f79768d87877f3ff7c8d6a3865842c0a3ee85ceb
[ccan] / ccan / failtest / failtest.c
1 /* Licensed under LGPL - see LICENSE file for details */
2 #include <ccan/failtest/failtest.h>
3 #include <stdarg.h>
4 #include <string.h>
5 #include <stdio.h>
6 #include <stdarg.h>
7 #include <ctype.h>
8 #include <err.h>
9 #include <unistd.h>
10 #include <poll.h>
11 #include <errno.h>
12 #include <sys/types.h>
13 #include <sys/wait.h>
14 #include <sys/stat.h>
15 #include <sys/time.h>
16 #include <signal.h>
17 #include <assert.h>
18 #include <ccan/time/time.h>
19 #include <ccan/read_write_all/read_write_all.h>
20 #include <ccan/failtest/failtest_proto.h>
21 #include <ccan/build_assert/build_assert.h>
22
23 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
24
25 static int tracefd = -1;
26
27 unsigned int failtest_timeout_ms = 20000;
28
29 const char *failpath;
30 const char *debugpath;
31
32 enum info_type {
33         WRITE,
34         RELEASE_LOCKS,
35         FAILURE,
36         SUCCESS,
37         UNEXPECTED
38 };
39
40 struct lock_info {
41         int fd;
42         /* end is inclusive: you can't have a 0-byte lock. */
43         off_t start, end;
44         int type;
45 };
46
47 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
48
49 static struct failtest_call *history = NULL;
50 static unsigned int history_num = 0;
51 static int control_fd = -1;
52 static struct timeval start;
53 static unsigned int probe_count = 0;
54
55 static struct write_call *child_writes = NULL;
56 static unsigned int child_writes_num = 0;
57
58 static pid_t lock_owner;
59 static struct lock_info *locks = NULL;
60 static unsigned int lock_num = 0;
61
62 static pid_t orig_pid;
63
64 static const char info_to_arg[] = "mceoxprwf";
65
66 /* Dummy call used for failtest_undo wrappers. */
67 static struct failtest_call unrecorded_call;
68
69 static struct failtest_call *add_history_(enum failtest_call_type type,
70                                           const char *file,
71                                           unsigned int line,
72                                           const void *elem,
73                                           size_t elem_size)
74 {
75         /* NULL file is how we suppress failure. */
76         if (!file)
77                 return &unrecorded_call;
78
79         history = realloc(history, (history_num + 1) * sizeof(*history));
80         history[history_num].type = type;
81         history[history_num].file = file;
82         history[history_num].line = line;
83         history[history_num].cleanup = NULL;
84         memcpy(&history[history_num].u, elem, elem_size);
85         return &history[history_num++];
86 }
87
88 #define add_history(type, file, line, elem) \
89         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
90
91 /* We do a fake call inside a sizeof(), to check types. */
92 #define set_cleanup(call, clean, type)                  \
93         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL),1), (clean))
94
95 static bool read_write_info(int fd)
96 {
97         struct write_call *w;
98         char *buf;
99
100         /* We don't need all of this, but it's simple. */
101         child_writes = realloc(child_writes,
102                                (child_writes_num+1) * sizeof(child_writes[0]));
103         w = &child_writes[child_writes_num];
104         if (!read_all(fd, w, sizeof(*w)))
105                 return false;
106
107         w->buf = buf = malloc(w->count);
108         if (!read_all(fd, buf, w->count))
109                 return false;
110
111         child_writes_num++;
112         return true;
113 }
114
115 static char *failpath_string(void)
116 {
117         unsigned int i;
118         char *ret = malloc(history_num + 1);
119
120         for (i = 0; i < history_num; i++) {
121                 ret[i] = info_to_arg[history[i].type];
122                 if (history[i].fail)
123                         ret[i] = toupper(ret[i]);
124         }
125         ret[i] = '\0';
126         return ret;
127 }
128
129 static void tell_parent(enum info_type type)
130 {
131         if (control_fd != -1)
132                 write_all(control_fd, &type, sizeof(type));
133 }
134
135 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
136 {
137         va_list ap;
138         char *path = failpath_string();
139
140         va_start(ap, fmt);
141         vfprintf(stderr, fmt, ap);
142         va_end(ap);
143
144         fprintf(stderr, "%.*s", (int)outlen, out);
145         printf("To reproduce: --failpath=%s\n", path);
146         free(path);
147         tell_parent(FAILURE);
148         exit(1);
149 }
150
151 static void trace(const char *fmt, ...)
152 {
153         va_list ap;
154
155         if (tracefd == -1)
156                 return;
157
158         va_start(ap, fmt);
159         vdprintf(tracefd, fmt, ap);
160         va_end(ap);
161 }
162
163 static pid_t child;
164
165 static void hand_down(int signum)
166 {
167         kill(child, signum);
168 }
169
170 static void release_locks(void)
171 {
172         /* Locks were never acquired/reacquired? */
173         if (lock_owner == 0)
174                 return;
175
176         /* We own them?  Release them all. */
177         if (lock_owner == getpid()) {
178                 unsigned int i;
179                 struct flock fl;
180                 fl.l_type = F_UNLCK;
181                 fl.l_whence = SEEK_SET;
182                 fl.l_start = 0;
183                 fl.l_len = 0;
184
185                 for (i = 0; i < lock_num; i++)
186                         fcntl(locks[i].fd, F_SETLK, &fl);
187         } else {
188                 /* Our parent must have them; pass request up. */
189                 enum info_type type = RELEASE_LOCKS;
190                 assert(control_fd != -1);
191                 write_all(control_fd, &type, sizeof(type));
192         }
193         lock_owner = 0;
194 }
195
196 /* off_t is a signed type.  Getting its max is non-trivial. */
197 static off_t off_max(void)
198 {
199         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
200         if (sizeof(off_t) == 4)
201                 return (off_t)0x7FFFFFF;
202         else
203                 return (off_t)0x7FFFFFFFFFFFFFFULL;
204 }
205
206 static void get_locks(void)
207 {
208         unsigned int i;
209         struct flock fl;
210
211         if (lock_owner == getpid())
212                 return;
213
214         if (lock_owner != 0) {
215                 enum info_type type = RELEASE_LOCKS;
216                 assert(control_fd != -1);
217                 write_all(control_fd, &type, sizeof(type));
218         }
219
220         fl.l_whence = SEEK_SET;
221
222         for (i = 0; i < lock_num; i++) {
223                 fl.l_type = locks[i].type;
224                 fl.l_start = locks[i].start;
225                 if (locks[i].end == off_max())
226                         fl.l_len = 0;
227                 else
228                         fl.l_len = locks[i].end - locks[i].start + 1;
229
230                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
231                         abort();
232         }
233         lock_owner = getpid();
234 }
235
236 struct saved_file {
237         struct saved_file *next;
238         int fd;
239         void *contents;
240         off_t off, len;
241 };
242
243 static struct saved_file *save_file(struct saved_file *next, int fd)
244 {
245         struct saved_file *s = malloc(sizeof(*s));
246
247         s->next = next;
248         s->fd = fd;
249         s->off = lseek(fd, 0, SEEK_CUR);
250         /* Special file?  Erk... */
251         assert(s->off != -1);
252         s->len = lseek(fd, 0, SEEK_END);
253         lseek(fd, 0, SEEK_SET);
254         s->contents = malloc(s->len);
255         if (read(fd, s->contents, s->len) != s->len)
256                 err(1, "Failed to save %zu bytes", (size_t)s->len);
257         lseek(fd, s->off, SEEK_SET);
258         return s;
259 }
260         
261 /* We have little choice but to save and restore open files: mmap means we
262  * can really intercept changes in the child.
263  *
264  * We could do non-mmap'ed files on demand, however. */
265 static struct saved_file *save_files(void)
266 {
267         struct saved_file *files = NULL;
268         int i;
269
270         /* Figure out the set of live fds. */
271         for (i = history_num - 2; i >= 0; i--) {
272                 if (history[i].type == FAILTEST_OPEN) {
273                         int fd = history[i].u.open.ret;
274                         /* Only do successful, writable fds. */
275                         if (fd < 0)
276                                 continue;
277
278                         /* If it was closed, cleanup == NULL. */
279                         if (!history[i].cleanup)
280                                 continue;
281
282                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
283                                 files = save_file(files, fd);
284                         } else if ((history[i].u.open.flags & O_WRONLY)
285                                    == O_WRONLY) {
286                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
287                                 abort();
288                         }
289                 }
290         }
291
292         return files;
293 }
294
295 static void restore_files(struct saved_file *s)
296 {
297         while (s) {
298                 struct saved_file *next = s->next;
299
300                 lseek(s->fd, 0, SEEK_SET);
301                 if (write(s->fd, s->contents, s->len) != s->len)
302                         err(1, "Failed to restore %zu bytes", (size_t)s->len);
303                 if (ftruncate(s->fd, s->len) != 0)
304                         err(1, "Failed to trim file to length %zu",
305                             (size_t)s->len);
306                 free(s->contents);
307                 lseek(s->fd, s->off, SEEK_SET);
308                 free(s);
309                 s = next;
310         }
311 }
312
313 static void free_files(struct saved_file *s)
314 {
315         while (s) {
316                 struct saved_file *next = s->next;
317                 free(s->contents);
318                 free(s);
319                 s = next;
320         }
321 }
322
323 /* Free up memory, so valgrind doesn't report leaks. */
324 static void free_everything(void)
325 {
326         unsigned int i;
327
328         /* We don't do this in cleanup: needed even for failed opens. */
329         for (i = 0; i < history_num; i++) {
330                 if (history[i].type == FAILTEST_OPEN)
331                         free((char *)history[i].u.open.pathname);
332         }
333         free(history);
334 }
335
336 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
337 {
338         int i;
339
340         /* For children, we don't care if they "failed" the testing. */
341         if (control_fd != -1)
342                 status = 0;
343
344         if (forced_cleanup)
345                 history_num--;
346
347         /* Cleanup everything, in reverse order. */
348         for (i = history_num - 1; i >= 0; i--) {
349                 if (!history[i].cleanup)
350                         continue;
351                 if (!forced_cleanup) {
352                         printf("Leak at %s:%u: --failpath=%s\n",
353                                history[i].file, history[i].line,
354                                failpath_string());
355                         status = 1;
356                 }
357                 history[i].cleanup(&history[i].u);
358         }
359
360         free_everything();
361         tell_parent(SUCCESS);
362         exit(status);
363 }
364
365 static bool should_fail(struct failtest_call *call)
366 {
367         int status;
368         int control[2], output[2];
369         enum info_type type = UNEXPECTED;
370         char *out = NULL;
371         size_t outlen = 0;
372         struct saved_file *files;
373
374         /* Are we probing? */
375         if (probe_count && --probe_count == 0 && control_fd != -1)
376                 failtest_cleanup(true, 0);
377
378         if (call == &unrecorded_call)
379                 return false;
380
381         if (failpath) {
382                 /* + means continue after end, like normal. */
383                 if (*failpath == '+')
384                         failpath = NULL;
385                 else if (*failpath == '\0') {
386                         /* Continue, but don't inject errors. */
387                         return call->fail = false;
388                 } else {
389                         if (tolower((unsigned char)*failpath)
390                             != info_to_arg[call->type])
391                                 errx(1, "Failpath expected '%c' got '%c'\n",
392                                      info_to_arg[call->type], *failpath);
393                         call->fail = isupper((unsigned char)*(failpath++));
394                         return call->fail;
395                 }
396         }
397
398         /* Attach debugger if they asked for it. */
399         if (debugpath && history_num == strlen(debugpath)) {
400                 unsigned int i;
401
402                 for (i = 0; i < history_num; i++) {
403                         unsigned char c = info_to_arg[history[i].type];
404                         if (history[i].fail)
405                                 c = toupper(c);
406                         if (c != debugpath[i])
407                                 break;
408                 }
409                 if (i == history_num) {
410                         char str[80];
411
412                         /* Don't timeout. */
413                         signal(SIGUSR1, SIG_IGN);
414                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
415                                 getpid(), getpid());
416                         if (system(str) == 0)
417                                 sleep(5);
418                 }
419         }
420
421         if (failtest_hook) {
422                 switch (failtest_hook(history, history_num)) {
423                 case FAIL_OK:
424                         break;
425                 case FAIL_PROBE:
426                         /* Already down probe path?  Stop now. */
427                         if (!probe_count) {
428                                 /* FIXME: We should run *parent* and
429                                  * run probe until calls match up again. */
430                                 probe_count = 3;
431                                 break;
432                         } else {
433                                 /* Child should give up now. */
434                                 if (control_fd != -1)
435                                         failtest_cleanup(true, 0);
436                                 /* Parent, don't fail again. */
437                         }
438                 case FAIL_DONT_FAIL:
439                         call->fail = false;
440                         return false;
441                 default:
442                         abort();
443                 }
444         }
445
446         files = save_files();
447
448         /* We're going to fail in the child. */
449         call->fail = true;
450         if (pipe(control) != 0 || pipe(output) != 0)
451                 err(1, "opening pipe");
452
453         /* Prevent double-printing (in child and parent) */
454         fflush(stdout);
455         child = fork();
456         if (child == -1)
457                 err(1, "forking failed");
458
459         if (child == 0) {
460                 if (tracefd != -1) {
461                         struct timeval diff;
462                         const char *p;
463
464                         diff = time_sub(time_now(), start);
465                         p = failpath_string();
466                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
467                               (int)diff.tv_sec, (int)diff.tv_usec / 10000, p);
468                         free((char *)p);
469                         p = strrchr(history[history_num-1].file, '/');
470                         if (p)
471                                 trace("%s", p+1);
472                         else
473                                 trace("%s", history[history_num-1].file);
474                         trace(":%u)\n", history[history_num-1].line);
475                 }
476                 close(control[0]);
477                 close(output[0]);
478                 dup2(output[1], STDOUT_FILENO);
479                 dup2(output[1], STDERR_FILENO);
480                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
481                         close(output[1]);
482                 control_fd = control[1];
483                 /* Valgrind spots the leak if we don't free these. */
484                 free_files(files);
485                 return true;
486         }
487
488         signal(SIGUSR1, hand_down);
489
490         close(control[1]);
491         close(output[1]);
492
493         /* We grab output so we can display it; we grab writes so we
494          * can compare. */
495         do {
496                 struct pollfd pfd[2];
497                 int ret;
498
499                 pfd[0].fd = output[0];
500                 pfd[0].events = POLLIN|POLLHUP;
501                 pfd[1].fd = control[0];
502                 pfd[1].events = POLLIN|POLLHUP;
503
504                 if (type == SUCCESS)
505                         ret = poll(pfd, 1, failtest_timeout_ms);
506                 else
507                         ret = poll(pfd, 2, failtest_timeout_ms);
508
509                 if (ret == 0)
510                         hand_down(SIGUSR1);
511                 if (ret < 0) {
512                         if (errno == EINTR)
513                                 continue;
514                         err(1, "Poll returned %i", ret);
515                 }
516
517                 if (pfd[0].revents & POLLIN) {
518                         ssize_t len;
519
520                         out = realloc(out, outlen + 8192);
521                         len = read(output[0], out + outlen, 8192);
522                         outlen += len;
523                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
524                         if (read_all(control[0], &type, sizeof(type))) {
525                                 if (type == WRITE) {
526                                         if (!read_write_info(control[0]))
527                                                 break;
528                                 } else if (type == RELEASE_LOCKS) {
529                                         release_locks();
530                                         /* FIXME: Tell them we're done... */
531                                 }
532                         }
533                 } else if (pfd[0].revents & POLLHUP) {
534                         break;
535                 }
536         } while (type != FAILURE);
537
538         close(output[0]);
539         close(control[0]);
540         waitpid(child, &status, 0);
541         if (!WIFEXITED(status)) {
542                 if (WTERMSIG(status) == SIGUSR1)
543                         child_fail(out, outlen, "Timed out");
544                 else
545                         child_fail(out, outlen, "Killed by signal %u: ",
546                                    WTERMSIG(status));
547         }
548         /* Child printed failure already, just pass up exit code. */
549         if (type == FAILURE) {
550                 fprintf(stderr, "%.*s", (int)outlen, out);
551                 tell_parent(type);
552                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
553         }
554         if (WEXITSTATUS(status) != 0)
555                 child_fail(out, outlen, "Exited with status %i: ",
556                            WEXITSTATUS(status));
557
558         free(out);
559         signal(SIGUSR1, SIG_DFL);
560
561         restore_files(files);
562
563         /* We continue onwards without failing. */
564         call->fail = false;
565         return false;
566 }
567
568 static void cleanup_calloc(struct calloc_call *call)
569 {
570         free(call->ret);
571 }
572
573 void *failtest_calloc(size_t nmemb, size_t size,
574                       const char *file, unsigned line)
575 {
576         struct failtest_call *p;
577         struct calloc_call call;
578         call.nmemb = nmemb;
579         call.size = size;
580         p = add_history(FAILTEST_CALLOC, file, line, &call);
581
582         if (should_fail(p)) {
583                 p->u.calloc.ret = NULL;
584                 p->error = ENOMEM;
585         } else {
586                 p->u.calloc.ret = calloc(nmemb, size);
587                 set_cleanup(p, cleanup_calloc, struct calloc_call);
588         }
589         errno = p->error;
590         return p->u.calloc.ret;
591 }
592
593 static void cleanup_malloc(struct malloc_call *call)
594 {
595         free(call->ret);
596 }
597
598 void *failtest_malloc(size_t size, const char *file, unsigned line)
599 {
600         struct failtest_call *p;
601         struct malloc_call call;
602         call.size = size;
603
604         p = add_history(FAILTEST_MALLOC, file, line, &call);
605         if (should_fail(p)) {
606                 p->u.calloc.ret = NULL;
607                 p->error = ENOMEM;
608         } else {
609                 p->u.calloc.ret = malloc(size);
610                 set_cleanup(p, cleanup_malloc, struct malloc_call);
611         }
612         errno = p->error;
613         return p->u.calloc.ret;
614 }
615
616 static void cleanup_realloc(struct realloc_call *call)
617 {
618         free(call->ret);
619 }
620
621 /* Walk back and find out if we got this ptr from a previous routine. */
622 static void fixup_ptr_history(void *ptr, unsigned int last)
623 {
624         int i;
625
626         /* Start at end of history, work back. */
627         for (i = last - 1; i >= 0; i--) {
628                 switch (history[i].type) {
629                 case FAILTEST_REALLOC:
630                         if (history[i].u.realloc.ret == ptr) {
631                                 history[i].cleanup = NULL;
632                                 return;
633                         }
634                         break;
635                 case FAILTEST_MALLOC:
636                         if (history[i].u.malloc.ret == ptr) {
637                                 history[i].cleanup = NULL;
638                                 return;
639                         }
640                         break;
641                 case FAILTEST_CALLOC:
642                         if (history[i].u.calloc.ret == ptr) {
643                                 history[i].cleanup = NULL;
644                                 return;
645                         }
646                         break;
647                 default:
648                         break;
649                 }
650         }
651 }
652
653 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
654 {
655         struct failtest_call *p;
656         struct realloc_call call;
657         call.size = size;
658         p = add_history(FAILTEST_REALLOC, file, line, &call);
659
660         /* FIXME: Try one child moving allocation, one not. */
661         if (should_fail(p)) {
662                 p->u.realloc.ret = NULL;
663                 p->error = ENOMEM;
664         } else {
665                 fixup_ptr_history(ptr, history_num-1);
666                 p->u.realloc.ret = realloc(ptr, size);
667                 set_cleanup(p, cleanup_realloc, struct realloc_call);
668         }
669         errno = p->error;
670         return p->u.realloc.ret;
671 }
672
673 void failtest_free(void *ptr)
674 {
675         fixup_ptr_history(ptr, history_num);
676         free(ptr);
677 }
678
679 static void cleanup_open(struct open_call *call)
680 {
681         close(call->ret);
682 }
683
684 int failtest_open(const char *pathname,
685                   const char *file, unsigned line, ...)
686 {
687         struct failtest_call *p;
688         struct open_call call;
689         va_list ap;
690
691         call.pathname = strdup(pathname);
692         va_start(ap, line);
693         call.flags = va_arg(ap, int);
694         if (call.flags & O_CREAT) {
695                 call.mode = va_arg(ap, int);
696                 va_end(ap);
697         }
698         p = add_history(FAILTEST_OPEN, file, line, &call);
699         /* Avoid memory leak! */
700         if (p == &unrecorded_call)
701                 free((char *)call.pathname);
702         p->u.open.ret = open(pathname, call.flags, call.mode);
703
704         if (p->u.open.ret == -1) {
705                 p->fail = false;
706                 p->error = errno;
707         } else if (should_fail(p)) {
708                 close(p->u.open.ret);
709                 p->u.open.ret = -1;
710                 /* FIXME: Play with error codes? */
711                 p->error = EACCES;
712         } else {
713                 set_cleanup(p, cleanup_open, struct open_call);
714         }
715         errno = p->error;
716         return p->u.open.ret;
717 }
718
719 static void cleanup_pipe(struct pipe_call *call)
720 {
721         if (!call->closed[0])
722                 close(call->fds[0]);
723         if (!call->closed[1])
724                 close(call->fds[1]);
725 }
726
727 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
728 {
729         struct failtest_call *p;
730         struct pipe_call call;
731
732         p = add_history(FAILTEST_PIPE, file, line, &call);
733         if (should_fail(p)) {
734                 p->u.open.ret = -1;
735                 /* FIXME: Play with error codes? */
736                 p->error = EMFILE;
737         } else {
738                 p->u.pipe.ret = pipe(p->u.pipe.fds);
739                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
740                 set_cleanup(p, cleanup_pipe, struct pipe_call);
741         }
742         /* This causes valgrind to notice if they use pipefd[] after failure */
743         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
744         errno = p->error;
745         return p->u.pipe.ret;
746 }
747
748 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
749                        const char *file, unsigned line)
750 {
751         struct failtest_call *p;
752         struct read_call call;
753         call.fd = fd;
754         call.buf = buf;
755         call.count = count;
756         call.off = off;
757         p = add_history(FAILTEST_READ, file, line, &call);
758
759         /* FIXME: Try partial read returns. */
760         if (should_fail(p)) {
761                 p->u.read.ret = -1;
762                 p->error = EIO;
763         } else {
764                 p->u.read.ret = pread(fd, buf, count, off);
765         }
766         errno = p->error;
767         return p->u.read.ret;
768 }
769
770 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
771                         const char *file, unsigned line)
772 {
773         struct failtest_call *p;
774         struct write_call call;
775
776         call.fd = fd;
777         call.buf = buf;
778         call.count = count;
779         call.off = off;
780         p = add_history(FAILTEST_WRITE, file, line, &call);
781
782         /* If we're a child, we need to make sure we write the same thing
783          * to non-files as the parent does, so tell it. */
784         if (control_fd != -1 && off == (off_t)-1) {
785                 enum info_type type = WRITE;
786
787                 write_all(control_fd, &type, sizeof(type));
788                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
789                 write_all(control_fd, buf, count);
790         }
791
792         /* FIXME: Try partial write returns. */
793         if (should_fail(p)) {
794                 p->u.write.ret = -1;
795                 p->error = EIO;
796         } else {
797                 /* FIXME: We assume same write order in parent and child */
798                 if (off == (off_t)-1 && child_writes_num != 0) {
799                         if (child_writes[0].fd != fd)
800                                 errx(1, "Child wrote to fd %u, not %u?",
801                                      child_writes[0].fd, fd);
802                         if (child_writes[0].off != p->u.write.off)
803                                 errx(1, "Child wrote to offset %zu, not %zu?",
804                                      (size_t)child_writes[0].off,
805                                      (size_t)p->u.write.off);
806                         if (child_writes[0].count != count)
807                                 errx(1, "Child wrote length %zu, not %zu?",
808                                      child_writes[0].count, count);
809                         if (memcmp(child_writes[0].buf, buf, count)) {
810                                 child_fail(NULL, 0,
811                                            "Child wrote differently to"
812                                            " fd %u than we did!\n", fd);
813                         }
814                         free((char *)child_writes[0].buf);
815                         child_writes_num--;
816                         memmove(&child_writes[0], &child_writes[1],
817                                 sizeof(child_writes[0]) * child_writes_num);
818
819                         /* Is this is a socket or pipe, child wrote it
820                            already. */
821                         if (p->u.write.off == (off_t)-1) {
822                                 p->u.write.ret = count;
823                                 errno = p->error;
824                                 return p->u.write.ret;
825                         }
826                 }
827                 p->u.write.ret = pwrite(fd, buf, count, off);
828         }
829         errno = p->error;
830         return p->u.write.ret;
831 }
832
833 ssize_t failtest_read(int fd, void *buf, size_t count,
834                       const char *file, unsigned line)
835 {
836         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
837                               file, line);
838 }
839
840 ssize_t failtest_write(int fd, const void *buf, size_t count,
841                        const char *file, unsigned line)
842 {
843         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
844                                file, line);
845 }
846
847 static struct lock_info *WARN_UNUSED_RESULT
848 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
849 {
850         unsigned int i;
851         struct lock_info *l;
852
853         for (i = 0; i < lock_num; i++) {
854                 l = &locks[i];
855
856                 if (l->fd != fd)
857                         continue;
858                 /* Four cases we care about:
859                  * Start overlap:
860                  *      l =    |      |
861                  *      new = |   |
862                  * Mid overlap:
863                  *      l =    |      |
864                  *      new =    |  |
865                  * End overlap:
866                  *      l =    |      |
867                  *      new =      |    |
868                  * Total overlap:
869                  *      l =    |      |
870                  *      new = |         |
871                  */
872                 if (start > l->start && end < l->end) {
873                         /* Mid overlap: trim entry, add new one. */
874                         off_t new_start, new_end;
875                         new_start = end + 1;
876                         new_end = l->end;
877                         l->end = start - 1;
878                         locks = add_lock(locks,
879                                          fd, new_start, new_end, l->type);
880                         l = &locks[i];
881                 } else if (start <= l->start && end >= l->end) {
882                         /* Total overlap: eliminate entry. */
883                         l->end = 0;
884                         l->start = 1;
885                 } else if (end >= l->start && end < l->end) {
886                         /* Start overlap: trim entry. */
887                         l->start = end + 1;
888                 } else if (start > l->start && start <= l->end) {
889                         /* End overlap: trim entry. */
890                         l->end = start-1;
891                 }
892                 /* Nothing left?  Remove it. */
893                 if (l->end < l->start) {
894                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
895                         i--;
896                 }
897         }
898
899         if (type != F_UNLCK) {
900                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
901                 l = &locks[lock_num++];
902                 l->fd = fd;
903                 l->start = start;
904                 l->end = end;
905                 l->type = type;
906         }
907         return locks;
908 }
909
910 /* We trap this so we can record it: we don't fail it. */
911 int failtest_close(int fd, const char *file, unsigned line)
912 {
913         int i;
914         struct close_call call;
915         struct failtest_call *p;
916
917         call.fd = fd;
918         p = add_history(FAILTEST_CLOSE, file, line, &call);
919         p->fail = false;
920
921         /* Consume close from failpath. */
922         if (failpath)
923                 if (should_fail(p))
924                         abort();
925
926         if (fd < 0)
927                 return close(fd);
928
929         /* Trace history to find source of fd. */
930         for (i = history_num-1; i >= 0; i--) {
931                 switch (history[i].type) {
932                 case FAILTEST_PIPE:
933                         /* From a pipe? */
934                         if (history[i].u.pipe.fds[0] == fd) {
935                                 assert(!history[i].u.pipe.closed[0]);
936                                 history[i].u.pipe.closed[0] = true;
937                                 if (history[i].u.pipe.closed[1])
938                                         history[i].cleanup = NULL;
939                                 goto out;
940                         }
941                         if (history[i].u.pipe.fds[1] == fd) {
942                                 assert(!history[i].u.pipe.closed[1]);
943                                 history[i].u.pipe.closed[1] = true;
944                                 if (history[i].u.pipe.closed[0])
945                                         history[i].cleanup = NULL;
946                                 goto out;
947                         }
948                         break;
949                 case FAILTEST_OPEN:
950                         if (history[i].u.open.ret == fd) {
951                                 assert((void *)history[i].cleanup
952                                        == (void *)cleanup_open);
953                                 history[i].cleanup = NULL;
954                                 goto out;
955                         }
956                         break;
957                 default:
958                         break;
959                 }
960         }
961
962 out:
963         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
964         return close(fd);
965 }
966
967 /* Zero length means "to end of file" */
968 static off_t end_of(off_t start, off_t len)
969 {
970         if (len == 0)
971                 return off_max();
972         return start + len - 1;
973 }
974
975 /* FIXME: This only handles locks, really. */
976 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
977 {
978         struct failtest_call *p;
979         struct fcntl_call call;
980         va_list ap;
981
982         call.fd = fd;
983         call.cmd = cmd;
984
985         /* Argument extraction. */
986         switch (cmd) {
987         case F_SETFL:
988         case F_SETFD:
989                 va_start(ap, cmd);
990                 call.arg.l = va_arg(ap, long);
991                 va_end(ap);
992                 return fcntl(fd, cmd, call.arg.l);
993         case F_GETFD:
994         case F_GETFL:
995                 return fcntl(fd, cmd);
996         case F_GETLK:
997                 get_locks();
998                 va_start(ap, cmd);
999                 call.arg.fl = *va_arg(ap, struct flock *);
1000                 va_end(ap);
1001                 return fcntl(fd, cmd, &call.arg.fl);
1002         case F_SETLK:
1003         case F_SETLKW:
1004                 va_start(ap, cmd);
1005                 call.arg.fl = *va_arg(ap, struct flock *);
1006                 va_end(ap);
1007                 break;
1008         default:
1009                 /* This means you need to implement it here. */
1010                 err(1, "failtest: unknown fcntl %u", cmd);
1011         }
1012
1013         p = add_history(FAILTEST_FCNTL, file, line, &call);
1014
1015         if (should_fail(p)) {
1016                 p->u.fcntl.ret = -1;
1017                 if (p->u.fcntl.cmd == F_SETLK)
1018                         p->error = EAGAIN;
1019                 else
1020                         p->error = EDEADLK;
1021         } else {
1022                 get_locks();
1023                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1024                                        &p->u.fcntl.arg.fl);
1025                 if (p->u.fcntl.ret == -1)
1026                         p->error = errno;
1027                 else {
1028                         /* We don't handle anything else yet. */
1029                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1030                         locks = add_lock(locks,
1031                                          p->u.fcntl.fd,
1032                                          p->u.fcntl.arg.fl.l_start,
1033                                          end_of(p->u.fcntl.arg.fl.l_start,
1034                                                 p->u.fcntl.arg.fl.l_len),
1035                                          p->u.fcntl.arg.fl.l_type);
1036                 }
1037         }
1038         errno = p->error;
1039         return p->u.fcntl.ret;
1040 }
1041
1042 pid_t failtest_getpid(const char *file, unsigned line)
1043 {
1044         /* You must call failtest_init first! */
1045         assert(orig_pid);
1046         return orig_pid;
1047 }
1048         
1049 void failtest_init(int argc, char *argv[])
1050 {
1051         unsigned int i;
1052
1053         orig_pid = getpid();
1054                 
1055         for (i = 1; i < argc; i++) {
1056                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1057                         failpath = argv[i] + strlen("--failpath=");
1058                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1059                         tracefd = dup(STDERR_FILENO);
1060                         failtest_timeout_ms = -1;
1061                 } else if (!strncmp(argv[i], "--debugpath=",
1062                                     strlen("--debugpath="))) {
1063                         debugpath = argv[i] + strlen("--debugpath=");
1064                 }
1065         }
1066         start = time_now();
1067 }
1068
1069 bool failtest_has_failed(void)
1070 {
1071         return control_fd != -1;
1072 }
1073
1074 void failtest_exit(int status)
1075 {
1076         if (failtest_exit_check) {
1077                 if (!failtest_exit_check(history, history_num))
1078                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1079         }
1080
1081         failtest_cleanup(false, status);
1082 }