#include <ccan/failtest/failtest.h>
#include <ccan/build_assert/build_assert.h>
-bool (*failtest_hook)(struct failtest_call *history, unsigned num)
-= failtest_default_hook;
+enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
static int tracefd = -1;
unsigned int failtest_timeout_ms = 20000;
const char *failpath;
+const char *debugpath;
enum info_type {
WRITE,
UNEXPECTED
};
-struct write_info_hdr {
- size_t len;
- off_t offset;
- int fd;
-};
-
-struct fd_orig {
- int fd;
- off_t offset;
- size_t size;
- bool dupped;
-};
-
-struct write_info {
- struct write_info_hdr hdr;
- char *data;
- size_t oldlen;
- char *olddata;
-};
-
struct lock_info {
int fd;
/* end is inclusive: you can't have a 0-byte lock. */
static unsigned int history_num = 0;
static int control_fd = -1;
static struct timeval start;
+static unsigned int probe_count = 0;
-static struct write_info *writes = NULL;
-static unsigned int writes_num = 0;
-
-static struct write_info *child_writes = NULL;
+static struct write_call *child_writes = NULL;
static unsigned int child_writes_num = 0;
-static struct fd_orig *fd_orig = NULL;
-static unsigned int fd_orig_num = 0;
-
static pid_t lock_owner;
static struct lock_info *locks = NULL;
static unsigned int lock_num = 0;
-static const char info_to_arg[] = "mceoprwf";
+static const char info_to_arg[] = "mceoxprwf";
/* Dummy call used for failtest_undo wrappers. */
static struct failtest_call unrecorded_call;
history[history_num].type = type;
history[history_num].file = file;
history[history_num].line = line;
+ history[history_num].cleanup = NULL;
memcpy(&history[history_num].u, elem, elem_size);
return &history[history_num++];
}
#define add_history(type, file, line, elem) \
add_history_((type), (file), (line), (elem), sizeof(*(elem)))
-static void save_fd_orig(int fd)
-{
- unsigned int i;
-
- for (i = 0; i < fd_orig_num; i++)
- if (fd_orig[i].fd == fd)
- return;
-
- fd_orig = realloc(fd_orig, (fd_orig_num + 1) * sizeof(*fd_orig));
- fd_orig[fd_orig_num].fd = fd;
- fd_orig[fd_orig_num].dupped = false;
- fd_orig[fd_orig_num].offset = lseek(fd, 0, SEEK_CUR);
- fd_orig[fd_orig_num].size = lseek(fd, 0, SEEK_END);
- lseek(fd, fd_orig[fd_orig_num].offset, SEEK_SET);
- fd_orig_num++;
-}
-
-bool failtest_default_hook(struct failtest_call *history, unsigned num)
-{
- return true;
-}
+#define set_cleanup(call, clean, type) \
+ (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
static bool read_write_info(int fd)
{
- struct write_info_hdr hdr;
-
- if (!read_all(fd, &hdr, sizeof(hdr)))
- return false;
+ struct write_call *w;
+ char *buf;
+ /* We don't need all of this, but it's simple. */
child_writes = realloc(child_writes,
(child_writes_num+1) * sizeof(child_writes[0]));
- child_writes[child_writes_num].hdr = hdr;
- child_writes[child_writes_num].data = malloc(hdr.len);
- if (!read_all(fd, child_writes[child_writes_num].data, hdr.len))
+ w = &child_writes[child_writes_num];
+ if (!read_all(fd, w, sizeof(*w)))
+ return false;
+
+ w->buf = buf = malloc(w->count);
+ if (!read_all(fd, buf, w->count))
return false;
child_writes_num++;
{
ssize_t ret;
- while ((ret = write(tracefd, str, strlen(str))) <= 0) {
+ while ((ret = write(tracefd, str, strlen(str))) > 0) {
str += ret;
if (!*str)
return;
err(1, "Writing trace.");
}
+struct saved_file {
+ struct saved_file *next;
+ int fd;
+ void *contents;
+ off_t off, len;
+};
+
+static struct saved_file *save_file(struct saved_file *next, int fd)
+{
+ struct saved_file *s = malloc(sizeof(*s));
+
+ s->next = next;
+ s->fd = fd;
+ s->off = lseek(fd, 0, SEEK_CUR);
+ /* Special file? Erk... */
+ assert(s->off != -1);
+ s->len = lseek(fd, 0, SEEK_END);
+ lseek(fd, 0, SEEK_SET);
+ s->contents = malloc(s->len);
+ read(fd, s->contents, s->len);
+ lseek(fd, s->off, SEEK_SET);
+ return s;
+}
+
+/* We have little choice but to save and restore open files: mmap means we
+ * can really intercept changes in the child.
+ *
+ * We could do non-mmap'ed files on demand, however. */
+static struct saved_file *save_files(void)
+{
+ struct saved_file *files = NULL;
+ int i;
+
+ /* Figure out the set of live fds. */
+ for (i = history_num - 2; i >= 0; i--) {
+ if (history[i].type == FAILTEST_OPEN) {
+ int fd = history[i].u.open.ret;
+ /* Only do successful, writable fds. */
+ if (fd < 0)
+ continue;
+
+ /* If it was closed, cleanup == NULL. */
+ if (!history[i].cleanup)
+ continue;
+
+ if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
+ files = save_file(files, fd);
+ } else if ((history[i].u.open.flags & O_WRONLY)
+ == O_WRONLY) {
+ /* FIXME: Handle O_WRONLY. Open with O_RDWR? */
+ abort();
+ }
+ }
+ }
+
+ return files;
+}
+
+static void restore_files(struct saved_file *s)
+{
+ while (s) {
+ struct saved_file *next = s->next;
+
+ lseek(s->fd, 0, SEEK_SET);
+ write(s->fd, s->contents, s->len);
+ ftruncate(s->fd, s->len);
+ free(s->contents);
+ lseek(s->fd, s->off, SEEK_SET);
+ free(s);
+ s = next;
+ }
+}
+
+/* Free up memory, so valgrind doesn't report leaks. */
+static void free_everything(void)
+{
+ unsigned int i;
+
+ /* We don't do this in cleanup: needed even for failed opens. */
+ for (i = 0; i < history_num; i++) {
+ if (history[i].type == FAILTEST_OPEN)
+ free((char *)history[i].u.open.pathname);
+ }
+ free(history);
+}
+
+static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
+{
+ int i;
+
+ /* For children, we don't care if they "failed" the testing. */
+ if (control_fd != -1)
+ status = 0;
+
+ if (forced_cleanup)
+ history_num--;
+
+ /* Cleanup everything, in reverse order. */
+ for (i = history_num - 1; i >= 0; i--) {
+ if (!history[i].cleanup)
+ continue;
+ if (!forced_cleanup) {
+ printf("Leak at %s:%u\n",
+ history[i].file, history[i].line);
+ status = 1;
+ }
+ history[i].cleanup(&history[i].u);
+ }
+
+ free_everything();
+ tell_parent(SUCCESS);
+ exit(status);
+}
+
static bool should_fail(struct failtest_call *call)
{
int status;
enum info_type type = UNEXPECTED;
char *out = NULL;
size_t outlen = 0;
+ struct saved_file *files;
+
+ /* Are we probing? */
+ if (probe_count && --probe_count == 0)
+ failtest_cleanup(true, 0);
if (call == &unrecorded_call)
return false;
}
}
- if (!failtest_hook(history, history_num)) {
- call->fail = false;
- return false;
+ /* Attach debugger if they asked for it. */
+ if (debugpath && history_num == strlen(debugpath)) {
+ unsigned int i;
+
+ for (i = 0; i < history_num; i++) {
+ char c = info_to_arg[history[i].type];
+ if (history[i].fail)
+ c = toupper(c);
+ if (c != debugpath[i])
+ break;
+ }
+ if (i == history_num) {
+ char str[80];
+
+ /* Don't timeout. */
+ signal(SIGUSR1, SIG_IGN);
+ sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
+ getpid(), getpid());
+ system(str);
+ sleep(5);
+ }
+ }
+
+ if (failtest_hook) {
+ switch (failtest_hook(history, history_num)) {
+ case FAIL_OK:
+ break;
+ case FAIL_DONT_FAIL:
+ call->fail = false;
+ return false;
+ case FAIL_PROBE:
+ /* Already down probe path? Stop now. */
+ if (probe_count)
+ failtest_cleanup(true, 0);
+ /* FIXME: We should run *parent* and run probe until
+ * calls match up again. */
+ probe_count = 3;
+ break;
+ default:
+ abort();
+ }
}
+ files = save_files();
+
/* We're going to fail in the child. */
call->fail = true;
if (pipe(control) != 0 || pipe(output) != 0)
close(output[0]);
close(control[0]);
waitpid(child, &status, 0);
- if (!WIFEXITED(status))
- child_fail(out, outlen, "Killed by signal %u: ",
- WTERMSIG(status));
+ if (!WIFEXITED(status)) {
+ if (WTERMSIG(status) == SIGUSR1)
+ child_fail(out, outlen, "Timed out");
+ else
+ child_fail(out, outlen, "Killed by signal %u: ",
+ WTERMSIG(status));
+ }
/* Child printed failure already, just pass up exit code. */
if (type == FAILURE) {
fprintf(stderr, "%.*s", (int)outlen, out);
free(out);
signal(SIGUSR1, SIG_DFL);
+ restore_files(files);
+
/* We continue onwards without failing. */
call->fail = false;
return false;
}
+static void cleanup_calloc(struct calloc_call *call)
+{
+ free(call->ret);
+}
+
void *failtest_calloc(size_t nmemb, size_t size,
const char *file, unsigned line)
{
p->error = ENOMEM;
} else {
p->u.calloc.ret = calloc(nmemb, size);
+ set_cleanup(p, cleanup_calloc, struct calloc_call);
}
errno = p->error;
return p->u.calloc.ret;
}
+static void cleanup_malloc(struct malloc_call *call)
+{
+ free(call->ret);
+}
+
void *failtest_malloc(size_t size, const char *file, unsigned line)
{
struct failtest_call *p;
p->error = ENOMEM;
} else {
p->u.calloc.ret = malloc(size);
+ set_cleanup(p, cleanup_malloc, struct malloc_call);
}
errno = p->error;
return p->u.calloc.ret;
}
+static void cleanup_realloc(struct realloc_call *call)
+{
+ free(call->ret);
+}
+
+/* Walk back and find out if we got this ptr from a previous routine. */
+static void fixup_ptr_history(void *ptr, unsigned int last)
+{
+ int i;
+
+ /* Start at end of history, work back. */
+ for (i = last - 1; i >= 0; i--) {
+ switch (history[i].type) {
+ case FAILTEST_REALLOC:
+ if (history[i].u.realloc.ret == ptr) {
+ history[i].cleanup = NULL;
+ return;
+ }
+ break;
+ case FAILTEST_MALLOC:
+ if (history[i].u.malloc.ret == ptr) {
+ history[i].cleanup = NULL;
+ return;
+ }
+ break;
+ case FAILTEST_CALLOC:
+ if (history[i].u.calloc.ret == ptr) {
+ history[i].cleanup = NULL;
+ return;
+ }
+ break;
+ default:
+ break;
+ }
+ }
+}
+
void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
{
struct failtest_call *p;
p->u.realloc.ret = NULL;
p->error = ENOMEM;
} else {
+ fixup_ptr_history(ptr, history_num-1);
p->u.realloc.ret = realloc(ptr, size);
+ set_cleanup(p, cleanup_realloc, struct realloc_call);
}
errno = p->error;
return p->u.realloc.ret;
}
+void failtest_free(void *ptr)
+{
+ fixup_ptr_history(ptr, history_num);
+ free(ptr);
+}
+
+static void cleanup_open(struct open_call *call)
+{
+ close(call->ret);
+}
+
int failtest_open(const char *pathname,
const char *file, unsigned line, ...)
{
/* Avoid memory leak! */
if (p == &unrecorded_call)
free((char *)call.pathname);
- if (should_fail(p)) {
+ p->u.open.ret = open(pathname, call.flags, call.mode);
+
+ if (!failpath && p->u.open.ret == -1) {
+ p->fail = false;
+ p->error = errno;
+ } else if (should_fail(p)) {
+ close(p->u.open.ret);
p->u.open.ret = -1;
/* FIXME: Play with error codes? */
p->error = EACCES;
} else {
- p->u.open.ret = open(pathname, call.flags, call.mode);
+ set_cleanup(p, cleanup_open, struct open_call);
}
errno = p->error;
return p->u.open.ret;
}
+static void cleanup_pipe(struct pipe_call *call)
+{
+ if (!call->closed[0])
+ close(call->fds[0]);
+ if (!call->closed[1])
+ close(call->fds[1]);
+}
+
int failtest_pipe(int pipefd[2], const char *file, unsigned line)
{
struct failtest_call *p;
p->error = EMFILE;
} else {
p->u.pipe.ret = pipe(p->u.pipe.fds);
+ p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
+ set_cleanup(p, cleanup_pipe, struct pipe_call);
}
/* This causes valgrind to notice if they use pipefd[] after failure */
memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
call.off = off;
p = add_history(FAILTEST_READ, file, line, &call);
- /* This is going to change seek offset, so save it. */
- if (control_fd != -1)
- save_fd_orig(fd);
-
/* FIXME: Try partial read returns. */
if (should_fail(p)) {
p->u.read.ret = -1;
return p->u.read.ret;
}
-static struct write_info *new_write(void)
-{
- writes = realloc(writes, (writes_num + 1) * sizeof(*writes));
- return &writes[writes_num++];
-}
-
ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
const char *file, unsigned line)
{
struct failtest_call *p;
struct write_call call;
- off_t offset;
call.fd = fd;
call.buf = buf;
call.off = off;
p = add_history(FAILTEST_WRITE, file, line, &call);
- offset = lseek(fd, 0, SEEK_CUR);
-
- /* If we're a child, save contents and tell parent about write. */
- if (control_fd != -1) {
- struct write_info *winfo = new_write();
+ /* If we're a child, we need to make sure we write the same thing
+ * to non-files as the parent does, so tell it. */
+ if (control_fd != -1 && off == (off_t)-1) {
enum info_type type = WRITE;
- save_fd_orig(fd);
-
- winfo->hdr.len = count;
- winfo->hdr.fd = fd;
- winfo->data = malloc(count);
- memcpy(winfo->data, buf, count);
- winfo->hdr.offset = offset;
- if (winfo->hdr.offset != (off_t)-1) {
- lseek(fd, offset, SEEK_SET);
- winfo->olddata = malloc(count);
- winfo->oldlen = read(fd, winfo->olddata, count);
- if (winfo->oldlen == -1)
- winfo->oldlen = 0;
- }
write_all(control_fd, &type, sizeof(type));
- write_all(control_fd, &winfo->hdr, sizeof(winfo->hdr));
- write_all(control_fd, winfo->data, count);
+ write_all(control_fd, &p->u.write, sizeof(p->u.write));
+ write_all(control_fd, buf, count);
}
/* FIXME: Try partial write returns. */
p->error = EIO;
} else {
/* FIXME: We assume same write order in parent and child */
- if (child_writes_num != 0) {
- if (child_writes[0].hdr.fd != fd)
+ if (off == (off_t)-1 && child_writes_num != 0) {
+ if (child_writes[0].fd != fd)
errx(1, "Child wrote to fd %u, not %u?",
- child_writes[0].hdr.fd, fd);
- if (child_writes[0].hdr.offset != offset)
+ child_writes[0].fd, fd);
+ if (child_writes[0].off != p->u.write.off)
errx(1, "Child wrote to offset %zu, not %zu?",
- (size_t)child_writes[0].hdr.offset,
- (size_t)offset);
- if (child_writes[0].hdr.len != count)
+ (size_t)child_writes[0].off,
+ (size_t)p->u.write.off);
+ if (child_writes[0].count != count)
errx(1, "Child wrote length %zu, not %zu?",
- child_writes[0].hdr.len, count);
- if (memcmp(child_writes[0].data, buf, count)) {
+ child_writes[0].count, count);
+ if (memcmp(child_writes[0].buf, buf, count)) {
child_fail(NULL, 0,
"Child wrote differently to"
" fd %u than we did!\n", fd);
}
- free(child_writes[0].data);
+ free((char *)child_writes[0].buf);
child_writes_num--;
memmove(&child_writes[0], &child_writes[1],
sizeof(child_writes[0]) * child_writes_num);
/* Is this is a socket or pipe, child wrote it
already. */
- if (offset == (off_t)-1) {
+ if (p->u.write.off == (off_t)-1) {
p->u.write.ret = count;
errno = p->error;
return p->u.write.ret;
return locks;
}
-/* We only trap this so we can dup fds in case we need to restore. */
-int failtest_close(int fd)
+/* We trap this so we can record it: we don't fail it. */
+int failtest_close(int fd, const char *file, unsigned line)
{
- unsigned int i;
- int newfd = -1;
+ int i;
+ struct close_call call;
+ struct failtest_call *p;
- for (i = 0; i < fd_orig_num; i++) {
- if (fd_orig[i].fd == fd) {
- fd_orig[i].fd = newfd = dup(fd);
- fd_orig[i].dupped = true;
- }
- }
+ call.fd = fd;
+ p = add_history(FAILTEST_CLOSE, file, line, &call);
+ p->fail = false;
- for (i = 0; i < writes_num; i++) {
- if (writes[i].hdr.fd == fd)
- writes[i].hdr.fd = newfd;
+ /* Consume close from failpath. */
+ if (failpath)
+ if (should_fail(p))
+ abort();
+
+ if (fd < 0)
+ return close(fd);
+
+ /* Trace history to find source of fd. */
+ for (i = history_num-1; i >= 0; i--) {
+ switch (history[i].type) {
+ case FAILTEST_PIPE:
+ /* From a pipe? */
+ if (history[i].u.pipe.fds[0] == fd) {
+ assert(!history[i].u.pipe.closed[0]);
+ history[i].u.pipe.closed[0] = true;
+ if (history[i].u.pipe.closed[1])
+ history[i].cleanup = NULL;
+ goto out;
+ }
+ if (history[i].u.pipe.fds[1] == fd) {
+ assert(!history[i].u.pipe.closed[1]);
+ history[i].u.pipe.closed[1] = true;
+ if (history[i].u.pipe.closed[0])
+ history[i].cleanup = NULL;
+ goto out;
+ }
+ break;
+ case FAILTEST_OPEN:
+ if (history[i].u.open.ret == fd) {
+ assert((void *)history[i].cleanup
+ == (void *)cleanup_open);
+ history[i].cleanup = NULL;
+ goto out;
+ }
+ break;
+ default:
+ break;
+ }
}
+out:
locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
return close(fd);
}
}
p = add_history(FAILTEST_FCNTL, file, line, &call);
- get_locks();
if (should_fail(p)) {
p->u.fcntl.ret = -1;
else
p->error = EDEADLK;
} else {
+ get_locks();
p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
&p->u.fcntl.arg.fl);
if (p->u.fcntl.ret == -1)
} else if (strcmp(argv[i], "--tracepath") == 0) {
tracefd = dup(STDERR_FILENO);
failtest_timeout_ms = -1;
+ } else if (!strncmp(argv[i], "--debugpath=",
+ strlen("--debugpath="))) {
+ debugpath = argv[i] + strlen("--debugpath=");
}
}
gettimeofday(&start, NULL);
}
-/* Free up memory, so valgrind doesn't report leaks. */
-static void free_everything(void)
-{
- unsigned int i;
-
- for (i = 0; i < writes_num; i++) {
- free(writes[i].data);
- if (writes[i].hdr.offset != (off_t)-1)
- free(writes[i].olddata);
- }
- free(writes);
- free(fd_orig);
- for (i = 0; i < history_num; i++) {
- if (history[i].type == FAILTEST_OPEN)
- free((char *)history[i].u.open.pathname);
- }
- free(history);
-}
-
void failtest_exit(int status)
{
- unsigned int i;
-
- if (control_fd == -1) {
- free_everything();
- exit(status);
- }
-
if (failtest_exit_check) {
if (!failtest_exit_check(history, history_num))
child_fail(NULL, 0, "failtest_exit_check failed\n");
}
- /* Restore any stuff we overwrote. */
- for (i = 0; i < writes_num; i++) {
- if (writes[i].hdr.offset == (off_t)-1)
- continue;
- if (writes[i].oldlen != 0) {
- lseek(writes[i].hdr.fd, writes[i].hdr.offset,
- SEEK_SET);
- write(writes[i].hdr.fd, writes[i].olddata,
- writes[i].oldlen);
- }
- }
-
- /* Fix up fd offsets, restore sizes. */
- for (i = 0; i < fd_orig_num; i++) {
- lseek(fd_orig[i].fd, fd_orig[i].offset, SEEK_SET);
- ftruncate(fd_orig[i].fd, fd_orig[i].size);
- /* Free up any file descriptors we dup'ed. */
- if (fd_orig[i].dupped)
- close(fd_orig[i].fd);
- }
-
- free_everything();
- tell_parent(SUCCESS);
- exit(0);
+ failtest_cleanup(false, status);
}