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

[*] Update fenceless notes in AuMutex.NT.cpp

Browse Source
This commit is contained in:
Reece Wilson 2023-05-08 13:58:08 +01:00
parent 1ec1d4d01e
commit 08f790c3af
1 changed files with 37 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
Source/Threading/Primitives/AuMutex.NT.cpp
View File

@ -1,4 +1,4 @@
/***
/***
Copyright (C) 2021 J Reece Wilson (a/k/a "Reece"). All rights reserved.
File: AuMutex.NT.cpp
@ -245,22 +245,44 @@ namespace Aurora::Threading::Primitives
auto &uValueRef = this->state_;
#if defined(AURORA_ARCH_X86) || defined(AURORA_ARCH_X64)
// Intel 64 and IA - 32 Architectures Software Developer's Manual, Volume 3A
// Section: 9.1.1
// Intel 64 and IA - 32 Architectures Software Developer's Manual, Volume 3A: Section: 9.1.1
*(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.
// 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, paragraph 3)
// Also note: mfence is far too expensive and the _ReadWriteBarrier() intrinsics do absolutely nothing
#else
AuAtomicAnd(&uValueRef, ~0xFFu);
InterlockedAndRelease((volatile LONG *)&uValueRef, ~0xFF);
#endif
while (true)
{
auto uValue = uValueRef;
//
if (uValue < kFutexBitWait)
{
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;
@ -268,7 +290,16 @@ namespace Aurora::Threading::Primitives
if (uValue & kFutexBitWake)
{
return;
// StoreLoad paranoia
if (AuAtomicCompareExchange(&uValueRef, uValue, uValue) == uValue)
{
return;
}
else
{
SMPPause();
continue;
}
}
if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait + kFutexBitWake, uValue) == uValue)