Based on [a patch](https://www.winehq.org/mailman3/hyperkitty/list/wine-devel@winehq.org/messag...) by Jinoh Kang (@iamahuman) from February 2022.
I removed the need for the event object and implemented fast paths for Linux. On macOS 10.14+ `thread_get_register_pointer_values` is used on every thread of the process. On Linux 4.14+ `membarrier(MEMBARRIER_CMD_GLOBAL_EXPEDITED, ...)` is used. On x86 Linux <= 4.13 `madvise(..., MADV_DONTNEED)` is used, which sends IPIs to all cores causing them to do a memory barrier. On non-x86 Linux <= 4.2 and on other platforms the fallback path using APCs is used.
-- v3: ntdll: Add thread_get_register_pointer_values-based fast path for NtFlushProcessWriteBuffers. ntdll: Add sys_membarrier-based fast path to NtFlushProcessWriteBuffers. ntdll: Add MADV_DONTNEED-based fast path for NtFlushProcessWriteBuffers. ntdll: Make server_select a memory barrier. ntdll: Implement NtFlushProcessWriteBuffers.
From: Torge Matthies tmatthies@codeweavers.com
Based on a patch by Jinoh Kang from February 2022 [1]. The following description is copied from said patch:
NtFlushProcessWriteBuffers is the NT equivalent of Linux membarrier() system call. The .NET Framework garbage collector uses it to synchronize with other threads, and thus is required to avoid silent memory corruption.
[1] https://www.winehq.org/mailman3/hyperkitty/list/wine-devel@winehq.org/messag... --- dlls/ntdll/unix/server.c | 6 ++ dlls/ntdll/unix/virtual.c | 27 ++++++++- server/protocol.def | 19 ++++++- server/thread.c | 116 ++++++++++++++++++++++++++++++++++++++ server/thread.h | 1 + 5 files changed, 165 insertions(+), 4 deletions(-)
diff --git a/dlls/ntdll/unix/server.c b/dlls/ntdll/unix/server.c index 77e8d5c7566..51a83f472e1 100644 --- a/dlls/ntdll/unix/server.c +++ b/dlls/ntdll/unix/server.c @@ -574,6 +574,12 @@ static void invoke_system_apc( const apc_call_t *call, apc_result_t *result, BOO if (!self) NtClose( wine_server_ptr_handle(call->dup_handle.dst_process) ); break; } + case APC_MEMORY_BARRIER: + { + MemoryBarrier(); + result->type = call->type; + break; + } default: server_protocol_error( "get_apc_request: bad type %d\n", call->type ); break; diff --git a/dlls/ntdll/unix/virtual.c b/dlls/ntdll/unix/virtual.c index 96a5e095d16..0d2ce328371 100644 --- a/dlls/ntdll/unix/virtual.c +++ b/dlls/ntdll/unix/virtual.c @@ -5020,8 +5020,31 @@ NTSTATUS WINAPI NtFlushInstructionCache( HANDLE handle, const void *addr, SIZE_T */ void WINAPI NtFlushProcessWriteBuffers(void) { - static int once = 0; - if (!once++) FIXME( "stub\n" ); + static pthread_mutex_t apc_memorybarrier_mutex = PTHREAD_MUTEX_INITIALIZER; + NTSTATUS status; + + pthread_mutex_lock( &apc_memorybarrier_mutex ); + + do + { + SERVER_START_REQ( flush_process_write_buffers ) + { + status = wine_server_call( req ); + } + SERVER_END_REQ; + } + while (status); + + do + { + select_op_t select_op; + select_op.membarrier.op = SELECT_MEMBARRIER; + status = server_select( &select_op, sizeof(select_op.membarrier), SELECT_INTERRUPTIBLE, + TIMEOUT_INFINITE, NULL, NULL ); + } + while (status); + + pthread_mutex_unlock( &apc_memorybarrier_mutex ); }
diff --git a/server/protocol.def b/server/protocol.def index d828d41d1f7..eb062d3c7a3 100644 --- a/server/protocol.def +++ b/server/protocol.def @@ -462,7 +462,8 @@ enum select_op SELECT_WAIT_ALL, SELECT_SIGNAL_AND_WAIT, SELECT_KEYED_EVENT_WAIT, - SELECT_KEYED_EVENT_RELEASE + SELECT_KEYED_EVENT_RELEASE, + SELECT_MEMBARRIER };
typedef union @@ -485,6 +486,10 @@ typedef union obj_handle_t handle; client_ptr_t key; } keyed_event; + struct + { + enum select_op op; /* SELECT_MEMBARRIER */ + } membarrier; } select_op_t;
enum apc_type @@ -502,7 +507,8 @@ enum apc_type APC_MAP_VIEW, APC_UNMAP_VIEW, APC_CREATE_THREAD, - APC_DUP_HANDLE + APC_DUP_HANDLE, + APC_MEMORY_BARRIER };
typedef struct @@ -611,6 +617,10 @@ typedef union unsigned int attributes; /* object attributes */ unsigned int options; /* duplicate options */ } dup_handle; + struct + { + enum apc_type type; /* APC_MEMORY_BARRIER */ + } memory_barrier; } apc_call_t;
typedef union @@ -1610,6 +1620,11 @@ enum server_fd_type @END
+/* Issue a memory barrier on other threads in the same process */ +@REQ(flush_process_write_buffers) +@END + + struct thread_info { timeout_t start_time; diff --git a/server/thread.c b/server/thread.c index f49fbf40b78..1f052092b56 100644 --- a/server/thread.c +++ b/server/thread.c @@ -112,6 +112,41 @@ static const struct object_ops thread_apc_ops = thread_apc_destroy /* destroy */ };
+/* process-wide memory barriers */ + +struct memory_barrier +{ + struct object obj; /* object header */ +}; + +static void dump_memory_barrier( struct object *obj, int verbose ); +static int memory_barrier_signaled( struct object *obj, struct wait_queue_entry *entry ); + +static const struct object_ops memory_barrier_ops = +{ + sizeof(struct memory_barrier), /* size */ + &no_type, /* type */ + dump_memory_barrier, /* dump */ + add_queue, /* add_queue */ + remove_queue, /* remove_queue */ + memory_barrier_signaled, /* signaled */ + no_satisfied, /* satisfied */ + no_signal, /* signal */ + no_get_fd, /* get_fd */ + default_map_access, /* map_access */ + default_get_sd, /* get_sd */ + default_set_sd, /* set_sd */ + no_get_full_name, /* get_full_name */ + no_lookup_name, /* lookup_name */ + no_link_name, /* link_name */ + NULL, /* unlink_name */ + no_open_file, /* open_file */ + no_kernel_obj_list, /* get_kernel_obj_list */ + no_close_handle, /* close_handle */ + no_destroy /* destroy */ +}; + +struct memory_barrier *memory_barrier_obj;
/* thread CPU context */
@@ -249,6 +284,7 @@ static inline void init_thread_structure( struct thread *thread ) thread->token = NULL; thread->desc = NULL; thread->desc_len = 0; + thread->mb_apcs_pending = 0;
thread->creation_time = current_time; thread->exit_time = 0; @@ -306,6 +342,11 @@ struct thread *create_thread( int fd, struct process *process, const struct secu struct thread *thread; int request_pipe[2];
+ if (memory_barrier_obj) + grab_object( &memory_barrier_obj->obj ); + else if (!(memory_barrier_obj = alloc_object( &memory_barrier_ops ))) + return NULL; + if (fd == -1) { if (pipe( request_pipe ) == -1) @@ -441,12 +482,14 @@ static void cleanup_thread( struct thread *thread ) thread->desktop = 0; thread->desc = NULL; thread->desc_len = 0; + thread->mb_apcs_pending = 0; }
/* destroy a thread when its refcount is 0 */ static void destroy_thread( struct object *obj ) { struct thread *thread = (struct thread *)obj; + struct memory_barrier *mb = memory_barrier_obj; assert( obj->ops == &thread_ops );
list_remove( &thread->entry ); @@ -454,6 +497,9 @@ static void destroy_thread( struct object *obj ) release_object( thread->process ); if (thread->id) free_ptid( thread->id ); if (thread->token) release_object( thread->token ); + if (mb->obj.refcount == 1) + memory_barrier_obj = NULL; + release_object( &mb->obj ); }
/* dump a thread on stdout for debugging purposes */ @@ -526,6 +572,18 @@ static struct thread_apc *create_apc( struct object *owner, const apc_call_t *ca return apc; }
+static void dump_memory_barrier( struct object *obj, int verbose ) +{ + assert( obj->ops == &memory_barrier_ops ); + fprintf( stderr, "Memory barrier\n" ); +} + +static int memory_barrier_signaled( struct object *obj, struct wait_queue_entry *entry ) +{ + struct thread *thread = entry->wait->thread; + return !thread->mb_apcs_pending; +} + /* get a thread pointer from a thread id (and increment the refcount) */ struct thread *get_thread_from_id( thread_id_t id ) { @@ -1029,6 +1087,13 @@ static int select_on( const select_op_t *select_op, data_size_t op_size, client_ current->wait->key = select_op->keyed_event.key; break;
+ case SELECT_MEMBARRIER: + object = &memory_barrier_obj->obj; + if (!object) return 1; + ret = wait_on( select_op, 1, &object, flags, when ); + if (!ret) return 1; + break; + default: set_error( STATUS_INVALID_PARAMETER ); return 1; @@ -1165,6 +1230,16 @@ int thread_queue_apc( struct process *process, struct thread *thread, struct obj return ret; }
+static void finish_membarrier_apc( struct thread_apc *apc ) +{ + struct thread *thread = (struct thread *)apc->owner; + + assert( thread ); + assert( thread->mb_apcs_pending > 0 ); + if (--thread->mb_apcs_pending) + wake_up( &memory_barrier_obj->obj, 1 ); +} + /* cancel the async procedure call owned by a specific object */ void thread_cancel_apc( struct thread *thread, struct object *owner, enum apc_type type ) { @@ -1176,6 +1251,8 @@ void thread_cancel_apc( struct thread *thread, struct object *owner, enum apc_ty if (apc->owner != owner) continue; list_remove( &apc->entry ); apc->executed = 1; + if (apc->call.type == APC_MEMORY_BARRIER) + finish_membarrier_apc( apc ); wake_up( &apc->obj, 0 ); release_object( apc ); return; @@ -1206,6 +1283,8 @@ static void clear_apc_queue( struct list *queue ) struct thread_apc *apc = LIST_ENTRY( ptr, struct thread_apc, entry ); list_remove( &apc->entry ); apc->executed = 1; + if (apc->call.type == APC_MEMORY_BARRIER) + finish_membarrier_apc( apc ); wake_up( &apc->obj, 0 ); release_object( apc ); } @@ -1650,6 +1729,8 @@ DECL_HANDLER(select) apc->result.create_thread.handle = handle; clear_error(); /* ignore errors from the above calls */ } + if (apc->call.type == APC_MEMORY_BARRIER) /* wake up caller if membarriers done */ + finish_membarrier_apc( apc ); wake_up( &apc->obj, 0 ); close_handle( current->process, req->prev_apc ); release_object( apc ); @@ -1671,6 +1752,8 @@ DECL_HANDLER(select) else { apc->executed = 1; + if (apc->call.type == APC_MEMORY_BARRIER) + finish_membarrier_apc( apc ); wake_up( &apc->obj, 0 ); } release_object( apc ); @@ -2012,3 +2095,36 @@ DECL_HANDLER(get_next_thread) set_error( STATUS_NO_MORE_ENTRIES ); release_object( process ); } + +DECL_HANDLER(flush_process_write_buffers) +{ + struct process *process = current->process; + struct thread *thread; + apc_call_t call; + + assert( memory_barrier_obj ); + + memset( &call, 0, sizeof(call) ); + call.memory_barrier.type = APC_MEMORY_BARRIER; + + LIST_FOR_EACH_ENTRY( thread, &process->thread_list, struct thread, proc_entry ) + { + struct thread_apc *apc; + int success; + + if (thread == current || thread->state == TERMINATED) continue; + + if (!(apc = create_apc( ¤t->obj, &call ))) break; + + if ((success = queue_apc( NULL, thread, apc ))) + thread->mb_apcs_pending++; + + release_object( apc ); + + if (!success) + { + set_error( STATUS_UNSUCCESSFUL ); + break; + } + } +} diff --git a/server/thread.h b/server/thread.h index 8dcf966a90a..c9040704700 100644 --- a/server/thread.h +++ b/server/thread.h @@ -90,6 +90,7 @@ struct thread struct list kernel_object; /* list of kernel object pointers */ data_size_t desc_len; /* thread description length in bytes */ WCHAR *desc; /* thread description string */ + int mb_apcs_pending; /* number of APCs left for the current memory barrier */ };
extern struct thread *current;
From: Torge Matthies tmatthies@codeweavers.com
--- dlls/ntdll/unix/server.c | 3 +++ server/thread.c | 2 +- 2 files changed, 4 insertions(+), 1 deletion(-)
diff --git a/dlls/ntdll/unix/server.c b/dlls/ntdll/unix/server.c index 51a83f472e1..362c7793cbc 100644 --- a/dlls/ntdll/unix/server.c +++ b/dlls/ntdll/unix/server.c @@ -602,6 +602,9 @@ unsigned int server_select( const select_op_t *select_op, data_size_t size, UINT sigset_t old_set; int signaled;
+ /* ensure writes so far are visible to other threads */ + MemoryBarrier(); + memset( &result, 0, sizeof(result) );
do diff --git a/server/thread.c b/server/thread.c index 1f052092b56..ca5c1c51502 100644 --- a/server/thread.c +++ b/server/thread.c @@ -2112,7 +2112,7 @@ DECL_HANDLER(flush_process_write_buffers) struct thread_apc *apc; int success;
- if (thread == current || thread->state == TERMINATED) continue; + if (thread == current || thread->state == TERMINATED || thread->wait) continue;
if (!(apc = create_apc( ¤t->obj, &call ))) break;
From: Torge Matthies tmatthies@codeweavers.com
Credits to Avi Kivity (scylladb) and Aliaksei Kandratsenka (gperftools) for this trick, see [1].
[1] https://github.com/scylladb/seastar/commit/77a58e4dc020233f66fccb8d9e8f7a8b7... --- dlls/ntdll/unix/virtual.c | 54 ++++++++++++++++++++++++++++++++++++--- 1 file changed, 50 insertions(+), 4 deletions(-)
diff --git a/dlls/ntdll/unix/virtual.c b/dlls/ntdll/unix/virtual.c index 0d2ce328371..0f473dc0476 100644 --- a/dlls/ntdll/unix/virtual.c +++ b/dlls/ntdll/unix/virtual.c @@ -214,6 +214,11 @@ struct range_entry static struct range_entry *free_ranges; static struct range_entry *free_ranges_end;
+#if defined(__linux__) && (defined(__i386__) || defined(__x86_64__)) +static void *dontneed_page; +static pthread_mutex_t dontneed_page_mutex = PTHREAD_MUTEX_INITIALIZER; +#endif +
static inline BOOL is_beyond_limit( const void *addr, size_t size, const void *limit ) { @@ -5015,10 +5020,40 @@ NTSTATUS WINAPI NtFlushInstructionCache( HANDLE handle, const void *addr, SIZE_T }
-/********************************************************************** - * NtFlushProcessWriteBuffers (NTDLL.@) - */ -void WINAPI NtFlushProcessWriteBuffers(void) +#if defined(__linux__) && (defined(__i386__) || defined(__x86_64__)) +static int try_madvise( void ) +{ + int ret = 0; + char *mem; + + pthread_mutex_lock(&dontneed_page_mutex); + /* Credits to Avi Kivity (scylladb) and Aliaksei Kandratsenka (gperftools) for this trick, + * see https://github.com/scylladb/seastar/commit/77a58e4dc020233f66fccb8d9e8f7a8b7... */ + mem = dontneed_page; + if (!mem) + { + mem = anon_mmap_alloc( page_size, PROT_READ | PROT_WRITE ); + if (mem == MAP_FAILED) + goto failed; + if (mlock( mem, page_size )) + { + munmap( mem, page_size ); + goto failed; + } + dontneed_page = mem; + } + *mem = 3; + ret = !madvise( mem, page_size, MADV_DONTNEED ); +failed: + pthread_mutex_unlock(&dontneed_page_mutex); + return ret; +} +#else +static int try_madvise( void ) { return 0; } +#endif + + +static void do_apc_memorybarrier( void ) { static pthread_mutex_t apc_memorybarrier_mutex = PTHREAD_MUTEX_INITIALIZER; NTSTATUS status; @@ -5048,6 +5083,17 @@ void WINAPI NtFlushProcessWriteBuffers(void) }
+/********************************************************************** + * NtFlushProcessWriteBuffers (NTDLL.@) + */ +void WINAPI NtFlushProcessWriteBuffers(void) +{ + if (try_madvise()) + return; + do_apc_memorybarrier(); +} + + /********************************************************************** * NtCreatePagingFile (NTDLL.@) */
From: Torge Matthies tmatthies@codeweavers.com
Uses the MEMBARRIER_CMD_PRIVATE_EXPEDITED membarrier command introduced in Linux 4.14. --- dlls/ntdll/unix/virtual.c | 49 ++++++++++++++++++++++++++++++++++++++- 1 file changed, 48 insertions(+), 1 deletion(-)
diff --git a/dlls/ntdll/unix/virtual.c b/dlls/ntdll/unix/virtual.c index 0f473dc0476..67d78fc3185 100644 --- a/dlls/ntdll/unix/virtual.c +++ b/dlls/ntdll/unix/virtual.c @@ -39,6 +39,9 @@ #ifdef HAVE_SYS_SYSINFO_H # include <sys/sysinfo.h> #endif +#ifdef HAVE_SYS_SYSCALL_H +# include <sys/syscall.h> +#endif #ifdef HAVE_SYS_SYSCTL_H # include <sys/sysctl.h> #endif @@ -214,10 +217,16 @@ struct range_entry static struct range_entry *free_ranges; static struct range_entry *free_ranges_end;
-#if defined(__linux__) && (defined(__i386__) || defined(__x86_64__)) +#ifdef __linux__ +#ifdef __NR_membarrier +static BOOL membarrier_exp_available; +static pthread_once_t membarrier_init_once = PTHREAD_ONCE_INIT; +#endif +#if defined(__i386__) || defined(__x86_64__) static void *dontneed_page; static pthread_mutex_t dontneed_page_mutex = PTHREAD_MUTEX_INITIALIZER; #endif +#endif
static inline BOOL is_beyond_limit( const void *addr, size_t size, const void *limit ) @@ -5020,6 +5029,42 @@ NTSTATUS WINAPI NtFlushInstructionCache( HANDLE handle, const void *addr, SIZE_T }
+#if defined(__linux__) && defined(__NR_membarrier) +#define MEMBARRIER_CMD_QUERY 0x00 +#define MEMBARRIER_CMD_PRIVATE_EXPEDITED 0x08 +#define MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED 0x10 + + +static int membarrier( int cmd, unsigned int flags, int cpu_id ) +{ + return syscall( __NR_membarrier, cmd, flags, cpu_id ); +} + + +static void membarrier_init( void ) +{ + static const int exp_required_cmds = + MEMBARRIER_CMD_PRIVATE_EXPEDITED | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED; + int available_cmds = membarrier( MEMBARRIER_CMD_QUERY, 0, 0 ); + if (available_cmds == -1) + return; + if ((available_cmds & exp_required_cmds) == exp_required_cmds) + membarrier_exp_available = !membarrier( MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0, 0 ); +} + + +static int try_exp_membarrier( void ) +{ + pthread_once(&membarrier_init_once, membarrier_init); + if (!membarrier_exp_available) + return 0; + return !membarrier( MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0 ); +} +#else +static int try_exp_membarrier( void ) { return 0; } +#endif + + #if defined(__linux__) && (defined(__i386__) || defined(__x86_64__)) static int try_madvise( void ) { @@ -5088,6 +5133,8 @@ static void do_apc_memorybarrier( void ) */ void WINAPI NtFlushProcessWriteBuffers(void) { + if (try_exp_membarrier()) + return; if (try_madvise()) return; do_apc_memorybarrier();
From: Torge Matthies tmatthies@codeweavers.com
--- dlls/ntdll/unix/virtual.c | 58 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 58 insertions(+)
diff --git a/dlls/ntdll/unix/virtual.c b/dlls/ntdll/unix/virtual.c index 67d78fc3185..0f181fdb094 100644 --- a/dlls/ntdll/unix/virtual.c +++ b/dlls/ntdll/unix/virtual.c @@ -65,6 +65,9 @@ #if defined(__APPLE__) # include <mach/mach_init.h> # include <mach/mach_vm.h> +# include <mach/task.h> +# include <mach/thread_state.h> +# include <mach/vm_map.h> #endif
#include "ntstatus.h" @@ -217,6 +220,11 @@ struct range_entry static struct range_entry *free_ranges; static struct range_entry *free_ranges_end;
+#ifdef __APPLE__ +static kern_return_t (*pthread_get_register_pointer_values)( thread_t, uintptr_t*, size_t*, uintptr_t* ); +static pthread_once_t tgrpvs_init_once = PTHREAD_ONCE_INIT; +#endif + #ifdef __linux__ #ifdef __NR_membarrier static BOOL membarrier_exp_available; @@ -5029,6 +5037,54 @@ NTSTATUS WINAPI NtFlushInstructionCache( HANDLE handle, const void *addr, SIZE_T }
+#ifdef __APPLE__ + +static void tgrpvs_init( void ) +{ + pthread_get_register_pointer_values = dlsym( RTLD_DEFAULT, "thread_get_register_pointer_values" ); +} + +static int try_mach_tgrpvs( void ) +{ + /* Taken from https://github.com/dotnet/runtime/blob/7be37908e5a1cbb83b1062768c1649827eeac... */ + mach_msg_type_number_t count, i; + thread_act_array_t threads; + kern_return_t kret; + int ret = 0; + + pthread_once(&tgrpvs_init_once, tgrpvs_init); + if (!pthread_get_register_pointer_values) + return 0; + + kret = task_threads( mach_task_self(), &threads, &count ); + if (kret) + return 0; + + for (i = 0; i < count; i++) + { + uintptr_t reg_values[128]; + size_t reg_count = ARRAY_SIZE( reg_values ); + uintptr_t sp; + + kret = pthread_get_register_pointer_values( threads[i], &sp, ®_count, reg_values ); + if (kret) + goto fail; + + mach_port_deallocate( mach_task_self(), threads[i] ); + } + ret = 1; +fail: + for (; i < count; i++) + mach_port_deallocate( mach_task_self(), threads[i] ); + vm_deallocate( mach_task_self(), (vm_address_t)threads, count * sizeof(threads[0]) ); + return ret; +} + +#else +static int try_mach_tgrpvs( void ) { return 0; } +#endif + + #if defined(__linux__) && defined(__NR_membarrier) #define MEMBARRIER_CMD_QUERY 0x00 #define MEMBARRIER_CMD_PRIVATE_EXPEDITED 0x08 @@ -5133,6 +5189,8 @@ static void do_apc_memorybarrier( void ) */ void WINAPI NtFlushProcessWriteBuffers(void) { + if (try_mach_tgrpvs()) + return; if (try_exp_membarrier()) return; if (try_madvise())
On Sun Sep 4 13:46:59 2022 +0000, Jinoh Kang wrote:
I don't think I can/should do that, since then if thread 2 issues a
memory barrier while thread 1 is already waiting on a memory barrier, then thread 1 will also wait for the second memory barrier to complete instead of just its own. We can serialize the barrier calls with mutex here, too. It will avoid excessive APCs in case multiple threads call NtFlushProcessWriteBuffers (e.g. RCU with concurrent writers, GC in multiple arena/isolates). If we don't serialize the barrier or coalesce APCs, the total number of simultaneous APCs will be `NM` where N = number of threads in the process, and M = concurrent calls to NtFlushProcessWriteBuffers. Since not all applications use membarrier in the first place, we can also avoid extra object allocation for threads that will never end up using the global memory barrier. (Yet another approach to solve this problem would be keeping track of generations. It will let us coalesce APCs, but this sounds like an overkill.) In general, we want to minimize the complexity and overhead of the fallback path since its use will not be very common: newest operating systems will just use mprotect/membarrier/mach calls, and the fallback is only used when all else fails.
I protected the APC path with a mutex and made the memory barrier object a global object that is only created once. This means that the `wake_up(...)` calls might do a little unnecessary work if multiple processes issue a memory barrier at the same time but I don't think that matters much and we don't have to create one object per process (or thread).
I have though up a way to coalesce the APCs too but it's more complex and doesn't easily allow reporting back errors to the origin thread. Not sure if I should implement it?
Is this patchset abandoned? I'm currently thinking of a contention-mitigating approach for the fallback path based on versioning, which won't wake up already sleeping threads as well. Perhaps I could continue the work.
Actually, I think using APCs or signals introduce too much complexity for a fallback path. We should just use `mprotect()` as *the* fallback mechanism instead.
https://github.com/dotnet/runtime/blob/7be37908e5a1cbb83b1062768c1649827eeac...
On Wed Oct 19 13:32:29 2022 +0000, Jinoh Kang wrote:
Is this patchset abandoned? I'm currently thinking of a contention-mitigating approach for the fallback path based on versioning, which won't wake up already sleeping threads as well. Perhaps I could continue the work.
Hello, this MR is basically ready except that I have not yet have anyone try out the macOS code path. If you have ideas how to improve the code, feel free.
On Wed Oct 19 13:33:59 2022 +0000, Jinoh Kang wrote:
Actually, I think using APCs or signals introduce too much complexity for a fallback path. We should just use `mprotect()` as *the* fallback mechanism instead. https://github.com/dotnet/runtime/blob/7be37908e5a1cbb83b1062768c1649827eeac...
Is this true for all architectures though? Or at least all that Wine intends to support? If so then yeah we can throw out the APC path.
On Wed Oct 19 13:32:29 2022 +0000, Torge Matthies wrote:
Hello, this MR is basically ready except that I have not yet have anyone try out the macOS code path. If you have ideas how to improve the code, feel free.
If it's ready, maybe it shouldn't be marked as "draft"?
On Wed Oct 19 15:12:28 2022 +0000, Zebediah Figura wrote:
If it's ready, maybe it shouldn't be marked as "draft"?
Well, it's not ready for merging since the macOS code is untested? But if that doesn't mean that it should be draft then I will remove that.
Brendan Shanks (@bshanks) commented about dlls/ntdll/unix/virtual.c:
- if (!pthread_get_register_pointer_values)
return 0;- kret = task_threads( mach_task_self(), &threads, &count );
- if (kret)
return 0;- for (i = 0; i < count; i++)
- {
uintptr_t reg_values[128];size_t reg_count = ARRAY_SIZE( reg_values );uintptr_t sp;kret = pthread_get_register_pointer_values( threads[i], &sp, ®_count, reg_values );if (kret)goto fail;
I tried this on macOS Monterey, both Intel and Apple Silicon (under Rosetta). It works correctly on Intel, but on Apple Silicon under Rosetta `thread_get_register_pointer_values` always fails with 0x10000003 (MACH_SEND_INVALID_DESTINATION) when querying Rosetta's exception handling thread, leading to the whole function failing.
The .NET code ignores errors from `thread_get_register_pointer_values` (other than `KERN_INSUFFICIENT_BUFFER_SIZE`) which allows this case to succeed, I think that would be the easiest fix.
(Rosetta's exception handling thread has the name 'com.apple.rosetta.exceptionserver' and is somewhat-hidden, for example it doesn't show up if you attach lldb to the process)
-
Brendan Shanks (＠bshanks) -
Jinoh Kang (＠iamahuman)
-
Torge Matthies
-
Torge Matthies (＠tmatthies)
-
Zebediah Figura (＠zfigura)