xwrite: poll on non-blocking FDs

write(2) can hit the same EAGAIN/EWOULDBLOCK errors as read(2),
so busy-looping on a non-blocking FD is a waste of resources.

Currently, I do not know of a way for this happen:

* the NonBlocking directive in systemd does not apply to stdin,
  stdout, or stderr.

* xinetd provides no way to set the non-blocking flag at all

But theoretically, it's possible a careless C10K HTTP server
could use pipe2(..., O_NONBLOCK) to setup a pipe for
git-http-backend with only the intent to use non-blocking reads;
but accidentally leave non-blocking set on the write end passed
as stdout to git-upload-pack.

Followup-to: 1079c4be0b ("xread: poll on non blocking fds")

Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

wrapper.c

@@ -271,8 +271,26 @@ ssize_t xwrite(int fd, const void *buf, size_t len)
 	    len = MAX_IO_SIZE;
 	while (1) {
 		nr = write(fd, buf, len);
-		if ((nr < 0) && (errno == EAGAIN || errno == EINTR))
+		if (nr < 0) {
+			if (errno == EINTR)
 				continue;
+			if (errno == EAGAIN || errno == EWOULDBLOCK) {
+				struct pollfd pfd;
+				pfd.events = POLLOUT;
+				pfd.fd = fd;
+				/*
+				 * it is OK if this poll() failed; we
+				 * want to leave this infinite loop
+				 * only when write() returns with
+				 * success, or an expected failure,
+				 * which would be checked by the next
+				 * call to write(2).
+				 */
+				poll(&pfd, 1, -1);
+				continue;
+			}
+		}
+
 		return nr;
 	}
 }