X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ffailtest%2Ffailtest.c;h=91420dc699b9c7fdf7f40171b1e39d5fbb7d7ef3;hp=1874d620286d697c9c063319f8435e52dd5823aa;hb=03f017e9aff901d5d80bd577732543961a13c497;hpb=479813329a1ca27f48877b2e872d813f85c957ab diff --git a/ccan/failtest/failtest.c b/ccan/failtest/failtest.c index 1874d620..91420dc6 100644 --- a/ccan/failtest/failtest.c +++ b/ccan/failtest/failtest.c @@ -1,3 +1,5 @@ +/* Licensed under LGPL - see LICENSE file for details */ +#include "config.h" #include #include #include @@ -11,20 +13,21 @@ #include #include #include +#include #include #include #include #include #include -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 +50,7 @@ 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 +59,9 @@ 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 pid_t orig_pid; + +static const char info_to_arg[] = "mceoxprwf"; /* Dummy call used for failtest_undo wrappers. */ static struct failtest_call unrecorded_call; @@ -82,13 +88,9 @@ static struct failtest_call *add_history_(enum failtest_call_type type, #define add_history(type, file, line, elem) \ add_history_((type), (file), (line), (elem), sizeof(*(elem))) +/* We do a fake call inside a sizeof(), to check types. */ #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; -} + (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL),1), (clean)) static bool read_write_info(int fd) { @@ -146,11 +148,23 @@ static void child_fail(const char *out, size_t outlen, const char *fmt, ...) exit(1); } +static void trace(const char *fmt, ...) +{ + va_list ap; + + if (tracefd == -1) + return; + + va_start(ap, fmt); + vdprintf(tracefd, fmt, ap); + va_end(ap); +} + static pid_t child; -static void hand_down(int signal) +static void hand_down(int signum) { - kill(child, signal); + kill(child, signum); } static void release_locks(void) @@ -219,18 +233,6 @@ static void get_locks(void) lock_owner = getpid(); } -static void trace_str(const char *str) -{ - ssize_t ret; - - 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; @@ -250,7 +252,8 @@ static struct saved_file *save_file(struct saved_file *next, int fd) s->len = lseek(fd, 0, SEEK_END); lseek(fd, 0, SEEK_SET); s->contents = malloc(s->len); - read(fd, s->contents, s->len); + if (read(fd, s->contents, s->len) != s->len) + err(1, "Failed to save %zu bytes", (size_t)s->len); lseek(fd, s->off, SEEK_SET); return s; } @@ -295,8 +298,11 @@ static void restore_files(struct saved_file *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); + if (write(s->fd, s->contents, s->len) != s->len) + err(1, "Failed to restore %zu bytes", (size_t)s->len); + if (ftruncate(s->fd, s->len) != 0) + err(1, "Failed to trim file to length %zu", + (size_t)s->len); free(s->contents); lseek(s->fd, s->off, SEEK_SET); free(s); @@ -304,6 +310,48 @@ 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: --failpath=%s\n", + history[i].file, history[i].line, + failpath_string()); + 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 +361,10 @@ 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 && control_fd != -1) + failtest_cleanup(true, 0); + if (call == &unrecorded_call) return false; @@ -320,18 +372,65 @@ static bool should_fail(struct failtest_call *call) /* + means continue after end, like normal. */ if (*failpath == '+') failpath = NULL; - else { - if (tolower(*failpath) != info_to_arg[call->type]) + else if (*failpath == '\0') { + /* Continue, but don't inject errors. */ + return call->fail = false; + } else { + if (tolower((unsigned char)*failpath) + != info_to_arg[call->type]) errx(1, "Failpath expected '%c' got '%c'\n", info_to_arg[call->type], *failpath); - call->fail = isupper(*(failpath++)); + call->fail = isupper((unsigned char)*(failpath++)); return call->fail; } } - 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++) { + unsigned 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()); + if (system(str) == 0) + sleep(5); + } + } + + if (failtest_hook) { + switch (failtest_hook(history, history_num)) { + case FAIL_OK: + break; + case FAIL_PROBE: + /* Already down probe path? Stop now. */ + if (!probe_count) { + /* FIXME: We should run *parent* and + * run probe until calls match up again. */ + probe_count = 3; + break; + } else { + /* Child should give up now. */ + if (control_fd != -1) + failtest_cleanup(true, 0); + /* Parent, don't fail again. */ + } + case FAIL_DONT_FAIL: + call->fail = false; + return false; + default: + abort(); + } } files = save_files(); @@ -350,7 +449,7 @@ static bool should_fail(struct failtest_call *call) if (child == 0) { if (tracefd != -1) { struct timeval now; - char str[50], *p; + const char *p; gettimeofday(&now, NULL); if (now.tv_usec < start.tv_usec) { now.tv_sec--; @@ -358,20 +457,16 @@ static bool should_fail(struct failtest_call *call) } now.tv_usec -= start.tv_usec; now.tv_sec -= start.tv_sec; - sprintf(str, "%u (%u.%02u): ", getpid(), - (int)now.tv_sec, (int)now.tv_usec / 10000); - trace_str(str); p = failpath_string(); - trace_str(p); - free(p); - trace_str("("); - p = strchr(history[history_num-1].file, '/'); + trace("%u->%u (%u.%02u): %s (", getppid(), getpid(), + (int)now.tv_sec, (int)now.tv_usec / 10000, p); + free((char *)p); + p = strrchr(history[history_num-1].file, '/'); if (p) - trace_str(p+1); + trace("%s", p+1); else - trace_str(history[history_num-1].file); - sprintf(str, ":%u)\n", history[history_num-1].line); - trace_str(str); + trace("%s", history[history_num-1].file); + trace(":%u)\n", history[history_num-1].line); } close(control[0]); close(output[0]); @@ -404,8 +499,13 @@ static bool should_fail(struct failtest_call *call) else ret = poll(pfd, 2, failtest_timeout_ms); - if (ret <= 0) + if (ret == 0) hand_down(SIGUSR1); + if (ret < 0) { + if (errno == EINTR) + continue; + err(1, "Poll returned %i", ret); + } if (pfd[0].revents & POLLIN) { ssize_t len; @@ -431,9 +531,13 @@ 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); @@ -581,7 +685,7 @@ int failtest_open(const char *pathname, va_start(ap, line); call.flags = va_arg(ap, int); if (call.flags & O_CREAT) { - call.mode = va_arg(ap, mode_t); + call.mode = va_arg(ap, int); va_end(ap); } p = add_history(FAILTEST_OPEN, file, line, &call); @@ -797,9 +901,20 @@ 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); @@ -889,7 +1004,6 @@ int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...) } p = add_history(FAILTEST_FCNTL, file, line, &call); - get_locks(); if (should_fail(p)) { p->u.fcntl.ret = -1; @@ -898,6 +1012,7 @@ int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...) 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) @@ -917,54 +1032,44 @@ int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...) return p->u.fcntl.ret; } +pid_t failtest_getpid(const char *file, unsigned line) +{ + /* You must call failtest_init first! */ + assert(orig_pid); + return orig_pid; +} + void failtest_init(int argc, char *argv[]) { unsigned int i; + orig_pid = getpid(); + for (i = 1; i < argc; i++) { if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) { failpath = argv[i] + strlen("--failpath="); } 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) +bool failtest_has_failed(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); + return control_fd != -1; } 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); }