]> git.ozlabs.org Git - ccan/blobdiff - ccan/failtest/failtest.c
failtest: report failpath problems correctly.
[ccan] / ccan / failtest / failtest.c
index 851f2eabf043e4e3f8fa8c53cfe2056431a27220..a0c825399d03881ab0338075a0a46fb2c4ec1898 100644 (file)
@@ -1,4 +1,5 @@
-#include "config.h"
+/* Licensed under LGPL - see LICENSE file for details */
+#include <ccan/failtest/failtest.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdio.h>
 #include <sys/wait.h>
 #include <sys/stat.h>
 #include <sys/time.h>
+#include <sys/mman.h>
 #include <signal.h>
 #include <assert.h>
+#include <ccan/time/time.h>
 #include <ccan/read_write_all/read_write_all.h>
 #include <ccan/failtest/failtest_proto.h>
-#include <ccan/failtest/failtest.h>
 #include <ccan/build_assert/build_assert.h>
+#include <ccan/hash/hash.h>
+#include <ccan/htable/htable_type.h>
+#include <ccan/str/str.h>
 
-enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
+enum failtest_result (*failtest_hook)(struct tlist_calls *);
 
 static int tracefd = -1;
+static int warnfd;
 
 unsigned int failtest_timeout_ms = 20000;
 
@@ -43,13 +49,44 @@ struct lock_info {
        int type;
 };
 
-bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
+/* We hash the call location together with its backtrace. */
+static size_t hash_call(const struct failtest_call *call)
+{
+       return hash(call->file, strlen(call->file),
+                   hash(&call->line, 1,
+                        hash(call->backtrace, call->backtrace_num,
+                             call->type)));
+}
+
+static bool call_eq(const struct failtest_call *call1,
+                   const struct failtest_call *call2)
+{
+       unsigned int i;
+
+       if (strcmp(call1->file, call2->file) != 0
+           || call1->line != call2->line
+           || call1->type != call2->type
+           || call1->backtrace_num != call2->backtrace_num)
+               return false;
+
+       for (i = 0; i < call1->backtrace_num; i++)
+               if (call1->backtrace[i] != call2->backtrace[i])
+                       return false;
+
+       return true;
+}
+
+/* Defines struct failtable. */
+HTABLE_DEFINE_TYPE(struct failtest_call, (struct failtest_call *), hash_call,
+                  call_eq, failtable);
 
-static struct failtest_call *history = NULL;
-static unsigned int history_num = 0;
+bool (*failtest_exit_check)(struct tlist_calls *history);
+
+static struct tlist_calls history = TLIST_INIT(history);
 static int control_fd = -1;
 static struct timeval start;
-static unsigned int probe_count = 0;
+static bool probing = false;
+static struct failtable failtable;
 
 static struct write_call *child_writes = NULL;
 static unsigned int child_writes_num = 0;
@@ -58,28 +95,62 @@ static pid_t lock_owner;
 static struct lock_info *locks = NULL;
 static unsigned int lock_num = 0;
 
-static const char info_to_arg[] = "mceoxprwf";
+static pid_t orig_pid;
+
+static const char info_to_arg[] = "mceoxprwfa";
 
 /* Dummy call used for failtest_undo wrappers. */
 static struct failtest_call unrecorded_call;
 
+#if HAVE_BACKTRACE
+#include <execinfo.h>
+
+static void **get_backtrace(unsigned int *num)
+{
+       static unsigned int max_back = 100;
+       void **ret;
+
+again:
+       ret = malloc(max_back * sizeof(void *));
+       *num = backtrace(ret, max_back);
+       if (*num == max_back) {
+               free(ret);
+               max_back *= 2;
+               goto again;
+       }
+       return ret;
+}
+#else
+/* This will test slightly less, since will consider all of the same
+ * calls as identical.  But, it's slightly faster! */
+static void **get_backtrace(unsigned int *num)
+{
+       *num = 0;
+       return NULL;
+}
+#endif /* HAVE_BACKTRACE */
+
 static struct failtest_call *add_history_(enum failtest_call_type type,
                                          const char *file,
                                          unsigned int line,
                                          const void *elem,
                                          size_t elem_size)
 {
+       struct failtest_call *call;
+
        /* NULL file is how we suppress failure. */
        if (!file)
                return &unrecorded_call;
 
-       history = realloc(history, (history_num + 1) * sizeof(*history));
-       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++];
+       call = malloc(sizeof *call);
+       call->type = type;
+       call->file = file;
+       call->line = line;
+       call->cleanup = NULL;
+       call->backtrace = get_backtrace(&call->backtrace_num);
+       memcpy(&call->u, elem, elem_size);
+       tlist_add_tail(&history, call, list);
+       return call;
 }
 
 #define add_history(type, file, line, elem) \
@@ -89,6 +160,24 @@ 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),1), (clean))
 
+
+/* Dup the fd to a high value (out of the way I hope!), and close the old fd. */
+static int move_fd_to_high(int fd)
+{
+       int i;
+
+       for (i = FD_SETSIZE - 1; i >= 0; i--) {
+               if (fcntl(i, F_GETFL) == -1 && errno == EBADF) {
+                       if (dup2(fd, i) == -1)
+                               err(1, "Failed to dup fd %i to %i", fd, i);
+                       close(fd);
+                       return i;
+               }
+       }
+       /* Nothing?  Really?  Er... ok? */
+       return fd;
+}
+
 static bool read_write_info(int fd)
 {
        struct write_call *w;
@@ -111,18 +200,52 @@ static bool read_write_info(int fd)
 
 static char *failpath_string(void)
 {
-       unsigned int i;
-       char *ret = malloc(history_num + 1);
-
-       for (i = 0; i < history_num; i++) {
-               ret[i] = info_to_arg[history[i].type];
-               if (history[i].fail)
-                       ret[i] = toupper(ret[i]);
+       struct failtest_call *i;
+       char *ret = strdup("");
+       unsigned len = 0;
+
+       /* Inefficient, but who cares? */
+       tlist_for_each(&history, i, list) {
+               ret = realloc(ret, len + 2);
+               ret[len] = info_to_arg[i->type];
+               if (i->fail)
+                       ret[len] = toupper(ret[len]);
+               ret[++len] = '\0';
        }
-       ret[i] = '\0';
        return ret;
 }
 
+static void warn_via_fd(int e, const char *fmt, va_list ap)
+{
+       char *p = failpath_string();
+
+       vdprintf(warnfd, fmt, ap);
+       if (e != -1)
+               dprintf(warnfd, ": %s", strerror(e));
+       dprintf(warnfd, " [%s]\n", p);
+       free(p);
+}
+
+static void fwarn(const char *fmt, ...)
+{
+       va_list ap;
+       int e = errno;
+
+       va_start(ap, fmt);
+       warn_via_fd(e, fmt, ap);
+       va_end(ap);
+}
+
+
+static void fwarnx(const char *fmt, ...)
+{
+       va_list ap;
+
+       va_start(ap, fmt);
+       warn_via_fd(-1, fmt, ap);
+       va_end(ap);
+}
+
 static void tell_parent(enum info_type type)
 {
        if (control_fd != -1)
@@ -262,23 +385,23 @@ static struct saved_file *save_file(struct saved_file *next, int fd)
 static struct saved_file *save_files(void)
 {
        struct saved_file *files = NULL;
-       int i;
+       struct failtest_call *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;
+       tlist_for_each_rev(&history, i, list) {
+               if (i->type == FAILTEST_OPEN) {
+                       int fd = i->u.open.ret;
                        /* Only do successful, writable fds. */
                        if (fd < 0)
                                continue;
 
                        /* If it was closed, cleanup == NULL. */
-                       if (!history[i].cleanup)
+                       if (!i->cleanup)
                                continue;
 
-                       if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
+                       if ((i->u.open.flags & O_RDWR) == O_RDWR) {
                                files = save_file(files, fd);
-                       } else if ((history[i].u.open.flags & O_WRONLY)
+                       } else if ((i->u.open.flags & O_WRONLY)
                                   == O_WRONLY) {
                                /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
                                abort();
@@ -307,45 +430,68 @@ static void restore_files(struct saved_file *s)
        }
 }
 
+static void free_files(struct saved_file *s)
+{
+       while (s) {
+               struct saved_file *next = s->next;
+               free(s->contents);
+               free(s);
+               s = next;
+       }
+}
+
+static void free_call(struct failtest_call *call)
+{
+       /* We don't do this in cleanup: needed even for failed opens. */
+       if (call->type == FAILTEST_OPEN)
+               free((char *)call->u.open.pathname);
+       free(call->backtrace);
+       tlist_del_from(&history, call, list);
+       free(call);
+}
+
 /* Free up memory, so valgrind doesn't report leaks. */
 static void free_everything(void)
 {
-       unsigned int i;
+       struct failtest_call *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);
+       while ((i = tlist_top(&history, struct failtest_call, list)) != NULL)
+               free_call(i);
+
+       failtable_clear(&failtable);
 }
 
 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 {
-       int i;
+       struct failtest_call *i;
 
        /* For children, we don't care if they "failed" the testing. */
        if (control_fd != -1)
                status = 0;
 
-       if (forced_cleanup)
-               history_num--;
+       if (forced_cleanup) {
+               /* We didn't actually do final operation: remove it. */
+               i = tlist_tail(&history, struct failtest_call, list);
+               free_call(i);
+       }
 
        /* Cleanup everything, in reverse order. */
-       for (i = history_num - 1; i >= 0; i--) {
-               if (!history[i].cleanup)
+       tlist_for_each_rev(&history, i, list) {
+               if (!i->cleanup)
                        continue;
                if (!forced_cleanup) {
                        printf("Leak at %s:%u: --failpath=%s\n",
-                              history[i].file, history[i].line,
-                              failpath_string());
+                              i->file, i->line, failpath_string());
                        status = 1;
                }
-               history[i].cleanup(&history[i].u);
+               i->cleanup(&i->u);
        }
 
        free_everything();
-       tell_parent(SUCCESS);
+       if (status == 0)
+               tell_parent(SUCCESS);
+       else
+               tell_parent(FAILURE);
        exit(status);
 }
 
@@ -357,10 +503,7 @@ static bool should_fail(struct failtest_call *call)
        char *out = NULL;
        size_t outlen = 0;
        struct saved_file *files;
-
-       /* Are we probing? */
-       if (probe_count && --probe_count == 0)
-               failtest_cleanup(true, 0);
+       struct failtest_call *dup;
 
        if (call == &unrecorded_call)
                return false;
@@ -375,25 +518,23 @@ static bool should_fail(struct failtest_call *call)
                } 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((unsigned char)*(failpath++));
+                               errx(1, "Failpath expected '%s' got '%c'\n",
+                                    failpath, info_to_arg[call->type]);
+                       call->fail = cisupper(*(failpath++));
                        return call->fail;
                }
        }
 
        /* Attach debugger if they asked for it. */
-       if (debugpath && history_num == strlen(debugpath)) {
-               unsigned int i;
+       if (debugpath) {
+               char *path;
 
-               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) {
+               /* Pretend this last call matches whatever path wanted:
+                * keeps valgrind happy. */
+               call->fail = cisupper(debugpath[strlen(debugpath)-1]);
+               path = failpath_string();
+
+               if (streq(path, debugpath)) {
                        char str[80];
 
                        /* Don't timeout. */
@@ -402,29 +543,45 @@ static bool should_fail(struct failtest_call *call)
                                getpid(), getpid());
                        if (system(str) == 0)
                                sleep(5);
+               } else {
+                       /* Ignore last character: could be upper or lower. */
+                       path[strlen(path)-1] = '\0';
+                       if (!strstarts(debugpath, path)) {
+                               fprintf(stderr,
+                                       "--debugpath not followed: %s\n", path);
+                               debugpath = NULL;
+                       }
                }
+               free(path);
        }
 
+       /* Are we probing?  If so, we never fail twice. */
+       if (probing)
+               return call->fail = false;
+
+       /* Don't more than once in the same place. */
+       dup = failtable_get(&failtable, call);
+       if (dup)
+               return call->fail = false;
+
        if (failtest_hook) {
-               switch (failtest_hook(history, history_num)) {
+               switch (failtest_hook(&history)) {
                case FAIL_OK:
                        break;
+               case FAIL_PROBE:
+                       probing = true;
+                       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();
                }
        }
 
+       /* Add it to our table of calls. */
+       failtable_add(&failtable, call);
+
        files = save_files();
 
        /* We're going to fail in the child. */
@@ -440,25 +597,24 @@ static bool should_fail(struct failtest_call *call)
 
        if (child == 0) {
                if (tracefd != -1) {
-                       struct timeval now;
+                       struct timeval diff;
                        const char *p;
-                       gettimeofday(&now, NULL);
-                       if (now.tv_usec < start.tv_usec) {
-                               now.tv_sec--;
-                               now.tv_usec += 1000000;
-                       }
-                       now.tv_usec -= start.tv_usec;
-                       now.tv_sec -= start.tv_sec;
-                       p = failpath_string();
+                       char *failpath;
+                       struct failtest_call *c;
+
+                       c = tlist_tail(&history, struct failtest_call, list);
+                       diff = time_sub(time_now(), start);
+                       failpath = failpath_string();
                        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, '/');
+                             (int)diff.tv_sec, (int)diff.tv_usec / 10000,
+                             failpath);
+                       free(failpath);
+                       p = strrchr(c->file, '/');
                        if (p)
                                trace("%s", p+1);
                        else
-                               trace("%s", history[history_num-1].file);
-                       trace(":%u)\n", history[history_num-1].line);
+                               trace("%s", c->file);
+                       trace(":%u)\n", c->line);
                }
                close(control[0]);
                close(output[0]);
@@ -466,7 +622,9 @@ static bool should_fail(struct failtest_call *call)
                dup2(output[1], STDERR_FILENO);
                if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
                        close(output[1]);
-               control_fd = control[1];
+               control_fd = move_fd_to_high(control[1]);
+               /* Valgrind spots the leak if we don't free these. */
+               free_files(files);
                return true;
        }
 
@@ -545,6 +703,9 @@ static bool should_fail(struct failtest_call *call)
 
        restore_files(files);
 
+       /* Only child does probe. */
+       probing = false;
+
        /* We continue onwards without failing. */
        call->fail = false;
        return false;
@@ -588,14 +749,14 @@ void *failtest_malloc(size_t size, const char *file, unsigned line)
 
        p = add_history(FAILTEST_MALLOC, file, line, &call);
        if (should_fail(p)) {
-               p->u.calloc.ret = NULL;
+               p->u.malloc.ret = NULL;
                p->error = ENOMEM;
        } else {
-               p->u.calloc.ret = malloc(size);
+               p->u.malloc.ret = malloc(size);
                set_cleanup(p, cleanup_malloc, struct malloc_call);
        }
        errno = p->error;
-       return p->u.calloc.ret;
+       return p->u.malloc.ret;
 }
 
 static void cleanup_realloc(struct realloc_call *call)
@@ -604,28 +765,28 @@ static void cleanup_realloc(struct realloc_call *call)
 }
 
 /* Walk back and find out if we got this ptr from a previous routine. */
-static void fixup_ptr_history(void *ptr, unsigned int last)
+static void fixup_ptr_history(void *ptr)
 {
-       int i;
+       struct failtest_call *i;
 
        /* Start at end of history, work back. */
-       for (i = last - 1; i >= 0; i--) {
-               switch (history[i].type) {
+       tlist_for_each_rev(&history, i, list) {
+               switch (i->type) {
                case FAILTEST_REALLOC:
-                       if (history[i].u.realloc.ret == ptr) {
-                               history[i].cleanup = NULL;
+                       if (i->u.realloc.ret == ptr) {
+                               i->cleanup = NULL;
                                return;
                        }
                        break;
                case FAILTEST_MALLOC:
-                       if (history[i].u.malloc.ret == ptr) {
-                               history[i].cleanup = NULL;
+                       if (i->u.malloc.ret == ptr) {
+                               i->cleanup = NULL;
                                return;
                        }
                        break;
                case FAILTEST_CALLOC:
-                       if (history[i].u.calloc.ret == ptr) {
-                               history[i].cleanup = NULL;
+                       if (i->u.calloc.ret == ptr) {
+                               i->cleanup = NULL;
                                return;
                        }
                        break;
@@ -647,7 +808,9 @@ void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
                p->u.realloc.ret = NULL;
                p->error = ENOMEM;
        } else {
-               fixup_ptr_history(ptr, history_num-1);
+               /* Don't catch this one in the history fixup... */
+               p->u.realloc.ret = NULL;
+               fixup_ptr_history(ptr);
                p->u.realloc.ret = realloc(ptr, size);
                set_cleanup(p, cleanup_realloc, struct realloc_call);
        }
@@ -657,7 +820,7 @@ void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 
 void failtest_free(void *ptr)
 {
-       fixup_ptr_history(ptr, history_num);
+       fixup_ptr_history(ptr);
        free(ptr);
 }
 
@@ -677,7 +840,7 @@ 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);
@@ -686,7 +849,7 @@ int failtest_open(const char *pathname,
                free((char *)call.pathname);
        p->u.open.ret = open(pathname, call.flags, call.mode);
 
-       if (!failpath && p->u.open.ret == -1) {
+       if (p->u.open.ret == -1) {
                p->fail = false;
                p->error = errno;
        } else if (should_fail(p)) {
@@ -701,6 +864,30 @@ int failtest_open(const char *pathname,
        return p->u.open.ret;
 }
 
+void *failtest_mmap(void *addr, size_t length, int prot, int flags,
+                   int fd, off_t offset, const char *file, unsigned line)
+{
+       struct failtest_call *p;
+       struct mmap_call call;
+
+       call.addr = addr;
+       call.length = length;
+       call.prot = prot;
+       call.flags = flags;
+       call.offset = offset;
+       call.fd = fd;
+
+       p = add_history(FAILTEST_MMAP, file, line, &call);
+       if (should_fail(p)) {
+               p->u.mmap.ret = MAP_FAILED;
+               p->error = ENOMEM;
+       } else {
+               p->u.mmap.ret = mmap(addr, length, prot, flags, fd, offset);
+       }
+       errno = p->error;
+       return p->u.mmap.ret;
+}
+
 static void cleanup_pipe(struct pipe_call *call)
 {
        if (!call->closed[0])
@@ -895,7 +1082,7 @@ 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, const char *file, unsigned line)
 {
-       int i;
+       struct failtest_call *i;
        struct close_call call;
        struct failtest_call *p;
 
@@ -912,30 +1099,30 @@ int failtest_close(int fd, const char *file, unsigned line)
                return close(fd);
 
        /* Trace history to find source of fd. */
-       for (i = history_num-1; i >= 0; i--) {
-               switch (history[i].type) {
+       tlist_for_each_rev(&history, i, list) {
+               switch (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;
+                       if (i->u.pipe.fds[0] == fd) {
+                               assert(!i->u.pipe.closed[0]);
+                               i->u.pipe.closed[0] = true;
+                               if (i->u.pipe.closed[1])
+                                       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;
+                       if (i->u.pipe.fds[1] == fd) {
+                               assert(!i->u.pipe.closed[1]);
+                               i->u.pipe.closed[1] = true;
+                               if (i->u.pipe.closed[0])
+                                       i->cleanup = NULL;
                                goto out;
                        }
                        break;
                case FAILTEST_OPEN:
-                       if (history[i].u.open.ret == fd) {
-                               assert((void *)history[i].cleanup
+                       if (i->u.open.ret == fd) {
+                               assert((void *)i->cleanup
                                       == (void *)cleanup_open);
-                               history[i].cleanup = NULL;
+                               i->cleanup = NULL;
                                goto out;
                        }
                        break;
@@ -1024,28 +1211,44 @@ 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();
+
+       warnfd = move_fd_to_high(dup(STDERR_FILENO));
        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);
+                       tracefd = warnfd;
                        failtest_timeout_ms = -1;
                } else if (!strncmp(argv[i], "--debugpath=",
                                    strlen("--debugpath="))) {
                        debugpath = argv[i] + strlen("--debugpath=");
                }
        }
-       gettimeofday(&start, NULL);
+       failtable_init(&failtable);
+       start = time_now();
+}
+
+bool failtest_has_failed(void)
+{
+       return control_fd != -1;
 }
 
 void failtest_exit(int status)
 {
        if (failtest_exit_check) {
-               if (!failtest_exit_check(history, history_num))
+               if (!failtest_exit_check(&history))
                        child_fail(NULL, 0, "failtest_exit_check failed\n");
        }