Merge pull request #453 from compnerd/windows-improvements

Windows improvements

Author: ktopley-apple

CMakeLists.txt

@@ -77,6 +77,8 @@ if(NOT ENABLE_SWIFT)
   set(INSTALL_OS_HEADERS_DIR "include/os" CACHE PATH "Path where the headers will be installed")
 endif()
 
+option(DISPATCH_ENABLE_ASSERTS "enable debug assertions" FALSE)
+
 option(ENABLE_DTRACE "enable dtrace support" "")
 
 option(ENABLE_TESTING "build libdispatch tests" ON)

private/private.h

@@ -177,13 +177,13 @@ void _dispatch_prohibit_transition_to_multithreaded(bool prohibit);
 
 #if TARGET_OS_MAC
 #define DISPATCH_COCOA_COMPAT 1
-#elif defined(__linux__) || defined(__FreeBSD__)
+#elif defined(__linux__) || defined(__FreeBSD__) || defined(_WIN32)
 #define DISPATCH_COCOA_COMPAT 1
 #else
 #define DISPATCH_COCOA_COMPAT 0
 #endif
 
-#if DISPATCH_COCOA_COMPAT || defined(_WIN32)
+#if DISPATCH_COCOA_COMPAT
 
 #define DISPATCH_CF_SPI_VERSION 20160712
 
@@ -197,12 +197,10 @@ typedef void *dispatch_runloop_handle_t;
 #error "runloop support not implemented on this platform"
 #endif
 
-#if TARGET_OS_MAC
 API_AVAILABLE(macos(10.6), ios(4.0))
 DISPATCH_EXPORT DISPATCH_CONST DISPATCH_WARN_RESULT DISPATCH_NOTHROW
 dispatch_runloop_handle_t
 _dispatch_get_main_queue_port_4CF(void);
-#endif
 
 API_AVAILABLE(macos(10.12), ios(10.0), tvos(10.0), watchos(3.0))
 DISPATCH_EXPORT DISPATCH_NOTHROW
@@ -221,12 +219,12 @@ dispatch_queue_serial_t
 _dispatch_runloop_root_queue_create_4CF(const char *_Nullable label,
 		unsigned long flags);
 
-#if TARGET_OS_MAC || defined(_WIN32)
 API_AVAILABLE(macos(10.9), ios(7.0))
 DISPATCH_EXPORT DISPATCH_WARN_RESULT DISPATCH_NOTHROW
 dispatch_runloop_handle_t
 _dispatch_runloop_root_queue_get_port_4CF(dispatch_queue_t queue);
 
+#if TARGET_OS_MAC
 API_AVAILABLE(macos(10.13.2), ios(11.2), tvos(11.2), watchos(4.2))
 DISPATCH_EXPORT DISPATCH_WARN_RESULT DISPATCH_NOTHROW
 bool
@@ -257,7 +255,7 @@ API_AVAILABLE(macos(10.6), ios(4.0))
 DISPATCH_EXPORT
 void (*_Nullable _dispatch_end_NSAutoReleasePool)(void *);
 
-#endif /* DISPATCH_COCOA_COMPAT || defined(_WIN32) */
+#endif /* DISPATCH_COCOA_COMPAT */
 
 API_AVAILABLE(macos(10.13), ios(11.0), tvos(11.0), watchos(4.0))
 DISPATCH_EXPORT DISPATCH_NOTHROW

src/allocator_internal.h

@@ -97,7 +97,7 @@
 // Use the largest type your platform is comfortable doing atomic ops with.
 // TODO: rdar://11477843
 typedef unsigned long bitmap_t;
-#if defined(__LP64__)
+#if DISPATCH_SIZEOF_PTR == 8
 #define BYTES_PER_BITMAP 8
 #else
 #define BYTES_PER_BITMAP 4
@@ -147,7 +147,7 @@ typedef unsigned long bitmap_t;
 
 #define PADDING_TO_CONTINUATION_SIZE(x) (ROUND_UP_TO_CONTINUATION_SIZE(x) - (x))
 
-#if defined(__LP64__)
+#if DISPATCH_SIZEOF_PTR == 8
 #define SIZEOF_HEADER 16
 #else
 #define SIZEOF_HEADER 8

src/benchmark.c

@@ -41,7 +41,7 @@ _dispatch_benchmark_init(void *context)
 	register size_t cnt = bdata->count;
 	size_t i = 0;
 	uint64_t start, delta;
-#if defined(__LP64__)
+#if DISPATCH_SIZEOF_PTR == 8 && !defined(_WIN32)
 	__uint128_t lcost;
 #else
 	long double lcost;
@@ -93,7 +93,7 @@ dispatch_benchmark_f(size_t count, register void *ctxt,
 	};
 	static dispatch_once_t pred;
 	uint64_t ns, start, delta;
-#if defined(__LP64__)
+#if DISPATCH_SIZEOF_PTR == 8 && !defined(_WIN32)
 	__uint128_t conversion, big_denom;
 #else
 	long double conversion, big_denom;

src/event/event_windows.c

@@ -21,59 +21,223 @@
 #include "internal.h"
 #if DISPATCH_EVENT_BACKEND_WINDOWS
 
+static HANDLE hPort = NULL;
+enum _dispatch_windows_port {
+	DISPATCH_PORT_POKE = 0,
+	DISPATCH_PORT_TIMER_CLOCK_WALL,
+	DISPATCH_PORT_TIMER_CLOCK_UPTIME,
+	DISPATCH_PORT_TIMER_CLOCK_MONOTONIC,
+};
+
 #pragma mark dispatch_unote_t
 
 bool
-_dispatch_unote_register(dispatch_unote_t du DISPATCH_UNUSED,
-		dispatch_wlh_t wlh DISPATCH_UNUSED,
-		dispatch_priority_t pri DISPATCH_UNUSED)
+_dispatch_unote_register_muxed(dispatch_unote_t du DISPATCH_UNUSED)
 {
 	WIN_PORT_ERROR();
 	return false;
 }
 
 void
-_dispatch_unote_resume(dispatch_unote_t du DISPATCH_UNUSED)
+_dispatch_unote_resume_muxed(dispatch_unote_t du DISPATCH_UNUSED)
 {
 	WIN_PORT_ERROR();
 }
 
 bool
-_dispatch_unote_unregister(dispatch_unote_t du DISPATCH_UNUSED,
-		uint32_t flags DISPATCH_UNUSED)
+_dispatch_unote_unregister_muxed(dispatch_unote_t du DISPATCH_UNUSED)
 {
 	WIN_PORT_ERROR();
 	return false;
 }
 
 #pragma mark timers
 
+typedef struct _dispatch_windows_timeout_s {
+	PTP_TIMER pTimer;
+	enum _dispatch_windows_port ullIdent;
+	bool bArmed;
+} *dispatch_windows_timeout_t;
+
+#define DISPATCH_WINDOWS_TIMEOUT_INITIALIZER(clock)                             \
+	[DISPATCH_CLOCK_##clock] = {                                            \
+		.pTimer = NULL,                                                 \
+		.ullIdent = DISPATCH_PORT_TIMER_CLOCK_##clock,                  \
+		.bArmed = FALSE,                                                \
+	}
+
+static struct _dispatch_windows_timeout_s _dispatch_windows_timeout[] = {
+	DISPATCH_WINDOWS_TIMEOUT_INITIALIZER(WALL),
+	DISPATCH_WINDOWS_TIMEOUT_INITIALIZER(UPTIME),
+	DISPATCH_WINDOWS_TIMEOUT_INITIALIZER(MONOTONIC),
+};
+
+static void
+_dispatch_event_merge_timer(dispatch_clock_t clock)
+{
+	uint32_t tidx = DISPATCH_TIMER_INDEX(clock, 0);
+
+	_dispatch_windows_timeout[clock].bArmed = FALSE;
+
+	_dispatch_timers_heap_dirty(_dispatch_timers_heap, tidx);
+	_dispatch_timers_heap[tidx].dth_needs_program = true;
+	_dispatch_timers_heap[tidx].dth_armed = false;
+}
+
+static void CALLBACK
+_dispatch_timer_callback(PTP_CALLBACK_INSTANCE Instance, PVOID Context,
+	PTP_TIMER Timer)
+{
+	BOOL bSuccess;
+
+	bSuccess = PostQueuedCompletionStatus(hPort, 0, (ULONG_PTR)Context,
+		NULL);
+	if (bSuccess == FALSE) {
+		DISPATCH_INTERNAL_CRASH(GetLastError(),
+			"PostQueuedCompletionStatus");
+	}
+}
+
 void
-_dispatch_event_loop_timer_arm(uint32_t tidx DISPATCH_UNUSED,
-		dispatch_timer_delay_s range DISPATCH_UNUSED,
-		dispatch_clock_now_cache_t nows DISPATCH_UNUSED)
+_dispatch_event_loop_timer_arm(dispatch_timer_heap_t dth DISPATCH_UNUSED,
+		uint32_t tidx, dispatch_timer_delay_s range,
+		dispatch_clock_now_cache_t nows)
 {
-	WIN_PORT_ERROR();
+	dispatch_windows_timeout_t timer;
+	FILETIME ftDueTime;
+	LARGE_INTEGER liTime;
+
+	switch (DISPATCH_TIMER_CLOCK(tidx)) {
+	case DISPATCH_CLOCK_WALL:
+		timer = &_dispatch_windows_timeout[DISPATCH_CLOCK_WALL];
+		liTime.QuadPart = range.delay +
+			_dispatch_time_now_cached(DISPATCH_TIMER_CLOCK(tidx), nows);
+		break;
+
+	case DISPATCH_CLOCK_UPTIME:
+	case DISPATCH_CLOCK_MONOTONIC:
+		timer = &_dispatch_windows_timeout[DISPATCH_TIMER_CLOCK(tidx)];
+		liTime.QuadPart = -((range.delay + 99) / 100);
+		break;
+	}
+
+	if (timer->pTimer == NULL) {
+		timer->pTimer = CreateThreadpoolTimer(_dispatch_timer_callback,
+			(LPVOID)timer->ullIdent, NULL);
+		if (timer->pTimer == NULL) {
+			DISPATCH_INTERNAL_CRASH(GetLastError(),
+				"CreateThreadpoolTimer");
+		}
+	}
+
+	ftDueTime.dwHighDateTime = liTime.HighPart;
+	ftDueTime.dwLowDateTime = liTime.LowPart;
+
+	SetThreadpoolTimer(timer->pTimer, &ftDueTime, /*msPeriod=*/0,
+		/*msWindowLength=*/0);
+	timer->bArmed = TRUE;
 }
 
 void
-_dispatch_event_loop_timer_delete(uint32_t tidx DISPATCH_UNUSED)
+_dispatch_event_loop_timer_delete(dispatch_timer_heap_t dth DISPATCH_UNUSED,
+		uint32_t tidx)
 {
-	WIN_PORT_ERROR();
+	dispatch_windows_timeout_t timer;
+
+	switch (DISPATCH_TIMER_CLOCK(tidx)) {
+	case DISPATCH_CLOCK_WALL:
+		timer = &_dispatch_windows_timeout[DISPATCH_CLOCK_WALL];
+		break;
+
+	case DISPATCH_CLOCK_UPTIME:
+	case DISPATCH_CLOCK_MONOTONIC:
+		timer = &_dispatch_windows_timeout[DISPATCH_TIMER_CLOCK(tidx)];
+		break;
+	}
+
+	SetThreadpoolTimer(timer->pTimer, NULL, /*msPeriod=*/0,
+		/*msWindowLength=*/0);
+	timer->bArmed = FALSE;
 }
 
 #pragma mark dispatch_loop
 
+static void
+_dispatch_windows_port_init(void *context DISPATCH_UNUSED)
+{
+	hPort = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
+	if (hPort == NULL) {
+		DISPATCH_INTERNAL_CRASH(GetLastError(),
+			"CreateIoCompletionPort");
+	}
+
+#if DISPATCH_USE_MGR_THREAD
+	_dispatch_trace_item_push(_dispatch_mgr_q.do_targetq, &_dispatch_mgr_q);
+	dx_push(_dispatch_mgr_q.do_targetq, &_dispatch_mgr_q, 0);
+#endif
+}
+
 void
 _dispatch_event_loop_poke(dispatch_wlh_t wlh DISPATCH_UNUSED,
 		uint64_t dq_state DISPATCH_UNUSED, uint32_t flags DISPATCH_UNUSED)
 {
-	WIN_PORT_ERROR();
+	static dispatch_once_t _dispatch_windows_port_init_pred;
+	BOOL bSuccess;
+
+	dispatch_once_f(&_dispatch_windows_port_init_pred, NULL,
+		_dispatch_windows_port_init);
+	bSuccess = PostQueuedCompletionStatus(hPort, 0, DISPATCH_PORT_POKE,
+		NULL);
+	(void)dispatch_assume(bSuccess);
 }
 
 DISPATCH_NOINLINE
 void
-_dispatch_event_loop_drain(uint32_t flags DISPATCH_UNUSED)
+_dispatch_event_loop_drain(uint32_t flags)
+{
+	DWORD dwNumberOfBytesTransferred;
+	ULONG_PTR ulCompletionKey;
+	LPOVERLAPPED pOV;
+	BOOL bSuccess;
+
+	pOV = (LPOVERLAPPED)&pOV;
+	bSuccess = GetQueuedCompletionStatus(hPort, &dwNumberOfBytesTransferred,
+		&ulCompletionKey, &pOV,
+		(flags & KEVENT_FLAG_IMMEDIATE) ? 0 : INFINITE);
+	while (bSuccess) {
+		switch (ulCompletionKey) {
+		case DISPATCH_PORT_POKE:
+			break;
+
+		case DISPATCH_PORT_TIMER_CLOCK_WALL:
+			_dispatch_event_merge_timer(DISPATCH_CLOCK_WALL);
+			break;
+
+		case DISPATCH_PORT_TIMER_CLOCK_UPTIME:
+			_dispatch_event_merge_timer(DISPATCH_CLOCK_UPTIME);
+			break;
+
+		case DISPATCH_PORT_TIMER_CLOCK_MONOTONIC:
+			_dispatch_event_merge_timer(DISPATCH_CLOCK_MONOTONIC);
+			break;
+
+		default:
+			DISPATCH_INTERNAL_CRASH(ulCompletionKey,
+				"unsupported completion key");
+		}
+
+		bSuccess = GetQueuedCompletionStatus(hPort,
+			&dwNumberOfBytesTransferred, &ulCompletionKey, &pOV, 0);
+	}
+
+	if (bSuccess == FALSE && pOV != NULL) {
+		DISPATCH_INTERNAL_CRASH(GetLastError(),
+			"GetQueuedCompletionStatus");
+	}
+}
+
+void
+_dispatch_event_loop_cancel_waiter(dispatch_sync_context_t dsc DISPATCH_UNUSED)
 {
 	WIN_PORT_ERROR();
 }
@@ -109,9 +273,9 @@ _dispatch_event_loop_assert_not_owned(dispatch_wlh_t wlh)
 #endif
 
 void
-_dispatch_event_loop_leave_immediate(dispatch_wlh_t wlh, uint64_t dq_state)
+_dispatch_event_loop_leave_immediate(uint64_t dq_state)
 {
-	(void)wlh; (void)dq_state;
+	(void)dq_state;
 }
 
 #endif // DISPATCH_EVENT_BACKEND_WINDOWS

src/event/workqueue.c

@@ -97,7 +97,6 @@ _dispatch_workq_worker_register(dispatch_queue_global_t root_q)
 	_dispatch_unfair_lock_unlock(&mon->registered_tid_lock);
 #else
 	(void)root_q;
-	(void)cls;
 #endif // HAVE_DISPATCH_WORKQ_MONITORING
 }
 
@@ -124,7 +123,6 @@ _dispatch_workq_worker_unregister(dispatch_queue_global_t root_q)
 	_dispatch_unfair_lock_unlock(&mon->registered_tid_lock);
 #else
 	(void)root_q;
-	(void)cls;
 #endif // HAVE_DISPATCH_WORKQ_MONITORING
 }
 

src/inline_internal.h

@@ -1294,7 +1294,7 @@ _dispatch_queue_drain_try_lock_wlh(dispatch_queue_t dq, uint64_t *dq_state)
 	if (unlikely(!_dq_state_is_base_wlh(old_state) ||
 			!_dq_state_is_enqueued_on_target(old_state) ||
 			_dq_state_is_enqueued_on_manager(old_state))) {
-#if !__LP64__
+#if DISPATCH_SIZEOF_PTR == 4
 		old_state >>= 32;
 #endif
 		DISPATCH_INTERNAL_CRASH(old_state, "Invalid wlh state");

src/internal.h

@@ -306,9 +306,6 @@ upcast(dispatch_object_t dou)
 #include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
-#if defined(_WIN32)
-#define _CRT_RAND_S
-#endif
 #include <stdlib.h>
 #include <string.h>
 #if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__))