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