Defer forwarding messages during the forwarding of flist data.
[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) 2003-2009 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 allowed_lull;
40 extern int am_server;
41 extern int am_daemon;
42 extern int am_sender;
43 extern int am_receiver;
44 extern int am_generator;
45 extern int inc_recurse;
46 extern int io_error;
47 extern int eol_nulls;
48 extern int flist_eof;
49 extern int list_only;
50 extern int read_batch;
51 extern int csum_length;
52 extern int protect_args;
53 extern int checksum_seed;
54 extern int protocol_version;
55 extern int remove_source_files;
56 extern int preserve_hard_links;
57 extern struct stats stats;
58 extern struct file_list *cur_flist;
59 #ifdef ICONV_OPTION
60 extern int filesfrom_convert;
61 extern iconv_t ic_send, ic_recv;
62 #endif
63
64 const char phase_unknown[] = "unknown";
65 int batch_fd = -1;
66 int msgdone_cnt = 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, keep_defer_forwarding = 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 check_timeout(void)
123 {
124         time_t t, chk;
125
126         /* On the receiving side, the generator is now handling timeouts, so
127          * the receiver ignores them.  Note that the am_receiver flag is not
128          * set until the receiver forks from the generator, so timeouts will be
129          * based on receiving data on the receiving side until that event. */
130         if (!io_timeout || am_receiver)
131                 return;
132
133         t = time(NULL);
134
135         if (!last_io_in)
136                 last_io_in = t;
137
138         chk = MAX(last_io_out, last_io_in);
139         if (t - chk >= io_timeout) {
140                 if (am_server || am_daemon)
141                         exit_cleanup(RERR_TIMEOUT);
142                 rprintf(FERROR, "[%s] io timeout after %d seconds -- exiting\n",
143                         who_am_i(), (int)(t-chk));
144                 exit_cleanup(RERR_TIMEOUT);
145         }
146 }
147
148 static void readfd(int fd, char *buffer, size_t N);
149 static void writefd(int fd, const char *buf, size_t len);
150 static void writefd_unbuffered(int fd, const char *buf, size_t len);
151 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert);
152
153 static flist_ndx_list redo_list, hlink_list;
154
155 struct msg_list_item {
156         struct msg_list_item *next;
157         char convert;
158         char buf[1];
159 };
160
161 struct msg_list {
162         struct msg_list_item *head, *tail;
163 };
164
165 static struct msg_list msg_queue;
166
167 static void got_flist_entry_status(enum festatus status, const char *buf)
168 {
169         int ndx = IVAL(buf, 0);
170         struct file_list *flist = flist_for_ndx(ndx, "got_flist_entry_status");
171
172         if (remove_source_files) {
173                 active_filecnt--;
174                 active_bytecnt -= F_LENGTH(flist->files[ndx - flist->ndx_start]);
175         }
176
177         if (inc_recurse)
178                 flist->in_progress--;
179
180         switch (status) {
181         case FES_SUCCESS:
182                 if (remove_source_files)
183                         send_msg(MSG_SUCCESS, buf, 4, 0);
184                 if (preserve_hard_links) {
185                         struct file_struct *file = flist->files[ndx - flist->ndx_start];
186                         if (F_IS_HLINKED(file)) {
187                                 flist_ndx_push(&hlink_list, ndx);
188                                 flist->in_progress++;
189                         }
190                 }
191                 break;
192         case FES_REDO:
193                 if (read_batch) {
194                         if (inc_recurse)
195                                 flist->in_progress++;
196                         break;
197                 }
198                 if (inc_recurse)
199                         flist->to_redo++;
200                 flist_ndx_push(&redo_list, ndx);
201                 break;
202         case FES_NO_SEND:
203                 break;
204         }
205 }
206
207 /* Note the fds used for the main socket (which might really be a pipe
208  * for a local transfer, but we can ignore that). */
209 void io_set_sock_fds(int f_in, int f_out)
210 {
211         sock_f_in = f_in;
212         sock_f_out = f_out;
213 }
214
215 void set_io_timeout(int secs)
216 {
217         io_timeout = secs;
218
219         if (!io_timeout || io_timeout > SELECT_TIMEOUT)
220                 select_timeout = SELECT_TIMEOUT;
221         else
222                 select_timeout = io_timeout;
223
224         allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2;
225 }
226
227 /* Setup the fd used to receive MSG_* messages.  Only needed during the
228  * early stages of being a local sender (up through the sending of the
229  * file list) or when we're the generator (to fetch the messages from
230  * the receiver). */
231 void set_msg_fd_in(int fd)
232 {
233         msg_fd_in = fd;
234 }
235
236 /* Setup the fd used to send our MSG_* messages.  Only needed when
237  * we're the receiver (to send our messages to the generator). */
238 void set_msg_fd_out(int fd)
239 {
240         msg_fd_out = fd;
241         set_nonblocking(msg_fd_out);
242 }
243
244 /* Add a message to the pending MSG_* list. */
245 static void msg_list_add(struct msg_list *lst, int code, const char *buf, int len, int convert)
246 {
247         struct msg_list_item *m;
248         int sz = len + 4 + sizeof m[0] - 1;
249
250         if (!(m = (struct msg_list_item *)new_array(char, sz)))
251                 out_of_memory("msg_list_add");
252         m->next = NULL;
253         m->convert = convert;
254         SIVAL(m->buf, 0, ((code+MPLEX_BASE)<<24) | len);
255         memcpy(m->buf + 4, buf, len);
256         if (lst->tail)
257                 lst->tail->next = m;
258         else
259                 lst->head = m;
260         lst->tail = m;
261 }
262
263 static inline int flush_a_msg(int fd)
264 {
265         struct msg_list_item *m = msg_queue.head;
266         int len = IVAL(m->buf, 0) & 0xFFFFFF;
267         int tag = *((uchar*)m->buf+3) - MPLEX_BASE;
268
269         if (!(msg_queue.head = m->next))
270                 msg_queue.tail = NULL;
271
272         defer_forwarding_messages++;
273         mplex_write(fd, tag, m->buf + 4, len, m->convert);
274         defer_forwarding_messages--;
275
276         free(m);
277
278         return len;
279 }
280
281 static void msg_flush(void)
282 {
283         if (am_generator) {
284                 while (msg_queue.head && io_multiplexing_out)
285                         stats.total_written += flush_a_msg(sock_f_out) + 4;
286         } else {
287                 while (msg_queue.head)
288                         (void)flush_a_msg(msg_fd_out);
289         }
290 }
291
292 static void check_for_d_option_error(const char *msg)
293 {
294         static char rsync263_opts[] = "BCDHIKLPRSTWabceghlnopqrtuvxz";
295         char *colon;
296         int saw_d = 0;
297
298         if (*msg != 'r'
299          || strncmp(msg, REMOTE_OPTION_ERROR, sizeof REMOTE_OPTION_ERROR - 1) != 0)
300                 return;
301
302         msg += sizeof REMOTE_OPTION_ERROR - 1;
303         if (*msg == '-' || (colon = strchr(msg, ':')) == NULL
304          || strncmp(colon, REMOTE_OPTION_ERROR2, sizeof REMOTE_OPTION_ERROR2 - 1) != 0)
305                 return;
306
307         for ( ; *msg != ':'; msg++) {
308                 if (*msg == 'd')
309                         saw_d = 1;
310                 else if (*msg == 'e')
311                         break;
312                 else if (strchr(rsync263_opts, *msg) == NULL)
313                         return;
314         }
315
316         if (saw_d) {
317                 rprintf(FWARNING,
318                     "*** Try using \"--old-d\" if remote rsync is <= 2.6.3 ***\n");
319         }
320 }
321
322 /* Read a message from the MSG_* fd and handle it.  This is called either
323  * during the early stages of being a local sender (up through the sending
324  * of the file list) or when we're the generator (to fetch the messages
325  * from the receiver). */
326 static void read_msg_fd(void)
327 {
328         char buf[2048];
329         size_t n;
330         struct file_list *flist;
331         int fd = msg_fd_in;
332         int tag, len;
333
334         /* Temporarily disable msg_fd_in.  This is needed to avoid looping back
335          * to this routine from writefd_unbuffered(). */
336         no_flush++;
337         msg_fd_in = -1;
338         defer_forwarding_messages++;
339
340         readfd(fd, buf, 4);
341         tag = IVAL(buf, 0);
342
343         len = tag & 0xFFFFFF;
344         tag = (tag >> 24) - MPLEX_BASE;
345
346         switch (tag) {
347         case MSG_DONE:
348                 if (len < 0 || len > 1 || !am_generator) {
349                   invalid_msg:
350                         rprintf(FERROR, "invalid message %d:%d [%s%s]\n",
351                                 tag, len, who_am_i(),
352                                 inc_recurse ? "/inc" : "");
353                         exit_cleanup(RERR_STREAMIO);
354                 }
355                 if (len) {
356                         readfd(fd, buf, len);
357                         stats.total_read = read_varlong(fd, 3);
358                 }
359                 msgdone_cnt++;
360                 break;
361         case MSG_REDO:
362                 if (len != 4 || !am_generator)
363                         goto invalid_msg;
364                 readfd(fd, buf, 4);
365                 got_flist_entry_status(FES_REDO, buf);
366                 break;
367         case MSG_FLIST:
368                 if (len != 4 || !am_generator || !inc_recurse)
369                         goto invalid_msg;
370                 readfd(fd, buf, 4);
371                 /* Read extra file list from receiver. */
372                 assert(iobuf_in != NULL);
373                 assert(iobuf_f_in == fd);
374                 if (verbose > 3) {
375                         rprintf(FINFO, "[%s] receiving flist for dir %d\n",
376                                 who_am_i(), IVAL(buf,0));
377                 }
378                 flist = recv_file_list(fd);
379                 flist->parent_ndx = IVAL(buf,0);
380 #ifdef SUPPORT_HARD_LINKS
381                 if (preserve_hard_links)
382                         match_hard_links(flist);
383 #endif
384                 break;
385         case MSG_FLIST_EOF:
386                 if (len != 0 || !am_generator || !inc_recurse)
387                         goto invalid_msg;
388                 flist_eof = 1;
389                 break;
390         case MSG_IO_ERROR:
391                 if (len != 4)
392                         goto invalid_msg;
393                 readfd(fd, buf, len);
394                 io_error |= IVAL(buf, 0);
395                 break;
396         case MSG_DELETED:
397                 if (len >= (int)sizeof buf || !am_generator)
398                         goto invalid_msg;
399                 readfd(fd, buf, len);
400                 send_msg(MSG_DELETED, buf, len, 1);
401                 break;
402         case MSG_SUCCESS:
403                 if (len != 4 || !am_generator)
404                         goto invalid_msg;
405                 readfd(fd, buf, 4);
406                 got_flist_entry_status(FES_SUCCESS, buf);
407                 break;
408         case MSG_NO_SEND:
409                 if (len != 4 || !am_generator)
410                         goto invalid_msg;
411                 readfd(fd, buf, 4);
412                 got_flist_entry_status(FES_NO_SEND, buf);
413                 break;
414         case MSG_ERROR_SOCKET:
415         case MSG_ERROR_UTF8:
416         case MSG_CLIENT:
417                 if (!am_generator)
418                         goto invalid_msg;
419                 if (tag == MSG_ERROR_SOCKET)
420                         io_end_multiplex_out();
421                 /* FALL THROUGH */
422         case MSG_INFO:
423         case MSG_ERROR:
424         case MSG_ERROR_XFER:
425         case MSG_WARNING:
426         case MSG_LOG:
427                 while (len) {
428                         n = len;
429                         if (n >= sizeof buf)
430                                 n = sizeof buf - 1;
431                         readfd(fd, buf, n);
432                         rwrite((enum logcode)tag, buf, n, !am_generator);
433                         len -= n;
434                 }
435                 break;
436         default:
437                 rprintf(FERROR, "unknown message %d:%d [%s]\n",
438                         tag, len, who_am_i());
439                 exit_cleanup(RERR_STREAMIO);
440         }
441
442         no_flush--;
443         msg_fd_in = fd;
444         if (!--defer_forwarding_messages && !no_flush)
445                 msg_flush();
446 }
447
448 /* This is used by the generator to limit how many file transfers can
449  * be active at once when --remove-source-files is specified.  Without
450  * this, sender-side deletions were mostly happening at the end. */
451 void increment_active_files(int ndx, int itemizing, enum logcode code)
452 {
453         while (1) {
454                 /* TODO: tune these limits? */
455                 int limit = active_bytecnt >= 128*1024 ? 10 : 50;
456                 if (active_filecnt < limit)
457                         break;
458                 check_for_finished_files(itemizing, code, 0);
459                 if (active_filecnt < limit)
460                         break;
461                 if (iobuf_out_cnt)
462                         io_flush(NORMAL_FLUSH);
463                 else
464                         read_msg_fd();
465         }
466
467         active_filecnt++;
468         active_bytecnt += F_LENGTH(cur_flist->files[ndx - cur_flist->ndx_start]);
469 }
470
471 /* Write an message to a multiplexed stream. If this fails, rsync exits. */
472 static void mplex_write(int fd, enum msgcode code, const char *buf, size_t len, int convert)
473 {
474         char buffer[BIGPATHBUFLEN]; /* Oversized for use by iconv code. */
475         size_t n = len;
476
477 #ifdef ICONV_OPTION
478         /* We need to convert buf before doing anything else so that we
479          * can include the (converted) byte length in the message header. */
480         if (convert && ic_send != (iconv_t)-1) {
481                 xbuf outbuf, inbuf;
482
483                 INIT_XBUF(outbuf, buffer + 4, 0, sizeof buffer - 4);
484                 INIT_XBUF(inbuf, (char*)buf, len, -1);
485
486                 iconvbufs(ic_send, &inbuf, &outbuf,
487                           ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE);
488                 if (inbuf.len > 0) {
489                         rprintf(FERROR, "overflowed conversion buffer in mplex_write");
490                         exit_cleanup(RERR_UNSUPPORTED);
491                 }
492
493                 n = len = outbuf.len;
494         } else
495 #endif
496         if (n > 1024 - 4) /* BIGPATHBUFLEN can handle 1024 bytes */
497                 n = 0;    /* We'd rather do 2 writes than too much memcpy(). */
498         else
499                 memcpy(buffer + 4, buf, n);
500
501         SIVAL(buffer, 0, ((MPLEX_BASE + (int)code)<<24) + len);
502
503         keep_defer_forwarding++; /* defer_forwarding_messages++ on return */
504         writefd_unbuffered(fd, buffer, n+4);
505         keep_defer_forwarding--;
506
507         if (len > n)
508                 writefd_unbuffered(fd, buf+n, len-n);
509
510         if (!--defer_forwarding_messages && !no_flush)
511                 msg_flush();
512 }
513
514 int send_msg(enum msgcode code, const char *buf, int len, int convert)
515 {
516         if (msg_fd_out < 0) {
517                 if (!defer_forwarding_messages)
518                         return io_multiplex_write(code, buf, len, convert);
519                 if (!io_multiplexing_out)
520                         return 0;
521                 msg_list_add(&msg_queue, code, buf, len, convert);
522                 return 1;
523         }
524         if (flist_forward_from >= 0)
525                 msg_list_add(&msg_queue, code, buf, len, convert);
526         else
527                 mplex_write(msg_fd_out, code, buf, len, convert);
528         return 1;
529 }
530
531 void send_msg_int(enum msgcode code, int num)
532 {
533         char numbuf[4];
534         SIVAL(numbuf, 0, num);
535         send_msg(code, numbuf, 4, 0);
536 }
537
538 void wait_for_receiver(void)
539 {
540         if (io_flush(NORMAL_FLUSH))
541                 return;
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 ff_buf to read a block of data from f_in (when it is ready, since
562  * it might be a pipe) and then blast it out f_out (when it is ready to
563  * 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         alloc_xbuf(&ff_buf, 2048);
570 #ifdef ICONV_OPTION
571         if (protect_args)
572                 alloc_xbuf(&iconv_buf, 1024);
573 #endif
574 }
575
576 /* It's almost always an error to get an EOF when we're trying to read from the
577  * network, because the protocol is (for the most part) self-terminating.
578  *
579  * There is one case for the receiver when it is at the end of the transfer
580  * (hanging around reading any keep-alive packets that might come its way): if
581  * the sender dies before the generator's kill-signal comes through, we can end
582  * up here needing to loop until the kill-signal arrives.  In this situation,
583  * kluge_around_eof will be < 0.
584  *
585  * There is another case for older protocol versions (< 24) where the module
586  * listing was not terminated, so we must ignore an EOF error in that case and
587  * exit.  In this situation, kluge_around_eof will be > 0. */
588 static void whine_about_eof(int fd)
589 {
590         if (kluge_around_eof && fd == sock_f_in) {
591                 int i;
592                 if (kluge_around_eof > 0)
593                         exit_cleanup(0);
594                 /* If we're still here after 10 seconds, exit with an error. */
595                 for (i = 10*1000/20; i--; )
596                         msleep(20);
597         }
598
599         rprintf(FERROR, RSYNC_NAME ": connection unexpectedly closed "
600                 "(%.0f bytes received so far) [%s]\n",
601                 (double)stats.total_read, who_am_i());
602
603         exit_cleanup(RERR_STREAMIO);
604 }
605
606 /**
607  * Read from a socket with I/O timeout. return the number of bytes
608  * read. If no bytes can be read then exit, never return a number <= 0.
609  *
610  * TODO: If the remote shell connection fails, then current versions
611  * actually report an "unexpected EOF" error here.  Since it's a
612  * fairly common mistake to try to use rsh when ssh is required, we
613  * should trap that: if we fail to read any data at all, we should
614  * give a better explanation.  We can tell whether the connection has
615  * started by looking e.g. at whether the remote version is known yet.
616  */
617 static int read_timeout(int fd, char *buf, size_t len)
618 {
619         int n, cnt = 0;
620
621         io_flush(FULL_FLUSH);
622
623         while (cnt == 0) {
624                 /* until we manage to read *something* */
625                 fd_set r_fds, w_fds;
626                 struct timeval tv;
627                 int maxfd = fd;
628                 int count;
629
630                 FD_ZERO(&r_fds);
631                 FD_ZERO(&w_fds);
632                 FD_SET(fd, &r_fds);
633                 if (io_filesfrom_f_out >= 0) {
634                         int new_fd;
635                         if (ff_buf.len == 0) {
636                                 if (io_filesfrom_f_in >= 0) {
637                                         FD_SET(io_filesfrom_f_in, &r_fds);
638                                         new_fd = io_filesfrom_f_in;
639                                 } else {
640                                         io_filesfrom_f_out = -1;
641                                         new_fd = -1;
642                                 }
643                         } else {
644                                 FD_SET(io_filesfrom_f_out, &w_fds);
645                                 new_fd = io_filesfrom_f_out;
646                         }
647                         if (new_fd > maxfd)
648                                 maxfd = new_fd;
649                 }
650
651                 tv.tv_sec = select_timeout;
652                 tv.tv_usec = 0;
653
654                 errno = 0;
655
656                 count = select(maxfd + 1, &r_fds, &w_fds, NULL, &tv);
657
658                 if (count <= 0) {
659                         if (errno == EBADF) {
660                                 defer_forwarding_messages = 0;
661                                 exit_cleanup(RERR_SOCKETIO);
662                         }
663                         check_timeout();
664                         continue;
665                 }
666
667                 if (io_filesfrom_f_out >= 0) {
668                         if (ff_buf.len) {
669                                 if (FD_ISSET(io_filesfrom_f_out, &w_fds)) {
670                                         int l = write(io_filesfrom_f_out,
671                                                       ff_buf.buf + ff_buf.pos,
672                                                       ff_buf.len);
673                                         if (l > 0) {
674                                                 if (!(ff_buf.len -= l))
675                                                         ff_buf.pos = 0;
676                                                 else
677                                                         ff_buf.pos += l;
678                                         } else if (errno != EINTR) {
679                                                 /* XXX should we complain? */
680                                                 io_filesfrom_f_out = -1;
681                                         }
682                                 }
683                         } else if (io_filesfrom_f_in >= 0) {
684                                 if (FD_ISSET(io_filesfrom_f_in, &r_fds)) {
685 #ifdef ICONV_OPTION
686                                         xbuf *ibuf = filesfrom_convert ? &iconv_buf : &ff_buf;
687 #else
688                                         xbuf *ibuf = &ff_buf;
689 #endif
690                                         int l = read(io_filesfrom_f_in, ibuf->buf, ibuf->size);
691                                         if (l <= 0) {
692                                                 if (l == 0 || errno != EINTR) {
693                                                         /* Send end-of-file marker */
694                                                         memcpy(ff_buf.buf, "\0\0", 2);
695                                                         ff_buf.len = ff_lastchar? 2 : 1;
696                                                         ff_buf.pos = 0;
697                                                         io_filesfrom_f_in = -1;
698                                                 }
699                                         } else {
700 #ifdef ICONV_OPTION
701                                                 if (filesfrom_convert) {
702                                                         iconv_buf.pos = 0;
703                                                         iconv_buf.len = l;
704                                                         iconvbufs(ic_send, &iconv_buf, &ff_buf,
705                                                             ICB_EXPAND_OUT|ICB_INCLUDE_BAD|ICB_INCLUDE_INCOMPLETE);
706                                                         l = ff_buf.len;
707                                                 }
708 #endif
709                                                 if (!eol_nulls) {
710                                                         char *s = ff_buf.buf + l;
711                                                         /* Transform CR and/or LF into '\0' */
712                                                         while (s-- > ff_buf.buf) {
713                                                                 if (*s == '\n' || *s == '\r')
714                                                                         *s = '\0';
715                                                         }
716                                                 }
717                                                 if (!ff_lastchar) {
718                                                         /* Last buf ended with a '\0', so don't
719                                                          * let this buf start with one. */
720                                                         while (l && ff_buf.buf[ff_buf.pos] == '\0')
721                                                                 ff_buf.pos++, l--;
722                                                 }
723                                                 if (!l)
724                                                         ff_buf.pos = 0;
725                                                 else {
726                                                         char *f = ff_buf.buf + ff_buf.pos;
727                                                         char *t = f;
728                                                         char *eob = f + l;
729                                                         /* Eliminate any multi-'\0' runs. */
730                                                         while (f != eob) {
731                                                                 if (!(*t++ = *f++)) {
732                                                                         while (f != eob && !*f)
733                                                                                 f++, l--;
734                                                                 }
735                                                         }
736                                                         ff_lastchar = f[-1];
737                                                 }
738                                                 ff_buf.len = l;
739                                         }
740                                 }
741                         }
742                 }
743
744                 if (!FD_ISSET(fd, &r_fds))
745                         continue;
746
747                 n = read(fd, buf, len);
748
749                 if (n <= 0) {
750                         if (n == 0)
751                                 whine_about_eof(fd); /* Doesn't return. */
752                         if (errno == EINTR || errno == EWOULDBLOCK
753                             || errno == EAGAIN)
754                                 continue;
755
756                         /* Don't write errors on a dead socket. */
757                         if (fd == sock_f_in) {
758                                 io_end_multiplex_out();
759                                 rsyserr(FERROR_SOCKET, errno, "read error");
760                         } else
761                                 rsyserr(FERROR, errno, "read error");
762                         exit_cleanup(RERR_STREAMIO);
763                 }
764
765                 buf += n;
766                 len -= n;
767                 cnt += n;
768
769                 if (fd == sock_f_in && io_timeout)
770                         last_io_in = time(NULL);
771         }
772
773         return cnt;
774 }
775
776 /* Read a line into the "buf" buffer. */
777 int read_line(int fd, char *buf, size_t bufsiz, int flags)
778 {
779         char ch, *s, *eob;
780         int cnt;
781
782 #ifdef ICONV_OPTION
783         if (flags & RL_CONVERT && iconv_buf.size < bufsiz)
784                 realloc_xbuf(&iconv_buf, bufsiz + 1024);
785 #endif
786
787   start:
788 #ifdef ICONV_OPTION
789         s = flags & RL_CONVERT ? iconv_buf.buf : buf;
790 #else
791         s = buf;
792 #endif
793         eob = s + bufsiz - 1;
794         while (1) {
795                 cnt = read(fd, &ch, 1);
796                 if (cnt < 0 && (errno == EWOULDBLOCK
797                   || errno == EINTR || errno == EAGAIN)) {
798                         struct timeval tv;
799                         fd_set r_fds, e_fds;
800                         FD_ZERO(&r_fds);
801                         FD_SET(fd, &r_fds);
802                         FD_ZERO(&e_fds);
803                         FD_SET(fd, &e_fds);
804                         tv.tv_sec = select_timeout;
805                         tv.tv_usec = 0;
806                         if (!select(fd+1, &r_fds, NULL, &e_fds, &tv))
807                                 check_timeout();
808                         /*if (FD_ISSET(fd, &e_fds))
809                                 rprintf(FINFO, "select exception on fd %d\n", fd); */
810                         continue;
811                 }
812                 if (cnt != 1)
813                         break;
814                 if (flags & RL_EOL_NULLS ? ch == '\0' : (ch == '\r' || ch == '\n')) {
815                         /* Skip empty lines if dumping comments. */
816                         if (flags & RL_DUMP_COMMENTS && s == buf)
817                                 continue;
818                         break;
819                 }
820                 if (s < eob)
821                         *s++ = ch;
822         }
823         *s = '\0';
824
825         if (flags & RL_DUMP_COMMENTS && (*buf == '#' || *buf == ';'))
826                 goto start;
827
828 #ifdef ICONV_OPTION
829         if (flags & RL_CONVERT) {
830                 xbuf outbuf;
831                 INIT_XBUF(outbuf, buf, 0, bufsiz);
832                 iconv_buf.pos = 0;
833                 iconv_buf.len = s - iconv_buf.buf;
834                 iconvbufs(ic_recv, &iconv_buf, &outbuf,
835                           ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE);
836                 outbuf.buf[outbuf.len] = '\0';
837                 return outbuf.len;
838         }
839 #endif
840
841         return s - buf;
842 }
843
844 void read_args(int f_in, char *mod_name, char *buf, size_t bufsiz, int rl_nulls,
845                char ***argv_p, int *argc_p, char **request_p)
846 {
847         int maxargs = MAX_ARGS;
848         int dot_pos = 0;
849         int argc = 0;
850         char **argv, *p;
851         int rl_flags = (rl_nulls ? RL_EOL_NULLS : 0);
852
853 #ifdef ICONV_OPTION
854         rl_flags |= (protect_args && ic_recv != (iconv_t)-1 ? RL_CONVERT : 0);
855 #endif
856
857         if (!(argv = new_array(char *, maxargs)))
858                 out_of_memory("read_args");
859         if (mod_name && !protect_args)
860                 argv[argc++] = "rsyncd";
861
862         while (1) {
863                 if (read_line(f_in, buf, bufsiz, rl_flags) == 0)
864                         break;
865
866                 if (argc == maxargs-1) {
867                         maxargs += MAX_ARGS;
868                         if (!(argv = realloc_array(argv, char *, maxargs)))
869                                 out_of_memory("read_args");
870                 }
871
872                 if (dot_pos) {
873                         if (request_p) {
874                                 *request_p = strdup(buf);
875                                 request_p = NULL;
876                         }
877                         if (mod_name)
878                                 glob_expand_module(mod_name, buf, &argv, &argc, &maxargs);
879                         else
880                                 glob_expand(buf, &argv, &argc, &maxargs);
881                 } else {
882                         if (!(p = strdup(buf)))
883                                 out_of_memory("read_args");
884                         argv[argc++] = p;
885                         if (*p == '.' && p[1] == '\0')
886                                 dot_pos = argc;
887                 }
888         }
889         argv[argc] = NULL;
890
891         glob_expand(NULL, NULL, NULL, NULL);
892
893         *argc_p = argc;
894         *argv_p = argv;
895 }
896
897 int io_start_buffering_out(int f_out)
898 {
899         if (iobuf_out) {
900                 assert(f_out == iobuf_f_out);
901                 return 0;
902         }
903         if (!(iobuf_out = new_array(char, IO_BUFFER_SIZE)))
904                 out_of_memory("io_start_buffering_out");
905         iobuf_out_cnt = 0;
906         iobuf_f_out = f_out;
907         return 1;
908 }
909
910 int io_start_buffering_in(int f_in)
911 {
912         if (iobuf_in) {
913                 assert(f_in == iobuf_f_in);
914                 return 0;
915         }
916         iobuf_in_siz = 2 * IO_BUFFER_SIZE;
917         if (!(iobuf_in = new_array(char, iobuf_in_siz)))
918                 out_of_memory("io_start_buffering_in");
919         iobuf_f_in = f_in;
920         return 1;
921 }
922
923 void io_end_buffering_in(void)
924 {
925         if (!iobuf_in)
926                 return;
927         free(iobuf_in);
928         iobuf_in = NULL;
929         iobuf_in_ndx = 0;
930         iobuf_in_remaining = 0;
931         iobuf_f_in = -1;
932 }
933
934 void io_end_buffering_out(void)
935 {
936         if (!iobuf_out)
937                 return;
938         io_flush(FULL_FLUSH);
939         free(iobuf_out);
940         iobuf_out = NULL;
941         iobuf_f_out = -1;
942 }
943
944 void maybe_flush_socket(int important)
945 {
946         if (iobuf_out && iobuf_out_cnt
947          && (important || time(NULL) - last_io_out >= 5))
948                 io_flush(NORMAL_FLUSH);
949 }
950
951 /* Older rsync versions used to send either a MSG_NOOP (protocol 30) or a
952  * raw-data-based keep-alive (protocol 29), both of which implied forwarding of
953  * the message through the sender.  Since the new timeout method does not need
954  * any forwarding, we just send an empty MSG_DATA message, which works with all
955  * rsync versions.  This avoids any message forwarding, and leaves the raw-data
956  * stream alone (since we can never be quite sure if that stream is in the
957  * right state for a keep-alive message). */
958 void maybe_send_keepalive(void)
959 {
960         if (time(NULL) - last_io_out >= allowed_lull) {
961                 if (!iobuf_out || !iobuf_out_cnt)
962                         send_msg(MSG_DATA, "", 0, 0);
963                 if (iobuf_out)
964                         io_flush(NORMAL_FLUSH);
965         }
966 }
967
968 void start_flist_forward(int f_in)
969 {
970         assert(iobuf_out != NULL);
971         assert(iobuf_f_out == msg_fd_out);
972         flist_forward_from = f_in;
973         defer_forwarding_messages++;
974 }
975
976 void stop_flist_forward(void)
977 {
978         flist_forward_from = -1;
979         defer_forwarding_messages--;
980         io_flush(FULL_FLUSH);
981 }
982
983 /**
984  * Continue trying to read len bytes - don't return until len has been
985  * read.
986  **/
987 static void read_loop(int fd, char *buf, size_t len)
988 {
989         while (len) {
990                 int n = read_timeout(fd, buf, len);
991
992                 buf += n;
993                 len -= n;
994         }
995 }
996
997 /**
998  * Read from the file descriptor handling multiplexing - return number
999  * of bytes read.
1000  *
1001  * Never returns <= 0.
1002  */
1003 static int readfd_unbuffered(int fd, char *buf, size_t len)
1004 {
1005         size_t msg_bytes;
1006         int tag, cnt = 0;
1007         char line[BIGPATHBUFLEN];
1008
1009         if (!iobuf_in || fd != iobuf_f_in)
1010                 return read_timeout(fd, buf, len);
1011
1012         if (!io_multiplexing_in && iobuf_in_remaining == 0) {
1013                 iobuf_in_remaining = read_timeout(fd, iobuf_in, iobuf_in_siz);
1014                 iobuf_in_ndx = 0;
1015         }
1016
1017         while (cnt == 0) {
1018                 if (iobuf_in_remaining) {
1019                         len = MIN(len, iobuf_in_remaining);
1020                         memcpy(buf, iobuf_in + iobuf_in_ndx, len);
1021                         iobuf_in_ndx += len;
1022                         iobuf_in_remaining -= len;
1023                         cnt = len;
1024                         break;
1025                 }
1026
1027                 read_loop(fd, line, 4);
1028                 tag = IVAL(line, 0);
1029
1030                 msg_bytes = tag & 0xFFFFFF;
1031                 tag = (tag >> 24) - MPLEX_BASE;
1032
1033                 switch (tag) {
1034                 case MSG_DATA:
1035                         if (msg_bytes > iobuf_in_siz) {
1036                                 if (!(iobuf_in = realloc_array(iobuf_in, char,
1037                                                                msg_bytes)))
1038                                         out_of_memory("readfd_unbuffered");
1039                                 iobuf_in_siz = msg_bytes;
1040                         }
1041                         read_loop(fd, iobuf_in, msg_bytes);
1042                         iobuf_in_remaining = msg_bytes;
1043                         iobuf_in_ndx = 0;
1044                         break;
1045                 case MSG_NOOP:
1046                         /* Support protocol-30 keep-alive method. */
1047                         if (msg_bytes != 0)
1048                                 goto invalid_msg;
1049                         if (am_sender)
1050                                 maybe_send_keepalive();
1051                         break;
1052                 case MSG_IO_ERROR:
1053                         if (msg_bytes != 4)
1054                                 goto invalid_msg;
1055                         read_loop(fd, line, msg_bytes);
1056                         send_msg_int(MSG_IO_ERROR, IVAL(line, 0));
1057                         io_error |= IVAL(line, 0);
1058                         break;
1059                 case MSG_DELETED:
1060                         if (msg_bytes >= sizeof line)
1061                                 goto overflow;
1062 #ifdef ICONV_OPTION
1063                         if (ic_recv != (iconv_t)-1) {
1064                                 xbuf outbuf, inbuf;
1065                                 char ibuf[512];
1066                                 int add_null = 0;
1067                                 int pos = 0;
1068
1069                                 INIT_CONST_XBUF(outbuf, line);
1070                                 INIT_XBUF(inbuf, ibuf, 0, -1);
1071
1072                                 while (msg_bytes) {
1073                                         inbuf.len = msg_bytes > sizeof ibuf
1074                                                   ? sizeof ibuf : msg_bytes;
1075                                         read_loop(fd, inbuf.buf, inbuf.len);
1076                                         if (!(msg_bytes -= inbuf.len)
1077                                          && !ibuf[inbuf.len-1])
1078                                                 inbuf.len--, add_null = 1;
1079                                         if (iconvbufs(ic_send, &inbuf, &outbuf,
1080                                             ICB_INCLUDE_BAD | ICB_INCLUDE_INCOMPLETE) < 0)
1081                                                 goto overflow;
1082                                         pos = -1;
1083                                 }
1084                                 if (add_null) {
1085                                         if (outbuf.len == outbuf.size)
1086                                                 goto overflow;
1087                                         outbuf.buf[outbuf.len++] = '\0';
1088                                 }
1089                                 msg_bytes = outbuf.len;
1090                         } else
1091 #endif
1092                                 read_loop(fd, line, msg_bytes);
1093                         /* A directory name was sent with the trailing null */
1094                         if (msg_bytes > 0 && !line[msg_bytes-1])
1095                                 log_delete(line, S_IFDIR);
1096                         else {
1097                                 line[msg_bytes] = '\0';
1098                                 log_delete(line, S_IFREG);
1099                         }
1100                         break;
1101                 case MSG_SUCCESS:
1102                         if (msg_bytes != 4) {
1103                           invalid_msg:
1104                                 rprintf(FERROR, "invalid multi-message %d:%ld [%s]\n",
1105                                         tag, (long)msg_bytes, who_am_i());
1106                                 exit_cleanup(RERR_STREAMIO);
1107                         }
1108                         read_loop(fd, line, msg_bytes);
1109                         successful_send(IVAL(line, 0));
1110                         break;
1111                 case MSG_NO_SEND:
1112                         if (msg_bytes != 4)
1113                                 goto invalid_msg;
1114                         read_loop(fd, line, msg_bytes);
1115                         send_msg_int(MSG_NO_SEND, IVAL(line, 0));
1116                         break;
1117                 case MSG_INFO:
1118                 case MSG_ERROR:
1119                 case MSG_ERROR_XFER:
1120                 case MSG_WARNING:
1121                         if (msg_bytes >= sizeof line) {
1122                             overflow:
1123                                 rprintf(FERROR,
1124                                         "multiplexing overflow %d:%ld [%s]\n",
1125                                         tag, (long)msg_bytes, who_am_i());
1126                                 exit_cleanup(RERR_STREAMIO);
1127                         }
1128                         read_loop(fd, line, msg_bytes);
1129                         rwrite((enum logcode)tag, line, msg_bytes, 1);
1130                         if (first_message) {
1131                                 if (list_only && !am_sender && tag == 1) {
1132                                         line[msg_bytes] = '\0';
1133                                         check_for_d_option_error(line);
1134                                 }
1135                                 first_message = 0;
1136                         }
1137                         break;
1138                 default:
1139                         rprintf(FERROR, "unexpected tag %d [%s]\n",
1140                                 tag, who_am_i());
1141                         exit_cleanup(RERR_STREAMIO);
1142                 }
1143         }
1144
1145         if (iobuf_in_remaining == 0)
1146                 io_flush(NORMAL_FLUSH);
1147
1148         return cnt;
1149 }
1150
1151 /* Do a buffered read from fd.  Don't return until all N bytes have
1152  * been read.  If all N can't be read then exit with an error. */
1153 static void readfd(int fd, char *buffer, size_t N)
1154 {
1155         int  cnt;
1156         size_t total = 0;
1157
1158         while (total < N) {
1159                 cnt = readfd_unbuffered(fd, buffer + total, N-total);
1160                 total += cnt;
1161         }
1162
1163         if (fd == write_batch_monitor_in) {
1164                 if ((size_t)write(batch_fd, buffer, total) != total)
1165                         exit_cleanup(RERR_FILEIO);
1166         }
1167
1168         if (fd == flist_forward_from)
1169                 writefd(iobuf_f_out, buffer, total);
1170
1171         if (fd == sock_f_in)
1172                 stats.total_read += total;
1173 }
1174
1175 unsigned short read_shortint(int f)
1176 {
1177         char b[2];
1178         readfd(f, b, 2);
1179         return (UVAL(b, 1) << 8) + UVAL(b, 0);
1180 }
1181
1182 int32 read_int(int f)
1183 {
1184         char b[4];
1185         int32 num;
1186
1187         readfd(f, b, 4);
1188         num = IVAL(b, 0);
1189 #if SIZEOF_INT32 > 4
1190         if (num & (int32)0x80000000)
1191                 num |= ~(int32)0xffffffff;
1192 #endif
1193         return num;
1194 }
1195
1196 int32 read_varint(int f)
1197 {
1198         union {
1199             char b[5];
1200             int32 x;
1201         } u;
1202         uchar ch;
1203         int extra;
1204
1205         u.x = 0;
1206         readfd(f, (char*)&ch, 1);
1207         extra = int_byte_extra[ch / 4];
1208         if (extra) {
1209                 uchar bit = ((uchar)1<<(8-extra));
1210                 if (extra >= (int)sizeof u.b) {
1211                         rprintf(FERROR, "Overflow in read_varint()\n");
1212                         exit_cleanup(RERR_STREAMIO);
1213                 }
1214                 readfd(f, u.b, extra);
1215                 u.b[extra] = ch & (bit-1);
1216         } else
1217                 u.b[0] = ch;
1218 #if CAREFUL_ALIGNMENT
1219         u.x = IVAL(u.b,0);
1220 #endif
1221 #if SIZEOF_INT32 > 4
1222         if (u.x & (int32)0x80000000)
1223                 u.x |= ~(int32)0xffffffff;
1224 #endif
1225         return u.x;
1226 }
1227
1228 int64 read_varlong(int f, uchar min_bytes)
1229 {
1230         union {
1231             char b[9];
1232             int64 x;
1233         } u;
1234         char b2[8];
1235         int extra;
1236
1237 #if SIZEOF_INT64 < 8
1238         memset(u.b, 0, 8);
1239 #else
1240         u.x = 0;
1241 #endif
1242         readfd(f, b2, min_bytes);
1243         memcpy(u.b, b2+1, min_bytes-1);
1244         extra = int_byte_extra[CVAL(b2, 0) / 4];
1245         if (extra) {
1246                 uchar bit = ((uchar)1<<(8-extra));
1247                 if (min_bytes + extra > (int)sizeof u.b) {
1248                         rprintf(FERROR, "Overflow in read_varlong()\n");
1249                         exit_cleanup(RERR_STREAMIO);
1250                 }
1251                 readfd(f, u.b + min_bytes - 1, extra);
1252                 u.b[min_bytes + extra - 1] = CVAL(b2, 0) & (bit-1);
1253 #if SIZEOF_INT64 < 8
1254                 if (min_bytes + extra > 5 || u.b[4] || CVAL(u.b,3) & 0x80) {
1255                         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1256                         exit_cleanup(RERR_UNSUPPORTED);
1257                 }
1258 #endif
1259         } else
1260                 u.b[min_bytes + extra - 1] = CVAL(b2, 0);
1261 #if SIZEOF_INT64 < 8
1262         u.x = IVAL(u.b,0);
1263 #elif CAREFUL_ALIGNMENT
1264         u.x = IVAL(u.b,0) | (((int64)IVAL(u.b,4))<<32);
1265 #endif
1266         return u.x;
1267 }
1268
1269 int64 read_longint(int f)
1270 {
1271 #if SIZEOF_INT64 >= 8
1272         char b[9];
1273 #endif
1274         int32 num = read_int(f);
1275
1276         if (num != (int32)0xffffffff)
1277                 return num;
1278
1279 #if SIZEOF_INT64 < 8
1280         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1281         exit_cleanup(RERR_UNSUPPORTED);
1282 #else
1283         readfd(f, b, 8);
1284         return IVAL(b,0) | (((int64)IVAL(b,4))<<32);
1285 #endif
1286 }
1287
1288 void read_buf(int f, char *buf, size_t len)
1289 {
1290         readfd(f,buf,len);
1291 }
1292
1293 void read_sbuf(int f, char *buf, size_t len)
1294 {
1295         readfd(f, buf, len);
1296         buf[len] = '\0';
1297 }
1298
1299 uchar read_byte(int f)
1300 {
1301         uchar c;
1302         readfd(f, (char *)&c, 1);
1303         return c;
1304 }
1305
1306 int read_vstring(int f, char *buf, int bufsize)
1307 {
1308         int len = read_byte(f);
1309
1310         if (len & 0x80)
1311                 len = (len & ~0x80) * 0x100 + read_byte(f);
1312
1313         if (len >= bufsize) {
1314                 rprintf(FERROR, "over-long vstring received (%d > %d)\n",
1315                         len, bufsize - 1);
1316                 return -1;
1317         }
1318
1319         if (len)
1320                 readfd(f, buf, len);
1321         buf[len] = '\0';
1322         return len;
1323 }
1324
1325 /* Populate a sum_struct with values from the socket.  This is
1326  * called by both the sender and the receiver. */
1327 void read_sum_head(int f, struct sum_struct *sum)
1328 {
1329         int32 max_blength = protocol_version < 30 ? OLD_MAX_BLOCK_SIZE : MAX_BLOCK_SIZE;
1330         sum->count = read_int(f);
1331         if (sum->count < 0) {
1332                 rprintf(FERROR, "Invalid checksum count %ld [%s]\n",
1333                         (long)sum->count, who_am_i());
1334                 exit_cleanup(RERR_PROTOCOL);
1335         }
1336         sum->blength = read_int(f);
1337         if (sum->blength < 0 || sum->blength > max_blength) {
1338                 rprintf(FERROR, "Invalid block length %ld [%s]\n",
1339                         (long)sum->blength, who_am_i());
1340                 exit_cleanup(RERR_PROTOCOL);
1341         }
1342         sum->s2length = protocol_version < 27 ? csum_length : (int)read_int(f);
1343         if (sum->s2length < 0 || sum->s2length > MAX_DIGEST_LEN) {
1344                 rprintf(FERROR, "Invalid checksum length %d [%s]\n",
1345                         sum->s2length, who_am_i());
1346                 exit_cleanup(RERR_PROTOCOL);
1347         }
1348         sum->remainder = read_int(f);
1349         if (sum->remainder < 0 || sum->remainder > sum->blength) {
1350                 rprintf(FERROR, "Invalid remainder length %ld [%s]\n",
1351                         (long)sum->remainder, who_am_i());
1352                 exit_cleanup(RERR_PROTOCOL);
1353         }
1354 }
1355
1356 /* Send the values from a sum_struct over the socket.  Set sum to
1357  * NULL if there are no checksums to send.  This is called by both
1358  * the generator and the sender. */
1359 void write_sum_head(int f, struct sum_struct *sum)
1360 {
1361         static struct sum_struct null_sum;
1362
1363         if (sum == NULL)
1364                 sum = &null_sum;
1365
1366         write_int(f, sum->count);
1367         write_int(f, sum->blength);
1368         if (protocol_version >= 27)
1369                 write_int(f, sum->s2length);
1370         write_int(f, sum->remainder);
1371 }
1372
1373 /**
1374  * Sleep after writing to limit I/O bandwidth usage.
1375  *
1376  * @todo Rather than sleeping after each write, it might be better to
1377  * use some kind of averaging.  The current algorithm seems to always
1378  * use a bit less bandwidth than specified, because it doesn't make up
1379  * for slow periods.  But arguably this is a feature.  In addition, we
1380  * ought to take the time used to write the data into account.
1381  *
1382  * During some phases of big transfers (file FOO is uptodate) this is
1383  * called with a small bytes_written every time.  As the kernel has to
1384  * round small waits up to guarantee that we actually wait at least the
1385  * requested number of microseconds, this can become grossly inaccurate.
1386  * We therefore keep track of the bytes we've written over time and only
1387  * sleep when the accumulated delay is at least 1 tenth of a second.
1388  **/
1389 static void sleep_for_bwlimit(int bytes_written)
1390 {
1391         static struct timeval prior_tv;
1392         static long total_written = 0;
1393         struct timeval tv, start_tv;
1394         long elapsed_usec, sleep_usec;
1395
1396 #define ONE_SEC 1000000L /* # of microseconds in a second */
1397
1398         if (!bwlimit_writemax)
1399                 return;
1400
1401         total_written += bytes_written;
1402
1403         gettimeofday(&start_tv, NULL);
1404         if (prior_tv.tv_sec) {
1405                 elapsed_usec = (start_tv.tv_sec - prior_tv.tv_sec) * ONE_SEC
1406                              + (start_tv.tv_usec - prior_tv.tv_usec);
1407                 total_written -= elapsed_usec * bwlimit / (ONE_SEC/1024);
1408                 if (total_written < 0)
1409                         total_written = 0;
1410         }
1411
1412         sleep_usec = total_written * (ONE_SEC/1024) / bwlimit;
1413         if (sleep_usec < ONE_SEC / 10) {
1414                 prior_tv = start_tv;
1415                 return;
1416         }
1417
1418         tv.tv_sec  = sleep_usec / ONE_SEC;
1419         tv.tv_usec = sleep_usec % ONE_SEC;
1420         select(0, NULL, NULL, NULL, &tv);
1421
1422         gettimeofday(&prior_tv, NULL);
1423         elapsed_usec = (prior_tv.tv_sec - start_tv.tv_sec) * ONE_SEC
1424                      + (prior_tv.tv_usec - start_tv.tv_usec);
1425         total_written = (sleep_usec - elapsed_usec) * bwlimit / (ONE_SEC/1024);
1426 }
1427
1428 static const char *what_fd_is(int fd)
1429 {
1430         static char buf[20];
1431
1432         if (fd == sock_f_out)
1433                 return "socket";
1434         else if (fd == msg_fd_out)
1435                 return "message fd";
1436         else if (fd == batch_fd)
1437                 return "batch file";
1438         else {
1439                 snprintf(buf, sizeof buf, "fd %d", fd);
1440                 return buf;
1441         }
1442 }
1443
1444 /* Write len bytes to the file descriptor fd, looping as necessary to get
1445  * the job done and also (in certain circumstances) reading any data on
1446  * msg_fd_in to avoid deadlock.
1447  *
1448  * This function underlies the multiplexing system.  The body of the
1449  * application never calls this function directly. */
1450 static void writefd_unbuffered(int fd, const char *buf, size_t len)
1451 {
1452         size_t n, total = 0;
1453         fd_set w_fds, r_fds, e_fds;
1454         int maxfd, count, cnt, using_r_fds;
1455         int defer_inc = 0;
1456         struct timeval tv;
1457
1458         if (no_flush++)
1459                 defer_forwarding_messages++, defer_inc++;
1460
1461         while (total < len) {
1462                 FD_ZERO(&w_fds);
1463                 FD_SET(fd, &w_fds);
1464                 FD_ZERO(&e_fds);
1465                 FD_SET(fd, &e_fds);
1466                 maxfd = fd;
1467
1468                 if (msg_fd_in >= 0) {
1469                         FD_ZERO(&r_fds);
1470                         FD_SET(msg_fd_in, &r_fds);
1471                         if (msg_fd_in > maxfd)
1472                                 maxfd = msg_fd_in;
1473                         using_r_fds = 1;
1474                 } else
1475                         using_r_fds = 0;
1476
1477                 tv.tv_sec = select_timeout;
1478                 tv.tv_usec = 0;
1479
1480                 errno = 0;
1481                 count = select(maxfd + 1, using_r_fds ? &r_fds : NULL,
1482                                &w_fds, &e_fds, &tv);
1483
1484                 if (count <= 0) {
1485                         if (count < 0 && errno == EBADF)
1486                                 exit_cleanup(RERR_SOCKETIO);
1487                         check_timeout();
1488                         continue;
1489                 }
1490
1491                 /*if (FD_ISSET(fd, &e_fds))
1492                         rprintf(FINFO, "select exception on fd %d\n", fd); */
1493
1494                 if (using_r_fds && FD_ISSET(msg_fd_in, &r_fds))
1495                         read_msg_fd();
1496
1497                 if (!FD_ISSET(fd, &w_fds))
1498                         continue;
1499
1500                 n = len - total;
1501                 if (bwlimit_writemax && n > bwlimit_writemax)
1502                         n = bwlimit_writemax;
1503                 cnt = write(fd, buf + total, n);
1504
1505                 if (cnt <= 0) {
1506                         if (cnt < 0) {
1507                                 if (errno == EINTR)
1508                                         continue;
1509                                 if (errno == EWOULDBLOCK || errno == EAGAIN) {
1510                                         msleep(1);
1511                                         continue;
1512                                 }
1513                         }
1514
1515                         /* Don't try to write errors back across the stream. */
1516                         if (fd == sock_f_out)
1517                                 io_end_multiplex_out();
1518                         /* Don't try to write errors down a failing msg pipe. */
1519                         if (am_server && fd == msg_fd_out)
1520                                 exit_cleanup(RERR_STREAMIO);
1521                         rsyserr(FERROR, errno,
1522                                 "writefd_unbuffered failed to write %ld bytes to %s [%s]",
1523                                 (long)len, what_fd_is(fd), who_am_i());
1524                         /* If the other side is sending us error messages, try
1525                          * to grab any messages they sent before they died. */
1526                         while (!am_server && fd == sock_f_out && io_multiplexing_in) {
1527                                 char buf[1024];
1528                                 set_io_timeout(30);
1529                                 readfd_unbuffered(sock_f_in, buf, sizeof buf);
1530                         }
1531                         exit_cleanup(RERR_STREAMIO);
1532                 }
1533
1534                 total += cnt;
1535                 defer_forwarding_messages++, defer_inc++;
1536
1537                 if (fd == sock_f_out) {
1538                         if (io_timeout || am_generator)
1539                                 last_io_out = time(NULL);
1540                         sleep_for_bwlimit(cnt);
1541                 }
1542         }
1543
1544         no_flush--;
1545         if (keep_defer_forwarding)
1546                 defer_inc--;
1547         if (!(defer_forwarding_messages -= defer_inc) && !no_flush)
1548                 msg_flush();
1549 }
1550
1551 int io_flush(int flush_it_all)
1552 {
1553         int flushed_something = 0;
1554
1555         if (no_flush)
1556                 return 0;
1557
1558         if (iobuf_out_cnt) {
1559                 if (io_multiplexing_out)
1560                         mplex_write(sock_f_out, MSG_DATA, iobuf_out, iobuf_out_cnt, 0);
1561                 else
1562                         writefd_unbuffered(iobuf_f_out, iobuf_out, iobuf_out_cnt);
1563                 iobuf_out_cnt = 0;
1564                 flushed_something = 1;
1565         }
1566
1567         if (flush_it_all && !defer_forwarding_messages && msg_queue.head) {
1568                 msg_flush();
1569                 flushed_something = 1;
1570         }
1571
1572         return flushed_something;
1573 }
1574
1575 static void writefd(int fd, const char *buf, size_t len)
1576 {
1577         if (fd == sock_f_out)
1578                 stats.total_written += len;
1579
1580         if (fd == write_batch_monitor_out)
1581                 writefd_unbuffered(batch_fd, buf, len);
1582
1583         if (!iobuf_out || fd != iobuf_f_out) {
1584                 writefd_unbuffered(fd, buf, len);
1585                 return;
1586         }
1587
1588         while (len) {
1589                 int n = MIN((int)len, IO_BUFFER_SIZE - iobuf_out_cnt);
1590                 if (n > 0) {
1591                         memcpy(iobuf_out+iobuf_out_cnt, buf, n);
1592                         buf += n;
1593                         len -= n;
1594                         iobuf_out_cnt += n;
1595                 }
1596
1597                 if (iobuf_out_cnt == IO_BUFFER_SIZE)
1598                         io_flush(NORMAL_FLUSH);
1599         }
1600 }
1601
1602 void write_shortint(int f, unsigned short x)
1603 {
1604         char b[2];
1605         b[0] = (char)x;
1606         b[1] = (char)(x >> 8);
1607         writefd(f, b, 2);
1608 }
1609
1610 void write_int(int f, int32 x)
1611 {
1612         char b[4];
1613         SIVAL(b, 0, x);
1614         writefd(f, b, 4);
1615 }
1616
1617 void write_varint(int f, int32 x)
1618 {
1619         char b[5];
1620         uchar bit;
1621         int cnt = 4;
1622
1623         SIVAL(b, 1, x);
1624
1625         while (cnt > 1 && b[cnt] == 0)
1626                 cnt--;
1627         bit = ((uchar)1<<(7-cnt+1));
1628         if (CVAL(b, cnt) >= bit) {
1629                 cnt++;
1630                 *b = ~(bit-1);
1631         } else if (cnt > 1)
1632                 *b = b[cnt] | ~(bit*2-1);
1633         else
1634                 *b = b[cnt];
1635
1636         writefd(f, b, cnt);
1637 }
1638
1639 void write_varlong(int f, int64 x, uchar min_bytes)
1640 {
1641         char b[9];
1642         uchar bit;
1643         int cnt = 8;
1644
1645         SIVAL(b, 1, x);
1646 #if SIZEOF_INT64 >= 8
1647         SIVAL(b, 5, x >> 32);
1648 #else
1649         if (x <= 0x7FFFFFFF && x >= 0)
1650                 memset(b + 5, 0, 4);
1651         else {
1652                 rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1653                 exit_cleanup(RERR_UNSUPPORTED);
1654         }
1655 #endif
1656
1657         while (cnt > min_bytes && b[cnt] == 0)
1658                 cnt--;
1659         bit = ((uchar)1<<(7-cnt+min_bytes));
1660         if (CVAL(b, cnt) >= bit) {
1661                 cnt++;
1662                 *b = ~(bit-1);
1663         } else if (cnt > min_bytes)
1664                 *b = b[cnt] | ~(bit*2-1);
1665         else
1666                 *b = b[cnt];
1667
1668         writefd(f, b, cnt);
1669 }
1670
1671 /*
1672  * Note: int64 may actually be a 32-bit type if ./configure couldn't find any
1673  * 64-bit types on this platform.
1674  */
1675 void write_longint(int f, int64 x)
1676 {
1677         char b[12], * const s = b+4;
1678
1679         SIVAL(s, 0, x);
1680         if (x <= 0x7FFFFFFF && x >= 0) {
1681                 writefd(f, s, 4);
1682                 return;
1683         }
1684
1685 #if SIZEOF_INT64 < 8
1686         rprintf(FERROR, "Integer overflow: attempted 64-bit offset\n");
1687         exit_cleanup(RERR_UNSUPPORTED);
1688 #else
1689         memset(b, 0xFF, 4);
1690         SIVAL(s, 4, x >> 32);
1691         writefd(f, b, 12);
1692 #endif
1693 }
1694
1695 void write_buf(int f, const char *buf, size_t len)
1696 {
1697         writefd(f,buf,len);
1698 }
1699
1700 /** Write a string to the connection */
1701 void write_sbuf(int f, const char *buf)
1702 {
1703         writefd(f, buf, strlen(buf));
1704 }
1705
1706 void write_byte(int f, uchar c)
1707 {
1708         writefd(f, (char *)&c, 1);
1709 }
1710
1711 void write_vstring(int f, const char *str, int len)
1712 {
1713         uchar lenbuf[3], *lb = lenbuf;
1714
1715         if (len > 0x7F) {
1716                 if (len > 0x7FFF) {
1717                         rprintf(FERROR,
1718                                 "attempting to send over-long vstring (%d > %d)\n",
1719                                 len, 0x7FFF);
1720                         exit_cleanup(RERR_PROTOCOL);
1721                 }
1722                 *lb++ = len / 0x100 + 0x80;
1723         }
1724         *lb = len;
1725
1726         writefd(f, (char*)lenbuf, lb - lenbuf + 1);
1727         if (len)
1728                 writefd(f, str, len);
1729 }
1730
1731 /* Send a file-list index using a byte-reduction method. */
1732 void write_ndx(int f, int32 ndx)
1733 {
1734         static int32 prev_positive = -1, prev_negative = 1;
1735         int32 diff, cnt = 0;
1736         char b[6];
1737
1738         if (protocol_version < 30 || read_batch) {
1739                 write_int(f, ndx);
1740                 return;
1741         }
1742
1743         /* Send NDX_DONE as a single-byte 0 with no side effects.  Send
1744          * negative nums as a positive after sending a leading 0xFF. */
1745         if (ndx >= 0) {
1746                 diff = ndx - prev_positive;
1747                 prev_positive = ndx;
1748         } else if (ndx == NDX_DONE) {
1749                 *b = 0;
1750                 writefd(f, b, 1);
1751                 return;
1752         } else {
1753                 b[cnt++] = (char)0xFF;
1754                 ndx = -ndx;
1755                 diff = ndx - prev_negative;
1756                 prev_negative = ndx;
1757         }
1758
1759         /* A diff of 1 - 253 is sent as a one-byte diff; a diff of 254 - 32767
1760          * or 0 is sent as a 0xFE + a two-byte diff; otherwise we send 0xFE
1761          * & all 4 bytes of the (non-negative) num with the high-bit set. */
1762         if (diff < 0xFE && diff > 0)
1763                 b[cnt++] = (char)diff;
1764         else if (diff < 0 || diff > 0x7FFF) {
1765                 b[cnt++] = (char)0xFE;
1766                 b[cnt++] = (char)((ndx >> 24) | 0x80);
1767                 b[cnt++] = (char)ndx;
1768                 b[cnt++] = (char)(ndx >> 8);
1769                 b[cnt++] = (char)(ndx >> 16);
1770         } else {
1771                 b[cnt++] = (char)0xFE;
1772                 b[cnt++] = (char)(diff >> 8);
1773                 b[cnt++] = (char)diff;
1774         }
1775         writefd(f, b, cnt);
1776 }
1777
1778 /* Receive a file-list index using a byte-reduction method. */
1779 int32 read_ndx(int f)
1780 {
1781         static int32 prev_positive = -1, prev_negative = 1;
1782         int32 *prev_ptr, num;
1783         char b[4];
1784
1785         if (protocol_version < 30)
1786                 return read_int(f);
1787
1788         readfd(f, b, 1);
1789         if (CVAL(b, 0) == 0xFF) {
1790                 readfd(f, b, 1);
1791                 prev_ptr = &prev_negative;
1792         } else if (CVAL(b, 0) == 0)
1793                 return NDX_DONE;
1794         else
1795                 prev_ptr = &prev_positive;
1796         if (CVAL(b, 0) == 0xFE) {
1797                 readfd(f, b, 2);
1798                 if (CVAL(b, 0) & 0x80) {
1799                         b[3] = CVAL(b, 0) & ~0x80;
1800                         b[0] = b[1];
1801                         readfd(f, b+1, 2);
1802                         num = IVAL(b, 0);
1803                 } else
1804                         num = (UVAL(b,0)<<8) + UVAL(b,1) + *prev_ptr;
1805         } else
1806                 num = UVAL(b, 0) + *prev_ptr;
1807         *prev_ptr = num;
1808         if (prev_ptr == &prev_negative)
1809                 num = -num;
1810         return num;
1811 }
1812
1813 /* Read a line of up to bufsiz-1 characters into buf.  Strips
1814  * the (required) trailing newline and all carriage returns.
1815  * Returns 1 for success; 0 for I/O error or truncation. */
1816 int read_line_old(int f, char *buf, size_t bufsiz)
1817 {
1818         bufsiz--; /* leave room for the null */
1819         while (bufsiz > 0) {
1820                 buf[0] = 0;
1821                 read_buf(f, buf, 1);
1822                 if (buf[0] == 0)
1823                         return 0;
1824                 if (buf[0] == '\n')
1825                         break;
1826                 if (buf[0] != '\r') {
1827                         buf++;
1828                         bufsiz--;
1829                 }
1830         }
1831         *buf = '\0';
1832         return bufsiz > 0;
1833 }
1834
1835 void io_printf(int fd, const char *format, ...)
1836 {
1837         va_list ap;
1838         char buf[BIGPATHBUFLEN];
1839         int len;
1840
1841         va_start(ap, format);
1842         len = vsnprintf(buf, sizeof buf, format, ap);
1843         va_end(ap);
1844
1845         if (len < 0)
1846                 exit_cleanup(RERR_STREAMIO);
1847
1848         if (len > (int)sizeof buf) {
1849                 rprintf(FERROR, "io_printf() was too long for the buffer.\n");
1850                 exit_cleanup(RERR_STREAMIO);
1851         }
1852
1853         write_sbuf(fd, buf);
1854 }
1855
1856 /** Setup for multiplexing a MSG_* stream with the data stream. */
1857 void io_start_multiplex_out(void)
1858 {
1859         io_flush(NORMAL_FLUSH);
1860         io_start_buffering_out(sock_f_out);
1861         io_multiplexing_out = 1;
1862 }
1863
1864 /** Setup for multiplexing a MSG_* stream with the data stream. */
1865 void io_start_multiplex_in(void)
1866 {
1867         io_flush(NORMAL_FLUSH);
1868         io_start_buffering_in(sock_f_in);
1869         io_multiplexing_in = 1;
1870 }
1871
1872 /** Write an message to the multiplexed data stream. */
1873 int io_multiplex_write(enum msgcode code, const char *buf, size_t len, int convert)
1874 {
1875         if (!io_multiplexing_out)
1876                 return 0;
1877         io_flush(NORMAL_FLUSH);
1878         stats.total_written += (len+4);
1879         mplex_write(sock_f_out, code, buf, len, convert);
1880         return 1;
1881 }
1882
1883 void io_end_multiplex_in(void)
1884 {
1885         io_multiplexing_in = 0;
1886         io_end_buffering_in();
1887 }
1888
1889 /** Stop output multiplexing. */
1890 void io_end_multiplex_out(void)
1891 {
1892         io_multiplexing_out = 0;
1893         io_end_buffering_out();
1894 }
1895
1896 void start_write_batch(int fd)
1897 {
1898         /* Some communication has already taken place, but we don't
1899          * enable batch writing until here so that we can write a
1900          * canonical record of the communication even though the
1901          * actual communication so far depends on whether a daemon
1902          * is involved. */
1903         write_int(batch_fd, protocol_version);
1904         if (protocol_version >= 30)
1905                 write_byte(batch_fd, inc_recurse);
1906         write_int(batch_fd, checksum_seed);
1907
1908         if (am_sender)
1909                 write_batch_monitor_out = fd;
1910         else
1911                 write_batch_monitor_in = fd;
1912 }
1913
1914 void stop_write_batch(void)
1915 {
1916         write_batch_monitor_out = -1;
1917         write_batch_monitor_in = -1;
1918 }