ntdll: finish ARM64EC cooperative suspend (fixes named-pipe stalls under FEX)

dlls/ntdll/ntdll_misc.h

@@ -181,7 +181,6 @@ extern IMAGE_ARM64EC_METADATA *arm64ec_get_module_metadata( HMODULE module );
 extern void arm64ec_update_hybrid_metadata( void *module, IMAGE_NT_HEADERS *nt,
                                             const IMAGE_ARM64EC_METADATA *metadata );
 extern void invoke_arm64ec_syscall(void);
-extern void arm64ec_suspend_point(void);
 
 extern void *__os_arm64x_check_call;
 extern void *__os_arm64x_check_icall;

dlls/ntdll/signal_arm64ec.c

@@ -90,7 +90,6 @@ static inline BOOL enter_syscall_callback(void)
 static inline void leave_syscall_callback(void)
 {
     get_arm64ec_cpu_area()->InSyscallCallback = 0;
-    if (get_arm64ec_cpu_area()->SuspendDoorbell && *get_arm64ec_cpu_area()->SuspendDoorbell) arm64ec_suspend_point();
 }
 
 /**********************************************************************

dlls/ntdll/signal_x86_64.c

@@ -74,14 +74,6 @@ __ASM_GLOBAL_FUNC( invoke_arm64ec_syscall,
                    "pushq %r10\n\t"         /* and return to syscall thunk */
                    "ret" )
 
-/**************************************************************************
- *		arm64ec_suspend_point
- *
- * x64 stub to support cooperative suspend when leaving a syscall callack.
- */
-__ASM_GLOBAL_FUNC( arm64ec_suspend_point,
-                   "ret" )
-
 /*******************************************************************
  *		KiUserExceptionDispatcher (NTDLL.@)
  */

dlls/ntdll/unix/signal_arm64.c

@@ -213,10 +213,6 @@ struct syscall_frame
 
 C_ASSERT( sizeof( struct syscall_frame ) == 0x330 );
 
-static BOOL is_arm64ec_suspend_doorbell_valid(void)
-{
-    return is_arm64ec() && NtCurrentTeb()->ChpeV2CpuAreaInfo && NtCurrentTeb()->ChpeV2CpuAreaInfo->SuspendDoorbell;
-}
 
 /***********************************************************************
  *           context_init_empty_xstate
@@ -241,6 +237,7 @@ void set_process_instrumentation_callback( void *callback )
     if (callback) FIXME( "Not supported.\n" );
 }
 
+
 /***********************************************************************
  *           syscall_frame_fixup_for_fastpath
  *
@@ -347,20 +344,22 @@ static void restore_context( const CONTEXT *context, ucontext_t *sigcontext )
 NTSTATUS signal_set_full_context( CONTEXT *context )
 {
     struct syscall_frame *frame = get_syscall_frame();
-    NTSTATUS status = NtSetContextThread( GetCurrentThread(), context );
+    struct arm64_thread_data *arm64_data = arm64_thread_data();
+    NTSTATUS status;
 
-    if (is_arm64ec_suspend_doorbell_valid() && arm64_thread_data()->suspend_pending)
+    if (arm64_data->suspend_pending)
     {
-        CONTEXT suspend_context;
+        sigset_t old_set;
+        pthread_sigmask( SIG_BLOCK, &server_block_set, &old_set );
         *NtCurrentTeb()->ChpeV2CpuAreaInfo->SuspendDoorbell = 0;
-        arm64_thread_data()->suspend_pending = FALSE;
-        suspend_context.ContextFlags = CONTEXT_FULL | CONTEXT_EXCEPTION_REPORTING; /* TODO: check */
-        NtGetContextThread( GetCurrentThread(), &suspend_context );
-        wait_suspend( &suspend_context );
-        NtSetContextThread( GetCurrentThread(), &suspend_context );
+        arm64_data->suspend_pending = FALSE;
+        wait_suspend( context );
+        status = NtSetContextThread( GetCurrentThread(), context );
+        pthread_sigmask( SIG_SETMASK, &old_set, NULL );
     }
+    else status = NtSetContextThread( GetCurrentThread(), context );
 
-    if (!status && (context->ContextFlags & CONTEXT_INTEGER) == CONTEXT_INTEGER) /* TODO: also check with susp */
+    if (!status && (context->ContextFlags & CONTEXT_INTEGER) == CONTEXT_INTEGER)
         frame->restore_flags |= CONTEXT_INTEGER;
 
     if (is_arm64ec() && !is_ec_code( frame->pc ))
@@ -451,20 +450,12 @@ NTSTATUS WINAPI NtGetContextThread( HANDLE handle, CONTEXT *context )
 {
     struct syscall_frame *frame = get_syscall_frame();
     DWORD needed_flags = context->ContextFlags & ~CONTEXT_ARM64;
-    THREAD_BASIC_INFORMATION info;
-    NTSTATUS ret;
-    BOOL self;
-
-    NtQueryInformationThread( handle, ThreadBasicInformation, &info, sizeof(info), NULL );
-    self = HandleToULong( info.ClientId.UniqueThread ) == GetCurrentThreadId();
+    BOOL self = (handle == GetCurrentThread());
 
     if (!self)
     {
-        /* Avoid exposing JIT code pointers to other processes on ARM64EC */
-        if (is_arm64ec()) NtSuspendThread( handle, NULL );
-        ret = get_thread_context( handle, context, &self, IMAGE_FILE_MACHINE_ARM64 );
-        if (is_arm64ec()) NtResumeThread( handle, NULL );
-        return ret;
+        NTSTATUS ret = get_thread_context( handle, context, &self, IMAGE_FILE_MACHINE_ARM64 );
+        if (ret || !self) return ret;
     }
 
     if (needed_flags & CONTEXT_INTEGER)
@@ -1345,17 +1336,20 @@ static void quit_handler( int signal, siginfo_t *siginfo, void *sigcontext )
 static void usr1_handler( int signal, siginfo_t *siginfo, void *sigcontext )
 {
     ucontext_t *ucontext = sigcontext;
+    CHPE_V2_CPU_AREA_INFO *chpe;
     CONTEXT context;
 
-    if (is_arm64ec_suspend_doorbell_valid() &&
-        (NtCurrentTeb()->ChpeV2CpuAreaInfo->InSimulation || NtCurrentTeb()->ChpeV2CpuAreaInfo->InSyscallCallback))
+    if ((chpe = NtCurrentTeb()->ChpeV2CpuAreaInfo) && chpe->InSimulation && chpe->SuspendDoorbell)
     {
-        *NtCurrentTeb()->ChpeV2CpuAreaInfo->SuspendDoorbell = 1;
-        arm64_thread_data()->suspend_pending = TRUE;
-        return;
+        NTSTATUS status = server_select( NULL, 0, SELECT_INTERRUPTIBLE | SELECT_COOPERATIVE_SUSPEND,
+                                         0, NULL, NULL );
+        if (status == STATUS_THREAD_WAS_SUSPENDED)
+        {
+            *chpe->SuspendDoorbell = -1;
+            arm64_thread_data()->suspend_pending = TRUE;
+        }
     }
-
-    if (is_inside_syscall( SP_sig(ucontext) ))
+    else if (is_inside_syscall( SP_sig(ucontext) ))
     {
         context.ContextFlags = CONTEXT_FULL | CONTEXT_EXCEPTION_REQUEST;
         NtGetContextThread( GetCurrentThread(), &context );
@@ -1510,14 +1504,11 @@ void syscall_dispatcher_return_slowpath(void)
  */
 void init_syscall_frame( LPTHREAD_START_ROUTINE entry, void *arg, BOOL suspend, TEB *teb )
 {
-    struct ntdll_thread_data *thread_data = (struct ntdll_thread_data *)&teb->GdiTebBatch;
-    struct syscall_frame *frame = thread_data->syscall_frame;
+    struct syscall_frame *frame = ((struct ntdll_thread_data *)&teb->GdiTebBatch)->syscall_frame;
     CONTEXT *ctx, context = { CONTEXT_ALL };
     I386_CONTEXT *i386_context;
     ARM_CONTEXT *arm_context;
 
-    ((struct arm64_thread_data *)(thread_data->cpu_data))->suspend_pending = FALSE;
-
     context.X0  = (DWORD64)entry;
     context.X1  = (DWORD64)arg;
     context.X18 = (DWORD64)teb;

server/protocol.def

@@ -1463,8 +1463,9 @@ struct cpu_topology_override
     VARARG(call,apc_call);     /* APC call arguments */
     VARARG(contexts,contexts); /* suspend context(s) */
 @END
-#define SELECT_ALERTABLE     1
-#define SELECT_INTERRUPTIBLE 2
+#define SELECT_ALERTABLE           1
+#define SELECT_INTERRUPTIBLE       2
+#define SELECT_COOPERATIVE_SUSPEND 4
 
 
 /* Create an event */

server/thread.c

@@ -134,6 +134,7 @@ struct context
     struct object           obj;        /* object header */
     struct object          *sync;       /* sync object for wait/signal */
     unsigned int            status;     /* status of the context */
+    int                     cooperative;/* waiting for the cooperative suspend */
     struct context_data     regs[2];    /* context data */
 };
 #define CTX_NATIVE  0  /* context for native machine */
@@ -446,6 +447,7 @@ static inline void init_thread_structure( struct thread *thread )
 
 static inline int is_thread_suspended( struct thread *thread )
 {
+    if (thread->context && thread->context->cooperative) return 0;
     if (thread->suspend) return 1;
     return !thread->bypass_proc_suspend && thread->process->suspend;
 }
@@ -486,8 +488,9 @@ static struct context *create_thread_context( struct thread *thread )
 {
     struct context *context;
     if (!(context = alloc_object( &context_ops ))) return NULL;
-    context->sync   = NULL;
-    context->status = STATUS_PENDING;
+    context->sync        = NULL;
+    context->status      = STATUS_PENDING;
+    context->cooperative = 0;
     memset( &context->regs, 0, sizeof(context->regs) );
     context->regs[CTX_NATIVE].machine = native_machine;
 
@@ -1193,6 +1196,12 @@ static int check_wait( struct thread *thread )
     if ((wait->flags & SELECT_INTERRUPTIBLE) && !list_empty( &thread->system_apc ))
         return STATUS_KERNEL_APC;
 
+    if ((wait->flags & SELECT_COOPERATIVE_SUSPEND) && thread->context)
+    {
+        thread->context->cooperative = 1;
+        return STATUS_THREAD_WAS_SUSPENDED;
+    }
+
     /* Suspended threads may not acquire locks, but they can run system APCs */
     if (is_thread_suspended( thread )) return -1;
 
@@ -2058,6 +2067,7 @@ DECL_HANDLER(select)
             copy_context( &ctx->regs[CTX_WOW], wow_context, wow_context->flags & ~ctx->regs[CTX_WOW].flags );
         }
         ctx->status = STATUS_SUCCESS;
+        ctx->cooperative = 0;
         current->suspend_cookie = req->cookie;
         signal_sync( ctx->sync );
     }