failtest: rely on the save/restore of files, don't use write cleanup.
[ccan] / ccan / failtest / failtest.c
1 #include <stdarg.h>
2 #include <string.h>
3 #include <stdio.h>
4 #include <stdarg.h>
5 #include <ctype.h>
6 #include <err.h>
7 #include <unistd.h>
8 #include <poll.h>
9 #include <errno.h>
10 #include <sys/types.h>
11 #include <sys/wait.h>
12 #include <sys/stat.h>
13 #include <sys/time.h>
14 #include <assert.h>
15 #include <ccan/read_write_all/read_write_all.h>
16 #include <ccan/failtest/failtest_proto.h>
17 #include <ccan/failtest/failtest.h>
18 #include <ccan/build_assert/build_assert.h>
19
20 bool (*failtest_hook)(struct failtest_call *history, unsigned num)
21 = failtest_default_hook;
22
23 static int tracefd = -1;
24
25 unsigned int failtest_timeout_ms = 20000;
26
27 const char *failpath;
28
29 enum info_type {
30         WRITE,
31         RELEASE_LOCKS,
32         FAILURE,
33         SUCCESS,
34         UNEXPECTED
35 };
36
37 struct lock_info {
38         int fd;
39         /* end is inclusive: you can't have a 0-byte lock. */
40         off_t start, end;
41         int type;
42 };
43
44 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
45
46 static struct failtest_call *history = NULL;
47 static unsigned int history_num = 0;
48 static int control_fd = -1;
49 static struct timeval start;
50
51 static struct write_call *child_writes = NULL;
52 static unsigned int child_writes_num = 0;
53
54 static pid_t lock_owner;
55 static struct lock_info *locks = NULL;
56 static unsigned int lock_num = 0;
57
58 static const char info_to_arg[] = "mceoprwf";
59
60 /* Dummy call used for failtest_undo wrappers. */
61 static struct failtest_call unrecorded_call;
62
63 static struct failtest_call *add_history_(enum failtest_call_type type,
64                                           const char *file,
65                                           unsigned int line,
66                                           const void *elem,
67                                           size_t elem_size)
68 {
69         /* NULL file is how we suppress failure. */
70         if (!file)
71                 return &unrecorded_call;
72
73         history = realloc(history, (history_num + 1) * sizeof(*history));
74         history[history_num].type = type;
75         history[history_num].file = file;
76         history[history_num].line = line;
77         history[history_num].cleanup = NULL;
78         memcpy(&history[history_num].u, elem, elem_size);
79         return &history[history_num++];
80 }
81
82 #define add_history(type, file, line, elem) \
83         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
84
85 #define set_cleanup(call, clean, type)                  \
86         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
87
88 bool failtest_default_hook(struct failtest_call *history, unsigned num)
89 {
90         return true;
91 }
92
93 static bool read_write_info(int fd)
94 {
95         struct write_call *w;
96         char *buf;
97
98         /* We don't need all of this, but it's simple. */
99         child_writes = realloc(child_writes,
100                                (child_writes_num+1) * sizeof(child_writes[0]));
101         w = &child_writes[child_writes_num];
102         if (!read_all(fd, w, sizeof(*w)))
103                 return false;
104
105         w->buf = buf = malloc(w->count);
106         if (!read_all(fd, buf, w->count))
107                 return false;
108
109         child_writes_num++;
110         return true;
111 }
112
113 static char *failpath_string(void)
114 {
115         unsigned int i;
116         char *ret = malloc(history_num + 1);
117
118         for (i = 0; i < history_num; i++) {
119                 ret[i] = info_to_arg[history[i].type];
120                 if (history[i].fail)
121                         ret[i] = toupper(ret[i]);
122         }
123         ret[i] = '\0';
124         return ret;
125 }
126
127 static void tell_parent(enum info_type type)
128 {
129         if (control_fd != -1)
130                 write_all(control_fd, &type, sizeof(type));
131 }
132
133 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
134 {
135         va_list ap;
136         char *path = failpath_string();
137
138         va_start(ap, fmt);
139         vfprintf(stderr, fmt, ap);
140         va_end(ap);
141
142         fprintf(stderr, "%.*s", (int)outlen, out);
143         printf("To reproduce: --failpath=%s\n", path);
144         free(path);
145         tell_parent(FAILURE);
146         exit(1);
147 }
148
149 static pid_t child;
150
151 static void hand_down(int signal)
152 {
153         kill(child, signal);
154 }
155
156 static void release_locks(void)
157 {
158         /* Locks were never acquired/reacquired? */
159         if (lock_owner == 0)
160                 return;
161
162         /* We own them?  Release them all. */
163         if (lock_owner == getpid()) {
164                 unsigned int i;
165                 struct flock fl;
166                 fl.l_type = F_UNLCK;
167                 fl.l_whence = SEEK_SET;
168                 fl.l_start = 0;
169                 fl.l_len = 0;
170
171                 for (i = 0; i < lock_num; i++)
172                         fcntl(locks[i].fd, F_SETLK, &fl);
173         } else {
174                 /* Our parent must have them; pass request up. */
175                 enum info_type type = RELEASE_LOCKS;
176                 assert(control_fd != -1);
177                 write_all(control_fd, &type, sizeof(type));
178         }
179         lock_owner = 0;
180 }
181
182 /* off_t is a signed type.  Getting its max is non-trivial. */
183 static off_t off_max(void)
184 {
185         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
186         if (sizeof(off_t) == 4)
187                 return (off_t)0x7FFFFFF;
188         else
189                 return (off_t)0x7FFFFFFFFFFFFFFULL;
190 }
191
192 static void get_locks(void)
193 {
194         unsigned int i;
195         struct flock fl;
196
197         if (lock_owner == getpid())
198                 return;
199
200         if (lock_owner != 0) {
201                 enum info_type type = RELEASE_LOCKS;
202                 assert(control_fd != -1);
203                 write_all(control_fd, &type, sizeof(type));
204         }
205
206         fl.l_whence = SEEK_SET;
207
208         for (i = 0; i < lock_num; i++) {
209                 fl.l_type = locks[i].type;
210                 fl.l_start = locks[i].start;
211                 if (locks[i].end == off_max())
212                         fl.l_len = 0;
213                 else
214                         fl.l_len = locks[i].end - locks[i].start + 1;
215
216                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
217                         abort();
218         }
219         lock_owner = getpid();
220 }
221
222 static void trace_str(const char *str)
223 {
224         ssize_t ret;
225
226         while ((ret = write(tracefd, str, strlen(str))) <= 0) {
227                 str += ret;
228                 if (!*str)
229                         return;
230         }
231         err(1, "Writing trace.");
232 }
233
234 struct saved_file {
235         struct saved_file *next;
236         int fd;
237         void *contents;
238         off_t off, len;
239 };
240
241 static struct saved_file *save_file(struct saved_file *next, int fd)
242 {
243         struct saved_file *s = malloc(sizeof(*s));
244
245         s->next = next;
246         s->fd = fd;
247         s->off = lseek(fd, 0, SEEK_CUR);
248         /* Special file?  Erk... */
249         assert(s->off != -1);
250         s->len = lseek(fd, 0, SEEK_END);
251         lseek(fd, 0, SEEK_SET);
252         s->contents = malloc(s->len);
253         read(fd, s->contents, s->len);
254         lseek(fd, s->off, SEEK_SET);
255         return s;
256 }
257         
258 /* We have little choice but to save and restore open files: mmap means we
259  * can really intercept changes in the child.
260  *
261  * We could do non-mmap'ed files on demand, however. */
262 static struct saved_file *save_files(void)
263 {
264         struct saved_file *files = NULL;
265         int i;
266
267         /* Figure out the set of live fds. */
268         for (i = history_num - 2; i >= 0; i--) {
269                 if (history[i].type == FAILTEST_OPEN) {
270                         int fd = history[i].u.open.ret;
271                         /* Only do successful, writable fds. */
272                         if (fd < 0)
273                                 continue;
274
275                         /* If it was closed, cleanup == NULL. */
276                         if (!history[i].cleanup)
277                                 continue;
278
279                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
280                                 files = save_file(files, fd);
281                         } else if ((history[i].u.open.flags & O_WRONLY)
282                                    == O_WRONLY) {
283                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
284                                 abort();
285                         }
286                 }
287         }
288
289         return files;
290 }
291
292 static void restore_files(struct saved_file *s)
293 {
294         while (s) {
295                 struct saved_file *next = s->next;
296
297                 lseek(s->fd, 0, SEEK_SET);
298                 write(s->fd, s->contents, s->len);
299                 ftruncate(s->fd, s->len);
300                 free(s->contents);
301                 lseek(s->fd, s->off, SEEK_SET);
302                 free(s);
303                 s = next;
304         }
305 }
306
307 static bool should_fail(struct failtest_call *call)
308 {
309         int status;
310         int control[2], output[2];
311         enum info_type type = UNEXPECTED;
312         char *out = NULL;
313         size_t outlen = 0;
314         struct saved_file *files;
315
316         if (call == &unrecorded_call)
317                 return false;
318
319         if (failpath) {
320                 /* + means continue after end, like normal. */
321                 if (*failpath == '+')
322                         failpath = NULL;
323                 else {
324                         if (tolower(*failpath) != info_to_arg[call->type])
325                                 errx(1, "Failpath expected '%c' got '%c'\n",
326                                      info_to_arg[call->type], *failpath);
327                         call->fail = isupper(*(failpath++));
328                         return call->fail;
329                 }
330         }
331
332         if (!failtest_hook(history, history_num)) {
333                 call->fail = false;
334                 return false;
335         }
336
337         files = save_files();
338
339         /* We're going to fail in the child. */
340         call->fail = true;
341         if (pipe(control) != 0 || pipe(output) != 0)
342                 err(1, "opening pipe");
343
344         /* Prevent double-printing (in child and parent) */
345         fflush(stdout);
346         child = fork();
347         if (child == -1)
348                 err(1, "forking failed");
349
350         if (child == 0) {
351                 if (tracefd != -1) {
352                         struct timeval now;
353                         char str[50], *p;
354                         gettimeofday(&now, NULL);
355                         if (now.tv_usec < start.tv_usec) {
356                                 now.tv_sec--;
357                                 now.tv_usec += 1000000;
358                         }
359                         now.tv_usec -= start.tv_usec;
360                         now.tv_sec -= start.tv_sec;
361                         sprintf(str, "%u (%u.%02u): ", getpid(),
362                                 (int)now.tv_sec, (int)now.tv_usec / 10000);
363                         trace_str(str);
364                         p = failpath_string();
365                         trace_str(p);
366                         free(p);
367                         trace_str("(");
368                         p = strchr(history[history_num-1].file, '/');
369                         if (p)
370                                 trace_str(p+1);
371                         else
372                                 trace_str(history[history_num-1].file);
373                         sprintf(str, ":%u)\n", history[history_num-1].line);
374                         trace_str(str);
375                 }
376                 close(control[0]);
377                 close(output[0]);
378                 dup2(output[1], STDOUT_FILENO);
379                 dup2(output[1], STDERR_FILENO);
380                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
381                         close(output[1]);
382                 control_fd = control[1];
383                 return true;
384         }
385
386         signal(SIGUSR1, hand_down);
387
388         close(control[1]);
389         close(output[1]);
390
391         /* We grab output so we can display it; we grab writes so we
392          * can compare. */
393         do {
394                 struct pollfd pfd[2];
395                 int ret;
396
397                 pfd[0].fd = output[0];
398                 pfd[0].events = POLLIN|POLLHUP;
399                 pfd[1].fd = control[0];
400                 pfd[1].events = POLLIN|POLLHUP;
401
402                 if (type == SUCCESS)
403                         ret = poll(pfd, 1, failtest_timeout_ms);
404                 else
405                         ret = poll(pfd, 2, failtest_timeout_ms);
406
407                 if (ret <= 0)
408                         hand_down(SIGUSR1);
409
410                 if (pfd[0].revents & POLLIN) {
411                         ssize_t len;
412
413                         out = realloc(out, outlen + 8192);
414                         len = read(output[0], out + outlen, 8192);
415                         outlen += len;
416                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
417                         if (read_all(control[0], &type, sizeof(type))) {
418                                 if (type == WRITE) {
419                                         if (!read_write_info(control[0]))
420                                                 break;
421                                 } else if (type == RELEASE_LOCKS) {
422                                         release_locks();
423                                         /* FIXME: Tell them we're done... */
424                                 }
425                         }
426                 } else if (pfd[0].revents & POLLHUP) {
427                         break;
428                 }
429         } while (type != FAILURE);
430
431         close(output[0]);
432         close(control[0]);
433         waitpid(child, &status, 0);
434         if (!WIFEXITED(status))
435                 child_fail(out, outlen, "Killed by signal %u: ",
436                            WTERMSIG(status));
437         /* Child printed failure already, just pass up exit code. */
438         if (type == FAILURE) {
439                 fprintf(stderr, "%.*s", (int)outlen, out);
440                 tell_parent(type);
441                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
442         }
443         if (WEXITSTATUS(status) != 0)
444                 child_fail(out, outlen, "Exited with status %i: ",
445                            WEXITSTATUS(status));
446
447         free(out);
448         signal(SIGUSR1, SIG_DFL);
449
450         restore_files(files);
451
452         /* We continue onwards without failing. */
453         call->fail = false;
454         return false;
455 }
456
457 static void cleanup_calloc(struct calloc_call *call)
458 {
459         free(call->ret);
460 }
461
462 void *failtest_calloc(size_t nmemb, size_t size,
463                       const char *file, unsigned line)
464 {
465         struct failtest_call *p;
466         struct calloc_call call;
467         call.nmemb = nmemb;
468         call.size = size;
469         p = add_history(FAILTEST_CALLOC, file, line, &call);
470
471         if (should_fail(p)) {
472                 p->u.calloc.ret = NULL;
473                 p->error = ENOMEM;
474         } else {
475                 p->u.calloc.ret = calloc(nmemb, size);
476                 set_cleanup(p, cleanup_calloc, struct calloc_call);
477         }
478         errno = p->error;
479         return p->u.calloc.ret;
480 }
481
482 static void cleanup_malloc(struct malloc_call *call)
483 {
484         free(call->ret);
485 }
486
487 void *failtest_malloc(size_t size, const char *file, unsigned line)
488 {
489         struct failtest_call *p;
490         struct malloc_call call;
491         call.size = size;
492
493         p = add_history(FAILTEST_MALLOC, file, line, &call);
494         if (should_fail(p)) {
495                 p->u.calloc.ret = NULL;
496                 p->error = ENOMEM;
497         } else {
498                 p->u.calloc.ret = malloc(size);
499                 set_cleanup(p, cleanup_malloc, struct malloc_call);
500         }
501         errno = p->error;
502         return p->u.calloc.ret;
503 }
504
505 static void cleanup_realloc(struct realloc_call *call)
506 {
507         free(call->ret);
508 }
509
510 /* Walk back and find out if we got this ptr from a previous routine. */
511 static void fixup_ptr_history(void *ptr, unsigned int last)
512 {
513         int i;
514
515         /* Start at end of history, work back. */
516         for (i = last - 1; i >= 0; i--) {
517                 switch (history[i].type) {
518                 case FAILTEST_REALLOC:
519                         if (history[i].u.realloc.ret == ptr) {
520                                 history[i].cleanup = NULL;
521                                 return;
522                         }
523                         break;
524                 case FAILTEST_MALLOC:
525                         if (history[i].u.malloc.ret == ptr) {
526                                 history[i].cleanup = NULL;
527                                 return;
528                         }
529                         break;
530                 case FAILTEST_CALLOC:
531                         if (history[i].u.calloc.ret == ptr) {
532                                 history[i].cleanup = NULL;
533                                 return;
534                         }
535                         break;
536                 default:
537                         break;
538                 }
539         }
540 }
541
542 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
543 {
544         struct failtest_call *p;
545         struct realloc_call call;
546         call.size = size;
547         p = add_history(FAILTEST_REALLOC, file, line, &call);
548
549         /* FIXME: Try one child moving allocation, one not. */
550         if (should_fail(p)) {
551                 p->u.realloc.ret = NULL;
552                 p->error = ENOMEM;
553         } else {
554                 fixup_ptr_history(ptr, history_num-1);
555                 p->u.realloc.ret = realloc(ptr, size);
556                 set_cleanup(p, cleanup_realloc, struct realloc_call);
557         }
558         errno = p->error;
559         return p->u.realloc.ret;
560 }
561
562 void failtest_free(void *ptr)
563 {
564         fixup_ptr_history(ptr, history_num);
565         free(ptr);
566 }
567
568 static void cleanup_open(struct open_call *call)
569 {
570         close(call->ret);
571 }
572
573 int failtest_open(const char *pathname,
574                   const char *file, unsigned line, ...)
575 {
576         struct failtest_call *p;
577         struct open_call call;
578         va_list ap;
579
580         call.pathname = strdup(pathname);
581         va_start(ap, line);
582         call.flags = va_arg(ap, int);
583         if (call.flags & O_CREAT) {
584                 call.mode = va_arg(ap, mode_t);
585                 va_end(ap);
586         }
587         p = add_history(FAILTEST_OPEN, file, line, &call);
588         /* Avoid memory leak! */
589         if (p == &unrecorded_call)
590                 free((char *)call.pathname);
591         if (should_fail(p)) {
592                 p->u.open.ret = -1;
593                 /* FIXME: Play with error codes? */
594                 p->error = EACCES;
595         } else {
596                 p->u.open.ret = open(pathname, call.flags, call.mode);
597                 set_cleanup(p, cleanup_open, struct open_call);
598         }
599         errno = p->error;
600         return p->u.open.ret;
601 }
602
603 static void cleanup_pipe(struct pipe_call *call)
604 {
605         if (!call->closed[0])
606                 close(call->fds[0]);
607         if (!call->closed[1])
608                 close(call->fds[1]);
609 }
610
611 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
612 {
613         struct failtest_call *p;
614         struct pipe_call call;
615
616         p = add_history(FAILTEST_PIPE, file, line, &call);
617         if (should_fail(p)) {
618                 p->u.open.ret = -1;
619                 /* FIXME: Play with error codes? */
620                 p->error = EMFILE;
621         } else {
622                 p->u.pipe.ret = pipe(p->u.pipe.fds);
623                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
624                 set_cleanup(p, cleanup_pipe, struct pipe_call);
625         }
626         /* This causes valgrind to notice if they use pipefd[] after failure */
627         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
628         errno = p->error;
629         return p->u.pipe.ret;
630 }
631
632 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
633                        const char *file, unsigned line)
634 {
635         struct failtest_call *p;
636         struct read_call call;
637         call.fd = fd;
638         call.buf = buf;
639         call.count = count;
640         call.off = off;
641         p = add_history(FAILTEST_READ, file, line, &call);
642
643         /* FIXME: Try partial read returns. */
644         if (should_fail(p)) {
645                 p->u.read.ret = -1;
646                 p->error = EIO;
647         } else {
648                 p->u.read.ret = pread(fd, buf, count, off);
649         }
650         errno = p->error;
651         return p->u.read.ret;
652 }
653
654 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
655                         const char *file, unsigned line)
656 {
657         struct failtest_call *p;
658         struct write_call call;
659
660         call.fd = fd;
661         call.buf = buf;
662         call.count = count;
663         call.off = off;
664         p = add_history(FAILTEST_WRITE, file, line, &call);
665
666         /* If we're a child, tell parent about write. */
667         if (control_fd != -1) {
668                 enum info_type type = WRITE;
669
670                 write_all(control_fd, &type, sizeof(type));
671                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
672                 write_all(control_fd, buf, count);
673         }
674
675         /* FIXME: Try partial write returns. */
676         if (should_fail(p)) {
677                 p->u.write.ret = -1;
678                 p->error = EIO;
679         } else {
680                 /* FIXME: We assume same write order in parent and child */
681                 if (child_writes_num != 0) {
682                         if (child_writes[0].fd != fd)
683                                 errx(1, "Child wrote to fd %u, not %u?",
684                                      child_writes[0].fd, fd);
685                         if (child_writes[0].off != p->u.write.off)
686                                 errx(1, "Child wrote to offset %zu, not %zu?",
687                                      (size_t)child_writes[0].off,
688                                      (size_t)p->u.write.off);
689                         if (child_writes[0].count != count)
690                                 errx(1, "Child wrote length %zu, not %zu?",
691                                      child_writes[0].count, count);
692                         if (memcmp(child_writes[0].buf, buf, count)) {
693                                 child_fail(NULL, 0,
694                                            "Child wrote differently to"
695                                            " fd %u than we did!\n", fd);
696                         }
697                         free((char *)child_writes[0].buf);
698                         child_writes_num--;
699                         memmove(&child_writes[0], &child_writes[1],
700                                 sizeof(child_writes[0]) * child_writes_num);
701
702                         /* Is this is a socket or pipe, child wrote it
703                            already. */
704                         if (p->u.write.off == (off_t)-1) {
705                                 p->u.write.ret = count;
706                                 errno = p->error;
707                                 return p->u.write.ret;
708                         }
709                 }
710                 p->u.write.ret = pwrite(fd, buf, count, off);
711         }
712         errno = p->error;
713         return p->u.write.ret;
714 }
715
716 ssize_t failtest_read(int fd, void *buf, size_t count,
717                       const char *file, unsigned line)
718 {
719         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
720                               file, line);
721 }
722
723 ssize_t failtest_write(int fd, const void *buf, size_t count,
724                        const char *file, unsigned line)
725 {
726         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
727                                file, line);
728 }
729
730 static struct lock_info *WARN_UNUSED_RESULT
731 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
732 {
733         unsigned int i;
734         struct lock_info *l;
735
736         for (i = 0; i < lock_num; i++) {
737                 l = &locks[i];
738
739                 if (l->fd != fd)
740                         continue;
741                 /* Four cases we care about:
742                  * Start overlap:
743                  *      l =    |      |
744                  *      new = |   |
745                  * Mid overlap:
746                  *      l =    |      |
747                  *      new =    |  |
748                  * End overlap:
749                  *      l =    |      |
750                  *      new =      |    |
751                  * Total overlap:
752                  *      l =    |      |
753                  *      new = |         |
754                  */
755                 if (start > l->start && end < l->end) {
756                         /* Mid overlap: trim entry, add new one. */
757                         off_t new_start, new_end;
758                         new_start = end + 1;
759                         new_end = l->end;
760                         l->end = start - 1;
761                         locks = add_lock(locks,
762                                          fd, new_start, new_end, l->type);
763                         l = &locks[i];
764                 } else if (start <= l->start && end >= l->end) {
765                         /* Total overlap: eliminate entry. */
766                         l->end = 0;
767                         l->start = 1;
768                 } else if (end >= l->start && end < l->end) {
769                         /* Start overlap: trim entry. */
770                         l->start = end + 1;
771                 } else if (start > l->start && start <= l->end) {
772                         /* End overlap: trim entry. */
773                         l->end = start-1;
774                 }
775                 /* Nothing left?  Remove it. */
776                 if (l->end < l->start) {
777                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
778                         i--;
779                 }
780         }
781
782         if (type != F_UNLCK) {
783                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
784                 l = &locks[lock_num++];
785                 l->fd = fd;
786                 l->start = start;
787                 l->end = end;
788                 l->type = type;
789         }
790         return locks;
791 }
792
793 /* We trap this so we can record it: we don't fail it. */
794 int failtest_close(int fd)
795 {
796         int i;
797
798         if (fd < 0)
799                 return close(fd);
800
801         /* Trace history to find source of fd. */
802         for (i = history_num-1; i >= 0; i--) {
803                 switch (history[i].type) {
804                 case FAILTEST_PIPE:
805                         /* From a pipe? */
806                         if (history[i].u.pipe.fds[0] == fd) {
807                                 assert(!history[i].u.pipe.closed[0]);
808                                 history[i].u.pipe.closed[0] = true;
809                                 if (history[i].u.pipe.closed[1])
810                                         history[i].cleanup = NULL;
811                                 goto out;
812                         }
813                         if (history[i].u.pipe.fds[1] == fd) {
814                                 assert(!history[i].u.pipe.closed[1]);
815                                 history[i].u.pipe.closed[1] = true;
816                                 if (history[i].u.pipe.closed[0])
817                                         history[i].cleanup = NULL;
818                                 goto out;
819                         }
820                         break;
821                 case FAILTEST_OPEN:
822                         if (history[i].u.open.ret == fd) {
823                                 assert((void *)history[i].cleanup
824                                        == (void *)cleanup_open);
825                                 history[i].cleanup = NULL;
826                                 goto out;
827                         }
828                         break;
829                 default:
830                         break;
831                 }
832         }
833
834 out:
835         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
836         return close(fd);
837 }
838
839 /* Zero length means "to end of file" */
840 static off_t end_of(off_t start, off_t len)
841 {
842         if (len == 0)
843                 return off_max();
844         return start + len - 1;
845 }
846
847 /* FIXME: This only handles locks, really. */
848 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
849 {
850         struct failtest_call *p;
851         struct fcntl_call call;
852         va_list ap;
853
854         call.fd = fd;
855         call.cmd = cmd;
856
857         /* Argument extraction. */
858         switch (cmd) {
859         case F_SETFL:
860         case F_SETFD:
861                 va_start(ap, cmd);
862                 call.arg.l = va_arg(ap, long);
863                 va_end(ap);
864                 return fcntl(fd, cmd, call.arg.l);
865         case F_GETFD:
866         case F_GETFL:
867                 return fcntl(fd, cmd);
868         case F_GETLK:
869                 get_locks();
870                 va_start(ap, cmd);
871                 call.arg.fl = *va_arg(ap, struct flock *);
872                 va_end(ap);
873                 return fcntl(fd, cmd, &call.arg.fl);
874         case F_SETLK:
875         case F_SETLKW:
876                 va_start(ap, cmd);
877                 call.arg.fl = *va_arg(ap, struct flock *);
878                 va_end(ap);
879                 break;
880         default:
881                 /* This means you need to implement it here. */
882                 err(1, "failtest: unknown fcntl %u", cmd);
883         }
884
885         p = add_history(FAILTEST_FCNTL, file, line, &call);
886         get_locks();
887
888         if (should_fail(p)) {
889                 p->u.fcntl.ret = -1;
890                 if (p->u.fcntl.cmd == F_SETLK)
891                         p->error = EAGAIN;
892                 else
893                         p->error = EDEADLK;
894         } else {
895                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
896                                        &p->u.fcntl.arg.fl);
897                 if (p->u.fcntl.ret == -1)
898                         p->error = errno;
899                 else {
900                         /* We don't handle anything else yet. */
901                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
902                         locks = add_lock(locks,
903                                          p->u.fcntl.fd,
904                                          p->u.fcntl.arg.fl.l_start,
905                                          end_of(p->u.fcntl.arg.fl.l_start,
906                                                 p->u.fcntl.arg.fl.l_len),
907                                          p->u.fcntl.arg.fl.l_type);
908                 }
909         }
910         errno = p->error;
911         return p->u.fcntl.ret;
912 }
913
914 void failtest_init(int argc, char *argv[])
915 {
916         unsigned int i;
917
918         for (i = 1; i < argc; i++) {
919                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
920                         failpath = argv[i] + strlen("--failpath=");
921                 } else if (strcmp(argv[i], "--tracepath") == 0) {
922                         tracefd = dup(STDERR_FILENO);
923                         failtest_timeout_ms = -1;
924                 }
925         }
926         gettimeofday(&start, NULL);
927 }
928
929 /* Free up memory, so valgrind doesn't report leaks. */
930 static void free_everything(void)
931 {
932         unsigned int i;
933
934         /* We don't do this in cleanup: needed even for failed opens. */
935         for (i = 0; i < history_num; i++) {
936                 if (history[i].type == FAILTEST_OPEN)
937                         free((char *)history[i].u.open.pathname);
938         }
939         free(history);
940 }
941
942 void failtest_exit(int status)
943 {
944         int i;
945
946         if (control_fd == -1) {
947                 free_everything();
948                 exit(status);
949         }
950
951         if (failtest_exit_check) {
952                 if (!failtest_exit_check(history, history_num))
953                         child_fail(NULL, 0, "failtest_exit_check failed\n");
954         }
955
956         /* Cleanup everything, in reverse order. */
957         for (i = history_num - 1; i >= 0; i--)
958                 if (history[i].cleanup)
959                         history[i].cleanup(&history[i].u);
960
961         free_everything();
962         tell_parent(SUCCESS);
963         exit(0);
964 }