Changed read_msg_fd() to always use send_msg()/rwrite() instead of
[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 along
20  * with this program; if not, write to the Free Software Foundation, Inc.,
21  * 51 Franklin Street - Fifth Floor, Boston, MA 02110-1301, 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                 send_msg(MSG_DELETED, buf, len);
271                 break;
272         case MSG_SUCCESS:
273                 if (len != 4 || !am_generator) {
274                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
275                         exit_cleanup(RERR_STREAMIO);
276                 }
277                 read_loop(fd, buf, len);
278                 if (remove_sent_files) {
279                         decrement_active_files(IVAL(buf,0));
280                         send_msg(MSG_SUCCESS, buf, len);
281                 }
282                 if (preserve_hard_links)
283                         flist_ndx_push(&hlink_list, IVAL(buf,0));
284                 break;
285         case MSG_SOCKERR:
286                 if (!am_generator) {
287                         rprintf(FERROR, "invalid message %d:%d\n", tag, len);
288                         exit_cleanup(RERR_STREAMIO);
289                 }
290                 close_multiplexing_out();
291                 /* FALL THROUGH */
292         case MSG_INFO:
293         case MSG_ERROR:
294         case MSG_LOG:
295                 while (len) {
296                         n = len;
297                         if (n >= sizeof buf)
298                                 n = sizeof buf - 1;
299                         read_loop(fd, buf, n);
300                         rwrite(tag, buf, n);
301                         len -= n;
302                 }
303                 break;
304         default:
305                 rprintf(FERROR, "unknown message %d:%d [%s]\n",
306                         tag, len, who_am_i());
307                 exit_cleanup(RERR_STREAMIO);
308         }
309
310         msg_fd_in = fd;
311 }
312
313 /* This is used by the generator to limit how many file transfers can
314  * be active at once when --remove-sent-files is specified.  Without
315  * this, sender-side deletions were mostly happening at the end. */
316 void increment_active_files(int ndx, int itemizing, enum logcode code)
317 {
318         /* TODO: tune these limits? */
319         while (active_filecnt >= (active_bytecnt >= 128*1024 ? 10 : 50)) {
320                 if (hlink_list.head)
321                         check_for_finished_hlinks(itemizing, code);
322                 read_msg_fd();
323         }
324
325         active_filecnt++;
326         active_bytecnt += the_file_list->files[ndx]->length;
327 }
328
329 void decrement_active_files(int ndx)
330 {
331         active_filecnt--;
332         active_bytecnt -= the_file_list->files[ndx]->length;
333 }
334
335 /* Try to push messages off the list onto the wire.  If we leave with more
336  * to do, return 0.  On error, return -1.  If everything flushed, return 1.
337  * This is only active in the receiver. */
338 static int msg2genr_flush(int flush_it_all)
339 {
340         static int written = 0;
341         struct timeval tv;
342         fd_set fds;
343
344         if (msg_fd_out < 0)
345                 return -1;
346
347         while (msg2genr.head) {
348                 struct msg_list_item *m = msg2genr.head;
349                 int n = write(msg_fd_out, m->buf + written, m->len - written);
350                 if (n < 0) {
351                         if (errno == EINTR)
352                                 continue;
353                         if (errno != EWOULDBLOCK && errno != EAGAIN)
354                                 return -1;
355                         if (!flush_it_all)
356                                 return 0;
357                         FD_ZERO(&fds);
358                         FD_SET(msg_fd_out, &fds);
359                         tv.tv_sec = select_timeout;
360                         tv.tv_usec = 0;
361                         if (!select(msg_fd_out+1, NULL, &fds, NULL, &tv))
362                                 check_timeout();
363                 } else if ((written += n) == m->len) {
364                         msg2genr.head = m->next;
365                         if (!msg2genr.head)
366                                 msg2genr.tail = NULL;
367                         free(m);
368                         written = 0;
369                 }
370         }
371         return 1;
372 }
373
374 int send_msg(enum msgcode code, char *buf, int len)
375 {
376         if (msg_fd_out < 0) {
377                 if (!defer_forwarding_messages)
378                         return io_multiplex_write(code, buf, len);
379                 if (!io_multiplexing_out)
380                         return 0;
381                 msg_list_add(&msg2sndr, code, buf, len);
382                 return 1;
383         }
384         msg_list_add(&msg2genr, code, buf, len);
385         msg2genr_flush(NORMAL_FLUSH);
386         return 1;
387 }
388
389 int get_redo_num(int itemizing, enum logcode code)
390 {
391         while (1) {
392                 if (hlink_list.head)
393                         check_for_finished_hlinks(itemizing, code);
394                 if (redo_list.head)
395                         break;
396                 read_msg_fd();
397         }
398
399         return flist_ndx_pop(&redo_list);
400 }
401
402 int get_hlink_num(void)
403 {
404         return flist_ndx_pop(&hlink_list);
405 }
406
407 /**
408  * When we're the receiver and we have a local --files-from list of names
409  * that needs to be sent over the socket to the sender, we have to do two
410  * things at the same time: send the sender a list of what files we're
411  * processing and read the incoming file+info list from the sender.  We do
412  * this by augmenting the read_timeout() function to copy this data.  It
413  * uses the io_filesfrom_buf to read a block of data from f_in (when it is
414  * ready, since it might be a pipe) and then blast it out f_out (when it
415  * is ready to receive more data).
416  */
417 void io_set_filesfrom_fds(int f_in, int f_out)
418 {
419         io_filesfrom_f_in = f_in;
420         io_filesfrom_f_out = f_out;
421         io_filesfrom_bp = io_filesfrom_buf;
422         io_filesfrom_lastchar = '\0';
423         io_filesfrom_buflen = 0;
424 }
425
426 /* It's almost always an error to get an EOF when we're trying to read from the
427  * network, because the protocol is (for the most part) self-terminating.
428  *
429  * There is one case for the receiver when it is at the end of the transfer
430  * (hanging around reading any keep-alive packets that might come its way): if
431  * the sender dies before the generator's kill-signal comes through, we can end
432  * up here needing to loop until the kill-signal arrives.  In this situation,
433  * kluge_around_eof will be < 0.
434  *
435  * There is another case for older protocol versions (< 24) where the module
436  * listing was not terminated, so we must ignore an EOF error in that case and
437  * exit.  In this situation, kluge_around_eof will be > 0. */
438 static void whine_about_eof(int fd)
439 {
440         if (kluge_around_eof && fd == sock_f_in) {
441                 int i;
442                 if (kluge_around_eof > 0)
443                         exit_cleanup(0);
444                 /* If we're still here after 10 seconds, exit with an error. */
445                 for (i = 10*1000/20; i--; )
446                         msleep(20);
447         }
448
449         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
450                 "(%.0f bytes received so far) [%s]\n",
451                 (double)stats.total_read, who_am_i());
452
453         exit_cleanup(RERR_STREAMIO);
454 }
455
456 /**
457  * Read from a socket with I/O timeout. return the number of bytes
458  * read. If no bytes can be read then exit, never return a number <= 0.
459  *
460  * TODO: If the remote shell connection fails, then current versions
461  * actually report an "unexpected EOF" error here.  Since it's a
462  * fairly common mistake to try to use rsh when ssh is required, we
463  * should trap that: if we fail to read any data at all, we should
464  * give a better explanation.  We can tell whether the connection has
465  * started by looking e.g. at whether the remote version is known yet.
466  */
467 static int read_timeout(int fd, char *buf, size_t len)
468 {
469         int n, cnt = 0;
470
471         io_flush(NORMAL_FLUSH);
472
473         while (cnt == 0) {
474                 /* until we manage to read *something* */
475                 fd_set r_fds, w_fds;
476                 struct timeval tv;
477                 int maxfd = fd;
478                 int count;
479
480                 FD_ZERO(&r_fds);
481                 FD_ZERO(&w_fds);
482                 FD_SET(fd, &r_fds);
483                 if (msg2genr.head) {
484                         FD_SET(msg_fd_out, &w_fds);
485                         if (msg_fd_out > maxfd)
486                                 maxfd = msg_fd_out;
487                 }
488                 if (io_filesfrom_f_out >= 0) {
489                         int new_fd;
490                         if (io_filesfrom_buflen == 0) {
491                                 if (io_filesfrom_f_in >= 0) {
492                                         FD_SET(io_filesfrom_f_in, &r_fds);
493                                         new_fd = io_filesfrom_f_in;
494                                 } else {
495                                         io_filesfrom_f_out = -1;
496                                         new_fd = -1;
497                                 }
498                         } else {
499                                 FD_SET(io_filesfrom_f_out, &w_fds);
500                                 new_fd = io_filesfrom_f_out;
501                         }
502                         if (new_fd > maxfd)
503                                 maxfd = new_fd;
504                 }
505
506                 tv.tv_sec = select_timeout;
507                 tv.tv_usec = 0;
508
509                 errno = 0;
510
511                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
512
513                 if (count <= 0) {
514                         if (errno == EBADF)
515                                 exit_cleanup(RERR_SOCKETIO);
516                         check_timeout();
517                         continue;
518                 }
519
520                 if (msg2genr.head && FD_ISSET(msg_fd_out, &w_fds))
521                         msg2genr_flush(NORMAL_FLUSH);
522
523                 if (io_filesfrom_f_out >= 0) {
524                         if (io_filesfrom_buflen) {
525                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
526                                         int l = write(io_filesfrom_f_out,
527                                                       io_filesfrom_bp,
528                                                       io_filesfrom_buflen);
529                                         if (l > 0) {
530                                                 if (!(io_filesfrom_buflen -= l))
531                                                         io_filesfrom_bp = io_filesfrom_buf;
532                                                 else
533                                                         io_filesfrom_bp += l;
534                                         } else {
535                                                 /* XXX should we complain? */
536                                                 io_filesfrom_f_out = -1;
537                                         }
538                                 }
539                         } else if (io_filesfrom_f_in >= 0) {
540                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
541                                         int l = read(io_filesfrom_f_in,
542                                                      io_filesfrom_buf,
543                                                      sizeof io_filesfrom_buf);
544                                         if (l <= 0) {
545                                                 /* Send end-of-file marker */
546                                                 io_filesfrom_buf[0] = '\0';
547                                                 io_filesfrom_buf[1] = '\0';
548                                                 io_filesfrom_buflen = io_filesfrom_lastchar? 2 : 1;
549                                                 io_filesfrom_f_in = -1;
550                                         } else {
551                                                 if (!eol_nulls) {
552                                                         char *s = io_filesfrom_buf + l;
553                                                         /* Transform CR and/or LF into '\0' */
554                                                         while (s-- > io_filesfrom_buf) {
555                                                                 if (*s == '\n' || *s == '\r')
556                                                                         *s = '\0';
557                                                         }
558                                                 }
559                                                 if (!io_filesfrom_lastchar) {
560                                                         /* Last buf ended with a '\0', so don't
561                                                          * let this buf start with one. */
562                                                         while (l && !*io_filesfrom_bp)
563                                                                 io_filesfrom_bp++, l--;
564                                                 }
565                                                 if (!l)
566                                                         io_filesfrom_bp = io_filesfrom_buf;
567                                                 else {
568                                                         char *f = io_filesfrom_bp;
569                                                         char *t = f;
570                                                         char *eob = f + l;
571                                                         /* Eliminate any multi-'\0' runs. */
572                                                         while (f != eob) {
573                                                                 if (!(*t++ = *f++)) {
574                                                                         while (f != eob && !*f)
575                                                                                 f++, l--;
576                                                                 }
577                                                         }
578                                                         io_filesfrom_lastchar = f[-1];
579                                                 }
580                                                 io_filesfrom_buflen = l;
581                                         }
582                                 }
583                         }
584                 }
585
586                 if (!FD_ISSET(fd, &r_fds))
587                         continue;
588
589                 n = read(fd, buf, len);
590
591                 if (n <= 0) {
592                         if (n == 0)
593                                 whine_about_eof(fd); /* Doesn't return. */
594                         if (errno == EINTR || errno == EWOULDBLOCK
595                             || errno == EAGAIN)
596                                 continue;
597
598                         /* Don't write errors on a dead socket. */
599                         if (fd == sock_f_in) {
600                                 close_multiplexing_out();
601                                 rsyserr(FSOCKERR, errno, "read error");
602                         } else
603                                 rsyserr(FERROR, errno, "read error");
604                         exit_cleanup(RERR_STREAMIO);
605                 }
606
607                 buf += n;
608                 len -= n;
609                 cnt += n;
610
611                 if (fd == sock_f_in && io_timeout)
612                         last_io_in = time(NULL);
613         }
614
615         return cnt;
616 }
617
618 /**
619  * Read a line into the "fname" buffer (which must be at least MAXPATHLEN
620  * characters long).
621  */
622 int read_filesfrom_line(int fd, char *fname)
623 {
624         char ch, *s, *eob = fname + MAXPATHLEN - 1;
625         int cnt;
626         int reading_remotely = filesfrom_host != NULL;
627         int nulls = eol_nulls || reading_remotely;
628
629   start:
630         s = fname;
631         while (1) {
632                 cnt = read(fd, &ch, 1);
633                 if (cnt < 0 && (errno == EWOULDBLOCK
634                   || errno == EINTR || errno == EAGAIN)) {
635                         struct timeval tv;
636                         fd_set fds;
637                         FD_ZERO(&fds);
638                         FD_SET(fd, &fds);
639                         tv.tv_sec = select_timeout;
640                         tv.tv_usec = 0;
641                         if (!select(fd+1, &fds, NULL, NULL, &tv))
642                                 check_timeout();
643                         continue;
644                 }
645                 if (cnt != 1)
646                         break;
647                 if (nulls? !ch : (ch == '\r' || ch == '\n')) {
648                         /* Skip empty lines if reading locally. */
649                         if (!reading_remotely && s == fname)
650                                 continue;
651                         break;
652                 }
653                 if (s < eob)
654                         *s++ = ch;
655         }
656         *s = '\0';
657
658         /* Dump comments. */
659         if (*fname == '#' || *fname == ';')
660                 goto start;
661
662         return s - fname;
663 }
664
665 static char *iobuf_out;
666 static int iobuf_out_cnt;
667
668 void io_start_buffering_out(void)
669 {
670         if (iobuf_out)
671                 return;
672         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
673                 out_of_memory("io_start_buffering_out");
674         iobuf_out_cnt = 0;
675 }
676
677 static char *iobuf_in;
678 static size_t iobuf_in_siz;
679
680 void io_start_buffering_in(void)
681 {
682         if (iobuf_in)
683                 return;
684         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
685         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
686                 out_of_memory("io_start_buffering_in");
687 }
688
689 void io_end_buffering(void)
690 {
691         io_flush(NORMAL_FLUSH);
692         if (!io_multiplexing_out) {
693                 free(iobuf_out);
694                 iobuf_out = NULL;
695         }
696 }
697
698 void maybe_flush_socket(void)
699 {
700         if (iobuf_out && iobuf_out_cnt && time(NULL) - last_io_out >= 5)
701                 io_flush(NORMAL_FLUSH);
702 }
703
704 void maybe_send_keepalive(void)
705 {
706         if (time(NULL) - last_io_out >= allowed_lull) {
707                 if (!iobuf_out || !iobuf_out_cnt) {
708                         if (protocol_version < 29)
709                                 return; /* there's nothing we can do */
710                         write_int(sock_f_out, the_file_list->count);
711                         write_shortint(sock_f_out, ITEM_IS_NEW);
712                 }
713                 if (iobuf_out)
714                         io_flush(NORMAL_FLUSH);
715         }
716 }
717
718 /**
719  * Continue trying to read len bytes - don't return until len has been
720  * read.
721  **/
722 static void read_loop(int fd, char *buf, size_t len)
723 {
724         while (len) {
725                 int n = read_timeout(fd, buf, len);
726
727                 buf += n;
728                 len -= n;
729         }
730 }
731
732 /**
733  * Read from the file descriptor handling multiplexing - return number
734  * of bytes read.
735  *
736  * Never returns <= 0.
737  */
738 static int readfd_unbuffered(int fd, char *buf, size_t len)
739 {
740         static size_t remaining;
741         static size_t iobuf_in_ndx;
742         size_t msg_bytes;
743         int tag, cnt = 0;
744         char line[BIGPATHBUFLEN];
745
746         if (!iobuf_in || fd != sock_f_in)
747                 return read_timeout(fd, buf, len);
748
749         if (!io_multiplexing_in && remaining == 0) {
750                 remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
751                 iobuf_in_ndx = 0;
752         }
753
754         while (cnt == 0) {
755                 if (remaining) {
756                         len = MIN(len, remaining);
757                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
758                         iobuf_in_ndx += len;
759                         remaining -= len;
760                         cnt = len;
761                         break;
762                 }
763
764                 read_loop(fd, line, 4);
765                 tag = IVAL(line, 0);
766
767                 msg_bytes = tag & 0xFFFFFF;
768                 tag = (tag >> 24) - MPLEX_BASE;
769
770                 switch (tag) {
771                 case MSG_DATA:
772                         if (msg_bytes > iobuf_in_siz) {
773                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
774                                                                msg_bytes)))
775                                         out_of_memory("readfd_unbuffered");
776                                 iobuf_in_siz = msg_bytes;
777                         }
778                         read_loop(fd, iobuf_in, msg_bytes);
779                         remaining = msg_bytes;
780                         iobuf_in_ndx = 0;
781                         break;
782                 case MSG_DELETED:
783                         if (msg_bytes >= sizeof line)
784                                 goto overflow;
785                         read_loop(fd, line, msg_bytes);
786                         /* A directory name was sent with the trailing null */
787                         if (msg_bytes > 0 && !line[msg_bytes-1])
788                                 log_delete(line, S_IFDIR);
789                         else {
790                                 line[msg_bytes] = '\0';
791                                 log_delete(line, S_IFREG);
792                         }
793                         break;
794                 case MSG_SUCCESS:
795                         if (msg_bytes != 4) {
796                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
797                                         tag, (long)msg_bytes, who_am_i());
798                                 exit_cleanup(RERR_STREAMIO);
799                         }
800                         read_loop(fd, line, msg_bytes);
801                         successful_send(IVAL(line, 0));
802                         break;
803                 case MSG_INFO:
804                 case MSG_ERROR:
805                         if (msg_bytes >= sizeof line) {
806                             overflow:
807                                 rprintf(FERROR,
808                                         "multiplexing overflow %d:%ld [%s]\n",
809                                         tag, (long)msg_bytes, who_am_i());
810                                 exit_cleanup(RERR_STREAMIO);
811                         }
812                         read_loop(fd, line, msg_bytes);
813                         rwrite((enum logcode)tag, line, msg_bytes);
814                         break;
815                 default:
816                         rprintf(FERROR, "unexpected tag %d [%s]\n",
817                                 tag, who_am_i());
818                         exit_cleanup(RERR_STREAMIO);
819                 }
820         }
821
822         if (remaining == 0)
823                 io_flush(NORMAL_FLUSH);
824
825         return cnt;
826 }
827
828 /**
829  * Do a buffered read from @p fd.  Don't return until all @p n bytes
830  * have been read.  If all @p n can't be read then exit with an
831  * error.
832  **/
833 static void readfd(int fd, char *buffer, size_t N)
834 {
835         int  cnt;
836         size_t total = 0;
837
838         while (total < N) {
839                 cnt = readfd_unbuffered(fd, buffer + total, N-total);
840                 total += cnt;
841         }
842
843         if (fd == write_batch_monitor_in) {
844                 if ((size_t)write(batch_fd, buffer, total) != total)
845                         exit_cleanup(RERR_FILEIO);
846         }
847
848         if (fd == sock_f_in)
849                 stats.total_read += total;
850 }
851
852 int read_shortint(int f)
853 {
854         uchar b[2];
855         readfd(f, (char *)b, 2);
856         return (b[1] << 8) + b[0];
857 }
858
859 int32 read_int(int f)
860 {
861         char b[4];
862         int32 num;
863
864         readfd(f,b,4);
865         num = IVAL(b,0);
866         if (num == (int32)0xffffffff)
867                 return -1;
868         return num;
869 }
870
871 int64 read_longint(int f)
872 {
873         int64 num;
874         char b[8];
875         num = read_int(f);
876
877         if ((int32)num != (int32)0xffffffff)
878                 return num;
879
880 #if SIZEOF_INT64 < 8
881         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
882         exit_cleanup(RERR_UNSUPPORTED);
883 #else
884         readfd(f,b,8);
885         num = IVAL(b,0) | (((int64)IVAL(b,4))<<32);
886 #endif
887
888         return num;
889 }
890
891 void read_buf(int f,char *buf,size_t len)
892 {
893         readfd(f,buf,len);
894 }
895
896 void read_sbuf(int f,char *buf,size_t len)
897 {
898         readfd(f, buf, len);
899         buf[len] = '\0';
900 }
901
902 uchar read_byte(int f)
903 {
904         uchar c;
905         readfd(f, (char *)&c, 1);
906         return c;
907 }
908
909 int read_vstring(int f, char *buf, int bufsize)
910 {
911         int len = read_byte(f);
912
913         if (len & 0x80)
914                 len = (len & ~0x80) * 0x100 + read_byte(f);
915
916         if (len >= bufsize) {
917                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
918                         len, bufsize - 1);
919                 return -1;
920         }
921
922         if (len)
923                 readfd(f, buf, len);
924         buf[len] = '\0';
925         return len;
926 }
927
928 /* Populate a sum_struct with values from the socket.  This is
929  * called by both the sender and the receiver. */
930 void read_sum_head(int f, struct sum_struct *sum)
931 {
932         sum->count = read_int(f);
933         if (sum->count < 0) {
934                 rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
935                         (long)sum->count, who_am_i());
936                 exit_cleanup(RERR_PROTOCOL);
937         }
938         sum->blength = read_int(f);
939         if (sum->blength < 0 || sum->blength > MAX_BLOCK_SIZE) {
940                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
941                         (long)sum->blength, who_am_i());
942                 exit_cleanup(RERR_PROTOCOL);
943         }
944         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
945         if (sum->s2length < 0 || sum->s2length > MD4_SUM_LENGTH) {
946                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
947                         sum->s2length, who_am_i());
948                 exit_cleanup(RERR_PROTOCOL);
949         }
950         sum->remainder = read_int(f);
951         if (sum->remainder < 0 || sum->remainder > sum->blength) {
952                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
953                         (long)sum->remainder, who_am_i());
954                 exit_cleanup(RERR_PROTOCOL);
955         }
956 }
957
958 /* Send the values from a sum_struct over the socket.  Set sum to
959  * NULL if there are no checksums to send.  This is called by both
960  * the generator and the sender. */
961 void write_sum_head(int f, struct sum_struct *sum)
962 {
963         static struct sum_struct null_sum;
964
965         if (sum == NULL)
966                 sum = &null_sum;
967
968         write_int(f, sum->count);
969         write_int(f, sum->blength);
970         if (protocol_version >= 27)
971                 write_int(f, sum->s2length);
972         write_int(f, sum->remainder);
973 }
974
975 /**
976  * Sleep after writing to limit I/O bandwidth usage.
977  *
978  * @todo Rather than sleeping after each write, it might be better to
979  * use some kind of averaging.  The current algorithm seems to always
980  * use a bit less bandwidth than specified, because it doesn't make up
981  * for slow periods.  But arguably this is a feature.  In addition, we
982  * ought to take the time used to write the data into account.
983  *
984  * During some phases of big transfers (file FOO is uptodate) this is
985  * called with a small bytes_written every time.  As the kernel has to
986  * round small waits up to guarantee that we actually wait at least the
987  * requested number of microseconds, this can become grossly inaccurate.
988  * We therefore keep track of the bytes we've written over time and only
989  * sleep when the accumulated delay is at least 1 tenth of a second.
990  **/
991 static void sleep_for_bwlimit(int bytes_written)
992 {
993         static struct timeval prior_tv;
994         static long total_written = 0;
995         struct timeval tv, start_tv;
996         long elapsed_usec, sleep_usec;
997
998 #define ONE_SEC 1000000L /* # of microseconds in a second */
999
1000         if (!bwlimit_writemax)
1001                 return;
1002
1003         total_written += bytes_written;
1004
1005         gettimeofday(&start_tv, NULL);
1006         if (prior_tv.tv_sec) {
1007                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
1008                              + (start_tv.tv_usec - prior_tv.tv_usec);
1009                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
1010                 if (total_written < 0)
1011                         total_written = 0;
1012         }
1013
1014         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
1015         if (sleep_usec < ONE_SEC / 10) {
1016                 prior_tv = start_tv;
1017                 return;
1018         }
1019
1020         tv.tv_sec  = sleep_usec / ONE_SEC;
1021         tv.tv_usec = sleep_usec % ONE_SEC;
1022         select(0, NULL, NULL, NULL, &tv);
1023
1024         gettimeofday(&prior_tv, NULL);
1025         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
1026                      + (prior_tv.tv_usec - start_tv.tv_usec);
1027         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
1028 }
1029
1030 /* Write len bytes to the file descriptor fd, looping as necessary to get
1031  * the job done and also (in certain circumstances) reading any data on
1032  * msg_fd_in to avoid deadlock.
1033  *
1034  * This function underlies the multiplexing system.  The body of the
1035  * application never calls this function directly. */
1036 static void writefd_unbuffered(int fd,char *buf,size_t len)
1037 {
1038         size_t n, total = 0;
1039         fd_set w_fds, r_fds;
1040         int maxfd, count, cnt, using_r_fds;
1041         int defer_save = defer_forwarding_messages;
1042         struct timeval tv;
1043
1044         no_flush++;
1045
1046         while (total < len) {
1047                 FD_ZERO(&w_fds);
1048                 FD_SET(fd,&w_fds);
1049                 maxfd = fd;
1050
1051                 if (msg_fd_in >= 0) {
1052                         FD_ZERO(&r_fds);
1053                         FD_SET(msg_fd_in,&r_fds);
1054                         if (msg_fd_in > maxfd)
1055                                 maxfd = msg_fd_in;
1056                         using_r_fds = 1;
1057                 } else
1058                         using_r_fds = 0;
1059
1060                 tv.tv_sec = select_timeout;
1061                 tv.tv_usec = 0;
1062
1063                 errno = 0;
1064                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1065                                &w_fds, NULL, &tv);
1066
1067                 if (count <= 0) {
1068                         if (count < 0 && errno == EBADF)
1069                                 exit_cleanup(RERR_SOCKETIO);
1070                         check_timeout();
1071                         continue;
1072                 }
1073
1074                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1075                         read_msg_fd();
1076
1077                 if (!FD_ISSET(fd, &w_fds))
1078                         continue;
1079
1080                 n = len - total;
1081                 if (bwlimit_writemax && n > bwlimit_writemax)
1082                         n = bwlimit_writemax;
1083                 cnt = write(fd, buf + total, n);
1084
1085                 if (cnt <= 0) {
1086                         if (cnt < 0) {
1087                                 if (errno == EINTR)
1088                                         continue;
1089                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1090                                         msleep(1);
1091                                         continue;
1092                                 }
1093                         }
1094
1095                         /* Don't try to write errors back across the stream. */
1096                         if (fd == sock_f_out)
1097                                 close_multiplexing_out();
1098                         rsyserr(FERROR, errno,
1099                                 "writefd_unbuffered failed to write %ld bytes [%s]",
1100                                 (long)len, who_am_i());
1101                         /* If the other side is sending us error messages, try
1102                          * to grab any messages they sent before they died. */
1103                         while (fd == sock_f_out && io_multiplexing_in) {
1104                                 set_io_timeout(30);
1105                                 ignore_timeout = 0;
1106                                 readfd_unbuffered(sock_f_in, io_filesfrom_buf,
1107                                                   sizeof io_filesfrom_buf);
1108                         }
1109                         exit_cleanup(RERR_STREAMIO);
1110                 }
1111
1112                 total += cnt;
1113                 defer_forwarding_messages = 1;
1114
1115                 if (fd == sock_f_out) {
1116                         if (io_timeout || am_generator)
1117                                 last_io_out = time(NULL);
1118                         sleep_for_bwlimit(cnt);
1119                 }
1120         }
1121
1122         defer_forwarding_messages = defer_save;
1123         no_flush--;
1124 }
1125
1126 static void msg2sndr_flush(void)
1127 {
1128         if (defer_forwarding_messages)
1129                 return;
1130
1131         while (msg2sndr.head && io_multiplexing_out) {
1132                 struct msg_list_item *m = msg2sndr.head;
1133                 if (!(msg2sndr.head = m->next))
1134                         msg2sndr.tail = NULL;
1135                 stats.total_written += m->len;
1136                 defer_forwarding_messages = 1;
1137                 writefd_unbuffered(sock_f_out, m->buf, m->len);
1138                 defer_forwarding_messages = 0;
1139                 free(m);
1140         }
1141 }
1142
1143 /**
1144  * Write an message to a multiplexed stream. If this fails then rsync
1145  * exits.
1146  **/
1147 static void mplex_write(enum msgcode code, char *buf, size_t len)
1148 {
1149         char buffer[1024];
1150         size_t n = len;
1151
1152         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
1153
1154         if (n > sizeof buffer - 4)
1155                 n = 0;
1156         else
1157                 memcpy(buffer + 4, buf, n);
1158
1159         writefd_unbuffered(sock_f_out, buffer, n+4);
1160
1161         len -= n;
1162         buf += n;
1163
1164         if (len) {
1165                 defer_forwarding_messages = 1;
1166                 writefd_unbuffered(sock_f_out, buf, len);
1167                 defer_forwarding_messages = 0;
1168                 msg2sndr_flush();
1169         }
1170 }
1171
1172 void io_flush(int flush_it_all)
1173 {
1174         msg2genr_flush(flush_it_all);
1175         msg2sndr_flush();
1176
1177         if (!iobuf_out_cnt || no_flush)
1178                 return;
1179
1180         if (io_multiplexing_out)
1181                 mplex_write(MSG_DATA, iobuf_out, iobuf_out_cnt);
1182         else
1183                 writefd_unbuffered(sock_f_out, iobuf_out, iobuf_out_cnt);
1184         iobuf_out_cnt = 0;
1185 }
1186
1187 static void writefd(int fd,char *buf,size_t len)
1188 {
1189         if (fd == msg_fd_out) {
1190                 rprintf(FERROR, "Internal error: wrong write used in receiver.\n");
1191                 exit_cleanup(RERR_PROTOCOL);
1192         }
1193
1194         if (fd == sock_f_out)
1195                 stats.total_written += len;
1196
1197         if (fd == write_batch_monitor_out) {
1198                 if ((size_t)write(batch_fd, buf, len) != len)
1199                         exit_cleanup(RERR_FILEIO);
1200         }
1201
1202         if (!iobuf_out || fd != sock_f_out) {
1203                 writefd_unbuffered(fd, buf, len);
1204                 return;
1205         }
1206
1207         while (len) {
1208                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1209                 if (n > 0) {
1210                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1211                         buf += n;
1212                         len -= n;
1213                         iobuf_out_cnt += n;
1214                 }
1215
1216                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1217                         io_flush(NORMAL_FLUSH);
1218         }
1219 }
1220
1221 void write_shortint(int f, int x)
1222 {
1223         uchar b[2];
1224         b[0] = x;
1225         b[1] = x >> 8;
1226         writefd(f, (char *)b, 2);
1227 }
1228
1229 void write_int(int f,int32 x)
1230 {
1231         char b[4];
1232         SIVAL(b,0,x);
1233         writefd(f,b,4);
1234 }
1235
1236 /*
1237  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1238  * 64-bit types on this platform.
1239  */
1240 void write_longint(int f, int64 x)
1241 {
1242         char b[8];
1243
1244         if (x <= 0x7FFFFFFF) {
1245                 write_int(f, (int)x);
1246                 return;
1247         }
1248
1249 #if SIZEOF_INT64 < 8
1250         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1251         exit_cleanup(RERR_UNSUPPORTED);
1252 #else
1253         write_int(f, (int32)0xFFFFFFFF);
1254         SIVAL(b,0,(x&0xFFFFFFFF));
1255         SIVAL(b,4,((x>>32)&0xFFFFFFFF));
1256
1257         writefd(f,b,8);
1258 #endif
1259 }
1260
1261 void write_buf(int f,char *buf,size_t len)
1262 {
1263         writefd(f,buf,len);
1264 }
1265
1266 /** Write a string to the connection */
1267 void write_sbuf(int f, char *buf)
1268 {
1269         writefd(f, buf, strlen(buf));
1270 }
1271
1272 void write_byte(int f, uchar c)
1273 {
1274         writefd(f, (char *)&c, 1);
1275 }
1276
1277 void write_vstring(int f, char *str, int len)
1278 {
1279         uchar lenbuf[3], *lb = lenbuf;
1280
1281         if (len > 0x7F) {
1282                 if (len > 0x7FFF) {
1283                         rprintf(FERROR,
1284                                 "attempting to send over-long vstring (%d > %d)\n",
1285                                 len, 0x7FFF);
1286                         exit_cleanup(RERR_PROTOCOL);
1287                 }
1288                 *lb++ = len / 0x100 + 0x80;
1289         }
1290         *lb = len;
1291
1292         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1293         if (len)
1294                 writefd(f, str, len);
1295 }
1296
1297 /**
1298  * Read a line of up to @p maxlen characters into @p buf (not counting
1299  * the trailing null).  Strips the (required) trailing newline and all
1300  * carriage returns.
1301  *
1302  * @return 1 for success; 0 for I/O error or truncation.
1303  **/
1304 int read_line(int f, char *buf, size_t maxlen)
1305 {
1306         while (maxlen) {
1307                 buf[0] = 0;
1308                 read_buf(f, buf, 1);
1309                 if (buf[0] == 0)
1310                         return 0;
1311                 if (buf[0] == '\n')
1312                         break;
1313                 if (buf[0] != '\r') {
1314                         buf++;
1315                         maxlen--;
1316                 }
1317         }
1318         *buf = '\0';
1319         return maxlen > 0;
1320 }
1321
1322 void io_printf(int fd, const char *format, ...)
1323 {
1324         va_list ap;
1325         char buf[BIGPATHBUFLEN];
1326         int len;
1327
1328         va_start(ap, format);
1329         len = vsnprintf(buf, sizeof buf, format, ap);
1330         va_end(ap);
1331
1332         if (len < 0)
1333                 exit_cleanup(RERR_STREAMIO);
1334
1335         if (len > (int)sizeof buf) {
1336                 rprintf(FERROR, "io_printf() was too long for the buffer.\n");
1337                 exit_cleanup(RERR_STREAMIO);
1338         }
1339
1340         write_sbuf(fd, buf);
1341 }
1342
1343 /** Setup for multiplexing a MSG_* stream with the data stream. */
1344 void io_start_multiplex_out(void)
1345 {
1346         io_flush(NORMAL_FLUSH);
1347         io_start_buffering_out();
1348         io_multiplexing_out = 1;
1349 }
1350
1351 /** Setup for multiplexing a MSG_* stream with the data stream. */
1352 void io_start_multiplex_in(void)
1353 {
1354         io_flush(NORMAL_FLUSH);
1355         io_start_buffering_in();
1356         io_multiplexing_in = 1;
1357 }
1358
1359 /** Write an message to the multiplexed data stream. */
1360 int io_multiplex_write(enum msgcode code, char *buf, size_t len)
1361 {
1362         if (!io_multiplexing_out)
1363                 return 0;
1364
1365         io_flush(NORMAL_FLUSH);
1366         stats.total_written += (len+4);
1367         mplex_write(code, buf, len);
1368         return 1;
1369 }
1370
1371 void close_multiplexing_in(void)
1372 {
1373         io_multiplexing_in = 0;
1374 }
1375
1376 /** Stop output multiplexing. */
1377 void close_multiplexing_out(void)
1378 {
1379         io_multiplexing_out = 0;
1380 }
1381
1382 void start_write_batch(int fd)
1383 {
1384         write_stream_flags(batch_fd);
1385
1386         /* Some communication has already taken place, but we don't
1387          * enable batch writing until here so that we can write a
1388          * canonical record of the communication even though the
1389          * actual communication so far depends on whether a daemon
1390          * is involved. */
1391         write_int(batch_fd, protocol_version);
1392         write_int(batch_fd, checksum_seed);
1393
1394         if (am_sender)
1395                 write_batch_monitor_out = fd;
1396         else
1397                 write_batch_monitor_in = fd;
1398 }
1399
1400 void stop_write_batch(void)
1401 {
1402         write_batch_monitor_out = -1;
1403         write_batch_monitor_in = -1;
1404 }