+ write_int(f, sum->count);
+ write_int(f, sum->blength);
+ if (protocol_version >= 27)
+ write_int(f, sum->s2length);
+ write_int(f, sum->remainder);
+}
+
+/**
+ * Sleep after writing to limit I/O bandwidth usage.
+ *
+ * @todo Rather than sleeping after each write, it might be better to
+ * use some kind of averaging. The current algorithm seems to always
+ * use a bit less bandwidth than specified, because it doesn't make up
+ * for slow periods. But arguably this is a feature. In addition, we
+ * ought to take the time used to write the data into account.
+ *
+ * During some phases of big transfers (file FOO is uptodate) this is
+ * called with a small bytes_written every time. As the kernel has to
+ * round small waits up to guarantee that we actually wait at least the
+ * requested number of microseconds, this can become grossly inaccurate.
+ * We therefore keep track of the bytes we've written over time and only
+ * sleep when the accumulated delay is at least 1 tenth of a second.
+ **/
+static void sleep_for_bwlimit(int bytes_written)
+{
+ static struct timeval prior_tv;
+ static long total_written = 0;
+ struct timeval tv, start_tv;
+ long elapsed_usec, sleep_usec;
+
+#define ONE_SEC 1000000L /* # of microseconds in a second */
+
+ if (!bwlimit)
+ return;
+
+ total_written += bytes_written;
+
+ gettimeofday(&start_tv, NULL);
+ if (prior_tv.tv_sec) {
+ elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
+ + (start_tv.tv_usec - prior_tv.tv_usec);
+ total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
+ if (total_written < 0)
+ total_written = 0;
+ }
+
+ sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
+ if (sleep_usec < ONE_SEC / 10) {
+ prior_tv = start_tv;
+ return;
+ }
+
+ tv.tv_sec = sleep_usec / ONE_SEC;
+ tv.tv_usec = sleep_usec % ONE_SEC;
+ select(0, NULL, NULL, NULL, &tv);
+
+ gettimeofday(&prior_tv, NULL);
+ elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
+ + (prior_tv.tv_usec - start_tv.tv_usec);
+ total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
+}
+
+/* Write len bytes to the file descriptor fd, looping as necessary to get
+ * the job done and also (in certain circumstances) reading any data on
+ * msg_fd_in to avoid deadlock.
+ *
+ * This function underlies the multiplexing system. The body of the
+ * application never calls this function directly. */
+static void writefd_unbuffered(int fd,char *buf,size_t len)
+{
+ size_t n, total = 0;
+ fd_set w_fds, r_fds;
+ int maxfd, count, cnt, using_r_fds;
+ struct timeval tv;
+
+ no_flush++;
+
+ while (total < len) {
+ FD_ZERO(&w_fds);
+ FD_SET(fd,&w_fds);
+ maxfd = fd;
+
+ if (msg_fd_in >= 0 && len-total >= contiguous_write_len) {
+ FD_ZERO(&r_fds);
+ FD_SET(msg_fd_in,&r_fds);
+ if (msg_fd_in > maxfd)
+ maxfd = msg_fd_in;
+ using_r_fds = 1;
+ } else
+ using_r_fds = 0;
+
+ tv.tv_sec = select_timeout;
+ tv.tv_usec = 0;
+
+ errno = 0;
+ count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
+ &w_fds, NULL, &tv);
+
+ if (count <= 0) {
+ if (count < 0 && errno == EBADF)
+ exit_cleanup(RERR_SOCKETIO);
+ check_timeout();
+ continue;
+ }
+
+ if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
+ read_msg_fd();
+
+ if (!FD_ISSET(fd, &w_fds))
+ continue;
+
+ n = len - total;
+ if (bwlimit && n > bwlimit_writemax)
+ n = bwlimit_writemax;
+ cnt = write(fd, buf + total, n);
+
+ if (cnt <= 0) {
+ if (cnt < 0) {
+ if (errno == EINTR)
+ continue;
+ if (errno == EWOULDBLOCK || errno == EAGAIN) {
+ msleep(1);
+ continue;
+ }
+ }
+
+ /* Don't try to write errors back across the stream. */
+ if (fd == sock_f_out)
+ close_multiplexing_out();
+ rsyserr(FERROR, errno,
+ "writefd_unbuffered failed to write %ld bytes: phase \"%s\" [%s]",
+ (long)len, io_write_phase, who_am_i());
+ /* If the other side is sending us error messages, try
+ * to grab any messages they sent before they died. */
+ while (fd == sock_f_out && io_multiplexing_in) {
+ set_io_timeout(30);
+ ignore_timeout = 0;
+ readfd_unbuffered(sock_f_in, io_filesfrom_buf,
+ sizeof io_filesfrom_buf);
+ }
+ exit_cleanup(RERR_STREAMIO);
+ }
+
+ total += cnt;
+
+ if (fd == sock_f_out) {
+ if (io_timeout || am_generator)
+ last_io_out = time(NULL);
+ sleep_for_bwlimit(cnt);
+ }
+ }
+
+ no_flush--;
+}
+
+/**
+ * Write an message to a multiplexed stream. If this fails then rsync
+ * exits.
+ **/
+static void mplex_write(enum msgcode code, char *buf, size_t len)
+{
+ char buffer[1024];
+ size_t n = len;
+
+ SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
+
+ /* When the generator reads messages from the msg_fd_in pipe, it can
+ * cause output to occur down the socket. Setting contiguous_write_len
+ * prevents the reading of msg_fd_in once we actually start to write
+ * this sequence of data (though we might read it before the start). */
+ if (am_generator && msg_fd_in >= 0)
+ contiguous_write_len = len + 4;