Changed flist_for_ndx() to optionally die with an error
[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) 2003-2008 Wayne Davison
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 3 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, visit the http://fsf.org website.
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 #include "ifuncs.h"
32
33 /** If no timeout is specified then use a 60 second select timeout */
34 #define SELECT_TIMEOUT 60
35
36 extern int bwlimit;
37 extern size_t bwlimit_writemax;
38 extern int io_timeout;
39 extern int 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 list_only;
48 extern int read_batch;
49 extern int protect_args;
50 extern int checksum_seed;
51 extern int protocol_version;
52 extern int remove_source_files;
53 extern int preserve_hard_links;
54 extern struct stats stats;
55 extern struct file_list *cur_flist;
56 #ifdef ICONV_OPTION
57 extern int filesfrom_convert;
58 extern iconv_t ic_send, ic_recv;
59 #endif
60
61 int csum_length = SHORT_SUM_LENGTH; /* initial value */
62 int allowed_lull = 0;
63 int ignore_timeout = 0;
64 int batch_fd = -1;
65 int msgdone_cnt = 0;
66 int check_for_io_err = 0;
67
68 /* Ignore an EOF error if non-zero. See whine_about_eof(). */
69 int kluge_around_eof = 0;
70
71 int msg_fd_in = -1;
72 int msg_fd_out = -1;
73 int sock_f_in = -1;
74 int sock_f_out = -1;
75
76 static int iobuf_f_in = -1;
77 static char *iobuf_in;
78 static size_t iobuf_in_siz;
79 static size_t iobuf_in_ndx;
80 static size_t iobuf_in_remaining;
81
82 static int iobuf_f_out = -1;
83 static char *iobuf_out;
84 static int iobuf_out_cnt;
85
86 int flist_forward_from = -1;
87
88 static int io_multiplexing_out;
89 static int io_multiplexing_in;
90 static time_t last_io_in;
91 static time_t last_io_out;
92 static int no_flush;
93
94 static int write_batch_monitor_in = -1;
95 static int write_batch_monitor_out = -1;
96
97 static int io_filesfrom_f_in = -1;
98 static int io_filesfrom_f_out = -1;
99 static xbuf ff_buf = EMPTY_XBUF;
100 static char ff_lastchar;
101 #ifdef ICONV_OPTION
102 static xbuf iconv_buf = EMPTY_XBUF;
103 #endif
104 static int defer_forwarding_messages = 0, defer_forwarding_keep = 0;
105 static int select_timeout = SELECT_TIMEOUT;
106 static int active_filecnt = 0;
107 static OFF_T active_bytecnt = 0;
108 static int first_message = 1;
109
110 static char int_byte_extra[64] = {
111         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* (00 - 3F)/4 */
112         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* (40 - 7F)/4 */
113         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* (80 - BF)/4 */
114         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 5, 6, /* (C0 - FF)/4 */
115 };
116
117 #define REMOTE_OPTION_ERROR "rsync: on remote machine: -"
118 #define REMOTE_OPTION_ERROR2 ": unknown option"
119
120 enum festatus { FES_SUCCESS, FES_REDO, FES_NO_SEND };
121
122 static void readfd(int fd, char *buffer, size_t N);
123 static void writefd(int fd, const char *buf, size_t len);
124 static void writefd_unbuffered(int fd, const char *buf, size_t len);
125 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert);
126
127 struct flist_ndx_item {
128         struct flist_ndx_item *next;
129         int ndx;
130 };
131
132 struct flist_ndx_list {
133         struct flist_ndx_item *head, *tail;
134 };
135
136 static struct flist_ndx_list redo_list, hlink_list;
137
138 struct msg_list_item {
139         struct msg_list_item *next;
140         char convert;
141         char buf[1];
142 };
143
144 struct msg_list {
145         struct msg_list_item *head, *tail;
146 };
147
148 static struct msg_list msg_queue;
149
150 static void flist_ndx_push(struct flist_ndx_list *lp, int ndx)
151 {
152         struct flist_ndx_item *item;
153
154         if (!(item = new(struct flist_ndx_item)))
155                 out_of_memory("flist_ndx_push");
156         item->next = NULL;
157         item->ndx = ndx;
158         if (lp->tail)
159                 lp->tail->next = item;
160         else
161                 lp->head = item;
162         lp->tail = item;
163 }
164
165 static int flist_ndx_pop(struct flist_ndx_list *lp)
166 {
167         struct flist_ndx_item *next;
168         int ndx;
169
170         if (!lp->head)
171                 return -1;
172
173         ndx = lp->head->ndx;
174         next = lp->head->next;
175         free(lp->head);
176         lp->head = next;
177         if (!next)
178                 lp->tail = NULL;
179
180         return ndx;
181 }
182
183 static void got_flist_entry_status(enum festatus status, const char *buf)
184 {
185         int ndx = IVAL(buf, 0);
186         struct file_list *flist = flist_for_ndx(ndx, "got_flist_entry_status");
187
188         if (remove_source_files) {
189                 active_filecnt--;
190                 active_bytecnt -= F_LENGTH(flist->files[ndx - flist->ndx_start]);
191         }
192
193         if (inc_recurse)
194                 flist->in_progress--;
195
196         switch (status) {
197         case FES_SUCCESS:
198                 if (remove_source_files)
199                         send_msg(MSG_SUCCESS, buf, 4, 0);
200                 if (preserve_hard_links) {
201                         struct file_struct *file = flist->files[ndx - flist->ndx_start];
202                         if (F_IS_HLINKED(file)) {
203                                 flist_ndx_push(&hlink_list, ndx);
204                                 flist->in_progress++;
205                         }
206                 }
207                 break;
208         case FES_REDO:
209                 if (inc_recurse)
210                         flist->to_redo++;
211                 flist_ndx_push(&redo_list, ndx);
212                 break;
213         case FES_NO_SEND:
214                 break;
215         }
216 }
217
218 static void check_timeout(void)
219 {
220         time_t t;
221
222         if (!io_timeout || ignore_timeout)
223                 return;
224
225         if (!last_io_in) {
226                 last_io_in = time(NULL);
227                 return;
228         }
229
230         t = time(NULL);
231
232         if (t - last_io_in >= io_timeout) {
233                 if (!am_server && !am_daemon) {
234                         rprintf(FERROR, "io timeout after %d seconds -- exiting\n",
235                                 (int)(t-last_io_in));
236                 }
237                 exit_cleanup(RERR_TIMEOUT);
238         }
239 }
240
241 /* Note the fds used for the main socket (which might really be a pipe
242  * for a local transfer, but we can ignore that). */
243 void io_set_sock_fds(int f_in, int f_out)
244 {
245         sock_f_in = f_in;
246         sock_f_out = f_out;
247 }
248
249 void set_io_timeout(int secs)
250 {
251         io_timeout = secs;
252
253         if (!io_timeout || io_timeout > SELECT_TIMEOUT)
254                 select_timeout = SELECT_TIMEOUT;
255         else
256                 select_timeout = io_timeout;
257
258         allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2;
259 }
260
261 /* Setup the fd used to receive MSG_* messages.  Only needed during the
262  * early stages of being a local sender (up through the sending of the
263  * file list) or when we're the generator (to fetch the messages from
264  * the receiver). */
265 void set_msg_fd_in(int fd)
266 {
267         msg_fd_in = fd;
268 }
269
270 /* Setup the fd used to send our MSG_* messages.  Only needed when
271  * we're the receiver (to send our messages to the generator). */
272 void set_msg_fd_out(int fd)
273 {
274         msg_fd_out = fd;
275         set_nonblocking(msg_fd_out);
276 }
277
278 /* Add a message to the pending MSG_* list. */
279 static void msg_list_add(struct msg_list *lst, int code, const char *buf, int len, int convert)
280 {
281         struct msg_list_item *m;
282         int sz = len + 4 + sizeof m[0] - 1;
283
284         if (!(m = (struct msg_list_item *)new_array(char, sz)))
285                 out_of_memory("msg_list_add");
286         m->next = NULL;
287         m->convert = convert;
288         SIVAL(m->buf, 0, ((code+MPLEX_BASE)<<24) | len);
289         memcpy(m->buf + 4, buf, len);
290         if (lst->tail)
291                 lst->tail->next = m;
292         else
293                 lst->head = m;
294         lst->tail = m;
295 }
296
297 static void msg_flush(void)
298 {
299         if (am_generator) {
300                 while (msg_queue.head && io_multiplexing_out) {
301                         struct msg_list_item *m = msg_queue.head;
302                         int len = IVAL(m->buf, 0) & 0xFFFFFF;
303                         int tag = *((uchar*)m->buf+3) - MPLEX_BASE;
304                         if (!(msg_queue.head = m->next))
305                                 msg_queue.tail = NULL;
306                         stats.total_written += len + 4;
307                         defer_forwarding_messages++;
308                         mplex_write(sock_f_out, tag, m->buf + 4, len, m->convert);
309                         defer_forwarding_messages--;
310                         free(m);
311                 }
312         } else {
313                 while (msg_queue.head) {
314                         struct msg_list_item *m = msg_queue.head;
315                         int len = IVAL(m->buf, 0) & 0xFFFFFF;
316                         int tag = *((uchar*)m->buf+3) - MPLEX_BASE;
317                         if (!(msg_queue.head = m->next))
318                                 msg_queue.tail = NULL;
319                         defer_forwarding_messages++;
320                         mplex_write(msg_fd_out, tag, m->buf + 4, len, m->convert);
321                         defer_forwarding_messages--;
322                         free(m);
323                 }
324         }
325 }
326
327 static void check_for_d_option_error(const char *msg)
328 {
329         static char rsync263_opts[] = "BCDHIKLPRSTWabceghlnopqrtuvxz";
330         char *colon;
331         int saw_d = 0;
332
333         if (*msg != 'r'
334          || strncmp(msg, REMOTE_OPTION_ERROR, sizeof REMOTE_OPTION_ERROR - 1) != 0)
335                 return;
336
337         msg += sizeof REMOTE_OPTION_ERROR - 1;
338         if (*msg == '-' || (colon = strchr(msg, ':')) == NULL
339          || strncmp(colon, REMOTE_OPTION_ERROR2, sizeof REMOTE_OPTION_ERROR2 - 1) != 0)
340                 return;
341
342         for ( ; *msg != ':'; msg++) {
343                 if (*msg == 'd')
344                         saw_d = 1;
345                 else if (*msg == 'e')
346                         break;
347                 else if (strchr(rsync263_opts, *msg) == NULL)
348                         return;
349         }
350
351         if (saw_d) {
352                 rprintf(FWARNING,
353                     "*** Try using \"--old-d\" if remote rsync is <= 2.6.3 ***\n");
354         }
355 }
356
357 /* Read a message from the MSG_* fd and handle it.  This is called either
358  * during the early stages of being a local sender (up through the sending
359  * of the file list) or when we're the generator (to fetch the messages
360  * from the receiver). */
361 static void read_msg_fd(void)
362 {
363         char buf[2048];
364         size_t n;
365         struct file_list *flist;
366         int fd = msg_fd_in;
367         int tag, len;
368
369         /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
370          * to this routine from writefd_unbuffered(). */
371         no_flush++;
372         msg_fd_in = -1;
373         defer_forwarding_messages++;
374
375         readfd(fd, buf, 4);
376         tag = IVAL(buf, 0);
377
378         len = tag & 0xFFFFFF;
379         tag = (tag >> 24) - MPLEX_BASE;
380
381         check_for_io_err = 0;
382
383         switch (tag) {
384         case MSG_DONE:
385                 if (len < 0 || len > 1 || !am_generator) {
386                   invalid_msg:
387                         rprintf(FERROR, "invalid message %d:%d [%s%s]\n",
388                                 tag, len, who_am_i(),
389                                 inc_recurse ? "/inc" : "");
390                         exit_cleanup(RERR_STREAMIO);
391                 }
392                 if (len) {
393                         readfd(fd, buf, len);
394                         stats.total_read = read_varlong(fd, 3);
395                 }
396                 msgdone_cnt++;
397                 break;
398         case MSG_REDO:
399                 if (len != 4 || !am_generator)
400                         goto invalid_msg;
401                 readfd(fd, buf, 4);
402                 got_flist_entry_status(FES_REDO, buf);
403                 break;
404         case MSG_FLIST:
405                 if (len != 4 || !am_generator || !inc_recurse)
406                         goto invalid_msg;
407                 readfd(fd, buf, 4);
408                 /* Read extra file list from receiver. */
409                 assert(iobuf_in != NULL);
410                 assert(iobuf_f_in == fd);
411                 if (DEBUG_GTE(FLIST, 2)) {
412                         rprintf(FINFO, "[%s] receiving flist for dir %d\n",
413                                 who_am_i(), IVAL(buf,0));
414                 }
415                 flist = recv_file_list(fd);
416                 flist->parent_ndx = IVAL(buf,0);
417                 /* If the sender is going to send us an MSG_IO_ERROR value, it
418                  * will always be the very next message following MSG_FLIST. */
419                 check_for_io_err = 1;
420 #ifdef SUPPORT_HARD_LINKS
421                 if (preserve_hard_links)
422                         match_hard_links(flist);
423 #endif
424                 break;
425         case MSG_FLIST_EOF:
426                 if (len != 0 || !am_generator || !inc_recurse)
427                         goto invalid_msg;
428                 flist_eof = 1;
429                 break;
430         case MSG_IO_ERROR:
431                 if (len != 4)
432                         goto invalid_msg;
433                 readfd(fd, buf, len);
434                 io_error |= IVAL(buf, 0);
435                 break;
436         case MSG_DELETED:
437                 if (len >= (int)sizeof buf || !am_generator)
438                         goto invalid_msg;
439                 readfd(fd, buf, len);
440                 send_msg(MSG_DELETED, buf, len, 1);
441                 break;
442         case MSG_SUCCESS:
443                 if (len != 4 || !am_generator)
444                         goto invalid_msg;
445                 readfd(fd, buf, 4);
446                 got_flist_entry_status(FES_SUCCESS, buf);
447                 break;
448         case MSG_NO_SEND:
449                 if (len != 4 || !am_generator)
450                         goto invalid_msg;
451                 readfd(fd, buf, 4);
452                 got_flist_entry_status(FES_NO_SEND, buf);
453                 break;
454         case MSG_ERROR_SOCKET:
455         case MSG_ERROR_UTF8:
456         case MSG_CLIENT:
457                 if (!am_generator)
458                         goto invalid_msg;
459                 if (tag == MSG_ERROR_SOCKET)
460                         io_end_multiplex_out();
461                 /* FALL THROUGH */
462         case MSG_INFO:
463         case MSG_ERROR:
464         case MSG_ERROR_XFER:
465         case MSG_WARNING:
466         case MSG_LOG:
467                 while (len) {
468                         n = len;
469                         if (n >= sizeof buf)
470                                 n = sizeof buf - 1;
471                         readfd(fd, buf, n);
472                         rwrite((enum logcode)tag, buf, n, !am_generator);
473                         len -= n;
474                 }
475                 break;
476         default:
477                 rprintf(FERROR, "unknown message %d:%d [%s]\n",
478                         tag, len, who_am_i());
479                 exit_cleanup(RERR_STREAMIO);
480         }
481
482         no_flush--;
483         msg_fd_in = fd;
484         if (!--defer_forwarding_messages && !no_flush)
485                 msg_flush();
486 }
487
488 /* This is used by the generator to limit how many file transfers can
489  * be active at once when --remove-source-files is specified.  Without
490  * this, sender-side deletions were mostly happening at the end. */
491 void increment_active_files(int ndx, int itemizing, enum logcode code)
492 {
493         /* TODO: tune these limits? */
494         while (active_filecnt >= (active_bytecnt >= 128*1024 ? 10 : 50)) {
495                 check_for_finished_files(itemizing, code, 0);
496                 if (iobuf_out_cnt)
497                         io_flush(NORMAL_FLUSH);
498                 else
499                         read_msg_fd();
500         }
501
502         active_filecnt++;
503         active_bytecnt += F_LENGTH(cur_flist->files[ndx - cur_flist->ndx_start]);
504 }
505
506 /* Write an message to a multiplexed stream. If this fails, rsync exits. */
507 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert)
508 {
509         char buffer[BIGPATHBUFLEN]; /* Oversized for use by iconv code. */
510         size_t n = len;
511
512 #ifdef ICONV_OPTION
513         /* We need to convert buf before doing anything else so that we
514          * can include the (converted) byte length in the message header. */
515         if (convert && ic_send != (iconv_t)-1) {
516                 xbuf outbuf, inbuf;
517
518                 INIT_XBUF(outbuf, buffer + 4, 0, sizeof buffer - 4);
519                 INIT_XBUF(inbuf, (char*)buf, len, (size_t)-1);
520
521                 iconvbufs(ic_send, &inbuf, &outbuf,
522                           ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE);
523                 if (inbuf.len > 0) {
524                         rprintf(FERROR, "overflowed conversion buffer in mplex_write");
525                         exit_cleanup(RERR_UNSUPPORTED);
526                 }
527
528                 n = len = outbuf.len;
529         } else
530 #endif
531         if (n > 1024 - 4) /* BIGPATHBUFLEN can handle 1024 bytes */
532                 n = 0;    /* We'd rather do 2 writes than too much memcpy(). */
533         else
534                 memcpy(buffer + 4, buf, n);
535
536         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
537
538         defer_forwarding_keep = 1; /* defer_forwarding_messages++ on return */
539         writefd_unbuffered(fd, buffer, n+4);
540         defer_forwarding_keep = 0;
541
542         if (len > n)
543                 writefd_unbuffered(fd, buf+n, len-n);
544
545         if (!--defer_forwarding_messages && !no_flush)
546                 msg_flush();
547 }
548
549 int send_msg(enum msgcode code, const char *buf, int len, int convert)
550 {
551         if (msg_fd_out < 0) {
552                 if (!defer_forwarding_messages)
553                         return io_multiplex_write(code, buf, len, convert);
554                 if (!io_multiplexing_out)
555                         return 0;
556                 msg_list_add(&msg_queue, code, buf, len, convert);
557                 return 1;
558         }
559         if (flist_forward_from >= 0)
560                 msg_list_add(&msg_queue, code, buf, len, convert);
561         else
562                 mplex_write(msg_fd_out, code, buf, len, convert);
563         return 1;
564 }
565
566 void send_msg_int(enum msgcode code, int num)
567 {
568         char numbuf[4];
569         SIVAL(numbuf, 0, num);
570         send_msg(code, numbuf, 4, 0);
571 }
572
573 void wait_for_receiver(void)
574 {
575         if (iobuf_out_cnt)
576                 io_flush(NORMAL_FLUSH);
577         else
578                 read_msg_fd();
579 }
580
581 int get_redo_num(void)
582 {
583         return flist_ndx_pop(&redo_list);
584 }
585
586 int get_hlink_num(void)
587 {
588         return flist_ndx_pop(&hlink_list);
589 }
590
591 /**
592  * When we're the receiver and we have a local --files-from list of names
593  * that needs to be sent over the socket to the sender, we have to do two
594  * things at the same time: send the sender a list of what files we're
595  * processing and read the incoming file+info list from the sender.  We do
596  * this by augmenting the read_timeout() function to copy this data.  It
597  * uses ff_buf to read a block of data from f_in (when it is ready, since
598  * it might be a pipe) and then blast it out f_out (when it is ready to
599  * receive more data).
600  */
601 void io_set_filesfrom_fds(int f_in, int f_out)
602 {
603         io_filesfrom_f_in = f_in;
604         io_filesfrom_f_out = f_out;
605         alloc_xbuf(&ff_buf, 2048);
606 #ifdef ICONV_OPTION
607         if (protect_args)
608                 alloc_xbuf(&iconv_buf, 1024);
609 #endif
610 }
611
612 /* It's almost always an error to get an EOF when we're trying to read from the
613  * network, because the protocol is (for the most part) self-terminating.
614  *
615  * There is one case for the receiver when it is at the end of the transfer
616  * (hanging around reading any keep-alive packets that might come its way): if
617  * the sender dies before the generator's kill-signal comes through, we can end
618  * up here needing to loop until the kill-signal arrives.  In this situation,
619  * kluge_around_eof will be < 0.
620  *
621  * There is another case for older protocol versions (< 24) where the module
622  * listing was not terminated, so we must ignore an EOF error in that case and
623  * exit.  In this situation, kluge_around_eof will be > 0. */
624 static void whine_about_eof(int fd)
625 {
626         if (kluge_around_eof && fd == sock_f_in) {
627                 int i;
628                 if (kluge_around_eof > 0)
629                         exit_cleanup(0);
630                 /* If we're still here after 10 seconds, exit with an error. */
631                 for (i = 10*1000/20; i--; )
632                         msleep(20);
633         }
634
635         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
636                 "(%s bytes received so far) [%s]\n",
637                 big_num(stats.total_read, 0), who_am_i());
638
639         exit_cleanup(RERR_STREAMIO);
640 }
641
642 /**
643  * Read from a socket with I/O timeout. return the number of bytes
644  * read. If no bytes can be read then exit, never return a number <= 0.
645  *
646  * TODO: If the remote shell connection fails, then current versions
647  * actually report an "unexpected EOF" error here.  Since it's a
648  * fairly common mistake to try to use rsh when ssh is required, we
649  * should trap that: if we fail to read any data at all, we should
650  * give a better explanation.  We can tell whether the connection has
651  * started by looking e.g. at whether the remote version is known yet.
652  */
653 static int read_timeout(int fd, char *buf, size_t len)
654 {
655         int n, cnt = 0;
656
657         io_flush(FULL_FLUSH);
658
659         while (cnt == 0) {
660                 /* until we manage to read *something* */
661                 fd_set r_fds, w_fds;
662                 struct timeval tv;
663                 int maxfd = fd;
664                 int count;
665
666                 FD_ZERO(&r_fds);
667                 FD_ZERO(&w_fds);
668                 FD_SET(fd, &r_fds);
669                 if (io_filesfrom_f_out >= 0) {
670                         int new_fd;
671                         if (ff_buf.len == 0) {
672                                 if (io_filesfrom_f_in >= 0) {
673                                         FD_SET(io_filesfrom_f_in, &r_fds);
674                                         new_fd = io_filesfrom_f_in;
675                                 } else {
676                                         io_filesfrom_f_out = -1;
677                                         new_fd = -1;
678                                 }
679                         } else {
680                                 FD_SET(io_filesfrom_f_out, &w_fds);
681                                 new_fd = io_filesfrom_f_out;
682                         }
683                         if (new_fd > maxfd)
684                                 maxfd = new_fd;
685                 }
686
687                 tv.tv_sec = select_timeout;
688                 tv.tv_usec = 0;
689
690                 errno = 0;
691
692                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
693
694                 if (count <= 0) {
695                         if (errno == EBADF) {
696                                 defer_forwarding_messages = 0;
697                                 exit_cleanup(RERR_SOCKETIO);
698                         }
699                         check_timeout();
700                         continue;
701                 }
702
703                 if (io_filesfrom_f_out >= 0) {
704                         if (ff_buf.len) {
705                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
706                                         int l = write(io_filesfrom_f_out,
707                                                       ff_buf.buf + ff_buf.pos,
708                                                       ff_buf.len);
709                                         if (l > 0) {
710                                                 if (!(ff_buf.len -= l))
711                                                         ff_buf.pos = 0;
712                                                 else
713                                                         ff_buf.pos += l;
714                                         } else if (errno != EINTR) {
715                                                 /* XXX should we complain? */
716                                                 io_filesfrom_f_out = -1;
717                                         }
718                                 }
719                         } else if (io_filesfrom_f_in >= 0) {
720                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
721 #ifdef ICONV_OPTION
722                                         xbuf *ibuf = filesfrom_convert ? &iconv_buf : &ff_buf;
723 #else
724                                         xbuf *ibuf = &ff_buf;
725 #endif
726                                         int l = read(io_filesfrom_f_in, ibuf->buf, ibuf->size);
727                                         if (l <= 0) {
728                                                 if (l == 0 || errno != EINTR) {
729                                                         /* Send end-of-file marker */
730                                                         memcpy(ff_buf.buf, "\0\0", 2);
731                                                         ff_buf.len = ff_lastchar? 2 : 1;
732                                                         ff_buf.pos = 0;
733                                                         io_filesfrom_f_in = -1;
734                                                 }
735                                         } else {
736 #ifdef ICONV_OPTION
737                                                 if (filesfrom_convert) {
738                                                         iconv_buf.pos = 0;
739                                                         iconv_buf.len = l;
740                                                         iconvbufs(ic_send, &iconv_buf, &ff_buf,
741                                                             ICB_EXPAND_OUT|ICB_INCLUDE_BAD|ICB_INCLUDE_INCOMPLETE);
742                                                         l = ff_buf.len;
743                                                 }
744 #endif
745                                                 if (!eol_nulls) {
746                                                         char *s = ff_buf.buf + l;
747                                                         /* Transform CR and/or LF into '\0' */
748                                                         while (s-- > ff_buf.buf) {
749                                                                 if (*s == '\n' || *s == '\r')
750                                                                         *s = '\0';
751                                                         }
752                                                 }
753                                                 if (!ff_lastchar) {
754                                                         /* Last buf ended with a '\0', so don't
755                                                          * let this buf start with one. */
756                                                         while (l && ff_buf.buf[ff_buf.pos] == '\0')
757                                                                 ff_buf.pos++, l--;
758                                                 }
759                                                 if (!l)
760                                                         ff_buf.pos = 0;
761                                                 else {
762                                                         char *f = ff_buf.buf + ff_buf.pos;
763                                                         char *t = f;
764                                                         char *eob = f + l;
765                                                         /* Eliminate any multi-'\0' runs. */
766                                                         while (f != eob) {
767                                                                 if (!(*t++ = *f++)) {
768                                                                         while (f != eob && !*f)
769                                                                                 f++, l--;
770                                                                 }
771                                                         }
772                                                         ff_lastchar = f[-1];
773                                                 }
774                                                 ff_buf.len = l;
775                                         }
776                                 }
777                         }
778                 }
779
780                 if (!FD_ISSET(fd, &r_fds))
781                         continue;
782
783                 n = read(fd, buf, len);
784
785                 if (n <= 0) {
786                         if (n == 0)
787                                 whine_about_eof(fd); /* Doesn't return. */
788                         if (errno == EINTR || errno == EWOULDBLOCK
789                             || errno == EAGAIN)
790                                 continue;
791
792                         /* Don't write errors on a dead socket. */
793                         if (fd == sock_f_in) {
794                                 io_end_multiplex_out();
795                                 rsyserr(FERROR_SOCKET, errno, "read error");
796                         } else
797                                 rsyserr(FERROR, errno, "read error");
798                         exit_cleanup(RERR_STREAMIO);
799                 }
800
801                 buf += n;
802                 len -= n;
803                 cnt += n;
804
805                 if (fd == sock_f_in && io_timeout)
806                         last_io_in = time(NULL);
807         }
808
809         return cnt;
810 }
811
812 /* Read a line into the "buf" buffer. */
813 int read_line(int fd, char *buf, size_t bufsiz, int flags)
814 {
815         char ch, *s, *eob;
816         int cnt;
817
818 #ifdef ICONV_OPTION
819         if (flags & RL_CONVERT && iconv_buf.size < bufsiz)
820                 realloc_xbuf(&iconv_buf, bufsiz + 1024);
821 #endif
822
823   start:
824 #ifdef ICONV_OPTION
825         s = flags & RL_CONVERT ? iconv_buf.buf : buf;
826 #else
827         s = buf;
828 #endif
829         eob = s + bufsiz - 1;
830         while (1) {
831                 cnt = read(fd, &ch, 1);
832                 if (cnt < 0 && (errno == EWOULDBLOCK
833                   || errno == EINTR || errno == EAGAIN)) {
834                         struct timeval tv;
835                         fd_set r_fds, e_fds;
836                         FD_ZERO(&r_fds);
837                         FD_SET(fd, &r_fds);
838                         FD_ZERO(&e_fds);
839                         FD_SET(fd, &e_fds);
840                         tv.tv_sec = select_timeout;
841                         tv.tv_usec = 0;
842                         if (!select(fd+1, &r_fds, NULL, &e_fds, &tv))
843                                 check_timeout();
844                         /*if (FD_ISSET(fd, &e_fds))
845                                 rprintf(FINFO, "select exception on fd %d\n", fd); */
846                         continue;
847                 }
848                 if (cnt != 1)
849                         break;
850                 if (flags & RL_EOL_NULLS ? ch == '\0' : (ch == '\r' || ch == '\n')) {
851                         /* Skip empty lines if dumping comments. */
852                         if (flags & RL_DUMP_COMMENTS && s == buf)
853                                 continue;
854                         break;
855                 }
856                 if (s < eob)
857                         *s++ = ch;
858         }
859         *s = '\0';
860
861         if (flags & RL_DUMP_COMMENTS && (*buf == '#' || *buf == ';'))
862                 goto start;
863
864 #ifdef ICONV_OPTION
865         if (flags & RL_CONVERT) {
866                 xbuf outbuf;
867                 INIT_XBUF(outbuf, buf, 0, bufsiz);
868                 iconv_buf.pos = 0;
869                 iconv_buf.len = s - iconv_buf.buf;
870                 iconvbufs(ic_recv, &iconv_buf, &outbuf,
871                           ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE);
872                 outbuf.buf[outbuf.len] = '\0';
873                 return outbuf.len;
874         }
875 #endif
876
877         return s - buf;
878 }
879
880 void read_args(int f_in, char *mod_name, char *buf, size_t bufsiz, int rl_nulls,
881                char ***argv_p, int *argc_p, char **request_p)
882 {
883         int maxargs = MAX_ARGS;
884         int dot_pos = 0;
885         int argc = 0;
886         char **argv, *p;
887         int rl_flags = (rl_nulls ? RL_EOL_NULLS : 0);
888
889 #ifdef ICONV_OPTION
890         rl_flags |= (protect_args && ic_recv != (iconv_t)-1 ? RL_CONVERT : 0);
891 #endif
892
893         if (!(argv = new_array(char *, maxargs)))
894                 out_of_memory("read_args");
895         if (mod_name && !protect_args)
896                 argv[argc++] = "rsyncd";
897
898         while (1) {
899                 if (read_line(f_in, buf, bufsiz, rl_flags) == 0)
900                         break;
901
902                 if (argc == maxargs-1) {
903                         maxargs += MAX_ARGS;
904                         if (!(argv = realloc_array(argv, char *, maxargs)))
905                                 out_of_memory("read_args");
906                 }
907
908                 if (dot_pos) {
909                         if (request_p) {
910                                 *request_p = strdup(buf);
911                                 request_p = NULL;
912                         }
913                         if (mod_name)
914                                 glob_expand_module(mod_name, buf, &argv, &argc, &maxargs);
915                         else
916                                 glob_expand(buf, &argv, &argc, &maxargs);
917                 } else {
918                         if (!(p = strdup(buf)))
919                                 out_of_memory("read_args");
920                         argv[argc++] = p;
921                         if (*p == '.' && p[1] == '\0')
922                                 dot_pos = argc;
923                 }
924         }
925         argv[argc] = NULL;
926
927         glob_expand(NULL, NULL, NULL, NULL);
928
929         *argc_p = argc;
930         *argv_p = argv;
931 }
932
933 int io_start_buffering_out(int f_out)
934 {
935         if (iobuf_out) {
936                 assert(f_out == iobuf_f_out);
937                 return 0;
938         }
939         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
940                 out_of_memory("io_start_buffering_out");
941         iobuf_out_cnt = 0;
942         iobuf_f_out = f_out;
943         return 1;
944 }
945
946 int io_start_buffering_in(int f_in)
947 {
948         if (iobuf_in) {
949                 assert(f_in == iobuf_f_in);
950                 return 0;
951         }
952         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
953         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
954                 out_of_memory("io_start_buffering_in");
955         iobuf_f_in = f_in;
956         return 1;
957 }
958
959 void io_end_buffering_in(void)
960 {
961         if (!iobuf_in)
962                 return;
963         free(iobuf_in);
964         iobuf_in = NULL;
965         iobuf_in_ndx = 0;
966         iobuf_in_remaining = 0;
967         iobuf_f_in = -1;
968 }
969
970 void io_end_buffering_out(void)
971 {
972         if (!iobuf_out)
973                 return;
974         io_flush(FULL_FLUSH);
975         free(iobuf_out);
976         iobuf_out = NULL;
977         iobuf_f_out = -1;
978 }
979
980 void maybe_flush_socket(int important)
981 {
982         if (iobuf_out && iobuf_out_cnt
983          && (important || time(NULL) - last_io_out >= 5))
984                 io_flush(NORMAL_FLUSH);
985 }
986
987 void maybe_send_keepalive(void)
988 {
989         if (time(NULL) - last_io_out >= allowed_lull) {
990                 if (!iobuf_out || !iobuf_out_cnt) {
991                         if (protocol_version < 29)
992                                 return; /* there's nothing we can do */
993                         if (protocol_version >= 30)
994                                 send_msg(MSG_NOOP, "", 0, 0);
995                         else {
996                                 write_int(sock_f_out, cur_flist->used);
997                                 write_shortint(sock_f_out, ITEM_IS_NEW);
998                         }
999                 }
1000                 if (iobuf_out)
1001                         io_flush(NORMAL_FLUSH);
1002         }
1003 }
1004
1005 void start_flist_forward(int f_in)
1006 {
1007         assert(iobuf_out != NULL);
1008         assert(iobuf_f_out == msg_fd_out);
1009         flist_forward_from = f_in;
1010 }
1011
1012 void stop_flist_forward()
1013 {
1014         flist_forward_from = -1;
1015         io_flush(FULL_FLUSH);
1016 }
1017
1018 /**
1019  * Continue trying to read len bytes - don't return until len has been
1020  * read.
1021  **/
1022 static void read_loop(int fd, char *buf, size_t len)
1023 {
1024         while (len) {
1025                 int n = read_timeout(fd, buf, len);
1026
1027                 buf += n;
1028                 len -= n;
1029         }
1030 }
1031
1032 /**
1033  * Read from the file descriptor handling multiplexing - return number
1034  * of bytes read.
1035  *
1036  * Never returns <= 0.
1037  */
1038 static int readfd_unbuffered(int fd, char *buf, size_t len)
1039 {
1040         size_t msg_bytes;
1041         int tag, cnt = 0;
1042         char line[BIGPATHBUFLEN];
1043
1044         if (!iobuf_in || fd != iobuf_f_in)
1045                 return read_timeout(fd, buf, len);
1046
1047         if (!io_multiplexing_in && iobuf_in_remaining == 0) {
1048                 iobuf_in_remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
1049                 iobuf_in_ndx = 0;
1050         }
1051
1052         while (cnt == 0) {
1053                 if (iobuf_in_remaining) {
1054                         len = MIN(len, iobuf_in_remaining);
1055                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
1056                         iobuf_in_ndx += len;
1057                         iobuf_in_remaining -= len;
1058                         cnt = len;
1059                         break;
1060                 }
1061
1062                 read_loop(fd, line, 4);
1063                 tag = IVAL(line, 0);
1064
1065                 msg_bytes = tag & 0xFFFFFF;
1066                 tag = (tag >> 24) - MPLEX_BASE;
1067
1068                 check_for_io_err = 0;
1069
1070                 switch (tag) {
1071                 case MSG_DATA:
1072                         if (msg_bytes > iobuf_in_siz) {
1073                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
1074                                                                msg_bytes)))
1075                                         out_of_memory("readfd_unbuffered");
1076                                 iobuf_in_siz = msg_bytes;
1077                         }
1078                         read_loop(fd, iobuf_in, msg_bytes);
1079                         iobuf_in_remaining = msg_bytes;
1080                         iobuf_in_ndx = 0;
1081                         break;
1082                 case MSG_NOOP:
1083                         if (am_sender)
1084                                 maybe_send_keepalive();
1085                         break;
1086                 case MSG_IO_ERROR:
1087                         if (msg_bytes != 4)
1088                                 goto invalid_msg;
1089                         read_loop(fd, line, msg_bytes);
1090                         send_msg_int(MSG_IO_ERROR, IVAL(line, 0));
1091                         io_error |= IVAL(line, 0);
1092                         break;
1093                 case MSG_DELETED:
1094                         if (msg_bytes >= sizeof line)
1095                                 goto overflow;
1096 #ifdef ICONV_OPTION
1097                         if (ic_recv != (iconv_t)-1) {
1098                                 xbuf outbuf, inbuf;
1099                                 char ibuf[512];
1100                                 int add_null = 0;
1101
1102                                 INIT_CONST_XBUF(outbuf, line);
1103                                 INIT_XBUF(inbuf, ibuf, 0, (size_t)-1);
1104
1105                                 while (msg_bytes) {
1106                                         inbuf.len = msg_bytes > sizeof ibuf
1107                                                   ? sizeof ibuf : msg_bytes;
1108                                         read_loop(fd, inbuf.buf, inbuf.len);
1109                                         if (!(msg_bytes -= inbuf.len)
1110                                          && !ibuf[inbuf.len-1])
1111                                                 inbuf.len--, add_null = 1;
1112                                         if (iconvbufs(ic_send, &inbuf, &outbuf,
1113                                             ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE) < 0)
1114                                                 goto overflow;
1115                                 }
1116                                 if (add_null) {
1117                                         if (outbuf.len == outbuf.size)
1118                                                 goto overflow;
1119                                         outbuf.buf[outbuf.len++] = '\0';
1120                                 }
1121                                 msg_bytes = outbuf.len;
1122                         } else
1123 #endif
1124                                 read_loop(fd, line, msg_bytes);
1125                         /* A directory name was sent with the trailing null */
1126                         if (msg_bytes > 0 && !line[msg_bytes-1])
1127                                 log_delete(line, S_IFDIR);
1128                         else {
1129                                 line[msg_bytes] = '\0';
1130                                 log_delete(line, S_IFREG);
1131                         }
1132                         break;
1133                 case MSG_SUCCESS:
1134                         if (msg_bytes != 4) {
1135                           invalid_msg:
1136                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
1137                                         tag, (long)msg_bytes, who_am_i());
1138                                 exit_cleanup(RERR_STREAMIO);
1139                         }
1140                         read_loop(fd, line, msg_bytes);
1141                         successful_send(IVAL(line, 0));
1142                         break;
1143                 case MSG_NO_SEND:
1144                         if (msg_bytes != 4)
1145                                 goto invalid_msg;
1146                         read_loop(fd, line, msg_bytes);
1147                         send_msg_int(MSG_NO_SEND, IVAL(line, 0));
1148                         break;
1149                 case MSG_INFO:
1150                 case MSG_ERROR:
1151                 case MSG_ERROR_XFER:
1152                 case MSG_WARNING:
1153                         if (msg_bytes >= sizeof line) {
1154                             overflow:
1155                                 rprintf(FERROR,
1156                                         "multiplexing overflow %d:%ld [%s]\n",
1157                                         tag, (long)msg_bytes, who_am_i());
1158                                 exit_cleanup(RERR_STREAMIO);
1159                         }
1160                         read_loop(fd, line, msg_bytes);
1161                         rwrite((enum logcode)tag, line, msg_bytes, 1);
1162                         if (first_message) {
1163                                 if (list_only && !am_sender && tag == 1) {
1164                                         line[msg_bytes] = '\0';
1165                                         check_for_d_option_error(line);
1166                                 }
1167                                 first_message = 0;
1168                         }
1169                         break;
1170                 default:
1171                         rprintf(FERROR, "unexpected tag %d [%s]\n",
1172                                 tag, who_am_i());
1173                         exit_cleanup(RERR_STREAMIO);
1174                 }
1175         }
1176
1177         if (iobuf_in_remaining == 0)
1178                 io_flush(NORMAL_FLUSH);
1179
1180         return cnt;
1181 }
1182
1183 /* Do a buffered read from fd.  Don't return until all N bytes have
1184  * been read.  If all N can't be read then exit with an error. */
1185 static void readfd(int fd, char *buffer, size_t N)
1186 {
1187         int  cnt;
1188         size_t total = 0;
1189
1190         while (total < N) {
1191                 cnt = readfd_unbuffered(fd, buffer + total, N-total);
1192                 total += cnt;
1193         }
1194
1195         if (fd == write_batch_monitor_in) {
1196                 if ((size_t)write(batch_fd, buffer, total) != total)
1197                         exit_cleanup(RERR_FILEIO);
1198         }
1199
1200         if (fd == flist_forward_from)
1201                 writefd(iobuf_f_out, buffer, total);
1202
1203         if (fd == sock_f_in)
1204                 stats.total_read += total;
1205 }
1206
1207 unsigned short read_shortint(int f)
1208 {
1209         char b[2];
1210         readfd(f, b, 2);
1211         return (UVAL(b, 1) << 8) + UVAL(b, 0);
1212 }
1213
1214 int32 read_int(int f)
1215 {
1216         char b[4];
1217         int32 num;
1218
1219         readfd(f, b, 4);
1220         num = IVAL(b, 0);
1221 #if SIZEOF_INT32 > 4
1222         if (num & (int32)0x80000000)
1223                 num |= ~(int32)0xffffffff;
1224 #endif
1225         return num;
1226 }
1227
1228 int32 read_varint(int f)
1229 {
1230         union {
1231             char b[5];
1232             int32 x;
1233         } u;
1234         uchar ch;
1235         int extra;
1236
1237         u.x = 0;
1238         readfd(f, (char*)&ch, 1);
1239         extra = int_byte_extra[ch / 4];
1240         if (extra) {
1241                 uchar bit = ((uchar)1<<(8-extra));
1242                 if (extra >= (int)sizeof u.b) {
1243                         rprintf(FERROR, "Overflow in read_varint()\n");
1244                         exit_cleanup(RERR_STREAMIO);
1245                 }
1246                 readfd(f, u.b, extra);
1247                 u.b[extra] = ch & (bit-1);
1248         } else
1249                 u.b[0] = ch;
1250 #if CAREFUL_ALIGNMENT
1251         u.x = IVAL(u.b,0);
1252 #endif
1253 #if SIZEOF_INT32 > 4
1254         if (u.x & (int32)0x80000000)
1255                 u.x |= ~(int32)0xffffffff;
1256 #endif
1257         return u.x;
1258 }
1259
1260 int64 read_varlong(int f, uchar min_bytes)
1261 {
1262         union {
1263             char b[9];
1264             int64 x;
1265         } u;
1266         char b2[8];
1267         int extra;
1268
1269 #if SIZEOF_INT64 < 8
1270         memset(u.b, 0, 8);
1271 #else
1272         u.x = 0;
1273 #endif
1274         readfd(f, b2, min_bytes);
1275         memcpy(u.b, b2+1, min_bytes-1);
1276         extra = int_byte_extra[CVAL(b2, 0) / 4];
1277         if (extra) {
1278                 uchar bit = ((uchar)1<<(8-extra));
1279                 if (min_bytes + extra > (int)sizeof u.b) {
1280                         rprintf(FERROR, "Overflow in read_varlong()\n");
1281                         exit_cleanup(RERR_STREAMIO);
1282                 }
1283                 readfd(f, u.b + min_bytes - 1, extra);
1284                 u.b[min_bytes + extra - 1] = CVAL(b2, 0) & (bit-1);
1285 #if SIZEOF_INT64 < 8
1286                 if (min_bytes + extra > 5 || u.b[4] || CVAL(u.b,3) & 0x80) {
1287                         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1288                         exit_cleanup(RERR_UNSUPPORTED);
1289                 }
1290 #endif
1291         } else
1292                 u.b[min_bytes + extra - 1] = CVAL(b2, 0);
1293 #if SIZEOF_INT64 < 8
1294         u.x = IVAL(u.b,0);
1295 #elif CAREFUL_ALIGNMENT
1296         u.x = IVAL(u.b,0) | (((int64)IVAL(u.b,4))<<32);
1297 #endif
1298         return u.x;
1299 }
1300
1301 int64 read_longint(int f)
1302 {
1303 #if SIZEOF_INT64 >= 8
1304         char b[9];
1305 #endif
1306         int32 num = read_int(f);
1307
1308         if (num != (int32)0xffffffff)
1309                 return num;
1310
1311 #if SIZEOF_INT64 < 8
1312         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1313         exit_cleanup(RERR_UNSUPPORTED);
1314 #else
1315         readfd(f, b, 8);
1316         return IVAL(b,0) | (((int64)IVAL(b,4))<<32);
1317 #endif
1318 }
1319
1320 void read_buf(int f, char *buf, size_t len)
1321 {
1322         readfd(f,buf,len);
1323 }
1324
1325 void read_sbuf(int f, char *buf, size_t len)
1326 {
1327         readfd(f, buf, len);
1328         buf[len] = '\0';
1329 }
1330
1331 uchar read_byte(int f)
1332 {
1333         uchar c;
1334         readfd(f, (char *)&c, 1);
1335         return c;
1336 }
1337
1338 int read_vstring(int f, char *buf, int bufsize)
1339 {
1340         int len = read_byte(f);
1341
1342         if (len & 0x80)
1343                 len = (len & ~0x80) * 0x100 + read_byte(f);
1344
1345         if (len >= bufsize) {
1346                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
1347                         len, bufsize - 1);
1348                 return -1;
1349         }
1350
1351         if (len)
1352                 readfd(f, buf, len);
1353         buf[len] = '\0';
1354         return len;
1355 }
1356
1357 /* Populate a sum_struct with values from the socket.  This is
1358  * called by both the sender and the receiver. */
1359 void read_sum_head(int f, struct sum_struct *sum)
1360 {
1361         int32 max_blength = protocol_version < 30 ? OLD_MAX_BLOCK_SIZE : MAX_BLOCK_SIZE;
1362         sum->count = read_int(f);
1363         if (sum->count < 0) {
1364                 rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
1365                         (long)sum->count, who_am_i());
1366                 exit_cleanup(RERR_PROTOCOL);
1367         }
1368         sum->blength = read_int(f);
1369         if (sum->blength < 0 || sum->blength > max_blength) {
1370                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
1371                         (long)sum->blength, who_am_i());
1372                 exit_cleanup(RERR_PROTOCOL);
1373         }
1374         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
1375         if (sum->s2length < 0 || sum->s2length > MAX_DIGEST_LEN) {
1376                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
1377                         sum->s2length, who_am_i());
1378                 exit_cleanup(RERR_PROTOCOL);
1379         }
1380         sum->remainder = read_int(f);
1381         if (sum->remainder < 0 || sum->remainder > sum->blength) {
1382                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
1383                         (long)sum->remainder, who_am_i());
1384                 exit_cleanup(RERR_PROTOCOL);
1385         }
1386 }
1387
1388 /* Send the values from a sum_struct over the socket.  Set sum to
1389  * NULL if there are no checksums to send.  This is called by both
1390  * the generator and the sender. */
1391 void write_sum_head(int f, struct sum_struct *sum)
1392 {
1393         static struct sum_struct null_sum;
1394
1395         if (sum == NULL)
1396                 sum = &null_sum;
1397
1398         write_int(f, sum->count);
1399         write_int(f, sum->blength);
1400         if (protocol_version >= 27)
1401                 write_int(f, sum->s2length);
1402         write_int(f, sum->remainder);
1403 }
1404
1405 /**
1406  * Sleep after writing to limit I/O bandwidth usage.
1407  *
1408  * @todo Rather than sleeping after each write, it might be better to
1409  * use some kind of averaging.  The current algorithm seems to always
1410  * use a bit less bandwidth than specified, because it doesn't make up
1411  * for slow periods.  But arguably this is a feature.  In addition, we
1412  * ought to take the time used to write the data into account.
1413  *
1414  * During some phases of big transfers (file FOO is uptodate) this is
1415  * called with a small bytes_written every time.  As the kernel has to
1416  * round small waits up to guarantee that we actually wait at least the
1417  * requested number of microseconds, this can become grossly inaccurate.
1418  * We therefore keep track of the bytes we've written over time and only
1419  * sleep when the accumulated delay is at least 1 tenth of a second.
1420  **/
1421 static void sleep_for_bwlimit(int bytes_written)
1422 {
1423         static struct timeval prior_tv;
1424         static long total_written = 0;
1425         struct timeval tv, start_tv;
1426         long elapsed_usec, sleep_usec;
1427
1428 #define ONE_SEC 1000000L /* # of microseconds in a second */
1429
1430         if (!bwlimit_writemax)
1431                 return;
1432
1433         total_written += bytes_written;
1434
1435         gettimeofday(&start_tv, NULL);
1436         if (prior_tv.tv_sec) {
1437                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
1438                              + (start_tv.tv_usec - prior_tv.tv_usec);
1439                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
1440                 if (total_written < 0)
1441                         total_written = 0;
1442         }
1443
1444         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
1445         if (sleep_usec < ONE_SEC / 10) {
1446                 prior_tv = start_tv;
1447                 return;
1448         }
1449
1450         tv.tv_sec  = sleep_usec / ONE_SEC;
1451         tv.tv_usec = sleep_usec % ONE_SEC;
1452         select(0, NULL, NULL, NULL, &tv);
1453
1454         gettimeofday(&prior_tv, NULL);
1455         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
1456                      + (prior_tv.tv_usec - start_tv.tv_usec);
1457         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
1458 }
1459
1460 /* Write len bytes to the file descriptor fd, looping as necessary to get
1461  * the job done and also (in certain circumstances) reading any data on
1462  * msg_fd_in to avoid deadlock.
1463  *
1464  * This function underlies the multiplexing system.  The body of the
1465  * application never calls this function directly. */
1466 static void writefd_unbuffered(int fd, const char *buf, size_t len)
1467 {
1468         size_t n, total = 0;
1469         fd_set w_fds, r_fds, e_fds;
1470         int maxfd, count, cnt, using_r_fds;
1471         int defer_inc = 0;
1472         struct timeval tv;
1473
1474         if (no_flush++)
1475                 defer_forwarding_messages++, defer_inc++;
1476
1477         while (total < len) {
1478                 FD_ZERO(&w_fds);
1479                 FD_SET(fd, &w_fds);
1480                 FD_ZERO(&e_fds);
1481                 FD_SET(fd, &e_fds);
1482                 maxfd = fd;
1483
1484                 if (msg_fd_in >= 0) {
1485                         FD_ZERO(&r_fds);
1486                         FD_SET(msg_fd_in, &r_fds);
1487                         if (msg_fd_in > maxfd)
1488                                 maxfd = msg_fd_in;
1489                         using_r_fds = 1;
1490                 } else
1491                         using_r_fds = 0;
1492
1493                 tv.tv_sec = select_timeout;
1494                 tv.tv_usec = 0;
1495
1496                 errno = 0;
1497                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1498                                &w_fds, &e_fds, &tv);
1499
1500                 if (count <= 0) {
1501                         if (count < 0 && errno == EBADF)
1502                                 exit_cleanup(RERR_SOCKETIO);
1503                         check_timeout();
1504                         continue;
1505                 }
1506
1507                 /*if (FD_ISSET(fd, &e_fds))
1508                         rprintf(FINFO, "select exception on fd %d\n", fd); */
1509
1510                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1511                         read_msg_fd();
1512
1513                 if (!FD_ISSET(fd, &w_fds))
1514                         continue;
1515
1516                 n = len - total;
1517                 if (bwlimit_writemax && n > bwlimit_writemax)
1518                         n = bwlimit_writemax;
1519                 cnt = write(fd, buf + total, n);
1520
1521                 if (cnt <= 0) {
1522                         if (cnt < 0) {
1523                                 if (errno == EINTR)
1524                                         continue;
1525                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1526                                         msleep(1);
1527                                         continue;
1528                                 }
1529                         }
1530
1531                         /* Don't try to write errors back across the stream. */
1532                         if (fd == sock_f_out)
1533                                 io_end_multiplex_out();
1534                         /* Don't try to write errors down a failing msg pipe. */
1535                         if (am_server && fd == msg_fd_out)
1536                                 exit_cleanup(RERR_STREAMIO);
1537                         rsyserr(FERROR, errno,
1538                                 "writefd_unbuffered failed to write %ld bytes [%s]",
1539                                 (long)len, who_am_i());
1540                         /* If the other side is sending us error messages, try
1541                          * to grab any messages they sent before they died. */
1542                         while (!am_server && fd == sock_f_out && io_multiplexing_in) {
1543                                 char buf[1024];
1544                                 set_io_timeout(30);
1545                                 ignore_timeout = 0;
1546                                 readfd_unbuffered(sock_f_in, buf, sizeof buf);
1547                         }
1548                         exit_cleanup(RERR_STREAMIO);
1549                 }
1550
1551                 total += cnt;
1552                 defer_forwarding_messages++, defer_inc++;
1553
1554                 if (fd == sock_f_out) {
1555                         if (io_timeout || am_generator)
1556                                 last_io_out = time(NULL);
1557                         sleep_for_bwlimit(cnt);
1558                 }
1559         }
1560
1561         no_flush--;
1562         defer_inc -= defer_forwarding_keep;
1563         if (!(defer_forwarding_messages -= defer_inc) && !no_flush)
1564                 msg_flush();
1565 }
1566
1567 void io_flush(int flush_it_all)
1568 {
1569         if (!iobuf_out_cnt || no_flush)
1570                 return;
1571
1572         if (io_multiplexing_out)
1573                 mplex_write(sock_f_out, MSG_DATA, iobuf_out, iobuf_out_cnt, 0);
1574         else
1575                 writefd_unbuffered(iobuf_f_out, iobuf_out, iobuf_out_cnt);
1576         iobuf_out_cnt = 0;
1577
1578         if (flush_it_all && !defer_forwarding_messages)
1579                 msg_flush();
1580 }
1581
1582 static void writefd(int fd, const char *buf, size_t len)
1583 {
1584         if (fd == sock_f_out)
1585                 stats.total_written += len;
1586
1587         if (fd == write_batch_monitor_out) {
1588                 if ((size_t)write(batch_fd, buf, len) != len)
1589                         exit_cleanup(RERR_FILEIO);
1590         }
1591
1592         if (!iobuf_out || fd != iobuf_f_out) {
1593                 writefd_unbuffered(fd, buf, len);
1594                 return;
1595         }
1596
1597         while (len) {
1598                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1599                 if (n > 0) {
1600                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1601                         buf += n;
1602                         len -= n;
1603                         iobuf_out_cnt += n;
1604                 }
1605
1606                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1607                         io_flush(NORMAL_FLUSH);
1608         }
1609 }
1610
1611 void write_shortint(int f, unsigned short x)
1612 {
1613         char b[2];
1614         b[0] = (char)x;
1615         b[1] = (char)(x >> 8);
1616         writefd(f, b, 2);
1617 }
1618
1619 void write_int(int f, int32 x)
1620 {
1621         char b[4];
1622         SIVAL(b, 0, x);
1623         writefd(f, b, 4);
1624 }
1625
1626 void write_varint(int f, int32 x)
1627 {
1628         char b[5];
1629         uchar bit;
1630         int cnt = 4;
1631
1632         SIVAL(b, 1, x);
1633
1634         while (cnt > 1 && b[cnt] == 0)
1635                 cnt--;
1636         bit = ((uchar)1<<(7-cnt+1));
1637         if (CVAL(b, cnt) >= bit) {
1638                 cnt++;
1639                 *b = ~(bit-1);
1640         } else if (cnt > 1)
1641                 *b = b[cnt] | ~(bit*2-1);
1642         else
1643                 *b = b[cnt];
1644
1645         writefd(f, b, cnt);
1646 }
1647
1648 void write_varlong(int f, int64 x, uchar min_bytes)
1649 {
1650         char b[9];
1651         uchar bit;
1652         int cnt = 8;
1653
1654         SIVAL(b, 1, x);
1655 #if SIZEOF_INT64 >= 8
1656         SIVAL(b, 5, x >> 32);
1657 #else
1658         if (x <= 0x7FFFFFFF && x >= 0)
1659                 memset(b + 5, 0, 4);
1660         else {
1661                 rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1662                 exit_cleanup(RERR_UNSUPPORTED);
1663         }
1664 #endif
1665
1666         while (cnt > min_bytes && b[cnt] == 0)
1667                 cnt--;
1668         bit = ((uchar)1<<(7-cnt+min_bytes));
1669         if (CVAL(b, cnt) >= bit) {
1670                 cnt++;
1671                 *b = ~(bit-1);
1672         } else if (cnt > min_bytes)
1673                 *b = b[cnt] | ~(bit*2-1);
1674         else
1675                 *b = b[cnt];
1676
1677         writefd(f, b, cnt);
1678 }
1679
1680 /*
1681  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1682  * 64-bit types on this platform.
1683  */
1684 void write_longint(int f, int64 x)
1685 {
1686         char b[12], * const s = b+4;
1687
1688         SIVAL(s, 0, x);
1689         if (x <= 0x7FFFFFFF && x >= 0) {
1690                 writefd(f, s, 4);
1691                 return;
1692         }
1693
1694 #if SIZEOF_INT64 < 8
1695         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1696         exit_cleanup(RERR_UNSUPPORTED);
1697 #else
1698         memset(b, 0xFF, 4);
1699         SIVAL(s, 4, x >> 32);
1700         writefd(f, b, 12);
1701 #endif
1702 }
1703
1704 void write_buf(int f, const char *buf, size_t len)
1705 {
1706         writefd(f,buf,len);
1707 }
1708
1709 /** Write a string to the connection */
1710 void write_sbuf(int f, const char *buf)
1711 {
1712         writefd(f, buf, strlen(buf));
1713 }
1714
1715 void write_byte(int f, uchar c)
1716 {
1717         writefd(f, (char *)&c, 1);
1718 }
1719
1720 void write_vstring(int f, const char *str, int len)
1721 {
1722         uchar lenbuf[3], *lb = lenbuf;
1723
1724         if (len > 0x7F) {
1725                 if (len > 0x7FFF) {
1726                         rprintf(FERROR,
1727                                 "attempting to send over-long vstring (%d > %d)\n",
1728                                 len, 0x7FFF);
1729                         exit_cleanup(RERR_PROTOCOL);
1730                 }
1731                 *lb++ = len / 0x100 + 0x80;
1732         }
1733         *lb = len;
1734
1735         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1736         if (len)
1737                 writefd(f, str, len);
1738 }
1739
1740 /* Send a file-list index using a byte-reduction method. */
1741 void write_ndx(int f, int32 ndx)
1742 {
1743         static int32 prev_positive = -1, prev_negative = 1;
1744         int32 diff, cnt = 0;
1745         char b[6];
1746
1747         if (protocol_version < 30 || read_batch) {
1748                 write_int(f, ndx);
1749                 return;
1750         }
1751
1752         /* Send NDX_DONE as a single-byte 0 with no side effects.  Send
1753          * negative nums as a positive after sending a leading 0xFF. */
1754         if (ndx >= 0) {
1755                 diff = ndx - prev_positive;
1756                 prev_positive = ndx;
1757         } else if (ndx == NDX_DONE) {
1758                 *b = 0;
1759                 writefd(f, b, 1);
1760                 return;
1761         } else {
1762                 b[cnt++] = (char)0xFF;
1763                 ndx = -ndx;
1764                 diff = ndx - prev_negative;
1765                 prev_negative = ndx;
1766         }
1767
1768         /* A diff of 1 - 253 is sent as a one-byte diff; a diff of 254 - 32767
1769          * or 0 is sent as a 0xFE + a two-byte diff; otherwise we send 0xFE
1770          * & all 4 bytes of the (non-negative) num with the high-bit set. */
1771         if (diff < 0xFE && diff > 0)
1772                 b[cnt++] = (char)diff;
1773         else if (diff < 0 || diff > 0x7FFF) {
1774                 b[cnt++] = (char)0xFE;
1775                 b[cnt++] = (char)((ndx >> 24) | 0x80);
1776                 b[cnt++] = (char)ndx;
1777                 b[cnt++] = (char)(ndx >> 8);
1778                 b[cnt++] = (char)(ndx >> 16);
1779         } else {
1780                 b[cnt++] = (char)0xFE;
1781                 b[cnt++] = (char)(diff >> 8);
1782                 b[cnt++] = (char)diff;
1783         }
1784         writefd(f, b, cnt);
1785 }
1786
1787 /* Receive a file-list index using a byte-reduction method. */
1788 int32 read_ndx(int f)
1789 {
1790         static int32 prev_positive = -1, prev_negative = 1;
1791         int32 *prev_ptr, num;
1792         char b[4];
1793
1794         if (protocol_version < 30)
1795                 return read_int(f);
1796
1797         readfd(f, b, 1);
1798         if (CVAL(b, 0) == 0xFF) {
1799                 readfd(f, b, 1);
1800                 prev_ptr = &prev_negative;
1801         } else if (CVAL(b, 0) == 0)
1802                 return NDX_DONE;
1803         else
1804                 prev_ptr = &prev_positive;
1805         if (CVAL(b, 0) == 0xFE) {
1806                 readfd(f, b, 2);
1807                 if (CVAL(b, 0) & 0x80) {
1808                         b[3] = CVAL(b, 0) & ~0x80;
1809                         b[0] = b[1];
1810                         readfd(f, b+1, 2);
1811                         num = IVAL(b, 0);
1812                 } else
1813                         num = (UVAL(b,0)<<8) + UVAL(b,1) + *prev_ptr;
1814         } else
1815                 num = UVAL(b, 0) + *prev_ptr;
1816         *prev_ptr = num;
1817         if (prev_ptr == &prev_negative)
1818                 num = -num;
1819         return num;
1820 }
1821
1822 /* Read a line of up to bufsiz-1 characters into buf.  Strips
1823  * the (required) trailing newline and all carriage returns.
1824  * Returns 1 for success; 0 for I/O error or truncation. */
1825 int read_line_old(int f, char *buf, size_t bufsiz)
1826 {
1827         bufsiz--; /* leave room for the null */
1828         while (bufsiz > 0) {
1829                 buf[0] = 0;
1830                 read_buf(f, buf, 1);
1831                 if (buf[0] == 0)
1832                         return 0;
1833                 if (buf[0] == '\n')
1834                         break;
1835                 if (buf[0] != '\r') {
1836                         buf++;
1837                         bufsiz--;
1838                 }
1839         }
1840         *buf = '\0';
1841         return bufsiz > 0;
1842 }
1843
1844 void io_printf(int fd, const char *format, ...)
1845 {
1846         va_list ap;
1847         char buf[BIGPATHBUFLEN];
1848         int len;
1849
1850         va_start(ap, format);
1851         len = vsnprintf(buf, sizeof buf, format, ap);
1852         va_end(ap);
1853
1854         if (len < 0)
1855                 exit_cleanup(RERR_STREAMIO);
1856
1857         if (len > (int)sizeof buf) {
1858                 rprintf(FERROR, "io_printf() was too long for the buffer.\n");
1859                 exit_cleanup(RERR_STREAMIO);
1860         }
1861
1862         write_sbuf(fd, buf);
1863 }
1864
1865 /** Setup for multiplexing a MSG_* stream with the data stream. */
1866 void io_start_multiplex_out(void)
1867 {
1868         io_flush(NORMAL_FLUSH);
1869         io_start_buffering_out(sock_f_out);
1870         io_multiplexing_out = 1;
1871 }
1872
1873 /** Setup for multiplexing a MSG_* stream with the data stream. */
1874 void io_start_multiplex_in(void)
1875 {
1876         io_flush(NORMAL_FLUSH);
1877         io_start_buffering_in(sock_f_in);
1878         io_multiplexing_in = 1;
1879 }
1880
1881 /** Write an message to the multiplexed data stream. */
1882 int io_multiplex_write(enum msgcode code, const char *buf, size_t len, int convert)
1883 {
1884         if (!io_multiplexing_out)
1885                 return 0;
1886         io_flush(NORMAL_FLUSH);
1887         stats.total_written += (len+4);
1888         mplex_write(sock_f_out, code, buf, len, convert);
1889         return 1;
1890 }
1891
1892 void io_end_multiplex_in(void)
1893 {
1894         io_multiplexing_in = 0;
1895         io_end_buffering_in();
1896 }
1897
1898 /** Stop output multiplexing. */
1899 void io_end_multiplex_out(void)
1900 {
1901         io_multiplexing_out = 0;
1902         io_end_buffering_out();
1903 }
1904
1905 void start_write_batch(int fd)
1906 {
1907         /* Some communication has already taken place, but we don't
1908          * enable batch writing until here so that we can write a
1909          * canonical record of the communication even though the
1910          * actual communication so far depends on whether a daemon
1911          * is involved. */
1912         write_int(batch_fd, protocol_version);
1913         if (protocol_version >= 30)
1914                 write_byte(batch_fd, inc_recurse);
1915         write_int(batch_fd, checksum_seed);
1916
1917         if (am_sender)
1918                 write_batch_monitor_out = fd;
1919         else
1920                 write_batch_monitor_in = fd;
1921 }
1922
1923 void stop_write_batch(void)
1924 {
1925         write_batch_monitor_out = -1;
1926         write_batch_monitor_in = -1;
1927 }