[Bug 46252] StarCitizen has bad performance

https://bugs.winehq.org/show_bug.cgi?id=46252

--- Comment #10 from rawfox  ---
Comment on attachment 63216
  --> https://bugs.winehq.org/attachment.cgi?id=63216
merged futex conditions patch for wine vanilla
> --- a/dlls/ntdll/sync.c	2019-01-07 22:06:29.876808084 +0100
> +++ b/dlls/ntdll/sync.c	2019-01-07 23:51:22.897154211 +0100
> @@ -26,6 +26,7 @@
>
> #include <assert.h>
> #include <errno.h>
> +#include <limits.h>
> #include <signal.h>
> #ifdef HAVE_SYS_TIME_H
> # include <sys/time.h>
> @@ -36,6 +37,9 @@
> #ifdef HAVE_SYS_POLL_H
> # include <sys/poll.h>
> #endif
> +#ifdef HAVE_SYS_SYSCALL_H
> +# include <sys/syscall.h>
> +#endif
> #ifdef HAVE_UNISTD_H
> # include <unistd.h>
> #endif
> @@ -61,7 +65,140 @@
>
> HANDLE keyed_event = NULL;
> -
>
> static const LARGE_INTEGER zero_timeout;
> +#define TICKSPERSEC        10000000
> -
>
> +#ifdef __linux__
> -
>
> +static int wait_op = 128; /*FUTEX_WAIT|FUTEX_PRIVATE_FLAG*/
> +static int wake_op = 129; /*FUTEX_WAKE|FUTEX_PRIVATE_FLAG*/
> -
>
> +static inline int futex_wait( int *addr, int val, struct timespec *timeout )
> +{
> - return syscall( __NR_futex, addr, wait_op, val, timeout, 0, 0 );
>
> +}
> -
>
> +static inline int futex_wake( int *addr, int val )
> +{
> - return syscall( __NR_futex, addr, wake_op, val, NULL, 0, 0 );
>
> +}
> -
>
> +static inline int use_futexes(void)
> +{
> - static int supported = -1;
> -
> - if (supported == -1)
> - {
> -
>    futex_wait( &supported, 10, NULL );
> -
>    if (errno == ENOSYS)
> -
>    {
> -
>        wait_op = 0; /*FUTEX_WAIT*/
> -
>        wake_op = 1; /*FUTEX_WAKE*/
> -
>        futex_wait( &supported, 10, NULL );
> -
>    }
> -
>    supported = (errno != ENOSYS);
> - }
> - return supported;
>
> +}
> -
>
> +static inline NTSTATUS fast_wait( RTL_CONDITION_VARIABLE *variable, int val,
> - const LARGE_INTEGER *timeout)
>
> +{
> - struct timespec timespec;
> - LONGLONG timeleft;
> - LARGE_INTEGER now;
> - int ret;
> -
> - if (!use_futexes()) return STATUS_NOT_IMPLEMENTED;
> -
> - if (timeout && timeout->QuadPart != TIMEOUT_INFINITE)
> - {
> -
>    if (timeout->QuadPart >= 0)
> -
>    {
> -
>        NtQuerySystemTime( &now );
> -
>        timeleft = timeout->QuadPart - now.QuadPart;
> -
>        if (timeleft < 0) timeleft = 0;
> -
>    }
> -
>    else
> -
>        timeleft = -timeout->QuadPart;
> -
> -
>    timespec.tv_sec = timeleft / TICKSPERSEC;
> -
>    timespec.tv_nsec = (timeleft % TICKSPERSEC) * 100;
> -
> -
>    ret = futex_wait( (int *)&variable->Ptr, val, &timespec );
> - }
> - else
> -
>    ret = futex_wait( (int *)&variable->Ptr, val, NULL );
> -
> - if (ret == -1 && errno == ETIMEDOUT) return STATUS_TIMEOUT;
> - return STATUS_WAIT_0;
>
> +}
> -
>
> +static inline NTSTATUS fast_sleep_cs( RTL_CONDITION_VARIABLE *variable,
> - RTL_CRITICAL_SECTION *crit, const LARGE_INTEGER *timeout )
>
> +{
> - int val = *(int *)&variable->Ptr;
> - NTSTATUS ret;
> -
> - RtlLeaveCriticalSection( crit );
> -
> - ret = fast_wait( variable, val, timeout );
> -
> - RtlEnterCriticalSection( crit );
> -
> - return ret;
>
> +}
> -
>
> +static inline NTSTATUS fast_sleep_srw( RTL_CONDITION_VARIABLE *variable,
> - RTL_SRWLOCK *lock, const LARGE_INTEGER *timeout, ULONG flags )
>
> +{
> - int val = *(int *)&variable->Ptr;
> - NTSTATUS ret;
> -
> - if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>    RtlReleaseSRWLockShared( lock );
> - else
> -
>    RtlReleaseSRWLockExclusive( lock );
> -
> - ret = fast_wait( variable, val, timeout );
> -
> - if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>    RtlAcquireSRWLockShared( lock );
> - else
> -
>    RtlAcquireSRWLockExclusive( lock );
> -
> - return ret;
>
> +}
> -
>
> +static inline NTSTATUS fast_wake( RTL_CONDITION_VARIABLE *variable, int val )
> +{
> - if (!use_futexes()) return STATUS_NOT_IMPLEMENTED;
> -
> - interlocked_xchg_add( (int *)&variable->Ptr, 1 );
> - futex_wake( (int *)&variable->Ptr, val );
> - return STATUS_SUCCESS;
>
> +}
> -
>
> +#else
> +static inline NTSTATUS fast_sleep_cs( RTL_CONDITION_VARIABLE *variable,
> - RTL_CRITICAL_SECTION *crit, const LARGE_INTEGER *timeout )
>
> +{
> - return STATUS_NOT_IMPLEMENTED;
>
> +}
> -
>
> +static inline NTSTATUS fast_sleep_srw( RTL_CONDITION_VARIABLE *variable,
> - RTL_SRWLOCK *lock, const LARGE_INTEGER *timeout, ULONG flags )
>
> +{
> - return STATUS_NOT_IMPLEMENTED;
>
> +}
> -
>
> +static inline NTSTATUS fast_wake( RTL_CONDITION_VARIABLE *variable, int val )
> +{
> - return STATUS_NOT_IMPLEMENTED;
>
> +}
> +#endif
> -
>
> static inline int interlocked_dec_if_nonzero( int *dest )
> {
> @@ -1868,8 +2005,11 @@
>  */
> void WINAPI RtlWakeConditionVariable( RTL_CONDITION_VARIABLE *variable )
> {
> - if (interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>    NtReleaseKeyedEvent( 0, &variable->Ptr, FALSE, NULL );
> - if (fast_wake( variable, 1 ) == STATUS_NOT_IMPLEMENTED)
> - {
> -
>    if (interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>        NtReleaseKeyedEvent( keyed_event, &variable->Ptr, FALSE, NULL );
> - }
>
> }
>
> /***********************************************************************
> @@ -1879,9 +2019,12 @@
>  */
> void WINAPI RtlWakeAllConditionVariable( RTL_CONDITION_VARIABLE *variable )
> {
> - int val = interlocked_xchg( (int *)&variable->Ptr, 0 );
> - while (val-- > 0)
> -
>    NtReleaseKeyedEvent( 0, &variable->Ptr, FALSE, NULL );
> - if (fast_wake( variable, INT_MAX ) == STATUS_NOT_IMPLEMENTED)
> - {
> -
>    int val = interlocked_xchg( (int *)&variable->Ptr, 0 );
> -
>    while (val-- > 0)
> -
>        NtReleaseKeyedEvent( keyed_event, &variable->Ptr, FALSE, NULL );
> - }
>
> }
>
> /***********************************************************************
> @@ -1903,17 +2046,24 @@
>                                              const LARGE_INTEGER *timeout )
> {
>     NTSTATUS status;
> -
>
> interlocked_xchg_add( (int *)&variable->Ptr, 1 );
> -
>
> RtlLeaveCriticalSection( crit );
> -
>
> status = NtWaitForKeyedEvent( 0, &variable->Ptr, FALSE, timeout );
> -
>
> if (status != STATUS_SUCCESS)
> - if ((status = fast_sleep_cs( variable, crit, timeout )) == STATUS_NOT_IMPLEMENTED)
> {
> -
>    if (!interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>        status = NtWaitForKeyedEvent( 0, &variable->Ptr, FALSE, NULL );
> - }
> -
>    interlocked_xchg_add( (int *)&variable->Ptr, 1 );
> -
>    RtlLeaveCriticalSection( crit );
>
> - RtlEnterCriticalSection( crit );
> -
>    status = NtWaitForKeyedEvent( keyed_event, &variable->Ptr, FALSE, timeout );
> -
>    if (status != STATUS_SUCCESS)
> -
>    {
> -
>        if (!interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>            status = NtWaitForKeyedEvent( keyed_event, &variable->Ptr, FALSE, NULL );
> -
>    }
> -
> -
>    else if (status != STATUS_SUCCESS)
> -
>        interlocked_dec_if_nonzero( (int *)&variable->Ptr );
> -
> -
>    RtlEnterCriticalSection( crit );
> - }
> return status;
>
> }
>
> @@ -1940,24 +2090,31 @@
>                                               const LARGE_INTEGER *timeout, ULONG flags )
> {
>     NTSTATUS status;
> -
>
> interlocked_xchg_add( (int *)&variable->Ptr, 1 );
> -
> -
>
> if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>    RtlReleaseSRWLockShared( lock );
> -
>
> else
> -
>    RtlReleaseSRWLockExclusive( lock );
>
> -
>
> status = NtWaitForKeyedEvent( 0, &variable->Ptr, FALSE, timeout );
> -
>
> if (status != STATUS_SUCCESS)
> - if ((status = fast_sleep_srw( variable, lock, timeout, flags )) == STATUS_NOT_IMPLEMENTED)
> {
> -
>    if (!interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>        status = NtWaitForKeyedEvent( 0, &variable->Ptr, FALSE, NULL );
> - }
> -
>    interlocked_xchg_add( (int *)&variable->Ptr, 1 );
>
> - if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>    RtlAcquireSRWLockShared( lock );
> - else
> -
>    RtlAcquireSRWLockExclusive( lock );
> -
>    if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>        RtlReleaseSRWLockShared( lock );
> -
>    else
> -
>        RtlReleaseSRWLockExclusive( lock );
> -
> -
>    status = NtWaitForKeyedEvent( keyed_event, &variable->Ptr, FALSE, timeout );
> -
>    if (status != STATUS_SUCCESS)
> -
>    {
> -
>        if (!interlocked_dec_if_nonzero( (int *)&variable->Ptr ))
> -
>            status = NtWaitForKeyedEvent( keyed_event, &variable->Ptr, FALSE, NULL );
> -
>    }
> -
> -
>    else if (status != STATUS_SUCCESS)
> -
>        interlocked_dec_if_nonzero( (int *)&variable->Ptr );
> -
> -
>    if (flags & RTL_CONDITION_VARIABLE_LOCKMODE_SHARED)
> -
>        RtlAcquireSRWLockShared( lock );
> -
>    else
> -
>        RtlAcquireSRWLockExclusive( lock );
> - }
> return status;
>
> }
>
> @@ -1987,6 +2144,89 @@
>     return FALSE;
> }
>
> +#ifdef __linux__
> -
>
> +static int addr_futex_table[256];
> -
>
> +static inline int *hash_addr( const void *addr )
> +{
> - ULONG_PTR val = (ULONG_PTR)addr;
> -
> - return &addr_futex_table[(val >> 2) & 255];
>
> +}
> -
>
> +static inline NTSTATUS fast_wait_addr( const void *addr, const void *cmp, SIZE_T size,
> -
>                                   const LARGE_INTEGER *timeout )
>
> +{
> - int *futex;
> - int val;
> - LARGE_INTEGER now;
> - timeout_t diff;
> - struct timespec timespec;
> - int ret;
> -
> - if (!use_futexes())
> -
>    return STATUS_NOT_IMPLEMENTED;
> -
> - futex = hash_addr( addr );
> -
> - /* We must read the previous value of the futex before checking the value
> -
> * of the address being waited on. That way, if we receive a wake between
> -
> * now and waiting on the futex, we know that val will have changed. */
> - val = interlocked_cmpxchg( futex, 0, 0 );
> - if (!compare_addr( addr, cmp, size ))
> -
>    return STATUS_SUCCESS;
> -
> - if (timeout)
> - {
> -
>    if (timeout->QuadPart > 0)
> -
>    {
> -
>        NtQuerySystemTime( &now );
> -
>        diff = timeout->QuadPart - now.QuadPart;
> -
>    }
> -
>    else
> -
>        diff = -timeout->QuadPart;
> -
> -
>    timespec.tv_sec  = diff / 1000000;
> -
>    timespec.tv_nsec = diff % 1000000;
> -
> -
>    ret = futex_wait( futex, val, &timespec );
> - }
> - else
> -
>    ret = futex_wait( futex, val, NULL );
> -
> - if (ret == -1 && errno == ETIMEDOUT)
> -
>    return STATUS_TIMEOUT;
> - return STATUS_SUCCESS;
>
> +}
> -
>
> +static inline NTSTATUS fast_wake_addr( const void *addr )
> +{
> - int *futex;
> -
> - if (!use_futexes())
> -
>    return STATUS_NOT_IMPLEMENTED;
> -
> - futex = hash_addr( addr );
> -
> - interlocked_xchg_add( futex, 1 );
> -
> - futex_wake( futex, INT_MAX );
> - return STATUS_SUCCESS;
>
> +}
> +#else
> +static inline NTSTATUS fast_wait_addr( const void *addr, const void *cmp, SIZE_T size,
> -
>                                   const LARGE_INTEGER *timeout )
>
> +{
> - return STATUS_NOT_IMPLEMENTED;
>
> +}
> -
>
> +static inline NTSTATUS fast_wake_addr( const void *addr )
> +{
> - return STATUS_NOT_IMPLEMENTED;
>
> +}
> +#endif
> -
>
> /***********************************************************************
> -
>       RtlWaitOnAddress   (NTDLL.@)
>
> */
> @@ -2005,6 +2245,9 @@
>     if (size != 1 && size != 2 && size != 4 && size != 8)
>         return STATUS_INVALID_PARAMETER;
> - if ((ret = fast_wait_addr( addr, cmp, size, timeout )) != STATUS_NOT_IMPLEMENTED)
> -
>    return ret;
> - select_op.keyed_event.op     = SELECT_KEYED_EVENT_WAIT;
>   select_op.keyed_event.handle = wine_server_obj_handle( keyed_event );
>   select_op.keyed_event.key    = wine_server_client_ptr( addr );
>
> @@ -2059,6 +2302,9 @@
>  */
> void WINAPI RtlWakeAddressAll( const void *addr )
> {
> - if (fast_wake_addr( addr ) != STATUS_NOT_IMPLEMENTED)
> -
>    return;
> - RtlEnterCriticalSection( &addr_section );
>   while (NtReleaseKeyedEvent( 0, addr, 0, &zero_timeout ) == STATUS_SUCCESS) {}
>   RtlLeaveCriticalSection( &addr_section );
>
> @@ -2069,6 +2315,9 @@
>  */
> void WINAPI RtlWakeAddressSingle( const void *addr )
> {
> - if (fast_wake_addr( addr ) != STATUS_NOT_IMPLEMENTED)
> -
>    return;
> - RtlEnterCriticalSection( &addr_section );
>   NtReleaseKeyedEvent( 0, addr, 0, &zero_timeout );
>   RtlLeaveCriticalSection( &addr_section );