Changed the *_abbbrevint() functions to *_varint().
[rsync/rsync.git] / io.c
diff --git a/io.c b/io.c
index 029f74b..2557440 100644 (file)
--- a/io.c
+++ b/io.c
@@ -1,38 +1,31 @@
-/* -*- c-file-style: "linux" -*-
+/*
+ * Socket and pipe I/O utilities used in rsync.
  *
- * Copyright (C) 1996-2001 by Andrew Tridgell
- * Copyright (C) Paul Mackerras 1996
- * Copyright (C) 2001, 2002 by Martin Pool <mbp@samba.org>
+ * Copyright (C) 1996-2001 Andrew Tridgell
+ * Copyright (C) 1996 Paul Mackerras
+ * Copyright (C) 2001, 2002 Martin Pool <mbp@samba.org>
+ * Copyright (C) 2003-2007 Wayne Davison
  *
  * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
  *
  * This program is distributed in the hope that it will be useful,
  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  * GNU General Public License for more details.
  *
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street - Fifth Floor, Boston, MA 02110-1301, USA.
  */
 
-/**
- * @file io.c
- *
- * Socket and pipe I/O utilities used in rsync.
- *
- * rsync provides its own multiplexing system, which is used to send
- * stderr and stdout over a single socket.  We need this because
- * stdout normally carries the binary data stream, and stderr all our
- * error messages.
+/* Rsync provides its own multiplexing system, which is used to send
+ * stderr and stdout over a single socket.
  *
  * For historical reasons this is off during the start of the
  * connection, but it's switched on quite early using
- * io_start_multiplex_out() and io_start_multiplex_in().
- **/
+ * io_start_multiplex_out() and io_start_multiplex_in(). */
 
 #include "rsync.h"
 
 
 extern int bwlimit;
 extern size_t bwlimit_writemax;
-extern int verbose;
 extern int io_timeout;
+extern int allowed_lull;
 extern int am_server;
 extern int am_daemon;
 extern int am_sender;
+extern int am_generator;
+extern int inc_recurse;
+extern int io_error;
 extern int eol_nulls;
+extern int flist_eof;
+extern int read_batch;
+extern int csum_length;
 extern int checksum_seed;
 extern int protocol_version;
-extern char *remote_filesfrom_file;
+extern int remove_source_files;
+extern int preserve_hard_links;
+extern char *filesfrom_host;
 extern struct stats stats;
+extern struct file_list *cur_flist, *first_flist;
 
 const char phase_unknown[] = "unknown";
-int select_timeout = SELECT_TIMEOUT;
+int ignore_timeout = 0;
 int batch_fd = -1;
-int batch_gen_fd = -1;
+int msgdone_cnt = 0;
 
-/**
- * The connection might be dropped at some point; perhaps because the
- * remote instance crashed.  Just giving the offset on the stream is
- * not very helpful.  So instead we try to make io_phase_name point to
- * something useful.
- *
- * For buffered/multiplexed I/O these names will be somewhat
- * approximate; perhaps for ease of support we would rather make the
- * buffer always flush when a single application-level I/O finishes.
- *
- * @todo Perhaps we want some simple stack functionality, but there's
- * no need to overdo it.
- **/
-const char *io_write_phase = phase_unknown;
-const char *io_read_phase = phase_unknown;
-
-/** Ignore EOF errors while reading a module listing if the remote
-    version is 24 or less. */
-int kludge_around_eof = False;
+/* Ignore an EOF error if non-zero. See whine_about_eof(). */
+int kluge_around_eof = 0;
 
 int msg_fd_in = -1;
 int msg_fd_out = -1;
+int sock_f_in = -1;
+int sock_f_out = -1;
+
+static int iobuf_f_in = -1;
+static char *iobuf_in;
+static size_t iobuf_in_siz;
+static size_t iobuf_in_ndx;
+static size_t iobuf_in_remaining;
+
+static int iobuf_f_out = -1;
+static char *iobuf_out;
+static int iobuf_out_cnt;
+
+int flist_forward_from = -1;
 
 static int io_multiplexing_out;
 static int io_multiplexing_in;
-static int sock_f_in = -1;
-static int sock_f_out = -1;
-static time_t last_io;
+static time_t last_io_in;
+static time_t last_io_out;
 static int no_flush;
 
 static int write_batch_monitor_in = -1;
@@ -96,59 +94,98 @@ static char io_filesfrom_buf[2048];
 static char *io_filesfrom_bp;
 static char io_filesfrom_lastchar;
 static int io_filesfrom_buflen;
+static int defer_forwarding_messages = 0;
+static int select_timeout = SELECT_TIMEOUT;
+static int active_filecnt = 0;
+static OFF_T active_bytecnt = 0;
+
+static char int_byte_cnt[64] = {
+       1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* (00 - 3F)/4 */
+       1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* (40 - 7F)/4 */
+       2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, /* (80 - BF)/4 */
+       3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 6, 7, /* (C0 - FF)/4 */
+};
 
-static void read_loop(int fd, char *buf, size_t len);
+static void readfd(int fd, char *buffer, size_t N);
+static void writefd(int fd, const char *buf, size_t len);
+static void writefd_unbuffered(int fd, const char *buf, size_t len);
+static void decrement_active_files(int ndx);
+static void decrement_flist_in_progress(int ndx, int redo);
 
-struct redo_list {
-       struct redo_list *next;
-       int num;
+struct flist_ndx_item {
+       struct flist_ndx_item *next;
+       int ndx;
 };
 
-static struct redo_list *redo_list_head;
-static struct redo_list *redo_list_tail;
+struct flist_ndx_list {
+       struct flist_ndx_item *head, *tail;
+};
 
-struct msg_list {
-       struct msg_list *next;
-       char *buf;
+static struct flist_ndx_list redo_list, hlink_list;
+
+struct msg_list_item {
+       struct msg_list_item *next;
        int len;
+       char buf[1];
+};
+
+struct msg_list {
+       struct msg_list_item *head, *tail;
 };
 
-static struct msg_list *msg_list_head;
-static struct msg_list *msg_list_tail;
+static struct msg_list msg2sndr;
 
-static void redo_list_add(int num)
+static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
 {
-       struct redo_list *rl;
-
-       if (!(rl = new(struct redo_list)))
-               exit_cleanup(RERR_MALLOC);
-       rl->next = NULL;
-       rl->num = num;
-       if (redo_list_tail)
-               redo_list_tail->next = rl;
+       struct flist_ndx_item *item;
+
+       if (!(item = new(struct flist_ndx_item)))
+               out_of_memory("flist_ndx_push");
+       item->next = NULL;
+       item->ndx = ndx;
+       if (lp->tail)
+               lp->tail->next = item;
        else
-               redo_list_head = rl;
-       redo_list_tail = rl;
+               lp->head = item;
+       lp->tail = item;
+}
+
+static int flist_ndx_pop(struct flist_ndx_list *lp)
+{
+       struct flist_ndx_item *next;
+       int ndx;
+
+       if (!lp->head)
+               return -1;
+
+       ndx = lp->head->ndx;
+       next = lp->head->next;
+       free(lp->head);
+       lp->head = next;
+       if (!next)
+               lp->tail = NULL;
+
+       return ndx;
 }
 
 static void check_timeout(void)
 {
        time_t t;
 
-       if (!io_timeout)
+       if (!io_timeout || ignore_timeout)
                return;
 
-       if (!last_io) {
-               last_io = time(NULL);
+       if (!last_io_in) {
+               last_io_in = time(NULL);
                return;
        }
 
        t = time(NULL);
 
-       if (t - last_io >= io_timeout) {
+       if (t - last_io_in >= io_timeout) {
                if (!am_server && !am_daemon) {
-                       rprintf(FERROR, "io timeout after %d seconds - exiting\n",
-                               (int)(t-last_io));
+                       rprintf(FERROR, "io timeout after %d seconds -- exiting\n",
+                               (int)(t-last_io_in));
                }
                exit_cleanup(RERR_TIMEOUT);
        }
@@ -162,17 +199,29 @@ void io_set_sock_fds(int f_in, int f_out)
        sock_f_out = f_out;
 }
 
-/** Setup the fd used to receive MSG_* messages.  Only needed when
- * we're the generator because the sender and receiver both use the
- * multiplexed I/O setup. */
+void set_io_timeout(int secs)
+{
+       io_timeout = secs;
+
+       if (!io_timeout || io_timeout > SELECT_TIMEOUT)
+               select_timeout = SELECT_TIMEOUT;
+       else
+               select_timeout = io_timeout;
+
+       allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2;
+}
+
+/* Setup the fd used to receive MSG_* messages.  Only needed during the
+ * early stages of being a local sender (up through the sending of the
+ * file list) or when we're the generator (to fetch the messages from
+ * the receiver). */
 void set_msg_fd_in(int fd)
 {
        msg_fd_in = fd;
 }
 
-/** Setup the fd used to send our MSG_* messages.  Only needed when
- * we're the receiver because the generator and the sender both use
- * the multiplexed I/O setup. */
+/* Setup the fd used to send our MSG_* messages.  Only needed when
+ * we're the receiver (to send our messages to the generator). */
 void set_msg_fd_out(int fd)
 {
        msg_fd_out = fd;
@@ -180,45 +229,57 @@ void set_msg_fd_out(int fd)
 }
 
 /* Add a message to the pending MSG_* list. */
-static void msg_list_add(int code, char *buf, int len)
-{
-       struct msg_list *ml;
-
-       if (!(ml = new(struct msg_list)))
-               exit_cleanup(RERR_MALLOC);
-       ml->next = NULL;
-       if (!(ml->buf = new_array(char, len+4)))
-               exit_cleanup(RERR_MALLOC);
-       SIVAL(ml->buf, 0, ((code+MPLEX_BASE)<<24) | len);
-       memcpy(ml->buf+4, buf, len);
-       ml->len = len+4;
-       if (msg_list_tail)
-               msg_list_tail->next = ml;
+static void msg_list_add(struct msg_list *lst, int code, const char *buf, int len)
+{
+       struct msg_list_item *m;
+       int sz = len + 4 + sizeof m[0] - 1;
+
+       if (!(m = (struct msg_list_item *)new_array(char, sz)))
+               out_of_memory("msg_list_add");
+       m->next = NULL;
+       m->len = len + 4;
+       SIVAL(m->buf, 0, ((code+MPLEX_BASE)<<24) | len);
+       memcpy(m->buf + 4, buf, len);
+       if (lst->tail)
+               lst->tail->next = m;
        else
-               msg_list_head = ml;
-       msg_list_tail = ml;
+               lst->head = m;
+       lst->tail = m;
 }
 
-void send_msg(enum msgcode code, char *buf, int len)
+static void msg2sndr_flush(void)
 {
-       msg_list_add(code, buf, len);
-       msg_list_push(NORMAL_FLUSH);
+       while (msg2sndr.head && io_multiplexing_out) {
+               struct msg_list_item *m = msg2sndr.head;
+               if (!(msg2sndr.head = m->next))
+                       msg2sndr.tail = NULL;
+               stats.total_written += m->len;
+               defer_forwarding_messages = 1;
+               writefd_unbuffered(sock_f_out, m->buf, m->len);
+               defer_forwarding_messages = 0;
+               free(m);
+       }
 }
 
-/** Read a message from the MSG_* fd and dispatch it.  This is only
- * called by the generator. */
+/* Read a message from the MSG_* fd and handle it.  This is called either
+ * during the early stages of being a local sender (up through the sending
+ * of the file list) or when we're the generator (to fetch the messages
+ * from the receiver). */
 static void read_msg_fd(void)
 {
        char buf[2048];
        size_t n;
+       struct file_list *flist;
        int fd = msg_fd_in;
        int tag, len;
 
        /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
-        * to this routine from read_timeout() and writefd_unbuffered(). */
+        * to this routine from writefd_unbuffered(). */
+       no_flush++;
        msg_fd_in = -1;
+       defer_forwarding_messages++;
 
-       read_loop(fd, buf, 4);
+       readfd(fd, buf, 4);
        tag = IVAL(buf, 0);
 
        len = tag & 0xFFFFFF;
@@ -226,20 +287,77 @@ static void read_msg_fd(void)
 
        switch (tag) {
        case MSG_DONE:
-               if (len != 0) {
-                       rprintf(FERROR, "invalid message %d:%d\n", tag, len);
+               if (len < 0 || len > 1 || !am_generator) {
+                 invalid_msg:
+                       rprintf(FERROR, "invalid message %d:%d [%s%s]\n",
+                               tag, len, who_am_i(),
+                               inc_recurse ? "/inc" : "");
                        exit_cleanup(RERR_STREAMIO);
                }
-               redo_list_add(-1);
+               if (len) {
+                       readfd(fd, buf, len);
+                       stats.total_read = read_longint(fd);
+               }
+               msgdone_cnt++;
                break;
        case MSG_REDO:
-               if (len != 4) {
-                       rprintf(FERROR, "invalid message %d:%d\n", tag, len);
-                       exit_cleanup(RERR_STREAMIO);
+               if (len != 4 || !am_generator)
+                       goto invalid_msg;
+               readfd(fd, buf, 4);
+               if (remove_source_files)
+                       decrement_active_files(IVAL(buf,0));
+               flist_ndx_push(&redo_list, IVAL(buf,0));
+               if (inc_recurse)
+                       decrement_flist_in_progress(IVAL(buf,0), 1);
+               break;
+       case MSG_FLIST:
+               if (len != 4 || !am_generator || !inc_recurse)
+                       goto invalid_msg;
+               readfd(fd, buf, 4);
+               /* Read extra file list from receiver. */
+               assert(iobuf_in != NULL);
+               assert(iobuf_f_in == fd);
+               flist = recv_file_list(fd);
+               flist->parent_ndx = IVAL(buf,0);
+               break;
+       case MSG_FLIST_EOF:
+               if (len != 0 || !am_generator || !inc_recurse)
+                       goto invalid_msg;
+               flist_eof = 1;
+               break;
+       case MSG_DELETED:
+               if (len >= (int)sizeof buf || !am_generator)
+                       goto invalid_msg;
+               readfd(fd, buf, len);
+               send_msg(MSG_DELETED, buf, len);
+               break;
+       case MSG_SUCCESS:
+               if (len != 4 || !am_generator)
+                       goto invalid_msg;
+               readfd(fd, buf, len);
+               if (remove_source_files) {
+                       decrement_active_files(IVAL(buf,0));
+                       send_msg(MSG_SUCCESS, buf, len);
                }
-               read_loop(fd, buf, 4);
-               redo_list_add(IVAL(buf,0));
+               if (preserve_hard_links)
+                       flist_ndx_push(&hlink_list, IVAL(buf,0));
+               if (inc_recurse)
+                       decrement_flist_in_progress(IVAL(buf,0), 0);
+               break;
+       case MSG_NO_SEND:
+               if (len != 4 || !am_generator)
+                       goto invalid_msg;
+               readfd(fd, buf, len);
+               if (inc_recurse)
+                       decrement_flist_in_progress(IVAL(buf,0), 0);
                break;
+       case MSG_SOCKERR:
+       case MSG_CLIENT:
+               if (!am_generator)
+                       goto invalid_msg;
+               if (tag == MSG_SOCKERR)
+                       io_end_multiplex_out();
+               /* FALL THROUGH */
        case MSG_INFO:
        case MSG_ERROR:
        case MSG_LOG:
@@ -247,75 +365,138 @@ static void read_msg_fd(void)
                        n = len;
                        if (n >= sizeof buf)
                                n = sizeof buf - 1;
-                       read_loop(fd, buf, n);
+                       readfd(fd, buf, n);
                        rwrite((enum logcode)tag, buf, n);
                        len -= n;
                }
                break;
        default:
-               rprintf(FERROR, "unknown message %d:%d\n", tag, len);
+               rprintf(FERROR, "unknown message %d:%d [%s]\n",
+                       tag, len, who_am_i());
                exit_cleanup(RERR_STREAMIO);
        }
 
+       no_flush--;
        msg_fd_in = fd;
+       if (!--defer_forwarding_messages)
+               msg2sndr_flush();
 }
 
-/* Try to push messages off the list onto the wire.  If we leave with more
- * to do, return 0.  On error, return -1.  If everything flushed, return 1.
- * This is only active in the receiver. */
-int msg_list_push(int flush_it_all)
+/* This is used by the generator to limit how many file transfers can
+ * be active at once when --remove-source-files is specified.  Without
+ * this, sender-side deletions were mostly happening at the end. */
+void increment_active_files(int ndx, int itemizing, enum logcode code)
 {
-       static int written = 0;
-       struct timeval tv;
-       fd_set fds;
+       /* TODO: tune these limits? */
+       while (active_filecnt >= (active_bytecnt >= 128*1024 ? 10 : 50)) {
+               check_for_finished_files(itemizing, code, 0);
+               if (iobuf_out_cnt)
+                       io_flush(NORMAL_FLUSH);
+               else
+                       read_msg_fd();
+       }
 
-       if (msg_fd_out < 0)
-               return -1;
+       active_filecnt++;
+       active_bytecnt += F_LENGTH(cur_flist->files[ndx - cur_flist->ndx_start]);
+}
 
-       while (msg_list_head) {
-               struct msg_list *ml = msg_list_head;
-               int n = write(msg_fd_out, ml->buf + written, ml->len - written);
-               if (n < 0) {
-                       if (errno == EINTR)
-                               continue;
-                       if (errno != EWOULDBLOCK && errno != EAGAIN)
-                               return -1;
-                       if (!flush_it_all)
-                               return 0;
-                       FD_ZERO(&fds);
-                       FD_SET(msg_fd_out, &fds);
-                       tv.tv_sec = select_timeout;
-                       tv.tv_usec = 0;
-                       if (!select(msg_fd_out+1, NULL, &fds, NULL, &tv))
-                               check_timeout();
-               } else if ((written += n) == ml->len) {
-                       free(ml->buf);
-                       msg_list_head = ml->next;
-                       if (!msg_list_head)
-                               msg_list_tail = NULL;
-                       free(ml);
-                       written = 0;
+static void decrement_active_files(int ndx)
+{
+       struct file_list *flist = flist_for_ndx(ndx);
+       assert(flist != NULL);
+       active_filecnt--;
+       active_bytecnt -= F_LENGTH(flist->files[ndx - flist->ndx_start]);
+}
+
+static void decrement_flist_in_progress(int ndx, int redo)
+{
+       struct file_list *flist = cur_flist ? cur_flist : first_flist;
+
+       while (ndx < flist->ndx_start) {
+               if (flist == first_flist) {
+                 invalid_ndx:
+                       rprintf(FERROR,
+                               "Invalid file index: %d (%d - %d) [%s]\n",
+                               ndx, first_flist->ndx_start,
+                               first_flist->prev->ndx_start + first_flist->prev->count - 1,
+                               who_am_i());
+                       exit_cleanup(RERR_PROTOCOL);
                }
+               flist = flist->prev;
+       }
+       while (ndx >= flist->ndx_start + flist->count) {
+               if (!(flist = flist->next))
+                       goto invalid_ndx;
+       }
+
+       flist->in_progress--;
+       if (redo)
+               flist->to_redo++;
+}
+
+/* Write an message to a multiplexed stream. If this fails, rsync exits. */
+static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len)
+{
+       char buffer[1024];
+       size_t n = len;
+
+       SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
+
+       if (n > sizeof buffer - 4)
+               n = 0;
+       else
+               memcpy(buffer + 4, buf, n);
+
+       writefd_unbuffered(fd, buffer, n+4);
+
+       len -= n;
+       buf += n;
+
+       if (len) {
+               defer_forwarding_messages++;
+               writefd_unbuffered(fd, buf, len);
+               if (!--defer_forwarding_messages)
+                       msg2sndr_flush();
        }
+}
+
+int send_msg(enum msgcode code, const char *buf, int len)
+{
+       if (msg_fd_out < 0) {
+               if (!defer_forwarding_messages)
+                       return io_multiplex_write(code, buf, len);
+               if (!io_multiplexing_out)
+                       return 0;
+               msg_list_add(&msg2sndr, code, buf, len);
+               return 1;
+       }
+       mplex_write(msg_fd_out, code, buf, len);
        return 1;
 }
 
-int get_redo_num(void)
+void send_msg_int(enum msgcode code, int num)
 {
-       struct redo_list *next;
-       int num;
+       char numbuf[4];
+       SIVAL(numbuf, 0, num);
+       send_msg(code, numbuf, 4);
+}
 
-       while (!redo_list_head)
+void wait_for_receiver(void)
+{
+       if (iobuf_out_cnt)
+               io_flush(NORMAL_FLUSH);
+       else
                read_msg_fd();
+}
 
-       num = redo_list_head->num;
-       next = redo_list_head->next;
-       free(redo_list_head);
-       redo_list_head = next;
-       if (!next)
-               redo_list_tail = NULL;
+int get_redo_num(void)
+{
+       return flist_ndx_pop(&redo_list);
+}
 
-       return num;
+int get_hlink_num(void)
+{
+       return flist_ndx_pop(&hlink_list);
 }
 
 /**
@@ -337,20 +518,28 @@ void io_set_filesfrom_fds(int f_in, int f_out)
        io_filesfrom_buflen = 0;
 }
 
-/**
- * It's almost always an error to get an EOF when we're trying to read
- * from the network, because the protocol is self-terminating.
+/* It's almost always an error to get an EOF when we're trying to read from the
+ * network, because the protocol is (for the most part) self-terminating.
  *
- * However, there is one unfortunate cases where it is not, which is
- * rsync <2.4.6 sending a list of modules on a server, since the list
- * is terminated by closing the socket. So, for the section of the
- * program where that is a problem (start_socket_client),
- * kludge_around_eof is True and we just exit.
- */
+ * There is one case for the receiver when it is at the end of the transfer
+ * (hanging around reading any keep-alive packets that might come its way): if
+ * the sender dies before the generator's kill-signal comes through, we can end
+ * up here needing to loop until the kill-signal arrives.  In this situation,
+ * kluge_around_eof will be < 0.
+ *
+ * There is another case for older protocol versions (< 24) where the module
+ * listing was not terminated, so we must ignore an EOF error in that case and
+ * exit.  In this situation, kluge_around_eof will be > 0. */
 static void whine_about_eof(int fd)
 {
-       if (kludge_around_eof && fd == sock_f_in)
-               exit_cleanup(0);
+       if (kluge_around_eof && fd == sock_f_in) {
+               int i;
+               if (kluge_around_eof > 0)
+                       exit_cleanup(0);
+               /* If we're still here after 10 seconds, exit with an error. */
+               for (i = 10*1000/20; i--; )
+                       msleep(20);
+       }
 
        rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
                "(%.0f bytes received so far) [%s]\n",
@@ -359,7 +548,6 @@ static void whine_about_eof(int fd)
        exit_cleanup(RERR_STREAMIO);
 }
 
-
 /**
  * Read from a socket with I/O timeout. return the number of bytes
  * read. If no bytes can be read then exit, never return a number <= 0.
@@ -373,11 +561,11 @@ static void whine_about_eof(int fd)
  */
 static int read_timeout(int fd, char *buf, size_t len)
 {
-       int n, ret = 0;
+       int n, cnt = 0;
 
-       io_flush(NORMAL_FLUSH);
+       io_flush(FULL_FLUSH);
 
-       while (ret == 0) {
+       while (cnt == 0) {
                /* until we manage to read *something* */
                fd_set r_fds, w_fds;
                struct timeval tv;
@@ -387,15 +575,6 @@ static int read_timeout(int fd, char *buf, size_t len)
                FD_ZERO(&r_fds);
                FD_ZERO(&w_fds);
                FD_SET(fd, &r_fds);
-               if (msg_fd_in >= 0) {
-                       FD_SET(msg_fd_in, &r_fds);
-                       if (msg_fd_in > maxfd)
-                               maxfd = msg_fd_in;
-               } else if (msg_list_head) {
-                       FD_SET(msg_fd_out, &w_fds);
-                       if (msg_fd_out > maxfd)
-                               maxfd = msg_fd_out;
-               }
                if (io_filesfrom_f_out >= 0) {
                        int new_fd;
                        if (io_filesfrom_buflen == 0) {
@@ -422,17 +601,14 @@ static int read_timeout(int fd, char *buf, size_t len)
                count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
 
                if (count <= 0) {
-                       if (errno == EBADF)
+                       if (errno == EBADF) {
+                               defer_forwarding_messages = 0;
                                exit_cleanup(RERR_SOCKETIO);
+                       }
                        check_timeout();
                        continue;
                }
 
-               if (msg_fd_in >= 0 && FD_ISSET(msg_fd_in, &r_fds))
-                       read_msg_fd();
-               else if (msg_list_head && FD_ISSET(msg_fd_out, &w_fds))
-                       msg_list_push(NORMAL_FLUSH);
-
                if (io_filesfrom_f_out >= 0) {
                        if (io_filesfrom_buflen) {
                                if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
@@ -509,21 +685,23 @@ static int read_timeout(int fd, char *buf, size_t len)
                                continue;
 
                        /* Don't write errors on a dead socket. */
-                       if (fd == sock_f_in)
-                               close_multiplexing_out();
-                       rsyserr(FERROR, errno, "read error");
+                       if (fd == sock_f_in) {
+                               io_end_multiplex_out();
+                               rsyserr(FSOCKERR, errno, "read error");
+                       } else
+                               rsyserr(FERROR, errno, "read error");
                        exit_cleanup(RERR_STREAMIO);
                }
 
                buf += n;
                len -= n;
-               ret += n;
+               cnt += n;
 
-               if (io_timeout && fd == sock_f_in)
-                       last_io = time(NULL);
+               if (fd == sock_f_in && io_timeout)
+                       last_io_in = time(NULL);
        }
 
-       return ret;
+       return cnt;
 }
 
 /**
@@ -534,7 +712,7 @@ int read_filesfrom_line(int fd, char *fname)
 {
        char ch, *s, *eob = fname + MAXPATHLEN - 1;
        int cnt;
-       int reading_remotely = remote_filesfrom_file != NULL;
+       int reading_remotely = filesfrom_host != NULL;
        int nulls = eol_nulls || reading_remotely;
 
   start:
@@ -544,13 +722,19 @@ int read_filesfrom_line(int fd, char *fname)
                if (cnt < 0 && (errno == EWOULDBLOCK
                  || errno == EINTR || errno == EAGAIN)) {
                        struct timeval tv;
-                       fd_set fds;
-                       FD_ZERO(&fds);
-                       FD_SET(fd, &fds);
+                       fd_set r_fds, e_fds;
+                       FD_ZERO(&r_fds);
+                       FD_SET(fd, &r_fds);
+                       FD_ZERO(&e_fds);
+                       FD_SET(fd, &e_fds);
                        tv.tv_sec = select_timeout;
                        tv.tv_usec = 0;
-                       if (!select(fd+1, &fds, NULL, NULL, &tv))
+                       if (!select(fd+1, &r_fds, NULL, &e_fds, &tv))
                                check_timeout();
+                       if (FD_ISSET(fd, &e_fds)) {
+                               rsyserr(FINFO, errno,
+                                       "select exception on fd %d", fd);
+                       }
                        continue;
                }
                if (cnt != 1)
@@ -573,42 +757,90 @@ int read_filesfrom_line(int fd, char *fname)
        return s - fname;
 }
 
-
-static char *iobuf_out;
-static int iobuf_out_cnt;
-
-void io_start_buffering_out(void)
+int io_start_buffering_out(int f_out)
 {
-       if (iobuf_out)
-               return;
+       if (iobuf_out) {
+               assert(f_out == iobuf_f_out);
+               return 0;
+       }
        if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
                out_of_memory("io_start_buffering_out");
        iobuf_out_cnt = 0;
+       iobuf_f_out = f_out;
+       return 1;
 }
 
-
-static char *iobuf_in;
-static size_t iobuf_in_siz;
-
-void io_start_buffering_in(void)
+int io_start_buffering_in(int f_in)
 {
-       if (iobuf_in)
-               return;
+       if (iobuf_in) {
+               assert(f_in == iobuf_f_in);
+               return 0;
+       }
        iobuf_in_siz = 2 * IO_BUFFER_SIZE;
        if (!(iobuf_in = new_array(char, iobuf_in_siz)))
                out_of_memory("io_start_buffering_in");
+       iobuf_f_in = f_in;
+       return 1;
 }
 
+void io_end_buffering_in(void)
+{
+       if (!iobuf_in)
+               return;
+       free(iobuf_in);
+       iobuf_in = NULL;
+       iobuf_in_ndx = 0;
+       iobuf_in_remaining = 0;
+       iobuf_f_in = -1;
+}
 
-void io_end_buffering(void)
+void io_end_buffering_out(void)
 {
-       io_flush(NORMAL_FLUSH);
-       if (!io_multiplexing_out) {
-               free(iobuf_out);
-               iobuf_out = NULL;
+       if (!iobuf_out)
+               return;
+       io_flush(FULL_FLUSH);
+       free(iobuf_out);
+       iobuf_out = NULL;
+       iobuf_f_out = -1;
+}
+
+void maybe_flush_socket(int important)
+{
+       if (iobuf_out && iobuf_out_cnt
+        && (important || time(NULL) - last_io_out >= 5))
+               io_flush(NORMAL_FLUSH);
+}
+
+void maybe_send_keepalive(void)
+{
+       if (time(NULL) - last_io_out >= allowed_lull) {
+               if (!iobuf_out || !iobuf_out_cnt) {
+                       if (protocol_version < 29)
+                               return; /* there's nothing we can do */
+                       if (protocol_version >= 30)
+                               send_msg(MSG_NOOP, "", 0);
+                       else {
+                               write_int(sock_f_out, cur_flist->count);
+                               write_shortint(sock_f_out, ITEM_IS_NEW);
+                       }
+               }
+               if (iobuf_out)
+                       io_flush(NORMAL_FLUSH);
        }
 }
 
+void start_flist_forward(int f_in)
+{
+       assert(iobuf_out != NULL);
+       assert(iobuf_f_out == msg_fd_out);
+       flist_forward_from = f_in;
+}
+
+void stop_flist_forward()
+{
+       flist_forward_from = -1;
+       io_flush(FULL_FLUSH);
+}
 
 /**
  * Continue trying to read len bytes - don't return until len has been
@@ -624,7 +856,6 @@ static void read_loop(int fd, char *buf, size_t len)
        }
 }
 
-
 /**
  * Read from the file descriptor handling multiplexing - return number
  * of bytes read.
@@ -633,84 +864,119 @@ static void read_loop(int fd, char *buf, size_t len)
  */
 static int readfd_unbuffered(int fd, char *buf, size_t len)
 {
-       static size_t remaining;
-       static size_t iobuf_in_ndx;
-       int tag, ret = 0;
-       char line[1024];
+       size_t msg_bytes;
+       int tag, cnt = 0;
+       char line[BIGPATHBUFLEN];
 
-       if (!iobuf_in || fd != sock_f_in)
+       if (!iobuf_in || fd != iobuf_f_in)
                return read_timeout(fd, buf, len);
 
-       if (!io_multiplexing_in && remaining == 0) {
-               remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
+       if (!io_multiplexing_in && iobuf_in_remaining == 0) {
+               iobuf_in_remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
                iobuf_in_ndx = 0;
        }
 
-       while (ret == 0) {
-               if (remaining) {
-                       len = MIN(len, remaining);
+       while (cnt == 0) {
+               if (iobuf_in_remaining) {
+                       len = MIN(len, iobuf_in_remaining);
                        memcpy(buf, iobuf_in + iobuf_in_ndx, len);
                        iobuf_in_ndx += len;
-                       remaining -= len;
-                       ret = len;
+                       iobuf_in_remaining -= len;
+                       cnt = len;
                        break;
                }
 
                read_loop(fd, line, 4);
                tag = IVAL(line, 0);
 
-               remaining = tag & 0xFFFFFF;
+               msg_bytes = tag & 0xFFFFFF;
                tag = (tag >> 24) - MPLEX_BASE;
 
                switch (tag) {
                case MSG_DATA:
-                       if (remaining > iobuf_in_siz) {
+                       if (msg_bytes > iobuf_in_siz) {
                                if (!(iobuf_in = realloc_array(iobuf_in, char,
-                                                              remaining)))
+                                                              msg_bytes)))
                                        out_of_memory("readfd_unbuffered");
-                               iobuf_in_siz = remaining;
+                               iobuf_in_siz = msg_bytes;
                        }
-                       read_loop(fd, iobuf_in, remaining);
+                       read_loop(fd, iobuf_in, msg_bytes);
+                       iobuf_in_remaining = msg_bytes;
                        iobuf_in_ndx = 0;
                        break;
+               case MSG_NOOP:
+                       if (am_sender)
+                               maybe_send_keepalive();
+                       break;
+               case MSG_IO_ERROR:
+                       if (msg_bytes != 4)
+                               goto invalid_msg;
+                       read_loop(fd, line, msg_bytes);
+                       io_error |= IVAL(line, 0);
+                       break;
+               case MSG_DELETED:
+                       if (msg_bytes >= sizeof line)
+                               goto overflow;
+                       read_loop(fd, line, msg_bytes);
+                       /* A directory name was sent with the trailing null */
+                       if (msg_bytes > 0 && !line[msg_bytes-1])
+                               log_delete(line, S_IFDIR);
+                       else {
+                               line[msg_bytes] = '\0';
+                               log_delete(line, S_IFREG);
+                       }
+                       break;
+               case MSG_SUCCESS:
+                       if (msg_bytes != 4) {
+                         invalid_msg:
+                               rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
+                                       tag, (long)msg_bytes, who_am_i());
+                               exit_cleanup(RERR_STREAMIO);
+                       }
+                       read_loop(fd, line, msg_bytes);
+                       successful_send(IVAL(line, 0));
+                       break;
+               case MSG_NO_SEND:
+                       if (msg_bytes != 4)
+                               goto invalid_msg;
+                       read_loop(fd, line, msg_bytes);
+                       send_msg_int(MSG_NO_SEND, IVAL(line, 0));
+                       break;
                case MSG_INFO:
                case MSG_ERROR:
-                       if (remaining >= sizeof line) {
-                               rprintf(FERROR, "multiplexing overflow %d:%ld\n\n",
-                                       tag, (long)remaining);
+                       if (msg_bytes >= sizeof line) {
+                           overflow:
+                               rprintf(FERROR,
+                                       "multiplexing overflow %d:%ld [%s]\n",
+                                       tag, (long)msg_bytes, who_am_i());
                                exit_cleanup(RERR_STREAMIO);
                        }
-                       read_loop(fd, line, remaining);
-                       rwrite((enum logcode)tag, line, remaining);
-                       remaining = 0;
+                       read_loop(fd, line, msg_bytes);
+                       rwrite((enum logcode)tag, line, msg_bytes);
                        break;
                default:
-                       rprintf(FERROR, "unexpected tag %d\n", tag);
+                       rprintf(FERROR, "unexpected tag %d [%s]\n",
+                               tag, who_am_i());
                        exit_cleanup(RERR_STREAMIO);
                }
        }
 
-       if (remaining == 0)
+       if (iobuf_in_remaining == 0)
                io_flush(NORMAL_FLUSH);
 
-       return ret;
+       return cnt;
 }
 
-
-
-/**
- * Do a buffered read from @p fd.  Don't return until all @p n bytes
- * have been read.  If all @p n can't be read then exit with an
- * error.
- **/
+/* Do a buffered read from fd.  Don't return until all N bytes have
+ * been read.  If all N can't be read then exit with an error. */
 static void readfd(int fd, char *buffer, size_t N)
 {
-       int  ret;
+       int  cnt;
        size_t total = 0;
 
        while (total < N) {
-               ret = readfd_unbuffered(fd, buffer + total, N-total);
-               total += ret;
+               cnt = readfd_unbuffered(fd, buffer + total, N-total);
+               total += cnt;
        }
 
        if (fd == write_batch_monitor_in) {
@@ -718,62 +984,210 @@ static void readfd(int fd, char *buffer, size_t N)
                        exit_cleanup(RERR_FILEIO);
        }
 
+       if (fd == flist_forward_from)
+               writefd(iobuf_f_out, buffer, total);
+
        if (fd == sock_f_in)
                stats.total_read += total;
 }
 
+unsigned short read_shortint(int f)
+{
+       char b[2];
+       readfd(f, b, 2);
+       return (UVAL(b, 1) << 8) + UVAL(b, 0);
+}
+
+int32 read_varint(int f)
+{
+       int32 num;
+       char b[5];
+       int cnt;
+       readfd(f, b, 1);
+       cnt = int_byte_cnt[CVAL(b, 0) / 4];
+       if (cnt > 1)
+               readfd(f, b + 1, cnt - 1);
+       switch (cnt) {
+       case 1:
+               num = NVAL1(b, 0);
+               break;
+       case 2:
+               num = NVAL2(b, 0x80);
+               break;
+       case 3:
+               num = NVAL3(b, 0xC0);
+               break;
+       case 4:
+               num = NVAL4(b, 0xE0);
+               break;
+       case 5:
+               num = NVAL4(b+1, 0);
+               break;
+       default:
+               exit_cleanup(RERR_PROTOCOL); /* impossible... */
+       }
+       return num;
+}
 
 int32 read_int(int f)
 {
        char b[4];
-       int32 ret;
+       int32 num;
 
-       readfd(f,b,4);
-       ret = IVAL(b,0);
-       if (ret == (int32)0xffffffff)
-               return -1;
-       return ret;
+       readfd(f, b, 4);
+       num = IVAL(b, 0);
+#if SIZEOF_INT32 > 4
+       if (num & (int32)0x80000000)
+               num |= ~(int32)0xffffffff;
+#endif
+       return num;
 }
 
 int64 read_longint(int f)
 {
-       int64 ret;
-       char b[8];
-       ret = read_int(f);
+       int64 num;
+       char b[9];
 
-       if ((int32)ret != (int32)0xffffffff)
-               return ret;
+       if (protocol_version < 30) {
+               num = read_int(f);
 
-#ifdef INT64_IS_OFF_T
-       if (sizeof (int64) < 8) {
+               if ((int32)num != (int32)0xffffffff)
+                       return num;
+
+#if SIZEOF_INT64 < 8
                rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
                exit_cleanup(RERR_UNSUPPORTED);
-       }
+#else
+               readfd(f, b, 8);
+               num = IVAL(b,0) | (((int64)IVAL(b,4))<<32);
+#endif
+       } else {
+               int cnt;
+               readfd(f, b, 3);
+               cnt = int_byte_cnt[CVAL(b, 0) / 4] + 2;
+#if SIZEOF_INT64 < 8
+               if (cnt > 5 || (cnt == 5 && (CVAL(b,0)&0x3F || CVAL(b,1)&0x80))) {
+                       rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
+                       exit_cleanup(RERR_UNSUPPORTED);
+               }
 #endif
-       readfd(f,b,8);
-       ret = IVAL(b,0) | (((int64)IVAL(b,4))<<32);
+               if (cnt > 3)
+                       readfd(f, b + 3, cnt - 3);
+               switch (cnt) {
+               case 3:
+                       num = NVAL3(b, 0);
+                       break;
+               case 4:
+                       num = NVAL4(b, 0x80);
+                       break;
+               case 5:
+                       num = NVAL5(b, 0xC0);
+                       break;
+#if SIZEOF_INT64 >= 8
+               case 6:
+                       num = NVAL6(b, 0xE0);
+                       break;
+               case 7:
+                       num = NVAL7(b, 0xF0);
+                       break;
+               case 8:
+                       num = NVAL8(b, 0xF8);
+                       break;
+               case 9:
+                       num = NVAL8(b+1, 0);
+                       break;
+#endif
+               default:
+                       exit_cleanup(RERR_PROTOCOL); /* impossible... */
+               }
+       }
 
-       return ret;
+       return num;
 }
 
-void read_buf(int f,char *buf,size_t len)
+void read_buf(int f, char *buf, size_t len)
 {
        readfd(f,buf,len);
 }
 
-void read_sbuf(int f,char *buf,size_t len)
+void read_sbuf(int f, char *buf, size_t len)
 {
        readfd(f, buf, len);
-       buf[len] = 0;
+       buf[len] = '\0';
 }
 
-unsigned char read_byte(int f)
+uchar read_byte(int f)
 {
-       unsigned char c;
+       uchar c;
        readfd(f, (char *)&c, 1);
        return c;
 }
 
+int read_vstring(int f, char *buf, int bufsize)
+{
+       int len = read_byte(f);
+
+       if (len & 0x80)
+               len = (len & ~0x80) * 0x100 + read_byte(f);
+
+       if (len >= bufsize) {
+               rprintf(FERROR, "over-long vstring received (%d > %d)\n",
+                       len, bufsize - 1);
+               return -1;
+       }
+
+       if (len)
+               readfd(f, buf, len);
+       buf[len] = '\0';
+       return len;
+}
+
+/* Populate a sum_struct with values from the socket.  This is
+ * called by both the sender and the receiver. */
+void read_sum_head(int f, struct sum_struct *sum)
+{
+       sum->count = read_int(f);
+       if (sum->count < 0) {
+               rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
+                       (long)sum->count, who_am_i());
+               exit_cleanup(RERR_PROTOCOL);
+       }
+       sum->blength = read_int(f);
+       if (sum->blength < 0 || sum->blength > MAX_BLOCK_SIZE) {
+               rprintf(FERROR, "Invalid block length %ld [%s]\n",
+                       (long)sum->blength, who_am_i());
+               exit_cleanup(RERR_PROTOCOL);
+       }
+       sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
+       if (sum->s2length < 0 || sum->s2length > MAX_DIGEST_LEN) {
+               rprintf(FERROR, "Invalid checksum length %d [%s]\n",
+                       sum->s2length, who_am_i());
+               exit_cleanup(RERR_PROTOCOL);
+       }
+       sum->remainder = read_int(f);
+       if (sum->remainder < 0 || sum->remainder > sum->blength) {
+               rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
+                       (long)sum->remainder, who_am_i());
+               exit_cleanup(RERR_PROTOCOL);
+       }
+}
+
+/* Send the values from a sum_struct over the socket.  Set sum to
+ * NULL if there are no checksums to send.  This is called by both
+ * the generator and the sender. */
+void write_sum_head(int f, struct sum_struct *sum)
+{
+       static struct sum_struct null_sum;
+
+       if (sum == NULL)
+               sum = &null_sum;
+
+       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.
@@ -794,16 +1208,16 @@ unsigned char read_byte(int f)
 static void sleep_for_bwlimit(int bytes_written)
 {
        static struct timeval prior_tv;
-       static long total_written = 0; 
+       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)
+       if (!bwlimit_writemax)
                return;
 
-       total_written += bytes_written; 
+       total_written += bytes_written;
 
        gettimeofday(&start_tv, NULL);
        if (prior_tv.tv_sec) {
@@ -830,40 +1244,45 @@ static void sleep_for_bwlimit(int bytes_written)
        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 the generator) reading any data on msg_fd_in
- * (to avoid deadlock).
+ * 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)
+static void writefd_unbuffered(int fd, const char *buf, size_t len)
 {
        size_t n, total = 0;
-       fd_set w_fds, r_fds;
-       int maxfd, count, ret;
+       fd_set w_fds, r_fds, e_fds;
+       int maxfd, count, cnt, using_r_fds;
+       int defer_save = defer_forwarding_messages;
        struct timeval tv;
 
-       no_flush++;
+       if (no_flush++)
+               defer_forwarding_messages = 1;
 
        while (total < len) {
                FD_ZERO(&w_fds);
-               FD_SET(fd,&w_fds);
+               FD_SET(fd, &w_fds);
+               FD_ZERO(&e_fds);
+               FD_SET(fd, &e_fds);
                maxfd = fd;
 
                if (msg_fd_in >= 0) {
                        FD_ZERO(&r_fds);
-                       FD_SET(msg_fd_in,&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, msg_fd_in >= 0 ? &r_fds : NULL,
-                              &w_fds, NULL, &tv);
+               count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
+                              &w_fds, &e_fds, &tv);
 
                if (count <= 0) {
                        if (count < 0 && errno == EBADF)
@@ -872,19 +1291,24 @@ static void writefd_unbuffered(int fd,char *buf,size_t len)
                        continue;
                }
 
-               if (msg_fd_in >= 0 && FD_ISSET(msg_fd_in, &r_fds))
+               if (FD_ISSET(fd, &e_fds)) {
+                       rsyserr(FINFO, errno,
+                               "select exception on fd %d", fd);
+               }
+
+               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)
+               if (bwlimit_writemax && n > bwlimit_writemax)
                        n = bwlimit_writemax;
-               ret = write(fd, buf + total, n);
+               cnt = write(fd, buf + total, n);
 
-               if (ret <= 0) {
-                       if (ret < 0) {
+               if (cnt <= 0) {
+                       if (cnt < 0) {
                                if (errno == EINTR)
                                        continue;
                                if (errno == EWOULDBLOCK || errno == EAGAIN) {
@@ -895,80 +1319,56 @@ static void writefd_unbuffered(int fd,char *buf,size_t len)
 
                        /* Don't try to write errors back across the stream. */
                        if (fd == sock_f_out)
-                               close_multiplexing_out();
+                               io_end_multiplex_out();
+                       /* Don't try to write errors down a failing msg pipe. */
+                       if (am_server && fd == msg_fd_out)
+                               exit_cleanup(RERR_STREAMIO);
                        rsyserr(FERROR, errno,
-                               "writefd_unbuffered failed to write %ld bytes: phase \"%s\" [%s]",
-                               (long)len, io_write_phase, who_am_i());
+                               "writefd_unbuffered failed to write %ld bytes [%s]",
+                               (long)len, 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) {
-                               io_timeout = 30;
+                               set_io_timeout(30);
+                               ignore_timeout = 0;
                                readfd_unbuffered(sock_f_in, io_filesfrom_buf,
                                                  sizeof io_filesfrom_buf);
                        }
                        exit_cleanup(RERR_STREAMIO);
                }
 
-               total += ret;
+               total += cnt;
+               defer_forwarding_messages = 1;
 
                if (fd == sock_f_out) {
-                       if (io_timeout)
-                               last_io = time(NULL);
-                       sleep_for_bwlimit(ret);
+                       if (io_timeout || am_generator)
+                               last_io_out = time(NULL);
+                       sleep_for_bwlimit(cnt);
                }
        }
 
        no_flush--;
+       if (!(defer_forwarding_messages = defer_save))
+               msg2sndr_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[4096];
-       size_t n = len;
-
-       SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
-
-       if (n > sizeof buffer - 4)
-               n = sizeof buffer - 4;
-
-       memcpy(&buffer[4], buf, n);
-       writefd_unbuffered(sock_f_out, buffer, n+4);
-
-       len -= n;
-       buf += n;
-
-       if (len)
-               writefd_unbuffered(sock_f_out, buf, len);
-}
-
-
 void io_flush(int flush_it_all)
 {
-       msg_list_push(flush_it_all);
+       if (flush_it_all && !defer_forwarding_messages)
+               msg2sndr_flush();
 
        if (!iobuf_out_cnt || no_flush)
                return;
 
        if (io_multiplexing_out)
-               mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
+               mplex_write(sock_f_out, MSG_DATA, iobuf_out, iobuf_out_cnt);
        else
-               writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
+               writefd_unbuffered(iobuf_f_out, iobuf_out, iobuf_out_cnt);
        iobuf_out_cnt = 0;
 }
 
-
-static void writefd(int fd,char *buf,size_t len)
+static void writefd(int fd, const char *buf, size_t len)
 {
-       if (fd == msg_fd_out) {
-               rprintf(FERROR, "Internal error: wrong write used in receiver.\n");
-               exit_cleanup(RERR_PROTOCOL);
-       }
-
        if (fd == sock_f_out)
                stats.total_written += len;
 
@@ -977,7 +1377,7 @@ static void writefd(int fd,char *buf,size_t len)
                        exit_cleanup(RERR_FILEIO);
        }
 
-       if (!iobuf_out || fd != sock_f_out) {
+       if (!iobuf_out || fd != iobuf_f_out) {
                writefd_unbuffered(fd, buf, len);
                return;
        }
@@ -996,22 +1396,51 @@ static void writefd(int fd,char *buf,size_t len)
        }
 }
 
-
-void write_int(int f,int32 x)
+void write_shortint(int f, unsigned short x)
 {
-       char b[4];
-       SIVAL(b,0,x);
-       writefd(f,b,4);
+       char b[2];
+       b[0] = (char)x;
+       b[1] = (char)(x >> 8);
+       writefd(f, b, 2);
 }
 
-
-void write_int_named(int f, int32 x, const char *phase)
+void write_varint(int f, int32 x)
 {
-       io_write_phase = phase;
-       write_int(f, x);
-       io_write_phase = phase_unknown;
+       char b[5];
+       if ((uint32)x < ((uint32)1<<(1*8-1))) {
+               b[0] = (char)x;
+               writefd(f, b, 1);
+       } else if ((uint32)x < ((uint32)1<<(2*8-2))) {
+               b[0] = (char)(x >> 8) | 0x80;
+               b[1] = (char)x;
+               writefd(f, b, 2);
+       } else if ((uint32)x < ((uint32)1<<(3*8-3))) {
+               b[0] = (char)(x >> 16) | 0xC0;
+               b[1] = (char)(x >> 8);
+               b[2] = (char)x;
+               writefd(f, b, 3);
+       } else if ((uint32)x < ((uint32)1<<(4*8-4))) {
+               b[0] = (char)(x >> 24) | 0xE0;
+               b[1] = (char)(x >> 16);
+               b[2] = (char)(x >> 8);
+               b[3] = (char)x;
+               writefd(f, b, 4);
+       } else {
+               b[0] = 0xF0;
+               b[1] = (char)(x >> 24);
+               b[2] = (char)(x >> 16);
+               b[3] = (char)(x >> 8);
+               b[4] = (char)x;
+               writefd(f, b, 5);
+       }
 }
 
+void write_int(int f, int32 x)
+{
+       char b[4];
+       SIVAL(b, 0, x);
+       writefd(f, b, 4);
+}
 
 /*
  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
@@ -1019,44 +1448,218 @@ void write_int_named(int f, int32 x, const char *phase)
  */
 void write_longint(int f, int64 x)
 {
-       char b[8];
-
-       if (x <= 0x7FFFFFFF) {
-               write_int(f, (int)x);
-               return;
-       }
+       char b[12];
 
-#ifdef INT64_IS_OFF_T
-       if (sizeof (int64) < 8) {
+#if SIZEOF_INT64 < 8
+       if (x < 0 || x > 0x7FFFFFFF) {
                rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
                exit_cleanup(RERR_UNSUPPORTED);
        }
 #endif
 
-       write_int(f, (int32)0xFFFFFFFF);
-       SIVAL(b,0,(x&0xFFFFFFFF));
-       SIVAL(b,4,((x>>32)&0xFFFFFFFF));
+       if (protocol_version < 30) {
+               char * const s = b+4;
+               SIVAL(s, 0, x);
+#if SIZEOF_INT64 < 8
+               writefd(f, s, 4);
+#else
+               if (x <= 0x7FFFFFFF && x >= 0) {
+                       writefd(f, s, 4);
+                       return;
+               }
 
-       writefd(f,b,8);
+               memset(b, 0xFF, 4);
+               SIVAL(s, 4, x >> 32);
+               writefd(f, b, 12);
+       } else if (x < 0) {
+               goto all_bits;
+#endif
+       } else if (x < ((int32)1<<(3*8-1))) {
+               b[0] = (char)(x >> 16);
+               b[1] = (char)(x >> 8);
+               b[2] = (char)x;
+               writefd(f, b, 3);
+       } else if (x < ((int64)1<<(4*8-2))) {
+               b[0] = (char)((x >> 24) | 0x80);
+               b[1] = (char)(x >> 16);
+               b[2] = (char)(x >> 8);
+               b[3] = (char)x;
+               writefd(f, b, 4);
+#if SIZEOF_INT64 < 8
+       } else {
+               b[0] = 0xC0;
+               b[1] = (char)(x >> 24);
+               b[2] = (char)(x >> 16);
+               b[3] = (char)(x >> 8);
+               b[4] = (char)x;
+               writefd(f, b, 5);
+       }
+#else
+       } else if (x < ((int64)1<<(5*8-3))) {
+               b[0] = (char)((x >> 32) | 0xC0);
+               b[1] = (char)(x >> 24);
+               b[2] = (char)(x >> 16);
+               b[3] = (char)(x >> 8);
+               b[4] = (char)x;
+               writefd(f, b, 5);
+       } else if (x < ((int64)1<<(6*8-4))) {
+               b[0] = (char)((x >> 40) | 0xE0);
+               b[1] = (char)(x >> 32);
+               b[2] = (char)(x >> 24);
+               b[3] = (char)(x >> 16);
+               b[4] = (char)(x >> 8);
+               b[5] = (char)x;
+               writefd(f, b, 6);
+       } else if (x < ((int64)1<<(7*8-5))) {
+               b[0] = (char)((x >> 48) | 0xF0);
+               b[1] = (char)(x >> 40);
+               b[2] = (char)(x >> 32);
+               b[3] = (char)(x >> 24);
+               b[4] = (char)(x >> 16);
+               b[5] = (char)(x >> 8);
+               b[6] = (char)x;
+               writefd(f, b, 7);
+       } else if (x < ((int64)1<<(8*8-6))) {
+               b[0] = (char)((x >> 56) | 0xF8);
+               b[1] = (char)(x >> 48);
+               b[2] = (char)(x >> 40);
+               b[3] = (char)(x >> 32);
+               b[4] = (char)(x >> 24);
+               b[5] = (char)(x >> 16);
+               b[6] = (char)(x >> 8);
+               b[7] = (char)x;
+               writefd(f, b, 8);
+       } else {
+         all_bits:
+               b[0] = (char)0xFC;
+               b[1] = (char)(x >> 56);
+               b[2] = (char)(x >> 48);
+               b[3] = (char)(x >> 40);
+               b[4] = (char)(x >> 32);
+               b[5] = (char)(x >> 24);
+               b[6] = (char)(x >> 16);
+               b[7] = (char)(x >> 8);
+               b[8] = (char)x;
+               writefd(f, b, 9);
+       }
+#endif
 }
 
-void write_buf(int f,char *buf,size_t len)
+void write_buf(int f, const char *buf, size_t len)
 {
        writefd(f,buf,len);
 }
 
 /** Write a string to the connection */
-void write_sbuf(int f, char *buf)
+void write_sbuf(int f, const char *buf)
 {
        writefd(f, buf, strlen(buf));
 }
 
-void write_byte(int f,unsigned char c)
+void write_byte(int f, uchar c)
 {
        writefd(f, (char *)&c, 1);
 }
 
+void write_vstring(int f, const char *str, int len)
+{
+       uchar lenbuf[3], *lb = lenbuf;
+
+       if (len > 0x7F) {
+               if (len > 0x7FFF) {
+                       rprintf(FERROR,
+                               "attempting to send over-long vstring (%d > %d)\n",
+                               len, 0x7FFF);
+                       exit_cleanup(RERR_PROTOCOL);
+               }
+               *lb++ = len / 0x100 + 0x80;
+       }
+       *lb = len;
 
+       writefd(f, (char*)lenbuf, lb - lenbuf + 1);
+       if (len)
+               writefd(f, str, len);
+}
+
+/* Send a file-list index using a byte-reduction method. */
+void write_ndx(int f, int32 ndx)
+{
+       static int32 prev_positive = -1, prev_negative = 1;
+       int32 diff, cnt = 0;
+       char b[6];
+
+       if (protocol_version < 30 || read_batch) {
+               write_int(f, ndx);
+               return;
+       }
+
+       /* Send NDX_DONE as a single-byte 0 with no side effects.  Send
+        * negative nums as a positive after sending a leading 0xFF. */
+       if (ndx >= 0) {
+               diff = ndx - prev_positive;
+               prev_positive = ndx;
+       } else if (ndx == NDX_DONE) {
+               *b = 0;
+               writefd(f, b, 1);
+               return;
+       } else {
+               b[cnt++] = (char)0xFF;
+               ndx = -ndx;
+               diff = ndx - prev_negative;
+               prev_negative = ndx;
+       }
+
+       /* A diff of 1 - 253 is sent as a one-byte diff; a diff of 254 - 32767
+        * or 0 is sent as a 0xFE + a two-byte diff; otherwise we send 0xFE
+        * & all 4 bytes of the (non-negative) num with the high-bit set. */
+       if (diff < 0xFE && diff > 0)
+               b[cnt++] = (char)diff;
+       else if (diff < 0 || diff > 0x7FFF) {
+               b[cnt++] = (char)0xFE;
+               b[cnt++] = (char)((ndx >> 24) | 0x80);
+               b[cnt++] = (char)(ndx >> 16);
+               b[cnt++] = (char)(ndx >> 8);
+               b[cnt++] = (char)ndx;
+       } else {
+               b[cnt++] = (char)0xFE;
+               b[cnt++] = (char)(diff >> 8);
+               b[cnt++] = (char)diff;
+       }
+       writefd(f, b, cnt);
+}
+
+/* Receive a file-list index using a byte-reduction method. */
+int32 read_ndx(int f)
+{
+       static int32 prev_positive = -1, prev_negative = 1;
+       int32 *prev_ptr, num;
+       char b[4];
+
+       if (protocol_version < 30)
+               return read_int(f);
+
+       readfd(f, b, 1);
+       if (CVAL(b, 0) == 0xFF) {
+               readfd(f, b, 1);
+               prev_ptr = &prev_negative;
+       } else if (CVAL(b, 0) == 0)
+               return NDX_DONE;
+       else
+               prev_ptr = &prev_positive;
+       if (CVAL(b, 0) == 0xFE) {
+               readfd(f, b, 2);
+               if (CVAL(b, 0) & 0x80) {
+                       readfd(f, b+2, 2);
+                       num = NVAL4(b, 0x80);
+               } else
+                       num = NVAL2(b, 0) + *prev_ptr;
+       } else
+               num = CVAL(b, 0) + *prev_ptr;
+       *prev_ptr = num;
+       if (prev_ptr == &prev_negative)
+               num = -num;
+       return num;
+}
 
 /**
  * Read a line of up to @p maxlen characters into @p buf (not counting
@@ -1083,11 +1686,10 @@ int read_line(int f, char *buf, size_t maxlen)
        return maxlen > 0;
 }
 
-
 void io_printf(int fd, const char *format, ...)
 {
        va_list ap;
-       char buf[1024];
+       char buf[BIGPATHBUFLEN];
        int len;
 
        va_start(ap, format);
@@ -1097,15 +1699,19 @@ void io_printf(int fd, const char *format, ...)
        if (len < 0)
                exit_cleanup(RERR_STREAMIO);
 
+       if (len > (int)sizeof buf) {
+               rprintf(FERROR, "io_printf() was too long for the buffer.\n");
+               exit_cleanup(RERR_STREAMIO);
+       }
+
        write_sbuf(fd, buf);
 }
 
-
 /** Setup for multiplexing a MSG_* stream with the data stream. */
 void io_start_multiplex_out(void)
 {
        io_flush(NORMAL_FLUSH);
-       io_start_buffering_out();
+       io_start_buffering_out(sock_f_out);
        io_multiplexing_out = 1;
 }
 
@@ -1113,37 +1719,36 @@ void io_start_multiplex_out(void)
 void io_start_multiplex_in(void)
 {
        io_flush(NORMAL_FLUSH);
-       io_start_buffering_in();
+       io_start_buffering_in(sock_f_in);
        io_multiplexing_in = 1;
 }
 
 /** Write an message to the multiplexed data stream. */
-int io_multiplex_write(enum msgcode code, char *buf, size_t len)
+int io_multiplex_write(enum msgcode code, const char *buf, size_t len)
 {
        if (!io_multiplexing_out)
                return 0;
-
        io_flush(NORMAL_FLUSH);
        stats.total_written += (len+4);
-       mplex_write(code, buf, len);
+       mplex_write(sock_f_out, code, buf, len);
        return 1;
 }
 
-void close_multiplexing_in(void)
+void io_end_multiplex_in(void)
 {
        io_multiplexing_in = 0;
+       io_end_buffering_in();
 }
 
 /** Stop output multiplexing. */
-void close_multiplexing_out(void)
+void io_end_multiplex_out(void)
 {
        io_multiplexing_out = 0;
+       io_end_buffering_out();
 }
 
 void start_write_batch(int fd)
 {
-       write_stream_flags(batch_fd);
-
        /* Some communication has already taken place, but we don't
         * enable batch writing until here so that we can write a
         * canonical record of the communication even though the