Fixes a Mutex Issue Preventing Some UE Titles From Booting (#451)

* Optimize guest import, memory, and pthread hot paths

* Fix UE adaptive mutex self-lock handling
This commit is contained in:
Spooks
2026-07-19 13:20:05 -06:00
committed by GitHub
parent 8ef5a54ee4
commit 90c72ebecf
9 changed files with 647 additions and 58 deletions
@@ -39,6 +39,60 @@ public sealed class PhysicalVirtualMemoryTests
Assert.Equal([(page, 0x1000UL, HostPageProtection.ReadWrite)], host.CommitCalls);
}
[Fact]
public void RepeatedLazyReadUsesCommittedRangeCache()
{
using var host = new LazyZeroedHostMemory();
using var memory = new PhysicalVirtualMemory(host);
var address = memory.AllocateAt(0, (4UL << 30) + 0x1000, executable: false);
host.CommitCalls.Clear();
Span<byte> buffer = stackalloc byte[16];
Assert.True(memory.TryRead(address + 0x100, buffer));
var queryCallsAfterFirstRead = host.QueryCalls;
Assert.True(memory.TryRead(address + 0x108, buffer[..8]));
Assert.Equal(queryCallsAfterFirstRead, host.QueryCalls);
Assert.Single(host.CommitCalls);
}
[Fact]
public void TryCopyHandlesOverlappingIdentityMappedRanges()
{
using var host = new LazyZeroedHostMemory();
using var memory = new PhysicalVirtualMemory(host);
var address = memory.AllocateAt(0, (4UL << 30) + 0x1000, executable: false);
Assert.True(memory.TryWrite(address, new byte[] { 1, 2, 3, 4, 5, 6 }));
Assert.True(memory.TryCopy(address + 2, address, 4));
Span<byte> result = stackalloc byte[6];
Assert.True(memory.TryRead(address, result));
Assert.Equal(new byte[] { 1, 2, 1, 2, 3, 4 }, result.ToArray());
}
[Fact]
public void RepeatedTryCopyKeepsSourceAndDestinationCommitRangesCached()
{
using var host = new LazyZeroedHostMemory();
using var memory = new PhysicalVirtualMemory(host);
var address = memory.AllocateAt(0, (4UL << 30) + 0x1000, executable: false);
var source = address + 0x100;
var destination = address + 0x1100;
Assert.True(memory.TryWrite(source, new byte[] { 1, 2, 3, 4 }));
Assert.True(memory.TryWrite(destination, new byte[4]));
host.CommitCalls.Clear();
Assert.True(memory.TryCopy(destination, source, 4));
var queryCallsAfterFirstCopy = host.QueryCalls;
Assert.True(memory.TryCopy(destination, source, 4));
Assert.Equal(queryCallsAfterFirstCopy, host.QueryCalls);
}
// 2. Reserve-only region: GetPointer commits the page before returning it,
// so callers receive a valid (non-null) pointer. An unmapped address yields null.
[Fact]
@@ -125,12 +179,14 @@ public sealed class PhysicalVirtualMemoryTests
public LazyZeroedHostMemory()
{
_allocation = System.Runtime.InteropServices.NativeMemory.AllocZeroed(0x2000);
_allocation = System.Runtime.InteropServices.NativeMemory.AllocZeroed(0x3000);
_address = ((ulong)_allocation + 0xFFF) & ~0xFFFUL;
}
public List<(ulong Address, ulong Size, HostPageProtection Protection)> CommitCalls { get; } = [];
public int QueryCalls { get; private set; }
public ulong Allocate(ulong desiredAddress, ulong size, HostPageProtection protection) => _address;
public ulong Reserve(ulong desiredAddress, ulong size, HostPageProtection protection) => _address;
@@ -162,6 +218,7 @@ public sealed class PhysicalVirtualMemoryTests
public bool Query(ulong address, out HostRegionInfo info)
{
QueryCalls++;
var pageAddress = address & ~0xFFFUL;
info = new HostRegionInfo(
pageAddress,
@@ -10,7 +10,7 @@ namespace SharpEmu.Libs.Tests.Pthread;
public sealed class PthreadMutexSemanticsTests
{
[Fact]
public void AdaptiveMutex_SelfLockUsesCompatibilityRecursion()
public void AdaptiveMutex_SelfLockIsIdempotent()
{
const ulong memoryBase = 0x1_0000_0000;
const ulong mutexAddress = memoryBase + 0x100;
@@ -22,7 +22,129 @@ public sealed class PthreadMutexSemanticsTests
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexLock(context));
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexLock(context));
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexUnlock(context));
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexUnlock(context));
Assert.NotEqual(0, KernelPthreadCompatExports.PthreadMutexUnlock(context));
}
[Fact]
public async Task ContendedMutex_HandsOffOneHostWaiterAtATime()
{
const ulong memoryBase = 0x2_0000_0000;
const ulong mutexAddress = memoryBase + 0x100;
var memory = new AllocatingCpuMemory(memoryBase, 0x4000);
var ownerContext = new CpuContext(memory, Generation.Gen5);
Assert.True(ownerContext.TryWriteUInt64(mutexAddress, 1));
ownerContext[CpuRegister.Rdi] = mutexAddress;
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexLock(ownerContext));
using var waitersStarted = new CountdownEvent(2);
using var firstAcquired = new ManualResetEventSlim(false);
using var secondAcquired = new ManualResetEventSlim(false);
using var releaseFirst = new ManualResetEventSlim(false);
var acquisitionCount = 0;
Task<(int LockResult, int UnlockResult)> StartWaiter() =>
Task.Factory.StartNew(
() =>
{
var waiterContext = new CpuContext(memory, Generation.Gen5);
waiterContext[CpuRegister.Rdi] = mutexAddress;
waitersStarted.Signal();
var lockResult = KernelPthreadCompatExports.PthreadMutexLock(waiterContext);
if (lockResult != 0)
{
return (lockResult, int.MinValue);
}
if (Interlocked.Increment(ref acquisitionCount) == 1)
{
firstAcquired.Set();
releaseFirst.Wait(TimeSpan.FromSeconds(5));
}
else
{
secondAcquired.Set();
}
var unlockResult = KernelPthreadCompatExports.PthreadMutexUnlock(waiterContext);
return (lockResult, unlockResult);
},
CancellationToken.None,
TaskCreationOptions.LongRunning,
TaskScheduler.Default);
var firstWaiter = StartWaiter();
var secondWaiter = StartWaiter();
Assert.True(waitersStarted.Wait(TimeSpan.FromSeconds(5)));
Thread.Sleep(50);
Assert.Equal(0, Volatile.Read(ref acquisitionCount));
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexUnlock(ownerContext));
Assert.True(firstAcquired.Wait(TimeSpan.FromSeconds(5)));
Assert.Equal(1, Volatile.Read(ref acquisitionCount));
releaseFirst.Set();
Assert.True(secondAcquired.Wait(TimeSpan.FromSeconds(5)));
var results = await Task.WhenAll(firstWaiter, secondWaiter).WaitAsync(TimeSpan.FromSeconds(5));
Assert.All(results, result => Assert.Equal((0, 0), result));
Assert.Equal(2, Volatile.Read(ref acquisitionCount));
}
[Fact]
public async Task ContendedMutex_PreservesMutualExclusionUnderLoad()
{
const ulong memoryBase = 0x3_0000_0000;
const ulong mutexAddress = memoryBase + 0x100;
const int workerCount = 4;
const int iterationsPerWorker = 250;
var memory = new AllocatingCpuMemory(memoryBase, 0x4000);
var initializationContext = new CpuContext(memory, Generation.Gen5);
Assert.True(initializationContext.TryWriteUInt64(mutexAddress, 1));
initializationContext[CpuRegister.Rdi] = mutexAddress;
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexLock(initializationContext));
Assert.Equal(0, KernelPthreadCompatExports.PthreadMutexUnlock(initializationContext));
using var start = new ManualResetEventSlim(false);
var insideCriticalSection = 0;
var mutualExclusionViolations = 0;
var protectedCounter = 0;
var workers = Enumerable.Range(0, workerCount)
.Select(_ => Task.Factory.StartNew(
() =>
{
var context = new CpuContext(memory, Generation.Gen5);
context[CpuRegister.Rdi] = mutexAddress;
start.Wait();
for (var iteration = 0; iteration < iterationsPerWorker; iteration++)
{
if (KernelPthreadCompatExports.PthreadMutexLock(context) != 0)
{
throw new InvalidOperationException("pthread mutex lock failed during contention stress.");
}
if (Interlocked.Increment(ref insideCriticalSection) != 1)
{
Interlocked.Increment(ref mutualExclusionViolations);
}
protectedCounter++;
Thread.SpinWait(20);
Interlocked.Decrement(ref insideCriticalSection);
if (KernelPthreadCompatExports.PthreadMutexUnlock(context) != 0)
{
throw new InvalidOperationException("pthread mutex unlock failed during contention stress.");
}
}
},
CancellationToken.None,
TaskCreationOptions.LongRunning,
TaskScheduler.Default))
.ToArray();
start.Set();
await Task.WhenAll(workers).WaitAsync(TimeSpan.FromSeconds(10));
Assert.Equal(0, Volatile.Read(ref mutualExclusionViolations));
Assert.Equal(workerCount * iterationsPerWorker, protectedCounter);
}
private sealed class AllocatingCpuMemory : ICpuMemory, IGuestMemoryAllocator