Display sys call names.

Only active if TRACY_SYSCALLS is defined, due to the volume of generated
data.

client/TracyCallstack.cpp

@@ -9,6 +9,7 @@
 #    define NOMINMAX
 #  endif
 #  include <windows.h>
+#  include <winternl.h>
 #  ifdef _MSC_VER
 #    pragma warning( push )
 #    pragma warning( disable : 4091 )
@@ -37,6 +38,29 @@ int cb_num;
 CallstackEntry cb_data[MaxCbTrace];
 
 extern "C" { t_RtlWalkFrameChain RtlWalkFrameChain = 0; }
+extern "C" typedef NTSTATUS (WINAPI *t_ZwQuerySystemInformation)( int, PVOID, ULONG, PULONG );
+
+extern "C" {
+typedef struct _SYSTEM_MODULE_ENTRY
+{
+    HANDLE Section;
+    PVOID MappedBase;
+    PVOID ImageBase;
+    ULONG ImageSize;
+    ULONG Flags;
+    USHORT LoadOrderIndex;
+    USHORT InitOrderIndex;
+    USHORT LoadCount;
+    USHORT OffsetToFileName;
+    UCHAR FullPathName[256];
+} SYSTEM_MODULE_ENTRY;
+
+typedef struct _SYSTEM_MODULE_INFORMATION
+{
+    ULONG Count;
+    SYSTEM_MODULE_ENTRY Module[1];
+} SYSTEM_MODULE_INFORMATION;
+}
 
 #if defined __MINGW32__ && API_VERSION_NUMBER < 12
 extern "C" {
@@ -51,28 +75,55 @@ BOOL IMAGEAPI SymGetLineFromInlineContext(HANDLE hProcess, DWORD64 qwAddr, ULONG
 };
 #endif
 
+static HANDLE currentProcess;
+
 void InitCallstack()
 {
+    currentProcess = GetCurrentProcess();
 #ifdef UNICODE
     RtlWalkFrameChain = (t_RtlWalkFrameChain)GetProcAddress( GetModuleHandle( L"ntdll.dll" ), "RtlWalkFrameChain" );
+    t_ZwQuerySystemInformation ZwQuerySystemInformation = (t_ZwQuerySystemInformation)GetProcAddress( GetModuleHandle( L"ntdll.dll" ), "ZwQuerySystemInformation" );
 #else
     RtlWalkFrameChain = (t_RtlWalkFrameChain)GetProcAddress( GetModuleHandle( "ntdll.dll" ), "RtlWalkFrameChain" );
+    t_ZwQuerySystemInformation ZwQuerySystemInformation = (t_ZwQuerySystemInformation)GetProcAddress( GetModuleHandle( "ntdll.dll" ), "ZwQuerySystemInformation" );
 #endif
-    SymInitialize( GetCurrentProcess(), nullptr, true );
+    SymInitialize( currentProcess, nullptr, true );
     SymSetOptions( SYMOPT_LOAD_LINES );
+
+    if( ZwQuerySystemInformation )
+    {
+        char sysroot[1024];
+        const auto rootlen = ExpandEnvironmentStringsA( "%SystemRoot%", sysroot, 1024 ) - 1;
+        ULONG size = 0;
+        ZwQuerySystemInformation( 11 /*SystemModuleInformation*/, 0, size, &size );
+        auto modules = (SYSTEM_MODULE_INFORMATION*)tracy_malloc( size );
+        ZwQuerySystemInformation( 11 /*SystemModuleInformation*/, modules, size, &size );
+        for( ULONG i=0; i<modules->Count; i++ )
+        {
+            char buf[1024];
+            const char* name = (char*)modules->Module[i].FullPathName;
+            if( strncmp( "\\SystemRoot", name, 11 ) == 0 )
+            {
+                memcpy( buf, sysroot, rootlen );
+                strcpy( buf + rootlen, name+11 );
+                name = buf;
+            }
+            auto res = SymLoadModuleEx( currentProcess, nullptr, name, nullptr, (DWORD64)modules->Module[i].ImageBase, modules->Module[i].ImageSize, nullptr, 0 );
+        }
+        tracy_free( modules );
+    }
 }
 
 const char* DecodeCallstackPtrFast( uint64_t ptr )
 {
     static char ret[1024];
-    const auto proc = GetCurrentProcess();
 
     char buf[sizeof( SYMBOL_INFO ) + 1024];
     auto si = (SYMBOL_INFO*)buf;
     si->SizeOfStruct = sizeof( SYMBOL_INFO );
     si->MaxNameLen = 1024;
 
-    if( SymFromAddr( proc, ptr, nullptr, si ) == 0 )
+    if( SymFromAddr( currentProcess, ptr, nullptr, si ) == 0 )
     {
         *ret = '\0';
     }
@@ -87,13 +138,12 @@ const char* DecodeCallstackPtrFast( uint64_t ptr )
 CallstackEntryData DecodeCallstackPtr( uint64_t ptr )
 {
     int write;
-    const auto proc = GetCurrentProcess();
 #ifndef __CYGWIN__
-    const auto inlineNum = std::min<DWORD>( MaxCbTrace - 1, SymAddrIncludeInlineTrace( proc, ptr ) );
+    const auto inlineNum = std::min<DWORD>( MaxCbTrace - 1, SymAddrIncludeInlineTrace( currentProcess, ptr ) );
     DWORD ctx = 0;
     DWORD idx;
     BOOL doInline = FALSE;
-    if( inlineNum != 0 ) doInline = SymQueryInlineTrace( proc, ptr, 0, ptr, ptr, &ctx, &idx );
+    if( inlineNum != 0 ) doInline = SymQueryInlineTrace( currentProcess, ptr, 0, ptr, ptr, &ctx, &idx );
     if( doInline )
     {
         write = inlineNum;
@@ -111,7 +161,7 @@ CallstackEntryData DecodeCallstackPtr( uint64_t ptr )
     si->SizeOfStruct = sizeof( SYMBOL_INFO );
     si->MaxNameLen = 1024;
 
-    if( SymFromAddr( proc, ptr, nullptr, si ) == 0 )
+    if( SymFromAddr( currentProcess, ptr, nullptr, si ) == 0 )
     {
         memcpy( si->Name, "[unknown]", 10 );
         si->NameLen = 9;
@@ -129,7 +179,7 @@ CallstackEntryData DecodeCallstackPtr( uint64_t ptr )
         cb_data[write].name = name;
 
         const char* filename;
-        if (SymGetLineFromAddr64(proc, ptr, &displacement, &line) == 0)
+        if( SymGetLineFromAddr64( currentProcess, ptr, &displacement, &line ) == 0 )
         {
             filename = "[unknown]";
             cb_data[write].line = 0;
@@ -155,7 +205,7 @@ CallstackEntryData DecodeCallstackPtr( uint64_t ptr )
         {
             auto& cb = cb_data[i];
 
-            if( SymFromInlineContext( proc, ptr, ctx, nullptr, si ) == 0 )
+            if( SymFromInlineContext( currentProcess, ptr, ctx, nullptr, si ) == 0 )
             {
                 memcpy( si->Name, "[unknown]", 10 );
                 si->NameLen = 9;
@@ -167,7 +217,7 @@ CallstackEntryData DecodeCallstackPtr( uint64_t ptr )
             cb.name = name;
 
             const char* filename;
-            if( SymGetLineFromInlineContext( proc, ptr, ctx, 0, &displacement, &line ) == 0 )
+            if( SymGetLineFromInlineContext( currentProcess, ptr, ctx, 0, &displacement, &line ) == 0 )
             {
                 filename = "[unknown]";
                 cb.line = 0;

client/TracyProfiler.cpp

@@ -1017,6 +1017,7 @@ Profiler::Profiler()
     , m_broadcast( nullptr )
     , m_noExit( false )
     , m_zoneId( 1 )
+    , m_pid( GetPid() )
     , m_stream( LZ4_createStream() )
     , m_buffer( (char*)tracy_malloc( TargetFrameSize*3 ) )
     , m_bufferOffset( 0 )
@@ -1172,8 +1173,6 @@ void Profiler::Worker()
     const auto hostinfo = GetHostInfo();
     const auto hisz = std::min<size_t>( strlen( hostinfo ), WelcomeMessageHostInfoSize - 1 );
 
-    const uint64_t pid = GetPid();
-
 #ifdef TRACY_ON_DEMAND
     uint8_t onDemand = 1;
 #else
@@ -1193,7 +1192,7 @@ void Profiler::Worker()
     MemWrite( &welcome.delay, m_delay );
     MemWrite( &welcome.resolution, m_resolution );
     MemWrite( &welcome.epoch, m_epoch );
-    MemWrite( &welcome.pid, pid );
+    MemWrite( &welcome.pid, m_pid );
     MemWrite( &welcome.onDemand, onDemand );
     MemWrite( &welcome.isApple, isApple );
     memcpy( welcome.programName, procname, pnsz );
@@ -1811,6 +1810,22 @@ Profiler::DequeueStatus Profiler::Dequeue( moodycamel::ConsumerToken& token )
                     MemWrite( &item->gpuTime.gpuTime, dt );
                     break;
                 }
+                case QueueType::SysCallEnter:
+                {
+                    int64_t t = MemRead<int64_t>( &item->sysCallEnter.time );
+                    int64_t dt = t - m_refTimeCtx;
+                    m_refTimeCtx = t;
+                    MemWrite( &item->sysCallEnter.time, dt );
+                    break;
+                }
+                case QueueType::SysCallExit:
+                {
+                    int64_t t = MemRead<int64_t>( &item->sysCallExit.time );
+                    int64_t dt = t - m_refTimeCtx;
+                    m_refTimeCtx = t;
+                    MemWrite( &item->sysCallExit.time, dt );
+                    break;
+                }
                 default:
                     assert( false );
                     break;
@@ -2039,7 +2054,9 @@ void Profiler::SendString( uint64_t str, const char* ptr, QueueType type )
             type == QueueType::PlotName ||
             type == QueueType::FrameName ||
             type == QueueType::ExternalName ||
-            type == QueueType::ExternalThreadName );
+            type == QueueType::ExternalThreadName ||
+            type == QueueType::CodeLocationString
+    );
 
     QueueItem item;
     MemWrite( &item.hdr.type, type );
@@ -2243,6 +2260,16 @@ bool Profiler::HandleServerQuery()
         SysTraceSendExternalName( ptr );
         break;
 #endif
+    case ServerQueryCodeLocationString:
+    {
+#ifdef TRACY_HAS_CALLSTACK
+        const auto name = DecodeCallstackPtrFast( ptr );
+#else
+        const auto name = "???";
+#endif
+        SendString( ptr, name, QueueType::CodeLocationString );
+        break;
+    }
     default:
         assert( false );
         break;
@@ -2482,6 +2509,11 @@ void Profiler::ProcessSysTime()
 }
 #endif
 
+uint64_t Profiler::GetProfilerTid() const
+{
+    return s_profilerThreadId;
+}
+
 }
 
 #ifdef __cplusplus

client/TracyProfiler.hpp

@@ -446,6 +446,9 @@ public:
 
     void SendString( uint64_t ptr, const char* str, QueueType type );
 
+    uint64_t Pid() const { return m_pid; }
+    uint64_t GetProfilerTid() const;
+
 private:
     enum class DequeueStatus { Success, ConnectionLost, QueueEmpty };
 
@@ -541,6 +544,7 @@ private:
     UdpBroadcast* m_broadcast;
     bool m_noExit;
     std::atomic<uint32_t> m_zoneId;
+    uint64_t m_pid;
 
     uint64_t m_threadCtx;
     int64_t m_refTimeThread;

client/TracySysTrace.cpp

@@ -54,6 +54,47 @@ struct ReadyThread
     int8_t      reserverd;
 };
 
+#ifdef TRACY_SYSCALLS
+struct SysCallEnter
+{
+    // Official MSDN documentation is bullshitting here
+    // This is corroborated by tracerpt displaying address as 64-bit
+#if defined __x86_64__ || defined _M_X64
+    uint64_t sysCallAddress;
+#else
+    uint32_t sysCallAddress;
+#endif
+};
+
+struct TraceThread
+{
+    uint32_t ProcessId;
+    uint32_t TThreadId;
+    uint32_t StackBase;
+    uint32_t StackLimit;
+    uint32_t UserStackBase;
+    uint32_t UserStackLimit;
+    uint32_t Affinity;
+    uint32_t Win32StartAddr;
+    uint32_t TebBase;
+    uint32_t SubProcessTag;
+    uint8_t  BasePriority;
+    uint8_t  PagePriority;
+    uint8_t  IoPriority;
+    uint8_t  ThreadFlags;
+};
+
+uint64_t coreThread[256] = {};
+
+// We need to know which thread ids are in our process. A hash map would be the best solution here,
+// but there is no solution ready to use right now. Instead, let's store a thread bit map. Note that
+// this is only an approximation, there are no guarantees about thread ids returned by the kernel, as
+// Raymond Chen said on multiple occasions. In the false positive case, we will be reporting system
+// calls belonging to some other process, and that's no big deal.
+enum { TidCapacity = 8192 };
+uint64_t processThreads[TidCapacity] = {};
+#endif
+
 void WINAPI EventRecordCallback( PEVENT_RECORD record )
 {
 #ifdef TRACY_ON_DEMAND
@@ -64,6 +105,7 @@ void WINAPI EventRecordCallback( PEVENT_RECORD record )
     if( hdr.EventDescriptor.Opcode == 36 )
     {
         const auto cswitch = (const CSwitch*)record->UserData;
+        const auto cpu = record->BufferContext.ProcessorNumber;
 
         Magic magic;
         auto token = GetToken();
@@ -75,10 +117,14 @@ void WINAPI EventRecordCallback( PEVENT_RECORD record )
         memcpy( &item->contextSwitch.newThread, &cswitch->newThreadId, sizeof( cswitch->newThreadId ) );
         memset( ((char*)&item->contextSwitch.oldThread)+4, 0, 4 );
         memset( ((char*)&item->contextSwitch.newThread)+4, 0, 4 );
-        MemWrite( &item->contextSwitch.cpu, record->BufferContext.ProcessorNumber );
+        MemWrite( &item->contextSwitch.cpu, cpu );
         MemWrite( &item->contextSwitch.reason, cswitch->oldThreadWaitReason );
         MemWrite( &item->contextSwitch.state, cswitch->oldThreadState );
         tail.store( magic + 1, std::memory_order_release );
+
+#ifdef TRACY_SYSCALLS
+        coreThread[cpu] = cswitch->newThreadId;
+#endif
     }
     else if( hdr.EventDescriptor.Opcode == 50 )
     {
@@ -94,6 +140,70 @@ void WINAPI EventRecordCallback( PEVENT_RECORD record )
         memset( ((char*)&item->threadWakeup.thread)+4, 0, 4 );
         tail.store( magic + 1, std::memory_order_release );
     }
+#ifdef TRACY_SYSCALLS
+    else if( hdr.EventDescriptor.Opcode == 51 )
+    {
+        const auto cpu = record->BufferContext.ProcessorNumber;
+        const auto thread = coreThread[cpu];
+
+        if( thread != 0 )
+        {
+            const auto tid = thread / 4;
+            const auto entry = ( tid / 64 ) % TidCapacity;
+            const auto bit = tid % 64;
+            if( processThreads[entry] & ( 1ull << bit ) )
+            {
+                const auto syscall = (const SysCallEnter*)record->UserData;
+                uint64_t addr = syscall->sysCallAddress;
+
+                Magic magic;
+                auto token = GetToken();
+                auto& tail = token->get_tail_index();
+                auto item = token->enqueue_begin( magic );
+                MemWrite( &item->hdr.type, QueueType::SysCallEnter );
+                MemWrite( &item->sysCallEnter.time, hdr.TimeStamp.QuadPart );
+                MemWrite( &item->sysCallEnter.thread, thread );
+                MemWrite( &item->sysCallEnter.address, addr );
+                tail.store( magic + 1, std::memory_order_release );
+            }
+        }
+    }
+    else if( hdr.EventDescriptor.Opcode == 52 )
+    {
+        const auto cpu = record->BufferContext.ProcessorNumber;
+        const auto thread = coreThread[cpu];
+
+        if( thread != 0 )
+        {
+            const auto tid = thread / 4;
+            const auto entry = ( tid / 64 ) % TidCapacity;
+            const auto bit = tid % 64;
+            if( processThreads[entry] & ( 1ull << bit ) )
+            {
+                Magic magic;
+                auto token = GetToken();
+                auto& tail = token->get_tail_index();
+                auto item = token->enqueue_begin( magic );
+                MemWrite( &item->hdr.type, QueueType::SysCallExit );
+                MemWrite( &item->sysCallExit.time, hdr.TimeStamp.QuadPart );
+                MemWrite( &item->sysCallExit.thread, thread );
+                tail.store( magic + 1, std::memory_order_release );
+            }
+        }
+    }
+    else if( hdr.EventDescriptor.Opcode == 1 || hdr.EventDescriptor.Opcode == 3 )
+    {
+        const auto tt = (const TraceThread*)record->UserData;
+        if( tt->ProcessId == GetProfiler().Pid() && tt->TThreadId != GetProfiler().GetProfilerTid() && tt->TThreadId != GetCurrentThreadId() )
+        {
+            // Thread ids are (currently) multiples of 4, but its not a part of the contract.
+            const uint32_t tid = tt->TThreadId / 4;
+            const auto entry = ( tid / 64 ) % TidCapacity;
+            const auto bit = tid % 64;
+            processThreads[entry] |= ( 1ull << bit );
+        }
+    }
+#endif
 }
 
 bool SysTraceStart()
@@ -114,7 +224,11 @@ bool SysTraceStart()
     const auto psz = sizeof( EVENT_TRACE_PROPERTIES ) + sizeof( KERNEL_LOGGER_NAME );
     s_prop = (EVENT_TRACE_PROPERTIES*)tracy_malloc( psz );
     memset( s_prop, 0, sizeof( EVENT_TRACE_PROPERTIES ) );
+#ifdef TRACY_SYSCALLS
+    s_prop->EnableFlags = EVENT_TRACE_FLAG_CSWITCH | EVENT_TRACE_FLAG_DISPATCHER | EVENT_TRACE_FLAG_SYSTEMCALL | EVENT_TRACE_FLAG_THREAD;
+#else
     s_prop->EnableFlags = EVENT_TRACE_FLAG_CSWITCH | EVENT_TRACE_FLAG_DISPATCHER;
+#endif
     s_prop->LogFileMode = EVENT_TRACE_REAL_TIME_MODE;
     s_prop->Wnode.BufferSize = psz;
     s_prop->Wnode.Flags = WNODE_FLAG_TRACED_GUID;

common/TracyProtocol.hpp

@@ -9,7 +9,7 @@
 namespace tracy
 {
 
-enum : uint32_t { ProtocolVersion = 21 };
+enum : uint32_t { ProtocolVersion = 22 };
 enum : uint32_t { BroadcastVersion = 0 };
 
 using lz4sz_t = uint32_t;
@@ -46,7 +46,8 @@ enum ServerQuery : uint8_t
     ServerQueryCallstackFrame,
     ServerQueryFrameName,
     ServerQueryDisconnect,
-    ServerQueryExternalName
+    ServerQueryExternalName,
+    ServerQueryCodeLocationString
 };
 
 struct ServerQueryPacket

common/TracyQueue.hpp

@@ -42,6 +42,8 @@ enum class QueueType : uint8_t
     ContextSwitch,
     ThreadWakeup,
     GpuTime,
+    SysCallEnter,
+    SysCallExit,
     Terminate,
     KeepAlive,
     ThreadContext,
@@ -73,6 +75,7 @@ enum class QueueType : uint8_t
     FrameImageData,
     ExternalName,
     ExternalThreadName,
+    CodeLocationString,
     NUM_TYPES
 };
 
@@ -330,6 +333,19 @@ struct QueueTidToPid
     uint64_t pid;
 };
 
+struct QueueSysCallEnter
+{
+    int64_t time;
+    uint64_t thread;
+    uint64_t address;
+};
+
+struct QueueSysCallExit
+{
+    int64_t time;
+    uint64_t thread;
+};
+
 struct QueueHeader
 {
     union
@@ -378,6 +394,8 @@ struct QueueItem
         QueueContextSwitch contextSwitch;
         QueueThreadWakeup threadWakeup;
         QueueTidToPid tidToPid;
+        QueueSysCallEnter sysCallEnter;
+        QueueSysCallExit sysCallExit;
     };
 };
 #pragma pack()
@@ -420,6 +438,8 @@ static const size_t QueueDataSize[] = {
     sizeof( QueueHeader ) + sizeof( QueueContextSwitch ),
     sizeof( QueueHeader ) + sizeof( QueueThreadWakeup ),
     sizeof( QueueHeader ) + sizeof( QueueGpuTime ),
+    sizeof( QueueHeader ) + sizeof( QueueSysCallEnter ),
+    sizeof( QueueHeader ) + sizeof( QueueSysCallExit ),
     // above items must be first
     sizeof( QueueHeader ),                                  // terminate
     sizeof( QueueHeader ),                                  // keep alive
@@ -453,6 +473,7 @@ static const size_t QueueDataSize[] = {
     sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // frame image data
     sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // external name
     sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // external thread name
+    sizeof( QueueHeader ) + sizeof( QueueStringTransfer ),  // code location string
 };
 
 static_assert( QueueItemSize == 32, "Queue item size not 32 bytes" );

profiler/build/win32/Tracy.sln

@@ -1,7 +1,7 @@
 
 Microsoft Visual Studio Solution File, Format Version 12.00
-# Visual Studio 15
+# Visual Studio Version 16
-VisualStudioVersion = 15.0.26730.15
+VisualStudioVersion = 16.0.29411.108
 MinimumVisualStudioVersion = 10.0.40219.1
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "Tracy", "Tracy.vcxproj", "{1C736F84-08DF-4A7A-A7FB-7BA3412B8C97}"
 EndProject

server/TracyEvent.hpp

@@ -339,6 +339,16 @@ struct MessageData
 
 enum { MessageDataSize = sizeof( MessageData ) };
 
+
+struct SysCall
+{
+    int64_t start;
+    int64_t end;
+    uint64_t ptr;
+};
+
+enum { SysCallSize = sizeof( SysCall ) };
+
 #pragma pack()
 
 
@@ -354,6 +364,7 @@ struct ThreadData
 #ifndef TRACY_NO_STATISTICS
     Vector<int64_t> childTimeStack;
 #endif
+    Vector<SysCall> sysCalls;
 };
 
 struct GpuCtxThreadData

server/TracyView.cpp

@@ -2337,6 +2337,13 @@ void View::DrawZones()
                 offset += ostep * lockDepth;
                 depth += lockDepth;
             }
+
+            if( !v->sysCalls.empty() )
+            {
+                const auto sysCallDepth = DrawSysCalls( v->sysCalls, hover, pxns, nspx, wpos, offset, yMin, yMax );
+                offset += ostep * sysCallDepth;
+                depth += sysCallDepth;
+            }
         }
         offset += ostep * 0.2f;
 
@@ -4371,6 +4378,119 @@ int View::DrawLocks( uint64_t tid, bool hover, double pxns, const ImVec2& wpos,
     return cnt;
 }
 
+int View::DrawSysCalls( const Vector<SysCall>& vec, bool hover, double pxns, int64_t nspx, const ImVec2& wpos, int offset, float yMin, float yMax )
+{
+    auto it = std::lower_bound( vec.begin(), vec.end(), std::max<int64_t>( 0, m_vd.zvStart ), [] ( const auto& l, const auto& r ) { return (uint64_t)l.end < (uint64_t)r; } );
+    if( it == vec.end() ) return 0;
+
+    const auto zitend = std::lower_bound( it, vec.end(), m_vd.zvEnd, [] ( const auto& l, const auto& r ) { return l.start < r; } );
+    if( it == zitend ) return 0;
+    if( it->end < 0 ) return 0;
+
+    const auto w = ImGui::GetWindowContentRegionWidth() - 1;
+    const auto ty = ImGui::GetFontSize();
+    const auto ostep = ty + 1;
+    auto draw = ImGui::GetWindowDrawList();
+
+    while( it < zitend )
+    {
+        const auto color = 0xFFDD7777;
+        const auto zsz = std::max( ( it->end - it->start ) * pxns, pxns * 0.5 );
+        if( zsz < MinVisSize )
+        {
+            const auto origStart = it->start;
+            int num = 0;
+            const auto px0 = ( it->start - m_vd.zvStart ) * pxns;
+            auto px1 = ( it->end - m_vd.zvStart ) * pxns;
+            auto rend = it->end;
+            auto nextTime = it->end + MinVisSize;
+            for(;;)
+            {
+                const auto prevIt = it;
+                it = std::lower_bound( it, zitend, nextTime, [] ( const auto& l, const auto& r ) { return (uint64_t)l.end < (uint64_t)r; } );
+                if( it == prevIt ) ++it;
+                num += std::distance( prevIt, it );
+                if( it == zitend ) break;
+                const auto nend = it->end < 0 ? m_worker.GetLastTime() : it->end;
+                const auto pxnext = ( nend - m_vd.zvStart ) * pxns;
+                if( pxnext - px1 >= MinVisSize * 2 ) break;
+                px1 = pxnext;
+                rend = nend;
+                nextTime = nend + nspx;
+            }
+            draw->AddRectFilled( wpos + ImVec2( std::max( px0, -10.0 ), offset ), wpos + ImVec2( std::min( std::max( px1, px0+MinVisSize ), double( w + 10 ) ), offset + ty ), color );
+            DrawZigZag( draw, wpos + ImVec2( 0, offset + ty/2 ), std::max( px0, -10.0 ), std::min( std::max( px1, px0+MinVisSize ), double( w + 10 ) ), ty/4, DarkenColor( color ) );
+            if( hover && ImGui::IsMouseHoveringRect( wpos + ImVec2( std::max( px0, -10.0 ), offset ), wpos + ImVec2( std::min( std::max( px1, px0+MinVisSize ), double( w + 10 ) ), offset + ty ) ) )
+            {
+                if( num > 1 )
+                {
+                    ImGui::BeginTooltip();
+                    TextFocused( "Sys calls too small to display:", RealToString( num, true ) );
+                    ImGui::Separator();
+                    TextFocused( "Execution time:", TimeToString( rend - origStart ) );
+                    ImGui::EndTooltip();
+
+                    if( ImGui::IsMouseClicked( 2 ) && rend - origStart > 0 )
+                    {
+                        ZoomToRange( origStart, rend );
+                    }
+                }
+                else if( ImGui::IsMouseClicked( 2 ) && rend - origStart > 0 )
+                {
+                    ZoomToRange( origStart, it->end );
+                }
+            }
+        }
+        else
+        {
+            const auto name = m_worker.GetCodeLocationName( it->ptr );
+            const auto tsz = ImGui::CalcTextSize( name );
+
+            const auto pr0 = ( it->start - m_vd.zvStart ) * pxns;
+            const auto pr1 = ( it->end - m_vd.zvStart ) * pxns;
+            const auto px0 = std::max( pr0, -10.0 );
+            const auto px1 = std::max( { std::min( pr1, double( w + 10 ) ), px0 + pxns * 0.5, px0 + MinVisSize } );
+            draw->AddRectFilled( wpos + ImVec2( px0, offset ), wpos + ImVec2( px1, offset + tsz.y ), color );
+            draw->AddRect( wpos + ImVec2( px0, offset ), wpos + ImVec2( px1, offset + tsz.y ), HighlightColor( color ) );
+            if( tsz.x < zsz )
+            {
+                const auto x = ( it->start - m_vd.zvStart ) * pxns + ( ( it->end - it->start ) * pxns - tsz.x ) / 2;
+                if( x < 0 || x > w - tsz.x )
+                {
+                    ImGui::PushClipRect( wpos + ImVec2( px0, offset ), wpos + ImVec2( px1, offset + tsz.y * 2 ), true );
+                    DrawTextContrast( draw, wpos + ImVec2( std::max( std::max( 0., px0 ), std::min( double( w - tsz.x ), x ) ), offset ), 0xFFFFFFFF, name );
+                    ImGui::PopClipRect();
+                }
+                else if( it->start == it->end )
+                {
+                    DrawTextContrast( draw, wpos + ImVec2( px0 + ( px1 - px0 - tsz.x ) * 0.5, offset ), 0xFFFFFFFF, name );
+                }
+                else
+                {
+                    DrawTextContrast( draw, wpos + ImVec2( x, offset ), 0xFFFFFFFF, name );
+                }
+            }
+            else
+            {
+                ImGui::PushClipRect( wpos + ImVec2( px0, offset ), wpos + ImVec2( px1, offset + tsz.y * 2 ), true );
+                DrawTextContrast( draw, wpos + ImVec2( ( it->start - m_vd.zvStart ) * pxns, offset ), 0xFFFFFFFF, name );
+                ImGui::PopClipRect();
+            }
+
+            if( hover && ImGui::IsMouseHoveringRect( wpos + ImVec2( px0, offset ), wpos + ImVec2( px1, offset + tsz.y ) ) )
+            {
+                if( !m_zoomAnim.active && ImGui::IsMouseClicked( 2 ) )
+                {
+                    ZoomToRange( it->start, it->end );
+                }
+            }
+
+            ++it;
+        }
+    }
+    return 1;
+}
+
 int View::DrawCpuData( int offset, double pxns, const ImVec2& wpos, bool hover, float yMin, float yMax )
 {
     const auto w = ImGui::GetWindowContentRegionWidth() - 1;
@@ -10508,6 +10628,7 @@ void View::DrawInfo()
         TextFocused( "Context switch regions:", RealToString( m_worker.GetContextSwitchCount(), true ) );
         ImGui::SameLine();
         TextFocused( "+", RealToString( m_worker.GetContextSwitchPerCpuCount(), true ) );
+        TextFocused( "System calls:", RealToString( m_worker.GetSysCallCount(), true ) );
         ImGui::TreePop();
     }
 

server/TracyView.hpp

@@ -126,6 +126,7 @@ private:
     int SkipGpuZoneLevel( const Vector<GpuEvent*>& vec, bool hover, double pxns, int64_t nspx, const ImVec2& wpos, int offset, int depth, uint64_t thread, float yMin, float yMax, int64_t begin, int drift );
     void DrawLockHeader( uint32_t id, const LockMap& lockmap, const SourceLocation& srcloc, bool hover, ImDrawList* draw, const ImVec2& wpos, float w, float ty, float offset, uint8_t tid );
     int DrawLocks( uint64_t tid, bool hover, double pxns, const ImVec2& wpos, int offset, LockHighlight& highlight, float yMin, float yMax );
+    int DrawSysCalls( const Vector<SysCall>& vec, bool hover, double pxns, int64_t nspx, const ImVec2& wpos, int offset, float yMin, float yMax );
     int DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, float yMin, float yMax );
     void DrawPlotPoint( const ImVec2& wpos, float x, float y, int offset, uint32_t color, bool hover, bool hasPrev, const PlotItem* item, double prev, bool merged, PlotType type, float PlotHeight );
     void DrawPlotPoint( const ImVec2& wpos, float x, float y, int offset, uint32_t color, bool hover, bool hasPrev, double val, double prev, bool merged, PlotType type, float PlotHeight );

server/TracyWorker.cpp

@@ -247,6 +247,7 @@ Worker::Worker( const char* addr, int port )
     , m_pendingSourceLocation( 0 )
     , m_pendingCallstackFrames( 0 )
     , m_pendingCallstackSubframes( 0 )
+    , m_pendingCodeLocationStrings( 0 )
     , m_callstackFrameStaging( nullptr )
     , m_traceVersion( CurrentVersion )
     , m_loadTime( 0 )
@@ -1840,6 +1841,16 @@ uint64_t Worker::GetContextSwitchPerCpuCount() const
     return cnt;
 }
 
+uint64_t Worker::GetSysCallCount() const
+{
+    uint64_t cnt = 0;
+    for( auto& td : m_data.threads )
+    {
+        cnt += td->sysCalls.size();
+    }
+    return cnt;
+}
+
 uint64_t Worker::GetPidFromTid( uint64_t tid ) const
 {
     auto it = m_data.tidToPid.find( tid );
@@ -2148,6 +2159,19 @@ std::pair<const char*, const char*> Worker::GetExternalName( uint64_t id ) const
     }
 }
 
+const char* Worker::GetCodeLocationName( uint64_t ptr ) const
+{
+    const auto it = m_data.codeLocationNames.find( ptr );
+    if( it == m_data.codeLocationNames.end() )
+    {
+        return "???";
+    }
+    else
+    {
+        return it->second;
+    }
+}
+
 const char* Worker::GetZoneName( const SourceLocation& srcloc ) const
 {
     if( srcloc.name.active )
@@ -2448,7 +2472,8 @@ void Worker::Exec()
         {
             if( m_pendingStrings != 0 || m_pendingThreads != 0 || m_pendingSourceLocation != 0 || m_pendingCallstackFrames != 0 ||
                 !m_pendingCustomStrings.empty() || m_data.plots.IsPending() || m_pendingCallstackPtr != 0 ||
-                m_pendingExternalNames != 0 || m_pendingCallstackSubframes != 0 || !m_pendingFrameImageData.empty() )
+                m_pendingExternalNames != 0 || m_pendingCallstackSubframes != 0 || !m_pendingFrameImageData.empty() ||
+                m_pendingCodeLocationStrings != 0 )
             {
                 continue;
             }
@@ -2549,6 +2574,9 @@ bool Worker::DispatchProcess( const QueueItem& ev, char*& ptr )
             case QueueType::ExternalThreadName:
                 AddExternalThreadName( ev.stringTransfer.ptr, ptr, sz );
                 break;
+            case QueueType::CodeLocationString:
+                AddCodeLocationString( ev.stringTransfer.ptr, ptr, sz );
+                break;
             default:
                 assert( false );
                 break;
@@ -2846,6 +2874,16 @@ void Worker::CheckExternalName( uint64_t id )
     Query( ServerQueryExternalName, id );
 }
 
+void Worker::CheckCodeLocationString( uint64_t id )
+{
+    if( m_data.codeLocationNames.find( id ) != m_data.codeLocationNames.end() ) return;
+
+    m_data.codeLocationNames.emplace( id, "???" );
+    m_pendingCodeLocationStrings++;
+
+    Query( ServerQueryCodeLocationString, id );
+}
+
 void Worker::AddSourceLocation( const QueueSourceLocation& srcloc )
 {
     assert( m_pendingSourceLocation > 0 );
@@ -2964,6 +3002,16 @@ void Worker::AddExternalThreadName( uint64_t ptr, char* str, size_t sz )
     it->second.second = sl.ptr;
 }
 
+void Worker::AddCodeLocationString( uint64_t ptr, char* str, size_t sz )
+{
+    assert( m_pendingCodeLocationStrings > 0 );
+    m_pendingCodeLocationStrings--;
+    auto it = m_data.codeLocationNames.find( ptr );
+    assert( it != m_data.codeLocationNames.end() && strcmp( it->second, "???" ) == 0 );
+    const auto sl = StoreString( str, sz );
+    it->second = sl.ptr;
+}
+
 static const uint8_t DxtcIndexTable[256] = {
     85,     87,     86,     84,     93,     95,     94,     92,     89,     91,     90,     88,     81,     83,     82,     80,
     117,    119,    118,    116,    125,    127,    126,    124,    121,    123,    122,    120,    113,    115,    114,    112,
@@ -3399,6 +3447,12 @@ bool Worker::Process( const QueueItem& ev )
     case QueueType::TidToPid:
         ProcessTidToPid( ev.tidToPid );
         break;
+    case QueueType::SysCallEnter:
+        ProcessSysCallEnter( ev.sysCallEnter );
+        break;
+    case QueueType::SysCallExit:
+        ProcessSysCallExit( ev.sysCallExit );
+        break;
     default:
         assert( false );
         break;
@@ -4641,6 +4695,40 @@ void Worker::ProcessTidToPid( const QueueTidToPid& ev )
     m_data.tidToPid.emplace( ev.tid, ev.pid );
 }
 
+void Worker::ProcessSysCallEnter( const QueueSysCallEnter& ev )
+{
+    auto td = RetrieveThread( ev.thread );
+    if( !td ) return;
+    if( td->sysCalls.empty() )
+    {
+        CheckCodeLocationString( ev.address );
+        const auto refTime = m_refTimeCtx + ev.time;
+        m_refTimeCtx = refTime;
+        const auto time = TscTime( refTime - m_data.baseTime );
+        td->sysCalls.push_back( SysCall { time, -1, ev.address } );
+    }
+    else
+    {
+        if( td->sysCalls.back().end < 0 ) return;
+        CheckCodeLocationString( ev.address );
+        const auto refTime = m_refTimeCtx + ev.time;
+        m_refTimeCtx = refTime;
+        const auto time = TscTime( refTime - m_data.baseTime );
+        td->sysCalls.push_back_non_empty( SysCall { time, -1, ev.address } );
+    }
+}
+
+void Worker::ProcessSysCallExit( const QueueSysCallExit& ev )
+{
+    auto td = RetrieveThread( ev.thread );
+    if( !td || td->sysCalls.empty() ) return;
+    if( td->sysCalls.back().end >= 0 ) return;
+    const auto refTime = m_refTimeCtx + ev.time;
+    m_refTimeCtx = refTime;
+    const auto time = TscTime( refTime - m_data.baseTime );
+    td->sysCalls.back().end = time;
+}
+
 void Worker::MemAllocChanged( int64_t time )
 {
     const auto val = (double)m_data.memory.usage;

server/TracyWorker.hpp

@@ -173,6 +173,7 @@ private:
         flat_hash_map<charutil::StringKey, uint32_t, charutil::StringKey::HasherPOT, charutil::StringKey::Comparator> stringMap;
         flat_hash_map<uint64_t, const char*, nohash<uint64_t>> threadNames;
         flat_hash_map<uint64_t, std::pair<const char*, const char*>, nohash<uint64_t>> externalNames;
+        flat_hash_map<uint64_t, const char*, nohash<uint64_t>> codeLocationNames;
 
         flat_hash_map<uint64_t, SourceLocation, nohash<uint64_t>> sourceLocation;
         Vector<SourceLocation*> sourceLocationPayload;
@@ -298,6 +299,7 @@ public:
     uint64_t GetContextSwitchPerCpuCount() const;
     bool HasContextSwitches() const { return !m_data.ctxSwitch.empty(); }
     uint64_t GetSrcLocCount() const { return m_data.sourceLocationPayload.size() + m_data.sourceLocation.size(); }
+    uint64_t GetSysCallCount() const;
     uint64_t GetCallstackPayloadCount() const { return m_data.callstackPayload.size() - 1; }
     uint64_t GetCallstackFrameCount() const { return m_data.callstackFrameMap.size(); }
     uint32_t GetFrameImageCount() const { return (uint32_t)m_data.frameImage.size(); }
@@ -354,6 +356,7 @@ public:
     bool IsThreadLocal( uint64_t id ) const;
     const SourceLocation& GetSourceLocation( int16_t srcloc ) const;
     std::pair<const char*, const char*> GetExternalName( uint64_t id ) const;
+    const char* GetCodeLocationName( uint64_t ptr ) const;
 
     const char* GetZoneName( const SourceLocation& srcloc ) const;
     const char* GetZoneName( const ZoneEvent& ev ) const;
@@ -459,6 +462,8 @@ private:
     tracy_force_inline void ProcessContextSwitch( const QueueContextSwitch& ev );
     tracy_force_inline void ProcessThreadWakeup( const QueueThreadWakeup& ev );
     tracy_force_inline void ProcessTidToPid( const QueueTidToPid& ev );
+    tracy_force_inline void ProcessSysCallEnter( const QueueSysCallEnter& ev );
+    tracy_force_inline void ProcessSysCallExit( const QueueSysCallExit& ev );
 
     tracy_force_inline void ProcessZoneBeginImpl( ZoneEvent* zone, const QueueZoneBegin& ev );
     tracy_force_inline void ProcessZoneBeginAllocSrcLocImpl( ZoneEvent* zone, const QueueZoneBegin& ev );
@@ -526,6 +531,7 @@ private:
     void CheckString( uint64_t ptr );
     void CheckThreadString( uint64_t id );
     void CheckExternalName( uint64_t id );
+    void CheckCodeLocationString( uint64_t id );
 
     void AddSourceLocation( const QueueSourceLocation& srcloc );
     void AddSourceLocationPayload( uint64_t ptr, char* data, size_t sz );
@@ -535,6 +541,7 @@ private:
     void AddCustomString( uint64_t ptr, char* str, size_t sz );
     void AddExternalName( uint64_t ptr, char* str, size_t sz );
     void AddExternalThreadName( uint64_t ptr, char* str, size_t sz );
+    void AddCodeLocationString( uint64_t ptr, char* str, size_t sz );
     void AddFrameImageData( uint64_t ptr, char* data, size_t sz );
 
     tracy_force_inline void AddCallstackPayload( uint64_t ptr, char* data, size_t sz );
@@ -615,6 +622,7 @@ private:
     uint32_t m_pendingSourceLocation;
     uint32_t m_pendingCallstackFrames;
     uint8_t m_pendingCallstackSubframes;
+    uint32_t m_pendingCodeLocationStrings;
 
     CallstackFrameData* m_callstackFrameStaging;
     uint64_t m_callstackFrameStagingPtr;