niansa-misc/wine
1
0
Fork 0
mirror of synced 2025-03-07 03:53:26 +01:00
Code Issues Projects Releases Packages Wiki Activity

ntdll: Don't hardcode xstate feature mask.

Browse source
This commit is contained in:
Paul Gofman 2024-01-31 14:24:24 -06:00 • committed by Alexandre Julliard
parent af4ef5fa02
commit 1e0728c5d4
4 changed files with 84 additions and 51 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

61
dlls/ntdll/unix/signal_i386.c
View file

@ -525,6 +525,7 @@ struct x86_thread_data
UINT dr7; /* 1f0 */
SYSTEM_SERVICE_TABLE *syscall_table; /* 1f4 syscall table */
struct syscall_frame *syscall_frame; /* 1f8 frame pointer on syscall entry */
UINT64 xstate_features_mask; /* 1fc */
};
C_ASSERT( sizeof(struct x86_thread_data) <= sizeof(((struct ntdll_thread_data *)0)->cpu_data) );
@ -609,8 +610,6 @@ struct xcontext
ULONG64 host_compaction_mask;
};
extern BOOL xstate_compaction_enabled;
static inline XSTATE *xstate_from_context( const CONTEXT *context )
{
CONTEXT_EX *xctx = (CONTEXT_EX *)(context + 1);
@ -832,7 +831,7 @@ static inline void save_context( struct xcontext *xcontext, const ucontext_t *si
context->ContextFlags |= CONTEXT_FLOATING_POINT | CONTEXT_EXTENDED_REGISTERS;
memcpy( context->ExtendedRegisters, fpux, sizeof(*fpux) );
if (!fpu) fpux_to_fpu( &context->FloatSave, fpux );
if ((cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX) && (xs = XState_sig(fpux)))
if (xstate_extended_features() && (xs = XState_sig(fpux)))
{
context_init_xstate( context, xs );
xcontext->host_compaction_mask = xs->CompactionMask;
@ -936,7 +935,7 @@ NTSTATUS WINAPI NtSetContextThread( HANDLE handle, const CONTEXT *context )
DWORD flags = context->ContextFlags & ~CONTEXT_i386;
BOOL self = (handle == GetCurrentThread());
if ((flags & CONTEXT_XSTATE) && (cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX))
if ((flags & CONTEXT_XSTATE) && xstate_extended_features())
{
CONTEXT_EX *context_ex = (CONTEXT_EX *)(context + 1);
XSTATE *xs = (XSTATE *)((char *)context_ex + context_ex->XState.Offset);
@ -944,7 +943,7 @@ NTSTATUS WINAPI NtSetContextThread( HANDLE handle, const CONTEXT *context )
if (context_ex->XState.Length < offsetof(XSTATE, YmmContext) ||
context_ex->XState.Length > sizeof(XSTATE))
return STATUS_INVALID_PARAMETER;
if ((xs->Mask & XSTATE_MASK_GSSE) && (context_ex->XState.Length < sizeof(XSTATE)))
if ((xs->Mask & xstate_extended_features()) && (context_ex->XState.Length < sizeof(XSTATE)))
return STATUS_BUFFER_OVERFLOW;
}
else flags &= ~CONTEXT_XSTATE;
@ -1138,7 +1137,7 @@ NTSTATUS WINAPI NtGetContextThread( HANDLE handle, CONTEXT *context )
context->ContextFlags |= CONTEXT_EXTENDED_REGISTERS;
}
if ((needed_flags & CONTEXT_XSTATE) && (cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX))
if ((needed_flags & CONTEXT_XSTATE) && xstate_extended_features())
{
CONTEXT_EX *context_ex = (CONTEXT_EX *)(context + 1);
XSTATE *xstate = (XSTATE *)((char *)context_ex + context_ex->XState.Offset);
@ -1148,7 +1147,7 @@ NTSTATUS WINAPI NtGetContextThread( HANDLE handle, CONTEXT *context )
|| context_ex->XState.Length > sizeof(XSTATE))
return STATUS_INVALID_PARAMETER;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & XSTATE_MASK_GSSE;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & xstate_extended_features();
xstate->Mask = frame->xstate.Mask & mask;
xstate->CompactionMask = xstate_compaction_enabled ? (0x8000000000000000 | mask) : 0;
memset( xstate->Reserved, 0, sizeof(xstate->Reserved) );
@ -1485,7 +1484,7 @@ static void setup_raise_exception( ucontext_t *sigcontext, void *stack_ptr,
context_init_xstate( &stack->context, dst_xs );
memset( dst_xs, 0, offsetof(XSTATE, YmmContext) );
dst_xs->CompactionMask = xstate_compaction_enabled ? 0x8000000000000004 : 0;
dst_xs->CompactionMask = xstate_compaction_enabled ? 0x8000000000000000 | xstate_extended_features() : 0;
if (src_xs->Mask & 4)
{
dst_xs->Mask = 4;
@ -1587,7 +1586,7 @@ NTSTATUS call_user_exception_dispatcher( EXCEPTION_RECORD *rec, CONTEXT *context
context_init_xstate( &stack->context, dst_xs );
memset( dst_xs, 0, offsetof(XSTATE, YmmContext) );
dst_xs->CompactionMask = xstate_compaction_enabled ? 0x8000000000000004 : 0;
dst_xs->CompactionMask = xstate_compaction_enabled ? 0x8000000000000000 | xstate_extended_features() : 0;
if (src_xs->Mask & 4)
{
dst_xs->Mask = 4;
@ -2481,6 +2480,7 @@ void call_init_thunk( LPTHREAD_START_ROUTINE entry, void *arg, BOOL suspend, TEB
ldt_set_fs( thread_data->fs, teb );
thread_data->gs = get_gs();
thread_data->syscall_table = KeServiceDescriptorTable;
thread_data->xstate_features_mask = xstate_supported_features_mask;
context.SegCs = get_cs();
context.SegDs = get_ds();
@ -2504,6 +2504,8 @@ void call_init_thunk( LPTHREAD_START_ROUTINE entry, void *arg, BOOL suspend, TEB
*ctx = context;
ctx->ContextFlags = CONTEXT_FULL | CONTEXT_FLOATING_POINT | CONTEXT_EXTENDED_REGISTERS;
memset( frame, 0, sizeof(*frame) );
if (xstate_compaction_enabled)
frame->xstate.CompactionMask = 0x8000000000000000 | xstate_supported_features_mask;
NtSetContextThread( GetCurrentThread(), ctx );
stack = (DWORD *)ctx;
@ -2605,26 +2607,27 @@ __ASM_GLOBAL_FUNC( __wine_syscall_dispatcher,
"addl %fs:0x1f4,%ebx\n\t" /* x86_thread_data()->syscall_table */
"testl $3,(%ecx)\n\t" /* frame->syscall_flags & (SYSCALL_HAVE_XSAVE | SYSCALL_HAVE_XSAVEC) */
"jz 2f\n\t"
"movl $7,%eax\n\t"
"xorl %edx,%edx\n\t"
"movl %edx,0x240(%ecx)\n\t"
"movl %edx,0x244(%ecx)\n\t"
"movl %edx,0x248(%ecx)\n\t"
"movl %edx,0x24c(%ecx)\n\t"
"movl %edx,0x250(%ecx)\n\t"
"movl %edx,0x254(%ecx)\n\t"
"movl %fs:0x1fc,%eax\n\t" /* x86_thread_data()->xstate_features_mask */
"movl %fs:0x200,%edx\n\t" /* x86_thread_data()->xstate_features_mask high dword */
"xorl %edi,%edi\n\t"
"movl %edi,0x240(%ecx)\n\t"
"movl %edi,0x244(%ecx)\n\t"
"movl %edi,0x248(%ecx)\n\t"
"movl %edi,0x24c(%ecx)\n\t"
"movl %edi,0x250(%ecx)\n\t"
"movl %edi,0x254(%ecx)\n\t"
"testl $2,(%ecx)\n\t" /* frame->syscall_flags & SYSCALL_HAVE_XSAVEC */
"jz 1f\n\t"
"movl %edx,0x258(%ecx)\n\t"
"movl %edx,0x25c(%ecx)\n\t"
"movl %edx,0x260(%ecx)\n\t"
"movl %edx,0x264(%ecx)\n\t"
"movl %edx,0x268(%ecx)\n\t"
"movl %edx,0x26c(%ecx)\n\t"
"movl %edx,0x270(%ecx)\n\t"
"movl %edx,0x274(%ecx)\n\t"
"movl %edx,0x278(%ecx)\n\t"
"movl %edx,0x27c(%ecx)\n\t"
"movl %edi,0x258(%ecx)\n\t"
"movl %edi,0x25c(%ecx)\n\t"
"movl %edi,0x260(%ecx)\n\t"
"movl %edi,0x264(%ecx)\n\t"
"movl %edi,0x268(%ecx)\n\t"
"movl %edi,0x26c(%ecx)\n\t"
"movl %edi,0x270(%ecx)\n\t"
"movl %edi,0x274(%ecx)\n\t"
"movl %edi,0x278(%ecx)\n\t"
"movl %edi,0x27c(%ecx)\n\t"
/* The xsavec instruction is not supported by
* binutils < 2.25. */
".byte 0x0f, 0xc7, 0x61, 0x40\n\t" /* xsavec 0x40(%ecx) */
@ -2669,8 +2672,8 @@ __ASM_GLOBAL_FUNC( __wine_syscall_dispatcher,
"testl $3,%ecx\n\t" /* SYSCALL_HAVE_XSAVE | SYSCALL_HAVE_XSAVEC */
"jz 1f\n\t"
"movl %eax,%esi\n\t"
"movl $7,%eax\n\t"
"xorl %edx,%edx\n\t"
"movl %fs:0x1fc,%eax\n\t" /* x86_thread_data()->xstate_features_mask */
"movl %fs:0x200,%edx\n\t" /* x86_thread_data()->xstate_features_mask high dword */
"xrstor 0x40(%esp)\n\t"
"movl %esi,%eax\n\t"
"jmp 3f\n"

61
dlls/ntdll/unix/signal_x86_64.c
View file

@ -441,6 +441,8 @@ struct amd64_thread_data
struct syscall_frame *syscall_frame; /* 0328 syscall frame pointer */
SYSTEM_SERVICE_TABLE *syscall_table; /* 0330 syscall table */
DWORD fs; /* 0338 WOW TEB selector */
DWORD align;
UINT64 xstate_features_mask; /* 0340 */
};
C_ASSERT( sizeof(struct amd64_thread_data) <= sizeof(((struct ntdll_thread_data *)0)->cpu_data) );
@ -448,6 +450,7 @@ C_ASSERT( offsetof( TEB, GdiTebBatch ) + offsetof( struct amd64_thread_data, pth
C_ASSERT( offsetof( TEB, GdiTebBatch ) + offsetof( struct amd64_thread_data, syscall_frame ) == 0x328 );
C_ASSERT( offsetof( TEB, GdiTebBatch ) + offsetof( struct amd64_thread_data, syscall_table ) == 0x330 );
C_ASSERT( offsetof( TEB, GdiTebBatch ) + offsetof( struct amd64_thread_data, fs ) == 0x338 );
C_ASSERT( offsetof( TEB, GdiTebBatch ) + offsetof( struct amd64_thread_data, xstate_features_mask ) == 0x340 );
static inline struct amd64_thread_data *amd64_thread_data(void)
{
@ -477,8 +480,6 @@ struct xcontext
ULONG64 host_compaction_mask;
};
extern BOOL xstate_compaction_enabled;
static inline XSTATE *xstate_from_context( const CONTEXT *context )
{
CONTEXT_EX *xctx = (CONTEXT_EX *)(context + 1);
@ -898,7 +899,7 @@ static void save_context( struct xcontext *xcontext, const ucontext_t *sigcontex
context->ContextFlags |= CONTEXT_FLOATING_POINT;
context->FltSave = *FPU_sig(sigcontext);
context->MxCsr = context->FltSave.MxCsr;
if ((cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX) && (xs = XState_sig(FPU_sig(sigcontext))))
if (xstate_extended_features() && (xs = XState_sig(FPU_sig(sigcontext))))
{
/* xcontext and sigcontext are both on the signal stack, so we can
* just reference sigcontext without overflowing 32 bit XState.Offset */
@ -928,7 +929,7 @@ static void restore_context( const struct xcontext *xcontext, ucontext_t *sigcon
amd64_thread_data()->dr7 = context->Dr7;
set_sigcontext( context, sigcontext );
if (FPU_sig(sigcontext)) *FPU_sig(sigcontext) = context->FltSave;
if ((cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX) && (xs = XState_sig(FPU_sig(sigcontext))))
if (xstate_extended_features() && (xs = XState_sig(FPU_sig(sigcontext))))
xs->CompactionMask = xcontext->host_compaction_mask;
leave_handler( sigcontext );
}
@ -977,7 +978,7 @@ NTSTATUS WINAPI NtSetContextThread( HANDLE handle, const CONTEXT *context )
BOOL self = (handle == GetCurrentThread());
struct syscall_frame *frame = amd64_thread_data()->syscall_frame;
if ((flags & CONTEXT_XSTATE) && (cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX))
if ((flags & CONTEXT_XSTATE) && xstate_extended_features())
{
CONTEXT_EX *context_ex = (CONTEXT_EX *)(context + 1);
XSTATE *xs = (XSTATE *)((char *)context_ex + context_ex->XState.Offset);
@ -985,7 +986,7 @@ NTSTATUS WINAPI NtSetContextThread( HANDLE handle, const CONTEXT *context )
if (context_ex->XState.Length < offsetof(XSTATE, YmmContext) ||
context_ex->XState.Length > sizeof(XSTATE))
return STATUS_INVALID_PARAMETER;
if ((xs->Mask & XSTATE_MASK_GSSE) && (context_ex->XState.Length < sizeof(XSTATE)))
if ((xs->Mask & xstate_extended_features()) && (context_ex->XState.Length < sizeof(XSTATE)))
return STATUS_BUFFER_OVERFLOW;
}
else flags &= ~CONTEXT_XSTATE;
@ -1155,7 +1156,7 @@ NTSTATUS WINAPI NtGetContextThread( HANDLE handle, CONTEXT *context )
context->MxCsr = context->FltSave.MxCsr;
context->ContextFlags |= CONTEXT_FLOATING_POINT;
}
if ((needed_flags & CONTEXT_XSTATE) && (cpu_info.ProcessorFeatureBits & CPU_FEATURE_AVX))
if ((needed_flags & CONTEXT_XSTATE) && xstate_extended_features())
{
CONTEXT_EX *context_ex = (CONTEXT_EX *)(context + 1);
XSTATE *xstate = (XSTATE *)((char *)context_ex + context_ex->XState.Offset);
@ -1165,7 +1166,7 @@ NTSTATUS WINAPI NtGetContextThread( HANDLE handle, CONTEXT *context )
|| context_ex->XState.Length > sizeof(XSTATE))
return STATUS_INVALID_PARAMETER;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & XSTATE_MASK_GSSE;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & xstate_extended_features();
xstate->Mask = frame->xstate.Mask & mask;
xstate->CompactionMask = xstate_compaction_enabled ? (0x8000000000000000 | mask) : 0;
memset( xstate->Reserved, 0, sizeof(xstate->Reserved) );
@ -1377,7 +1378,7 @@ NTSTATUS get_thread_wow64_context( HANDLE handle, void *ctx, ULONG size )
context_ex->XState.Length > sizeof(XSTATE))
return STATUS_INVALID_PARAMETER;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & XSTATE_MASK_GSSE;
mask = (xstate_compaction_enabled ? xstate->CompactionMask : xstate->Mask) & xstate_extended_features();
xstate->Mask = frame->xstate.Mask & mask;
xstate->CompactionMask = xstate_compaction_enabled ? (0x8000000000000000 | mask) : 0;
memset( xstate->Reserved, 0, sizeof(xstate->Reserved) );
@ -1442,7 +1443,7 @@ static void setup_raise_exception( ucontext_t *sigcontext, EXCEPTION_RECORD *rec
assert( !((ULONG_PTR)&stack->xstate & 63) );
context_init_xstate( &stack->context, &stack->xstate );
memset( &stack->xstate, 0, offsetof(XSTATE, YmmContext) );
stack->xstate.CompactionMask = xstate_compaction_enabled ? 0x8000000000000004 : 0;
stack->xstate.CompactionMask = xstate_compaction_enabled ? 0x8000000000000000 | xstate_extended_features() : 0;
if (src_xs->Mask & 4)
{
stack->xstate.Mask = 4;
@ -2479,6 +2480,7 @@ void call_init_thunk( LPTHREAD_START_ROUTINE entry, void *arg, BOOL suspend, TEB
I386_CONTEXT *wow_context;
thread_data->syscall_table = KeServiceDescriptorTable;
thread_data->xstate_features_mask = xstate_supported_features_mask;
#if defined __linux__
arch_prctl( ARCH_SET_GS, teb );
@ -2539,6 +2541,8 @@ void call_init_thunk( LPTHREAD_START_ROUTINE entry, void *arg, BOOL suspend, TEB
*ctx = context;
ctx->ContextFlags = CONTEXT_FULL;
memset( frame, 0, sizeof(*frame) );
if (xstate_compaction_enabled)
frame->xstate.CompactionMask = 0x8000000000000000 | xstate_supported_features_mask;
NtSetContextThread( GetCurrentThread(), ctx );
frame->cs = cs64_sel;
@ -2636,18 +2640,25 @@ __ASM_GLOBAL_FUNC( __wine_syscall_dispatcher,
"movl 0xb0(%rcx),%r14d\n\t" /* frame->syscall_flags */
"testl $3,%r14d\n\t" /* SYSCALL_HAVE_XSAVE | SYSCALL_HAVE_XSAVEC */
"jz 2f\n\t"
"movl $7,%eax\n\t"
"xorl %edx,%edx\n\t"
"movq %rdx,0x2c0(%rcx)\n\t"
"movq %rdx,0x2c8(%rcx)\n\t"
"movq %rdx,0x2d0(%rcx)\n\t"
#ifdef __APPLE__
"movq %gs:0x30,%rdx\n\t"
"movl 0x340(%rdx),%eax\n\t"
"movl 0x344(%rdx),%edx\n\t"
#else
"movl %gs:0x340,%eax\n\t" /* amd64_thread_data()->xstate_features_mask */
"movl %gs:0x344,%edx\n\t" /* amd64_thread_data()->xstate_features_mask high dword */
#endif
"xorq %rbp,%rbp\n\t"
"movq %rbp,0x2c0(%rcx)\n\t"
"movq %rbp,0x2c8(%rcx)\n\t"
"movq %rbp,0x2d0(%rcx)\n\t"
"testl $2,%r14d\n\t" /* SYSCALL_HAVE_XSAVEC */
"jz 1f\n\t"
"movq %rdx,0x2d8(%rcx)\n\t"
"movq %rdx,0x2e0(%rcx)\n\t"
"movq %rdx,0x2e8(%rcx)\n\t"
"movq %rdx,0x2f0(%rcx)\n\t"
"movq %rdx,0x2f8(%rcx)\n\t"
"movq %rbp,0x2d8(%rcx)\n\t"
"movq %rbp,0x2e0(%rcx)\n\t"
"movq %rbp,0x2e8(%rcx)\n\t"
"movq %rbp,0x2f0(%rcx)\n\t"
"movq %rbp,0x2f8(%rcx)\n\t"
/* The xsavec instruction is not supported by
* binutils < 2.25. */
".byte 0x48, 0x0f, 0xc7, 0xa1, 0xc0, 0x00, 0x00, 0x00\n\t" /* xsavec64 0xc0(%rcx) */
@ -2749,8 +2760,14 @@ __ASM_GLOBAL_FUNC( __wine_syscall_dispatcher,
"2:\ttestl $3,%r14d\n\t" /* SYSCALL_HAVE_XSAVE | SYSCALL_HAVE_XSAVEC */
"jz 3f\n\t"
"movq %rax,%r11\n\t"
"movl $7,%eax\n\t"
"xorl %edx,%edx\n\t"
#ifdef __APPLE__
"movq %gs:0x30,%rdx\n\t"
"movl 0x340(%rdx),%eax\n\t"
"movl 0x344(%rdx),%edx\n\t"
#else
"movl %gs:0x340,%eax\n\t" /* amd64_thread_data()->xstate_features_mask */
"movl %gs:0x344,%edx\n\t" /* amd64_thread_data()->xstate_features_mask high dword */
#endif
"xrstor64 0xc0(%rcx)\n\t"
"movq %r11,%rax\n\t"
"movl 0xb4(%rcx),%edx\n\t" /* frame->restore_flags */

4
dlls/ntdll/unix/system.c
View file

@ -247,6 +247,7 @@ static pthread_mutex_t timezone_mutex = PTHREAD_MUTEX_INITIALIZER;
#if defined(__i386__) || defined(__x86_64__)
BOOL xstate_compaction_enabled = FALSE;
UINT64 xstate_supported_features_mask;
#define AUTH 0x68747541 /* "Auth" */
#define ENTI 0x69746e65 /* "enti" */
@ -396,6 +397,9 @@ static void get_cpuinfo( SYSTEM_CPU_INFORMATION *info )
{
do_cpuid( 0x0000000d, 1, regs3 ); /* get XSAVE details */
if (regs3[0] & 2) xstate_compaction_enabled = TRUE;
xstate_supported_features_mask = 3;
if (features & CPU_FEATURE_AVX)
xstate_supported_features_mask |= (UINT64)1 << XSTATE_AVX;
}
if (regs[1] == AUTH && regs[3] == ENTI && regs[2] == CAMD)

9
dlls/ntdll/unix/unix_private.h
View file

@ -214,6 +214,15 @@ extern int server_pipe( int fd[2] );
extern void fpux_to_fpu( I386_FLOATING_SAVE_AREA *fpu, const XSAVE_FORMAT *fpux );
extern void fpu_to_fpux( XSAVE_FORMAT *fpux, const I386_FLOATING_SAVE_AREA *fpu );
extern BOOL xstate_compaction_enabled;
extern UINT64 xstate_supported_features_mask;
static inline UINT64 xstate_extended_features(void)
{
return xstate_supported_features_mask & ~(UINT64)3;
}
extern void *get_cpu_area( USHORT machine );
extern void set_thread_id( TEB *teb, DWORD pid, DWORD tid );
extern NTSTATUS init_thread_stack( TEB *teb, ULONG_PTR limit, SIZE_T reserve_size, SIZE_T commit_size );