1 /* Licensed under LGPL - see LICENSE file for details */
12 #include <sys/types.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>
23 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
25 static int tracefd = -1;
27 unsigned int failtest_timeout_ms = 20000;
30 const char *debugpath;
42 /* end is inclusive: you can't have a 0-byte lock. */
47 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
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;
55 static struct write_call *child_writes = NULL;
56 static unsigned int child_writes_num = 0;
58 static pid_t lock_owner;
59 static struct lock_info *locks = NULL;
60 static unsigned int lock_num = 0;
62 static pid_t orig_pid;
64 static const char info_to_arg[] = "mceoxprwf";
66 /* Dummy call used for failtest_undo wrappers. */
67 static struct failtest_call unrecorded_call;
69 static struct failtest_call *add_history_(enum failtest_call_type type,
75 /* NULL file is how we suppress failure. */
77 return &unrecorded_call;
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++];
88 #define add_history(type, file, line, elem) \
89 add_history_((type), (file), (line), (elem), sizeof(*(elem)))
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))
95 static bool read_write_info(int fd)
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)))
107 w->buf = buf = malloc(w->count);
108 if (!read_all(fd, buf, w->count))
115 static char *failpath_string(void)
118 char *ret = malloc(history_num + 1);
120 for (i = 0; i < history_num; i++) {
121 ret[i] = info_to_arg[history[i].type];
123 ret[i] = toupper(ret[i]);
129 static void tell_parent(enum info_type type)
131 if (control_fd != -1)
132 write_all(control_fd, &type, sizeof(type));
135 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
138 char *path = failpath_string();
141 vfprintf(stderr, fmt, ap);
144 fprintf(stderr, "%.*s", (int)outlen, out);
145 printf("To reproduce: --failpath=%s\n", path);
147 tell_parent(FAILURE);
151 static void trace(const char *fmt, ...)
159 vdprintf(tracefd, fmt, ap);
165 static void hand_down(int signum)
170 static void release_locks(void)
172 /* Locks were never acquired/reacquired? */
176 /* We own them? Release them all. */
177 if (lock_owner == getpid()) {
181 fl.l_whence = SEEK_SET;
185 for (i = 0; i < lock_num; i++)
186 fcntl(locks[i].fd, F_SETLK, &fl);
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));
196 /* off_t is a signed type. Getting its max is non-trivial. */
197 static off_t off_max(void)
199 BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
200 if (sizeof(off_t) == 4)
201 return (off_t)0x7FFFFFF;
203 return (off_t)0x7FFFFFFFFFFFFFFULL;
206 static void get_locks(void)
211 if (lock_owner == getpid())
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));
220 fl.l_whence = SEEK_SET;
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())
228 fl.l_len = locks[i].end - locks[i].start + 1;
230 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
233 lock_owner = getpid();
237 struct saved_file *next;
243 static struct saved_file *save_file(struct saved_file *next, int fd)
245 struct saved_file *s = malloc(sizeof(*s));
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);
261 /* We have little choice but to save and restore open files: mmap means we
262 * can really intercept changes in the child.
264 * We could do non-mmap'ed files on demand, however. */
265 static struct saved_file *save_files(void)
267 struct saved_file *files = NULL;
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. */
278 /* If it was closed, cleanup == NULL. */
279 if (!history[i].cleanup)
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)
286 /* FIXME: Handle O_WRONLY. Open with O_RDWR? */
295 static void restore_files(struct saved_file *s)
298 struct saved_file *next = s->next;
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",
307 lseek(s->fd, s->off, SEEK_SET);
313 /* Free up memory, so valgrind doesn't report leaks. */
314 static void free_everything(void)
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);
326 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
330 /* For children, we don't care if they "failed" the testing. */
331 if (control_fd != -1)
337 /* Cleanup everything, in reverse order. */
338 for (i = history_num - 1; i >= 0; i--) {
339 if (!history[i].cleanup)
341 if (!forced_cleanup) {
342 printf("Leak at %s:%u: --failpath=%s\n",
343 history[i].file, history[i].line,
347 history[i].cleanup(&history[i].u);
351 tell_parent(SUCCESS);
355 static bool should_fail(struct failtest_call *call)
358 int control[2], output[2];
359 enum info_type type = UNEXPECTED;
362 struct saved_file *files;
364 /* Are we probing? */
365 if (probe_count && --probe_count == 0 && control_fd != -1)
366 failtest_cleanup(true, 0);
368 if (call == &unrecorded_call)
372 /* + means continue after end, like normal. */
373 if (*failpath == '+')
375 else if (*failpath == '\0') {
376 /* Continue, but don't inject errors. */
377 return call->fail = false;
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++));
388 /* Attach debugger if they asked for it. */
389 if (debugpath && history_num == strlen(debugpath)) {
392 for (i = 0; i < history_num; i++) {
393 unsigned char c = info_to_arg[history[i].type];
396 if (c != debugpath[i])
399 if (i == history_num) {
403 signal(SIGUSR1, SIG_IGN);
404 sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
406 if (system(str) == 0)
412 switch (failtest_hook(history, history_num)) {
416 /* Already down probe path? Stop now. */
418 /* FIXME: We should run *parent* and
419 * run probe until calls match up again. */
423 /* Child should give up now. */
424 if (control_fd != -1)
425 failtest_cleanup(true, 0);
426 /* Parent, don't fail again. */
436 files = save_files();
438 /* We're going to fail in the child. */
440 if (pipe(control) != 0 || pipe(output) != 0)
441 err(1, "opening pipe");
443 /* Prevent double-printing (in child and parent) */
447 err(1, "forking failed");
453 gettimeofday(&now, NULL);
454 if (now.tv_usec < start.tv_usec) {
456 now.tv_usec += 1000000;
458 now.tv_usec -= start.tv_usec;
459 now.tv_sec -= start.tv_sec;
460 p = failpath_string();
461 trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
462 (int)now.tv_sec, (int)now.tv_usec / 10000, p);
464 p = strrchr(history[history_num-1].file, '/');
468 trace("%s", history[history_num-1].file);
469 trace(":%u)\n", history[history_num-1].line);
473 dup2(output[1], STDOUT_FILENO);
474 dup2(output[1], STDERR_FILENO);
475 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
477 control_fd = control[1];
481 signal(SIGUSR1, hand_down);
486 /* We grab output so we can display it; we grab writes so we
489 struct pollfd pfd[2];
492 pfd[0].fd = output[0];
493 pfd[0].events = POLLIN|POLLHUP;
494 pfd[1].fd = control[0];
495 pfd[1].events = POLLIN|POLLHUP;
498 ret = poll(pfd, 1, failtest_timeout_ms);
500 ret = poll(pfd, 2, failtest_timeout_ms);
507 err(1, "Poll returned %i", ret);
510 if (pfd[0].revents & POLLIN) {
513 out = realloc(out, outlen + 8192);
514 len = read(output[0], out + outlen, 8192);
516 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
517 if (read_all(control[0], &type, sizeof(type))) {
519 if (!read_write_info(control[0]))
521 } else if (type == RELEASE_LOCKS) {
523 /* FIXME: Tell them we're done... */
526 } else if (pfd[0].revents & POLLHUP) {
529 } while (type != FAILURE);
533 waitpid(child, &status, 0);
534 if (!WIFEXITED(status)) {
535 if (WTERMSIG(status) == SIGUSR1)
536 child_fail(out, outlen, "Timed out");
538 child_fail(out, outlen, "Killed by signal %u: ",
541 /* Child printed failure already, just pass up exit code. */
542 if (type == FAILURE) {
543 fprintf(stderr, "%.*s", (int)outlen, out);
545 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
547 if (WEXITSTATUS(status) != 0)
548 child_fail(out, outlen, "Exited with status %i: ",
549 WEXITSTATUS(status));
552 signal(SIGUSR1, SIG_DFL);
554 restore_files(files);
556 /* We continue onwards without failing. */
561 static void cleanup_calloc(struct calloc_call *call)
566 void *failtest_calloc(size_t nmemb, size_t size,
567 const char *file, unsigned line)
569 struct failtest_call *p;
570 struct calloc_call call;
573 p = add_history(FAILTEST_CALLOC, file, line, &call);
575 if (should_fail(p)) {
576 p->u.calloc.ret = NULL;
579 p->u.calloc.ret = calloc(nmemb, size);
580 set_cleanup(p, cleanup_calloc, struct calloc_call);
583 return p->u.calloc.ret;
586 static void cleanup_malloc(struct malloc_call *call)
591 void *failtest_malloc(size_t size, const char *file, unsigned line)
593 struct failtest_call *p;
594 struct malloc_call call;
597 p = add_history(FAILTEST_MALLOC, file, line, &call);
598 if (should_fail(p)) {
599 p->u.calloc.ret = NULL;
602 p->u.calloc.ret = malloc(size);
603 set_cleanup(p, cleanup_malloc, struct malloc_call);
606 return p->u.calloc.ret;
609 static void cleanup_realloc(struct realloc_call *call)
614 /* Walk back and find out if we got this ptr from a previous routine. */
615 static void fixup_ptr_history(void *ptr, unsigned int last)
619 /* Start at end of history, work back. */
620 for (i = last - 1; i >= 0; i--) {
621 switch (history[i].type) {
622 case FAILTEST_REALLOC:
623 if (history[i].u.realloc.ret == ptr) {
624 history[i].cleanup = NULL;
628 case FAILTEST_MALLOC:
629 if (history[i].u.malloc.ret == ptr) {
630 history[i].cleanup = NULL;
634 case FAILTEST_CALLOC:
635 if (history[i].u.calloc.ret == ptr) {
636 history[i].cleanup = NULL;
646 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
648 struct failtest_call *p;
649 struct realloc_call call;
651 p = add_history(FAILTEST_REALLOC, file, line, &call);
653 /* FIXME: Try one child moving allocation, one not. */
654 if (should_fail(p)) {
655 p->u.realloc.ret = NULL;
658 fixup_ptr_history(ptr, history_num-1);
659 p->u.realloc.ret = realloc(ptr, size);
660 set_cleanup(p, cleanup_realloc, struct realloc_call);
663 return p->u.realloc.ret;
666 void failtest_free(void *ptr)
668 fixup_ptr_history(ptr, history_num);
672 static void cleanup_open(struct open_call *call)
677 int failtest_open(const char *pathname,
678 const char *file, unsigned line, ...)
680 struct failtest_call *p;
681 struct open_call call;
684 call.pathname = strdup(pathname);
686 call.flags = va_arg(ap, int);
687 if (call.flags & O_CREAT) {
688 call.mode = va_arg(ap, int);
691 p = add_history(FAILTEST_OPEN, file, line, &call);
692 /* Avoid memory leak! */
693 if (p == &unrecorded_call)
694 free((char *)call.pathname);
695 p->u.open.ret = open(pathname, call.flags, call.mode);
697 if (!failpath && p->u.open.ret == -1) {
700 } else if (should_fail(p)) {
701 close(p->u.open.ret);
703 /* FIXME: Play with error codes? */
706 set_cleanup(p, cleanup_open, struct open_call);
709 return p->u.open.ret;
712 static void cleanup_pipe(struct pipe_call *call)
714 if (!call->closed[0])
716 if (!call->closed[1])
720 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
722 struct failtest_call *p;
723 struct pipe_call call;
725 p = add_history(FAILTEST_PIPE, file, line, &call);
726 if (should_fail(p)) {
728 /* FIXME: Play with error codes? */
731 p->u.pipe.ret = pipe(p->u.pipe.fds);
732 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
733 set_cleanup(p, cleanup_pipe, struct pipe_call);
735 /* This causes valgrind to notice if they use pipefd[] after failure */
736 memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
738 return p->u.pipe.ret;
741 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
742 const char *file, unsigned line)
744 struct failtest_call *p;
745 struct read_call call;
750 p = add_history(FAILTEST_READ, file, line, &call);
752 /* FIXME: Try partial read returns. */
753 if (should_fail(p)) {
757 p->u.read.ret = pread(fd, buf, count, off);
760 return p->u.read.ret;
763 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
764 const char *file, unsigned line)
766 struct failtest_call *p;
767 struct write_call call;
773 p = add_history(FAILTEST_WRITE, file, line, &call);
775 /* If we're a child, we need to make sure we write the same thing
776 * to non-files as the parent does, so tell it. */
777 if (control_fd != -1 && off == (off_t)-1) {
778 enum info_type type = WRITE;
780 write_all(control_fd, &type, sizeof(type));
781 write_all(control_fd, &p->u.write, sizeof(p->u.write));
782 write_all(control_fd, buf, count);
785 /* FIXME: Try partial write returns. */
786 if (should_fail(p)) {
790 /* FIXME: We assume same write order in parent and child */
791 if (off == (off_t)-1 && child_writes_num != 0) {
792 if (child_writes[0].fd != fd)
793 errx(1, "Child wrote to fd %u, not %u?",
794 child_writes[0].fd, fd);
795 if (child_writes[0].off != p->u.write.off)
796 errx(1, "Child wrote to offset %zu, not %zu?",
797 (size_t)child_writes[0].off,
798 (size_t)p->u.write.off);
799 if (child_writes[0].count != count)
800 errx(1, "Child wrote length %zu, not %zu?",
801 child_writes[0].count, count);
802 if (memcmp(child_writes[0].buf, buf, count)) {
804 "Child wrote differently to"
805 " fd %u than we did!\n", fd);
807 free((char *)child_writes[0].buf);
809 memmove(&child_writes[0], &child_writes[1],
810 sizeof(child_writes[0]) * child_writes_num);
812 /* Is this is a socket or pipe, child wrote it
814 if (p->u.write.off == (off_t)-1) {
815 p->u.write.ret = count;
817 return p->u.write.ret;
820 p->u.write.ret = pwrite(fd, buf, count, off);
823 return p->u.write.ret;
826 ssize_t failtest_read(int fd, void *buf, size_t count,
827 const char *file, unsigned line)
829 return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
833 ssize_t failtest_write(int fd, const void *buf, size_t count,
834 const char *file, unsigned line)
836 return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
840 static struct lock_info *WARN_UNUSED_RESULT
841 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
846 for (i = 0; i < lock_num; i++) {
851 /* Four cases we care about:
865 if (start > l->start && end < l->end) {
866 /* Mid overlap: trim entry, add new one. */
867 off_t new_start, new_end;
871 locks = add_lock(locks,
872 fd, new_start, new_end, l->type);
874 } else if (start <= l->start && end >= l->end) {
875 /* Total overlap: eliminate entry. */
878 } else if (end >= l->start && end < l->end) {
879 /* Start overlap: trim entry. */
881 } else if (start > l->start && start <= l->end) {
882 /* End overlap: trim entry. */
885 /* Nothing left? Remove it. */
886 if (l->end < l->start) {
887 memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
892 if (type != F_UNLCK) {
893 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
894 l = &locks[lock_num++];
903 /* We trap this so we can record it: we don't fail it. */
904 int failtest_close(int fd, const char *file, unsigned line)
907 struct close_call call;
908 struct failtest_call *p;
911 p = add_history(FAILTEST_CLOSE, file, line, &call);
914 /* Consume close from failpath. */
922 /* Trace history to find source of fd. */
923 for (i = history_num-1; i >= 0; i--) {
924 switch (history[i].type) {
927 if (history[i].u.pipe.fds[0] == fd) {
928 assert(!history[i].u.pipe.closed[0]);
929 history[i].u.pipe.closed[0] = true;
930 if (history[i].u.pipe.closed[1])
931 history[i].cleanup = NULL;
934 if (history[i].u.pipe.fds[1] == fd) {
935 assert(!history[i].u.pipe.closed[1]);
936 history[i].u.pipe.closed[1] = true;
937 if (history[i].u.pipe.closed[0])
938 history[i].cleanup = NULL;
943 if (history[i].u.open.ret == fd) {
944 assert((void *)history[i].cleanup
945 == (void *)cleanup_open);
946 history[i].cleanup = NULL;
956 locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
960 /* Zero length means "to end of file" */
961 static off_t end_of(off_t start, off_t len)
965 return start + len - 1;
968 /* FIXME: This only handles locks, really. */
969 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
971 struct failtest_call *p;
972 struct fcntl_call call;
978 /* Argument extraction. */
983 call.arg.l = va_arg(ap, long);
985 return fcntl(fd, cmd, call.arg.l);
988 return fcntl(fd, cmd);
992 call.arg.fl = *va_arg(ap, struct flock *);
994 return fcntl(fd, cmd, &call.arg.fl);
998 call.arg.fl = *va_arg(ap, struct flock *);
1002 /* This means you need to implement it here. */
1003 err(1, "failtest: unknown fcntl %u", cmd);
1006 p = add_history(FAILTEST_FCNTL, file, line, &call);
1008 if (should_fail(p)) {
1009 p->u.fcntl.ret = -1;
1010 if (p->u.fcntl.cmd == F_SETLK)
1016 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1017 &p->u.fcntl.arg.fl);
1018 if (p->u.fcntl.ret == -1)
1021 /* We don't handle anything else yet. */
1022 assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1023 locks = add_lock(locks,
1025 p->u.fcntl.arg.fl.l_start,
1026 end_of(p->u.fcntl.arg.fl.l_start,
1027 p->u.fcntl.arg.fl.l_len),
1028 p->u.fcntl.arg.fl.l_type);
1032 return p->u.fcntl.ret;
1035 pid_t failtest_getpid(const char *file, unsigned line)
1037 /* You must call failtest_init first! */
1042 void failtest_init(int argc, char *argv[])
1046 orig_pid = getpid();
1048 for (i = 1; i < argc; i++) {
1049 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1050 failpath = argv[i] + strlen("--failpath=");
1051 } else if (strcmp(argv[i], "--tracepath") == 0) {
1052 tracefd = dup(STDERR_FILENO);
1053 failtest_timeout_ms = -1;
1054 } else if (!strncmp(argv[i], "--debugpath=",
1055 strlen("--debugpath="))) {
1056 debugpath = argv[i] + strlen("--debugpath=");
1059 gettimeofday(&start, NULL);
1062 bool failtest_has_failed(void)
1064 return control_fd != -1;
1067 void failtest_exit(int status)
1069 if (failtest_exit_check) {
1070 if (!failtest_exit_check(history, history_num))
1071 child_fail(NULL, 0, "failtest_exit_check failed\n");
1074 failtest_cleanup(false, status);