various: add LICENSE comments.
[ccan] / ccan / failtest / failtest.c
1 /* Licensed under LGPL - see LICENSE file for details */
2 #include "config.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/read_write_all/read_write_all.h>
19 #include <ccan/failtest/failtest_proto.h>
20 #include <ccan/failtest/failtest.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 /* Free up memory, so valgrind doesn't report leaks. */
314 static void free_everything(void)
315 {
316         unsigned int i;
317
318         /* We don't do this in cleanup: needed even for failed opens. */
319         for (i = 0; i < history_num; i++) {
320                 if (history[i].type == FAILTEST_OPEN)
321                         free((char *)history[i].u.open.pathname);
322         }
323         free(history);
324 }
325
326 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
327 {
328         int i;
329
330         /* For children, we don't care if they "failed" the testing. */
331         if (control_fd != -1)
332                 status = 0;
333
334         if (forced_cleanup)
335                 history_num--;
336
337         /* Cleanup everything, in reverse order. */
338         for (i = history_num - 1; i >= 0; i--) {
339                 if (!history[i].cleanup)
340                         continue;
341                 if (!forced_cleanup) {
342                         printf("Leak at %s:%u: --failpath=%s\n",
343                                history[i].file, history[i].line,
344                                failpath_string());
345                         status = 1;
346                 }
347                 history[i].cleanup(&history[i].u);
348         }
349
350         free_everything();
351         tell_parent(SUCCESS);
352         exit(status);
353 }
354
355 static bool should_fail(struct failtest_call *call)
356 {
357         int status;
358         int control[2], output[2];
359         enum info_type type = UNEXPECTED;
360         char *out = NULL;
361         size_t outlen = 0;
362         struct saved_file *files;
363
364         /* Are we probing? */
365         if (probe_count && --probe_count == 0)
366                 failtest_cleanup(true, 0);
367
368         if (call == &unrecorded_call)
369                 return false;
370
371         if (failpath) {
372                 /* + means continue after end, like normal. */
373                 if (*failpath == '+')
374                         failpath = NULL;
375                 else if (*failpath == '\0') {
376                         /* Continue, but don't inject errors. */
377                         return call->fail = false;
378                 } else {
379                         if (tolower((unsigned char)*failpath)
380                             != info_to_arg[call->type])
381                                 errx(1, "Failpath expected '%c' got '%c'\n",
382                                      info_to_arg[call->type], *failpath);
383                         call->fail = isupper((unsigned char)*(failpath++));
384                         return call->fail;
385                 }
386         }
387
388         /* Attach debugger if they asked for it. */
389         if (debugpath && history_num == strlen(debugpath)) {
390                 unsigned int i;
391
392                 for (i = 0; i < history_num; i++) {
393                         unsigned char c = info_to_arg[history[i].type];
394                         if (history[i].fail)
395                                 c = toupper(c);
396                         if (c != debugpath[i])
397                                 break;
398                 }
399                 if (i == history_num) {
400                         char str[80];
401
402                         /* Don't timeout. */
403                         signal(SIGUSR1, SIG_IGN);
404                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
405                                 getpid(), getpid());
406                         if (system(str) == 0)
407                                 sleep(5);
408                 }
409         }
410
411         if (failtest_hook) {
412                 switch (failtest_hook(history, history_num)) {
413                 case FAIL_OK:
414                         break;
415                 case FAIL_DONT_FAIL:
416                         call->fail = false;
417                         return false;
418                 case FAIL_PROBE:
419                         /* Already down probe path?  Stop now. */
420                         if (probe_count)
421                                 failtest_cleanup(true, 0);
422                         /* FIXME: We should run *parent* and run probe until
423                          * calls match up again. */
424                         probe_count = 3;
425                         break;
426                 default:
427                         abort();
428                 }
429         }
430
431         files = save_files();
432
433         /* We're going to fail in the child. */
434         call->fail = true;
435         if (pipe(control) != 0 || pipe(output) != 0)
436                 err(1, "opening pipe");
437
438         /* Prevent double-printing (in child and parent) */
439         fflush(stdout);
440         child = fork();
441         if (child == -1)
442                 err(1, "forking failed");
443
444         if (child == 0) {
445                 if (tracefd != -1) {
446                         struct timeval now;
447                         const char *p;
448                         gettimeofday(&now, NULL);
449                         if (now.tv_usec < start.tv_usec) {
450                                 now.tv_sec--;
451                                 now.tv_usec += 1000000;
452                         }
453                         now.tv_usec -= start.tv_usec;
454                         now.tv_sec -= start.tv_sec;
455                         p = failpath_string();
456                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
457                               (int)now.tv_sec, (int)now.tv_usec / 10000, p);
458                         free((char *)p);
459                         p = strrchr(history[history_num-1].file, '/');
460                         if (p)
461                                 trace("%s", p+1);
462                         else
463                                 trace("%s", history[history_num-1].file);
464                         trace(":%u)\n", history[history_num-1].line);
465                 }
466                 close(control[0]);
467                 close(output[0]);
468                 dup2(output[1], STDOUT_FILENO);
469                 dup2(output[1], STDERR_FILENO);
470                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
471                         close(output[1]);
472                 control_fd = control[1];
473                 return true;
474         }
475
476         signal(SIGUSR1, hand_down);
477
478         close(control[1]);
479         close(output[1]);
480
481         /* We grab output so we can display it; we grab writes so we
482          * can compare. */
483         do {
484                 struct pollfd pfd[2];
485                 int ret;
486
487                 pfd[0].fd = output[0];
488                 pfd[0].events = POLLIN|POLLHUP;
489                 pfd[1].fd = control[0];
490                 pfd[1].events = POLLIN|POLLHUP;
491
492                 if (type == SUCCESS)
493                         ret = poll(pfd, 1, failtest_timeout_ms);
494                 else
495                         ret = poll(pfd, 2, failtest_timeout_ms);
496
497                 if (ret == 0)
498                         hand_down(SIGUSR1);
499                 if (ret < 0) {
500                         if (errno == EINTR)
501                                 continue;
502                         err(1, "Poll returned %i", ret);
503                 }
504
505                 if (pfd[0].revents & POLLIN) {
506                         ssize_t len;
507
508                         out = realloc(out, outlen + 8192);
509                         len = read(output[0], out + outlen, 8192);
510                         outlen += len;
511                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
512                         if (read_all(control[0], &type, sizeof(type))) {
513                                 if (type == WRITE) {
514                                         if (!read_write_info(control[0]))
515                                                 break;
516                                 } else if (type == RELEASE_LOCKS) {
517                                         release_locks();
518                                         /* FIXME: Tell them we're done... */
519                                 }
520                         }
521                 } else if (pfd[0].revents & POLLHUP) {
522                         break;
523                 }
524         } while (type != FAILURE);
525
526         close(output[0]);
527         close(control[0]);
528         waitpid(child, &status, 0);
529         if (!WIFEXITED(status)) {
530                 if (WTERMSIG(status) == SIGUSR1)
531                         child_fail(out, outlen, "Timed out");
532                 else
533                         child_fail(out, outlen, "Killed by signal %u: ",
534                                    WTERMSIG(status));
535         }
536         /* Child printed failure already, just pass up exit code. */
537         if (type == FAILURE) {
538                 fprintf(stderr, "%.*s", (int)outlen, out);
539                 tell_parent(type);
540                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
541         }
542         if (WEXITSTATUS(status) != 0)
543                 child_fail(out, outlen, "Exited with status %i: ",
544                            WEXITSTATUS(status));
545
546         free(out);
547         signal(SIGUSR1, SIG_DFL);
548
549         restore_files(files);
550
551         /* We continue onwards without failing. */
552         call->fail = false;
553         return false;
554 }
555
556 static void cleanup_calloc(struct calloc_call *call)
557 {
558         free(call->ret);
559 }
560
561 void *failtest_calloc(size_t nmemb, size_t size,
562                       const char *file, unsigned line)
563 {
564         struct failtest_call *p;
565         struct calloc_call call;
566         call.nmemb = nmemb;
567         call.size = size;
568         p = add_history(FAILTEST_CALLOC, file, line, &call);
569
570         if (should_fail(p)) {
571                 p->u.calloc.ret = NULL;
572                 p->error = ENOMEM;
573         } else {
574                 p->u.calloc.ret = calloc(nmemb, size);
575                 set_cleanup(p, cleanup_calloc, struct calloc_call);
576         }
577         errno = p->error;
578         return p->u.calloc.ret;
579 }
580
581 static void cleanup_malloc(struct malloc_call *call)
582 {
583         free(call->ret);
584 }
585
586 void *failtest_malloc(size_t size, const char *file, unsigned line)
587 {
588         struct failtest_call *p;
589         struct malloc_call call;
590         call.size = size;
591
592         p = add_history(FAILTEST_MALLOC, file, line, &call);
593         if (should_fail(p)) {
594                 p->u.calloc.ret = NULL;
595                 p->error = ENOMEM;
596         } else {
597                 p->u.calloc.ret = malloc(size);
598                 set_cleanup(p, cleanup_malloc, struct malloc_call);
599         }
600         errno = p->error;
601         return p->u.calloc.ret;
602 }
603
604 static void cleanup_realloc(struct realloc_call *call)
605 {
606         free(call->ret);
607 }
608
609 /* Walk back and find out if we got this ptr from a previous routine. */
610 static void fixup_ptr_history(void *ptr, unsigned int last)
611 {
612         int i;
613
614         /* Start at end of history, work back. */
615         for (i = last - 1; i >= 0; i--) {
616                 switch (history[i].type) {
617                 case FAILTEST_REALLOC:
618                         if (history[i].u.realloc.ret == ptr) {
619                                 history[i].cleanup = NULL;
620                                 return;
621                         }
622                         break;
623                 case FAILTEST_MALLOC:
624                         if (history[i].u.malloc.ret == ptr) {
625                                 history[i].cleanup = NULL;
626                                 return;
627                         }
628                         break;
629                 case FAILTEST_CALLOC:
630                         if (history[i].u.calloc.ret == ptr) {
631                                 history[i].cleanup = NULL;
632                                 return;
633                         }
634                         break;
635                 default:
636                         break;
637                 }
638         }
639 }
640
641 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
642 {
643         struct failtest_call *p;
644         struct realloc_call call;
645         call.size = size;
646         p = add_history(FAILTEST_REALLOC, file, line, &call);
647
648         /* FIXME: Try one child moving allocation, one not. */
649         if (should_fail(p)) {
650                 p->u.realloc.ret = NULL;
651                 p->error = ENOMEM;
652         } else {
653                 fixup_ptr_history(ptr, history_num-1);
654                 p->u.realloc.ret = realloc(ptr, size);
655                 set_cleanup(p, cleanup_realloc, struct realloc_call);
656         }
657         errno = p->error;
658         return p->u.realloc.ret;
659 }
660
661 void failtest_free(void *ptr)
662 {
663         fixup_ptr_history(ptr, history_num);
664         free(ptr);
665 }
666
667 static void cleanup_open(struct open_call *call)
668 {
669         close(call->ret);
670 }
671
672 int failtest_open(const char *pathname,
673                   const char *file, unsigned line, ...)
674 {
675         struct failtest_call *p;
676         struct open_call call;
677         va_list ap;
678
679         call.pathname = strdup(pathname);
680         va_start(ap, line);
681         call.flags = va_arg(ap, int);
682         if (call.flags & O_CREAT) {
683                 call.mode = va_arg(ap, int);
684                 va_end(ap);
685         }
686         p = add_history(FAILTEST_OPEN, file, line, &call);
687         /* Avoid memory leak! */
688         if (p == &unrecorded_call)
689                 free((char *)call.pathname);
690         p->u.open.ret = open(pathname, call.flags, call.mode);
691
692         if (!failpath && p->u.open.ret == -1) {
693                 p->fail = false;
694                 p->error = errno;
695         } else if (should_fail(p)) {
696                 close(p->u.open.ret);
697                 p->u.open.ret = -1;
698                 /* FIXME: Play with error codes? */
699                 p->error = EACCES;
700         } else {
701                 set_cleanup(p, cleanup_open, struct open_call);
702         }
703         errno = p->error;
704         return p->u.open.ret;
705 }
706
707 static void cleanup_pipe(struct pipe_call *call)
708 {
709         if (!call->closed[0])
710                 close(call->fds[0]);
711         if (!call->closed[1])
712                 close(call->fds[1]);
713 }
714
715 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
716 {
717         struct failtest_call *p;
718         struct pipe_call call;
719
720         p = add_history(FAILTEST_PIPE, file, line, &call);
721         if (should_fail(p)) {
722                 p->u.open.ret = -1;
723                 /* FIXME: Play with error codes? */
724                 p->error = EMFILE;
725         } else {
726                 p->u.pipe.ret = pipe(p->u.pipe.fds);
727                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
728                 set_cleanup(p, cleanup_pipe, struct pipe_call);
729         }
730         /* This causes valgrind to notice if they use pipefd[] after failure */
731         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
732         errno = p->error;
733         return p->u.pipe.ret;
734 }
735
736 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
737                        const char *file, unsigned line)
738 {
739         struct failtest_call *p;
740         struct read_call call;
741         call.fd = fd;
742         call.buf = buf;
743         call.count = count;
744         call.off = off;
745         p = add_history(FAILTEST_READ, file, line, &call);
746
747         /* FIXME: Try partial read returns. */
748         if (should_fail(p)) {
749                 p->u.read.ret = -1;
750                 p->error = EIO;
751         } else {
752                 p->u.read.ret = pread(fd, buf, count, off);
753         }
754         errno = p->error;
755         return p->u.read.ret;
756 }
757
758 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
759                         const char *file, unsigned line)
760 {
761         struct failtest_call *p;
762         struct write_call call;
763
764         call.fd = fd;
765         call.buf = buf;
766         call.count = count;
767         call.off = off;
768         p = add_history(FAILTEST_WRITE, file, line, &call);
769
770         /* If we're a child, we need to make sure we write the same thing
771          * to non-files as the parent does, so tell it. */
772         if (control_fd != -1 && off == (off_t)-1) {
773                 enum info_type type = WRITE;
774
775                 write_all(control_fd, &type, sizeof(type));
776                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
777                 write_all(control_fd, buf, count);
778         }
779
780         /* FIXME: Try partial write returns. */
781         if (should_fail(p)) {
782                 p->u.write.ret = -1;
783                 p->error = EIO;
784         } else {
785                 /* FIXME: We assume same write order in parent and child */
786                 if (off == (off_t)-1 && child_writes_num != 0) {
787                         if (child_writes[0].fd != fd)
788                                 errx(1, "Child wrote to fd %u, not %u?",
789                                      child_writes[0].fd, fd);
790                         if (child_writes[0].off != p->u.write.off)
791                                 errx(1, "Child wrote to offset %zu, not %zu?",
792                                      (size_t)child_writes[0].off,
793                                      (size_t)p->u.write.off);
794                         if (child_writes[0].count != count)
795                                 errx(1, "Child wrote length %zu, not %zu?",
796                                      child_writes[0].count, count);
797                         if (memcmp(child_writes[0].buf, buf, count)) {
798                                 child_fail(NULL, 0,
799                                            "Child wrote differently to"
800                                            " fd %u than we did!\n", fd);
801                         }
802                         free((char *)child_writes[0].buf);
803                         child_writes_num--;
804                         memmove(&child_writes[0], &child_writes[1],
805                                 sizeof(child_writes[0]) * child_writes_num);
806
807                         /* Is this is a socket or pipe, child wrote it
808                            already. */
809                         if (p->u.write.off == (off_t)-1) {
810                                 p->u.write.ret = count;
811                                 errno = p->error;
812                                 return p->u.write.ret;
813                         }
814                 }
815                 p->u.write.ret = pwrite(fd, buf, count, off);
816         }
817         errno = p->error;
818         return p->u.write.ret;
819 }
820
821 ssize_t failtest_read(int fd, void *buf, size_t count,
822                       const char *file, unsigned line)
823 {
824         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
825                               file, line);
826 }
827
828 ssize_t failtest_write(int fd, const void *buf, size_t count,
829                        const char *file, unsigned line)
830 {
831         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
832                                file, line);
833 }
834
835 static struct lock_info *WARN_UNUSED_RESULT
836 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
837 {
838         unsigned int i;
839         struct lock_info *l;
840
841         for (i = 0; i < lock_num; i++) {
842                 l = &locks[i];
843
844                 if (l->fd != fd)
845                         continue;
846                 /* Four cases we care about:
847                  * Start overlap:
848                  *      l =    |      |
849                  *      new = |   |
850                  * Mid overlap:
851                  *      l =    |      |
852                  *      new =    |  |
853                  * End overlap:
854                  *      l =    |      |
855                  *      new =      |    |
856                  * Total overlap:
857                  *      l =    |      |
858                  *      new = |         |
859                  */
860                 if (start > l->start && end < l->end) {
861                         /* Mid overlap: trim entry, add new one. */
862                         off_t new_start, new_end;
863                         new_start = end + 1;
864                         new_end = l->end;
865                         l->end = start - 1;
866                         locks = add_lock(locks,
867                                          fd, new_start, new_end, l->type);
868                         l = &locks[i];
869                 } else if (start <= l->start && end >= l->end) {
870                         /* Total overlap: eliminate entry. */
871                         l->end = 0;
872                         l->start = 1;
873                 } else if (end >= l->start && end < l->end) {
874                         /* Start overlap: trim entry. */
875                         l->start = end + 1;
876                 } else if (start > l->start && start <= l->end) {
877                         /* End overlap: trim entry. */
878                         l->end = start-1;
879                 }
880                 /* Nothing left?  Remove it. */
881                 if (l->end < l->start) {
882                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
883                         i--;
884                 }
885         }
886
887         if (type != F_UNLCK) {
888                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
889                 l = &locks[lock_num++];
890                 l->fd = fd;
891                 l->start = start;
892                 l->end = end;
893                 l->type = type;
894         }
895         return locks;
896 }
897
898 /* We trap this so we can record it: we don't fail it. */
899 int failtest_close(int fd, const char *file, unsigned line)
900 {
901         int i;
902         struct close_call call;
903         struct failtest_call *p;
904
905         call.fd = fd;
906         p = add_history(FAILTEST_CLOSE, file, line, &call);
907         p->fail = false;
908
909         /* Consume close from failpath. */
910         if (failpath)
911                 if (should_fail(p))
912                         abort();
913
914         if (fd < 0)
915                 return close(fd);
916
917         /* Trace history to find source of fd. */
918         for (i = history_num-1; i >= 0; i--) {
919                 switch (history[i].type) {
920                 case FAILTEST_PIPE:
921                         /* From a pipe? */
922                         if (history[i].u.pipe.fds[0] == fd) {
923                                 assert(!history[i].u.pipe.closed[0]);
924                                 history[i].u.pipe.closed[0] = true;
925                                 if (history[i].u.pipe.closed[1])
926                                         history[i].cleanup = NULL;
927                                 goto out;
928                         }
929                         if (history[i].u.pipe.fds[1] == fd) {
930                                 assert(!history[i].u.pipe.closed[1]);
931                                 history[i].u.pipe.closed[1] = true;
932                                 if (history[i].u.pipe.closed[0])
933                                         history[i].cleanup = NULL;
934                                 goto out;
935                         }
936                         break;
937                 case FAILTEST_OPEN:
938                         if (history[i].u.open.ret == fd) {
939                                 assert((void *)history[i].cleanup
940                                        == (void *)cleanup_open);
941                                 history[i].cleanup = NULL;
942                                 goto out;
943                         }
944                         break;
945                 default:
946                         break;
947                 }
948         }
949
950 out:
951         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
952         return close(fd);
953 }
954
955 /* Zero length means "to end of file" */
956 static off_t end_of(off_t start, off_t len)
957 {
958         if (len == 0)
959                 return off_max();
960         return start + len - 1;
961 }
962
963 /* FIXME: This only handles locks, really. */
964 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
965 {
966         struct failtest_call *p;
967         struct fcntl_call call;
968         va_list ap;
969
970         call.fd = fd;
971         call.cmd = cmd;
972
973         /* Argument extraction. */
974         switch (cmd) {
975         case F_SETFL:
976         case F_SETFD:
977                 va_start(ap, cmd);
978                 call.arg.l = va_arg(ap, long);
979                 va_end(ap);
980                 return fcntl(fd, cmd, call.arg.l);
981         case F_GETFD:
982         case F_GETFL:
983                 return fcntl(fd, cmd);
984         case F_GETLK:
985                 get_locks();
986                 va_start(ap, cmd);
987                 call.arg.fl = *va_arg(ap, struct flock *);
988                 va_end(ap);
989                 return fcntl(fd, cmd, &call.arg.fl);
990         case F_SETLK:
991         case F_SETLKW:
992                 va_start(ap, cmd);
993                 call.arg.fl = *va_arg(ap, struct flock *);
994                 va_end(ap);
995                 break;
996         default:
997                 /* This means you need to implement it here. */
998                 err(1, "failtest: unknown fcntl %u", cmd);
999         }
1000
1001         p = add_history(FAILTEST_FCNTL, file, line, &call);
1002
1003         if (should_fail(p)) {
1004                 p->u.fcntl.ret = -1;
1005                 if (p->u.fcntl.cmd == F_SETLK)
1006                         p->error = EAGAIN;
1007                 else
1008                         p->error = EDEADLK;
1009         } else {
1010                 get_locks();
1011                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1012                                        &p->u.fcntl.arg.fl);
1013                 if (p->u.fcntl.ret == -1)
1014                         p->error = errno;
1015                 else {
1016                         /* We don't handle anything else yet. */
1017                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1018                         locks = add_lock(locks,
1019                                          p->u.fcntl.fd,
1020                                          p->u.fcntl.arg.fl.l_start,
1021                                          end_of(p->u.fcntl.arg.fl.l_start,
1022                                                 p->u.fcntl.arg.fl.l_len),
1023                                          p->u.fcntl.arg.fl.l_type);
1024                 }
1025         }
1026         errno = p->error;
1027         return p->u.fcntl.ret;
1028 }
1029
1030 pid_t failtest_getpid(const char *file, unsigned line)
1031 {
1032         /* You must call failtest_init first! */
1033         assert(orig_pid);
1034         return orig_pid;
1035 }
1036         
1037 void failtest_init(int argc, char *argv[])
1038 {
1039         unsigned int i;
1040
1041         orig_pid = getpid();
1042                 
1043         for (i = 1; i < argc; i++) {
1044                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1045                         failpath = argv[i] + strlen("--failpath=");
1046                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1047                         tracefd = dup(STDERR_FILENO);
1048                         failtest_timeout_ms = -1;
1049                 } else if (!strncmp(argv[i], "--debugpath=",
1050                                     strlen("--debugpath="))) {
1051                         debugpath = argv[i] + strlen("--debugpath=");
1052                 }
1053         }
1054         gettimeofday(&start, NULL);
1055 }
1056
1057 bool failtest_has_failed(void)
1058 {
1059         return control_fd != -1;
1060 }
1061
1062 void failtest_exit(int status)
1063 {
1064         if (failtest_exit_check) {
1065                 if (!failtest_exit_check(history, history_num))
1066                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1067         }
1068
1069         failtest_cleanup(false, status);
1070 }