failtest: detect leaks in children.
[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 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
21
22 static int tracefd = -1;
23
24 unsigned int failtest_timeout_ms = 20000;
25
26 const char *failpath;
27 const char *debugpath;
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 static unsigned int probe_count = 0;
51
52 static struct write_call *child_writes = NULL;
53 static unsigned int child_writes_num = 0;
54
55 static pid_t lock_owner;
56 static struct lock_info *locks = NULL;
57 static unsigned int lock_num = 0;
58
59 static const char info_to_arg[] = "mceoxprwf";
60
61 /* Dummy call used for failtest_undo wrappers. */
62 static struct failtest_call unrecorded_call;
63
64 static struct failtest_call *add_history_(enum failtest_call_type type,
65                                           const char *file,
66                                           unsigned int line,
67                                           const void *elem,
68                                           size_t elem_size)
69 {
70         /* NULL file is how we suppress failure. */
71         if (!file)
72                 return &unrecorded_call;
73
74         history = realloc(history, (history_num + 1) * sizeof(*history));
75         history[history_num].type = type;
76         history[history_num].file = file;
77         history[history_num].line = line;
78         history[history_num].cleanup = NULL;
79         memcpy(&history[history_num].u, elem, elem_size);
80         return &history[history_num++];
81 }
82
83 #define add_history(type, file, line, elem) \
84         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
85
86 #define set_cleanup(call, clean, type)                  \
87         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
88
89 static bool read_write_info(int fd)
90 {
91         struct write_call *w;
92         char *buf;
93
94         /* We don't need all of this, but it's simple. */
95         child_writes = realloc(child_writes,
96                                (child_writes_num+1) * sizeof(child_writes[0]));
97         w = &child_writes[child_writes_num];
98         if (!read_all(fd, w, sizeof(*w)))
99                 return false;
100
101         w->buf = buf = malloc(w->count);
102         if (!read_all(fd, buf, w->count))
103                 return false;
104
105         child_writes_num++;
106         return true;
107 }
108
109 static char *failpath_string(void)
110 {
111         unsigned int i;
112         char *ret = malloc(history_num + 1);
113
114         for (i = 0; i < history_num; i++) {
115                 ret[i] = info_to_arg[history[i].type];
116                 if (history[i].fail)
117                         ret[i] = toupper(ret[i]);
118         }
119         ret[i] = '\0';
120         return ret;
121 }
122
123 static void tell_parent(enum info_type type)
124 {
125         if (control_fd != -1)
126                 write_all(control_fd, &type, sizeof(type));
127 }
128
129 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
130 {
131         va_list ap;
132         char *path = failpath_string();
133
134         va_start(ap, fmt);
135         vfprintf(stderr, fmt, ap);
136         va_end(ap);
137
138         fprintf(stderr, "%.*s", (int)outlen, out);
139         printf("To reproduce: --failpath=%s\n", path);
140         free(path);
141         tell_parent(FAILURE);
142         exit(1);
143 }
144
145 static pid_t child;
146
147 static void hand_down(int signal)
148 {
149         kill(child, signal);
150 }
151
152 static void release_locks(void)
153 {
154         /* Locks were never acquired/reacquired? */
155         if (lock_owner == 0)
156                 return;
157
158         /* We own them?  Release them all. */
159         if (lock_owner == getpid()) {
160                 unsigned int i;
161                 struct flock fl;
162                 fl.l_type = F_UNLCK;
163                 fl.l_whence = SEEK_SET;
164                 fl.l_start = 0;
165                 fl.l_len = 0;
166
167                 for (i = 0; i < lock_num; i++)
168                         fcntl(locks[i].fd, F_SETLK, &fl);
169         } else {
170                 /* Our parent must have them; pass request up. */
171                 enum info_type type = RELEASE_LOCKS;
172                 assert(control_fd != -1);
173                 write_all(control_fd, &type, sizeof(type));
174         }
175         lock_owner = 0;
176 }
177
178 /* off_t is a signed type.  Getting its max is non-trivial. */
179 static off_t off_max(void)
180 {
181         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
182         if (sizeof(off_t) == 4)
183                 return (off_t)0x7FFFFFF;
184         else
185                 return (off_t)0x7FFFFFFFFFFFFFFULL;
186 }
187
188 static void get_locks(void)
189 {
190         unsigned int i;
191         struct flock fl;
192
193         if (lock_owner == getpid())
194                 return;
195
196         if (lock_owner != 0) {
197                 enum info_type type = RELEASE_LOCKS;
198                 assert(control_fd != -1);
199                 write_all(control_fd, &type, sizeof(type));
200         }
201
202         fl.l_whence = SEEK_SET;
203
204         for (i = 0; i < lock_num; i++) {
205                 fl.l_type = locks[i].type;
206                 fl.l_start = locks[i].start;
207                 if (locks[i].end == off_max())
208                         fl.l_len = 0;
209                 else
210                         fl.l_len = locks[i].end - locks[i].start + 1;
211
212                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
213                         abort();
214         }
215         lock_owner = getpid();
216 }
217
218 static void trace_str(const char *str)
219 {
220         ssize_t ret;
221
222         while ((ret = write(tracefd, str, strlen(str))) <= 0) {
223                 str += ret;
224                 if (!*str)
225                         return;
226         }
227         err(1, "Writing trace.");
228 }
229
230 struct saved_file {
231         struct saved_file *next;
232         int fd;
233         void *contents;
234         off_t off, len;
235 };
236
237 static struct saved_file *save_file(struct saved_file *next, int fd)
238 {
239         struct saved_file *s = malloc(sizeof(*s));
240
241         s->next = next;
242         s->fd = fd;
243         s->off = lseek(fd, 0, SEEK_CUR);
244         /* Special file?  Erk... */
245         assert(s->off != -1);
246         s->len = lseek(fd, 0, SEEK_END);
247         lseek(fd, 0, SEEK_SET);
248         s->contents = malloc(s->len);
249         read(fd, s->contents, s->len);
250         lseek(fd, s->off, SEEK_SET);
251         return s;
252 }
253         
254 /* We have little choice but to save and restore open files: mmap means we
255  * can really intercept changes in the child.
256  *
257  * We could do non-mmap'ed files on demand, however. */
258 static struct saved_file *save_files(void)
259 {
260         struct saved_file *files = NULL;
261         int i;
262
263         /* Figure out the set of live fds. */
264         for (i = history_num - 2; i >= 0; i--) {
265                 if (history[i].type == FAILTEST_OPEN) {
266                         int fd = history[i].u.open.ret;
267                         /* Only do successful, writable fds. */
268                         if (fd < 0)
269                                 continue;
270
271                         /* If it was closed, cleanup == NULL. */
272                         if (!history[i].cleanup)
273                                 continue;
274
275                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
276                                 files = save_file(files, fd);
277                         } else if ((history[i].u.open.flags & O_WRONLY)
278                                    == O_WRONLY) {
279                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
280                                 abort();
281                         }
282                 }
283         }
284
285         return files;
286 }
287
288 static void restore_files(struct saved_file *s)
289 {
290         while (s) {
291                 struct saved_file *next = s->next;
292
293                 lseek(s->fd, 0, SEEK_SET);
294                 write(s->fd, s->contents, s->len);
295                 ftruncate(s->fd, s->len);
296                 free(s->contents);
297                 lseek(s->fd, s->off, SEEK_SET);
298                 free(s);
299                 s = next;
300         }
301 }
302
303 /* Free up memory, so valgrind doesn't report leaks. */
304 static void free_everything(void)
305 {
306         unsigned int i;
307
308         /* We don't do this in cleanup: needed even for failed opens. */
309         for (i = 0; i < history_num; i++) {
310                 if (history[i].type == FAILTEST_OPEN)
311                         free((char *)history[i].u.open.pathname);
312         }
313         free(history);
314 }
315
316 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
317 {
318         int i;
319
320         /* For children, we don't care if they "failed" the testing. */
321         if (control_fd != -1)
322                 status = 0;
323
324         if (forced_cleanup)
325                 history_num--;
326
327         /* Cleanup everything, in reverse order. */
328         for (i = history_num - 1; i >= 0; i--) {
329                 if (!history[i].cleanup)
330                         continue;
331                 if (!forced_cleanup) {
332                         printf("Leak at %s:%u\n",
333                                history[i].file, history[i].line);
334                         status = 1;
335                 }
336                 history[i].cleanup(&history[i].u);
337         }
338
339         free_everything();
340         tell_parent(SUCCESS);
341         exit(status);
342 }
343
344 static bool should_fail(struct failtest_call *call)
345 {
346         int status;
347         int control[2], output[2];
348         enum info_type type = UNEXPECTED;
349         char *out = NULL;
350         size_t outlen = 0;
351         struct saved_file *files;
352
353         /* Are we probing? */
354         if (probe_count && --probe_count == 0)
355                 failtest_cleanup(true, 0);
356
357         if (call == &unrecorded_call)
358                 return false;
359
360         if (failpath) {
361                 /* + means continue after end, like normal. */
362                 if (*failpath == '+')
363                         failpath = NULL;
364                 else {
365                         if (tolower(*failpath) != info_to_arg[call->type])
366                                 errx(1, "Failpath expected '%c' got '%c'\n",
367                                      info_to_arg[call->type], *failpath);
368                         call->fail = isupper(*(failpath++));
369                         return call->fail;
370                 }
371         }
372
373         /* Attach debugger if they asked for it. */
374         if (debugpath && history_num == strlen(debugpath)) {
375                 unsigned int i;
376
377                 for (i = 0; i < history_num; i++) {
378                         char c = info_to_arg[history[i].type];
379                         if (history[i].fail)
380                                 c = toupper(c);
381                         if (c != debugpath[i])
382                                 break;
383                 }
384                 if (i == history_num) {
385                         char str[80];
386
387                         /* Don't timeout. */
388                         signal(SIGUSR1, SIG_IGN);
389                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
390                                 getpid(), getpid());
391                         system(str);
392                         sleep(5);
393                 }
394         }
395
396         if (failtest_hook) {
397                 switch (failtest_hook(history, history_num)) {
398                 case FAIL_OK:
399                         break;
400                 case FAIL_DONT_FAIL:
401                         call->fail = false;
402                         return false;
403                 case FAIL_PROBE:
404                         /* Already down probe path?  Stop now. */
405                         if (probe_count)
406                                 failtest_cleanup(true, 0);
407                         /* FIXME: We should run *parent* and run probe until
408                          * calls match up again. */
409                         probe_count = 3;
410                         break;
411                 default:
412                         abort();
413                 }
414         }
415
416         files = save_files();
417
418         /* We're going to fail in the child. */
419         call->fail = true;
420         if (pipe(control) != 0 || pipe(output) != 0)
421                 err(1, "opening pipe");
422
423         /* Prevent double-printing (in child and parent) */
424         fflush(stdout);
425         child = fork();
426         if (child == -1)
427                 err(1, "forking failed");
428
429         if (child == 0) {
430                 if (tracefd != -1) {
431                         struct timeval now;
432                         char str[50], *p;
433                         gettimeofday(&now, NULL);
434                         if (now.tv_usec < start.tv_usec) {
435                                 now.tv_sec--;
436                                 now.tv_usec += 1000000;
437                         }
438                         now.tv_usec -= start.tv_usec;
439                         now.tv_sec -= start.tv_sec;
440                         sprintf(str, "%u (%u.%02u): ", getpid(),
441                                 (int)now.tv_sec, (int)now.tv_usec / 10000);
442                         trace_str(str);
443                         p = failpath_string();
444                         trace_str(p);
445                         free(p);
446                         trace_str("(");
447                         p = strchr(history[history_num-1].file, '/');
448                         if (p)
449                                 trace_str(p+1);
450                         else
451                                 trace_str(history[history_num-1].file);
452                         sprintf(str, ":%u)\n", history[history_num-1].line);
453                         trace_str(str);
454                 }
455                 close(control[0]);
456                 close(output[0]);
457                 dup2(output[1], STDOUT_FILENO);
458                 dup2(output[1], STDERR_FILENO);
459                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
460                         close(output[1]);
461                 control_fd = control[1];
462                 return true;
463         }
464
465         signal(SIGUSR1, hand_down);
466
467         close(control[1]);
468         close(output[1]);
469
470         /* We grab output so we can display it; we grab writes so we
471          * can compare. */
472         do {
473                 struct pollfd pfd[2];
474                 int ret;
475
476                 pfd[0].fd = output[0];
477                 pfd[0].events = POLLIN|POLLHUP;
478                 pfd[1].fd = control[0];
479                 pfd[1].events = POLLIN|POLLHUP;
480
481                 if (type == SUCCESS)
482                         ret = poll(pfd, 1, failtest_timeout_ms);
483                 else
484                         ret = poll(pfd, 2, failtest_timeout_ms);
485
486                 if (ret <= 0)
487                         hand_down(SIGUSR1);
488
489                 if (pfd[0].revents & POLLIN) {
490                         ssize_t len;
491
492                         out = realloc(out, outlen + 8192);
493                         len = read(output[0], out + outlen, 8192);
494                         outlen += len;
495                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
496                         if (read_all(control[0], &type, sizeof(type))) {
497                                 if (type == WRITE) {
498                                         if (!read_write_info(control[0]))
499                                                 break;
500                                 } else if (type == RELEASE_LOCKS) {
501                                         release_locks();
502                                         /* FIXME: Tell them we're done... */
503                                 }
504                         }
505                 } else if (pfd[0].revents & POLLHUP) {
506                         break;
507                 }
508         } while (type != FAILURE);
509
510         close(output[0]);
511         close(control[0]);
512         waitpid(child, &status, 0);
513         if (!WIFEXITED(status)) {
514                 if (WTERMSIG(status) == SIGUSR1)
515                         child_fail(out, outlen, "Timed out");
516                 else
517                         child_fail(out, outlen, "Killed by signal %u: ",
518                                    WTERMSIG(status));
519         }
520         /* Child printed failure already, just pass up exit code. */
521         if (type == FAILURE) {
522                 fprintf(stderr, "%.*s", (int)outlen, out);
523                 tell_parent(type);
524                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
525         }
526         if (WEXITSTATUS(status) != 0)
527                 child_fail(out, outlen, "Exited with status %i: ",
528                            WEXITSTATUS(status));
529
530         free(out);
531         signal(SIGUSR1, SIG_DFL);
532
533         restore_files(files);
534
535         /* We continue onwards without failing. */
536         call->fail = false;
537         return false;
538 }
539
540 static void cleanup_calloc(struct calloc_call *call)
541 {
542         free(call->ret);
543 }
544
545 void *failtest_calloc(size_t nmemb, size_t size,
546                       const char *file, unsigned line)
547 {
548         struct failtest_call *p;
549         struct calloc_call call;
550         call.nmemb = nmemb;
551         call.size = size;
552         p = add_history(FAILTEST_CALLOC, file, line, &call);
553
554         if (should_fail(p)) {
555                 p->u.calloc.ret = NULL;
556                 p->error = ENOMEM;
557         } else {
558                 p->u.calloc.ret = calloc(nmemb, size);
559                 set_cleanup(p, cleanup_calloc, struct calloc_call);
560         }
561         errno = p->error;
562         return p->u.calloc.ret;
563 }
564
565 static void cleanup_malloc(struct malloc_call *call)
566 {
567         free(call->ret);
568 }
569
570 void *failtest_malloc(size_t size, const char *file, unsigned line)
571 {
572         struct failtest_call *p;
573         struct malloc_call call;
574         call.size = size;
575
576         p = add_history(FAILTEST_MALLOC, file, line, &call);
577         if (should_fail(p)) {
578                 p->u.calloc.ret = NULL;
579                 p->error = ENOMEM;
580         } else {
581                 p->u.calloc.ret = malloc(size);
582                 set_cleanup(p, cleanup_malloc, struct malloc_call);
583         }
584         errno = p->error;
585         return p->u.calloc.ret;
586 }
587
588 static void cleanup_realloc(struct realloc_call *call)
589 {
590         free(call->ret);
591 }
592
593 /* Walk back and find out if we got this ptr from a previous routine. */
594 static void fixup_ptr_history(void *ptr, unsigned int last)
595 {
596         int i;
597
598         /* Start at end of history, work back. */
599         for (i = last - 1; i >= 0; i--) {
600                 switch (history[i].type) {
601                 case FAILTEST_REALLOC:
602                         if (history[i].u.realloc.ret == ptr) {
603                                 history[i].cleanup = NULL;
604                                 return;
605                         }
606                         break;
607                 case FAILTEST_MALLOC:
608                         if (history[i].u.malloc.ret == ptr) {
609                                 history[i].cleanup = NULL;
610                                 return;
611                         }
612                         break;
613                 case FAILTEST_CALLOC:
614                         if (history[i].u.calloc.ret == ptr) {
615                                 history[i].cleanup = NULL;
616                                 return;
617                         }
618                         break;
619                 default:
620                         break;
621                 }
622         }
623 }
624
625 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
626 {
627         struct failtest_call *p;
628         struct realloc_call call;
629         call.size = size;
630         p = add_history(FAILTEST_REALLOC, file, line, &call);
631
632         /* FIXME: Try one child moving allocation, one not. */
633         if (should_fail(p)) {
634                 p->u.realloc.ret = NULL;
635                 p->error = ENOMEM;
636         } else {
637                 fixup_ptr_history(ptr, history_num-1);
638                 p->u.realloc.ret = realloc(ptr, size);
639                 set_cleanup(p, cleanup_realloc, struct realloc_call);
640         }
641         errno = p->error;
642         return p->u.realloc.ret;
643 }
644
645 void failtest_free(void *ptr)
646 {
647         fixup_ptr_history(ptr, history_num);
648         free(ptr);
649 }
650
651 static void cleanup_open(struct open_call *call)
652 {
653         close(call->ret);
654 }
655
656 int failtest_open(const char *pathname,
657                   const char *file, unsigned line, ...)
658 {
659         struct failtest_call *p;
660         struct open_call call;
661         va_list ap;
662
663         call.pathname = strdup(pathname);
664         va_start(ap, line);
665         call.flags = va_arg(ap, int);
666         if (call.flags & O_CREAT) {
667                 call.mode = va_arg(ap, mode_t);
668                 va_end(ap);
669         }
670         p = add_history(FAILTEST_OPEN, file, line, &call);
671         /* Avoid memory leak! */
672         if (p == &unrecorded_call)
673                 free((char *)call.pathname);
674         p->u.open.ret = open(pathname, call.flags, call.mode);
675
676         if (!failpath && p->u.open.ret == -1) {
677                 p->fail = false;
678                 p->error = errno;
679         } else if (should_fail(p)) {
680                 close(p->u.open.ret);
681                 p->u.open.ret = -1;
682                 /* FIXME: Play with error codes? */
683                 p->error = EACCES;
684         } else {
685                 set_cleanup(p, cleanup_open, struct open_call);
686         }
687         errno = p->error;
688         return p->u.open.ret;
689 }
690
691 static void cleanup_pipe(struct pipe_call *call)
692 {
693         if (!call->closed[0])
694                 close(call->fds[0]);
695         if (!call->closed[1])
696                 close(call->fds[1]);
697 }
698
699 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
700 {
701         struct failtest_call *p;
702         struct pipe_call call;
703
704         p = add_history(FAILTEST_PIPE, file, line, &call);
705         if (should_fail(p)) {
706                 p->u.open.ret = -1;
707                 /* FIXME: Play with error codes? */
708                 p->error = EMFILE;
709         } else {
710                 p->u.pipe.ret = pipe(p->u.pipe.fds);
711                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
712                 set_cleanup(p, cleanup_pipe, struct pipe_call);
713         }
714         /* This causes valgrind to notice if they use pipefd[] after failure */
715         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
716         errno = p->error;
717         return p->u.pipe.ret;
718 }
719
720 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
721                        const char *file, unsigned line)
722 {
723         struct failtest_call *p;
724         struct read_call call;
725         call.fd = fd;
726         call.buf = buf;
727         call.count = count;
728         call.off = off;
729         p = add_history(FAILTEST_READ, file, line, &call);
730
731         /* FIXME: Try partial read returns. */
732         if (should_fail(p)) {
733                 p->u.read.ret = -1;
734                 p->error = EIO;
735         } else {
736                 p->u.read.ret = pread(fd, buf, count, off);
737         }
738         errno = p->error;
739         return p->u.read.ret;
740 }
741
742 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
743                         const char *file, unsigned line)
744 {
745         struct failtest_call *p;
746         struct write_call call;
747
748         call.fd = fd;
749         call.buf = buf;
750         call.count = count;
751         call.off = off;
752         p = add_history(FAILTEST_WRITE, file, line, &call);
753
754         /* If we're a child, we need to make sure we write the same thing
755          * to non-files as the parent does, so tell it. */
756         if (control_fd != -1 && off == (off_t)-1) {
757                 enum info_type type = WRITE;
758
759                 write_all(control_fd, &type, sizeof(type));
760                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
761                 write_all(control_fd, buf, count);
762         }
763
764         /* FIXME: Try partial write returns. */
765         if (should_fail(p)) {
766                 p->u.write.ret = -1;
767                 p->error = EIO;
768         } else {
769                 /* FIXME: We assume same write order in parent and child */
770                 if (off == (off_t)-1 && child_writes_num != 0) {
771                         if (child_writes[0].fd != fd)
772                                 errx(1, "Child wrote to fd %u, not %u?",
773                                      child_writes[0].fd, fd);
774                         if (child_writes[0].off != p->u.write.off)
775                                 errx(1, "Child wrote to offset %zu, not %zu?",
776                                      (size_t)child_writes[0].off,
777                                      (size_t)p->u.write.off);
778                         if (child_writes[0].count != count)
779                                 errx(1, "Child wrote length %zu, not %zu?",
780                                      child_writes[0].count, count);
781                         if (memcmp(child_writes[0].buf, buf, count)) {
782                                 child_fail(NULL, 0,
783                                            "Child wrote differently to"
784                                            " fd %u than we did!\n", fd);
785                         }
786                         free((char *)child_writes[0].buf);
787                         child_writes_num--;
788                         memmove(&child_writes[0], &child_writes[1],
789                                 sizeof(child_writes[0]) * child_writes_num);
790
791                         /* Is this is a socket or pipe, child wrote it
792                            already. */
793                         if (p->u.write.off == (off_t)-1) {
794                                 p->u.write.ret = count;
795                                 errno = p->error;
796                                 return p->u.write.ret;
797                         }
798                 }
799                 p->u.write.ret = pwrite(fd, buf, count, off);
800         }
801         errno = p->error;
802         return p->u.write.ret;
803 }
804
805 ssize_t failtest_read(int fd, void *buf, size_t count,
806                       const char *file, unsigned line)
807 {
808         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
809                               file, line);
810 }
811
812 ssize_t failtest_write(int fd, const void *buf, size_t count,
813                        const char *file, unsigned line)
814 {
815         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
816                                file, line);
817 }
818
819 static struct lock_info *WARN_UNUSED_RESULT
820 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
821 {
822         unsigned int i;
823         struct lock_info *l;
824
825         for (i = 0; i < lock_num; i++) {
826                 l = &locks[i];
827
828                 if (l->fd != fd)
829                         continue;
830                 /* Four cases we care about:
831                  * Start overlap:
832                  *      l =    |      |
833                  *      new = |   |
834                  * Mid overlap:
835                  *      l =    |      |
836                  *      new =    |  |
837                  * End overlap:
838                  *      l =    |      |
839                  *      new =      |    |
840                  * Total overlap:
841                  *      l =    |      |
842                  *      new = |         |
843                  */
844                 if (start > l->start && end < l->end) {
845                         /* Mid overlap: trim entry, add new one. */
846                         off_t new_start, new_end;
847                         new_start = end + 1;
848                         new_end = l->end;
849                         l->end = start - 1;
850                         locks = add_lock(locks,
851                                          fd, new_start, new_end, l->type);
852                         l = &locks[i];
853                 } else if (start <= l->start && end >= l->end) {
854                         /* Total overlap: eliminate entry. */
855                         l->end = 0;
856                         l->start = 1;
857                 } else if (end >= l->start && end < l->end) {
858                         /* Start overlap: trim entry. */
859                         l->start = end + 1;
860                 } else if (start > l->start && start <= l->end) {
861                         /* End overlap: trim entry. */
862                         l->end = start-1;
863                 }
864                 /* Nothing left?  Remove it. */
865                 if (l->end < l->start) {
866                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
867                         i--;
868                 }
869         }
870
871         if (type != F_UNLCK) {
872                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
873                 l = &locks[lock_num++];
874                 l->fd = fd;
875                 l->start = start;
876                 l->end = end;
877                 l->type = type;
878         }
879         return locks;
880 }
881
882 /* We trap this so we can record it: we don't fail it. */
883 int failtest_close(int fd, const char *file, unsigned line)
884 {
885         int i;
886         struct close_call call;
887         struct failtest_call *p;
888
889         call.fd = fd;
890         p = add_history(FAILTEST_CLOSE, file, line, &call);
891         p->fail = false;
892
893         /* Consume close from failpath. */
894         if (failpath)
895                 if (should_fail(p))
896                         abort();
897
898         if (fd < 0)
899                 return close(fd);
900
901         /* Trace history to find source of fd. */
902         for (i = history_num-1; i >= 0; i--) {
903                 switch (history[i].type) {
904                 case FAILTEST_PIPE:
905                         /* From a pipe? */
906                         if (history[i].u.pipe.fds[0] == fd) {
907                                 assert(!history[i].u.pipe.closed[0]);
908                                 history[i].u.pipe.closed[0] = true;
909                                 if (history[i].u.pipe.closed[1])
910                                         history[i].cleanup = NULL;
911                                 goto out;
912                         }
913                         if (history[i].u.pipe.fds[1] == fd) {
914                                 assert(!history[i].u.pipe.closed[1]);
915                                 history[i].u.pipe.closed[1] = true;
916                                 if (history[i].u.pipe.closed[0])
917                                         history[i].cleanup = NULL;
918                                 goto out;
919                         }
920                         break;
921                 case FAILTEST_OPEN:
922                         if (history[i].u.open.ret == fd) {
923                                 assert((void *)history[i].cleanup
924                                        == (void *)cleanup_open);
925                                 history[i].cleanup = NULL;
926                                 goto out;
927                         }
928                         break;
929                 default:
930                         break;
931                 }
932         }
933
934 out:
935         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
936         return close(fd);
937 }
938
939 /* Zero length means "to end of file" */
940 static off_t end_of(off_t start, off_t len)
941 {
942         if (len == 0)
943                 return off_max();
944         return start + len - 1;
945 }
946
947 /* FIXME: This only handles locks, really. */
948 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
949 {
950         struct failtest_call *p;
951         struct fcntl_call call;
952         va_list ap;
953
954         call.fd = fd;
955         call.cmd = cmd;
956
957         /* Argument extraction. */
958         switch (cmd) {
959         case F_SETFL:
960         case F_SETFD:
961                 va_start(ap, cmd);
962                 call.arg.l = va_arg(ap, long);
963                 va_end(ap);
964                 return fcntl(fd, cmd, call.arg.l);
965         case F_GETFD:
966         case F_GETFL:
967                 return fcntl(fd, cmd);
968         case F_GETLK:
969                 get_locks();
970                 va_start(ap, cmd);
971                 call.arg.fl = *va_arg(ap, struct flock *);
972                 va_end(ap);
973                 return fcntl(fd, cmd, &call.arg.fl);
974         case F_SETLK:
975         case F_SETLKW:
976                 va_start(ap, cmd);
977                 call.arg.fl = *va_arg(ap, struct flock *);
978                 va_end(ap);
979                 break;
980         default:
981                 /* This means you need to implement it here. */
982                 err(1, "failtest: unknown fcntl %u", cmd);
983         }
984
985         p = add_history(FAILTEST_FCNTL, file, line, &call);
986         get_locks();
987
988         if (should_fail(p)) {
989                 p->u.fcntl.ret = -1;
990                 if (p->u.fcntl.cmd == F_SETLK)
991                         p->error = EAGAIN;
992                 else
993                         p->error = EDEADLK;
994         } else {
995                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
996                                        &p->u.fcntl.arg.fl);
997                 if (p->u.fcntl.ret == -1)
998                         p->error = errno;
999                 else {
1000                         /* We don't handle anything else yet. */
1001                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1002                         locks = add_lock(locks,
1003                                          p->u.fcntl.fd,
1004                                          p->u.fcntl.arg.fl.l_start,
1005                                          end_of(p->u.fcntl.arg.fl.l_start,
1006                                                 p->u.fcntl.arg.fl.l_len),
1007                                          p->u.fcntl.arg.fl.l_type);
1008                 }
1009         }
1010         errno = p->error;
1011         return p->u.fcntl.ret;
1012 }
1013
1014 void failtest_init(int argc, char *argv[])
1015 {
1016         unsigned int i;
1017
1018         for (i = 1; i < argc; i++) {
1019                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1020                         failpath = argv[i] + strlen("--failpath=");
1021                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1022                         tracefd = dup(STDERR_FILENO);
1023                         failtest_timeout_ms = -1;
1024                 } else if (!strncmp(argv[i], "--debugpath=",
1025                                     strlen("--debugpath="))) {
1026                         debugpath = argv[i] + strlen("--debugpath=");
1027                 }
1028         }
1029         gettimeofday(&start, NULL);
1030 }
1031
1032 void failtest_exit(int status)
1033 {
1034         if (failtest_exit_check) {
1035                 if (!failtest_exit_check(history, history_num))
1036                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1037         }
1038
1039         failtest_cleanup(false, status);
1040 }