m404 = header.has("404");
     }
+private:
+
+    enum EDiskCacheFlags {
+        FLAG_SKIN = 1 << LLModel::NUM_LODS,
+        FLAG_PHYSCONVEX = 1 << (LLModel::NUM_LODS + 1),
+        FLAG_PHYSMESH = 1 << (LLModel::NUM_LODS + 2),
+    };
+public:
+    U32 getFlags()
+    {
+        U32 flags = 0;
+        for (U32 i = 0; i < LLModel::NUM_LODS; i++)
+        {
+            if (mLodInCache[i])
+            {
+                flags |= 1 << i;
+            }
+        }
+        if (mSkinInCache)
+        {
+            flags |= FLAG_SKIN;
+        }
+        if (mPhysicsConvexInCache)
+        {
+            flags |= FLAG_PHYSCONVEX;
+        }
+        if (mPhysicsMeshInCache)
+        {
+            flags |= FLAG_PHYSMESH;
+        }
+        return flags;
+    }
+
+    void setFromFlags(U32 flags)
+    {
+        for (U32 i = 0; i < LLModel::NUM_LODS; i++)
+        {
+            mLodInCache[i] = (flags & (1 << i)) != 0;
+        }
+        mSkinInCache          = (flags & FLAG_SKIN) != 0;
+        mPhysicsConvexInCache = (flags & FLAG_PHYSCONVEX) != 0;
+        mPhysicsMeshInCache   = (flags & FLAG_PHYSMESH) != 0;
+    }
 
     S32 mVersion = -1;
     S32 mSkinOffset = -1;
     S32 mSkinSize = -1;
+    bool mSkinInCache = false;
 
     S32 mPhysicsConvexOffset = -1;
     S32 mPhysicsConvexSize = -1;
+    bool mPhysicsConvexInCache = false;
 
     S32 mPhysicsMeshOffset = -1;
     S32 mPhysicsMeshSize = -1;
+    bool mPhysicsMeshInCache = false;
 
-    S32 mLodOffset[4] = { -1 };
-    S32 mLodSize[4] = { -1 };
+    S32 mLodOffset[LLModel::NUM_LODS] = { -1 };
+    S32 mLodSize[LLModel::NUM_LODS] = { -1 };
+    bool mLodInCache[LLModel::NUM_LODS] = { false };
+    S32 mHeaderSize = -1;
 
     bool m404 = false;
 };
@@ -258,6 +378,7 @@ public:
 
     static std::atomic<S32> sActiveHeaderRequests;
     static std::atomic<S32> sActiveLODRequests;
+    static std::atomic<S32> sActiveSkinRequests;
     static U32 sMaxConcurrentRequests;
     static S32 sRequestLowWater;
     static S32 sRequestHighWater;
@@ -265,10 +386,13 @@ public:
 
     LLMutex*    mMutex;
     LLMutex*    mHeaderMutex;
+    LLMutex*    mLoadedMutex;
+    LLMutex*    mPendingMutex;
+    LLMutex*    mSkinMapMutex;
     LLCondition* mSignal;
 
     //map of known mesh headers
-    typedef boost::unordered_map<LLUUID, std::pair<U32, LLMeshHeader>> mesh_header_map; // pair is header_size and data
+    typedef boost::unordered_map<LLUUID, LLMeshHeader> mesh_header_map; // pair is header_size and data
     mesh_header_map mMeshHeader;
 
     class HeaderRequest : public RequestStats
@@ -292,19 +416,10 @@ public:
     public:
         LLVolumeParams  mMeshParams;
         S32 mLOD;
-        F32 mScore;
 
         LODRequest(const LLVolumeParams&  mesh_params, S32 lod)
-            : RequestStats(), mMeshParams(mesh_params), mLOD(lod), mScore(0.f)
-        {
-        }
-    };
-
-    struct CompareScoreGreater
-    {
-        bool operator()(const LODRequest& lhs, const LODRequest& rhs)
+            : RequestStats(), mMeshParams(mesh_params), mLOD(lod)
         {
-            return lhs.mScore > rhs.mScore; // greatest = first
         }
     };
 
@@ -369,7 +484,7 @@ public:
     std::deque<LoadedMesh> mLoadedQ;
 
     //map of pending header requests and currently desired LODs
-    typedef std::unordered_map<LLUUID, std::vector<S32> > pending_lod_map;
+    typedef std::unordered_map<LLUUID, std::array<S32, LLModel::NUM_LODS> > pending_lod_map;
     pending_lod_map mPendingLOD;
 
     // map of mesh ID to skin info (mirrors LLMeshRepository::mSkinMap)
@@ -379,6 +494,8 @@ public:
 
     // workqueue for processing generic requests
     LL::WorkQueue mWorkQueue;
+    // lods have their own thread due to costly cacheOptimize() calls
+    std::unique_ptr<LL::ThreadPool> mMeshThreadPool;
 
     // llcorehttp library interface objects.
     LLCore::HttpStatus                  mHttpStatus;
@@ -402,16 +519,16 @@ public:
     void lockAndLoadMeshLOD(const LLVolumeParams& mesh_params, S32 lod);
     void loadMeshLOD(const LLVolumeParams& mesh_params, S32 lod);
 
-    bool fetchMeshHeader(const LLVolumeParams& mesh_params, bool can_retry = true);
-    bool fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod, bool can_retry = true);
-    EMeshProcessingResult headerReceived(const LLVolumeParams& mesh_params, U8* data, S32 data_size);
+    bool fetchMeshHeader(const LLVolumeParams& mesh_params);
+    bool fetchMeshLOD(const LLVolumeParams& mesh_params, S32 lod);
+    EMeshProcessingResult headerReceived(const LLVolumeParams& mesh_params, U8* data, S32 data_size, U32 flags = 0);
     EMeshProcessingResult lodReceived(const LLVolumeParams& mesh_params, S32 lod, U8* data, S32 data_size);
     bool skinInfoReceived(const LLUUID& mesh_id, U8* data, S32 data_size);
     bool decompositionReceived(const LLUUID& mesh_id, U8* data, S32 data_size);
     EMeshProcessingResult physicsShapeReceived(const LLUUID& mesh_id, U8* data, S32 data_size);
-    bool hasPhysicsShapeInHeader(const LLUUID& mesh_id);
-    bool hasSkinInfoInHeader(const LLUUID& mesh_id);
-    bool hasHeader(const LLUUID& mesh_id);
+    bool hasPhysicsShapeInHeader(const LLUUID& mesh_id) const;
+    bool hasSkinInfoInHeader(const LLUUID& mesh_id) const;
+    bool hasHeader(const LLUUID& mesh_id) const;
 
     void notifyLoadedMeshes();
     S32 getActualMeshLOD(const LLVolumeParams& mesh_params, S32 lod);
@@ -422,7 +539,7 @@ public:
 
     //send request for skin info, returns true if header info exists
     //  (should hold onto mesh_id and try again later if header info does not exist)
-    bool fetchMeshSkinInfo(const LLUUID& mesh_id, bool can_retry = true);
+    bool fetchMeshSkinInfo(const LLUUID& mesh_id);
 
     //send request for decomposition, returns true if header info exists
     //  (should hold onto mesh_id and try again later if header info does not exist)
@@ -436,6 +553,8 @@ public:
     static void decActiveLODRequests();
     static void incActiveHeaderRequests();
     static void decActiveHeaderRequests();
+    static void incActiveSkinRequests();
+    static void decActiveSkinRequests();
 
     // Set the caps strings and preferred version for constructing
     // mesh fetch URLs.
@@ -456,6 +575,14 @@ private:
     LLCore::HttpHandle getByteRange(const std::string & url,
                                     size_t offset, size_t len,
                                     const LLCore::HttpHandler::ptr_t &handler);
+
+    // Mutex: acquires mPendingMutex, mMutex and mHeaderMutex as needed
+    void loadMeshLOD(const LLUUID &mesh_id, const LLVolumeParams& mesh_params, S32 lod);
+
+    // Threads:  Repo thread only
+    U8* getDiskCacheBuffer(S32 size);
+    S32 mDiskCacheBufferSize = 0;
+    U8* mDiskCacheBuffer = nullptr;
 };
 
 
@@ -568,35 +695,35 @@ public:
     bool init(const LLMeshHeader& header);
 
     // Size for given LOD
-    S32 getSizeByLOD(S32 lod);
+    S32 getSizeByLOD(S32 lod) const;
 
     // Sum of all LOD sizes.
-    S32 getSizeTotal();
+    S32 getSizeTotal() const;
 
     // Estimated triangle counts for the given LOD.
-    F32 getEstTrisByLOD(S32 lod);
+    F32 getEstTrisByLOD(S32 lod) const;
 
     // Estimated triangle counts for the largest LOD. Typically this
     // is also the "high" LOD, but not necessarily.
-    F32 getEstTrisMax();
+    F32 getEstTrisMax() const;
 
     // Triangle count as computed by original streaming cost
     // formula. Triangles in each LOD are weighted based on how
     // frequently they will be seen.
     // This was called "unscaled_value" in the original getStreamingCost() functions.
-    F32 getRadiusWeightedTris(F32 radius);
+    F32 getRadiusWeightedTris(F32 radius) const;
 
     // Triangle count used by triangle-based cost formula. Based on
     // triangles in highest LOD plus potentially partial charges for
     // lower LODs depending on complexity.
-    F32 getEstTrisForStreamingCost();
+    F32 getEstTrisForStreamingCost() const;
 
     // Streaming cost. This should match the server-side calculation
     // for the corresponding volume.
-    F32 getRadiusBasedStreamingCost(F32 radius);
+    F32 getRadiusBasedStreamingCost(F32 radius) const;
 
     // New streaming cost formula, currently only used for animated objects.
-    F32 getTriangleBasedStreamingCost();
+    F32 getTriangleBasedStreamingCost() const;
 
 private:
     // From the "size" field of the mesh header. LOD 0=lowest, 3=highest.
@@ -620,12 +747,12 @@ public:
     static U32 sLODPending;
     static U32 sLODProcessing;
     static U32 sCacheBytesRead;
-    static U32 sCacheBytesWritten;
+    static std::atomic<U32> sCacheBytesWritten;
     static U32 sCacheBytesHeaders;
     static U32 sCacheBytesSkins;
     static U32 sCacheBytesDecomps;
     static U32 sCacheReads;
-    static U32 sCacheWrites;
+    static std::atomic<U32> sCacheWrites;
     static U32 sMaxLockHoldoffs;                // Maximum sequential locking failures
 
     static LLDeadmanTimer sQuiescentTimer;      // Time-to-complete-mesh-downloads after significant events
@@ -646,11 +773,11 @@ public:
 
     void unregisterMesh(LLVOVolume* volume);
     //mesh management functions
-    S32 loadMesh(LLVOVolume* volume, const LLVolumeParams& mesh_params, S32 detail = 0, S32 last_lod = -1);
+    S32 loadMesh(LLVOVolume* volume, const LLVolumeParams& mesh_params, S32 new_lod = 0, S32 last_lod = -1);
 
     void notifyLoadedMeshes();
-    void notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVolume* volume);
-    void notifyMeshUnavailable(const LLVolumeParams& mesh_params, S32 lod);
+    void notifyMeshLoaded(const LLVolumeParams& mesh_params, LLVolume* volume, S32 lod);
+    void notifyMeshUnavailable(const LLVolumeParams& mesh_params, S32 request_lod, S32 volume_lod);
     void notifySkinInfoReceived(LLMeshSkinInfo* info);
     void notifySkinInfoUnavailable(const LLUUID& info);
     void notifyDecompositionReceived(LLModel::Decomposition* info);
@@ -662,7 +789,7 @@ public:
     void fetchPhysicsShape(const LLUUID& mesh_id);
     bool hasPhysicsShape(const LLUUID& mesh_id);
     bool hasSkinInfo(const LLUUID& mesh_id);
-    bool hasHeader(const LLUUID& mesh_id);
+    bool hasHeader(const LLUUID& mesh_id) const;
 
     void buildHull(const LLVolumeParams& params, S32 detail);
     void buildPhysicsMesh(LLModel::Decomposition& decomp);
@@ -676,7 +803,7 @@ public:
                      LLHandle<LLWholeModelFeeObserver> fee_observer= (LLHandle<LLWholeModelFeeObserver>()),
                      LLHandle<LLWholeModelUploadObserver> upload_observer = (LLHandle<LLWholeModelUploadObserver>()));
 
-    S32 getMeshSize(const LLUUID& mesh_id, S32 lod);
+    S32 getMeshSize(const LLUUID& mesh_id, S32 lod) const;
 
     // Quiescent timer management, main thread only.
     static void metricsStart();
@@ -684,7 +811,7 @@ public:
     static void metricsProgress(unsigned int count);
     static void metricsUpdate();
 
-    typedef boost::unordered_map<LLUUID, std::vector<LLVOVolume*> > mesh_load_map;
+    typedef std::unordered_map<LLUUID, std::vector<LLVOVolume*> > mesh_load_map;
     mesh_load_map mLoadingMeshes[4];
 
     typedef std::unordered_map<LLUUID, LLPointer<LLMeshSkinInfo>> skin_map;
@@ -695,15 +822,13 @@ public:
 
     LLMutex*                    mMeshMutex;
 
-    std::vector<LLMeshRepoThread::LODRequest> mPendingRequests;
+    typedef std::vector <std::unique_ptr<PendingRequestBase> > pending_requests_vec;
+    pending_requests_vec mPendingRequests;
 
     //list of mesh ids awaiting skin info
-    typedef boost::unordered_map<LLUUID, std::vector<LLVOVolume*> > skin_load_map;
+    typedef std::unordered_map<LLUUID, std::vector<LLVOVolume*> > skin_load_map;
     skin_load_map mLoadingSkins;
 
-    //list of mesh ids that need to send skin info fetch requests
-    std::queue<LLUUID> mPendingSkinRequests;
-
     //list of mesh ids awaiting decompositions
     std::unordered_set<LLUUID> mLoadingDecompositions;
 
diff --git a/indra/newview/llstartup.cpp b/indra/newview/llstartup.cpp
index 6c3f3849ce..eb0e9ef4bc 100644
--- a/indra/newview/llstartup.cpp
+++ b/indra/newview/llstartup.cpp
@@ -297,16 +297,62 @@ void callback_cache_name(const LLUUID& id, const std::string& full_name, bool is
 // exported functionality
 //
 
+void do_startup_frame()
+{
+    // Until after STATE_AGENT_SEND we don't get very many UDP packets to poll the socket,
+    // and after STATE_PRECACHE the LLAppViewer::idleNetwork() will do UDP processing,
+    // so we only bother to process between those two states.
+    EStartupState state = LLStartUp::getStartupState();
+    if (state > STATE_AGENT_SEND && state < STATE_PRECACHE)
+    {
+        // drain the UDP socket...
+        U64 t0 = totalTime();
+        constexpr U64 MAX_STARTUP_FRAME_TIME = 2000; // usec
+        constexpr U64 MAX_STARTUP_FRAME_MESSAGES = 100;
+        S32 num_messages = 0;
+        bool needs_drain = false;
+        LockMessageChecker lmc(gMessageSystem);
+        while (lmc.checkAllMessages(gFrameCount, gServicePump))
+        {
+            if (gDoDisconnect)
+            {
+                // We're disconnecting, don't process any more messages from the server
+                // We're usually disconnecting due to either network corruption or a
+                // server going down, so this is OK.
+                break;
+            }
+            if (++num_messages >= MAX_STARTUP_FRAME_MESSAGES
+                || (totalTime() - t0) > MAX_STARTUP_FRAME_TIME)
+            {
+                needs_drain = true;
+                break;
+            }
+        }
+        if (needs_drain || gMessageSystem->mPacketRing.getNumBufferedPackets() > 0)
+        {
+             gMessageSystem->drainUdpSocket();
+        }
+        lmc.processAcks();
+    }
+    // ...then call display_startup()
+    display_startup();
+}
+
 void pump_idle_startup_network(void)
 {
+    // while there are message to process:
+    //     process one then call display_startup()
+    S32 num_messages = 0;
     {
         LockMessageChecker lmc(gMessageSystem);
         while (lmc.checkAllMessages(gFrameCount, gServicePump))
         {
             display_startup();
+            ++num_messages;
         }
         lmc.processAcks();
     }
+    // finally call one last display_startup()
     display_startup();
 }
 
@@ -623,21 +669,6 @@ bool idle_startup()
 
             F32 dropPercent = gSavedSettings.getF32("PacketDropPercentage");
             msg->mPacketRing.setDropPercentage(dropPercent);
-
-            F32 inBandwidth = gSavedSettings.getF32("InBandwidth");
-            F32 outBandwidth = gSavedSettings.getF32("OutBandwidth");
-            if (inBandwidth != 0.f)
-            {
-                LL_DEBUGS("AppInit") << "Setting packetring incoming bandwidth to " << inBandwidth << LL_ENDL;
-                msg->mPacketRing.setUseInThrottle(true);
-                msg->mPacketRing.setInBandwidth(inBandwidth);
-            }
-            if (outBandwidth != 0.f)
-            {
-                LL_DEBUGS("AppInit") << "Setting packetring outgoing bandwidth to " << outBandwidth << LL_ENDL;
-                msg->mPacketRing.setUseOutThrottle(true);
-                msg->mPacketRing.setOutBandwidth(outBandwidth);
-            }
         }
 
         LL_INFOS("AppInit") << "Message System Initialized." << LL_ENDL;
@@ -743,7 +774,7 @@ bool idle_startup()
         LL_DEBUGS("AppInit") << "STATE_BROWSER_INIT" << LL_ENDL;
         std::string msg = LLTrans::getString("LoginInitializingBrowser");
         set_startup_status(0.03f, msg.c_str(), gAgent.mMOTD.c_str());
-        display_startup();
+        do_startup_frame();
         // LLViewerMedia::initBrowser();
         LLStartUp::setStartupState( STATE_LOGIN_SHOW );
         return false;
@@ -808,7 +839,7 @@ bool idle_startup()
                     LL_DEBUGS("AppInit") << "FirstLoginThisInstall off" << LL_ENDL;
                 }
             }
-            display_startup();
+            do_startup_frame();
             LLStartUp::setStartupState( STATE_LOGIN_WAIT );     // Wait for user input
         }
         else
@@ -839,7 +870,7 @@ bool idle_startup()
         }
         LL_DEBUGS("AppInit") << "PeekMessage processed" << LL_ENDL;
 #endif
-        display_startup();
+        do_startup_frame();
         timeout.reset();
         return false;
     }
@@ -854,7 +885,7 @@ bool idle_startup()
         // Don't do anything.  Wait for the login view to call the login_callback,
         // which will push us to the next state.
 
-        // display() function will be the one to run display_startup()
+        // display() function will be the one to run do_startup_frame()
         // Sleep so we don't spin the CPU
         ms_sleep(1);
         return false;
@@ -1036,7 +1067,7 @@ bool idle_startup()
         auth_desc = LLTrans::getString("LoginInProgress");
         set_startup_status(progress, auth_desc, auth_message);
         progress += 0.02f;
-        display_startup();
+        do_startup_frame();
 
         // Setting initial values...
         LLLoginInstance* login = LLLoginInstance::getInstance();
@@ -1073,7 +1104,7 @@ bool idle_startup()
         emsg << LLTrans::getString("LoginFailedHeader") << "\n";
         if(LLLoginInstance::getInstance()->authFailure())
         {
-            LL_INFOS("LLStartup") << "Login failed, LLLoginInstance::getResponse(): "
+            LL_INFOS("LLStartUp") << "Login failed, LLLoginInstance::getResponse(): "
                                   << LLLoginInstance::getInstance()->getResponse() << LL_ENDL;
             LLSD response = LLLoginInstance::getInstance()->getResponse();
             // Still have error conditions that may need some
@@ -1145,7 +1176,7 @@ bool idle_startup()
                 // If optional was skipped this case shouldn't
                 // be reached.
 
-                LL_INFOS("LLStartup") << "Forcing a quit due to update." << LL_ENDL;
+                LL_INFOS("LLStartUp") << "Forcing a quit due to update." << LL_ENDL;
                 LLLoginInstance::getInstance()->disconnect();
                 LLAppViewer::instance()->forceQuit();
             }
@@ -1173,7 +1204,7 @@ bool idle_startup()
                         }
                         catch (LLCertException &cert_exception)
                         {
-                            LL_WARNS("LLStartup", "SECAPI") << "Caught " << cert_exception.what() << " certificate expception on getCertificate("<< response["certificate"] << ")" << LL_ENDL;
+                            LL_WARNS("LLStartUp", "SECAPI") << "Caught " << cert_exception.what() << " certificate expception on getCertificate("<< response["certificate"] << ")" << LL_ENDL;
                             LLSD args;
                             args["REASON"] = LLTrans::getString(cert_exception.what());
 
@@ -1225,7 +1256,7 @@ bool idle_startup()
                         // notificatioin message.
                         LLSD args;
                         args["ERROR_MESSAGE"] = emsg.str();
-                        LL_INFOS("LLStartup") << "Notification: " << args << LL_ENDL;
+                        LL_INFOS("LLStartUp") << "Notification: " << args << LL_ENDL;
                         LLNotificationsUtil::add("ErrorMessage", args, LLSD(), login_alert_done);
                     }
                 }
@@ -1248,7 +1279,7 @@ bool idle_startup()
             {
                 LLSD args;
                 args["ERROR_MESSAGE"] = emsg.str();
-                LL_INFOS("LLStartup") << "Notification: " << args << LL_ENDL;
+                LL_INFOS("LLStartUp") << "Notification: " << args << LL_ENDL;
                 LLNotificationsUtil::add("ErrorMessage", args, LLSD(), login_alert_done);
                 transition_back_to_login_panel(emsg.str());
                 show_connect_box = true;
@@ -1264,71 +1295,71 @@ bool idle_startup()
     if (STATE_WORLD_INIT == LLStartUp::getStartupState())
     {
         set_startup_status(0.30f, LLTrans::getString("LoginInitializingWorld"), gAgent.mMOTD);
-        display_startup();
+        do_startup_frame();
         // We should have an agent id by this point.
         llassert(!(gAgentID == LLUUID::null));
 
         // Finish agent initialization.  (Requires gSavedSettings, builds camera)
         gAgent.init();
-        display_startup();
+        do_startup_frame();
         gAgentCamera.init();
-        display_startup();
-        display_startup();
+        do_startup_frame();
+        do_startup_frame();
 
         // Since we connected, save off the settings so the user doesn't have to
         // type the name/password again if we crash.
         gSavedSettings.saveToFile(gSavedSettings.getString("ClientSettingsFile"), true);
         LLUIColorTable::instance().saveUserSettings();
 
-        display_startup();
+        do_startup_frame();
 
         //
         // Initialize classes w/graphics stuff.
         //
         LLSurface::initClasses();
-        display_startup();
+        do_startup_frame();
 
-        display_startup();
+        do_startup_frame();
 
         LLDrawable::initClass();
-        display_startup();
+        do_startup_frame();
 
         // init the shader managers
         LLPostProcess::initClass();
-        display_startup();
+        do_startup_frame();
 
         LLAvatarAppearance::initClass("avatar_lad.xml","avatar_skeleton.xml");
-        display_startup();
+        do_startup_frame();
 
         LLViewerObject::initVOClasses();
-        display_startup();
+        do_startup_frame();
 
         // Initialize all our tools.  Must be done after saved settings loaded.
         // NOTE: This also is where gToolMgr used to be instantiated before being turned into a singleton.
         LLToolMgr::getInstance()->initTools();
-        display_startup();
+        do_startup_frame();
 
         // Pre-load floaters, like the world map, that are slow to spawn
         // due to XML complexity.
         gViewerWindow->initWorldUI();
 
-        display_startup();
+        do_startup_frame();
 
         // This is where we used to initialize gWorldp. Original comment said:
         // World initialization must be done after above window init
 
         // User might have overridden far clip
         LLWorld::getInstance()->setLandFarClip(gAgentCamera.mDrawDistance);
-        display_startup();
+        do_startup_frame();
         // Before we create the first region, we need to set the agent's mOriginGlobal
         // This is necessary because creating objects before this is set will result in a
         // bad mPositionAgent cache.
 
         gAgent.initOriginGlobal(from_region_handle(gFirstSimHandle));
-        display_startup();
+        do_startup_frame();
 
         LLWorld::getInstance()->addRegion(gFirstSimHandle, gFirstSim);
-        display_startup();
+        do_startup_frame();
 
         LLViewerRegion *regionp = LLWorld::getInstance()->getRegionFromHandle(gFirstSimHandle);
         LL_INFOS("AppInit") << "Adding initial simulator " << regionp->getOriginGlobal() << LL_ENDL;
@@ -1337,18 +1368,18 @@ bool idle_startup()
         << gFirstSimSeedCap << LL_ENDL;
         regionp->setSeedCapability(gFirstSimSeedCap);
         LL_DEBUGS("AppInit") << "Waiting for seed grant ...." << LL_ENDL;
-        display_startup();
+        do_startup_frame();
         // Set agent's initial region to be the one we just created.
         gAgent.setRegion(regionp);
-        display_startup();
+        do_startup_frame();
         // Set agent's initial position, which will be read by LLVOAvatar when the avatar
         // object is created.  I think this must be done after setting the region.  JC
         gAgent.setPositionAgent(agent_start_position_region);
 
-        display_startup();
+        do_startup_frame();
         LLStartUp::initExperiences();
 
-        display_startup();
+        do_startup_frame();
 
         // If logging should be enebled, turns it on and loads history from disk
         // Note: does not happen on init of singleton because preferences can use
@@ -1369,7 +1400,7 @@ bool idle_startup()
     {
         LLStartUp::multimediaInit();
         LLStartUp::setStartupState( STATE_FONT_INIT );
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1378,7 +1409,7 @@ bool idle_startup()
     {
         LLStartUp::fontInit();
         LLStartUp::setStartupState( STATE_SEED_GRANTED_WAIT );
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1450,7 +1481,7 @@ bool idle_startup()
                 set_startup_status(0.4f, LLTrans::getString("LoginRequestSeedCapGrant"), gAgent.mMOTD.c_str());
             }
         }
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1461,7 +1492,7 @@ bool idle_startup()
     //---------------------------------------------------------------------
     if (STATE_SEED_CAP_GRANTED == LLStartUp::getStartupState())
     {
-        display_startup();
+        do_startup_frame();
 
         // These textures are not warrantied to be cached, so needs
         // to hapen with caps granted
@@ -1470,9 +1501,9 @@ bool idle_startup()
         // will init images, should be done with caps, but before gSky.init()
         LLEnvironment::getInstance()->initSingleton();
 
-        display_startup();
+        do_startup_frame();
         update_texture_fetch();
-        display_startup();
+        do_startup_frame();
 
         if ( gViewerWindow != NULL)
         {   // This isn't the first logon attempt, so show the UI
@@ -1480,15 +1511,15 @@ bool idle_startup()
         }
         gLoginMenuBarView->setVisible( false );
         gLoginMenuBarView->setEnabled( false );
-        display_startup();
+        do_startup_frame();
 
         // direct logging to the debug console's line buffer
         LLError::logToFixedBuffer(gDebugView->mDebugConsolep);
-        display_startup();
+        do_startup_frame();
 
         // set initial visibility of debug console
         gDebugView->mDebugConsolep->setVisible(gSavedSettings.getBOOL("ShowDebugConsole"));
-        display_startup();
+        do_startup_frame();
 
         //
         // Set message handlers
@@ -1497,23 +1528,23 @@ bool idle_startup()
 
         // register callbacks for messages. . . do this after initial handshake to make sure that we don't catch any unwanted
         register_viewer_callbacks(gMessageSystem);
-        display_startup();
+        do_startup_frame();
 
         // Debugging info parameters
         gMessageSystem->setMaxMessageTime( 0.5f );          // Spam if decoding all msgs takes more than 500 ms
-        display_startup();
+        do_startup_frame();
 
         #ifndef LL_RELEASE_FOR_DOWNLOAD
             gMessageSystem->setTimeDecodes( true );             // Time the decode of each msg
             gMessageSystem->setTimeDecodesSpamThreshold( 0.05f );  // Spam if a single msg takes over 50ms to decode
         #endif
-        display_startup();
+        do_startup_frame();
 
         gXferManager->registerCallbacks(gMessageSystem);
-        display_startup();
+        do_startup_frame();
 
         LLStartUp::initNameCache();
-        display_startup();
+        do_startup_frame();
 
         // update the voice settings *after* gCacheName initialization
         // so that we can construct voice UI that relies on the name cache
@@ -1521,7 +1552,7 @@ bool idle_startup()
         {
             LLVoiceClient::getInstance()->updateSettings();
         }
-        display_startup();
+        do_startup_frame();
 
         // create a container's instance for start a controlling conversation windows
         // by the voice's events
@@ -1542,12 +1573,12 @@ bool idle_startup()
         // register null callbacks for audio until the audio system is initialized
         gMessageSystem->setHandlerFuncFast(_PREHASH_SoundTrigger, null_message_callback, NULL);
         gMessageSystem->setHandlerFuncFast(_PREHASH_AttachedSound, null_message_callback, NULL);
-        display_startup();
+        do_startup_frame();
 
         //reset statistics
         LLViewerStats::instance().resetStats();
 
-        display_startup();
+        do_startup_frame();
         //
         // Set up region and surface defaults
         //
@@ -1572,7 +1603,7 @@ bool idle_startup()
         LLViewerCamera::getInstance()->setAspect(gViewerWindow->getWorldViewAspectRatio());
         // Initialize FOV
         LLViewerCamera::getInstance()->setDefaultFOV(gSavedSettings.getF32("CameraAngle"));
-        display_startup();
+        do_startup_frame();
 
         // Move agent to starting location. The position handed to us by
         // the space server is in global coordinates, but the agent frame
@@ -1583,7 +1614,7 @@ bool idle_startup()
         gAgent.resetAxes(gAgentStartLookAt);
         gAgentCamera.stopCameraAnimation();
         gAgentCamera.resetCamera();
-        display_startup();
+        do_startup_frame();
 
         // Initialize global class data needed for surfaces (i.e. textures)
         LL_DEBUGS("AppInit") << "Initializing sky..." << LL_ENDL;
@@ -1594,7 +1625,7 @@ bool idle_startup()
 
         LLGLState::checkStates();
 
-        display_startup();
+        do_startup_frame();
 
         LL_DEBUGS("AppInit") << "Decoding images..." << LL_ENDL;
         // For all images pre-loaded into viewer cache, init
@@ -1608,12 +1639,12 @@ bool idle_startup()
         {
             F32 frac = (F32)i / (F32)DECODE_TIME_SEC;
             set_startup_status(0.45f + frac*0.1f, LLTrans::getString("LoginDecodingImages"), gAgent.mMOTD);
-            display_startup();
+            do_startup_frame();
             gTextureList.decodeAllImages(1.f);
         }
         LLStartUp::setStartupState( STATE_WORLD_WAIT );
 
-        display_startup();
+        do_startup_frame();
 
         // JC - Do this as late as possible to increase likelihood Purify
         // will run.
@@ -1642,7 +1673,7 @@ bool idle_startup()
             NULL);
 
         timeout.reset();
-        display_startup();
+        do_startup_frame();
 
         return false;
     }
@@ -1669,7 +1700,7 @@ bool idle_startup()
     {
         LL_DEBUGS("AppInit") << "Connecting to region..." << LL_ENDL;
         set_startup_status(0.60f, LLTrans::getString("LoginConnectingToRegion"), gAgent.mMOTD);
-        display_startup();
+        do_startup_frame();
         // register with the message system so it knows we're
         // expecting this message
         LLMessageSystem* msg = gMessageSystem;
@@ -1683,7 +1714,7 @@ bool idle_startup()
             gAssetStorage->setUpstream(regionp->getHost());
             gCacheName->setUpstream(regionp->getHost());
         }
-        display_startup();
+        do_startup_frame();
 
         // Create login effect
         // But not on first login, because you can't see your avatar then
@@ -1698,7 +1729,7 @@ bool idle_startup()
         LLStartUp::setStartupState( STATE_AGENT_WAIT );     // Go to STATE_AGENT_WAIT
 
         timeout.reset();
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1707,35 +1738,13 @@ bool idle_startup()
     //---------------------------------------------------------------------
     if (STATE_AGENT_WAIT == LLStartUp::getStartupState())
     {
-        {
-            LockMessageChecker lmc(gMessageSystem);
-            while (lmc.checkAllMessages(gFrameCount, gServicePump))
-            {
-                if (gAgentMovementCompleted)
-                {
-                    // Sometimes we have more than one message in the
-                    // queue. break out of this loop and continue
-                    // processing. If we don't, then this could skip one
-                    // or more login steps.
-                    break;
-                }
-                else
-                {
-                    LL_DEBUGS("AppInit") << "Awaiting AvatarInitComplete, got "
-                                         << gMessageSystem->getMessageName() << LL_ENDL;
-                }
-                display_startup();
-            }
-            lmc.processAcks();
-        }
-
-        display_startup();
+        do_startup_frame();
 
         if (gAgentMovementCompleted)
         {
             LLStartUp::setStartupState( STATE_INVENTORY_SEND );
         }
-        display_startup();
+        do_startup_frame();
 
         if (!gAgentMovementCompleted && timeout.getElapsedTimeF32() > STATE_AGENT_WAIT_TIMEOUT)
         {
@@ -1768,7 +1777,7 @@ bool idle_startup()
     if (STATE_INVENTORY_SEND == LLStartUp::getStartupState())
     {
         LL_PROFILE_ZONE_NAMED("State inventory send")
-        display_startup();
+        do_startup_frame();
 
         // request mute list
         LL_INFOS() << "Requesting Mute List" << LL_ENDL;
@@ -1778,12 +1787,12 @@ bool idle_startup()
         LL_INFOS() << "Requesting Money Balance" << LL_ENDL;
         LLStatusBar::sendMoneyBalanceRequest();
 
-        display_startup();
+        do_startup_frame();
 
         // Inform simulator of our language preference
         LLAgentLanguage::update();
 
-        display_startup();
+        do_startup_frame();
         // unpack thin inventory
         LLSD response = LLLoginInstance::getInstance()->getResponse();
         //bool dump_buffer = false;
@@ -1798,7 +1807,7 @@ bool idle_startup()
                 gInventory.setLibraryRootFolderID(id.asUUID());
             }
         }
-        display_startup();
+        do_startup_frame();
 
         LLSD inv_lib_owner = response["inventory-lib-owner"];
         if(inv_lib_owner.isDefined())
@@ -1810,9 +1819,9 @@ bool idle_startup()
                 gInventory.setLibraryOwnerID(LLUUID(id.asUUID()));
             }
         }
-        display_startup();
+        do_startup_frame();
         LLStartUp::setStartupState(STATE_INVENTORY_SKEL);
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1831,7 +1840,7 @@ bool idle_startup()
                 LL_WARNS("AppInit") << "Problem loading inventory-skel-lib" << LL_ENDL;
             }
         }
-        display_startup();
+        do_startup_frame();
 
         LLSD inv_skeleton = response["inventory-skeleton"];
         if (inv_skeleton.isDefined())
@@ -1842,9 +1851,9 @@ bool idle_startup()
                 LL_WARNS("AppInit") << "Problem loading inventory-skel-targets" << LL_ENDL;
             }
         }
-        display_startup();
+        do_startup_frame();
         LLStartUp::setStartupState(STATE_INVENTORY_SEND2);
-        display_startup();
+        do_startup_frame();
         return false;
     }
 
@@ -1890,7 +1899,7 @@ bool idle_startup()
                 list[agent_id] = new LLRelationship(given_rights, has_rights, false);
             }
             LLAvatarTracker::instance().addBuddyList(list);
-            display_startup();
+            do_startup_frame();
         }
 
         bool show_hud = false;
@@ -1918,7 +1927,7 @@ bool idle_startup()
                 //}
             }
         }
-        display_startup();
+        do_startup_frame();
 
         // Either we want to show tutorial because this is the first login
         // to a Linden Help Island or the user quit with the tutorial
@@ -1927,21 +1936,21 @@ bool idle_startup()
         {
             LLFloaterReg::showInstance("hud", LLSD(), false);
         }
-        display_startup();
+        do_startup_frame();
 
         LLSD event_notifications = response["event_notifications"];
         if(event_notifications.isDefined())
         {
             gEventNotifier.load(event_notifications);
         }
-        display_startup();
+        do_startup_frame();
 
         LLSD classified_categories = response["classified_categories"];
         if(classified_categories.isDefined())
         {
             LLClassifiedInfo::loadCategories(classified_categories);
         }
-        display_startup();
+        do_startup_frame();
 
         // This method MUST be called before gInventory.findCategoryUUIDForType because of
         // gInventory.mIsAgentInvUsable is set to true in the gInventory.buildParentChildMap.
@@ -1960,7 +1969,7 @@ bool idle_startup()
         LLInventoryModelBackgroundFetch::instance().start();
         gInventory.createCommonSystemCategories();
         LLStartUp::setStartupState(STATE_INVENTORY_CALLBACKS );
-        display_startup();
+        do_startup_frame();
 
         return false;
     }
@@ -1972,7 +1981,7 @@ bool idle_startup()
     {
         if (!LLInventoryModel::isSysFoldersReady())
         {
-            display_startup();
+            do_startup_frame();
             return false;
         }
 
@@ -1997,7 +2006,7 @@ bool idle_startup()
         gInventory.addChangedMask(LLInventoryObserver::ALL, LLUUID::null);
         gInventory.notifyObservers();
 
-        display_startup();
+        do_startup_frame();
 
         // set up callbacks
         LL_INFOS() << "Registering Callbacks" << LL_ENDL;
@@ -2008,18 +2017,18 @@ bool idle_startup()
         LLAvatarTracker::instance().registerCallbacks(msg);
         LL_INFOS() << " Landmark" << LL_ENDL;
         LLLandmark::registerCallbacks(msg);
-        display_startup();
+        do_startup_frame();
 
         // request all group information
         LL_INFOS() << "Requesting Agent Data" << LL_ENDL;
         gAgent.sendAgentDataUpdateRequest();
-        display_startup();
+        do_startup_frame();
         // Create the inventory views
         LL_INFOS() << "Creating Inventory Views" << LL_ENDL;
         LLFloaterReg::getInstance("inventory");
-        display_startup();
+        do_startup_frame();
         LLStartUp::setStartupState( STATE_MISC );
-        display_startup();
+        do_startup_frame();
 
         return false;
     }
@@ -2069,7 +2078,7 @@ bool idle_startup()
             gSavedSettings.setBOOL("ShowStartLocation", true);
         }
 
-        display_startup();
+        do_startup_frame();
 
         // Load stored local environment if needed.
         LLEnvironment::instance().loadFromSettings();
@@ -2077,7 +2086,7 @@ bool idle_startup()
         // *TODO : Uncomment that line once the whole grid migrated to SLM and suppress it from LLAgent::handleTeleportFinished() (llagent.cpp)
         //check_merchant_status();
 
-        display_startup();
+        do_startup_frame();
 
         if (gSavedSettings.getBOOL("HelpFloaterOpen"))
         {
@@ -2085,7 +2094,7 @@ bool idle_startup()
             LLViewerHelp::instance().showTopic("");
         }
 
-        display_startup();
+        do_startup_frame();
 
         // We're successfully logged in.
         gSavedSettings.setBOOL("FirstLoginThisInstall", false);
@@ -2094,12 +2103,12 @@ bool idle_startup()
 
         LLFloaterGridStatus::getInstance()->startGridStatusTimer();
 
-        display_startup();
+        do_startup_frame();
 
-        display_startup();
+        do_startup_frame();
         // JC: Initializing audio requests many sounds for download.
         init_audio();
-        display_startup();
+        do_startup_frame();
 
         // JC: Initialize "active" gestures.  This may also trigger
         // many gesture downloads, if this is the user's first
@@ -2137,7 +2146,7 @@ bool idle_startup()
             LLGestureMgr::instance().startFetch();
         }
         gDisplaySwapBuffers = true;
-        display_startup();
+        do_startup_frame();
 
         LLMessageSystem* msg = gMessageSystem;
         msg->setHandlerFuncFast(_PREHASH_SoundTrigger,              process_sound_trigger);
@@ -2215,10 +2224,10 @@ bool idle_startup()
             }
         }
 
-        display_startup();
+        do_startup_frame();
         //DEV-17797.  get null folder.  Any items found here moved to Lost and Found
         LLInventoryModelBackgroundFetch::instance().findLostItems();
-        display_startup();
+        do_startup_frame();
 
         LLStartUp::setStartupState( STATE_PRECACHE );
         timeout.reset();
@@ -2227,7 +2236,7 @@ bool idle_startup()
 
     if (STATE_PRECACHE == LLStartUp::getStartupState())
     {
-        display_startup();
+        do_startup_frame();
         F32 timeout_frac = timeout.getElapsedTimeF32()/PRECACHING_DELAY;
 
         // We now have an inventory skeleton, so if this is a user's first
@@ -2253,7 +2262,7 @@ bool idle_startup()
             callAfterCOFFetch(set_flags_and_update_appearance);
         }
 
-        display_startup();
+        do_startup_frame();
 
         // wait precache-delay and for agent's avatar or a lot longer.
         if ((timeout_frac > 1.f) && isAgentAvatarValid())
@@ -2278,7 +2287,7 @@ bool idle_startup()
             set_startup_status(0.60f + 0.30f * timeout_frac,
                 LLTrans::getString("LoginPrecaching"),
                     gAgent.mMOTD.c_str());
-            display_startup();
+            do_startup_frame();
         }
 
         return true;
@@ -2306,7 +2315,7 @@ bool idle_startup()
             LLStartUp::setStartupState( STATE_CLEANUP );
         }
 
-        display_startup();
+        do_startup_frame();
 
         if (gAgent.isOutfitChosen() && (wearables_time > MAX_WEARABLES_TIME))
         {
@@ -2347,7 +2356,7 @@ bool idle_startup()
     if (STATE_CLEANUP == LLStartUp::getStartupState())
     {
         set_startup_status(1.0, "", "");
-        display_startup();
+        do_startup_frame();
 
         if (!mBenefitsSuccessfullyInit)
         {
@@ -2368,7 +2377,7 @@ bool idle_startup()
         //gViewerWindow->revealIntroPanel();
         gViewerWindow->setStartupComplete();
         gViewerWindow->setProgressCancelButtonVisible(false);
-        display_startup();
+        do_startup_frame();
 
         // We're not away from keyboard, even though login might have taken
         // a while. JC
@@ -2400,7 +2409,7 @@ bool idle_startup()
         // LLUserAuth::getInstance()->reset();
 
         LLStartUp::setStartupState( STATE_STARTED );
-        display_startup();
+        do_startup_frame();
 
         // Unmute audio if desired and setup volumes.
         // This is a not-uncommon crash site, so surround it with
@@ -2416,7 +2425,7 @@ bool idle_startup()
 
         LLAgentPicksInfo::getInstance()->requestNumberOfPicks();
 
-        display_startup();
+        do_startup_frame();
 
         llassert(LLPathfindingManager::getInstance() != NULL);
         LLPathfindingManager::getInstance()->initSystem();
@@ -3021,9 +3030,7 @@ std::string LLStartUp::startupStateToString(EStartupState state)
 // static
 void LLStartUp::setStartupState( EStartupState state )
 {
-    LL_INFOS("AppInit") << "Startup state changing from " <<
-        getStartupStateString() << " to " <<
-        startupStateToString(state) << LL_ENDL;
+    LL_INFOS("AppInit") << getStartupStateString() << " --> " << startupStateToString(state) << LL_ENDL;
 
     getPhases().stopPhase(getStartupStateString());
     gStartupState = state;
@@ -3087,7 +3094,7 @@ void LLStartUp::multimediaInit()
     LL_DEBUGS("AppInit") << "Initializing Multimedia...." << LL_ENDL;
     std::string msg = LLTrans::getString("LoginInitializingMultimedia");
     set_startup_status(0.42f, msg.c_str(), gAgent.mMOTD.c_str());
-    display_startup();
+    do_startup_frame();
 }
 
 void LLStartUp::fontInit()
@@ -3095,7 +3102,7 @@ void LLStartUp::fontInit()
     LL_DEBUGS("AppInit") << "Initializing fonts...." << LL_ENDL;
     std::string msg = LLTrans::getString("LoginInitializingFonts");
     set_startup_status(0.45f, msg.c_str(), gAgent.mMOTD.c_str());
-    display_startup();
+    do_startup_frame();
 
     LLFontGL::loadDefaultFonts();
 }
@@ -3792,14 +3799,14 @@ bool process_login_success_response()
     {
         // We got an answer from the grid -> use that for map for the current session
         gSavedSettings.setString("CurrentMapServerURL", map_server_url);
-        LL_INFOS("LLStartup") << "map-server-url : we got an answer from the grid : " << map_server_url << LL_ENDL;
+        LL_INFOS("LLStartUp") << "map-server-url : we got an answer from the grid : " << map_server_url << LL_ENDL;
     }
     else
     {
         // No answer from the grid -> use the default setting for current session
         map_server_url = gSavedSettings.getString("MapServerURL");
         gSavedSettings.setString("CurrentMapServerURL", map_server_url);
-        LL_INFOS("LLStartup") << "map-server-url : no map-server-url answer, we use the default setting for the map : " << map_server_url << LL_ENDL;
+        LL_INFOS("LLStartUp") << "map-server-url : no map-server-url answer, we use the default setting for the map : " << map_server_url << LL_ENDL;
     }
 
     // Default male and female avatars allowing the user to choose their avatar on first login.
diff --git a/indra/newview/lltextureview.cpp b/indra/newview/lltextureview.cpp
index bda53f66eb..78d930c05c 100644
--- a/indra/newview/lltextureview.cpp
+++ b/indra/newview/lltextureview.cpp
@@ -642,7 +642,7 @@ void LLGLTexMemBar::draw()
     text = llformat("Mesh: Reqs(Tot/Htp/Big): %u/%u/%u Rtr/Err: %u/%u Cread/Cwrite: %u/%u Low/At/High: %d/%d/%d",
                     LLMeshRepository::sMeshRequestCount, LLMeshRepository::sHTTPRequestCount, LLMeshRepository::sHTTPLargeRequestCount,
                     LLMeshRepository::sHTTPRetryCount, LLMeshRepository::sHTTPErrorCount,
-                    LLMeshRepository::sCacheReads, LLMeshRepository::sCacheWrites,
+                    (U32)LLMeshRepository::sCacheReads, (U32)LLMeshRepository::sCacheWrites,
                     LLMeshRepoThread::sRequestLowWater, LLMeshRepoThread::sRequestWaterLevel, LLMeshRepoThread::sRequestHighWater);
     LLFontGL::getFontMonospace()->renderUTF8(text, 0, 0, v_offset + line_height*2,
                                              text_color, LLFontGL::LEFT, LLFontGL::TOP);
diff --git a/indra/newview/llviewermessage.cpp b/indra/newview/llviewermessage.cpp
index 1d4828fd33..5fd820f91d 100644
--- a/indra/newview/llviewermessage.cpp
+++ b/indra/newview/llviewermessage.cpp
@@ -3147,7 +3147,9 @@ void send_agent_update(bool force_send, bool send_reliable)
     LL_PROFILE_ZONE_SCOPED;
     llassert(!gCubeSnapshot);
 
-    if (gAgent.getTeleportState() != LLAgent::TELEPORT_NONE)
+    LLAgent::ETeleportState tp_state = gAgent.getTeleportState();
+    if (tp_state != LLAgent::TELEPORT_NONE
+        && tp_state != LLAgent::TELEPORT_ARRIVING)
     {
         // We don't care if they want to send an agent update, they're not allowed
         // until the target simulator is ready to receive them
@@ -3322,7 +3324,36 @@ void send_agent_update(bool force_send, bool send_reliable)
     msg->addVector3Fast(_PREHASH_CameraAtAxis, camera_at);
     msg->addVector3Fast(_PREHASH_CameraLeftAxis, LLViewerCamera::getInstance()->getLeftAxis());
     msg->addVector3Fast(_PREHASH_CameraUpAxis, LLViewerCamera::getInstance()->getUpAxis());
-    msg->addF32Fast(_PREHASH_Far, gAgentCamera.mDrawDistance);
+
+    static F32 last_draw_disatance_step = 1024;
+    if (tp_state == LLAgent::TELEPORT_ARRIVING || LLStartUp::getStartupState() < STATE_MISC)
+    {
+        // Inform interest list, prioritize closer area.
+        // Reason: currently server doesn't distance sort attachments, by restricting range
+        // we reduce the number of attachments sent to the viewer, thus prioritizing
+        // closer ones.
+        // Todo: revise and remove once server gets distance sorting.
+        last_draw_disatance_step = llmax((F32)(gAgentCamera.mDrawDistance / 2.f), 50.f);
+        msg->addF32Fast(_PREHASH_Far, last_draw_disatance_step);
+    }
+    else if (last_draw_disatance_step < gAgentCamera.mDrawDistance)
+    {
+        static LLFrameTimer last_step_time;
+        if (last_step_time.getElapsedTimeF32() > 1.f)
+        {
+            // gradually increase draw distance
+            // Idealy this should be not per second, but based on how loaded
+            // mesh thread is, but hopefully this is temporary.
+            last_step_time.reset();
+            F32 step = gAgentCamera.mDrawDistance * 0.1f;
+            last_draw_disatance_step = llmin(last_draw_disatance_step + step, gAgentCamera.mDrawDistance);
+        }
+        msg->addF32Fast(_PREHASH_Far, last_draw_disatance_step);
+    }
+    else
+    {
+        msg->addF32Fast(_PREHASH_Far, gAgentCamera.mDrawDistance);
+    }
 
     msg->addU32Fast(_PREHASH_ControlFlags, control_flags);
 
diff --git a/indra/newview/llviewertexturelist.cpp b/indra/newview/llviewertexturelist.cpp
index 12543089c9..f5bbff144b 100644
--- a/indra/newview/llviewertexturelist.cpp
+++ b/indra/newview/llviewertexturelist.cpp
@@ -30,6 +30,7 @@
 
 #include "llviewertexturelist.h"
 
+#include "llagent.h"
 #include "llgl.h" // fot gathering stats from GL
 #include "llimagegl.h"
 #include "llimagebmp.h"
@@ -815,10 +816,19 @@ void LLViewerTextureList::updateImages(F32 max_time)
             clearFetchingRequests();
             gPipeline.clearRebuildGroups();
             cleared = true;
+            return;
         }
-        return;
+        // ARRIVING is a delay to let things decode, cache and process,
+        // so process textures like normal despite gTeleportDisplay
+        if (gAgent.getTeleportState() != LLAgent::TELEPORT_ARRIVING)
+        {
+            return;
+        }
+    }
+    else
+    {
+        cleared = false;
     }
-    cleared = false;
 
     LLAppViewer::getTextureFetch()->setTextureBandwidth((F32)LLTrace::get_frame_recording().getPeriodMeanPerSec(LLStatViewer::TEXTURE_NETWORK_DATA_RECEIVED).value());
 
diff --git a/indra/newview/llvoavatar.cpp b/indra/newview/llvoavatar.cpp
index 7bf9c88b99..40312b7f4e 100644
--- a/indra/newview/llvoavatar.cpp
+++ b/indra/newview/llvoavatar.cpp
@@ -682,7 +682,7 @@ LLVOAvatar::LLVOAvatar(const LLUUID& id,
     mFullyLoaded(false),
     mPreviousFullyLoaded(false),
     mFullyLoadedInitialized(false),
-    mLastCloudAttachmentCount(0),
+    mLastCloudAttachmentCount(-1),
     mVisualComplexity(VISUAL_COMPLEXITY_UNKNOWN),
     mLoadedCallbacksPaused(false),
     mLoadedCallbackTextures(0),
@@ -2794,7 +2794,10 @@ void LLVOAvatar::idleUpdate(LLAgent &agent, const F64 &time)
     }
 
     // attach objects that were waiting for a drawable
-    lazyAttach();
+    if (!mPendingAttachment.empty())
+    {
+        lazyAttach();
+    }
 
     // animate the character
     // store off last frame's root position to be consistent with camera position
@@ -3226,6 +3229,7 @@ void LLVOAvatar::idleUpdateLoadingEffect()
             if (mFirstFullyVisible)
             {
                 mFirstFullyVisible = false;
+                mLastCloudAttachmentCount = (S32)mSimAttachments.size();
                 mFirstDecloudTime = mFirstAppearanceMessageTimer.getElapsedTimeF32();
                 if (isSelf())
                 {
@@ -7710,6 +7714,64 @@ void LLVOAvatar::cleanupAttachedMesh( LLViewerObject* pVO )
     }
 }
 
+bool check_object_for_mesh_loading(LLViewerObject* objectp)
+{
+    if (!objectp || !objectp->getVolume())
+    {
+        return false;
+    }
+    LLVolume* volp = objectp->getVolume();
+    const LLUUID& mesh_id = volp->getParams().getSculptID();
+    if (mesh_id.isNull())
+    {
+        // No mesh nor skin info needed
+        return false;
+    }
+
+    if (volp->isMeshAssetUnavaliable())
+    {
+        // Mesh failed to load, do not expect it
+        return false;
+    }
+
+    if (!objectp->mDrawable)
+    {
+        return false;
+    }
+
+    LLVOVolume* pvobj = objectp->mDrawable->getVOVolume();
+    if (pvobj)
+    {
+        if (!pvobj->isMesh())
+        {
+            // Not a mesh
+            return false;
+        }
+
+        if (!volp->isMeshAssetLoaded())
+        {
+            // Waiting for mesh
+            return true;
+        }
+
+        const LLMeshSkinInfo* skin_data = pvobj->getSkinInfo();
+        if (skin_data)
+        {
+            // Skin info present, done
+            return false;
+        }
+
+        if (pvobj->isSkinInfoUnavaliable())
+        {
+            // Load failed or info not present, don't expect it
+            return false;
+        }
+    }
+
+    // object is not ready
+    return true;
+}
+
 bool LLVOAvatar::hasPendingAttachedMeshes()
 {
     for (attachment_map_t::iterator iter = mAttachmentPoints.begin();
@@ -7724,62 +7786,20 @@ bool LLVOAvatar::hasPendingAttachedMeshes()
                  ++attachment_iter)
             {
                 LLViewerObject* objectp = attachment_iter->get();
-                if (objectp)
+                if (objectp && !objectp->isDead())
                 {
+                    if (check_object_for_mesh_loading(objectp))
+                    {
+                        return true;
+                    }
                     LLViewerObject::const_child_list_t& child_list = objectp->getChildren();
                     for (LLViewerObject::child_list_t::const_iterator iter1 = child_list.begin();
                          iter1 != child_list.end(); ++iter1)
                     {
                         LLViewerObject* objectchild = *iter1;
-                        if (objectchild && objectchild->getVolume())
+                        if (check_object_for_mesh_loading(objectchild))
                         {
-                            const LLUUID& mesh_id = objectchild->getVolume()->getParams().getSculptID();
-                            if (mesh_id.isNull())
-                            {
-                                // No mesh nor skin info needed
-                                continue;
-                            }
-
-                            if (objectchild->getVolume()->isMeshAssetUnavaliable())
-                            {
-                                // Mesh failed to load, do not expect it
-                                continue;
-                            }
-
-                            if (objectchild->mDrawable)
-                            {
-                                LLVOVolume* pvobj = objectchild->mDrawable->getVOVolume();
-                                if (pvobj)
-                                {
-                                    if (!pvobj->isMesh())
-                                    {
-                                        // Not a mesh
-                                        continue;
-                                    }
-
-                                    if (!objectchild->getVolume()->isMeshAssetLoaded())
-                                    {
-                                        // Waiting for mesh
-                                        return true;
-                                    }
-
-                                    const LLMeshSkinInfo* skin_data = pvobj->getSkinInfo();
-                                    if (skin_data)
-                                    {
-                                        // Skin info present, done
-                                        continue;
-                                    }
-
-                                    if (pvobj->isSkinInfoUnavaliable())
-                                    {
-                                        // Load failed or info not present, don't expect it
-                                        continue;
-                                    }
-                                }
-
-                                // objectchild is not ready
-                                return true;
-                            }
+                            return true;
                         }
                     }
                 }
@@ -8382,7 +8402,7 @@ bool LLVOAvatar::updateIsFullyLoaded()
                   );
 
         // compare amount of attachments to one reported by simulator
-        if (!loading && !isSelf() && rez_status < 4 && mLastCloudAttachmentCount < mSimAttachments.size())
+        if (!isSelf() && mLastCloudAttachmentCount < mSimAttachments.size() && mSimAttachments.size() > 0)
         {
             S32 attachment_count = getAttachmentCount();
             if (mLastCloudAttachmentCount != attachment_count)
@@ -8400,6 +8420,11 @@ bool LLVOAvatar::updateIsFullyLoaded()
                 // waiting
                 loading = true;
             }
+            else if (!loading)
+            {
+                // for hasFirstFullAttachmentData
+                mLastCloudAttachmentCount = (S32)mSimAttachments.size();
+            }
         }
     }
     updateRezzedStatusTimers(rez_status);
@@ -8513,6 +8538,12 @@ bool LLVOAvatar::isFullyLoaded() const
     return (mRenderUnloadedAvatar || mFullyLoaded);
 }
 
+bool LLVOAvatar::hasFirstFullAttachmentData() const
+{
+    return !mFirstFullyVisible // Avatar is fully visible, have all data
+        || mLastCloudAttachmentCount >= (S32)mSimAttachments.size();
+}
+
 bool LLVOAvatar::isTooComplex() const
 {
     bool too_complex;
diff --git a/indra/newview/llvoavatar.h b/indra/newview/llvoavatar.h
index dd1725c322..263c3dadf6 100644
--- a/indra/newview/llvoavatar.h
+++ b/indra/newview/llvoavatar.h
@@ -385,6 +385,7 @@ public:
     //--------------------------------------------------------------------
 public:
     bool            isFullyLoaded() const;
+    bool            hasFirstFullAttachmentData() const;
     F32             getFirstDecloudTime() const {return mFirstDecloudTime;}
 
     // check and return current state relative to limits