42a6bf0c87ee163152592b5e6751737013457dd8
[rsync.git] / io.c
1 /* -*- c-file-style: "linux" -*-
2  *
3  * Copyright (C) 1996-2001 by Andrew Tridgell
4  * Copyright (C) Paul Mackerras 1996
5  * Copyright (C) 2001, 2002 by Martin Pool <mbp@samba.org>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 2 of the License, or
10  * (at your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  */
21
22 /**
23  * @file io.c
24  *
25  * Socket and pipe I/O utilities used in rsync.
26  *
27  * rsync provides its own multiplexing system, which is used to send
28  * stderr and stdout over a single socket.  We need this because
29  * stdout normally carries the binary data stream, and stderr all our
30  * error messages.
31  *
32  * For historical reasons this is off during the start of the
33  * connection, but it's switched on quite early using
34  * io_start_multiplex_out() and io_start_multiplex_in().
35  **/
36
37 #include "rsync.h"
38
39 /** If no timeout is specified then use a 60 second select timeout */
40 #define SELECT_TIMEOUT 60
41
42 extern int bwlimit;
43 extern size_t bwlimit_writemax;
44 extern int verbose;
45 extern int io_timeout;
46 extern int allowed_lull;
47 extern int am_server;
48 extern int am_daemon;
49 extern int am_sender;
50 extern int am_generator;
51 extern int eol_nulls;
52 extern int csum_length;
53 extern int checksum_seed;
54 extern int protocol_version;
55 extern int remove_sent_files;
56 extern int preserve_hard_links;
57 extern char *filesfrom_host;
58 extern struct stats stats;
59 extern struct file_list *the_file_list;
60
61 const char phase_unknown[] = "unknown";
62 int select_timeout = SELECT_TIMEOUT;
63 int ignore_timeout = 0;
64 int batch_fd = -1;
65 int batch_gen_fd = -1;
66
67 /**
68  * The connection might be dropped at some point; perhaps because the
69  * remote instance crashed.  Just giving the offset on the stream is
70  * not very helpful.  So instead we try to make io_phase_name point to
71  * something useful.
72  *
73  * For buffered/multiplexed I/O these names will be somewhat
74  * approximate; perhaps for ease of support we would rather make the
75  * buffer always flush when a single application-level I/O finishes.
76  *
77  * @todo Perhaps we want some simple stack functionality, but there's
78  * no need to overdo it.
79  **/
80 const char *io_write_phase = phase_unknown;
81 const char *io_read_phase = phase_unknown;
82
83 /* Ignore an EOF error if non-zero. See whine_about_eof(). */
84 int kluge_around_eof = 0;
85
86 int msg_fd_in = -1;
87 int msg_fd_out = -1;
88 int sock_f_in = -1;
89 int sock_f_out = -1;
90
91 static int io_multiplexing_out;
92 static int io_multiplexing_in;
93 static time_t last_io;
94 static int no_flush;
95
96 static int write_batch_monitor_in = -1;
97 static int write_batch_monitor_out = -1;
98
99 static int io_filesfrom_f_in = -1;
100 static int io_filesfrom_f_out = -1;
101 static char io_filesfrom_buf[2048];
102 static char *io_filesfrom_bp;
103 static char io_filesfrom_lastchar;
104 static int io_filesfrom_buflen;
105 static size_t contiguous_write_len = 0;
106
107 static void read_loop(int fd, char *buf, size_t len);
108
109 struct flist_ndx_item {
110         struct flist_ndx_item *next;
111         int ndx;
112 };
113
114 struct flist_ndx_list {
115         struct flist_ndx_item *head, *tail;
116 };
117
118 static struct flist_ndx_list redo_list, hlink_list;
119
120 struct msg_list {
121         struct msg_list *next;
122         char *buf;
123         int len;
124 };
125
126 static struct msg_list *msg_list_head;
127 static struct msg_list *msg_list_tail;
128
129 static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
130 {
131         struct flist_ndx_item *item;
132
133         if (!(item = new(struct flist_ndx_item)))
134                 out_of_memory("flist_ndx_push");
135         item->next = NULL;
136         item->ndx = ndx;
137         if (lp->tail)
138                 lp->tail->next = item;
139         else
140                 lp->head = item;
141         lp->tail = item;
142 }
143
144 static int flist_ndx_pop(struct flist_ndx_list *lp)
145 {
146         struct flist_ndx_item *next;
147         int ndx;
148
149         if (!lp->head)
150                 return -1;
151
152         ndx = lp->head->ndx;
153         next = lp->head->next;
154         free(lp->head);
155         lp->head = next;
156         if (!next)
157                 lp->tail = NULL;
158
159         return ndx;
160 }
161
162 static void check_timeout(void)
163 {
164         time_t t;
165
166         if (!io_timeout || ignore_timeout)
167                 return;
168
169         if (!last_io) {
170                 last_io = time(NULL);
171                 return;
172         }
173
174         t = time(NULL);
175
176         if (t - last_io >= io_timeout) {
177                 if (!am_server && !am_daemon) {
178                         rprintf(FERROR, "io timeout after %d seconds -- exiting\n",
179                                 (int)(t-last_io));
180                 }
181                 exit_cleanup(RERR_TIMEOUT);
182         }
183 }
184
185 /* Note the fds used for the main socket (which might really be a pipe
186  * for a local transfer, but we can ignore that). */
187 void io_set_sock_fds(int f_in, int f_out)
188 {
189         sock_f_in = f_in;
190         sock_f_out = f_out;
191 }
192
193 /* Setup the fd used to receive MSG_* messages.  Only needed during the
194  * early stages of being a local sender (up through the sending of the
195  * file list) or when we're the generator (to fetch the messages from
196  * the receiver). */
197 void set_msg_fd_in(int fd)
198 {
199         msg_fd_in = fd;
200 }
201
202 /* Setup the fd used to send our MSG_* messages.  Only needed when
203  * we're the receiver (to send our messages to the generator). */
204 void set_msg_fd_out(int fd)
205 {
206         msg_fd_out = fd;
207         set_nonblocking(msg_fd_out);
208 }
209
210 /* Add a message to the pending MSG_* list. */
211 static void msg_list_add(int code, char *buf, int len)
212 {
213         struct msg_list *ml;
214
215         if (!(ml = new(struct msg_list)))
216                 out_of_memory("msg_list_add");
217         ml->next = NULL;
218         if (!(ml->buf = new_array(char, len+4)))
219                 out_of_memory("msg_list_add");
220         SIVAL(ml->buf, 0, ((code+MPLEX_BASE)<<24) | len);
221         memcpy(ml->buf+4, buf, len);
222         ml->len = len+4;
223         if (msg_list_tail)
224                 msg_list_tail->next = ml;
225         else
226                 msg_list_head = ml;
227         msg_list_tail = ml;
228 }
229
230 void send_msg(enum msgcode code, char *buf, int len)
231 {
232         if (msg_fd_out < 0) {
233                 io_multiplex_write(code, buf, len);
234                 return;
235         }
236         msg_list_add(code, buf, len);
237         msg_list_push(NORMAL_FLUSH);
238 }
239
240 /* Read a message from the MSG_* fd and handle it.  This is called either
241  * during the early stages of being a local sender (up through the sending
242  * of the file list) or when we're the generator (to fetch the messages
243  * from the receiver). */
244 static void read_msg_fd(void)
245 {
246         char buf[2048];
247         size_t n;
248         int fd = msg_fd_in;
249         int tag, len;
250
251         /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
252          * to this routine from writefd_unbuffered(). */
253         msg_fd_in = -1;
254
255         read_loop(fd, buf, 4);
256         tag = IVAL(buf, 0);
257
258         len = tag & 0xFFFFFF;
259         tag = (tag >> 24) - MPLEX_BASE;
260
261         switch (tag) {
262         case MSG_DONE:
263                 if (len != 0 || !am_generator) {
264                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
265                         exit_cleanup(RERR_STREAMIO);
266                 }
267                 flist_ndx_push(&redo_list, -1);
268                 break;
269         case MSG_REDO:
270                 if (len != 4 || !am_generator) {
271                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
272                         exit_cleanup(RERR_STREAMIO);
273                 }
274                 read_loop(fd, buf, 4);
275                 flist_ndx_push(&redo_list, IVAL(buf,0));
276                 break;
277         case MSG_DELETED:
278                 if (len >= (int)sizeof buf || !am_generator) {
279                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
280                         exit_cleanup(RERR_STREAMIO);
281                 }
282                 read_loop(fd, buf, len);
283                 io_multiplex_write(MSG_DELETED, buf, len);
284                 break;
285         case MSG_SUCCESS:
286                 if (len != 4 || !am_generator) {
287                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
288                         exit_cleanup(RERR_STREAMIO);
289                 }
290                 read_loop(fd, buf, len);
291                 if (remove_sent_files)
292                         io_multiplex_write(MSG_SUCCESS, buf, len);
293                 if (preserve_hard_links)
294                         flist_ndx_push(&hlink_list, IVAL(buf,0));
295                 break;
296         case MSG_INFO:
297         case MSG_ERROR:
298         case MSG_LOG:
299                 while (len) {
300                         n = len;
301                         if (n >= sizeof buf)
302                                 n = sizeof buf - 1;
303                         read_loop(fd, buf, n);
304                         rwrite((enum logcode)tag, buf, n);
305                         len -= n;
306                 }
307                 break;
308         default:
309                 rprintf(FERROR, "unknown message %d:%d\n", tag, len);
310                 exit_cleanup(RERR_STREAMIO);
311         }
312
313         msg_fd_in = fd;
314 }
315
316 /* Try to push messages off the list onto the wire.  If we leave with more
317  * to do, return 0.  On error, return -1.  If everything flushed, return 1.
318  * This is only active in the receiver. */
319 int msg_list_push(int flush_it_all)
320 {
321         static int written = 0;
322         struct timeval tv;
323         fd_set fds;
324
325         if (msg_fd_out < 0)
326                 return -1;
327
328         while (msg_list_head) {
329                 struct msg_list *ml = msg_list_head;
330                 int n = write(msg_fd_out, ml->buf + written, ml->len - written);
331                 if (n < 0) {
332                         if (errno == EINTR)
333                                 continue;
334                         if (errno != EWOULDBLOCK && errno != EAGAIN)
335                                 return -1;
336                         if (!flush_it_all)
337                                 return 0;
338                         FD_ZERO(&fds);
339                         FD_SET(msg_fd_out, &fds);
340                         tv.tv_sec = select_timeout;
341                         tv.tv_usec = 0;
342                         if (!select(msg_fd_out+1, NULL, &fds, NULL, &tv))
343                                 check_timeout();
344                 } else if ((written += n) == ml->len) {
345                         free(ml->buf);
346                         msg_list_head = ml->next;
347                         if (!msg_list_head)
348                                 msg_list_tail = NULL;
349                         free(ml);
350                         written = 0;
351                 }
352         }
353         return 1;
354 }
355
356 int get_redo_num(int itemizing, enum logcode code)
357 {
358         while (1) {
359                 if (hlink_list.head)
360                         check_for_finished_hlinks(itemizing, code);
361                 if (redo_list.head)
362                         break;
363                 read_msg_fd();
364         }
365
366         return flist_ndx_pop(&redo_list);
367 }
368
369 int get_hlink_num(void)
370 {
371         return flist_ndx_pop(&hlink_list);
372 }
373
374 /**
375  * When we're the receiver and we have a local --files-from list of names
376  * that needs to be sent over the socket to the sender, we have to do two
377  * things at the same time: send the sender a list of what files we're
378  * processing and read the incoming file+info list from the sender.  We do
379  * this by augmenting the read_timeout() function to copy this data.  It
380  * uses the io_filesfrom_buf to read a block of data from f_in (when it is
381  * ready, since it might be a pipe) and then blast it out f_out (when it
382  * is ready to receive more data).
383  */
384 void io_set_filesfrom_fds(int f_in, int f_out)
385 {
386         io_filesfrom_f_in = f_in;
387         io_filesfrom_f_out = f_out;
388         io_filesfrom_bp = io_filesfrom_buf;
389         io_filesfrom_lastchar = '\0';
390         io_filesfrom_buflen = 0;
391 }
392
393 /* It's almost always an error to get an EOF when we're trying to read from the
394  * network, because the protocol is (for the most part) self-terminating.
395  *
396  * There is one case for the receiver when it is at the end of the transfer
397  * (hanging around reading any keep-alive packets that might come its way): if
398  * the sender dies before the generator's kill-signal comes through, we can end
399  * up here needing to loop until the kill-signal arrives.  In this situation,
400  * kluge_around_eof will be < 0.
401  *
402  * There is another case for older protocol versions (< 24) where the module
403  * listing was not terminated, so we must ignore an EOF error in that case and
404  * exit.  In this situation, kluge_around_eof will be > 0. */
405 static void whine_about_eof(int fd)
406 {
407         if (kluge_around_eof && fd == sock_f_in) {
408                 int i;
409                 if (kluge_around_eof > 0)
410                         exit_cleanup(0);
411                 /* If we're still here after 10 seconds, exit with an error. */
412                 for (i = 10*1000/20; i--; )
413                         msleep(20);
414         }
415
416         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
417                 "(%.0f bytes received so far) [%s]\n",
418                 (double)stats.total_read, who_am_i());
419
420         exit_cleanup(RERR_STREAMIO);
421 }
422
423
424 /**
425  * Read from a socket with I/O timeout. return the number of bytes
426  * read. If no bytes can be read then exit, never return a number <= 0.
427  *
428  * TODO: If the remote shell connection fails, then current versions
429  * actually report an "unexpected EOF" error here.  Since it's a
430  * fairly common mistake to try to use rsh when ssh is required, we
431  * should trap that: if we fail to read any data at all, we should
432  * give a better explanation.  We can tell whether the connection has
433  * started by looking e.g. at whether the remote version is known yet.
434  */
435 static int read_timeout(int fd, char *buf, size_t len)
436 {
437         int n, ret = 0;
438
439         io_flush(NORMAL_FLUSH);
440
441         while (ret == 0) {
442                 /* until we manage to read *something* */
443                 fd_set r_fds, w_fds;
444                 struct timeval tv;
445                 int maxfd = fd;
446                 int count;
447
448                 FD_ZERO(&r_fds);
449                 FD_ZERO(&w_fds);
450                 FD_SET(fd, &r_fds);
451                 if (msg_list_head) {
452                         FD_SET(msg_fd_out, &w_fds);
453                         if (msg_fd_out > maxfd)
454                                 maxfd = msg_fd_out;
455                 }
456                 if (io_filesfrom_f_out >= 0) {
457                         int new_fd;
458                         if (io_filesfrom_buflen == 0) {
459                                 if (io_filesfrom_f_in >= 0) {
460                                         FD_SET(io_filesfrom_f_in, &r_fds);
461                                         new_fd = io_filesfrom_f_in;
462                                 } else {
463                                         io_filesfrom_f_out = -1;
464                                         new_fd = -1;
465                                 }
466                         } else {
467                                 FD_SET(io_filesfrom_f_out, &w_fds);
468                                 new_fd = io_filesfrom_f_out;
469                         }
470                         if (new_fd > maxfd)
471                                 maxfd = new_fd;
472                 }
473
474                 tv.tv_sec = select_timeout;
475                 tv.tv_usec = 0;
476
477                 errno = 0;
478
479                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
480
481                 if (count <= 0) {
482                         if (errno == EBADF)
483                                 exit_cleanup(RERR_SOCKETIO);
484                         check_timeout();
485                         continue;
486                 }
487
488                 if (msg_list_head && FD_ISSET(msg_fd_out, &w_fds))
489                         msg_list_push(NORMAL_FLUSH);
490
491                 if (io_filesfrom_f_out >= 0) {
492                         if (io_filesfrom_buflen) {
493                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
494                                         int l = write(io_filesfrom_f_out,
495                                                       io_filesfrom_bp,
496                                                       io_filesfrom_buflen);
497                                         if (l > 0) {
498                                                 if (!(io_filesfrom_buflen -= l))
499                                                         io_filesfrom_bp = io_filesfrom_buf;
500                                                 else
501                                                         io_filesfrom_bp += l;
502                                         } else {
503                                                 /* XXX should we complain? */
504                                                 io_filesfrom_f_out = -1;
505                                         }
506                                 }
507                         } else if (io_filesfrom_f_in >= 0) {
508                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
509                                         int l = read(io_filesfrom_f_in,
510                                                      io_filesfrom_buf,
511                                                      sizeof io_filesfrom_buf);
512                                         if (l <= 0) {
513                                                 /* Send end-of-file marker */
514                                                 io_filesfrom_buf[0] = '\0';
515                                                 io_filesfrom_buf[1] = '\0';
516                                                 io_filesfrom_buflen = io_filesfrom_lastchar? 2 : 1;
517                                                 io_filesfrom_f_in = -1;
518                                         } else {
519                                                 if (!eol_nulls) {
520                                                         char *s = io_filesfrom_buf + l;
521                                                         /* Transform CR and/or LF into '\0' */
522                                                         while (s-- > io_filesfrom_buf) {
523                                                                 if (*s == '\n' || *s == '\r')
524                                                                         *s = '\0';
525                                                         }
526                                                 }
527                                                 if (!io_filesfrom_lastchar) {
528                                                         /* Last buf ended with a '\0', so don't
529                                                          * let this buf start with one. */
530                                                         while (l && !*io_filesfrom_bp)
531                                                                 io_filesfrom_bp++, l--;
532                                                 }
533                                                 if (!l)
534                                                         io_filesfrom_bp = io_filesfrom_buf;
535                                                 else {
536                                                         char *f = io_filesfrom_bp;
537                                                         char *t = f;
538                                                         char *eob = f + l;
539                                                         /* Eliminate any multi-'\0' runs. */
540                                                         while (f != eob) {
541                                                                 if (!(*t++ = *f++)) {
542                                                                         while (f != eob && !*f)
543                                                                                 f++, l--;
544                                                                 }
545                                                         }
546                                                         io_filesfrom_lastchar = f[-1];
547                                                 }
548                                                 io_filesfrom_buflen = l;
549                                         }
550                                 }
551                         }
552                 }
553
554                 if (!FD_ISSET(fd, &r_fds))
555                         continue;
556
557                 n = read(fd, buf, len);
558
559                 if (n <= 0) {
560                         if (n == 0)
561                                 whine_about_eof(fd); /* Doesn't return. */
562                         if (errno == EINTR || errno == EWOULDBLOCK
563                             || errno == EAGAIN)
564                                 continue;
565
566                         /* Don't write errors on a dead socket. */
567                         if (fd == sock_f_in)
568                                 close_multiplexing_out();
569                         rsyserr(FERROR, errno, "read error");
570                         exit_cleanup(RERR_STREAMIO);
571                 }
572
573                 buf += n;
574                 len -= n;
575                 ret += n;
576
577                 if (io_timeout && fd == sock_f_in)
578                         last_io = time(NULL);
579         }
580
581         return ret;
582 }
583
584 /**
585  * Read a line into the "fname" buffer (which must be at least MAXPATHLEN
586  * characters long).
587  */
588 int read_filesfrom_line(int fd, char *fname)
589 {
590         char ch, *s, *eob = fname + MAXPATHLEN - 1;
591         int cnt;
592         int reading_remotely = filesfrom_host != NULL;
593         int nulls = eol_nulls || reading_remotely;
594
595   start:
596         s = fname;
597         while (1) {
598                 cnt = read(fd, &ch, 1);
599                 if (cnt < 0 && (errno == EWOULDBLOCK
600                   || errno == EINTR || errno == EAGAIN)) {
601                         struct timeval tv;
602                         fd_set fds;
603                         FD_ZERO(&fds);
604                         FD_SET(fd, &fds);
605                         tv.tv_sec = select_timeout;
606                         tv.tv_usec = 0;
607                         if (!select(fd+1, &fds, NULL, NULL, &tv))
608                                 check_timeout();
609                         continue;
610                 }
611                 if (cnt != 1)
612                         break;
613                 if (nulls? !ch : (ch == '\r' || ch == '\n')) {
614                         /* Skip empty lines if reading locally. */
615                         if (!reading_remotely && s == fname)
616                                 continue;
617                         break;
618                 }
619                 if (s < eob)
620                         *s++ = ch;
621         }
622         *s = '\0';
623
624         /* Dump comments. */
625         if (*fname == '#' || *fname == ';')
626                 goto start;
627
628         return s - fname;
629 }
630
631
632 static char *iobuf_out;
633 static int iobuf_out_cnt;
634
635 void io_start_buffering_out(void)
636 {
637         if (iobuf_out)
638                 return;
639         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
640                 out_of_memory("io_start_buffering_out");
641         iobuf_out_cnt = 0;
642 }
643
644
645 static char *iobuf_in;
646 static size_t iobuf_in_siz;
647
648 void io_start_buffering_in(void)
649 {
650         if (iobuf_in)
651                 return;
652         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
653         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
654                 out_of_memory("io_start_buffering_in");
655 }
656
657
658 void io_end_buffering(void)
659 {
660         io_flush(NORMAL_FLUSH);
661         if (!io_multiplexing_out) {
662                 free(iobuf_out);
663                 iobuf_out = NULL;
664         }
665 }
666
667
668 void maybe_send_keepalive(void)
669 {
670         if (time(NULL) - last_io >= allowed_lull) {
671                 if (!iobuf_out || !iobuf_out_cnt) {
672                         if (protocol_version < 29)
673                                 return; /* there's nothing we can do */
674                         write_int(sock_f_out, the_file_list->count);
675                         write_shortint(sock_f_out, ITEM_IS_NEW);
676                 }
677                 if (iobuf_out)
678                         io_flush(NORMAL_FLUSH);
679         }
680 }
681
682
683 /**
684  * Continue trying to read len bytes - don't return until len has been
685  * read.
686  **/
687 static void read_loop(int fd, char *buf, size_t len)
688 {
689         while (len) {
690                 int n = read_timeout(fd, buf, len);
691
692                 buf += n;
693                 len -= n;
694         }
695 }
696
697
698 /**
699  * Read from the file descriptor handling multiplexing - return number
700  * of bytes read.
701  *
702  * Never returns <= 0.
703  */
704 static int readfd_unbuffered(int fd, char *buf, size_t len)
705 {
706         static size_t remaining;
707         static size_t iobuf_in_ndx;
708         int tag, ret = 0;
709         char line[MAXPATHLEN+1];
710
711         if (!iobuf_in || fd != sock_f_in)
712                 return read_timeout(fd, buf, len);
713
714         if (!io_multiplexing_in && remaining == 0) {
715                 remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
716                 iobuf_in_ndx = 0;
717         }
718
719         while (ret == 0) {
720                 if (remaining) {
721                         len = MIN(len, remaining);
722                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
723                         iobuf_in_ndx += len;
724                         remaining -= len;
725                         ret = len;
726                         break;
727                 }
728
729                 read_loop(fd, line, 4);
730                 tag = IVAL(line, 0);
731
732                 remaining = tag & 0xFFFFFF;
733                 tag = (tag >> 24) - MPLEX_BASE;
734
735                 switch (tag) {
736                 case MSG_DATA:
737                         if (remaining > iobuf_in_siz) {
738                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
739                                                                remaining)))
740                                         out_of_memory("readfd_unbuffered");
741                                 iobuf_in_siz = remaining;
742                         }
743                         read_loop(fd, iobuf_in, remaining);
744                         iobuf_in_ndx = 0;
745                         break;
746                 case MSG_DELETED:
747                         if (remaining >= sizeof line) {
748                                 rprintf(FERROR, "invalid multi-message %d:%ld\n",
749                                         tag, (long)remaining);
750                                 exit_cleanup(RERR_STREAMIO);
751                         }
752                         read_loop(fd, line, remaining);
753                         line[remaining] = '\0';
754                         /* A directory name was sent with the trailing null */
755                         if (remaining > 0 && !line[remaining-1])
756                                 log_delete(line, S_IFDIR);
757                         else
758                                 log_delete(line, S_IFREG);
759                         remaining = 0;
760                         break;
761                 case MSG_SUCCESS:
762                         if (remaining != 4) {
763                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
764                                         tag, (long)remaining, who_am_i());
765                                 exit_cleanup(RERR_STREAMIO);
766                         }
767                         read_loop(fd, line, remaining);
768                         successful_send(IVAL(line, 0));
769                         remaining = 0;
770                         break;
771                 case MSG_INFO:
772                 case MSG_ERROR:
773                         if (remaining >= sizeof line) {
774                                 rprintf(FERROR,
775                                         "multiplexing overflow %d:%ld [%s]\n",
776                                         tag, (long)remaining, who_am_i());
777                                 exit_cleanup(RERR_STREAMIO);
778                         }
779                         read_loop(fd, line, remaining);
780                         rwrite((enum logcode)tag, line, remaining);
781                         remaining = 0;
782                         break;
783                 default:
784                         rprintf(FERROR, "unexpected tag %d [%s]\n",
785                                 tag, who_am_i());
786                         exit_cleanup(RERR_STREAMIO);
787                 }
788         }
789
790         if (remaining == 0)
791                 io_flush(NORMAL_FLUSH);
792
793         return ret;
794 }
795
796
797
798 /**
799  * Do a buffered read from @p fd.  Don't return until all @p n bytes
800  * have been read.  If all @p n can't be read then exit with an
801  * error.
802  **/
803 static void readfd(int fd, char *buffer, size_t N)
804 {
805         int  ret;
806         size_t total = 0;
807
808         while (total < N) {
809                 ret = readfd_unbuffered(fd, buffer + total, N-total);
810                 total += ret;
811         }
812
813         if (fd == write_batch_monitor_in) {
814                 if ((size_t)write(batch_fd, buffer, total) != total)
815                         exit_cleanup(RERR_FILEIO);
816         }
817
818         if (fd == sock_f_in)
819                 stats.total_read += total;
820 }
821
822
823 int read_shortint(int f)
824 {
825         uchar b[2];
826         readfd(f, (char *)b, 2);
827         return (b[1] << 8) + b[0];
828 }
829
830
831 int32 read_int(int f)
832 {
833         char b[4];
834         int32 ret;
835
836         readfd(f,b,4);
837         ret = IVAL(b,0);
838         if (ret == (int32)0xffffffff)
839                 return -1;
840         return ret;
841 }
842
843 int64 read_longint(int f)
844 {
845         int64 ret;
846         char b[8];
847         ret = read_int(f);
848
849         if ((int32)ret != (int32)0xffffffff)
850                 return ret;
851
852 #if SIZEOF_INT64 < 8
853         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
854         exit_cleanup(RERR_UNSUPPORTED);
855 #else
856         readfd(f,b,8);
857         ret = IVAL(b,0) | (((int64)IVAL(b,4))<<32);
858 #endif
859
860         return ret;
861 }
862
863 void read_buf(int f,char *buf,size_t len)
864 {
865         readfd(f,buf,len);
866 }
867
868 void read_sbuf(int f,char *buf,size_t len)
869 {
870         readfd(f, buf, len);
871         buf[len] = '\0';
872 }
873
874 uchar read_byte(int f)
875 {
876         uchar c;
877         readfd(f, (char *)&c, 1);
878         return c;
879 }
880
881 int read_vstring(int f, char *buf, int bufsize)
882 {
883         int len = read_byte(f);
884
885         if (len & 0x80)
886                 len = (len & ~0x80) * 0x100 + read_byte(f);
887
888         if (len >= bufsize) {
889                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
890                         len, bufsize - 1);
891                 return -1;
892         }
893
894         if (len)
895                 readfd(f, buf, len);
896         buf[len] = '\0';
897         return len;
898 }
899
900 /* Populate a sum_struct with values from the socket.  This is
901  * called by both the sender and the receiver. */
902 void read_sum_head(int f, struct sum_struct *sum)
903 {
904         sum->count = read_int(f);
905         sum->blength = read_int(f);
906         if (sum->blength < 0 || sum->blength > MAX_BLOCK_SIZE) {
907                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
908                         (long)sum->blength, who_am_i());
909                 exit_cleanup(RERR_PROTOCOL);
910         }
911         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
912         if (sum->s2length < 0 || sum->s2length > MD4_SUM_LENGTH) {
913                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
914                         sum->s2length, who_am_i());
915                 exit_cleanup(RERR_PROTOCOL);
916         }
917         sum->remainder = read_int(f);
918         if (sum->remainder < 0 || sum->remainder > sum->blength) {
919                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
920                         (long)sum->remainder, who_am_i());
921                 exit_cleanup(RERR_PROTOCOL);
922         }
923 }
924
925 /* Send the values from a sum_struct over the socket.  Set sum to
926  * NULL if there are no checksums to send.  This is called by both
927  * the generator and the sender. */
928 void write_sum_head(int f, struct sum_struct *sum)
929 {
930         static struct sum_struct null_sum;
931
932         if (sum == NULL)
933                 sum = &null_sum;
934
935         write_int(f, sum->count);
936         write_int(f, sum->blength);
937         if (protocol_version >= 27)
938                 write_int(f, sum->s2length);
939         write_int(f, sum->remainder);
940 }
941
942
943 /**
944  * Sleep after writing to limit I/O bandwidth usage.
945  *
946  * @todo Rather than sleeping after each write, it might be better to
947  * use some kind of averaging.  The current algorithm seems to always
948  * use a bit less bandwidth than specified, because it doesn't make up
949  * for slow periods.  But arguably this is a feature.  In addition, we
950  * ought to take the time used to write the data into account.
951  *
952  * During some phases of big transfers (file FOO is uptodate) this is
953  * called with a small bytes_written every time.  As the kernel has to
954  * round small waits up to guarantee that we actually wait at least the
955  * requested number of microseconds, this can become grossly inaccurate.
956  * We therefore keep track of the bytes we've written over time and only
957  * sleep when the accumulated delay is at least 1 tenth of a second.
958  **/
959 static void sleep_for_bwlimit(int bytes_written)
960 {
961         static struct timeval prior_tv;
962         static long total_written = 0; 
963         struct timeval tv, start_tv;
964         long elapsed_usec, sleep_usec;
965
966 #define ONE_SEC 1000000L /* # of microseconds in a second */
967
968         if (!bwlimit)
969                 return;
970
971         total_written += bytes_written; 
972
973         gettimeofday(&start_tv, NULL);
974         if (prior_tv.tv_sec) {
975                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
976                              + (start_tv.tv_usec - prior_tv.tv_usec);
977                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
978                 if (total_written < 0)
979                         total_written = 0;
980         }
981
982         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
983         if (sleep_usec < ONE_SEC / 10) {
984                 prior_tv = start_tv;
985                 return;
986         }
987
988         tv.tv_sec  = sleep_usec / ONE_SEC;
989         tv.tv_usec = sleep_usec % ONE_SEC;
990         select(0, NULL, NULL, NULL, &tv);
991
992         gettimeofday(&prior_tv, NULL);
993         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
994                      + (prior_tv.tv_usec - start_tv.tv_usec);
995         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
996 }
997
998
999 /* Write len bytes to the file descriptor fd, looping as necessary to get
1000  * the job done and also (in certain circumstnces) reading any data on
1001  * msg_fd_in to avoid deadlock.
1002  *
1003  * This function underlies the multiplexing system.  The body of the
1004  * application never calls this function directly. */
1005 static void writefd_unbuffered(int fd,char *buf,size_t len)
1006 {
1007         size_t n, total = 0;
1008         fd_set w_fds, r_fds;
1009         int maxfd, count, ret, using_r_fds;
1010         struct timeval tv;
1011
1012         no_flush++;
1013
1014         while (total < len) {
1015                 FD_ZERO(&w_fds);
1016                 FD_SET(fd,&w_fds);
1017                 maxfd = fd;
1018
1019                 if (msg_fd_in >= 0 && len-total >= contiguous_write_len) {
1020                         FD_ZERO(&r_fds);
1021                         FD_SET(msg_fd_in,&r_fds);
1022                         if (msg_fd_in > maxfd)
1023                                 maxfd = msg_fd_in;
1024                         using_r_fds = 1;
1025                 } else
1026                         using_r_fds = 0;
1027                 if (fd != sock_f_out && iobuf_out_cnt && no_flush == 1) {
1028                         FD_SET(sock_f_out, &w_fds);
1029                         if (sock_f_out > maxfd)
1030                                 maxfd = sock_f_out;
1031                 }
1032
1033                 tv.tv_sec = select_timeout;
1034                 tv.tv_usec = 0;
1035
1036                 errno = 0;
1037                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1038                                &w_fds, NULL, &tv);
1039
1040                 if (count <= 0) {
1041                         if (count < 0 && errno == EBADF)
1042                                 exit_cleanup(RERR_SOCKETIO);
1043                         check_timeout();
1044                         continue;
1045                 }
1046
1047                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1048                         read_msg_fd();
1049
1050                 if (!FD_ISSET(fd, &w_fds))
1051                         continue;
1052
1053                 n = len - total;
1054                 if (bwlimit && n > bwlimit_writemax)
1055                         n = bwlimit_writemax;
1056                 ret = write(fd, buf + total, n);
1057
1058                 if (ret <= 0) {
1059                         if (ret < 0) {
1060                                 if (errno == EINTR)
1061                                         continue;
1062                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1063                                         msleep(1);
1064                                         continue;
1065                                 }
1066                         }
1067
1068                         /* Don't try to write errors back across the stream. */
1069                         if (fd == sock_f_out)
1070                                 close_multiplexing_out();
1071                         rsyserr(FERROR, errno,
1072                                 "writefd_unbuffered failed to write %ld bytes: phase \"%s\" [%s]",
1073                                 (long)len, io_write_phase, who_am_i());
1074                         /* If the other side is sending us error messages, try
1075                          * to grab any messages they sent before they died. */
1076                         while (fd == sock_f_out && io_multiplexing_in) {
1077                                 io_timeout = select_timeout = 30;
1078                                 ignore_timeout = 0;
1079                                 readfd_unbuffered(sock_f_in, io_filesfrom_buf,
1080                                                   sizeof io_filesfrom_buf);
1081                         }
1082                         exit_cleanup(RERR_STREAMIO);
1083                 }
1084
1085                 total += ret;
1086
1087                 if (fd == sock_f_out) {
1088                         if (io_timeout)
1089                                 last_io = time(NULL);
1090                         sleep_for_bwlimit(ret);
1091                 }
1092         }
1093
1094         no_flush--;
1095 }
1096
1097
1098 /**
1099  * Write an message to a multiplexed stream. If this fails then rsync
1100  * exits.
1101  **/
1102 static void mplex_write(enum msgcode code, char *buf, size_t len)
1103 {
1104         char buffer[4096];
1105         size_t n = len;
1106
1107         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
1108
1109         /* When the generator reads messages from the msg_fd_in pipe, it can
1110          * cause output to occur down the socket.  Setting contiguous_write_len
1111          * prevents the reading of msg_fd_in once we actually start to write
1112          * this sequence of data (though we might read it before the start). */
1113         if (am_generator && msg_fd_in >= 0)
1114                 contiguous_write_len = len + 4;
1115
1116         if (n > sizeof buffer - 4)
1117                 n = sizeof buffer - 4;
1118
1119         memcpy(&buffer[4], buf, n);
1120         writefd_unbuffered(sock_f_out, buffer, n+4);
1121
1122         len -= n;
1123         buf += n;
1124
1125         if (len)
1126                 writefd_unbuffered(sock_f_out, buf, len);
1127
1128         if (am_generator && msg_fd_in >= 0)
1129                 contiguous_write_len = 0;
1130 }
1131
1132
1133 void io_flush(int flush_it_all)
1134 {
1135         msg_list_push(flush_it_all);
1136
1137         if (!iobuf_out_cnt || no_flush)
1138                 return;
1139
1140         if (io_multiplexing_out)
1141                 mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
1142         else
1143                 writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
1144         iobuf_out_cnt = 0;
1145 }
1146
1147
1148 static void writefd(int fd,char *buf,size_t len)
1149 {
1150         if (fd == msg_fd_out) {
1151                 rprintf(FERROR, "Internal error: wrong write used in receiver.\n");
1152                 exit_cleanup(RERR_PROTOCOL);
1153         }
1154
1155         if (fd == sock_f_out)
1156                 stats.total_written += len;
1157
1158         if (fd == write_batch_monitor_out) {
1159                 if ((size_t)write(batch_fd, buf, len) != len)
1160                         exit_cleanup(RERR_FILEIO);
1161         }
1162
1163         if (!iobuf_out || fd != sock_f_out) {
1164                 writefd_unbuffered(fd, buf, len);
1165                 return;
1166         }
1167
1168         while (len) {
1169                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1170                 if (n > 0) {
1171                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1172                         buf += n;
1173                         len -= n;
1174                         iobuf_out_cnt += n;
1175                 }
1176
1177                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1178                         io_flush(NORMAL_FLUSH);
1179         }
1180 }
1181
1182
1183 void write_shortint(int f, int x)
1184 {
1185         uchar b[2];
1186         b[0] = x;
1187         b[1] = x >> 8;
1188         writefd(f, (char *)b, 2);
1189 }
1190
1191
1192 void write_int(int f,int32 x)
1193 {
1194         char b[4];
1195         SIVAL(b,0,x);
1196         writefd(f,b,4);
1197 }
1198
1199
1200 void write_int_named(int f, int32 x, const char *phase)
1201 {
1202         io_write_phase = phase;
1203         write_int(f, x);
1204         io_write_phase = phase_unknown;
1205 }
1206
1207
1208 /*
1209  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1210  * 64-bit types on this platform.
1211  */
1212 void write_longint(int f, int64 x)
1213 {
1214         char b[8];
1215
1216         if (x <= 0x7FFFFFFF) {
1217                 write_int(f, (int)x);
1218                 return;
1219         }
1220
1221 #if SIZEOF_INT64 < 8
1222         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1223         exit_cleanup(RERR_UNSUPPORTED);
1224 #else
1225         write_int(f, (int32)0xFFFFFFFF);
1226         SIVAL(b,0,(x&0xFFFFFFFF));
1227         SIVAL(b,4,((x>>32)&0xFFFFFFFF));
1228
1229         writefd(f,b,8);
1230 #endif
1231 }
1232
1233 void write_buf(int f,char *buf,size_t len)
1234 {
1235         writefd(f,buf,len);
1236 }
1237
1238 /** Write a string to the connection */
1239 void write_sbuf(int f, char *buf)
1240 {
1241         writefd(f, buf, strlen(buf));
1242 }
1243
1244 void write_byte(int f, uchar c)
1245 {
1246         writefd(f, (char *)&c, 1);
1247 }
1248
1249 void write_vstring(int f, char *str, int len)
1250 {
1251         uchar lenbuf[3], *lb = lenbuf;
1252
1253         if (len > 0x7F) {
1254                 if (len > 0x7FFF) {
1255                         rprintf(FERROR,
1256                                 "attempting to send over-long vstring (%d > %d)\n",
1257                                 len, 0x7FFF);
1258                         exit_cleanup(RERR_PROTOCOL);
1259                 }
1260                 *lb++ = len / 0x100 + 0x80;
1261         }
1262         *lb = len;
1263
1264         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1265         if (len)
1266                 writefd(f, str, len);
1267 }
1268
1269
1270 /**
1271  * Read a line of up to @p maxlen characters into @p buf (not counting
1272  * the trailing null).  Strips the (required) trailing newline and all
1273  * carriage returns.
1274  *
1275  * @return 1 for success; 0 for I/O error or truncation.
1276  **/
1277 int read_line(int f, char *buf, size_t maxlen)
1278 {
1279         while (maxlen) {
1280                 buf[0] = 0;
1281                 read_buf(f, buf, 1);
1282                 if (buf[0] == 0)
1283                         return 0;
1284                 if (buf[0] == '\n')
1285                         break;
1286                 if (buf[0] != '\r') {
1287                         buf++;
1288                         maxlen--;
1289                 }
1290         }
1291         *buf = '\0';
1292         return maxlen > 0;
1293 }
1294
1295
1296 void io_printf(int fd, const char *format, ...)
1297 {
1298         va_list ap;
1299         char buf[1024];
1300         int len;
1301
1302         va_start(ap, format);
1303         len = vsnprintf(buf, sizeof buf, format, ap);
1304         va_end(ap);
1305
1306         if (len < 0)
1307                 exit_cleanup(RERR_STREAMIO);
1308
1309         write_sbuf(fd, buf);
1310 }
1311
1312
1313 /** Setup for multiplexing a MSG_* stream with the data stream. */
1314 void io_start_multiplex_out(void)
1315 {
1316         io_flush(NORMAL_FLUSH);
1317         io_start_buffering_out();
1318         io_multiplexing_out = 1;
1319 }
1320
1321 /** Setup for multiplexing a MSG_* stream with the data stream. */
1322 void io_start_multiplex_in(void)
1323 {
1324         io_flush(NORMAL_FLUSH);
1325         io_start_buffering_in();
1326         io_multiplexing_in = 1;
1327 }
1328
1329 /** Write an message to the multiplexed data stream. */
1330 int io_multiplex_write(enum msgcode code, char *buf, size_t len)
1331 {
1332         if (!io_multiplexing_out)
1333                 return 0;
1334
1335         io_flush(NORMAL_FLUSH);
1336         stats.total_written += (len+4);
1337         mplex_write(code, buf, len);
1338         return 1;
1339 }
1340
1341 void close_multiplexing_in(void)
1342 {
1343         io_multiplexing_in = 0;
1344 }
1345
1346 /** Stop output multiplexing. */
1347 void close_multiplexing_out(void)
1348 {
1349         io_multiplexing_out = 0;
1350 }
1351
1352 void start_write_batch(int fd)
1353 {
1354         write_stream_flags(batch_fd);
1355
1356         /* Some communication has already taken place, but we don't
1357          * enable batch writing until here so that we can write a
1358          * canonical record of the communication even though the
1359          * actual communication so far depends on whether a daemon
1360          * is involved. */
1361         write_int(batch_fd, protocol_version);
1362         write_int(batch_fd, checksum_seed);
1363
1364         if (am_sender)
1365                 write_batch_monitor_out = fd;
1366         else
1367                 write_batch_monitor_in = fd;
1368 }
1369
1370 void stop_write_batch(void)
1371 {
1372         write_batch_monitor_out = -1;
1373         write_batch_monitor_in = -1;
1374 }