wine-staging/patches/ntdll-NtAlertThreadByThreadId/0009-ntdll-Reimplement-the-critical-section-fast-path-on-.patch

From 098b4dc3d374b5ed26dbcd0be643c961f1fa156f Mon Sep 17 00:00:00 2001
From: Zebediah Figura <z.figura12@gmail.com>
Date: Mon, 31 Aug 2020 23:38:09 -0500
Subject: [PATCH] ntdll: Reimplement the critical section fast path on top of
 Win32 futexes.

Signed-off-by: Zebediah Figura <z.figura12@gmail.com>
---
 dlls/ntdll/sync.c              |  35 +++++++----
 dlls/ntdll/unix/loader.c       |   3 -
 dlls/ntdll/unix/sync.c         | 109 ---------------------------------
 dlls/ntdll/unix/unix_private.h |   3 -
 dlls/ntdll/unixlib.h           |   5 +-
 5 files changed, 24 insertions(+), 131 deletions(-)

diff --git a/dlls/ntdll/sync.c b/dlls/ntdll/sync.c
index c73fb09da47..6edf104c5e9 100644
--- a/dlls/ntdll/sync.c
+++ b/dlls/ntdll/sync.c
@@ -738,19 +738,26 @@ static inline HANDLE get_semaphore( RTL_CRITICAL_SECTION *crit )
 
 static inline NTSTATUS wait_semaphore( RTL_CRITICAL_SECTION *crit, int timeout )
 {
-    NTSTATUS ret;
+    LARGE_INTEGER time = {.QuadPart = timeout * (LONGLONG)-10000000};
 
     /* debug info is cleared by MakeCriticalSectionGlobal */
-    if (!crit_section_has_debuginfo( crit ) ||
-        ((ret = unix_funcs->fast_RtlpWaitForCriticalSection( crit, timeout )) == STATUS_NOT_IMPLEMENTED))
+    if (!crit_section_has_debuginfo( crit ))
     {
         HANDLE sem = get_semaphore( crit );
-        LARGE_INTEGER time;
-
-        time.QuadPart = timeout * (LONGLONG)-10000000;
-        ret = NtWaitForSingleObject( sem, FALSE, &time );
+        return NtWaitForSingleObject( sem, FALSE, &time );
+    }
+    else
+    {
+        int *lock = (int *)&crit->LockSemaphore;
+        while (!InterlockedCompareExchange( lock, 0, 1 ))
+        {
+            static const int zero;
+            /* this may wait longer than specified in case of multiple wake-ups */
+            if (RtlWaitOnAddress( lock, &zero, sizeof(int), &time ) == STATUS_TIMEOUT)
+                return STATUS_TIMEOUT;
+        }
+        return STATUS_WAIT_0;
     }
-    return ret;
 }
 
 /******************************************************************************
@@ -840,8 +847,6 @@ NTSTATUS WINAPI RtlDeleteCriticalSection( RTL_CRITICAL_SECTION *crit )
             RtlFreeHeap( GetProcessHeap(), 0, crit->DebugInfo );
             crit->DebugInfo = NULL;
         }
-        if (unix_funcs->fast_RtlDeleteCriticalSection( crit ) == STATUS_NOT_IMPLEMENTED)
-            NtClose( crit->LockSemaphore );
     }
     else NtClose( crit->LockSemaphore );
     crit->LockSemaphore = 0;
@@ -917,12 +922,18 @@ NTSTATUS WINAPI RtlpUnWaitCriticalSection( RTL_CRITICAL_SECTION *crit )
     NTSTATUS ret;
 
     /* debug info is cleared by MakeCriticalSectionGlobal */
-    if (!crit_section_has_debuginfo( crit ) ||
-        ((ret = unix_funcs->fast_RtlpUnWaitCriticalSection( crit )) == STATUS_NOT_IMPLEMENTED))
+    if (!crit_section_has_debuginfo( crit ))
     {
         HANDLE sem = get_semaphore( crit );
         ret = NtReleaseSemaphore( sem, 1, NULL );
     }
+    else
+    {
+        int *lock = (int *)&crit->LockSemaphore;
+        InterlockedExchange( lock, 1 );
+        RtlWakeAddressSingle( lock );
+        ret = STATUS_SUCCESS;
+    }
     if (ret) RtlRaiseStatus( ret );
     return ret;
 }
diff --git a/dlls/ntdll/unix/loader.c b/dlls/ntdll/unix/loader.c
index 3a4a943c65f..579c73b2ad7 100644
--- a/dlls/ntdll/unix/loader.c
+++ b/dlls/ntdll/unix/loader.c
@@ -1816,9 +1816,6 @@ static struct unix_funcs unix_funcs =
 #endif
     DbgUiIssueRemoteBreakin,
     RtlGetSystemTimePrecise,
-    fast_RtlpWaitForCriticalSection,
-    fast_RtlpUnWaitCriticalSection,
-    fast_RtlDeleteCriticalSection,
     fast_RtlTryAcquireSRWLockExclusive,
     fast_RtlAcquireSRWLockExclusive,
     fast_RtlTryAcquireSRWLockShared,
diff --git a/dlls/ntdll/unix/sync.c b/dlls/ntdll/unix/sync.c
index 8394a1b8601..a7cce945e97 100644
--- a/dlls/ntdll/unix/sync.c
+++ b/dlls/ntdll/unix/sync.c
@@ -2484,115 +2484,6 @@ NTSTATUS WINAPI NtWaitForAlertByThreadId( const void *address, const LARGE_INTEG
 }
 
 
-#ifdef __linux__
-
-NTSTATUS CDECL fast_RtlpWaitForCriticalSection( RTL_CRITICAL_SECTION *crit, int timeout )
-{
-    int val;
-    struct timespec timespec;
-
-    if (!use_futexes()) return STATUS_NOT_IMPLEMENTED;
-
-    timespec.tv_sec  = timeout;
-    timespec.tv_nsec = 0;
-    while ((val = InterlockedCompareExchange( (int *)&crit->LockSemaphore, 0, 1 )) != 1)
-    {
-        /* note: this may wait longer than specified in case of signals or */
-        /*       multiple wake-ups, but that shouldn't be a problem */
-        if (futex_wait( (int *)&crit->LockSemaphore, val, &timespec ) == -1 && errno == ETIMEDOUT)
-            return STATUS_TIMEOUT;
-    }
-    return STATUS_WAIT_0;
-}
-
-NTSTATUS CDECL fast_RtlpUnWaitCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    if (!use_futexes()) return STATUS_NOT_IMPLEMENTED;
-
-    *(int *)&crit->LockSemaphore = 1;
-    futex_wake( (int *)&crit->LockSemaphore, 1 );
-    return STATUS_SUCCESS;
-}
-
-NTSTATUS CDECL fast_RtlDeleteCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    if (!use_futexes()) return STATUS_NOT_IMPLEMENTED;
-    return STATUS_SUCCESS;
-}
-
-#elif defined(__APPLE__)
-
-static inline semaphore_t get_mach_semaphore( RTL_CRITICAL_SECTION *crit )
-{
-    semaphore_t ret = *(int *)&crit->LockSemaphore;
-    if (!ret)
-    {
-        semaphore_t sem;
-        if (semaphore_create( mach_task_self(), &sem, SYNC_POLICY_FIFO, 0 )) return 0;
-        if (!(ret = InterlockedCompareExchange( (int *)&crit->LockSemaphore, sem, 0 )))
-            ret = sem;
-        else
-            semaphore_destroy( mach_task_self(), sem );  /* somebody beat us to it */
-    }
-    return ret;
-}
-
-NTSTATUS CDECL fast_RtlpWaitForCriticalSection( RTL_CRITICAL_SECTION *crit, int timeout )
-{
-    mach_timespec_t timespec;
-    semaphore_t sem = get_mach_semaphore( crit );
-
-    timespec.tv_sec = timeout;
-    timespec.tv_nsec = 0;
-    for (;;)
-    {
-        switch( semaphore_timedwait( sem, timespec ))
-        {
-        case KERN_SUCCESS:
-            return STATUS_WAIT_0;
-        case KERN_ABORTED:
-            continue;  /* got a signal, restart */
-        case KERN_OPERATION_TIMED_OUT:
-            return STATUS_TIMEOUT;
-        default:
-            return STATUS_INVALID_HANDLE;
-        }
-    }
-}
-
-NTSTATUS CDECL fast_RtlpUnWaitCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    semaphore_t sem = get_mach_semaphore( crit );
-    semaphore_signal( sem );
-    return STATUS_SUCCESS;
-}
-
-NTSTATUS CDECL fast_RtlDeleteCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    semaphore_destroy( mach_task_self(), *(int *)&crit->LockSemaphore );
-    return STATUS_SUCCESS;
-}
-
-#else  /* __APPLE__ */
-
-NTSTATUS CDECL fast_RtlpWaitForCriticalSection( RTL_CRITICAL_SECTION *crit, int timeout )
-{
-    return STATUS_NOT_IMPLEMENTED;
-}
-
-NTSTATUS CDECL fast_RtlpUnWaitCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    return STATUS_NOT_IMPLEMENTED;
-}
-
-NTSTATUS CDECL fast_RtlDeleteCriticalSection( RTL_CRITICAL_SECTION *crit )
-{
-    return STATUS_NOT_IMPLEMENTED;
-}
-
-#endif
-
-
 #ifdef __linux__
 
 /* Futex-based SRW lock implementation:
diff --git a/dlls/ntdll/unix/unix_private.h b/dlls/ntdll/unix/unix_private.h
index a5bade02b8a..bd7c4a8068c 100644
--- a/dlls/ntdll/unix/unix_private.h
+++ b/dlls/ntdll/unix/unix_private.h
@@ -82,9 +82,6 @@ extern void     (WINAPI *pKiUserApcDispatcher)(CONTEXT*,ULONG_PTR,ULONG_PTR,ULON
 extern NTSTATUS (WINAPI *pKiUserExceptionDispatcher)(EXCEPTION_RECORD*,CONTEXT*) DECLSPEC_HIDDEN;
 extern void     (WINAPI *pLdrInitializeThunk)(CONTEXT*,void**,ULONG_PTR,ULONG_PTR) DECLSPEC_HIDDEN;
 extern void     (WINAPI *pRtlUserThreadStart)( PRTL_THREAD_START_ROUTINE entry, void *arg ) DECLSPEC_HIDDEN;
-extern NTSTATUS CDECL fast_RtlpWaitForCriticalSection( RTL_CRITICAL_SECTION *crit, int timeout ) DECLSPEC_HIDDEN;
-extern NTSTATUS CDECL fast_RtlpUnWaitCriticalSection( RTL_CRITICAL_SECTION *crit ) DECLSPEC_HIDDEN;
-extern NTSTATUS CDECL fast_RtlDeleteCriticalSection( RTL_CRITICAL_SECTION *crit ) DECLSPEC_HIDDEN;
 extern NTSTATUS CDECL fast_RtlTryAcquireSRWLockExclusive( RTL_SRWLOCK *lock ) DECLSPEC_HIDDEN;
 extern NTSTATUS CDECL fast_RtlAcquireSRWLockExclusive( RTL_SRWLOCK *lock ) DECLSPEC_HIDDEN;
 extern NTSTATUS CDECL fast_RtlTryAcquireSRWLockShared( RTL_SRWLOCK *lock ) DECLSPEC_HIDDEN;
diff --git a/dlls/ntdll/unixlib.h b/dlls/ntdll/unixlib.h
index 21e34842778..c67daa389a3 100644
--- a/dlls/ntdll/unixlib.h
+++ b/dlls/ntdll/unixlib.h
@@ -26,7 +26,7 @@
 struct _DISPATCHER_CONTEXT;
 
 /* increment this when you change the function table */
-#define NTDLL_UNIXLIB_VERSION 121
+#define NTDLL_UNIXLIB_VERSION 122
 
 struct unix_funcs
 {
@@ -40,9 +40,6 @@ struct unix_funcs
     LONGLONG      (WINAPI *RtlGetSystemTimePrecise)(void);
 
     /* fast locks */
-    NTSTATUS      (CDECL *fast_RtlpWaitForCriticalSection)( RTL_CRITICAL_SECTION *crit, int timeout );
-    NTSTATUS      (CDECL *fast_RtlpUnWaitCriticalSection)( RTL_CRITICAL_SECTION *crit );
-    NTSTATUS      (CDECL *fast_RtlDeleteCriticalSection)( RTL_CRITICAL_SECTION *crit );
     NTSTATUS      (CDECL *fast_RtlTryAcquireSRWLockExclusive)( RTL_SRWLOCK *lock );
     NTSTATUS      (CDECL *fast_RtlAcquireSRWLockExclusive)( RTL_SRWLOCK *lock );
     NTSTATUS      (CDECL *fast_RtlTryAcquireSRWLockShared)( RTL_SRWLOCK *lock );
-- 
2.30.2