10 #include <sys/types.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>
20 bool (*failtest_hook)(struct failtest_call *history, unsigned num)
21 = failtest_default_hook;
23 static int tracefd = -1;
25 unsigned int failtest_timeout_ms = 20000;
39 /* end is inclusive: you can't have a 0-byte lock. */
44 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
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;
51 static struct write_call *child_writes = NULL;
52 static unsigned int child_writes_num = 0;
54 static pid_t lock_owner;
55 static struct lock_info *locks = NULL;
56 static unsigned int lock_num = 0;
58 static const char info_to_arg[] = "mceoprwf";
60 /* Dummy call used for failtest_undo wrappers. */
61 static struct failtest_call unrecorded_call;
63 static struct failtest_call *add_history_(enum failtest_call_type type,
69 /* NULL file is how we suppress failure. */
71 return &unrecorded_call;
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++];
82 #define add_history(type, file, line, elem) \
83 add_history_((type), (file), (line), (elem), sizeof(*(elem)))
85 #define set_cleanup(call, clean, type) \
86 (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
88 bool failtest_default_hook(struct failtest_call *history, unsigned num)
93 static bool read_write_info(int fd)
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)))
105 w->buf = buf = malloc(w->count);
106 if (!read_all(fd, buf, w->count))
113 static char *failpath_string(void)
116 char *ret = malloc(history_num + 1);
118 for (i = 0; i < history_num; i++) {
119 ret[i] = info_to_arg[history[i].type];
121 ret[i] = toupper(ret[i]);
127 static void tell_parent(enum info_type type)
129 if (control_fd != -1)
130 write_all(control_fd, &type, sizeof(type));
133 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
136 char *path = failpath_string();
139 vfprintf(stderr, fmt, ap);
142 fprintf(stderr, "%.*s", (int)outlen, out);
143 printf("To reproduce: --failpath=%s\n", path);
145 tell_parent(FAILURE);
151 static void hand_down(int signal)
156 static void release_locks(void)
158 /* Locks were never acquired/reacquired? */
162 /* We own them? Release them all. */
163 if (lock_owner == getpid()) {
167 fl.l_whence = SEEK_SET;
171 for (i = 0; i < lock_num; i++)
172 fcntl(locks[i].fd, F_SETLK, &fl);
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));
182 /* off_t is a signed type. Getting its max is non-trivial. */
183 static off_t off_max(void)
185 BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
186 if (sizeof(off_t) == 4)
187 return (off_t)0x7FFFFFF;
189 return (off_t)0x7FFFFFFFFFFFFFFULL;
192 static void get_locks(void)
197 if (lock_owner == getpid())
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));
206 fl.l_whence = SEEK_SET;
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())
214 fl.l_len = locks[i].end - locks[i].start + 1;
216 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
219 lock_owner = getpid();
222 static void trace_str(const char *str)
226 while ((ret = write(tracefd, str, strlen(str))) <= 0) {
231 err(1, "Writing trace.");
235 struct saved_file *next;
241 static struct saved_file *save_file(struct saved_file *next, int fd)
243 struct saved_file *s = malloc(sizeof(*s));
244 off_t orig = lseek(fd, 0, SEEK_CUR);
246 /* Special file? Erk... */
251 s->len = lseek(fd, 0, SEEK_END);
252 lseek(fd, 0, SEEK_SET);
253 s->contents = malloc(s->len);
254 read(fd, s->contents, s->len);
255 lseek(fd, orig, SEEK_SET);
259 /* We have little choice but to save and restore open files: mmap means we
260 * can really intercept changes in the child.
262 * We could do non-mmap'ed files on demand, however. */
263 static struct saved_file *save_files(void)
265 struct saved_file *files = NULL;
269 /* Figure out the set of live fds. */
271 for (i = history_num - 2; i >= 0; i--) {
272 /* FIXME: Handle dup. */
273 if (history[i].type == FAILTEST_CLOSE) {
274 assert(!FD_ISSET(history[i].u.close.fd, &closed));
275 FD_SET(history[i].u.close.fd, &closed);
276 } else if (history[i].type == FAILTEST_OPEN) {
277 int fd = history[i].u.open.ret;
278 /* Only do successful, writable fds. */
282 /* If it wasn't closed again... */
283 if (!FD_ISSET(fd, &closed)) {
284 if ((history[i].u.open.flags & O_RDWR)
286 files = save_file(files, fd);
287 } else if ((history[i].u.open.flags & O_WRONLY)
289 /* FIXME: Handle O_WRONLY. Open with
294 FD_CLR(history[i].u.open.ret, &closed);
301 static void restore_files(struct saved_file *s)
304 struct saved_file *next = s->next;
305 off_t orig = lseek(s->fd, 0, SEEK_CUR);
307 lseek(s->fd, 0, SEEK_SET);
308 write(s->fd, s->contents, s->len);
310 lseek(s->fd, orig, SEEK_SET);
316 static bool should_fail(struct failtest_call *call)
319 int control[2], output[2];
320 enum info_type type = UNEXPECTED;
323 struct saved_file *files;
325 if (call == &unrecorded_call)
329 /* + means continue after end, like normal. */
330 if (*failpath == '+')
333 if (tolower(*failpath) != info_to_arg[call->type])
334 errx(1, "Failpath expected '%c' got '%c'\n",
335 info_to_arg[call->type], *failpath);
336 call->fail = isupper(*(failpath++));
341 if (!failtest_hook(history, history_num)) {
346 files = save_files();
348 /* We're going to fail in the child. */
350 if (pipe(control) != 0 || pipe(output) != 0)
351 err(1, "opening pipe");
353 /* Prevent double-printing (in child and parent) */
357 err(1, "forking failed");
363 gettimeofday(&now, NULL);
364 if (now.tv_usec < start.tv_usec) {
366 now.tv_usec += 1000000;
368 now.tv_usec -= start.tv_usec;
369 now.tv_sec -= start.tv_sec;
370 sprintf(str, "%u (%u.%02u): ", getpid(),
371 (int)now.tv_sec, (int)now.tv_usec / 10000);
373 p = failpath_string();
377 p = strchr(history[history_num-1].file, '/');
381 trace_str(history[history_num-1].file);
382 sprintf(str, ":%u)\n", history[history_num-1].line);
387 dup2(output[1], STDOUT_FILENO);
388 dup2(output[1], STDERR_FILENO);
389 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
391 control_fd = control[1];
395 signal(SIGUSR1, hand_down);
400 /* We grab output so we can display it; we grab writes so we
403 struct pollfd pfd[2];
406 pfd[0].fd = output[0];
407 pfd[0].events = POLLIN|POLLHUP;
408 pfd[1].fd = control[0];
409 pfd[1].events = POLLIN|POLLHUP;
412 ret = poll(pfd, 1, failtest_timeout_ms);
414 ret = poll(pfd, 2, failtest_timeout_ms);
419 if (pfd[0].revents & POLLIN) {
422 out = realloc(out, outlen + 8192);
423 len = read(output[0], out + outlen, 8192);
425 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
426 if (read_all(control[0], &type, sizeof(type))) {
428 if (!read_write_info(control[0]))
430 } else if (type == RELEASE_LOCKS) {
432 /* FIXME: Tell them we're done... */
435 } else if (pfd[0].revents & POLLHUP) {
438 } while (type != FAILURE);
442 waitpid(child, &status, 0);
443 if (!WIFEXITED(status))
444 child_fail(out, outlen, "Killed by signal %u: ",
446 /* Child printed failure already, just pass up exit code. */
447 if (type == FAILURE) {
448 fprintf(stderr, "%.*s", (int)outlen, out);
450 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
452 if (WEXITSTATUS(status) != 0)
453 child_fail(out, outlen, "Exited with status %i: ",
454 WEXITSTATUS(status));
457 signal(SIGUSR1, SIG_DFL);
459 restore_files(files);
461 /* We continue onwards without failing. */
466 static void cleanup_calloc(struct calloc_call *call)
471 void *failtest_calloc(size_t nmemb, size_t size,
472 const char *file, unsigned line)
474 struct failtest_call *p;
475 struct calloc_call call;
478 p = add_history(FAILTEST_CALLOC, file, line, &call);
480 if (should_fail(p)) {
481 p->u.calloc.ret = NULL;
484 p->u.calloc.ret = calloc(nmemb, size);
485 set_cleanup(p, cleanup_calloc, struct calloc_call);
488 return p->u.calloc.ret;
491 static void cleanup_malloc(struct malloc_call *call)
496 void *failtest_malloc(size_t size, const char *file, unsigned line)
498 struct failtest_call *p;
499 struct malloc_call call;
502 p = add_history(FAILTEST_MALLOC, file, line, &call);
503 if (should_fail(p)) {
504 p->u.calloc.ret = NULL;
507 p->u.calloc.ret = malloc(size);
508 set_cleanup(p, cleanup_malloc, struct malloc_call);
511 return p->u.calloc.ret;
514 static void cleanup_realloc(struct realloc_call *call)
519 /* Walk back and find out if we got this ptr from a previous routine. */
520 static void fixup_ptr_history(void *ptr, unsigned int last)
524 /* Start at end of history, work back. */
525 for (i = last - 1; i >= 0; i--) {
526 switch (history[i].type) {
527 case FAILTEST_REALLOC:
528 if (history[i].u.realloc.ret == ptr) {
529 history[i].cleanup = NULL;
533 case FAILTEST_MALLOC:
534 if (history[i].u.malloc.ret == ptr) {
535 history[i].cleanup = NULL;
539 case FAILTEST_CALLOC:
540 if (history[i].u.calloc.ret == ptr) {
541 history[i].cleanup = NULL;
551 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
553 struct failtest_call *p;
554 struct realloc_call call;
556 p = add_history(FAILTEST_REALLOC, file, line, &call);
558 /* FIXME: Try one child moving allocation, one not. */
559 if (should_fail(p)) {
560 p->u.realloc.ret = NULL;
563 fixup_ptr_history(ptr, history_num-1);
564 p->u.realloc.ret = realloc(ptr, size);
565 set_cleanup(p, cleanup_realloc, struct realloc_call);
568 return p->u.realloc.ret;
571 void failtest_free(void *ptr)
573 fixup_ptr_history(ptr, history_num);
577 static void cleanup_open(struct open_call *call)
582 int failtest_open(const char *pathname,
583 const char *file, unsigned line, ...)
585 struct failtest_call *p;
586 struct open_call call;
589 call.pathname = strdup(pathname);
591 call.flags = va_arg(ap, int);
592 if (call.flags & O_CREAT) {
593 call.mode = va_arg(ap, mode_t);
596 p = add_history(FAILTEST_OPEN, file, line, &call);
597 /* Avoid memory leak! */
598 if (p == &unrecorded_call)
599 free((char *)call.pathname);
600 if (should_fail(p)) {
602 /* FIXME: Play with error codes? */
605 p->u.open.ret = open(pathname, call.flags, call.mode);
606 set_cleanup(p, cleanup_open, struct open_call);
607 p->u.open.dup_fd = p->u.open.ret;
610 return p->u.open.ret;
613 static void cleanup_pipe(struct pipe_call *call)
615 if (!call->closed[0])
617 if (!call->closed[1])
621 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
623 struct failtest_call *p;
624 struct pipe_call call;
626 p = add_history(FAILTEST_PIPE, file, line, &call);
627 if (should_fail(p)) {
629 /* FIXME: Play with error codes? */
632 p->u.pipe.ret = pipe(p->u.pipe.fds);
633 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
634 set_cleanup(p, cleanup_pipe, struct pipe_call);
636 /* This causes valgrind to notice if they use pipefd[] after failure */
637 memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
639 return p->u.pipe.ret;
642 static void cleanup_read(struct read_call *call)
644 lseek(call->fd, call->off, SEEK_SET);
647 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
648 const char *file, unsigned line)
650 struct failtest_call *p;
651 struct read_call call;
656 p = add_history(FAILTEST_READ, file, line, &call);
658 /* FIXME: Try partial read returns. */
659 if (should_fail(p)) {
663 p->u.read.ret = pread(fd, buf, count, off);
664 set_cleanup(p, cleanup_read, struct read_call);
667 return p->u.read.ret;
670 static void cleanup_write(struct write_call *call)
672 lseek(call->dup_fd, call->off, SEEK_SET);
673 write(call->dup_fd, call->saved_contents, call->saved_len);
674 lseek(call->dup_fd, call->off, SEEK_SET);
675 ftruncate(call->dup_fd, call->old_filelen);
676 free(call->saved_contents);
679 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
680 const char *file, unsigned line)
682 struct failtest_call *p;
683 struct write_call call;
685 call.fd = call.dup_fd = fd;
689 p = add_history(FAILTEST_WRITE, file, line, &call);
691 /* Save old contents if we can */
692 if (p->u.write.off != -1) {
694 p->u.write.old_filelen = lseek(fd, 0, SEEK_END);
696 /* Write past end of file? Nothing to save.*/
697 if (p->u.write.old_filelen <= p->u.write.off)
698 p->u.write.saved_len = 0;
699 /* Write which goes over end of file? Partial save. */
700 else if (p->u.write.off + count > p->u.write.old_filelen)
701 p->u.write.saved_len = p->u.write.old_filelen
705 p->u.write.saved_len = count;
707 p->u.write.saved_contents = malloc(p->u.write.saved_len);
708 lseek(fd, p->u.write.off, SEEK_SET);
709 ret = read(fd, p->u.write.saved_contents, p->u.write.saved_len);
710 if (ret != p->u.write.saved_len)
711 err(1, "Expected %i bytes, got %i",
712 (int)p->u.write.saved_len, (int)ret);
713 lseek(fd, p->u.write.off, SEEK_SET);
714 set_cleanup(p, cleanup_write, struct write_call);
717 /* If we're a child, tell parent about write. */
718 if (control_fd != -1) {
719 enum info_type type = WRITE;
721 write_all(control_fd, &type, sizeof(type));
722 write_all(control_fd, &p->u.write, sizeof(p->u.write));
723 write_all(control_fd, buf, count);
726 /* FIXME: Try partial write returns. */
727 if (should_fail(p)) {
731 /* FIXME: We assume same write order in parent and child */
732 if (child_writes_num != 0) {
733 if (child_writes[0].fd != fd)
734 errx(1, "Child wrote to fd %u, not %u?",
735 child_writes[0].fd, fd);
736 if (child_writes[0].off != p->u.write.off)
737 errx(1, "Child wrote to offset %zu, not %zu?",
738 (size_t)child_writes[0].off,
739 (size_t)p->u.write.off);
740 if (child_writes[0].count != count)
741 errx(1, "Child wrote length %zu, not %zu?",
742 child_writes[0].count, count);
743 if (memcmp(child_writes[0].buf, buf, count)) {
745 "Child wrote differently to"
746 " fd %u than we did!\n", fd);
748 free((char *)child_writes[0].buf);
750 memmove(&child_writes[0], &child_writes[1],
751 sizeof(child_writes[0]) * child_writes_num);
753 /* Is this is a socket or pipe, child wrote it
755 if (p->u.write.off == (off_t)-1) {
756 p->u.write.ret = count;
758 return p->u.write.ret;
761 p->u.write.ret = pwrite(fd, buf, count, off);
764 return p->u.write.ret;
767 ssize_t failtest_read(int fd, void *buf, size_t count,
768 const char *file, unsigned line)
770 return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
774 ssize_t failtest_write(int fd, const void *buf, size_t count,
775 const char *file, unsigned line)
777 return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
781 static struct lock_info *WARN_UNUSED_RESULT
782 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
787 for (i = 0; i < lock_num; i++) {
792 /* Four cases we care about:
806 if (start > l->start && end < l->end) {
807 /* Mid overlap: trim entry, add new one. */
808 off_t new_start, new_end;
812 locks = add_lock(locks,
813 fd, new_start, new_end, l->type);
815 } else if (start <= l->start && end >= l->end) {
816 /* Total overlap: eliminate entry. */
819 } else if (end >= l->start && end < l->end) {
820 /* Start overlap: trim entry. */
822 } else if (start > l->start && start <= l->end) {
823 /* End overlap: trim entry. */
826 /* Nothing left? Remove it. */
827 if (l->end < l->start) {
828 memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
833 if (type != F_UNLCK) {
834 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
835 l = &locks[lock_num++];
844 /* We only trap this so we can dup fds in case we need to restore. */
845 int failtest_close(int fd)
852 /* Trace history to find source of fd, and if we need to cleanup writes. */
853 for (i = history_num-1; i >= 0; i--) {
854 switch (history[i].type) {
856 if (history[i].u.write.fd != fd)
858 if (!history[i].cleanup)
860 /* We need to save fd so we can restore file. */
863 history[i].u.write.dup_fd = new_fd;
866 /* We don't need to cleanup reads on closed fds. */
867 if (history[i].u.read.fd != fd)
869 history[i].cleanup = NULL;
872 /* From a pipe? We don't ever restore pipes... */
873 if (history[i].u.pipe.fds[0] == fd) {
874 assert(new_fd == -1);
875 history[i].u.pipe.closed[0] = true;
878 if (history[i].u.pipe.fds[1] == fd) {
879 assert(new_fd == -1);
880 history[i].u.pipe.closed[1] = true;
885 if (history[i].u.open.ret == fd) {
886 history[i].u.open.dup_fd = new_fd;
896 locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
900 /* Zero length means "to end of file" */
901 static off_t end_of(off_t start, off_t len)
905 return start + len - 1;
908 /* FIXME: This only handles locks, really. */
909 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
911 struct failtest_call *p;
912 struct fcntl_call call;
918 /* Argument extraction. */
923 call.arg.l = va_arg(ap, long);
925 return fcntl(fd, cmd, call.arg.l);
928 return fcntl(fd, cmd);
932 call.arg.fl = *va_arg(ap, struct flock *);
934 return fcntl(fd, cmd, &call.arg.fl);
938 call.arg.fl = *va_arg(ap, struct flock *);
942 /* This means you need to implement it here. */
943 err(1, "failtest: unknown fcntl %u", cmd);
946 p = add_history(FAILTEST_FCNTL, file, line, &call);
949 if (should_fail(p)) {
951 if (p->u.fcntl.cmd == F_SETLK)
956 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
958 if (p->u.fcntl.ret == -1)
961 /* We don't handle anything else yet. */
962 assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
963 locks = add_lock(locks,
965 p->u.fcntl.arg.fl.l_start,
966 end_of(p->u.fcntl.arg.fl.l_start,
967 p->u.fcntl.arg.fl.l_len),
968 p->u.fcntl.arg.fl.l_type);
972 return p->u.fcntl.ret;
975 void failtest_init(int argc, char *argv[])
979 for (i = 1; i < argc; i++) {
980 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
981 failpath = argv[i] + strlen("--failpath=");
982 } else if (strcmp(argv[i], "--tracepath") == 0) {
983 tracefd = dup(STDERR_FILENO);
984 failtest_timeout_ms = -1;
987 gettimeofday(&start, NULL);
990 /* Free up memory, so valgrind doesn't report leaks. */
991 static void free_everything(void)
995 /* We don't do this in cleanup: needed even for failed opens. */
996 for (i = 0; i < history_num; i++) {
997 if (history[i].type == FAILTEST_OPEN)
998 free((char *)history[i].u.open.pathname);
1003 void failtest_exit(int status)
1007 if (control_fd == -1) {
1012 if (failtest_exit_check) {
1013 if (!failtest_exit_check(history, history_num))
1014 child_fail(NULL, 0, "failtest_exit_check failed\n");
1017 /* Cleanup everything, in reverse order. */
1018 for (i = history_num - 1; i >= 0; i--)
1019 if (history[i].cleanup)
1020 history[i].cleanup(&history[i].u);
1023 tell_parent(SUCCESS);