Merge pull request #866 from compnerd/6.2-pipes

IO: switch Win32 pipes to `PIPE_WAIT` and use a sentinel byte

Author: compnerd

src/event/event_windows.c

@@ -277,6 +277,9 @@ _dispatch_pipe_monitor_thread(void *context)
 		char cBuffer[1];
 		DWORD dwNumberOfBytesTransferred;
 		OVERLAPPED ov = {0};
+		// Block on a 0-byte read; this will only resume when data is
+		// available in the pipe. The pipe must be PIPE_WAIT or this thread
+		// will spin.
 		BOOL bSuccess = ReadFile(hPipe, cBuffer, /* nNumberOfBytesToRead */ 0,
 				&dwNumberOfBytesTransferred, &ov);
 		DWORD dwBytesAvailable;

src/io.c

@@ -1463,20 +1463,20 @@ _dispatch_fd_entry_create_with_fd(dispatch_fd_t fd, uintptr_t hash)
 				int result = ioctlsocket((SOCKET)fd, (long)FIONBIO, &value);
 				(void)dispatch_assume_zero(result);
 			} else {
-				// Try to make writing nonblocking, although pipes not coming
-				// from Foundation.Pipe may not have FILE_WRITE_ATTRIBUTES.
+				// The _dispatch_pipe_monitor_thread expects pipes to be
+				// PIPE_WAIT and exploits this assumption by using a blocking
+				// 0-byte read as a synchronization mechanism.
 				DWORD dwPipeMode = 0;
 				if (GetNamedPipeHandleState((HANDLE)fd, &dwPipeMode, NULL,
-						NULL, NULL, NULL, 0) && !(dwPipeMode & PIPE_NOWAIT)) {
-					dwPipeMode |= PIPE_NOWAIT;
+						NULL, NULL, NULL, 0) && !(dwPipeMode & PIPE_WAIT)) {
+					dwPipeMode |= PIPE_WAIT;
 					if (!SetNamedPipeHandleState((HANDLE)fd, &dwPipeMode,
 							NULL, NULL)) {
-						// We may end up blocking on subsequent writes, but we
-						// don't have a good alternative.
-						// The WriteQuotaAvailable from NtQueryInformationFile
-						// erroneously returns 0 when there is a blocking read
-						// on the other end of the pipe.
-						_dispatch_fd_entry_debug("failed to set PIPE_NOWAIT",
+						// If setting the pipe to PIPE_WAIT fails, the
+						// monitoring thread will spin constantly, saturating
+						// a core, which is undesirable but non-fatal.
+						// The semantics will still be correct in this case.
+						_dispatch_fd_entry_debug("failed to set PIPE_WAIT",
 								fd_entry);
 					}
 				}
@@ -2548,13 +2548,40 @@ _dispatch_operation_perform(dispatch_operation_t op)
 				NTSTATUS status = _dispatch_NtQueryInformationFile(hFile,
 						&iosb, &fpli, sizeof(fpli), FilePipeLocalInformation);
 				if (NT_SUCCESS(status)) {
-					// WriteQuotaAvailable is unreliable in the presence
-					// of a blocking reader, when it can return zero, so only
-					// account for it otherwise
-					if (fpli.WriteQuotaAvailable > 0) {
-						len = MIN(len, fpli.WriteQuotaAvailable);
+					// WriteQuotaAvailable is the free space in the output buffer
+					// that has not already been reserved for reading. In other words,
+					// WriteQuotaAvailable =
+					//    OutboundQuota - WriteQuotaUsed - QueuedReadSize.
+					// It is not documented that QueuedReadSize is part of this
+					// calculation, but this behavior has been observed experimentally.
+					// Unfortunately, this means that it is not possible to distinguish
+					// between a full output buffer and a reader blocked waiting for a
+					// full buffer's worth of data. This is a problem because if the
+					// output buffer is full and no reader is waiting for data, then
+					// attempting to write to the buffer of a PIPE_WAIT, non-
+					// overlapped I/O pipe will block the dispatch queue thread.
+					//
+					// In order to work around this idiosyncrasy, we bound the size of
+					// the write to be OutboundQuota - 1. This affords us a sentinel value
+					// in WriteQuotaAvailable that can be used to detect if a reader is
+					// making progress or not.
+					// WriteQuotaAvailable = 0 => a reader is blocked waiting for data.
+					// WriteQuotaAvailable = 1 => the pipe has been written to, but no
+					//   reader is making progress.
+					// When we detect that WriteQuotaAvailable == 1, we write 0 bytes to
+					// avoid blocking the dispatch queue thread.
+					if (fpli.WriteQuotaAvailable == 0) {
+						// This condition can only occur when we have a reader blocked
+						// waiting for data on the pipe. In this case, write a full
+						// buffer's worth of data (less one byte to preserve this
+						// sentinel value of WriteQuotaAvailable == 0).
+						len = MIN(len, fpli.OutboundQuota - 1);
+					} else {
+						// Subtract 1 from WriteQuotaAvailable to ensure we do not fill
+						// the pipe and preserve the sentinel value of
+						// WriteQuotaAvailable == 1.
+						len = MIN(len, fpli.WriteQuotaAvailable - 1);
 					}
-					len = MIN(len, fpli.OutboundQuota);
 				}
 
 				OVERLAPPED ovlOverlapped = {};

tests/dispatch_io_pipe.c

@@ -404,7 +404,12 @@ test_dispatch_write(int kind, int delay)
 	dispatch_group_t g = dispatch_group_create();
 	dispatch_group_enter(g);
 
-	const size_t bufsize = test_get_pipe_buffer_size(kind);
+	// The libdispatch implementation writes at most bufsize-1 bytes
+	// before requiring a reader to start making progress. Because
+	// these tests operate serially, the reader will not make progress
+	// until the write finishes, and a write of >= bufsize will not
+	// finish until the reader starts draining the pipe.
+	const size_t bufsize = test_get_pipe_buffer_size(kind) - 1;
 
 	char *buf = calloc(bufsize, 1);
 	assert(buf);