synchronization block: make all value assignments atomic.

2 files changed, 65 insertions, 22 deletions

diff --git a/include/ntapi/nt_sync.h b/include/ntapi/nt_sync.h
index 5b28833..99508dd 100644
--- a/include/ntapi/nt_sync.h
+++ b/include/ntapi/nt_sync.h
@@ -122,7 +122,7 @@ typedef struct _nt_io_completion_basic_information {
 
 
 typedef union __attr_aligned__(NT_SYNC_BLOCK_SIZE) _nt_sync_block {
-	char		cache_line[NT_SYNC_BLOCK_SIZE];
+	intptr_t	cache_line[NT_SYNC_BLOCK_SIZE/sizeof(intptr_t)];
 	struct {
 		int32_t		tid;
 		int32_t		pid;
diff --git a/src/sync/ntapi_tt_sync_block.c b/src/sync/ntapi_tt_sync_block.c
index b37659a..0e2a9a4 100644
--- a/src/sync/ntapi_tt_sync_block.c
+++ b/src/sync/ntapi_tt_sync_block.c
@@ -11,6 +11,33 @@
 #include <ntapi/ntapi.h>
 #include "ntapi_impl.h"
 
+static void __sync_block_memset(
+	__in	nt_sync_block *	sync_block,
+	__in	intptr_t	value)
+{
+	intptr_t * sptr = sync_block->cache_line;
+
+	at_store(&sptr[0x0],value);
+	at_store(&sptr[0x1],value);
+	at_store(&sptr[0x2],value);
+	at_store(&sptr[0x3],value);
+	at_store(&sptr[0x4],value);
+	at_store(&sptr[0x5],value);
+	at_store(&sptr[0x6],value);
+	at_store(&sptr[0x7],value);
+
+	if (sizeof(intptr_t) == 4) {
+		at_store(&sptr[0x8],value);
+		at_store(&sptr[0x9],value);
+		at_store(&sptr[0xa],value);
+		at_store(&sptr[0xb],value);
+		at_store(&sptr[0xc],value);
+		at_store(&sptr[0xd],value);
+		at_store(&sptr[0xe],value);
+		at_store(&sptr[0xf],value);
+	}
+}
+
 void __stdcall __ntapi_tt_sync_block_init(
 	__in	nt_sync_block *	sync_block,
 	__in	uint32_t	flags			__optional,
@@ -19,23 +46,32 @@ void __stdcall __ntapi_tt_sync_block_init(
 	__in	int64_t		default_lock_wait	__optional,
 	__in	void *		hsignal			__optional)
 {
-	__ntapi->tt_aligned_block_memset(
-		sync_block,
-		0,sizeof(*sync_block));
-
-	sync_block->lock_tries = default_lock_tries
-		? default_lock_tries
-		: __NT_SYNC_BLOCK_LOCK_TRIES;
-
-	sync_block->lock_wait.quad = default_lock_wait
-		? default_lock_wait
-		: (-1);
-
-	sync_block->flags   = flags;
-	sync_block->srvtid  = srvtid;
-	sync_block->hsignal = hsignal;
-
-	return;
+	__sync_block_memset(
+		sync_block,0);
+
+	at_store_32(
+		&sync_block->lock_tries,
+		default_lock_tries
+			? default_lock_tries
+			: __NT_SYNC_BLOCK_LOCK_TRIES);
+
+	at_store_64(
+		&sync_block->lock_wait.quad,
+		default_lock_wait
+			? default_lock_wait
+			: (-1));
+
+	at_store_32(
+		(int32_t *)&sync_block->flags,
+		flags);
+
+	at_store_32(
+		(int32_t *)&sync_block->srvtid,
+		srvtid);
+
+	at_store(
+		(intptr_t *)&sync_block->hsignal,
+		(intptr_t)hsignal);
 }
 
 
@@ -148,7 +184,9 @@ int32_t __stdcall __ntapi_tt_sync_block_lock(
 	if (lock) return NT_STATUS_NOT_LOCKED;
 
 	/* shared section support */
-	sync_block->pid = pe_get_current_process_id();
+	at_store_32(
+		&sync_block->pid,
+		pe_get_current_process_id());
 
 	return NT_STATUS_SUCCESS;
 }
@@ -196,7 +234,9 @@ int32_t __stdcall __ntapi_tt_sync_block_server_lock(
 	}
 
 	/* yield request: set */
-	sync_block->flags |= NT_SYNC_BLOCK_YIELD_TO_SERVER;
+	at_locked_or_32(
+		(int32_t *)&sync_block->flags,
+		NT_SYNC_BLOCK_YIELD_TO_SERVER);
 
 	/* try again */
 	status = __ntapi_tt_sync_block_lock(
@@ -206,7 +246,9 @@ int32_t __stdcall __ntapi_tt_sync_block_server_lock(
 		sig_flag);
 
 	/* yield request: unset */
-	sync_block->flags ^= NT_SYNC_BLOCK_YIELD_TO_SERVER;
+	at_locked_xor_32(
+		(int32_t *)&sync_block->flags,
+		NT_SYNC_BLOCK_YIELD_TO_SERVER);
 
 	__ntapi->zw_set_event(
 		sync_block->hserver,
@@ -281,7 +323,8 @@ int32_t __stdcall __ntapi_tt_sync_block_discard(
 	if (sync_block->hserver)
 		__ntapi->zw_close(sync_block->hserver);
 
-	__ntapi->tt_aligned_block_memset(sync_block,-1,sizeof(*sync_block));
+	__sync_block_memset(
+		sync_block,-1);
 
 	return NT_STATUS_SUCCESS;
 }