Output a -vvv(erbose) message when receiving an incremental file list.
[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) 2003-2007 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 version 2 as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License along
19  * with this program; if not, write to the Free Software Foundation, Inc.,
20  * 51 Franklin Street - Fifth Floor, Boston, MA 02110-1301, USA.
21  */
22
23 /* Rsync provides its own multiplexing system, which is used to send
24  * stderr and stdout over a single socket.
25  *
26  * For historical reasons this is off during the start of the
27  * connection, but it's switched on quite early using
28  * io_start_multiplex_out() and io_start_multiplex_in(). */
29
30 #include "rsync.h"
31
32 /** If no timeout is specified then use a 60 second select timeout */
33 #define SELECT_TIMEOUT 60
34
35 extern int bwlimit;
36 extern size_t bwlimit_writemax;
37 extern int io_timeout;
38 extern int allowed_lull;
39 extern int am_server;
40 extern int am_daemon;
41 extern int am_sender;
42 extern int am_generator;
43 extern int inc_recurse;
44 extern int io_error;
45 extern int eol_nulls;
46 extern int flist_eof;
47 extern int read_batch;
48 extern int csum_length;
49 extern int checksum_seed;
50 extern int protocol_version;
51 extern int remove_source_files;
52 extern int preserve_hard_links;
53 extern char *filesfrom_host;
54 extern struct stats stats;
55 extern struct file_list *cur_flist, *first_flist;
56 #ifdef ICONV_OPTION
57 extern iconv_t ic_send, ic_recv;
58 #endif
59
60 const char phase_unknown[] = "unknown";
61 int ignore_timeout = 0;
62 int batch_fd = -1;
63 int msgdone_cnt = 0;
64
65 /* Ignore an EOF error if non-zero. See whine_about_eof(). */
66 int kluge_around_eof = 0;
67
68 int msg_fd_in = -1;
69 int msg_fd_out = -1;
70 int sock_f_in = -1;
71 int sock_f_out = -1;
72
73 static int iobuf_f_in = -1;
74 static char *iobuf_in;
75 static size_t iobuf_in_siz;
76 static size_t iobuf_in_ndx;
77 static size_t iobuf_in_remaining;
78
79 static int iobuf_f_out = -1;
80 static char *iobuf_out;
81 static int iobuf_out_cnt;
82
83 int flist_forward_from = -1;
84
85 static int io_multiplexing_out;
86 static int io_multiplexing_in;
87 static time_t last_io_in;
88 static time_t last_io_out;
89 static int no_flush;
90
91 static int write_batch_monitor_in = -1;
92 static int write_batch_monitor_out = -1;
93
94 static int io_filesfrom_f_in = -1;
95 static int io_filesfrom_f_out = -1;
96 static char io_filesfrom_buf[2048];
97 #ifdef ICONV_OPTION
98 static char iconv_buf[sizeof io_filesfrom_buf / 2];
99 #endif
100 static char *io_filesfrom_bp;
101 static char io_filesfrom_lastchar;
102 static int io_filesfrom_buflen;
103 static int defer_forwarding_messages = 0;
104 static int select_timeout = SELECT_TIMEOUT;
105 static int active_filecnt = 0;
106 static OFF_T active_bytecnt = 0;
107
108 static char int_byte_extra[64] = {
109         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* (00 - 3F)/4 */
110         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* (40 - 7F)/4 */
111         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* (80 - BF)/4 */
112         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 5, 6, /* (C0 - FF)/4 */
113 };
114
115 static void readfd(int fd, char *buffer, size_t N);
116 static void writefd(int fd, const char *buf, size_t len);
117 static void writefd_unbuffered(int fd, const char *buf, size_t len);
118 static void decrement_active_files(int ndx);
119 static void decrement_flist_in_progress(int ndx, int redo);
120 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert);
121
122 struct flist_ndx_item {
123         struct flist_ndx_item *next;
124         int ndx;
125 };
126
127 struct flist_ndx_list {
128         struct flist_ndx_item *head, *tail;
129 };
130
131 static struct flist_ndx_list redo_list, hlink_list;
132
133 struct msg_list_item {
134         struct msg_list_item *next;
135         char convert;
136         char buf[1];
137 };
138
139 struct msg_list {
140         struct msg_list_item *head, *tail;
141 };
142
143 static struct msg_list msg_queue;
144
145 static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
146 {
147         struct flist_ndx_item *item;
148
149         if (!(item = new(struct flist_ndx_item)))
150                 out_of_memory("flist_ndx_push");
151         item->next = NULL;
152         item->ndx = ndx;
153         if (lp->tail)
154                 lp->tail->next = item;
155         else
156                 lp->head = item;
157         lp->tail = item;
158 }
159
160 static int flist_ndx_pop(struct flist_ndx_list *lp)
161 {
162         struct flist_ndx_item *next;
163         int ndx;
164
165         if (!lp->head)
166                 return -1;
167
168         ndx = lp->head->ndx;
169         next = lp->head->next;
170         free(lp->head);
171         lp->head = next;
172         if (!next)
173                 lp->tail = NULL;
174
175         return ndx;
176 }
177
178 static void check_timeout(void)
179 {
180         time_t t;
181
182         if (!io_timeout || ignore_timeout)
183                 return;
184
185         if (!last_io_in) {
186                 last_io_in = time(NULL);
187                 return;
188         }
189
190         t = time(NULL);
191
192         if (t - last_io_in >= io_timeout) {
193                 if (!am_server && !am_daemon) {
194                         rprintf(FERROR, "io timeout after %d seconds -- exiting\n",
195                                 (int)(t-last_io_in));
196                 }
197                 exit_cleanup(RERR_TIMEOUT);
198         }
199 }
200
201 /* Note the fds used for the main socket (which might really be a pipe
202  * for a local transfer, but we can ignore that). */
203 void io_set_sock_fds(int f_in, int f_out)
204 {
205         sock_f_in = f_in;
206         sock_f_out = f_out;
207 }
208
209 void set_io_timeout(int secs)
210 {
211         io_timeout = secs;
212
213         if (!io_timeout || io_timeout > SELECT_TIMEOUT)
214                 select_timeout = SELECT_TIMEOUT;
215         else
216                 select_timeout = io_timeout;
217
218         allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2;
219 }
220
221 /* Setup the fd used to receive MSG_* messages.  Only needed during the
222  * early stages of being a local sender (up through the sending of the
223  * file list) or when we're the generator (to fetch the messages from
224  * the receiver). */
225 void set_msg_fd_in(int fd)
226 {
227         msg_fd_in = fd;
228 }
229
230 /* Setup the fd used to send our MSG_* messages.  Only needed when
231  * we're the receiver (to send our messages to the generator). */
232 void set_msg_fd_out(int fd)
233 {
234         msg_fd_out = fd;
235         set_nonblocking(msg_fd_out);
236 }
237
238 /* Add a message to the pending MSG_* list. */
239 static void msg_list_add(struct msg_list *lst, int code, const char *buf, int len, int convert)
240 {
241         struct msg_list_item *m;
242         int sz = len + 4 + sizeof m[0] - 1;
243
244         if (!(m = (struct msg_list_item *)new_array(char, sz)))
245                 out_of_memory("msg_list_add");
246         m->next = NULL;
247         m->convert = convert;
248         SIVAL(m->buf, 0, ((code+MPLEX_BASE)<<24) | len);
249         memcpy(m->buf + 4, buf, len);
250         if (lst->tail)
251                 lst->tail->next = m;
252         else
253                 lst->head = m;
254         lst->tail = m;
255 }
256
257 static void msg_flush(void)
258 {
259         if (am_generator) {
260                 while (msg_queue.head && io_multiplexing_out) {
261                         struct msg_list_item *m = msg_queue.head;
262                         int len = IVAL(m->buf, 0) & 0xFFFFFF;
263                         int tag = *((uchar*)m->buf+3) - MPLEX_BASE;
264                         if (!(msg_queue.head = m->next))
265                                 msg_queue.tail = NULL;
266                         stats.total_written += len + 4;
267                         defer_forwarding_messages++;
268                         mplex_write(sock_f_out, tag, m->buf + 4, len, m->convert);
269                         defer_forwarding_messages--;
270                         free(m);
271                 }
272         } else {
273                 while (msg_queue.head) {
274                         struct msg_list_item *m = msg_queue.head;
275                         int len = IVAL(m->buf, 0) & 0xFFFFFF;
276                         int tag = *((uchar*)m->buf+3) - MPLEX_BASE;
277                         if (!(msg_queue.head = m->next))
278                                 msg_queue.tail = NULL;
279                         defer_forwarding_messages++;
280                         mplex_write(msg_fd_out, tag, m->buf + 4, len, m->convert);
281                         defer_forwarding_messages--;
282                         free(m);
283                 }
284         }
285 }
286
287 /* Read a message from the MSG_* fd and handle it.  This is called either
288  * during the early stages of being a local sender (up through the sending
289  * of the file list) or when we're the generator (to fetch the messages
290  * from the receiver). */
291 static void read_msg_fd(void)
292 {
293         char buf[2048];
294         size_t n;
295         struct file_list *flist;
296         int fd = msg_fd_in;
297         int tag, len;
298
299         /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
300          * to this routine from writefd_unbuffered(). */
301         no_flush++;
302         msg_fd_in = -1;
303         defer_forwarding_messages++;
304
305         readfd(fd, buf, 4);
306         tag = IVAL(buf, 0);
307
308         len = tag & 0xFFFFFF;
309         tag = (tag >> 24) - MPLEX_BASE;
310
311         switch (tag) {
312         case MSG_DONE:
313                 if (len < 0 || len > 1 || !am_generator) {
314                   invalid_msg:
315                         rprintf(FERROR, "invalid message %d:%d [%s%s]\n",
316                                 tag, len, who_am_i(),
317                                 inc_recurse ? "/inc" : "");
318                         exit_cleanup(RERR_STREAMIO);
319                 }
320                 if (len) {
321                         readfd(fd, buf, len);
322                         stats.total_read = read_varlong(fd, 3);
323                 }
324                 msgdone_cnt++;
325                 break;
326         case MSG_REDO:
327                 if (len != 4 || !am_generator)
328                         goto invalid_msg;
329                 readfd(fd, buf, 4);
330                 if (remove_source_files)
331                         decrement_active_files(IVAL(buf,0));
332                 flist_ndx_push(&redo_list, IVAL(buf,0));
333                 if (inc_recurse)
334                         decrement_flist_in_progress(IVAL(buf,0), 1);
335                 break;
336         case MSG_FLIST:
337                 if (len != 4 || !am_generator || !inc_recurse)
338                         goto invalid_msg;
339                 readfd(fd, buf, 4);
340                 /* Read extra file list from receiver. */
341                 assert(iobuf_in != NULL);
342                 assert(iobuf_f_in == fd);
343                 if (verbose > 3) {
344                         rprintf(FINFO, "[%s] receiving flist for dir %d\n",
345                                 who_am_i(), IVAL(buf,0));
346                 }
347                 flist = recv_file_list(fd);
348                 flist->parent_ndx = IVAL(buf,0);
349                 break;
350         case MSG_FLIST_EOF:
351                 if (len != 0 || !am_generator || !inc_recurse)
352                         goto invalid_msg;
353                 flist_eof = 1;
354                 break;
355         case MSG_DELETED:
356                 if (len >= (int)sizeof buf || !am_generator)
357                         goto invalid_msg;
358                 readfd(fd, buf, len);
359                 send_msg(MSG_DELETED, buf, len, 1);
360                 break;
361         case MSG_SUCCESS:
362                 if (len != 4 || !am_generator)
363                         goto invalid_msg;
364                 readfd(fd, buf, len);
365                 if (remove_source_files) {
366                         decrement_active_files(IVAL(buf,0));
367                         send_msg(MSG_SUCCESS, buf, len, 0);
368                 }
369                 if (preserve_hard_links)
370                         flist_ndx_push(&hlink_list, IVAL(buf,0));
371                 if (inc_recurse)
372                         decrement_flist_in_progress(IVAL(buf,0), 0);
373                 break;
374         case MSG_NO_SEND:
375                 if (len != 4 || !am_generator)
376                         goto invalid_msg;
377                 readfd(fd, buf, len);
378                 if (inc_recurse)
379                         decrement_flist_in_progress(IVAL(buf,0), 0);
380                 break;
381         case MSG_SOCKERR:
382         case MSG_CLIENT:
383                 if (!am_generator)
384                         goto invalid_msg;
385                 if (tag == MSG_SOCKERR)
386                         io_end_multiplex_out();
387                 /* FALL THROUGH */
388         case MSG_INFO:
389         case MSG_ERROR:
390         case MSG_LOG:
391                 while (len) {
392                         n = len;
393                         if (n >= sizeof buf)
394                                 n = sizeof buf - 1;
395                         readfd(fd, buf, n);
396                         rwrite((enum logcode)tag, buf, n, !am_generator);
397                         len -= n;
398                 }
399                 break;
400         default:
401                 rprintf(FERROR, "unknown message %d:%d [%s]\n",
402                         tag, len, who_am_i());
403                 exit_cleanup(RERR_STREAMIO);
404         }
405
406         no_flush--;
407         msg_fd_in = fd;
408         if (!--defer_forwarding_messages)
409                 msg_flush();
410 }
411
412 /* This is used by the generator to limit how many file transfers can
413  * be active at once when --remove-source-files is specified.  Without
414  * this, sender-side deletions were mostly happening at the end. */
415 void increment_active_files(int ndx, int itemizing, enum logcode code)
416 {
417         /* TODO: tune these limits? */
418         while (active_filecnt >= (active_bytecnt >= 128*1024 ? 10 : 50)) {
419                 check_for_finished_files(itemizing, code, 0);
420                 if (iobuf_out_cnt)
421                         io_flush(NORMAL_FLUSH);
422                 else
423                         read_msg_fd();
424         }
425
426         active_filecnt++;
427         active_bytecnt += F_LENGTH(cur_flist->files[ndx - cur_flist->ndx_start]);
428 }
429
430 static void decrement_active_files(int ndx)
431 {
432         struct file_list *flist = flist_for_ndx(ndx);
433         assert(flist != NULL);
434         active_filecnt--;
435         active_bytecnt -= F_LENGTH(flist->files[ndx - flist->ndx_start]);
436 }
437
438 static void decrement_flist_in_progress(int ndx, int redo)
439 {
440         struct file_list *flist = cur_flist ? cur_flist : first_flist;
441
442         while (ndx < flist->ndx_start) {
443                 if (flist == first_flist) {
444                   invalid_ndx:
445                         rprintf(FERROR,
446                                 "Invalid file index: %d (%d - %d) [%s]\n",
447                                 ndx, first_flist->ndx_start,
448                                 first_flist->prev->ndx_start + first_flist->prev->count - 1,
449                                 who_am_i());
450                         exit_cleanup(RERR_PROTOCOL);
451                 }
452                 flist = flist->prev;
453         }
454         while (ndx >= flist->ndx_start + flist->count) {
455                 if (!(flist = flist->next))
456                         goto invalid_ndx;
457         }
458
459         flist->in_progress--;
460         if (redo)
461                 flist->to_redo++;
462 }
463
464 /* Write an message to a multiplexed stream. If this fails, rsync exits. */
465 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert)
466 {
467         char buffer[BIGPATHBUFLEN]; /* Oversized for use by iconv code. */
468         size_t n = len;
469
470         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
471
472 #ifdef ICONV_OPTION
473         if (convert && ic_send == (iconv_t)-1)
474 #endif
475                 convert = 0;
476
477         if (convert || n > 1024 - 4) /* BIGPATHBUFLEN can handle 1024 bytes */
478                 n = 0;
479         else
480                 memcpy(buffer + 4, buf, n);
481
482         writefd_unbuffered(fd, buffer, n+4);
483
484         len -= n;
485         buf += n;
486
487 #ifdef ICONV_OPTION
488         if (convert) {
489                 iconv(ic_send, NULL, 0, NULL, 0);
490                 defer_forwarding_messages++;
491                 while (len) {
492                         ICONV_CONST char *ibuf = (ICONV_CONST char *)buf;
493                         char *obuf = buffer;
494                         size_t ocnt = sizeof buffer;
495                         while (len && iconv(ic_send, &ibuf,&len,
496                                             &obuf,&ocnt) == (size_t)-1) {
497                                 if (errno == E2BIG || !ocnt)
498                                         break;
499                                 *obuf++ = *ibuf++;
500                                 ocnt--, len--;
501                         }
502                         n = obuf - buffer;
503                         writefd_unbuffered(fd, buffer, n);
504                 }
505                 if (!--defer_forwarding_messages)
506                         msg_flush();
507         } else
508 #endif
509         if (len) {
510                 defer_forwarding_messages++;
511                 writefd_unbuffered(fd, buf, len);
512                 if (!--defer_forwarding_messages)
513                         msg_flush();
514         }
515 }
516
517 int send_msg(enum msgcode code, const char *buf, int len, int convert)
518 {
519         if (msg_fd_out < 0) {
520                 if (!defer_forwarding_messages)
521                         return io_multiplex_write(code, buf, len, convert);
522                 if (!io_multiplexing_out)
523                         return 0;
524                 msg_list_add(&msg_queue, code, buf, len, convert);
525                 return 1;
526         }
527         if (flist_forward_from >= 0)
528                 msg_list_add(&msg_queue, code, buf, len, convert);
529         else
530                 mplex_write(msg_fd_out, code, buf, len, convert);
531         return 1;
532 }
533
534 void send_msg_int(enum msgcode code, int num)
535 {
536         char numbuf[4];
537         SIVAL(numbuf, 0, num);
538         send_msg(code, numbuf, 4, 0);
539 }
540
541 void wait_for_receiver(void)
542 {
543         if (iobuf_out_cnt)
544                 io_flush(NORMAL_FLUSH);
545         else
546                 read_msg_fd();
547 }
548
549 int get_redo_num(void)
550 {
551         return flist_ndx_pop(&redo_list);
552 }
553
554 int get_hlink_num(void)
555 {
556         return flist_ndx_pop(&hlink_list);
557 }
558
559 /**
560  * When we're the receiver and we have a local --files-from list of names
561  * that needs to be sent over the socket to the sender, we have to do two
562  * things at the same time: send the sender a list of what files we're
563  * processing and read the incoming file+info list from the sender.  We do
564  * this by augmenting the read_timeout() function to copy this data.  It
565  * uses the io_filesfrom_buf to read a block of data from f_in (when it is
566  * ready, since it might be a pipe) and then blast it out f_out (when it
567  * is ready to receive more data).
568  */
569 void io_set_filesfrom_fds(int f_in, int f_out)
570 {
571         io_filesfrom_f_in = f_in;
572         io_filesfrom_f_out = f_out;
573         io_filesfrom_bp = io_filesfrom_buf;
574         io_filesfrom_lastchar = '\0';
575         io_filesfrom_buflen = 0;
576 }
577
578 /* It's almost always an error to get an EOF when we're trying to read from the
579  * network, because the protocol is (for the most part) self-terminating.
580  *
581  * There is one case for the receiver when it is at the end of the transfer
582  * (hanging around reading any keep-alive packets that might come its way): if
583  * the sender dies before the generator's kill-signal comes through, we can end
584  * up here needing to loop until the kill-signal arrives.  In this situation,
585  * kluge_around_eof will be < 0.
586  *
587  * There is another case for older protocol versions (< 24) where the module
588  * listing was not terminated, so we must ignore an EOF error in that case and
589  * exit.  In this situation, kluge_around_eof will be > 0. */
590 static void whine_about_eof(int fd)
591 {
592         if (kluge_around_eof && fd == sock_f_in) {
593                 int i;
594                 if (kluge_around_eof > 0)
595                         exit_cleanup(0);
596                 /* If we're still here after 10 seconds, exit with an error. */
597                 for (i = 10*1000/20; i--; )
598                         msleep(20);
599         }
600
601         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
602                 "(%.0f bytes received so far) [%s]\n",
603                 (double)stats.total_read, who_am_i());
604
605         exit_cleanup(RERR_STREAMIO);
606 }
607
608 /**
609  * Read from a socket with I/O timeout. return the number of bytes
610  * read. If no bytes can be read then exit, never return a number <= 0.
611  *
612  * TODO: If the remote shell connection fails, then current versions
613  * actually report an "unexpected EOF" error here.  Since it's a
614  * fairly common mistake to try to use rsh when ssh is required, we
615  * should trap that: if we fail to read any data at all, we should
616  * give a better explanation.  We can tell whether the connection has
617  * started by looking e.g. at whether the remote version is known yet.
618  */
619 static int read_timeout(int fd, char *buf, size_t len)
620 {
621         int n, cnt = 0;
622
623         io_flush(FULL_FLUSH);
624
625         while (cnt == 0) {
626                 /* until we manage to read *something* */
627                 fd_set r_fds, w_fds;
628                 struct timeval tv;
629                 int maxfd = fd;
630                 int count;
631
632                 FD_ZERO(&r_fds);
633                 FD_ZERO(&w_fds);
634                 FD_SET(fd, &r_fds);
635                 if (io_filesfrom_f_out >= 0) {
636                         int new_fd;
637                         if (io_filesfrom_buflen == 0) {
638                                 if (io_filesfrom_f_in >= 0) {
639                                         FD_SET(io_filesfrom_f_in, &r_fds);
640                                         new_fd = io_filesfrom_f_in;
641                                 } else {
642                                         io_filesfrom_f_out = -1;
643                                         new_fd = -1;
644                                 }
645                         } else {
646                                 FD_SET(io_filesfrom_f_out, &w_fds);
647                                 new_fd = io_filesfrom_f_out;
648                         }
649                         if (new_fd > maxfd)
650                                 maxfd = new_fd;
651                 }
652
653                 tv.tv_sec = select_timeout;
654                 tv.tv_usec = 0;
655
656                 errno = 0;
657
658                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
659
660                 if (count <= 0) {
661                         if (errno == EBADF) {
662                                 defer_forwarding_messages = 0;
663                                 exit_cleanup(RERR_SOCKETIO);
664                         }
665                         check_timeout();
666                         continue;
667                 }
668
669                 if (io_filesfrom_f_out >= 0) {
670                         if (io_filesfrom_buflen) {
671                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
672                                         int l = write(io_filesfrom_f_out,
673                                                       io_filesfrom_bp,
674                                                       io_filesfrom_buflen);
675                                         if (l > 0) {
676                                                 if (!(io_filesfrom_buflen -= l))
677                                                         io_filesfrom_bp = io_filesfrom_buf;
678                                                 else
679                                                         io_filesfrom_bp += l;
680                                         } else {
681                                                 /* XXX should we complain? */
682                                                 io_filesfrom_f_out = -1;
683                                         }
684                                 }
685                         } else if (io_filesfrom_f_in >= 0) {
686                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
687                                         int l = read(io_filesfrom_f_in,
688                                                      io_filesfrom_buf,
689                                                      sizeof io_filesfrom_buf);
690                                         if (l <= 0) {
691                                                 /* Send end-of-file marker */
692                                                 io_filesfrom_buf[0] = '\0';
693                                                 io_filesfrom_buf[1] = '\0';
694                                                 io_filesfrom_buflen = io_filesfrom_lastchar? 2 : 1;
695                                                 io_filesfrom_f_in = -1;
696                                         } else {
697                                                 if (!eol_nulls) {
698                                                         char *s = io_filesfrom_buf + l;
699                                                         /* Transform CR and/or LF into '\0' */
700                                                         while (s-- > io_filesfrom_buf) {
701                                                                 if (*s == '\n' || *s == '\r')
702                                                                         *s = '\0';
703                                                         }
704                                                 }
705                                                 if (!io_filesfrom_lastchar) {
706                                                         /* Last buf ended with a '\0', so don't
707                                                          * let this buf start with one. */
708                                                         while (l && !*io_filesfrom_bp)
709                                                                 io_filesfrom_bp++, l--;
710                                                 }
711                                                 if (!l)
712                                                         io_filesfrom_bp = io_filesfrom_buf;
713                                                 else {
714                                                         char *f = io_filesfrom_bp;
715                                                         char *t = f;
716                                                         char *eob = f + l;
717                                                         /* Eliminate any multi-'\0' runs. */
718                                                         while (f != eob) {
719                                                                 if (!(*t++ = *f++)) {
720                                                                         while (f != eob && !*f)
721                                                                                 f++, l--;
722                                                                 }
723                                                         }
724                                                         io_filesfrom_lastchar = f[-1];
725                                                 }
726                                                 io_filesfrom_buflen = l;
727                                         }
728                                 }
729                         }
730                 }
731
732                 if (!FD_ISSET(fd, &r_fds))
733                         continue;
734
735                 n = read(fd, buf, len);
736
737                 if (n <= 0) {
738                         if (n == 0)
739                                 whine_about_eof(fd); /* Doesn't return. */
740                         if (errno == EINTR || errno == EWOULDBLOCK
741                             || errno == EAGAIN)
742                                 continue;
743
744                         /* Don't write errors on a dead socket. */
745                         if (fd == sock_f_in) {
746                                 io_end_multiplex_out();
747                                 rsyserr(FSOCKERR, errno, "read error");
748                         } else
749                                 rsyserr(FERROR, errno, "read error");
750                         exit_cleanup(RERR_STREAMIO);
751                 }
752
753                 buf += n;
754                 len -= n;
755                 cnt += n;
756
757                 if (fd == sock_f_in && io_timeout)
758                         last_io_in = time(NULL);
759         }
760
761         return cnt;
762 }
763
764 /* Read a line into the "fname" buffer (which must be at least MAXPATHLEN
765  * characters long). */
766 int read_filesfrom_line(int fd, char *fname)
767 {
768         char ch, *s, *eob = fname + MAXPATHLEN - 1;
769         int cnt;
770         int reading_remotely = filesfrom_host != NULL;
771         int nulls = eol_nulls || reading_remotely;
772
773   start:
774         s = fname;
775         while (1) {
776                 cnt = read(fd, &ch, 1);
777                 if (cnt < 0 && (errno == EWOULDBLOCK
778                   || errno == EINTR || errno == EAGAIN)) {
779                         struct timeval tv;
780                         fd_set r_fds, e_fds;
781                         FD_ZERO(&r_fds);
782                         FD_SET(fd, &r_fds);
783                         FD_ZERO(&e_fds);
784                         FD_SET(fd, &e_fds);
785                         tv.tv_sec = select_timeout;
786                         tv.tv_usec = 0;
787                         if (!select(fd+1, &r_fds, NULL, &e_fds, &tv))
788                                 check_timeout();
789                         if (FD_ISSET(fd, &e_fds)) {
790                                 rsyserr(FINFO, errno,
791                                         "select exception on fd %d", fd);
792                         }
793                         continue;
794                 }
795                 if (cnt != 1)
796                         break;
797                 if (nulls? !ch : (ch == '\r' || ch == '\n')) {
798                         /* Skip empty lines if reading locally. */
799                         if (!reading_remotely && s == fname)
800                                 continue;
801                         break;
802                 }
803                 if (s < eob)
804                         *s++ = ch;
805         }
806         *s = '\0';
807
808         /* Dump comments. */
809         if (*fname == '#' || *fname == ';')
810                 goto start;
811
812         return s - fname;
813 }
814
815 int io_start_buffering_out(int f_out)
816 {
817         if (iobuf_out) {
818                 assert(f_out == iobuf_f_out);
819                 return 0;
820         }
821         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
822                 out_of_memory("io_start_buffering_out");
823         iobuf_out_cnt = 0;
824         iobuf_f_out = f_out;
825         return 1;
826 }
827
828 int io_start_buffering_in(int f_in)
829 {
830         if (iobuf_in) {
831                 assert(f_in == iobuf_f_in);
832                 return 0;
833         }
834         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
835         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
836                 out_of_memory("io_start_buffering_in");
837         iobuf_f_in = f_in;
838         return 1;
839 }
840
841 void io_end_buffering_in(void)
842 {
843         if (!iobuf_in)
844                 return;
845         free(iobuf_in);
846         iobuf_in = NULL;
847         iobuf_in_ndx = 0;
848         iobuf_in_remaining = 0;
849         iobuf_f_in = -1;
850 }
851
852 void io_end_buffering_out(void)
853 {
854         if (!iobuf_out)
855                 return;
856         io_flush(FULL_FLUSH);
857         free(iobuf_out);
858         iobuf_out = NULL;
859         iobuf_f_out = -1;
860 }
861
862 void maybe_flush_socket(int important)
863 {
864         if (iobuf_out && iobuf_out_cnt
865          && (important || time(NULL) - last_io_out >= 5))
866                 io_flush(NORMAL_FLUSH);
867 }
868
869 void maybe_send_keepalive(void)
870 {
871         if (time(NULL) - last_io_out >= allowed_lull) {
872                 if (!iobuf_out || !iobuf_out_cnt) {
873                         if (protocol_version < 29)
874                                 return; /* there's nothing we can do */
875                         if (protocol_version >= 30)
876                                 send_msg(MSG_NOOP, "", 0, 0);
877                         else {
878                                 write_int(sock_f_out, cur_flist->count);
879                                 write_shortint(sock_f_out, ITEM_IS_NEW);
880                         }
881                 }
882                 if (iobuf_out)
883                         io_flush(NORMAL_FLUSH);
884         }
885 }
886
887 void start_flist_forward(int f_in)
888 {
889         assert(iobuf_out != NULL);
890         assert(iobuf_f_out == msg_fd_out);
891         flist_forward_from = f_in;
892 }
893
894 void stop_flist_forward()
895 {
896         flist_forward_from = -1;
897         io_flush(FULL_FLUSH);
898 }
899
900 /**
901  * Continue trying to read len bytes - don't return until len has been
902  * read.
903  **/
904 static void read_loop(int fd, char *buf, size_t len)
905 {
906         while (len) {
907                 int n = read_timeout(fd, buf, len);
908
909                 buf += n;
910                 len -= n;
911         }
912 }
913
914 /**
915  * Read from the file descriptor handling multiplexing - return number
916  * of bytes read.
917  *
918  * Never returns <= 0.
919  */
920 static int readfd_unbuffered(int fd, char *buf, size_t len)
921 {
922         size_t msg_bytes;
923         int tag, cnt = 0;
924         char line[BIGPATHBUFLEN];
925
926         if (!iobuf_in || fd != iobuf_f_in)
927                 return read_timeout(fd, buf, len);
928
929         if (!io_multiplexing_in && iobuf_in_remaining == 0) {
930                 iobuf_in_remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
931                 iobuf_in_ndx = 0;
932         }
933
934         while (cnt == 0) {
935                 if (iobuf_in_remaining) {
936                         len = MIN(len, iobuf_in_remaining);
937                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
938                         iobuf_in_ndx += len;
939                         iobuf_in_remaining -= len;
940                         cnt = len;
941                         break;
942                 }
943
944                 read_loop(fd, line, 4);
945                 tag = IVAL(line, 0);
946
947                 msg_bytes = tag & 0xFFFFFF;
948                 tag = (tag >> 24) - MPLEX_BASE;
949
950                 switch (tag) {
951                 case MSG_DATA:
952                         if (msg_bytes > iobuf_in_siz) {
953                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
954                                                                msg_bytes)))
955                                         out_of_memory("readfd_unbuffered");
956                                 iobuf_in_siz = msg_bytes;
957                         }
958                         read_loop(fd, iobuf_in, msg_bytes);
959                         iobuf_in_remaining = msg_bytes;
960                         iobuf_in_ndx = 0;
961                         break;
962                 case MSG_NOOP:
963                         if (am_sender)
964                                 maybe_send_keepalive();
965                         break;
966                 case MSG_IO_ERROR:
967                         if (msg_bytes != 4)
968                                 goto invalid_msg;
969                         read_loop(fd, line, msg_bytes);
970                         io_error |= IVAL(line, 0);
971                         break;
972                 case MSG_DELETED:
973                         if (msg_bytes >= sizeof line)
974                                 goto overflow;
975 #ifdef ICONV_OPTION
976                         if (ic_recv != (iconv_t)-1) {
977                                 ICONV_CONST char *ibuf;
978                                 char *obuf = line;
979                                 size_t icnt, ocnt = sizeof line - 1;
980                                 int add_null = 0;
981                                 iconv(ic_send, NULL, 0, NULL, 0);
982                                 while (msg_bytes) {
983                                         icnt = msg_bytes > sizeof iconv_buf
984                                              ? sizeof iconv_buf : msg_bytes;
985                                         read_loop(fd, iconv_buf, icnt);
986                                         if (!(msg_bytes -= icnt) && !iconv_buf[icnt-1])
987                                                 icnt--, add_null = 1;
988                                         ibuf = (ICONV_CONST char *)iconv_buf;
989                                         if (iconv(ic_send, &ibuf,&icnt,
990                                                   &obuf,&ocnt) == (size_t)-1)
991                                                 goto overflow; // XXX
992                                 }
993                                 if (add_null)
994                                         *obuf++ = '\0';
995                                 msg_bytes = obuf - line;
996                         } else
997 #endif
998                                 read_loop(fd, line, msg_bytes);
999                         /* A directory name was sent with the trailing null */
1000                         if (msg_bytes > 0 && !line[msg_bytes-1])
1001                                 log_delete(line, S_IFDIR);
1002                         else {
1003                                 line[msg_bytes] = '\0';
1004                                 log_delete(line, S_IFREG);
1005                         }
1006                         break;
1007                 case MSG_SUCCESS:
1008                         if (msg_bytes != 4) {
1009                           invalid_msg:
1010                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
1011                                         tag, (long)msg_bytes, who_am_i());
1012                                 exit_cleanup(RERR_STREAMIO);
1013                         }
1014                         read_loop(fd, line, msg_bytes);
1015                         successful_send(IVAL(line, 0));
1016                         break;
1017                 case MSG_NO_SEND:
1018                         if (msg_bytes != 4)
1019                                 goto invalid_msg;
1020                         read_loop(fd, line, msg_bytes);
1021                         send_msg_int(MSG_NO_SEND, IVAL(line, 0));
1022                         break;
1023                 case MSG_INFO:
1024                 case MSG_ERROR:
1025                         if (msg_bytes >= sizeof line) {
1026                             overflow:
1027                                 rprintf(FERROR,
1028                                         "multiplexing overflow %d:%ld [%s]\n",
1029                                         tag, (long)msg_bytes, who_am_i());
1030                                 exit_cleanup(RERR_STREAMIO);
1031                         }
1032                         read_loop(fd, line, msg_bytes);
1033                         rwrite((enum logcode)tag, line, msg_bytes, 1);
1034                         break;
1035                 default:
1036                         rprintf(FERROR, "unexpected tag %d [%s]\n",
1037                                 tag, who_am_i());
1038                         exit_cleanup(RERR_STREAMIO);
1039                 }
1040         }
1041
1042         if (iobuf_in_remaining == 0)
1043                 io_flush(NORMAL_FLUSH);
1044
1045         return cnt;
1046 }
1047
1048 /* Do a buffered read from fd.  Don't return until all N bytes have
1049  * been read.  If all N can't be read then exit with an error. */
1050 static void readfd(int fd, char *buffer, size_t N)
1051 {
1052         int  cnt;
1053         size_t total = 0;
1054
1055         while (total < N) {
1056                 cnt = readfd_unbuffered(fd, buffer + total, N-total);
1057                 total += cnt;
1058         }
1059
1060         if (fd == write_batch_monitor_in) {
1061                 if ((size_t)write(batch_fd, buffer, total) != total)
1062                         exit_cleanup(RERR_FILEIO);
1063         }
1064
1065         if (fd == flist_forward_from)
1066                 writefd(iobuf_f_out, buffer, total);
1067
1068         if (fd == sock_f_in)
1069                 stats.total_read += total;
1070 }
1071
1072 unsigned short read_shortint(int f)
1073 {
1074         char b[2];
1075         readfd(f, b, 2);
1076         return (UVAL(b, 1) << 8) + UVAL(b, 0);
1077 }
1078
1079 int32 read_int(int f)
1080 {
1081         char b[4];
1082         int32 num;
1083
1084         readfd(f, b, 4);
1085         num = IVAL(b, 0);
1086 #if SIZEOF_INT32 > 4
1087         if (num & (int32)0x80000000)
1088                 num |= ~(int32)0xffffffff;
1089 #endif
1090         return num;
1091 }
1092
1093 int32 read_varint(int f)
1094 {
1095         union {
1096             char b[5];
1097             int32 x;
1098         } u;
1099         uchar ch;
1100         int extra;
1101
1102         u.x = 0;
1103         readfd(f, (char*)&ch, 1);
1104         extra = int_byte_extra[ch / 4];
1105         if (extra) {
1106                 uchar bit = ((uchar)1<<(8-extra));
1107                 if (extra >= (int)sizeof u.b) {
1108                         rprintf(FERROR, "Overflow in read_varint()\n");
1109                         exit_cleanup(RERR_STREAMIO);
1110                 }
1111                 readfd(f, u.b, extra);
1112                 u.b[extra] = ch & (bit-1);
1113         } else
1114                 u.b[0] = ch;
1115 #if CAREFUL_ALIGNMENT
1116         u.x = IVAL(u.b,0);
1117 #endif
1118 #if SIZEOF_INT32 > 4
1119         if (u.x & (int32)0x80000000)
1120                 u.x |= ~(int32)0xffffffff;
1121 #endif
1122         return u.x;
1123 }
1124
1125 int64 read_varlong(int f, uchar min_bytes)
1126 {
1127         union {
1128             char b[9];
1129             int64 x;
1130         } u;
1131         char b2[8];
1132         int extra;
1133
1134 #if SIZEOF_INT64 < 8
1135         memset(u.b, 0, 8);
1136 #else
1137         u.x = 0;
1138 #endif
1139         readfd(f, b2, min_bytes);
1140         memcpy(u.b, b2+1, min_bytes-1);
1141         extra = int_byte_extra[CVAL(b2, 0) / 4];
1142         if (extra) {
1143                 uchar bit = ((uchar)1<<(8-extra));
1144                 if (min_bytes + extra > (int)sizeof u.b) {
1145                         rprintf(FERROR, "Overflow in read_varlong()\n");
1146                         exit_cleanup(RERR_STREAMIO);
1147                 }
1148                 readfd(f, u.b + min_bytes - 1, extra);
1149                 u.b[min_bytes + extra - 1] = CVAL(b2, 0) & (bit-1);
1150 #if SIZEOF_INT64 < 8
1151                 if (min_bytes + extra > 5 || u.b[4] || CVAL(u.b,3) & 0x80) {
1152                         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1153                         exit_cleanup(RERR_UNSUPPORTED);
1154                 }
1155 #endif
1156         } else
1157                 u.b[min_bytes + extra - 1] = CVAL(b2, 0);
1158 #if SIZEOF_INT64 < 8
1159         u.x = IVAL(u.b,0);
1160 #elif CAREFUL_ALIGNMENT
1161         u.x = IVAL(u.b,0) | (((int64)IVAL(u.b,4))<<32);
1162 #endif
1163         return u.x;
1164 }
1165
1166 int64 read_longint(int f)
1167 {
1168 #if SIZEOF_INT64 >= 8
1169         char b[9];
1170 #endif
1171         int32 num = read_int(f);
1172
1173         if (num != (int32)0xffffffff)
1174                 return num;
1175
1176 #if SIZEOF_INT64 < 8
1177         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1178         exit_cleanup(RERR_UNSUPPORTED);
1179 #else
1180         readfd(f, b, 8);
1181         return IVAL(b,0) | (((int64)IVAL(b,4))<<32);
1182 #endif
1183 }
1184
1185 void read_buf(int f, char *buf, size_t len)
1186 {
1187         readfd(f,buf,len);
1188 }
1189
1190 void read_sbuf(int f, char *buf, size_t len)
1191 {
1192         readfd(f, buf, len);
1193         buf[len] = '\0';
1194 }
1195
1196 uchar read_byte(int f)
1197 {
1198         uchar c;
1199         readfd(f, (char *)&c, 1);
1200         return c;
1201 }
1202
1203 int read_vstring(int f, char *buf, int bufsize)
1204 {
1205         int len = read_byte(f);
1206
1207         if (len & 0x80)
1208                 len = (len & ~0x80) * 0x100 + read_byte(f);
1209
1210         if (len >= bufsize) {
1211                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
1212                         len, bufsize - 1);
1213                 return -1;
1214         }
1215
1216         if (len)
1217                 readfd(f, buf, len);
1218         buf[len] = '\0';
1219         return len;
1220 }
1221
1222 /* Populate a sum_struct with values from the socket.  This is
1223  * called by both the sender and the receiver. */
1224 void read_sum_head(int f, struct sum_struct *sum)
1225 {
1226         sum->count = read_int(f);
1227         if (sum->count < 0) {
1228                 rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
1229                         (long)sum->count, who_am_i());
1230                 exit_cleanup(RERR_PROTOCOL);
1231         }
1232         sum->blength = read_int(f);
1233         if (sum->blength < 0 || sum->blength > MAX_BLOCK_SIZE) {
1234                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
1235                         (long)sum->blength, who_am_i());
1236                 exit_cleanup(RERR_PROTOCOL);
1237         }
1238         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
1239         if (sum->s2length < 0 || sum->s2length > MAX_DIGEST_LEN) {
1240                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
1241                         sum->s2length, who_am_i());
1242                 exit_cleanup(RERR_PROTOCOL);
1243         }
1244         sum->remainder = read_int(f);
1245         if (sum->remainder < 0 || sum->remainder > sum->blength) {
1246                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
1247                         (long)sum->remainder, who_am_i());
1248                 exit_cleanup(RERR_PROTOCOL);
1249         }
1250 }
1251
1252 /* Send the values from a sum_struct over the socket.  Set sum to
1253  * NULL if there are no checksums to send.  This is called by both
1254  * the generator and the sender. */
1255 void write_sum_head(int f, struct sum_struct *sum)
1256 {
1257         static struct sum_struct null_sum;
1258
1259         if (sum == NULL)
1260                 sum = &null_sum;
1261
1262         write_int(f, sum->count);
1263         write_int(f, sum->blength);
1264         if (protocol_version >= 27)
1265                 write_int(f, sum->s2length);
1266         write_int(f, sum->remainder);
1267 }
1268
1269 /**
1270  * Sleep after writing to limit I/O bandwidth usage.
1271  *
1272  * @todo Rather than sleeping after each write, it might be better to
1273  * use some kind of averaging.  The current algorithm seems to always
1274  * use a bit less bandwidth than specified, because it doesn't make up
1275  * for slow periods.  But arguably this is a feature.  In addition, we
1276  * ought to take the time used to write the data into account.
1277  *
1278  * During some phases of big transfers (file FOO is uptodate) this is
1279  * called with a small bytes_written every time.  As the kernel has to
1280  * round small waits up to guarantee that we actually wait at least the
1281  * requested number of microseconds, this can become grossly inaccurate.
1282  * We therefore keep track of the bytes we've written over time and only
1283  * sleep when the accumulated delay is at least 1 tenth of a second.
1284  **/
1285 static void sleep_for_bwlimit(int bytes_written)
1286 {
1287         static struct timeval prior_tv;
1288         static long total_written = 0;
1289         struct timeval tv, start_tv;
1290         long elapsed_usec, sleep_usec;
1291
1292 #define ONE_SEC 1000000L /* # of microseconds in a second */
1293
1294         if (!bwlimit_writemax)
1295                 return;
1296
1297         total_written += bytes_written;
1298
1299         gettimeofday(&start_tv, NULL);
1300         if (prior_tv.tv_sec) {
1301                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
1302                              + (start_tv.tv_usec - prior_tv.tv_usec);
1303                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
1304                 if (total_written < 0)
1305                         total_written = 0;
1306         }
1307
1308         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
1309         if (sleep_usec < ONE_SEC / 10) {
1310                 prior_tv = start_tv;
1311                 return;
1312         }
1313
1314         tv.tv_sec  = sleep_usec / ONE_SEC;
1315         tv.tv_usec = sleep_usec % ONE_SEC;
1316         select(0, NULL, NULL, NULL, &tv);
1317
1318         gettimeofday(&prior_tv, NULL);
1319         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
1320                      + (prior_tv.tv_usec - start_tv.tv_usec);
1321         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
1322 }
1323
1324 /* Write len bytes to the file descriptor fd, looping as necessary to get
1325  * the job done and also (in certain circumstances) reading any data on
1326  * msg_fd_in to avoid deadlock.
1327  *
1328  * This function underlies the multiplexing system.  The body of the
1329  * application never calls this function directly. */
1330 static void writefd_unbuffered(int fd, const char *buf, size_t len)
1331 {
1332         size_t n, total = 0;
1333         fd_set w_fds, r_fds, e_fds;
1334         int maxfd, count, cnt, using_r_fds;
1335         int defer_inc = 0;
1336         struct timeval tv;
1337
1338         if (no_flush++)
1339                 defer_forwarding_messages++, defer_inc++;
1340
1341         while (total < len) {
1342                 FD_ZERO(&w_fds);
1343                 FD_SET(fd, &w_fds);
1344                 FD_ZERO(&e_fds);
1345                 FD_SET(fd, &e_fds);
1346                 maxfd = fd;
1347
1348                 if (msg_fd_in >= 0) {
1349                         FD_ZERO(&r_fds);
1350                         FD_SET(msg_fd_in, &r_fds);
1351                         if (msg_fd_in > maxfd)
1352                                 maxfd = msg_fd_in;
1353                         using_r_fds = 1;
1354                 } else
1355                         using_r_fds = 0;
1356
1357                 tv.tv_sec = select_timeout;
1358                 tv.tv_usec = 0;
1359
1360                 errno = 0;
1361                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1362                                &w_fds, &e_fds, &tv);
1363
1364                 if (count <= 0) {
1365                         if (count < 0 && errno == EBADF)
1366                                 exit_cleanup(RERR_SOCKETIO);
1367                         check_timeout();
1368                         continue;
1369                 }
1370
1371                 if (FD_ISSET(fd, &e_fds)) {
1372                         rsyserr(FINFO, errno,
1373                                 "select exception on fd %d", fd);
1374                 }
1375
1376                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1377                         read_msg_fd();
1378
1379                 if (!FD_ISSET(fd, &w_fds))
1380                         continue;
1381
1382                 n = len - total;
1383                 if (bwlimit_writemax && n > bwlimit_writemax)
1384                         n = bwlimit_writemax;
1385                 cnt = write(fd, buf + total, n);
1386
1387                 if (cnt <= 0) {
1388                         if (cnt < 0) {
1389                                 if (errno == EINTR)
1390                                         continue;
1391                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1392                                         msleep(1);
1393                                         continue;
1394                                 }
1395                         }
1396
1397                         /* Don't try to write errors back across the stream. */
1398                         if (fd == sock_f_out)
1399                                 io_end_multiplex_out();
1400                         /* Don't try to write errors down a failing msg pipe. */
1401                         if (am_server && fd == msg_fd_out)
1402                                 exit_cleanup(RERR_STREAMIO);
1403                         rsyserr(FERROR, errno,
1404                                 "writefd_unbuffered failed to write %ld bytes [%s]",
1405                                 (long)len, who_am_i());
1406                         /* If the other side is sending us error messages, try
1407                          * to grab any messages they sent before they died. */
1408                         while (fd == sock_f_out && io_multiplexing_in) {
1409                                 set_io_timeout(30);
1410                                 ignore_timeout = 0;
1411                                 readfd_unbuffered(sock_f_in, io_filesfrom_buf,
1412                                                   sizeof io_filesfrom_buf);
1413                         }
1414                         exit_cleanup(RERR_STREAMIO);
1415                 }
1416
1417                 total += cnt;
1418                 defer_forwarding_messages++, defer_inc++;
1419
1420                 if (fd == sock_f_out) {
1421                         if (io_timeout || am_generator)
1422                                 last_io_out = time(NULL);
1423                         sleep_for_bwlimit(cnt);
1424                 }
1425         }
1426
1427         no_flush--;
1428         if (!(defer_forwarding_messages -= defer_inc))
1429                 msg_flush();
1430 }
1431
1432 void io_flush(int flush_it_all)
1433 {
1434         if (!iobuf_out_cnt || no_flush)
1435                 return;
1436
1437         if (io_multiplexing_out)
1438                 mplex_write(sock_f_out, MSG_DATA, iobuf_out, iobuf_out_cnt, 0);
1439         else
1440                 writefd_unbuffered(iobuf_f_out, iobuf_out, iobuf_out_cnt);
1441         iobuf_out_cnt = 0;
1442
1443         if (flush_it_all && !defer_forwarding_messages)
1444                 msg_flush();
1445 }
1446
1447 static void writefd(int fd, const char *buf, size_t len)
1448 {
1449         if (fd == sock_f_out)
1450                 stats.total_written += len;
1451
1452         if (fd == write_batch_monitor_out) {
1453                 if ((size_t)write(batch_fd, buf, len) != len)
1454                         exit_cleanup(RERR_FILEIO);
1455         }
1456
1457         if (!iobuf_out || fd != iobuf_f_out) {
1458                 writefd_unbuffered(fd, buf, len);
1459                 return;
1460         }
1461
1462         while (len) {
1463                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1464                 if (n > 0) {
1465                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1466                         buf += n;
1467                         len -= n;
1468                         iobuf_out_cnt += n;
1469                 }
1470
1471                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1472                         io_flush(NORMAL_FLUSH);
1473         }
1474 }
1475
1476 void write_shortint(int f, unsigned short x)
1477 {
1478         char b[2];
1479         b[0] = (char)x;
1480         b[1] = (char)(x >> 8);
1481         writefd(f, b, 2);
1482 }
1483
1484 void write_int(int f, int32 x)
1485 {
1486         char b[4];
1487         SIVAL(b, 0, x);
1488         writefd(f, b, 4);
1489 }
1490
1491 void write_varint(int f, int32 x)
1492 {
1493         char b[5];
1494         uchar bit;
1495         int cnt = 4;
1496
1497         SIVAL(b, 1, x);
1498
1499         while (cnt > 1 && b[cnt] == 0)
1500                 cnt--;
1501         bit = ((uchar)1<<(7-cnt+1));
1502         if (CVAL(b, cnt) >= bit) {
1503                 cnt++;
1504                 *b = ~(bit-1);
1505         } else if (cnt > 1)
1506                 *b = b[cnt] | ~(bit*2-1);
1507         else
1508                 *b = b[cnt];
1509
1510         writefd(f, b, cnt);
1511 }
1512
1513 void write_varlong(int f, int64 x, uchar min_bytes)
1514 {
1515         char b[9];
1516         uchar bit;
1517         int cnt = 8;
1518
1519         SIVAL(b, 1, x);
1520 #if SIZEOF_INT64 >= 8
1521         SIVAL(b, 5, x >> 32);
1522 #else
1523         if (x <= 0x7FFFFFFF && x >= 0)
1524                 memset(b + 5, 0, 4);
1525         else {
1526                 rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1527                 exit_cleanup(RERR_UNSUPPORTED);
1528         }
1529 #endif
1530
1531         while (cnt > min_bytes && b[cnt] == 0)
1532                 cnt--;
1533         bit = ((uchar)1<<(7-cnt+min_bytes));
1534         if (CVAL(b, cnt) >= bit) {
1535                 cnt++;
1536                 *b = ~(bit-1);
1537         } else if (cnt > min_bytes)
1538                 *b = b[cnt] | ~(bit*2-1);
1539         else
1540                 *b = b[cnt];
1541
1542         writefd(f, b, cnt);
1543 }
1544
1545 /*
1546  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1547  * 64-bit types on this platform.
1548  */
1549 void write_longint(int f, int64 x)
1550 {
1551         char b[12], * const s = b+4;
1552
1553         SIVAL(s, 0, x);
1554         if (x <= 0x7FFFFFFF && x >= 0) {
1555                 writefd(f, s, 4);
1556                 return;
1557         }
1558
1559 #if SIZEOF_INT64 < 8
1560         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1561         exit_cleanup(RERR_UNSUPPORTED);
1562 #else
1563         memset(b, 0xFF, 4);
1564         SIVAL(s, 4, x >> 32);
1565         writefd(f, b, 12);
1566 #endif
1567 }
1568
1569 void write_buf(int f, const char *buf, size_t len)
1570 {
1571         writefd(f,buf,len);
1572 }
1573
1574 /** Write a string to the connection */
1575 void write_sbuf(int f, const char *buf)
1576 {
1577         writefd(f, buf, strlen(buf));
1578 }
1579
1580 void write_byte(int f, uchar c)
1581 {
1582         writefd(f, (char *)&c, 1);
1583 }
1584
1585 void write_vstring(int f, const char *str, int len)
1586 {
1587         uchar lenbuf[3], *lb = lenbuf;
1588
1589         if (len > 0x7F) {
1590                 if (len > 0x7FFF) {
1591                         rprintf(FERROR,
1592                                 "attempting to send over-long vstring (%d > %d)\n",
1593                                 len, 0x7FFF);
1594                         exit_cleanup(RERR_PROTOCOL);
1595                 }
1596                 *lb++ = len / 0x100 + 0x80;
1597         }
1598         *lb = len;
1599
1600         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1601         if (len)
1602                 writefd(f, str, len);
1603 }
1604
1605 /* Send a file-list index using a byte-reduction method. */
1606 void write_ndx(int f, int32 ndx)
1607 {
1608         static int32 prev_positive = -1, prev_negative = 1;
1609         int32 diff, cnt = 0;
1610         char b[6];
1611
1612         if (protocol_version < 30 || read_batch) {
1613                 write_int(f, ndx);
1614                 return;
1615         }
1616
1617         /* Send NDX_DONE as a single-byte 0 with no side effects.  Send
1618          * negative nums as a positive after sending a leading 0xFF. */
1619         if (ndx >= 0) {
1620                 diff = ndx - prev_positive;
1621                 prev_positive = ndx;
1622         } else if (ndx == NDX_DONE) {
1623                 *b = 0;
1624                 writefd(f, b, 1);
1625                 return;
1626         } else {
1627                 b[cnt++] = (char)0xFF;
1628                 ndx = -ndx;
1629                 diff = ndx - prev_negative;
1630                 prev_negative = ndx;
1631         }
1632
1633         /* A diff of 1 - 253 is sent as a one-byte diff; a diff of 254 - 32767
1634          * or 0 is sent as a 0xFE + a two-byte diff; otherwise we send 0xFE
1635          * & all 4 bytes of the (non-negative) num with the high-bit set. */
1636         if (diff < 0xFE && diff > 0)
1637                 b[cnt++] = (char)diff;
1638         else if (diff < 0 || diff > 0x7FFF) {
1639                 b[cnt++] = (char)0xFE;
1640                 b[cnt++] = (char)((ndx >> 24) | 0x80);
1641                 b[cnt++] = (char)ndx;
1642                 b[cnt++] = (char)(ndx >> 8);
1643                 b[cnt++] = (char)(ndx >> 16);
1644         } else {
1645                 b[cnt++] = (char)0xFE;
1646                 b[cnt++] = (char)(diff >> 8);
1647                 b[cnt++] = (char)diff;
1648         }
1649         writefd(f, b, cnt);
1650 }
1651
1652 /* Receive a file-list index using a byte-reduction method. */
1653 int32 read_ndx(int f)
1654 {
1655         static int32 prev_positive = -1, prev_negative = 1;
1656         int32 *prev_ptr, num;
1657         char b[4];
1658
1659         if (protocol_version < 30)
1660                 return read_int(f);
1661
1662         readfd(f, b, 1);
1663         if (CVAL(b, 0) == 0xFF) {
1664                 readfd(f, b, 1);
1665                 prev_ptr = &prev_negative;
1666         } else if (CVAL(b, 0) == 0)
1667                 return NDX_DONE;
1668         else
1669                 prev_ptr = &prev_positive;
1670         if (CVAL(b, 0) == 0xFE) {
1671                 readfd(f, b, 2);
1672                 if (CVAL(b, 0) & 0x80) {
1673                         b[3] = CVAL(b, 0) & ~0x80;
1674                         b[0] = b[1];
1675                         readfd(f, b+1, 2);
1676                         num = IVAL(b, 0);
1677                 } else
1678                         num = (UVAL(b,0)<<8) + UVAL(b,1) + *prev_ptr;
1679         } else
1680                 num = UVAL(b, 0) + *prev_ptr;
1681         *prev_ptr = num;
1682         if (prev_ptr == &prev_negative)
1683                 num = -num;
1684         return num;
1685 }
1686
1687 /**
1688  * Read a line of up to @p maxlen characters into @p buf (not counting
1689  * the trailing null).  Strips the (required) trailing newline and all
1690  * carriage returns.
1691  *
1692  * @return 1 for success; 0 for I/O error or truncation.
1693  **/
1694 int read_line(int f, char *buf, size_t maxlen)
1695 {
1696         while (maxlen) {
1697                 buf[0] = 0;
1698                 read_buf(f, buf, 1);
1699                 if (buf[0] == 0)
1700                         return 0;
1701                 if (buf[0] == '\n')
1702                         break;
1703                 if (buf[0] != '\r') {
1704                         buf++;
1705                         maxlen--;
1706                 }
1707         }
1708         *buf = '\0';
1709         return maxlen > 0;
1710 }
1711
1712 void io_printf(int fd, const char *format, ...)
1713 {
1714         va_list ap;
1715         char buf[BIGPATHBUFLEN];
1716         int len;
1717
1718         va_start(ap, format);
1719         len = vsnprintf(buf, sizeof buf, format, ap);
1720         va_end(ap);
1721
1722         if (len < 0)
1723                 exit_cleanup(RERR_STREAMIO);
1724
1725         if (len > (int)sizeof buf) {
1726                 rprintf(FERROR, "io_printf() was too long for the buffer.\n");
1727                 exit_cleanup(RERR_STREAMIO);
1728         }
1729
1730         write_sbuf(fd, buf);
1731 }
1732
1733 /** Setup for multiplexing a MSG_* stream with the data stream. */
1734 void io_start_multiplex_out(void)
1735 {
1736         io_flush(NORMAL_FLUSH);
1737         io_start_buffering_out(sock_f_out);
1738         io_multiplexing_out = 1;
1739 }
1740
1741 /** Setup for multiplexing a MSG_* stream with the data stream. */
1742 void io_start_multiplex_in(void)
1743 {
1744         io_flush(NORMAL_FLUSH);
1745         io_start_buffering_in(sock_f_in);
1746         io_multiplexing_in = 1;
1747 }
1748
1749 /** Write an message to the multiplexed data stream. */
1750 int io_multiplex_write(enum msgcode code, const char *buf, size_t len, int convert)
1751 {
1752         if (!io_multiplexing_out)
1753                 return 0;
1754         io_flush(NORMAL_FLUSH);
1755         stats.total_written += (len+4);
1756         mplex_write(sock_f_out, code, buf, len, convert);
1757         return 1;
1758 }
1759
1760 void io_end_multiplex_in(void)
1761 {
1762         io_multiplexing_in = 0;
1763         io_end_buffering_in();
1764 }
1765
1766 /** Stop output multiplexing. */
1767 void io_end_multiplex_out(void)
1768 {
1769         io_multiplexing_out = 0;
1770         io_end_buffering_out();
1771 }
1772
1773 void start_write_batch(int fd)
1774 {
1775         /* Some communication has already taken place, but we don't
1776          * enable batch writing until here so that we can write a
1777          * canonical record of the communication even though the
1778          * actual communication so far depends on whether a daemon
1779          * is involved. */
1780         write_int(batch_fd, protocol_version);
1781         write_int(batch_fd, checksum_seed);
1782
1783         if (am_sender)
1784                 write_batch_monitor_out = fd;
1785         else
1786                 write_batch_monitor_in = fd;
1787 }
1788
1789 void stop_write_batch(void)
1790 {
1791         write_batch_monitor_out = -1;
1792         write_batch_monitor_in = -1;
1793 }