- Updated the address for the FSF in the opening comment.
[rsync/rsync.git] / io.c
1 /*
2  * Socket and pipe I/O utilities used in rsync.
3  *
4  * Copyright (C) 1996-2001 Andrew Tridgell
5  * Copyright (C) 1996 Paul Mackerras
6  * Copyright (C) 2001, 2002 Martin Pool <mbp@samba.org>
7  * Copyright (C) 2003, 2004, 2005, 2006 Wayne Davison
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; if not, write to the Free Software
21  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  */
23
24 /* Rsync provides its own multiplexing system, which is used to send
25  * stderr and stdout over a single socket.
26  *
27  * For historical reasons this is off during the start of the
28  * connection, but it's switched on quite early using
29  * io_start_multiplex_out() and io_start_multiplex_in(). */
30
31 #include "rsync.h"
32
33 /** If no timeout is specified then use a 60 second select timeout */
34 #define SELECT_TIMEOUT 60
35
36 extern int bwlimit;
37 extern size_t bwlimit_writemax;
38 extern int io_timeout;
39 extern int allowed_lull;
40 extern int am_server;
41 extern int am_daemon;
42 extern int am_sender;
43 extern int am_generator;
44 extern int eol_nulls;
45 extern int read_batch;
46 extern int csum_length;
47 extern int checksum_seed;
48 extern int protocol_version;
49 extern int remove_sent_files;
50 extern int preserve_hard_links;
51 extern char *filesfrom_host;
52 extern struct stats stats;
53 extern struct file_list *the_file_list;
54
55 const char phase_unknown[] = "unknown";
56 int ignore_timeout = 0;
57 int batch_fd = -1;
58 int batch_gen_fd = -1;
59
60 /* Ignore an EOF error if non-zero. See whine_about_eof(). */
61 int kluge_around_eof = 0;
62
63 int msg_fd_in = -1;
64 int msg_fd_out = -1;
65 int sock_f_in = -1;
66 int sock_f_out = -1;
67
68 static int io_multiplexing_out;
69 static int io_multiplexing_in;
70 static time_t last_io_in;
71 static time_t last_io_out;
72 static int no_flush;
73
74 static int write_batch_monitor_in = -1;
75 static int write_batch_monitor_out = -1;
76
77 static int io_filesfrom_f_in = -1;
78 static int io_filesfrom_f_out = -1;
79 static char io_filesfrom_buf[2048];
80 static char *io_filesfrom_bp;
81 static char io_filesfrom_lastchar;
82 static int io_filesfrom_buflen;
83 static int defer_forwarding_messages = 0;
84 static int select_timeout = SELECT_TIMEOUT;
85 static int active_filecnt = 0;
86 static OFF_T active_bytecnt = 0;
87
88 static void read_loop(int fd, char *buf, size_t len);
89
90 struct flist_ndx_item {
91         struct flist_ndx_item *next;
92         int ndx;
93 };
94
95 struct flist_ndx_list {
96         struct flist_ndx_item *head, *tail;
97 };
98
99 static struct flist_ndx_list redo_list, hlink_list;
100
101 struct msg_list_item {
102         struct msg_list_item *next;
103         int len;
104         char buf[1];
105 };
106
107 struct msg_list {
108         struct msg_list_item *head, *tail;
109 };
110
111 static struct msg_list msg2genr, msg2sndr;
112
113 static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
114 {
115         struct flist_ndx_item *item;
116
117         if (!(item = new(struct flist_ndx_item)))
118                 out_of_memory("flist_ndx_push");
119         item->next = NULL;
120         item->ndx = ndx;
121         if (lp->tail)
122                 lp->tail->next = item;
123         else
124                 lp->head = item;
125         lp->tail = item;
126 }
127
128 static int flist_ndx_pop(struct flist_ndx_list *lp)
129 {
130         struct flist_ndx_item *next;
131         int ndx;
132
133         if (!lp->head)
134                 return -1;
135
136         ndx = lp->head->ndx;
137         next = lp->head->next;
138         free(lp->head);
139         lp->head = next;
140         if (!next)
141                 lp->tail = NULL;
142
143         return ndx;
144 }
145
146 static void check_timeout(void)
147 {
148         time_t t;
149
150         if (!io_timeout || ignore_timeout)
151                 return;
152
153         if (!last_io_in) {
154                 last_io_in = time(NULL);
155                 return;
156         }
157
158         t = time(NULL);
159
160         if (t - last_io_in >= io_timeout) {
161                 if (!am_server && !am_daemon) {
162                         rprintf(FERROR, "io timeout after %d seconds -- exiting\n",
163                                 (int)(t-last_io_in));
164                 }
165                 exit_cleanup(RERR_TIMEOUT);
166         }
167 }
168
169 /* Note the fds used for the main socket (which might really be a pipe
170  * for a local transfer, but we can ignore that). */
171 void io_set_sock_fds(int f_in, int f_out)
172 {
173         sock_f_in = f_in;
174         sock_f_out = f_out;
175 }
176
177 void set_io_timeout(int secs)
178 {
179         io_timeout = secs;
180
181         if (!io_timeout || io_timeout > SELECT_TIMEOUT)
182                 select_timeout = SELECT_TIMEOUT;
183         else
184                 select_timeout = io_timeout;
185
186         allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2;
187 }
188
189 /* Setup the fd used to receive MSG_* messages.  Only needed during the
190  * early stages of being a local sender (up through the sending of the
191  * file list) or when we're the generator (to fetch the messages from
192  * the receiver). */
193 void set_msg_fd_in(int fd)
194 {
195         msg_fd_in = fd;
196 }
197
198 /* Setup the fd used to send our MSG_* messages.  Only needed when
199  * we're the receiver (to send our messages to the generator). */
200 void set_msg_fd_out(int fd)
201 {
202         msg_fd_out = fd;
203         set_nonblocking(msg_fd_out);
204 }
205
206 /* Add a message to the pending MSG_* list. */
207 static void msg_list_add(struct msg_list *lst, int code, char *buf, int len)
208 {
209         struct msg_list_item *m;
210         int sz = len + 4 + sizeof m[0] - 1;
211
212         if (!(m = (struct msg_list_item *)new_array(char, sz)))
213                 out_of_memory("msg_list_add");
214         m->next = NULL;
215         m->len = len + 4;
216         SIVAL(m->buf, 0, ((code+MPLEX_BASE)<<24) | len);
217         memcpy(m->buf + 4, buf, len);
218         if (lst->tail)
219                 lst->tail->next = m;
220         else
221                 lst->head = m;
222         lst->tail = m;
223 }
224
225 /* Read a message from the MSG_* fd and handle it.  This is called either
226  * during the early stages of being a local sender (up through the sending
227  * of the file list) or when we're the generator (to fetch the messages
228  * from the receiver). */
229 static void read_msg_fd(void)
230 {
231         char buf[2048];
232         size_t n;
233         int fd = msg_fd_in;
234         int tag, len;
235
236         /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
237          * to this routine from writefd_unbuffered(). */
238         msg_fd_in = -1;
239
240         read_loop(fd, buf, 4);
241         tag = IVAL(buf, 0);
242
243         len = tag & 0xFFFFFF;
244         tag = (tag >> 24) - MPLEX_BASE;
245
246         switch (tag) {
247         case MSG_DONE:
248                 if (len != 0 || !am_generator) {
249                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
250                         exit_cleanup(RERR_STREAMIO);
251                 }
252                 flist_ndx_push(&redo_list, -1);
253                 break;
254         case MSG_REDO:
255                 if (len != 4 || !am_generator) {
256                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
257                         exit_cleanup(RERR_STREAMIO);
258                 }
259                 read_loop(fd, buf, 4);
260                 if (remove_sent_files)
261                         decrement_active_files(IVAL(buf,0));
262                 flist_ndx_push(&redo_list, IVAL(buf,0));
263                 break;
264         case MSG_DELETED:
265                 if (len >= (int)sizeof buf || !am_generator) {
266                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
267                         exit_cleanup(RERR_STREAMIO);
268                 }
269                 read_loop(fd, buf, len);
270                 if (defer_forwarding_messages)
271                         msg_list_add(&msg2sndr, MSG_DELETED, buf, len);
272                 else
273                         io_multiplex_write(MSG_DELETED, buf, len);
274                 break;
275         case MSG_SUCCESS:
276                 if (len != 4 || !am_generator) {
277                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
278                         exit_cleanup(RERR_STREAMIO);
279                 }
280                 read_loop(fd, buf, len);
281                 if (remove_sent_files) {
282                         decrement_active_files(IVAL(buf,0));
283                         if (defer_forwarding_messages)
284                                 msg_list_add(&msg2sndr, MSG_SUCCESS, buf, len);
285                         else
286                                 io_multiplex_write(MSG_SUCCESS, buf, len);
287                 }
288                 if (preserve_hard_links)
289                         flist_ndx_push(&hlink_list, IVAL(buf,0));
290                 break;
291         case MSG_SOCKERR:
292                 if (!am_generator) {
293                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
294                         exit_cleanup(RERR_STREAMIO);
295                 }
296                 close_multiplexing_out();
297                 /* FALL THROUGH */
298         case MSG_INFO:
299         case MSG_ERROR:
300         case MSG_LOG:
301                 while (len) {
302                         n = len;
303                         if (n >= sizeof buf)
304                                 n = sizeof buf - 1;
305                         read_loop(fd, buf, n);
306                         if (am_generator && am_server && defer_forwarding_messages)
307                                 msg_list_add(&msg2sndr, tag, buf, n);
308                         else
309                                 rwrite((enum logcode)tag, buf, n);
310                         len -= n;
311                 }
312                 break;
313         default:
314                 rprintf(FERROR, "unknown message %d:%d [%s]\n",
315                         tag, len, who_am_i());
316                 exit_cleanup(RERR_STREAMIO);
317         }
318
319         msg_fd_in = fd;
320 }
321
322 /* This is used by the generator to limit how many file transfers can
323  * be active at once when --remove-sent-files is specified.  Without
324  * this, sender-side deletions were mostly happening at the end. */
325 void increment_active_files(int ndx, int itemizing, enum logcode code)
326 {
327         /* TODO: tune these limits? */
328         while (active_filecnt >= (active_bytecnt >= 128*1024 ? 10 : 50)) {
329                 if (hlink_list.head)
330                         check_for_finished_hlinks(itemizing, code);
331                 read_msg_fd();
332         }
333
334         active_filecnt++;
335         active_bytecnt += the_file_list->files[ndx]->length;
336 }
337
338 void decrement_active_files(int ndx)
339 {
340         active_filecnt--;
341         active_bytecnt -= the_file_list->files[ndx]->length;
342 }
343
344 /* Try to push messages off the list onto the wire.  If we leave with more
345  * to do, return 0.  On error, return -1.  If everything flushed, return 1.
346  * This is only active in the receiver. */
347 static int msg2genr_flush(int flush_it_all)
348 {
349         static int written = 0;
350         struct timeval tv;
351         fd_set fds;
352
353         if (msg_fd_out < 0)
354                 return -1;
355
356         while (msg2genr.head) {
357                 struct msg_list_item *m = msg2genr.head;
358                 int n = write(msg_fd_out, m->buf + written, m->len - written);
359                 if (n < 0) {
360                         if (errno == EINTR)
361                                 continue;
362                         if (errno != EWOULDBLOCK && errno != EAGAIN)
363                                 return -1;
364                         if (!flush_it_all)
365                                 return 0;
366                         FD_ZERO(&fds);
367                         FD_SET(msg_fd_out, &fds);
368                         tv.tv_sec = select_timeout;
369                         tv.tv_usec = 0;
370                         if (!select(msg_fd_out+1, NULL, &fds, NULL, &tv))
371                                 check_timeout();
372                 } else if ((written += n) == m->len) {
373                         msg2genr.head = m->next;
374                         if (!msg2genr.head)
375                                 msg2genr.tail = NULL;
376                         free(m);
377                         written = 0;
378                 }
379         }
380         return 1;
381 }
382
383 void send_msg(enum msgcode code, char *buf, int len)
384 {
385         if (msg_fd_out < 0) {
386                 io_multiplex_write(code, buf, len);
387                 return;
388         }
389         msg_list_add(&msg2genr, code, buf, len);
390         msg2genr_flush(NORMAL_FLUSH);
391 }
392
393 int get_redo_num(int itemizing, enum logcode code)
394 {
395         while (1) {
396                 if (hlink_list.head)
397                         check_for_finished_hlinks(itemizing, code);
398                 if (redo_list.head)
399                         break;
400                 read_msg_fd();
401         }
402
403         return flist_ndx_pop(&redo_list);
404 }
405
406 int get_hlink_num(void)
407 {
408         return flist_ndx_pop(&hlink_list);
409 }
410
411 /**
412  * When we're the receiver and we have a local --files-from list of names
413  * that needs to be sent over the socket to the sender, we have to do two
414  * things at the same time: send the sender a list of what files we're
415  * processing and read the incoming file+info list from the sender.  We do
416  * this by augmenting the read_timeout() function to copy this data.  It
417  * uses the io_filesfrom_buf to read a block of data from f_in (when it is
418  * ready, since it might be a pipe) and then blast it out f_out (when it
419  * is ready to receive more data).
420  */
421 void io_set_filesfrom_fds(int f_in, int f_out)
422 {
423         io_filesfrom_f_in = f_in;
424         io_filesfrom_f_out = f_out;
425         io_filesfrom_bp = io_filesfrom_buf;
426         io_filesfrom_lastchar = '\0';
427         io_filesfrom_buflen = 0;
428 }
429
430 /* It's almost always an error to get an EOF when we're trying to read from the
431  * network, because the protocol is (for the most part) self-terminating.
432  *
433  * There is one case for the receiver when it is at the end of the transfer
434  * (hanging around reading any keep-alive packets that might come its way): if
435  * the sender dies before the generator's kill-signal comes through, we can end
436  * up here needing to loop until the kill-signal arrives.  In this situation,
437  * kluge_around_eof will be < 0.
438  *
439  * There is another case for older protocol versions (< 24) where the module
440  * listing was not terminated, so we must ignore an EOF error in that case and
441  * exit.  In this situation, kluge_around_eof will be > 0. */
442 static void whine_about_eof(int fd)
443 {
444         if (kluge_around_eof && fd == sock_f_in) {
445                 int i;
446                 if (kluge_around_eof > 0)
447                         exit_cleanup(0);
448                 /* If we're still here after 10 seconds, exit with an error. */
449                 for (i = 10*1000/20; i--; )
450                         msleep(20);
451         }
452
453         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
454                 "(%.0f bytes received so far) [%s]\n",
455                 (double)stats.total_read, who_am_i());
456
457         exit_cleanup(RERR_STREAMIO);
458 }
459
460 /**
461  * Read from a socket with I/O timeout. return the number of bytes
462  * read. If no bytes can be read then exit, never return a number <= 0.
463  *
464  * TODO: If the remote shell connection fails, then current versions
465  * actually report an "unexpected EOF" error here.  Since it's a
466  * fairly common mistake to try to use rsh when ssh is required, we
467  * should trap that: if we fail to read any data at all, we should
468  * give a better explanation.  We can tell whether the connection has
469  * started by looking e.g. at whether the remote version is known yet.
470  */
471 static int read_timeout(int fd, char *buf, size_t len)
472 {
473         int n, cnt = 0;
474
475         io_flush(NORMAL_FLUSH);
476
477         while (cnt == 0) {
478                 /* until we manage to read *something* */
479                 fd_set r_fds, w_fds;
480                 struct timeval tv;
481                 int maxfd = fd;
482                 int count;
483
484                 FD_ZERO(&r_fds);
485                 FD_ZERO(&w_fds);
486                 FD_SET(fd, &r_fds);
487                 if (msg2genr.head) {
488                         FD_SET(msg_fd_out, &w_fds);
489                         if (msg_fd_out > maxfd)
490                                 maxfd = msg_fd_out;
491                 }
492                 if (io_filesfrom_f_out >= 0) {
493                         int new_fd;
494                         if (io_filesfrom_buflen == 0) {
495                                 if (io_filesfrom_f_in >= 0) {
496                                         FD_SET(io_filesfrom_f_in, &r_fds);
497                                         new_fd = io_filesfrom_f_in;
498                                 } else {
499                                         io_filesfrom_f_out = -1;
500                                         new_fd = -1;
501                                 }
502                         } else {
503                                 FD_SET(io_filesfrom_f_out, &w_fds);
504                                 new_fd = io_filesfrom_f_out;
505                         }
506                         if (new_fd > maxfd)
507                                 maxfd = new_fd;
508                 }
509
510                 tv.tv_sec = select_timeout;
511                 tv.tv_usec = 0;
512
513                 errno = 0;
514
515                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
516
517                 if (count <= 0) {
518                         if (errno == EBADF)
519                                 exit_cleanup(RERR_SOCKETIO);
520                         check_timeout();
521                         continue;
522                 }
523
524                 if (msg2genr.head && FD_ISSET(msg_fd_out, &w_fds))
525                         msg2genr_flush(NORMAL_FLUSH);
526
527                 if (io_filesfrom_f_out >= 0) {
528                         if (io_filesfrom_buflen) {
529                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
530                                         int l = write(io_filesfrom_f_out,
531                                                       io_filesfrom_bp,
532                                                       io_filesfrom_buflen);
533                                         if (l > 0) {
534                                                 if (!(io_filesfrom_buflen -= l))
535                                                         io_filesfrom_bp = io_filesfrom_buf;
536                                                 else
537                                                         io_filesfrom_bp += l;
538                                         } else {
539                                                 /* XXX should we complain? */
540                                                 io_filesfrom_f_out = -1;
541                                         }
542                                 }
543                         } else if (io_filesfrom_f_in >= 0) {
544                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
545                                         int l = read(io_filesfrom_f_in,
546                                                      io_filesfrom_buf,
547                                                      sizeof io_filesfrom_buf);
548                                         if (l <= 0) {
549                                                 /* Send end-of-file marker */
550                                                 io_filesfrom_buf[0] = '\0';
551                                                 io_filesfrom_buf[1] = '\0';
552                                                 io_filesfrom_buflen = io_filesfrom_lastchar? 2 : 1;
553                                                 io_filesfrom_f_in = -1;
554                                         } else {
555                                                 if (!eol_nulls) {
556                                                         char *s = io_filesfrom_buf + l;
557                                                         /* Transform CR and/or LF into '\0' */
558                                                         while (s-- > io_filesfrom_buf) {
559                                                                 if (*s == '\n' || *s == '\r')
560                                                                         *s = '\0';
561                                                         }
562                                                 }
563                                                 if (!io_filesfrom_lastchar) {
564                                                         /* Last buf ended with a '\0', so don't
565                                                          * let this buf start with one. */
566                                                         while (l && !*io_filesfrom_bp)
567                                                                 io_filesfrom_bp++, l--;
568                                                 }
569                                                 if (!l)
570                                                         io_filesfrom_bp = io_filesfrom_buf;
571                                                 else {
572                                                         char *f = io_filesfrom_bp;
573                                                         char *t = f;
574                                                         char *eob = f + l;
575                                                         /* Eliminate any multi-'\0' runs. */
576                                                         while (f != eob) {
577                                                                 if (!(*t++ = *f++)) {
578                                                                         while (f != eob && !*f)
579                                                                                 f++, l--;
580                                                                 }
581                                                         }
582                                                         io_filesfrom_lastchar = f[-1];
583                                                 }
584                                                 io_filesfrom_buflen = l;
585                                         }
586                                 }
587                         }
588                 }
589
590                 if (!FD_ISSET(fd, &r_fds))
591                         continue;
592
593                 n = read(fd, buf, len);
594
595                 if (n <= 0) {
596                         if (n == 0)
597                                 whine_about_eof(fd); /* Doesn't return. */
598                         if (errno == EINTR || errno == EWOULDBLOCK
599                             || errno == EAGAIN)
600                                 continue;
601
602                         /* Don't write errors on a dead socket. */
603                         if (fd == sock_f_in) {
604                                 close_multiplexing_out();
605                                 rsyserr(FSOCKERR, errno, "read error");
606                         } else
607                                 rsyserr(FERROR, errno, "read error");
608                         exit_cleanup(RERR_STREAMIO);
609                 }
610
611                 buf += n;
612                 len -= n;
613                 cnt += n;
614
615                 if (fd == sock_f_in && io_timeout)
616                         last_io_in = time(NULL);
617         }
618
619         return cnt;
620 }
621
622 /**
623  * Read a line into the "fname" buffer (which must be at least MAXPATHLEN
624  * characters long).
625  */
626 int read_filesfrom_line(int fd, char *fname)
627 {
628         char ch, *s, *eob = fname + MAXPATHLEN - 1;
629         int cnt;
630         int reading_remotely = filesfrom_host != NULL;
631         int nulls = eol_nulls || reading_remotely;
632
633   start:
634         s = fname;
635         while (1) {
636                 cnt = read(fd, &ch, 1);
637                 if (cnt < 0 && (errno == EWOULDBLOCK
638                   || errno == EINTR || errno == EAGAIN)) {
639                         struct timeval tv;
640                         fd_set fds;
641                         FD_ZERO(&fds);
642                         FD_SET(fd, &fds);
643                         tv.tv_sec = select_timeout;
644                         tv.tv_usec = 0;
645                         if (!select(fd+1, &fds, NULL, NULL, &tv))
646                                 check_timeout();
647                         continue;
648                 }
649                 if (cnt != 1)
650                         break;
651                 if (nulls? !ch : (ch == '\r' || ch == '\n')) {
652                         /* Skip empty lines if reading locally. */
653                         if (!reading_remotely && s == fname)
654                                 continue;
655                         break;
656                 }
657                 if (s < eob)
658                         *s++ = ch;
659         }
660         *s = '\0';
661
662         /* Dump comments. */
663         if (*fname == '#' || *fname == ';')
664                 goto start;
665
666         return s - fname;
667 }
668
669 static char *iobuf_out;
670 static int iobuf_out_cnt;
671
672 void io_start_buffering_out(void)
673 {
674         if (iobuf_out)
675                 return;
676         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
677                 out_of_memory("io_start_buffering_out");
678         iobuf_out_cnt = 0;
679 }
680
681 static char *iobuf_in;
682 static size_t iobuf_in_siz;
683
684 void io_start_buffering_in(void)
685 {
686         if (iobuf_in)
687                 return;
688         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
689         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
690                 out_of_memory("io_start_buffering_in");
691 }
692
693 void io_end_buffering(void)
694 {
695         io_flush(NORMAL_FLUSH);
696         if (!io_multiplexing_out) {
697                 free(iobuf_out);
698                 iobuf_out = NULL;
699         }
700 }
701
702 void maybe_flush_socket(void)
703 {
704         if (iobuf_out && iobuf_out_cnt && time(NULL) - last_io_out >= 5)
705                 io_flush(NORMAL_FLUSH);
706 }
707
708 void maybe_send_keepalive(void)
709 {
710         if (time(NULL) - last_io_out >= allowed_lull) {
711                 if (!iobuf_out || !iobuf_out_cnt) {
712                         if (protocol_version < 29)
713                                 return; /* there's nothing we can do */
714                         write_int(sock_f_out, the_file_list->count);
715                         write_shortint(sock_f_out, ITEM_IS_NEW);
716                 }
717                 if (iobuf_out)
718                         io_flush(NORMAL_FLUSH);
719         }
720 }
721
722 /**
723  * Continue trying to read len bytes - don't return until len has been
724  * read.
725  **/
726 static void read_loop(int fd, char *buf, size_t len)
727 {
728         while (len) {
729                 int n = read_timeout(fd, buf, len);
730
731                 buf += n;
732                 len -= n;
733         }
734 }
735
736 /**
737  * Read from the file descriptor handling multiplexing - return number
738  * of bytes read.
739  *
740  * Never returns <= 0.
741  */
742 static int readfd_unbuffered(int fd, char *buf, size_t len)
743 {
744         static size_t remaining;
745         static size_t iobuf_in_ndx;
746         size_t msg_bytes;
747         int tag, cnt = 0;
748         char line[BIGPATHBUFLEN];
749
750         if (!iobuf_in || fd != sock_f_in)
751                 return read_timeout(fd, buf, len);
752
753         if (!io_multiplexing_in && remaining == 0) {
754                 remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
755                 iobuf_in_ndx = 0;
756         }
757
758         while (cnt == 0) {
759                 if (remaining) {
760                         len = MIN(len, remaining);
761                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
762                         iobuf_in_ndx += len;
763                         remaining -= len;
764                         cnt = len;
765                         break;
766                 }
767
768                 read_loop(fd, line, 4);
769                 tag = IVAL(line, 0);
770
771                 msg_bytes = tag & 0xFFFFFF;
772                 tag = (tag >> 24) - MPLEX_BASE;
773
774                 switch (tag) {
775                 case MSG_DATA:
776                         if (msg_bytes > iobuf_in_siz) {
777                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
778                                                                msg_bytes)))
779                                         out_of_memory("readfd_unbuffered");
780                                 iobuf_in_siz = msg_bytes;
781                         }
782                         read_loop(fd, iobuf_in, msg_bytes);
783                         remaining = msg_bytes;
784                         iobuf_in_ndx = 0;
785                         break;
786                 case MSG_DELETED:
787                         if (msg_bytes >= sizeof line)
788                                 goto overflow;
789                         read_loop(fd, line, msg_bytes);
790                         /* A directory name was sent with the trailing null */
791                         if (msg_bytes > 0 && !line[msg_bytes-1])
792                                 log_delete(line, S_IFDIR);
793                         else {
794                                 line[msg_bytes] = '\0';
795                                 log_delete(line, S_IFREG);
796                         }
797                         break;
798                 case MSG_SUCCESS:
799                         if (msg_bytes != 4) {
800                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
801                                         tag, (long)msg_bytes, who_am_i());
802                                 exit_cleanup(RERR_STREAMIO);
803                         }
804                         read_loop(fd, line, msg_bytes);
805                         successful_send(IVAL(line, 0));
806                         break;
807                 case MSG_INFO:
808                 case MSG_ERROR:
809                         if (msg_bytes >= sizeof line) {
810                             overflow:
811                                 rprintf(FERROR,
812                                         "multiplexing overflow %d:%ld [%s]\n",
813                                         tag, (long)msg_bytes, who_am_i());
814                                 exit_cleanup(RERR_STREAMIO);
815                         }
816                         read_loop(fd, line, msg_bytes);
817                         rwrite((enum logcode)tag, line, msg_bytes);
818                         break;
819                 default:
820                         rprintf(FERROR, "unexpected tag %d [%s]\n",
821                                 tag, who_am_i());
822                         exit_cleanup(RERR_STREAMIO);
823                 }
824         }
825
826         if (remaining == 0)
827                 io_flush(NORMAL_FLUSH);
828
829         return cnt;
830 }
831
832 /**
833  * Do a buffered read from @p fd.  Don't return until all @p n bytes
834  * have been read.  If all @p n can't be read then exit with an
835  * error.
836  **/
837 static void readfd(int fd, char *buffer, size_t N)
838 {
839         int  cnt;
840         size_t total = 0;
841
842         while (total < N) {
843                 cnt = readfd_unbuffered(fd, buffer + total, N-total);
844                 total += cnt;
845         }
846
847         if (fd == write_batch_monitor_in) {
848                 if ((size_t)write(batch_fd, buffer, total) != total)
849                         exit_cleanup(RERR_FILEIO);
850         }
851
852         if (fd == sock_f_in)
853                 stats.total_read += total;
854 }
855
856 int read_shortint(int f)
857 {
858         uchar b[2];
859         readfd(f, (char *)b, 2);
860         return (b[1] << 8) + b[0];
861 }
862
863 int32 read_int(int f)
864 {
865         char b[4];
866         int32 num;
867
868         readfd(f,b,4);
869         num = IVAL(b,0);
870         if (num == (int32)0xffffffff)
871                 return -1;
872         return num;
873 }
874
875 int64 read_longint(int f)
876 {
877         int64 num;
878         char b[8];
879         num = read_int(f);
880
881         if ((int32)num != (int32)0xffffffff)
882                 return num;
883
884 #if SIZEOF_INT64 < 8
885         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
886         exit_cleanup(RERR_UNSUPPORTED);
887 #else
888         readfd(f,b,8);
889         num = IVAL(b,0) | (((int64)IVAL(b,4))<<32);
890 #endif
891
892         return num;
893 }
894
895 void read_buf(int f,char *buf,size_t len)
896 {
897         readfd(f,buf,len);
898 }
899
900 void read_sbuf(int f,char *buf,size_t len)
901 {
902         readfd(f, buf, len);
903         buf[len] = '\0';
904 }
905
906 uchar read_byte(int f)
907 {
908         uchar c;
909         readfd(f, (char *)&c, 1);
910         return c;
911 }
912
913 int read_vstring(int f, char *buf, int bufsize)
914 {
915         int len = read_byte(f);
916
917         if (len & 0x80)
918                 len = (len & ~0x80) * 0x100 + read_byte(f);
919
920         if (len >= bufsize) {
921                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
922                         len, bufsize - 1);
923                 return -1;
924         }
925
926         if (len)
927                 readfd(f, buf, len);
928         buf[len] = '\0';
929         return len;
930 }
931
932 /* Populate a sum_struct with values from the socket.  This is
933  * called by both the sender and the receiver. */
934 void read_sum_head(int f, struct sum_struct *sum)
935 {
936         sum->count = read_int(f);
937         if (sum->count < 0) {
938                 rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
939                         (long)sum->count, who_am_i());
940                 exit_cleanup(RERR_PROTOCOL);
941         }
942         sum->blength = read_int(f);
943         if (sum->blength < 0 || sum->blength > MAX_BLOCK_SIZE) {
944                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
945                         (long)sum->blength, who_am_i());
946                 exit_cleanup(RERR_PROTOCOL);
947         }
948         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
949         if (sum->s2length < 0 || sum->s2length > MD4_SUM_LENGTH) {
950                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
951                         sum->s2length, who_am_i());
952                 exit_cleanup(RERR_PROTOCOL);
953         }
954         sum->remainder = read_int(f);
955         if (sum->remainder < 0 || sum->remainder > sum->blength) {
956                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
957                         (long)sum->remainder, who_am_i());
958                 exit_cleanup(RERR_PROTOCOL);
959         }
960 }
961
962 /* Send the values from a sum_struct over the socket.  Set sum to
963  * NULL if there are no checksums to send.  This is called by both
964  * the generator and the sender. */
965 void write_sum_head(int f, struct sum_struct *sum)
966 {
967         static struct sum_struct null_sum;
968
969         if (sum == NULL)
970                 sum = &null_sum;
971
972         write_int(f, sum->count);
973         write_int(f, sum->blength);
974         if (protocol_version >= 27)
975                 write_int(f, sum->s2length);
976         write_int(f, sum->remainder);
977 }
978
979 /**
980  * Sleep after writing to limit I/O bandwidth usage.
981  *
982  * @todo Rather than sleeping after each write, it might be better to
983  * use some kind of averaging.  The current algorithm seems to always
984  * use a bit less bandwidth than specified, because it doesn't make up
985  * for slow periods.  But arguably this is a feature.  In addition, we
986  * ought to take the time used to write the data into account.
987  *
988  * During some phases of big transfers (file FOO is uptodate) this is
989  * called with a small bytes_written every time.  As the kernel has to
990  * round small waits up to guarantee that we actually wait at least the
991  * requested number of microseconds, this can become grossly inaccurate.
992  * We therefore keep track of the bytes we've written over time and only
993  * sleep when the accumulated delay is at least 1 tenth of a second.
994  **/
995 static void sleep_for_bwlimit(int bytes_written)
996 {
997         static struct timeval prior_tv;
998         static long total_written = 0;
999         struct timeval tv, start_tv;
1000         long elapsed_usec, sleep_usec;
1001
1002 #define ONE_SEC 1000000L /* # of microseconds in a second */
1003
1004         if (!bwlimit_writemax)
1005                 return;
1006
1007         total_written += bytes_written;
1008
1009         gettimeofday(&start_tv, NULL);
1010         if (prior_tv.tv_sec) {
1011                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
1012                              + (start_tv.tv_usec - prior_tv.tv_usec);
1013                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
1014                 if (total_written < 0)
1015                         total_written = 0;
1016         }
1017
1018         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
1019         if (sleep_usec < ONE_SEC / 10) {
1020                 prior_tv = start_tv;
1021                 return;
1022         }
1023
1024         tv.tv_sec  = sleep_usec / ONE_SEC;
1025         tv.tv_usec = sleep_usec % ONE_SEC;
1026         select(0, NULL, NULL, NULL, &tv);
1027
1028         gettimeofday(&prior_tv, NULL);
1029         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
1030                      + (prior_tv.tv_usec - start_tv.tv_usec);
1031         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
1032 }
1033
1034 /* Write len bytes to the file descriptor fd, looping as necessary to get
1035  * the job done and also (in certain circumstances) reading any data on
1036  * msg_fd_in to avoid deadlock.
1037  *
1038  * This function underlies the multiplexing system.  The body of the
1039  * application never calls this function directly. */
1040 static void writefd_unbuffered(int fd,char *buf,size_t len)
1041 {
1042         size_t n, total = 0;
1043         fd_set w_fds, r_fds;
1044         int maxfd, count, cnt, using_r_fds;
1045         int defer_save = defer_forwarding_messages;
1046         struct timeval tv;
1047
1048         no_flush++;
1049
1050         while (total < len) {
1051                 FD_ZERO(&w_fds);
1052                 FD_SET(fd,&w_fds);
1053                 maxfd = fd;
1054
1055                 if (msg_fd_in >= 0) {
1056                         FD_ZERO(&r_fds);
1057                         FD_SET(msg_fd_in,&r_fds);
1058                         if (msg_fd_in > maxfd)
1059                                 maxfd = msg_fd_in;
1060                         using_r_fds = 1;
1061                 } else
1062                         using_r_fds = 0;
1063
1064                 tv.tv_sec = select_timeout;
1065                 tv.tv_usec = 0;
1066
1067                 errno = 0;
1068                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1069                                &w_fds, NULL, &tv);
1070
1071                 if (count <= 0) {
1072                         if (count < 0 && errno == EBADF)
1073                                 exit_cleanup(RERR_SOCKETIO);
1074                         check_timeout();
1075                         continue;
1076                 }
1077
1078                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1079                         read_msg_fd();
1080
1081                 if (!FD_ISSET(fd, &w_fds))
1082                         continue;
1083
1084                 n = len - total;
1085                 if (bwlimit_writemax && n > bwlimit_writemax)
1086                         n = bwlimit_writemax;
1087                 cnt = write(fd, buf + total, n);
1088
1089                 if (cnt <= 0) {
1090                         if (cnt < 0) {
1091                                 if (errno == EINTR)
1092                                         continue;
1093                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1094                                         msleep(1);
1095                                         continue;
1096                                 }
1097                         }
1098
1099                         /* Don't try to write errors back across the stream. */
1100                         if (fd == sock_f_out)
1101                                 close_multiplexing_out();
1102                         rsyserr(FERROR, errno,
1103                                 "writefd_unbuffered failed to write %ld bytes [%s]",
1104                                 (long)len, who_am_i());
1105                         /* If the other side is sending us error messages, try
1106                          * to grab any messages they sent before they died. */
1107                         while (fd == sock_f_out && io_multiplexing_in) {
1108                                 set_io_timeout(30);
1109                                 ignore_timeout = 0;
1110                                 readfd_unbuffered(sock_f_in, io_filesfrom_buf,
1111                                                   sizeof io_filesfrom_buf);
1112                         }
1113                         exit_cleanup(RERR_STREAMIO);
1114                 }
1115
1116                 total += cnt;
1117                 defer_forwarding_messages = 1;
1118
1119                 if (fd == sock_f_out) {
1120                         if (io_timeout || am_generator)
1121                                 last_io_out = time(NULL);
1122                         sleep_for_bwlimit(cnt);
1123                 }
1124         }
1125
1126         defer_forwarding_messages = defer_save;
1127         no_flush--;
1128 }
1129
1130 static void msg2sndr_flush(void)
1131 {
1132         if (defer_forwarding_messages)
1133                 return;
1134
1135         while (msg2sndr.head && io_multiplexing_out) {
1136                 struct msg_list_item *m = msg2sndr.head;
1137                 if (!(msg2sndr.head = m->next))
1138                         msg2sndr.tail = NULL;
1139                 stats.total_written += m->len;
1140                 defer_forwarding_messages = 1;
1141                 writefd_unbuffered(sock_f_out, m->buf, m->len);
1142                 defer_forwarding_messages = 0;
1143                 free(m);
1144         }
1145 }
1146
1147 /**
1148  * Write an message to a multiplexed stream. If this fails then rsync
1149  * exits.
1150  **/
1151 static void mplex_write(enum msgcode code, char *buf, size_t len)
1152 {
1153         char buffer[1024];
1154         size_t n = len;
1155
1156         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
1157
1158         if (n > sizeof buffer - 4)
1159                 n = 0;
1160         else
1161                 memcpy(buffer + 4, buf, n);
1162
1163         writefd_unbuffered(sock_f_out, buffer, n+4);
1164
1165         len -= n;
1166         buf += n;
1167
1168         if (len) {
1169                 defer_forwarding_messages = 1;
1170                 writefd_unbuffered(sock_f_out, buf, len);
1171                 defer_forwarding_messages = 0;
1172                 msg2sndr_flush();
1173         }
1174 }
1175
1176 void io_flush(int flush_it_all)
1177 {
1178         msg2genr_flush(flush_it_all);
1179         msg2sndr_flush();
1180
1181         if (!iobuf_out_cnt || no_flush)
1182                 return;
1183
1184         if (io_multiplexing_out)
1185                 mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
1186         else
1187                 writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
1188         iobuf_out_cnt = 0;
1189 }
1190
1191 static void writefd(int fd,char *buf,size_t len)
1192 {
1193         if (fd == msg_fd_out) {
1194                 rprintf(FERROR, "Internal error: wrong write used in receiver.\n");
1195                 exit_cleanup(RERR_PROTOCOL);
1196         }
1197
1198         if (fd == sock_f_out)
1199                 stats.total_written += len;
1200
1201         if (fd == write_batch_monitor_out) {
1202                 if ((size_t)write(batch_fd, buf, len) != len)
1203                         exit_cleanup(RERR_FILEIO);
1204         }
1205
1206         if (!iobuf_out || fd != sock_f_out) {
1207                 writefd_unbuffered(fd, buf, len);
1208                 return;
1209         }
1210
1211         while (len) {
1212                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1213                 if (n > 0) {
1214                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1215                         buf += n;
1216                         len -= n;
1217                         iobuf_out_cnt += n;
1218                 }
1219
1220                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1221                         io_flush(NORMAL_FLUSH);
1222         }
1223 }
1224
1225 void write_shortint(int f, int x)
1226 {
1227         uchar b[2];
1228         b[0] = x;
1229         b[1] = x >> 8;
1230         writefd(f, (char *)b, 2);
1231 }
1232
1233 void write_int(int f,int32 x)
1234 {
1235         char b[4];
1236         SIVAL(b,0,x);
1237         writefd(f,b,4);
1238 }
1239
1240 /*
1241  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1242  * 64-bit types on this platform.
1243  */
1244 void write_longint(int f, int64 x)
1245 {
1246         char b[8];
1247
1248         if (x <= 0x7FFFFFFF) {
1249                 write_int(f, (int)x);
1250                 return;
1251         }
1252
1253 #if SIZEOF_INT64 < 8
1254         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1255         exit_cleanup(RERR_UNSUPPORTED);
1256 #else
1257         write_int(f, (int32)0xFFFFFFFF);
1258         SIVAL(b,0,(x&0xFFFFFFFF));
1259         SIVAL(b,4,((x>>32)&0xFFFFFFFF));
1260
1261         writefd(f,b,8);
1262 #endif
1263 }
1264
1265 void write_buf(int f,char *buf,size_t len)
1266 {
1267         writefd(f,buf,len);
1268 }
1269
1270 /** Write a string to the connection */
1271 void write_sbuf(int f, char *buf)
1272 {
1273         writefd(f, buf, strlen(buf));
1274 }
1275
1276 void write_byte(int f, uchar c)
1277 {
1278         writefd(f, (char *)&c, 1);
1279 }
1280
1281 void write_vstring(int f, char *str, int len)
1282 {
1283         uchar lenbuf[3], *lb = lenbuf;
1284
1285         if (len > 0x7F) {
1286                 if (len > 0x7FFF) {
1287                         rprintf(FERROR,
1288                                 "attempting to send over-long vstring (%d > %d)\n",
1289                                 len, 0x7FFF);
1290                         exit_cleanup(RERR_PROTOCOL);
1291                 }
1292                 *lb++ = len / 0x100 + 0x80;
1293         }
1294         *lb = len;
1295
1296         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1297         if (len)
1298                 writefd(f, str, len);
1299 }
1300
1301 /**
1302  * Read a line of up to @p maxlen characters into @p buf (not counting
1303  * the trailing null).  Strips the (required) trailing newline and all
1304  * carriage returns.
1305  *
1306  * @return 1 for success; 0 for I/O error or truncation.
1307  **/
1308 int read_line(int f, char *buf, size_t maxlen)
1309 {
1310         while (maxlen) {
1311                 buf[0] = 0;
1312                 read_buf(f, buf, 1);
1313                 if (buf[0] == 0)
1314                         return 0;
1315                 if (buf[0] == '\n')
1316                         break;
1317                 if (buf[0] != '\r') {
1318                         buf++;
1319                         maxlen--;
1320                 }
1321         }
1322         *buf = '\0';
1323         return maxlen > 0;
1324 }
1325
1326 void io_printf(int fd, const char *format, ...)
1327 {
1328         va_list ap;
1329         char buf[BIGPATHBUFLEN];
1330         int len;
1331
1332         va_start(ap, format);
1333         len = vsnprintf(buf, sizeof buf, format, ap);
1334         va_end(ap);
1335
1336         if (len < 0)
1337                 exit_cleanup(RERR_STREAMIO);
1338
1339         if (len > (int)sizeof buf) {
1340                 rprintf(FERROR, "io_printf() was too long for the buffer.\n");
1341                 exit_cleanup(RERR_STREAMIO);
1342         }
1343
1344         write_sbuf(fd, buf);
1345 }
1346
1347 /** Setup for multiplexing a MSG_* stream with the data stream. */
1348 void io_start_multiplex_out(void)
1349 {
1350         io_flush(NORMAL_FLUSH);
1351         io_start_buffering_out();
1352         io_multiplexing_out = 1;
1353 }
1354
1355 /** Setup for multiplexing a MSG_* stream with the data stream. */
1356 void io_start_multiplex_in(void)
1357 {
1358         io_flush(NORMAL_FLUSH);
1359         io_start_buffering_in();
1360         io_multiplexing_in = 1;
1361 }
1362
1363 /** Write an message to the multiplexed data stream. */
1364 int io_multiplex_write(enum msgcode code, char *buf, size_t len)
1365 {
1366         if (!io_multiplexing_out)
1367                 return 0;
1368
1369         io_flush(NORMAL_FLUSH);
1370         stats.total_written += (len+4);
1371         mplex_write(code, buf, len);
1372         return 1;
1373 }
1374
1375 void close_multiplexing_in(void)
1376 {
1377         io_multiplexing_in = 0;
1378 }
1379
1380 /** Stop output multiplexing. */
1381 void close_multiplexing_out(void)
1382 {
1383         io_multiplexing_out = 0;
1384 }
1385
1386 void start_write_batch(int fd)
1387 {
1388         write_stream_flags(batch_fd);
1389
1390         /* Some communication has already taken place, but we don't
1391          * enable batch writing until here so that we can write a
1392          * canonical record of the communication even though the
1393          * actual communication so far depends on whether a daemon
1394          * is involved. */
1395         write_int(batch_fd, protocol_version);
1396         write_int(batch_fd, checksum_seed);
1397
1398         if (am_sender)
1399                 write_batch_monitor_out = fd;
1400         else
1401                 write_batch_monitor_in = fd;
1402 }
1403
1404 void stop_write_batch(void)
1405 {
1406         write_batch_monitor_out = -1;
1407         write_batch_monitor_in = -1;
1408 }