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 static void free_files(struct saved_file *s)
316 struct saved_file *next = s->next;
323 /* Free up memory, so valgrind doesn't report leaks. */
324 static void free_everything(void)
328 /* We don't do this in cleanup: needed even for failed opens. */
329 for (i = 0; i < history_num; i++) {
330 if (history[i].type == FAILTEST_OPEN)
331 free((char *)history[i].u.open.pathname);
336 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
340 /* For children, we don't care if they "failed" the testing. */
341 if (control_fd != -1)
347 /* Cleanup everything, in reverse order. */
348 for (i = history_num - 1; i >= 0; i--) {
349 if (!history[i].cleanup)
351 if (!forced_cleanup) {
352 printf("Leak at %s:%u: --failpath=%s\n",
353 history[i].file, history[i].line,
357 history[i].cleanup(&history[i].u);
361 tell_parent(SUCCESS);
365 static bool should_fail(struct failtest_call *call)
368 int control[2], output[2];
369 enum info_type type = UNEXPECTED;
372 struct saved_file *files;
374 /* Are we probing? */
375 if (probe_count && --probe_count == 0 && control_fd != -1)
376 failtest_cleanup(true, 0);
378 if (call == &unrecorded_call)
382 /* + means continue after end, like normal. */
383 if (*failpath == '+')
385 else if (*failpath == '\0') {
386 /* Continue, but don't inject errors. */
387 return call->fail = false;
389 if (tolower((unsigned char)*failpath)
390 != info_to_arg[call->type])
391 errx(1, "Failpath expected '%c' got '%c'\n",
392 info_to_arg[call->type], *failpath);
393 call->fail = isupper((unsigned char)*(failpath++));
398 /* Attach debugger if they asked for it. */
399 if (debugpath && history_num == strlen(debugpath)) {
402 for (i = 0; i < history_num; i++) {
403 unsigned char c = info_to_arg[history[i].type];
406 if (c != debugpath[i])
409 if (i == history_num) {
413 signal(SIGUSR1, SIG_IGN);
414 sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
416 if (system(str) == 0)
422 switch (failtest_hook(history, history_num)) {
426 /* Already down probe path? Stop now. */
428 /* FIXME: We should run *parent* and
429 * run probe until calls match up again. */
433 /* Child should give up now. */
434 if (control_fd != -1)
435 failtest_cleanup(true, 0);
436 /* Parent, don't fail again. */
446 files = save_files();
448 /* We're going to fail in the child. */
450 if (pipe(control) != 0 || pipe(output) != 0)
451 err(1, "opening pipe");
453 /* Prevent double-printing (in child and parent) */
457 err(1, "forking failed");
463 gettimeofday(&now, NULL);
464 if (now.tv_usec < start.tv_usec) {
466 now.tv_usec += 1000000;
468 now.tv_usec -= start.tv_usec;
469 now.tv_sec -= start.tv_sec;
470 p = failpath_string();
471 trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
472 (int)now.tv_sec, (int)now.tv_usec / 10000, p);
474 p = strrchr(history[history_num-1].file, '/');
478 trace("%s", history[history_num-1].file);
479 trace(":%u)\n", history[history_num-1].line);
483 dup2(output[1], STDOUT_FILENO);
484 dup2(output[1], STDERR_FILENO);
485 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
487 control_fd = control[1];
488 /* Valgrind spots the leak if we don't free these. */
493 signal(SIGUSR1, hand_down);
498 /* We grab output so we can display it; we grab writes so we
501 struct pollfd pfd[2];
504 pfd[0].fd = output[0];
505 pfd[0].events = POLLIN|POLLHUP;
506 pfd[1].fd = control[0];
507 pfd[1].events = POLLIN|POLLHUP;
510 ret = poll(pfd, 1, failtest_timeout_ms);
512 ret = poll(pfd, 2, failtest_timeout_ms);
519 err(1, "Poll returned %i", ret);
522 if (pfd[0].revents & POLLIN) {
525 out = realloc(out, outlen + 8192);
526 len = read(output[0], out + outlen, 8192);
528 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
529 if (read_all(control[0], &type, sizeof(type))) {
531 if (!read_write_info(control[0]))
533 } else if (type == RELEASE_LOCKS) {
535 /* FIXME: Tell them we're done... */
538 } else if (pfd[0].revents & POLLHUP) {
541 } while (type != FAILURE);
545 waitpid(child, &status, 0);
546 if (!WIFEXITED(status)) {
547 if (WTERMSIG(status) == SIGUSR1)
548 child_fail(out, outlen, "Timed out");
550 child_fail(out, outlen, "Killed by signal %u: ",
553 /* Child printed failure already, just pass up exit code. */
554 if (type == FAILURE) {
555 fprintf(stderr, "%.*s", (int)outlen, out);
557 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
559 if (WEXITSTATUS(status) != 0)
560 child_fail(out, outlen, "Exited with status %i: ",
561 WEXITSTATUS(status));
564 signal(SIGUSR1, SIG_DFL);
566 restore_files(files);
568 /* We continue onwards without failing. */
573 static void cleanup_calloc(struct calloc_call *call)
578 void *failtest_calloc(size_t nmemb, size_t size,
579 const char *file, unsigned line)
581 struct failtest_call *p;
582 struct calloc_call call;
585 p = add_history(FAILTEST_CALLOC, file, line, &call);
587 if (should_fail(p)) {
588 p->u.calloc.ret = NULL;
591 p->u.calloc.ret = calloc(nmemb, size);
592 set_cleanup(p, cleanup_calloc, struct calloc_call);
595 return p->u.calloc.ret;
598 static void cleanup_malloc(struct malloc_call *call)
603 void *failtest_malloc(size_t size, const char *file, unsigned line)
605 struct failtest_call *p;
606 struct malloc_call call;
609 p = add_history(FAILTEST_MALLOC, file, line, &call);
610 if (should_fail(p)) {
611 p->u.calloc.ret = NULL;
614 p->u.calloc.ret = malloc(size);
615 set_cleanup(p, cleanup_malloc, struct malloc_call);
618 return p->u.calloc.ret;
621 static void cleanup_realloc(struct realloc_call *call)
626 /* Walk back and find out if we got this ptr from a previous routine. */
627 static void fixup_ptr_history(void *ptr, unsigned int last)
631 /* Start at end of history, work back. */
632 for (i = last - 1; i >= 0; i--) {
633 switch (history[i].type) {
634 case FAILTEST_REALLOC:
635 if (history[i].u.realloc.ret == ptr) {
636 history[i].cleanup = NULL;
640 case FAILTEST_MALLOC:
641 if (history[i].u.malloc.ret == ptr) {
642 history[i].cleanup = NULL;
646 case FAILTEST_CALLOC:
647 if (history[i].u.calloc.ret == ptr) {
648 history[i].cleanup = NULL;
658 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
660 struct failtest_call *p;
661 struct realloc_call call;
663 p = add_history(FAILTEST_REALLOC, file, line, &call);
665 /* FIXME: Try one child moving allocation, one not. */
666 if (should_fail(p)) {
667 p->u.realloc.ret = NULL;
670 fixup_ptr_history(ptr, history_num-1);
671 p->u.realloc.ret = realloc(ptr, size);
672 set_cleanup(p, cleanup_realloc, struct realloc_call);
675 return p->u.realloc.ret;
678 void failtest_free(void *ptr)
680 fixup_ptr_history(ptr, history_num);
684 static void cleanup_open(struct open_call *call)
689 int failtest_open(const char *pathname,
690 const char *file, unsigned line, ...)
692 struct failtest_call *p;
693 struct open_call call;
696 call.pathname = strdup(pathname);
698 call.flags = va_arg(ap, int);
699 if (call.flags & O_CREAT) {
700 call.mode = va_arg(ap, int);
703 p = add_history(FAILTEST_OPEN, file, line, &call);
704 /* Avoid memory leak! */
705 if (p == &unrecorded_call)
706 free((char *)call.pathname);
707 p->u.open.ret = open(pathname, call.flags, call.mode);
709 if (!failpath && p->u.open.ret == -1) {
712 } else if (should_fail(p)) {
713 close(p->u.open.ret);
715 /* FIXME: Play with error codes? */
718 set_cleanup(p, cleanup_open, struct open_call);
721 return p->u.open.ret;
724 static void cleanup_pipe(struct pipe_call *call)
726 if (!call->closed[0])
728 if (!call->closed[1])
732 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
734 struct failtest_call *p;
735 struct pipe_call call;
737 p = add_history(FAILTEST_PIPE, file, line, &call);
738 if (should_fail(p)) {
740 /* FIXME: Play with error codes? */
743 p->u.pipe.ret = pipe(p->u.pipe.fds);
744 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
745 set_cleanup(p, cleanup_pipe, struct pipe_call);
747 /* This causes valgrind to notice if they use pipefd[] after failure */
748 memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
750 return p->u.pipe.ret;
753 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
754 const char *file, unsigned line)
756 struct failtest_call *p;
757 struct read_call call;
762 p = add_history(FAILTEST_READ, file, line, &call);
764 /* FIXME: Try partial read returns. */
765 if (should_fail(p)) {
769 p->u.read.ret = pread(fd, buf, count, off);
772 return p->u.read.ret;
775 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
776 const char *file, unsigned line)
778 struct failtest_call *p;
779 struct write_call call;
785 p = add_history(FAILTEST_WRITE, file, line, &call);
787 /* If we're a child, we need to make sure we write the same thing
788 * to non-files as the parent does, so tell it. */
789 if (control_fd != -1 && off == (off_t)-1) {
790 enum info_type type = WRITE;
792 write_all(control_fd, &type, sizeof(type));
793 write_all(control_fd, &p->u.write, sizeof(p->u.write));
794 write_all(control_fd, buf, count);
797 /* FIXME: Try partial write returns. */
798 if (should_fail(p)) {
802 /* FIXME: We assume same write order in parent and child */
803 if (off == (off_t)-1 && child_writes_num != 0) {
804 if (child_writes[0].fd != fd)
805 errx(1, "Child wrote to fd %u, not %u?",
806 child_writes[0].fd, fd);
807 if (child_writes[0].off != p->u.write.off)
808 errx(1, "Child wrote to offset %zu, not %zu?",
809 (size_t)child_writes[0].off,
810 (size_t)p->u.write.off);
811 if (child_writes[0].count != count)
812 errx(1, "Child wrote length %zu, not %zu?",
813 child_writes[0].count, count);
814 if (memcmp(child_writes[0].buf, buf, count)) {
816 "Child wrote differently to"
817 " fd %u than we did!\n", fd);
819 free((char *)child_writes[0].buf);
821 memmove(&child_writes[0], &child_writes[1],
822 sizeof(child_writes[0]) * child_writes_num);
824 /* Is this is a socket or pipe, child wrote it
826 if (p->u.write.off == (off_t)-1) {
827 p->u.write.ret = count;
829 return p->u.write.ret;
832 p->u.write.ret = pwrite(fd, buf, count, off);
835 return p->u.write.ret;
838 ssize_t failtest_read(int fd, void *buf, size_t count,
839 const char *file, unsigned line)
841 return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
845 ssize_t failtest_write(int fd, const void *buf, size_t count,
846 const char *file, unsigned line)
848 return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
852 static struct lock_info *WARN_UNUSED_RESULT
853 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
858 for (i = 0; i < lock_num; i++) {
863 /* Four cases we care about:
877 if (start > l->start && end < l->end) {
878 /* Mid overlap: trim entry, add new one. */
879 off_t new_start, new_end;
883 locks = add_lock(locks,
884 fd, new_start, new_end, l->type);
886 } else if (start <= l->start && end >= l->end) {
887 /* Total overlap: eliminate entry. */
890 } else if (end >= l->start && end < l->end) {
891 /* Start overlap: trim entry. */
893 } else if (start > l->start && start <= l->end) {
894 /* End overlap: trim entry. */
897 /* Nothing left? Remove it. */
898 if (l->end < l->start) {
899 memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
904 if (type != F_UNLCK) {
905 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
906 l = &locks[lock_num++];
915 /* We trap this so we can record it: we don't fail it. */
916 int failtest_close(int fd, const char *file, unsigned line)
919 struct close_call call;
920 struct failtest_call *p;
923 p = add_history(FAILTEST_CLOSE, file, line, &call);
926 /* Consume close from failpath. */
934 /* Trace history to find source of fd. */
935 for (i = history_num-1; i >= 0; i--) {
936 switch (history[i].type) {
939 if (history[i].u.pipe.fds[0] == fd) {
940 assert(!history[i].u.pipe.closed[0]);
941 history[i].u.pipe.closed[0] = true;
942 if (history[i].u.pipe.closed[1])
943 history[i].cleanup = NULL;
946 if (history[i].u.pipe.fds[1] == fd) {
947 assert(!history[i].u.pipe.closed[1]);
948 history[i].u.pipe.closed[1] = true;
949 if (history[i].u.pipe.closed[0])
950 history[i].cleanup = NULL;
955 if (history[i].u.open.ret == fd) {
956 assert((void *)history[i].cleanup
957 == (void *)cleanup_open);
958 history[i].cleanup = NULL;
968 locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
972 /* Zero length means "to end of file" */
973 static off_t end_of(off_t start, off_t len)
977 return start + len - 1;
980 /* FIXME: This only handles locks, really. */
981 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
983 struct failtest_call *p;
984 struct fcntl_call call;
990 /* Argument extraction. */
995 call.arg.l = va_arg(ap, long);
997 return fcntl(fd, cmd, call.arg.l);
1000 return fcntl(fd, cmd);
1004 call.arg.fl = *va_arg(ap, struct flock *);
1006 return fcntl(fd, cmd, &call.arg.fl);
1010 call.arg.fl = *va_arg(ap, struct flock *);
1014 /* This means you need to implement it here. */
1015 err(1, "failtest: unknown fcntl %u", cmd);
1018 p = add_history(FAILTEST_FCNTL, file, line, &call);
1020 if (should_fail(p)) {
1021 p->u.fcntl.ret = -1;
1022 if (p->u.fcntl.cmd == F_SETLK)
1028 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1029 &p->u.fcntl.arg.fl);
1030 if (p->u.fcntl.ret == -1)
1033 /* We don't handle anything else yet. */
1034 assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1035 locks = add_lock(locks,
1037 p->u.fcntl.arg.fl.l_start,
1038 end_of(p->u.fcntl.arg.fl.l_start,
1039 p->u.fcntl.arg.fl.l_len),
1040 p->u.fcntl.arg.fl.l_type);
1044 return p->u.fcntl.ret;
1047 pid_t failtest_getpid(const char *file, unsigned line)
1049 /* You must call failtest_init first! */
1054 void failtest_init(int argc, char *argv[])
1058 orig_pid = getpid();
1060 for (i = 1; i < argc; i++) {
1061 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1062 failpath = argv[i] + strlen("--failpath=");
1063 } else if (strcmp(argv[i], "--tracepath") == 0) {
1064 tracefd = dup(STDERR_FILENO);
1065 failtest_timeout_ms = -1;
1066 } else if (!strncmp(argv[i], "--debugpath=",
1067 strlen("--debugpath="))) {
1068 debugpath = argv[i] + strlen("--debugpath=");
1071 gettimeofday(&start, NULL);
1074 bool failtest_has_failed(void)
1076 return control_fd != -1;
1079 void failtest_exit(int status)
1081 if (failtest_exit_check) {
1082 if (!failtest_exit_check(history, history_num))
1083 child_fail(NULL, 0, "failtest_exit_check failed\n");
1086 failtest_cleanup(false, status);