]> git.ozlabs.org Git - ccan/blobdiff - ccan/failtest/failtest.c
failtest: detect leaks in children.
[ccan] / ccan / failtest / failtest.c
index 9102ba0a88b884591ef766bdfa9582e4de72d424..77aea9a7de6e185226f07a63b7ecb47dadf3a93b 100644 (file)
 #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,
@@ -47,6 +47,7 @@ static struct failtest_call *history = NULL;
 static unsigned int history_num = 0;
 static int control_fd = -1;
 static struct timeval start;
+static unsigned int probe_count = 0;
 
 static struct write_call *child_writes = NULL;
 static unsigned int child_writes_num = 0;
@@ -55,7 +56,7 @@ 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;
@@ -85,11 +86,6 @@ static struct failtest_call *add_history_(enum failtest_call_type type,
 #define set_cleanup(call, clean, type)                 \
        (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
 
-bool failtest_default_hook(struct failtest_call *history, unsigned num)
-{
-       return true;
-}
-
 static bool read_write_info(int fd)
 {
        struct write_call *w;
@@ -304,6 +300,47 @@ static void restore_files(struct saved_file *s)
        }
 }
 
+/* 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;
@@ -313,6 +350,10 @@ static bool should_fail(struct failtest_call *call)
        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;
 
@@ -329,9 +370,47 @@ static bool should_fail(struct failtest_call *call)
                }
        }
 
-       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();
@@ -431,9 +510,13 @@ static bool should_fail(struct failtest_call *call)
        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);
@@ -588,12 +671,17 @@ int failtest_open(const char *pathname,
        /* 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;
@@ -792,9 +880,20 @@ add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 }
 
 /* We trap this so we can record it: we don't fail it. */
-int failtest_close(int fd)
+int failtest_close(int fd, const char *file, unsigned line)
 {
        int i;
+       struct close_call call;
+       struct failtest_call *p;
+
+       call.fd = fd;
+       p = add_history(FAILTEST_CLOSE, file, line, &call);
+       p->fail = false;
+
+       /* Consume close from failpath. */
+       if (failpath)
+               if (should_fail(p))
+                       abort();
 
        if (fd < 0)
                return close(fd);
@@ -922,44 +1021,20 @@ void failtest_init(int argc, char *argv[])
                } 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;
-
-       /* 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);
-}
-
 void failtest_exit(int status)
 {
-       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");
        }
 
-       /* Cleanup everything, in reverse order. */
-       for (i = history_num - 1; i >= 0; i--)
-               if (history[i].cleanup)
-                       history[i].cleanup(&history[i].u);
-
-       free_everything();
-       tell_parent(SUCCESS);
-       exit(0);
+       failtest_cleanup(false, status);
 }