AuroraSupport/AuroraRuntime
2
1
Fork 0
Code Issues 27 Pull Requests Releases Wiki Activity

[*] Minor NT optimization: move branch

Browse Source
This commit is contained in:
Reece Wilson 2023-07-10 18:51:28 +01:00
parent bdec6ff8ba
commit c90a13ad95
2 changed files with 76 additions and 78 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
Source/Threading/Primitives/AuConditionMutex.NT.cpp
View File

@ -111,44 +111,32 @@ namespace Aurora::Threading::Primitives
#else
InterlockedAndRelease((volatile LONG *)&uValueRef, ~0xFF);
#endif
if (gUseNativeWaitCondvar)
while (true)
{
while (true)
auto uOld = uValueRef;
auto uValue = uOld;
if (uValue & 1)
{
auto uValue = uValueRef;
return;
}
if (uValue < kFutexBitWait)
{
return;
}
if (uValue < kFutexBitWait)
{
return;
}
if (gUseNativeWaitCondvar)
{
if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait, uValue) == uValue)
{
pWakeByAddressSingle((void *)&this->lock_.uWaitCount);
pWakeByAddressSingle((void *)&uValueRef);
return;
}
SMPPause();
}
}
else
{
while (true)
else
{
auto uOld = uValueRef;
auto uValue = uOld;
if (uValue & 1)
{
return;
}
if (uValue < kFutexBitWait)
{
return;
}
if (uValue & kFutexBitWake)
{
if (AuAtomicCompareExchange(&uValueRef, uValue, uValue) == uValue)
@ -167,9 +155,9 @@ namespace Aurora::Threading::Primitives
pNtReleaseKeyedEvent(gKeyedEventHandle, (void *)&uValueRef, 0, NULL);
return;
}
SMPPause();
}
SMPPause();
}
#endif
}

106
Source/Threading/Primitives/AuMutex.NT.cpp
View File

@ -70,12 +70,12 @@ namespace Aurora::Threading::Primitives
bool MutexImpl::LockMS(AuUInt64 uTimeout)
{
if (AuAtomicTestAndSet(&this->state_, 0) == 0)
if (this->TryLockNoSpin())
{
return true;
}
return LockNS(AuMSToNS<AuUInt64>(uTimeout));
return this->LockNS(AuMSToNS<AuUInt64>(uTimeout));
}
bool MutexImpl::LockNS(AuUInt64 uTimeout)
@ -224,6 +224,8 @@ namespace Aurora::Threading::Primitives
void MutexImpl::Unlock()
{
#if defined(AURORA_FORCE_SRW_LOCKS)
if (gUseNativeWaitMutex)
{
auto &uValueRef = this->state_;
@ -250,61 +252,69 @@ namespace Aurora::Threading::Primitives
return;
}
else
{
#if defined(AURORA_FORCE_SRW_LOCKS)
::AcquireSRWLockExclusive(&this->atomicHolder_);
this->state_ = 0;
::ReleaseSRWLockExclusive(&this->atomicHolder_);
::WakeAllConditionVariable(&this->wakeup_);
#else
auto &uValueRef = this->state_;
::AcquireSRWLockExclusive(&this->atomicHolder_);
this->state_ = 0;
::ReleaseSRWLockExclusive(&this->atomicHolder_);
::WakeAllConditionVariable(&this->wakeup_);
#else
#if defined(AURORA_ARCH_X86) || defined(AURORA_ARCH_X64)
// Intel 64 and IA - 32 Architectures Software Developer's Manual, Volume 3A: Section: 8.2.3.1
*(AuUInt8 *)&uValueRef = 0;
auto &uValueRef = this->state_;
// From this point onwards, our thread could be subject to StoreLoad re-ordering
// ...but it should not matter.
#if defined(AURORA_ARCH_X86) || defined(AURORA_ARCH_X64)
// Intel 64 and IA - 32 Architectures Software Developer's Manual, Volume 3A: Section: 8.2.3.1
*(AuUInt8 *)&uValueRef = 0;
// From this point onwards, our thread could be subject to StoreLoad re-ordering
// ...but it should not matter.
// Given the memory model of x86[64], we can only really expect to be out of order during an unfenced load operation, which in this class, can only be expected under this function before the CAS.
// No other place reads.
// Given the memory model of x86[64], we can only really expect to be out of order during an unfenced load operation, which in this class, can only be expected under this function before the CAS.
// No other place reads.
// Re-ordering race condition 1: one thread wins an atomic bit set, that we dont catch until the CAS, resulting in: a slow implicit fence under the cas, a mm_pause stall, a compare, and a return
// alt: uValueRef reads zero, resulting in a preemptive return while no threads need to be awoken
// Re-ordering race condition 2: we unlock, multiple threads enter ::Lock(), we somehow read `uValue = uValueRef` as zero, and then the first atomic bitsetandtest winner thread signals the keyed mutex
// I fail to see how:
// *byte = 0; | |
// | interlocked atomicbitset | interlocked atomicbitset fail
// | [logic] | interlocked atomic set kFutexBitWait
// | *byte = 0; | yield
// | auto uValue =[acquire]= uValueRef
// ...would result in the second thread missing the third threads atomic set kFutexBitWait (cst (?) on the account of 8.2.3.1, 8.2.3.8, etc)
// Re-ordering race condition 1: one thread wins an atomic bit set, that we dont catch until the CAS, resulting in: a slow implicit fence under the cas, a mm_pause stall, a compare, and a return
// alt: uValueRef reads zero, resulting in a preemptive return while no threads need to be awoken
// Re-ordering race condition 2: we unlock, multiple threads enter ::Lock(), we somehow read `uValue = uValueRef` as zero, and then the first atomic bitsetandtest winner thread signals the keyed mutex
// I fail to see how:
// *byte = 0; | |
// | interlocked atomicbitset | interlocked atomicbitset fail
// | [logic] | interlocked atomic set kFutexBitWait
// | *byte = 0; | yield
// | auto uValue =[acquire]= uValueRef
// ...would result in the second thread missing the third threads atomic set kFutexBitWait (cst (?) on the account of 8.2.3.1, 8.2.3.8, etc)
// Also note: mfence is far too expensive and the _ReadWriteBarrier() intrinsics do absolutely nothing
#else
InterlockedAndRelease((volatile LONG *)&uValueRef, ~0xFF);
#endif
// Also note: mfence is far too expensive and the _ReadWriteBarrier() intrinsics do absolutely nothing
_ReadWriteBarrier();
#else
InterlockedAndRelease((volatile LONG *)&uValueRef, ~0xFF);
#endif
while (true)
while (true)
{
auto uValue = uValueRef;
if (uValue < kFutexBitWait)
{
auto uValue = uValueRef;
return;
}
//
if (uValue < kFutexBitWait)
// StoreLoad race-conditions here cannot result in a return
// We should see StoreLoads of at least our *pByte = 0
// or we should at least see the CST of kFutexBitWait being applied
if (uValue & 1)
{
return;
}
if (gUseNativeWaitMutex)
{
if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait, uValue) == uValue)
{
pWakeByAddressSingle((void *)&this->state_);
return;
}
// StoreLoad race-conditions here cannot result in a return
// We should see StoreLoads of at least our *pByte = 0
// or we should at least see the CST of kFutexBitWait being applied
if (uValue & 1)
{
return;
}
}
else
{
if (uValue & kFutexBitWake)
{
// StoreLoad paranoia
@ -324,12 +334,12 @@ namespace Aurora::Threading::Primitives
pNtReleaseKeyedEvent(gKeyedEventHandle, (void *)&uValueRef, 0, NULL);
return;
}
SMPPause();
}
#endif
SMPPause();
}
#endif
}
AUKN_SYM IHyperWaitable *MutexNew()