mirror of
https://github.com/wolfpld/tracy.git
synced 2026-09-24 11:14:30 +00:00
Rewrite server-side lock tracking as an append-only event engine.
Capture has serialized lock events sinced6f32a083and3e3aa80fa, so the out-of-order reconstruction machinery has nothing left to repair. Mark now targets the recorded Wait/Obtain event index; the old reverse scan could walk off the start of the timeline for a thread that never waited.
This commit is contained in:
Bartosz Taudul
359
server/TracyLocks.cpp
Normal file
359
server/TracyLocks.cpp
Normal file
@@ -0,0 +1,359 @@
|
||||
#include <utility>
|
||||
|
||||
#include "TracyLocks.hpp"
|
||||
|
||||
namespace tracy
|
||||
{
|
||||
|
||||
void InitLockMap( LockMap& map, int16_t srcloc, LockType type, int64_t announce )
|
||||
{
|
||||
map.srcloc = srcloc;
|
||||
map.type = type;
|
||||
map.timeAnnounce = announce;
|
||||
map.timeTerminate = 0;
|
||||
map.valid = true;
|
||||
map.isContended = false;
|
||||
}
|
||||
|
||||
void ReserveLockSlots( LockMap& map, const uint64_t* threadIds, size_t count )
|
||||
{
|
||||
const size_t base = map.threads.size();
|
||||
assert( base + count <= 0xFFFF );
|
||||
map.threadMap.reserve( base + count );
|
||||
map.threads.reserve( base + count );
|
||||
for( size_t i=0; i<count; i++ )
|
||||
{
|
||||
map.threadMap.emplace( threadIds[i], ( uint16_t )( base + i ) );
|
||||
LockThreadInfo ti;
|
||||
ti.thread = threadIds[i];
|
||||
map.threads.push_back( std::move( ti ) );
|
||||
}
|
||||
}
|
||||
|
||||
uint16_t GetLockSlot( LockMap& map, uint64_t thread )
|
||||
{
|
||||
auto it = map.threadMap.find( thread );
|
||||
if( it != map.threadMap.end() ) return it->second;
|
||||
if( map.threads.size() >= LockEvent::NoThread ) return LockEvent::NoThread;
|
||||
const auto slot = ( uint16_t )map.threads.size();
|
||||
map.threadMap.emplace( thread, slot );
|
||||
LockThreadInfo ti;
|
||||
ti.thread = thread;
|
||||
map.threads.push_back( std::move( ti ) );
|
||||
return slot;
|
||||
}
|
||||
|
||||
static tracy_force_inline void EraseSlot( Vector<uint16_t>& vec, uint16_t slot )
|
||||
{
|
||||
for( size_t i=0; i<vec.size(); i++ )
|
||||
{
|
||||
if( vec[i] == slot )
|
||||
{
|
||||
vec[i] = vec.back();
|
||||
vec.pop_back();
|
||||
return;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// a pure waiter settled at WaitLock is invariant at foreign events: its flags can only change at
|
||||
// its own events (self is always visited) and drain-time flag clears reach holders
|
||||
static tracy_force_inline bool PinnedWait( uint8_t state, uint8_t flags )
|
||||
{
|
||||
return state == LockEventState::WaitLock && ( flags & ( LockEventFlags::Waiting | LockEventFlags::SharedWaiting ) ) != 0 &&
|
||||
( flags & LockEventFlags::SharedHolding ) == 0;
|
||||
}
|
||||
|
||||
static void CacheSeverity( LockThreadInfo& ti, uint32_t sidx, uint8_t state )
|
||||
{
|
||||
if( state == LockEventState::HasBlockingLock ) ti.yellowSegs.push_back( sidx );
|
||||
else if( state == LockEventState::WaitLock ) ti.redSegs.push_back( sidx );
|
||||
}
|
||||
|
||||
LockEventState::Type ResolveLockState( const LockMap& map, uint16_t slot )
|
||||
{
|
||||
const auto& ti = map.threads[slot];
|
||||
const bool waiting = ( ti.flags & LockEventFlags::Waiting ) != 0;
|
||||
const bool sharedWaiting = ( ti.flags & LockEventFlags::SharedWaiting ) != 0;
|
||||
const bool sharedHolding = ( ti.flags & LockEventFlags::SharedHolding ) != 0;
|
||||
const bool holding = map.curLockCount > 0 && map.curLockingThread == slot;
|
||||
const bool exclOthers = ( map.curWaitCount - ( waiting ? 1 : 0 ) ) > 0;
|
||||
const bool sharedOthers = ( map.curWaitSharedCount - ( sharedWaiting ? 1 : 0 ) ) > 0;
|
||||
|
||||
if( map.type == LockType::Lockable )
|
||||
{
|
||||
if( holding ) return exclOthers ? LockEventState::HasBlockingLock : LockEventState::HasLock;
|
||||
if( map.curLockCount > 0 && waiting ) return LockEventState::WaitLock;
|
||||
return LockEventState::Nothing;
|
||||
}
|
||||
else
|
||||
{
|
||||
if( holding ) return ( exclOthers || sharedOthers ) ? LockEventState::HasBlockingLock : LockEventState::HasLock;
|
||||
if( map.curLockCount > 0 && ( waiting || sharedWaiting ) ) return LockEventState::WaitLock;
|
||||
// a shared holder's own exclusive wait is an upgrade request - a deadlock - so the check is self-inclusive
|
||||
if( map.curLockCount == 0 && sharedHolding ) return map.curWaitCount > 0 ? LockEventState::HasBlockingLock : LockEventState::HasLock;
|
||||
if( map.curLockCount == 0 && map.curSharedCount > 0 && waiting ) return LockEventState::WaitLock;
|
||||
return LockEventState::Nothing;
|
||||
}
|
||||
}
|
||||
|
||||
static void SegmentPass( LockMap& map, uint16_t self, LockEvent::Type type, uint32_t idx )
|
||||
{
|
||||
const bool acquire = type == LockEvent::Type::Obtain || type == LockEvent::Type::ObtainShared;
|
||||
|
||||
map.passScratch.clear();
|
||||
map.passScratch.push_back( self );
|
||||
for( auto s : map.activeSlots )
|
||||
{
|
||||
if( s != self ) map.passScratch.push_back( s );
|
||||
}
|
||||
if( acquire || ( type == LockEvent::Type::Release && map.curLockCount == 0 ) )
|
||||
{
|
||||
for( auto s : map.pendingStarts )
|
||||
{
|
||||
if( s != self ) map.passScratch.push_back( s );
|
||||
}
|
||||
}
|
||||
|
||||
for( auto T : map.passScratch )
|
||||
{
|
||||
auto& ti = map.threads[T];
|
||||
const bool open = ti.curState != LockEventState::Nothing;
|
||||
auto desired = ResolveLockState( map, T );
|
||||
if( open && PinnedWait( ti.curState, ti.flags ) ) desired = LockEventState::WaitLock;
|
||||
|
||||
const uint8_t snapState = ( uint8_t )desired;
|
||||
const uint8_t snapFlags = ti.flags;
|
||||
|
||||
if( !open )
|
||||
{
|
||||
if( desired != LockEventState::Nothing )
|
||||
{
|
||||
LockSegment seg;
|
||||
seg.evStart = idx;
|
||||
seg.nextEv = LockEvent::NoEvent;
|
||||
seg.state = snapState;
|
||||
seg.flags = snapFlags;
|
||||
ti.segments.push_back( seg );
|
||||
CacheSeverity( ti, ( uint32_t )ti.segments.size() - 1, snapState );
|
||||
ti.curState = snapState;
|
||||
if( !PinnedWait( snapState, snapFlags ) ) map.activeSlots.push_back( T );
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
auto& seg = ti.segments.back();
|
||||
if( seg.state != snapState || seg.flags != snapFlags )
|
||||
{
|
||||
seg.nextEv = idx;
|
||||
ti.curState = LockEventState::Nothing;
|
||||
EraseSlot( map.activeSlots, T );
|
||||
if( desired != LockEventState::Nothing )
|
||||
{
|
||||
LockSegment ns;
|
||||
ns.evStart = idx;
|
||||
ns.nextEv = LockEvent::NoEvent;
|
||||
ns.state = snapState;
|
||||
ns.flags = snapFlags;
|
||||
ti.segments.push_back( ns );
|
||||
CacheSeverity( ti, ( uint32_t )ti.segments.size() - 1, snapState );
|
||||
ti.curState = snapState;
|
||||
if( !PinnedWait( snapState, snapFlags ) ) map.activeSlots.push_back( T );
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// inversion handling (traces from clients that emit the release event after the unlock):
|
||||
// an ObtainShared arriving before the exclusive Release resolves to Nothing until
|
||||
// the exclusive drains - pending so the drain event opens the hold. With the release
|
||||
// event ordered before the unlock, SharedHolding under an exclusive holder is unreachable.
|
||||
const bool need = ti.curState == LockEventState::Nothing && ( ( ti.flags & ( LockEventFlags::Waiting | LockEventFlags::SharedWaiting ) ) != 0 ||
|
||||
( map.legacyInversions && map.type == LockType::SharedLockable && ( ti.flags & LockEventFlags::SharedHolding ) != 0 && map.curLockCount > 0 ) );
|
||||
if( need )
|
||||
{
|
||||
if( !ti.inPending )
|
||||
{
|
||||
ti.inPending = true;
|
||||
map.pendingStarts.push_back( T );
|
||||
}
|
||||
}
|
||||
else if( ti.inPending )
|
||||
{
|
||||
ti.inPending = false;
|
||||
EraseSlot( map.pendingStarts, T );
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void AppendLockEvent( LockMap& map, int64_t time, uint16_t slot, LockEvent::Type type, int16_t srcloc )
|
||||
{
|
||||
if( slot == LockEvent::NoThread ) return;
|
||||
auto& ti = map.threads[slot];
|
||||
|
||||
switch( type )
|
||||
{
|
||||
case LockEvent::Type::Wait:
|
||||
if( !( ti.flags & LockEventFlags::Waiting ) )
|
||||
{
|
||||
ti.flags |= LockEventFlags::Waiting;
|
||||
map.curWaitCount++;
|
||||
}
|
||||
break;
|
||||
case LockEvent::Type::WaitShared:
|
||||
if( !( ti.flags & LockEventFlags::SharedWaiting ) )
|
||||
{
|
||||
ti.flags |= LockEventFlags::SharedWaiting;
|
||||
map.curWaitSharedCount++;
|
||||
}
|
||||
break;
|
||||
case LockEvent::Type::Obtain:
|
||||
assert( map.curLockCount < UINT16_MAX );
|
||||
if( ti.flags & LockEventFlags::Waiting )
|
||||
{
|
||||
ti.flags &= ~LockEventFlags::Waiting;
|
||||
map.curWaitCount--;
|
||||
}
|
||||
if( map.curLockingThread != slot )
|
||||
{
|
||||
assert( map.legacyInversions || map.curLockCount == 0 );
|
||||
if( map.curLockCount > 0 ) map.threads[map.curLockingThread].flags &= ~LockEventFlags::LockHolding;
|
||||
}
|
||||
ti.flags |= LockEventFlags::LockHolding;
|
||||
map.curLockingThread = slot;
|
||||
map.curLockCount++;
|
||||
break;
|
||||
case LockEvent::Type::Release:
|
||||
if( map.curLockCount != 0 )
|
||||
{
|
||||
map.curLockCount--;
|
||||
if( map.curLockCount == 0 ) map.threads[map.curLockingThread].flags &= ~LockEventFlags::LockHolding;
|
||||
}
|
||||
break;
|
||||
case LockEvent::Type::ObtainShared:
|
||||
if( ti.flags & LockEventFlags::SharedWaiting )
|
||||
{
|
||||
ti.flags &= ~LockEventFlags::SharedWaiting;
|
||||
map.curWaitSharedCount--;
|
||||
}
|
||||
if( !( ti.flags & LockEventFlags::SharedHolding ) )
|
||||
{
|
||||
ti.flags |= LockEventFlags::SharedHolding;
|
||||
map.curSharedCount++;
|
||||
}
|
||||
break;
|
||||
case LockEvent::Type::ReleaseShared:
|
||||
if( ti.flags & LockEventFlags::SharedHolding )
|
||||
{
|
||||
ti.flags &= ~LockEventFlags::SharedHolding;
|
||||
map.curSharedCount--;
|
||||
}
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
assert( map.timeline.empty() || map.timeline.back().Time() <= time );
|
||||
assert( map.timeline.size() < LockEvent::NoEvent );
|
||||
const uint32_t idx = ( uint32_t )map.timeline.size();
|
||||
|
||||
LockEvent ev;
|
||||
ev.SetTime( time );
|
||||
ev.SetSrcLoc( srcloc );
|
||||
ev.thread = slot;
|
||||
ev.type = ( uint8_t )type;
|
||||
map.timeline.push_back( ev );
|
||||
if( type == LockEvent::Type::Obtain || type == LockEvent::Type::Release )
|
||||
map.holderChanges.push_back( { idx, map.curLockingThread, map.curLockCount } );
|
||||
if( srcloc != 0 && ( ti.marks.empty() || ti.marks.back() != idx ) ) ti.marks.push_back( idx );
|
||||
|
||||
if( ti.firstTime > time ) ti.firstTime = time;
|
||||
if( ti.lastTime < time ) ti.lastTime = time;
|
||||
|
||||
switch( type )
|
||||
{
|
||||
case LockEvent::Type::Wait:
|
||||
case LockEvent::Type::WaitShared:
|
||||
ti.openWaitStart = time;
|
||||
ti.hasOpenWait = true;
|
||||
break;
|
||||
case LockEvent::Type::Obtain:
|
||||
case LockEvent::Type::ObtainShared:
|
||||
if( ti.hasOpenWait )
|
||||
{
|
||||
ti.waitTotal += time - ti.openWaitStart;
|
||||
ti.waitCount++;
|
||||
ti.hasOpenWait = false;
|
||||
}
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
switch( type )
|
||||
{
|
||||
case LockEvent::Type::Wait:
|
||||
case LockEvent::Type::Obtain:
|
||||
case LockEvent::Type::WaitShared:
|
||||
case LockEvent::Type::ObtainShared:
|
||||
ti.lastWaitObtain = idx;
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
|
||||
if( map.curLockCount != 0 )
|
||||
{
|
||||
if( !map.holdOpen )
|
||||
{
|
||||
map.holdOpen = true;
|
||||
map.openHoldStart = time;
|
||||
}
|
||||
}
|
||||
else if( map.holdOpen )
|
||||
{
|
||||
map.holdTotal += time - map.openHoldStart;
|
||||
map.holdOpen = false;
|
||||
}
|
||||
|
||||
if( map.curWaitCount != 0 )
|
||||
{
|
||||
if( !map.waitAggOpen )
|
||||
{
|
||||
map.waitAggOpen = true;
|
||||
map.openWaitAggStart = time;
|
||||
}
|
||||
if( map.curWaitCount > map.maxWaiting ) map.maxWaiting = map.curWaitCount;
|
||||
}
|
||||
else if( map.waitAggOpen )
|
||||
{
|
||||
map.waitTotalAgg += time - map.openWaitAggStart;
|
||||
map.waitAggOpen = false;
|
||||
}
|
||||
|
||||
if( !map.isContended )
|
||||
{
|
||||
if( map.type == LockType::Lockable )
|
||||
{
|
||||
map.isContended = map.curLockCount != 0 && map.curWaitCount != 0;
|
||||
}
|
||||
else
|
||||
{
|
||||
map.isContended = ( map.curLockCount != 0 && ( map.curWaitCount != 0 || map.curWaitSharedCount != 0 ) ) || ( map.curSharedCount != 0 && map.curWaitCount != 0 );
|
||||
}
|
||||
}
|
||||
|
||||
SegmentPass( map, slot, type, idx );
|
||||
}
|
||||
|
||||
bool ApplyLockMark( LockMap& map, uint16_t slot, int16_t srcloc )
|
||||
{
|
||||
if( slot == LockEvent::NoThread ) return false;
|
||||
auto& ti = map.threads[slot];
|
||||
const auto idx = ti.lastWaitObtain;
|
||||
if( idx == LockEvent::NoEvent ) return false;
|
||||
map.timeline[idx].SetSrcLoc( srcloc );
|
||||
if( ti.marks.empty() || ti.marks.back() != idx ) ti.marks.push_back( idx );
|
||||
return true;
|
||||
}
|
||||
|
||||
}
|
||||
Reference in New Issue
Block a user