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 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
22 static int tracefd = -1;
24 unsigned int failtest_timeout_ms = 20000;
27 const char *debugpath;
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;
50 static unsigned int probe_count = 0;
52 static struct write_call *child_writes = NULL;
53 static unsigned int child_writes_num = 0;
55 static pid_t lock_owner;
56 static struct lock_info *locks = NULL;
57 static unsigned int lock_num = 0;
59 static const char info_to_arg[] = "mceoxprwf";
61 /* Dummy call used for failtest_undo wrappers. */
62 static struct failtest_call unrecorded_call;
64 static struct failtest_call *add_history_(enum failtest_call_type type,
70 /* NULL file is how we suppress failure. */
72 return &unrecorded_call;
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++];
83 #define add_history(type, file, line, elem) \
84 add_history_((type), (file), (line), (elem), sizeof(*(elem)))
86 #define set_cleanup(call, clean, type) \
87 (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
89 static bool read_write_info(int fd)
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)))
101 w->buf = buf = malloc(w->count);
102 if (!read_all(fd, buf, w->count))
109 static char *failpath_string(void)
112 char *ret = malloc(history_num + 1);
114 for (i = 0; i < history_num; i++) {
115 ret[i] = info_to_arg[history[i].type];
117 ret[i] = toupper(ret[i]);
123 static void tell_parent(enum info_type type)
125 if (control_fd != -1)
126 write_all(control_fd, &type, sizeof(type));
129 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
132 char *path = failpath_string();
135 vfprintf(stderr, fmt, ap);
138 fprintf(stderr, "%.*s", (int)outlen, out);
139 printf("To reproduce: --failpath=%s\n", path);
141 tell_parent(FAILURE);
147 static void hand_down(int signal)
152 static void release_locks(void)
154 /* Locks were never acquired/reacquired? */
158 /* We own them? Release them all. */
159 if (lock_owner == getpid()) {
163 fl.l_whence = SEEK_SET;
167 for (i = 0; i < lock_num; i++)
168 fcntl(locks[i].fd, F_SETLK, &fl);
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));
178 /* off_t is a signed type. Getting its max is non-trivial. */
179 static off_t off_max(void)
181 BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
182 if (sizeof(off_t) == 4)
183 return (off_t)0x7FFFFFF;
185 return (off_t)0x7FFFFFFFFFFFFFFULL;
188 static void get_locks(void)
193 if (lock_owner == getpid())
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));
202 fl.l_whence = SEEK_SET;
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())
210 fl.l_len = locks[i].end - locks[i].start + 1;
212 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
215 lock_owner = getpid();
218 static void trace_str(const char *str)
222 while ((ret = write(tracefd, str, strlen(str))) <= 0) {
227 err(1, "Writing trace.");
231 struct saved_file *next;
237 static struct saved_file *save_file(struct saved_file *next, int fd)
239 struct saved_file *s = malloc(sizeof(*s));
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);
254 /* We have little choice but to save and restore open files: mmap means we
255 * can really intercept changes in the child.
257 * We could do non-mmap'ed files on demand, however. */
258 static struct saved_file *save_files(void)
260 struct saved_file *files = NULL;
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. */
271 /* If it was closed, cleanup == NULL. */
272 if (!history[i].cleanup)
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)
279 /* FIXME: Handle O_WRONLY. Open with O_RDWR? */
288 static void restore_files(struct saved_file *s)
291 struct saved_file *next = s->next;
293 lseek(s->fd, 0, SEEK_SET);
294 write(s->fd, s->contents, s->len);
295 ftruncate(s->fd, s->len);
297 lseek(s->fd, s->off, SEEK_SET);
303 /* Free up memory, so valgrind doesn't report leaks. */
304 static void free_everything(void)
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);
316 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
323 /* Cleanup everything, in reverse order. */
324 for (i = history_num - 1; i >= 0; i--)
325 if (history[i].cleanup)
326 history[i].cleanup(&history[i].u);
330 if (control_fd == -1)
333 tell_parent(SUCCESS);
337 static bool should_fail(struct failtest_call *call)
340 int control[2], output[2];
341 enum info_type type = UNEXPECTED;
344 struct saved_file *files;
346 /* Are we probing? */
347 if (probe_count && --probe_count == 0)
348 failtest_cleanup(true, 0);
350 if (call == &unrecorded_call)
354 /* + means continue after end, like normal. */
355 if (*failpath == '+')
358 if (tolower(*failpath) != info_to_arg[call->type])
359 errx(1, "Failpath expected '%c' got '%c'\n",
360 info_to_arg[call->type], *failpath);
361 call->fail = isupper(*(failpath++));
366 /* Attach debugger if they asked for it. */
367 if (debugpath && history_num == strlen(debugpath)) {
370 for (i = 0; i < history_num; i++) {
371 char c = info_to_arg[history[i].type];
374 if (c != debugpath[i])
377 if (i == history_num) {
381 signal(SIGUSR1, SIG_IGN);
382 sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
390 switch (failtest_hook(history, history_num)) {
397 /* Already down probe path? Stop now. */
399 failtest_cleanup(true, 0);
400 /* FIXME: We should run *parent* and run probe until
401 * calls match up again. */
409 files = save_files();
411 /* We're going to fail in the child. */
413 if (pipe(control) != 0 || pipe(output) != 0)
414 err(1, "opening pipe");
416 /* Prevent double-printing (in child and parent) */
420 err(1, "forking failed");
426 gettimeofday(&now, NULL);
427 if (now.tv_usec < start.tv_usec) {
429 now.tv_usec += 1000000;
431 now.tv_usec -= start.tv_usec;
432 now.tv_sec -= start.tv_sec;
433 sprintf(str, "%u (%u.%02u): ", getpid(),
434 (int)now.tv_sec, (int)now.tv_usec / 10000);
436 p = failpath_string();
440 p = strchr(history[history_num-1].file, '/');
444 trace_str(history[history_num-1].file);
445 sprintf(str, ":%u)\n", history[history_num-1].line);
450 dup2(output[1], STDOUT_FILENO);
451 dup2(output[1], STDERR_FILENO);
452 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
454 control_fd = control[1];
458 signal(SIGUSR1, hand_down);
463 /* We grab output so we can display it; we grab writes so we
466 struct pollfd pfd[2];
469 pfd[0].fd = output[0];
470 pfd[0].events = POLLIN|POLLHUP;
471 pfd[1].fd = control[0];
472 pfd[1].events = POLLIN|POLLHUP;
475 ret = poll(pfd, 1, failtest_timeout_ms);
477 ret = poll(pfd, 2, failtest_timeout_ms);
482 if (pfd[0].revents & POLLIN) {
485 out = realloc(out, outlen + 8192);
486 len = read(output[0], out + outlen, 8192);
488 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
489 if (read_all(control[0], &type, sizeof(type))) {
491 if (!read_write_info(control[0]))
493 } else if (type == RELEASE_LOCKS) {
495 /* FIXME: Tell them we're done... */
498 } else if (pfd[0].revents & POLLHUP) {
501 } while (type != FAILURE);
505 waitpid(child, &status, 0);
506 if (!WIFEXITED(status)) {
507 if (WTERMSIG(status) == SIGUSR1)
508 child_fail(out, outlen, "Timed out");
510 child_fail(out, outlen, "Killed by signal %u: ",
513 /* Child printed failure already, just pass up exit code. */
514 if (type == FAILURE) {
515 fprintf(stderr, "%.*s", (int)outlen, out);
517 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
519 if (WEXITSTATUS(status) != 0)
520 child_fail(out, outlen, "Exited with status %i: ",
521 WEXITSTATUS(status));
524 signal(SIGUSR1, SIG_DFL);
526 restore_files(files);
528 /* We continue onwards without failing. */
533 static void cleanup_calloc(struct calloc_call *call)
538 void *failtest_calloc(size_t nmemb, size_t size,
539 const char *file, unsigned line)
541 struct failtest_call *p;
542 struct calloc_call call;
545 p = add_history(FAILTEST_CALLOC, file, line, &call);
547 if (should_fail(p)) {
548 p->u.calloc.ret = NULL;
551 p->u.calloc.ret = calloc(nmemb, size);
552 set_cleanup(p, cleanup_calloc, struct calloc_call);
555 return p->u.calloc.ret;
558 static void cleanup_malloc(struct malloc_call *call)
563 void *failtest_malloc(size_t size, const char *file, unsigned line)
565 struct failtest_call *p;
566 struct malloc_call call;
569 p = add_history(FAILTEST_MALLOC, file, line, &call);
570 if (should_fail(p)) {
571 p->u.calloc.ret = NULL;
574 p->u.calloc.ret = malloc(size);
575 set_cleanup(p, cleanup_malloc, struct malloc_call);
578 return p->u.calloc.ret;
581 static void cleanup_realloc(struct realloc_call *call)
586 /* Walk back and find out if we got this ptr from a previous routine. */
587 static void fixup_ptr_history(void *ptr, unsigned int last)
591 /* Start at end of history, work back. */
592 for (i = last - 1; i >= 0; i--) {
593 switch (history[i].type) {
594 case FAILTEST_REALLOC:
595 if (history[i].u.realloc.ret == ptr) {
596 history[i].cleanup = NULL;
600 case FAILTEST_MALLOC:
601 if (history[i].u.malloc.ret == ptr) {
602 history[i].cleanup = NULL;
606 case FAILTEST_CALLOC:
607 if (history[i].u.calloc.ret == ptr) {
608 history[i].cleanup = NULL;
618 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
620 struct failtest_call *p;
621 struct realloc_call call;
623 p = add_history(FAILTEST_REALLOC, file, line, &call);
625 /* FIXME: Try one child moving allocation, one not. */
626 if (should_fail(p)) {
627 p->u.realloc.ret = NULL;
630 fixup_ptr_history(ptr, history_num-1);
631 p->u.realloc.ret = realloc(ptr, size);
632 set_cleanup(p, cleanup_realloc, struct realloc_call);
635 return p->u.realloc.ret;
638 void failtest_free(void *ptr)
640 fixup_ptr_history(ptr, history_num);
644 static void cleanup_open(struct open_call *call)
649 int failtest_open(const char *pathname,
650 const char *file, unsigned line, ...)
652 struct failtest_call *p;
653 struct open_call call;
656 call.pathname = strdup(pathname);
658 call.flags = va_arg(ap, int);
659 if (call.flags & O_CREAT) {
660 call.mode = va_arg(ap, mode_t);
663 p = add_history(FAILTEST_OPEN, file, line, &call);
664 /* Avoid memory leak! */
665 if (p == &unrecorded_call)
666 free((char *)call.pathname);
667 p->u.open.ret = open(pathname, call.flags, call.mode);
669 if (!failpath && p->u.open.ret == -1) {
672 } else if (should_fail(p)) {
673 close(p->u.open.ret);
675 /* FIXME: Play with error codes? */
678 set_cleanup(p, cleanup_open, struct open_call);
681 return p->u.open.ret;
684 static void cleanup_pipe(struct pipe_call *call)
686 if (!call->closed[0])
688 if (!call->closed[1])
692 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
694 struct failtest_call *p;
695 struct pipe_call call;
697 p = add_history(FAILTEST_PIPE, file, line, &call);
698 if (should_fail(p)) {
700 /* FIXME: Play with error codes? */
703 p->u.pipe.ret = pipe(p->u.pipe.fds);
704 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
705 set_cleanup(p, cleanup_pipe, struct pipe_call);
707 /* This causes valgrind to notice if they use pipefd[] after failure */
708 memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
710 return p->u.pipe.ret;
713 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
714 const char *file, unsigned line)
716 struct failtest_call *p;
717 struct read_call call;
722 p = add_history(FAILTEST_READ, file, line, &call);
724 /* FIXME: Try partial read returns. */
725 if (should_fail(p)) {
729 p->u.read.ret = pread(fd, buf, count, off);
732 return p->u.read.ret;
735 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
736 const char *file, unsigned line)
738 struct failtest_call *p;
739 struct write_call call;
745 p = add_history(FAILTEST_WRITE, file, line, &call);
747 /* If we're a child, we need to make sure we write the same thing
748 * to non-files as the parent does, so tell it. */
749 if (control_fd != -1 && off == (off_t)-1) {
750 enum info_type type = WRITE;
752 write_all(control_fd, &type, sizeof(type));
753 write_all(control_fd, &p->u.write, sizeof(p->u.write));
754 write_all(control_fd, buf, count);
757 /* FIXME: Try partial write returns. */
758 if (should_fail(p)) {
762 /* FIXME: We assume same write order in parent and child */
763 if (off == (off_t)-1 && child_writes_num != 0) {
764 if (child_writes[0].fd != fd)
765 errx(1, "Child wrote to fd %u, not %u?",
766 child_writes[0].fd, fd);
767 if (child_writes[0].off != p->u.write.off)
768 errx(1, "Child wrote to offset %zu, not %zu?",
769 (size_t)child_writes[0].off,
770 (size_t)p->u.write.off);
771 if (child_writes[0].count != count)
772 errx(1, "Child wrote length %zu, not %zu?",
773 child_writes[0].count, count);
774 if (memcmp(child_writes[0].buf, buf, count)) {
776 "Child wrote differently to"
777 " fd %u than we did!\n", fd);
779 free((char *)child_writes[0].buf);
781 memmove(&child_writes[0], &child_writes[1],
782 sizeof(child_writes[0]) * child_writes_num);
784 /* Is this is a socket or pipe, child wrote it
786 if (p->u.write.off == (off_t)-1) {
787 p->u.write.ret = count;
789 return p->u.write.ret;
792 p->u.write.ret = pwrite(fd, buf, count, off);
795 return p->u.write.ret;
798 ssize_t failtest_read(int fd, void *buf, size_t count,
799 const char *file, unsigned line)
801 return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
805 ssize_t failtest_write(int fd, const void *buf, size_t count,
806 const char *file, unsigned line)
808 return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
812 static struct lock_info *WARN_UNUSED_RESULT
813 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
818 for (i = 0; i < lock_num; i++) {
823 /* Four cases we care about:
837 if (start > l->start && end < l->end) {
838 /* Mid overlap: trim entry, add new one. */
839 off_t new_start, new_end;
843 locks = add_lock(locks,
844 fd, new_start, new_end, l->type);
846 } else if (start <= l->start && end >= l->end) {
847 /* Total overlap: eliminate entry. */
850 } else if (end >= l->start && end < l->end) {
851 /* Start overlap: trim entry. */
853 } else if (start > l->start && start <= l->end) {
854 /* End overlap: trim entry. */
857 /* Nothing left? Remove it. */
858 if (l->end < l->start) {
859 memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
864 if (type != F_UNLCK) {
865 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
866 l = &locks[lock_num++];
875 /* We trap this so we can record it: we don't fail it. */
876 int failtest_close(int fd, const char *file, unsigned line)
879 struct close_call call;
880 struct failtest_call *p;
883 p = add_history(FAILTEST_CLOSE, file, line, &call);
886 /* Consume close from failpath. */
894 /* Trace history to find source of fd. */
895 for (i = history_num-1; i >= 0; i--) {
896 switch (history[i].type) {
899 if (history[i].u.pipe.fds[0] == fd) {
900 assert(!history[i].u.pipe.closed[0]);
901 history[i].u.pipe.closed[0] = true;
902 if (history[i].u.pipe.closed[1])
903 history[i].cleanup = NULL;
906 if (history[i].u.pipe.fds[1] == fd) {
907 assert(!history[i].u.pipe.closed[1]);
908 history[i].u.pipe.closed[1] = true;
909 if (history[i].u.pipe.closed[0])
910 history[i].cleanup = NULL;
915 if (history[i].u.open.ret == fd) {
916 assert((void *)history[i].cleanup
917 == (void *)cleanup_open);
918 history[i].cleanup = NULL;
928 locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
932 /* Zero length means "to end of file" */
933 static off_t end_of(off_t start, off_t len)
937 return start + len - 1;
940 /* FIXME: This only handles locks, really. */
941 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
943 struct failtest_call *p;
944 struct fcntl_call call;
950 /* Argument extraction. */
955 call.arg.l = va_arg(ap, long);
957 return fcntl(fd, cmd, call.arg.l);
960 return fcntl(fd, cmd);
964 call.arg.fl = *va_arg(ap, struct flock *);
966 return fcntl(fd, cmd, &call.arg.fl);
970 call.arg.fl = *va_arg(ap, struct flock *);
974 /* This means you need to implement it here. */
975 err(1, "failtest: unknown fcntl %u", cmd);
978 p = add_history(FAILTEST_FCNTL, file, line, &call);
981 if (should_fail(p)) {
983 if (p->u.fcntl.cmd == F_SETLK)
988 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
990 if (p->u.fcntl.ret == -1)
993 /* We don't handle anything else yet. */
994 assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
995 locks = add_lock(locks,
997 p->u.fcntl.arg.fl.l_start,
998 end_of(p->u.fcntl.arg.fl.l_start,
999 p->u.fcntl.arg.fl.l_len),
1000 p->u.fcntl.arg.fl.l_type);
1004 return p->u.fcntl.ret;
1007 void failtest_init(int argc, char *argv[])
1011 for (i = 1; i < argc; i++) {
1012 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1013 failpath = argv[i] + strlen("--failpath=");
1014 } else if (strcmp(argv[i], "--tracepath") == 0) {
1015 tracefd = dup(STDERR_FILENO);
1016 failtest_timeout_ms = -1;
1017 } else if (!strncmp(argv[i], "--debugpath=",
1018 strlen("--debugpath="))) {
1019 debugpath = argv[i] + strlen("--debugpath=");
1022 gettimeofday(&start, NULL);
1025 void failtest_exit(int status)
1027 if (failtest_exit_check) {
1028 if (!failtest_exit_check(history, history_num))
1029 child_fail(NULL, 0, "failtest_exit_check failed\n");
1032 failtest_cleanup(false, status);