]> git.ozlabs.org Git - ccan/blobdiff - ccan/failtest/failtest.c
failtest: report failpath problems correctly.
[ccan] / ccan / failtest / failtest.c
index 1118466ebc5d963bb78c8a218e73689a37c164bc..a0c825399d03881ab0338075a0a46fb2c4ec1898 100644 (file)
@@ -1,3 +1,5 @@
+/* Licensed under LGPL - see LICENSE file for details */
+#include <ccan/failtest/failtest.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdio.h>
 #include <sys/types.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>
 
-bool (*failtest_hook)(struct failtest_call *history, unsigned num)
-= failtest_default_hook;
+enum failtest_result (*failtest_hook)(struct tlist_calls *);
+
+static int tracefd = -1;
+static int warnfd;
 
 unsigned int failtest_timeout_ms = 20000;
 
 const char *failpath;
+const char *debugpath;
 
 enum info_type {
        WRITE,
@@ -31,26 +42,6 @@ enum info_type {
        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. */
@@ -58,103 +49,201 @@ 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 struct failtest_call *history = NULL;
-static unsigned int history_num = 0;
-static int control_fd = -1;
+static bool call_eq(const struct failtest_call *call1,
+                   const struct failtest_call *call2)
+{
+       unsigned int i;
 
-static struct write_info *writes = NULL;
-static unsigned int writes_num = 0;
+       if (strcmp(call1->file, call2->file) != 0
+           || call1->line != call2->line
+           || call1->type != call2->type
+           || call1->backtrace_num != call2->backtrace_num)
+               return false;
 
-static struct write_info *child_writes = NULL;
-static unsigned int child_writes_num = 0;
+       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 fd_orig *fd_orig = NULL;
-static unsigned int fd_orig_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 bool probing = false;
+static struct failtable failtable;
+
+static struct write_call *child_writes = NULL;
+static unsigned int child_writes_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 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;
-       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) \
        add_history_((type), (file), (line), (elem), sizeof(*(elem)))
 
-static void save_fd_orig(int fd)
-{
-       unsigned int i;
+/* 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),1), (clean))
 
-       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)
+/* 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)
 {
-       return true;
+       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_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++;
        return true;
 }
 
-static void print_reproduce(void)
+static char *failpath_string(void)
 {
-       unsigned int i;
-
-       printf("To reproduce: --failpath=");
-       for (i = 0; i < history_num; i++) {
-               if (history[i].fail)
-                       printf("%c", toupper(info_to_arg[history[i].type]));
-               else
-                       printf("%c", info_to_arg[history[i].type]);
+       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';
        }
-       printf("\n");
+       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)
@@ -166,22 +255,36 @@ static void tell_parent(enum info_type type)
 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 {
        va_list ap;
+       char *path = failpath_string();
 
        va_start(ap, fmt);
        vfprintf(stderr, fmt, ap);
        va_end(ap);
 
        fprintf(stderr, "%.*s", (int)outlen, out);
-       print_reproduce();
+       printf("To reproduce: --failpath=%s\n", path);
+       free(path);
        tell_parent(FAILURE);
        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)
@@ -250,6 +353,148 @@ static void get_locks(void)
        lock_owner = getpid();
 }
 
+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);
+       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;
+}
+       
+/* 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;
+       struct failtest_call *i;
+
+       /* Figure out the set of live fds. */
+       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 (!i->cleanup)
+                               continue;
+
+                       if ((i->u.open.flags & O_RDWR) == O_RDWR) {
+                               files = save_file(files, fd);
+                       } else if ((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);
+               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);
+               s = next;
+       }
+}
+
+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)
+{
+       struct failtest_call *i;
+
+       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)
+{
+       struct failtest_call *i;
+
+       /* For children, we don't care if they "failed" the testing. */
+       if (control_fd != -1)
+               status = 0;
+
+       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. */
+       tlist_for_each_rev(&history, i, list) {
+               if (!i->cleanup)
+                       continue;
+               if (!forced_cleanup) {
+                       printf("Leak at %s:%u: --failpath=%s\n",
+                              i->file, i->line, failpath_string());
+                       status = 1;
+               }
+               i->cleanup(&i->u);
+       }
+
+       free_everything();
+       if (status == 0)
+               tell_parent(SUCCESS);
+       else
+               tell_parent(FAILURE);
+       exit(status);
+}
+
 static bool should_fail(struct failtest_call *call)
 {
        int status;
@@ -257,22 +502,88 @@ static bool should_fail(struct failtest_call *call)
        enum info_type type = UNEXPECTED;
        char *out = NULL;
        size_t outlen = 0;
+       struct saved_file *files;
+       struct failtest_call *dup;
 
        if (call == &unrecorded_call)
                return false;
 
        if (failpath) {
-               if (tolower(*failpath) != info_to_arg[call->type])
-                       errx(1, "Failpath expected '%c' got '%c'\n",
-                            info_to_arg[call->type], *failpath);
-               return isupper(*(failpath++));
+               /* + means continue after end, like normal. */
+               if (*failpath == '+')
+                       failpath = NULL;
+               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 '%s' got '%c'\n",
+                                    failpath, info_to_arg[call->type]);
+                       call->fail = cisupper(*(failpath++));
+                       return call->fail;
+               }
        }
 
-       if (!failtest_hook(history, history_num)) {
-               call->fail = false;
-               return false;
+       /* Attach debugger if they asked for it. */
+       if (debugpath) {
+               char *path;
+
+               /* 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. */
+                       signal(SIGUSR1, SIG_IGN);
+                       sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
+                               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)) {
+               case FAIL_OK:
+                       break;
+               case FAIL_PROBE:
+                       probing = true;
+                       break;
+               case FAIL_DONT_FAIL:
+                       call->fail = false;
+                       return false;
+               default:
+                       abort();
+               }
        }
 
+       /* Add it to our table of calls. */
+       failtable_add(&failtable, call);
+
+       files = save_files();
+
        /* We're going to fail in the child. */
        call->fail = true;
        if (pipe(control) != 0 || pipe(output) != 0)
@@ -285,13 +596,35 @@ static bool should_fail(struct failtest_call *call)
                err(1, "forking failed");
 
        if (child == 0) {
+               if (tracefd != -1) {
+                       struct timeval diff;
+                       const char *p;
+                       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)diff.tv_sec, (int)diff.tv_usec / 10000,
+                             failpath);
+                       free(failpath);
+                       p = strrchr(c->file, '/');
+                       if (p)
+                               trace("%s", p+1);
+                       else
+                               trace("%s", c->file);
+                       trace(":%u)\n", c->line);
+               }
                close(control[0]);
                close(output[0]);
                dup2(output[1], STDOUT_FILENO);
                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;
        }
 
@@ -316,8 +649,13 @@ 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;
@@ -343,9 +681,13 @@ 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);
@@ -359,11 +701,21 @@ static bool should_fail(struct failtest_call *call)
        free(out);
        signal(SIGUSR1, SIG_DFL);
 
+       restore_files(files);
+
+       /* Only child does probe. */
+       probing = false;
+
        /* 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)
 {
@@ -378,11 +730,17 @@ void *failtest_calloc(size_t nmemb, size_t size,
                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;
@@ -391,13 +749,51 @@ 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)
+{
+       free(call->ret);
+}
+
+/* Walk back and find out if we got this ptr from a previous routine. */
+static void fixup_ptr_history(void *ptr)
+{
+       struct failtest_call *i;
+
+       /* Start at end of history, work back. */
+       tlist_for_each_rev(&history, i, list) {
+               switch (i->type) {
+               case FAILTEST_REALLOC:
+                       if (i->u.realloc.ret == ptr) {
+                               i->cleanup = NULL;
+                               return;
+                       }
+                       break;
+               case FAILTEST_MALLOC:
+                       if (i->u.malloc.ret == ptr) {
+                               i->cleanup = NULL;
+                               return;
+                       }
+                       break;
+               case FAILTEST_CALLOC:
+                       if (i->u.calloc.ret == ptr) {
+                               i->cleanup = NULL;
+                               return;
+                       }
+                       break;
+               default:
+                       break;
+               }
+       }
 }
 
 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
@@ -412,12 +808,27 @@ void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
                p->u.realloc.ret = NULL;
                p->error = ENOMEM;
        } else {
+               /* 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);
        }
        errno = p->error;
        return p->u.realloc.ret;
 }
 
+void failtest_free(void *ptr)
+{
+       fixup_ptr_history(ptr);
+       free(ptr);
+}
+
+static void cleanup_open(struct open_call *call)
+{
+       close(call->ret);
+}
+
 int failtest_open(const char *pathname,
                  const char *file, unsigned line, ...)
 {
@@ -429,24 +840,62 @@ 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);
        /* 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 (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;
 }
 
+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])
+               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;
@@ -459,6 +908,8 @@ int failtest_pipe(int pipefd[2], const char *file, unsigned line)
                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));
@@ -466,73 +917,48 @@ int failtest_pipe(int pipefd[2], const char *file, unsigned line)
        return p->u.pipe.ret;
 }
 
-ssize_t failtest_read(int fd, void *buf, size_t count,
-                     const char *file, unsigned line)
+ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
+                      const char *file, unsigned line)
 {
        struct failtest_call *p;
        struct read_call call;
        call.fd = fd;
        call.buf = buf;
        call.count = count;
+       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;
                p->error = EIO;
        } else {
-               p->u.read.ret = read(fd, buf, count);
+               p->u.read.ret = pread(fd, buf, count, off);
        }
        errno = p->error;
        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_write(int fd, const void *buf, size_t count,
-                      const char *file, unsigned line)
+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.count = count;
+       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. */
@@ -541,41 +967,55 @@ ssize_t failtest_write(int fd, const void *buf, size_t count,
                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;
                        }
                }
-               p->u.write.ret = write(fd, buf, count);
+               p->u.write.ret = pwrite(fd, buf, count, off);
        }
        errno = p->error;
        return p->u.write.ret;
 }
 
+ssize_t failtest_read(int fd, void *buf, size_t count,
+                     const char *file, unsigned line)
+{
+       return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
+                             file, line);
+}
+
+ssize_t failtest_write(int fd, const void *buf, size_t count,
+                      const char *file, unsigned line)
+{
+       return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
+                              file, line);
+}
+
 static struct lock_info *WARN_UNUSED_RESULT
 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 {
@@ -639,24 +1079,59 @@ add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
        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;
+       struct failtest_call *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;
+
+       /* Consume close from failpath. */
+       if (failpath)
+               if (should_fail(p))
+                       abort();
 
-       for (i = 0; i < writes_num; i++) {
-               if (writes[i].hdr.fd == fd)
-                       writes[i].hdr.fd = newfd;
+       if (fd < 0)
+               return close(fd);
+
+       /* Trace history to find source of fd. */
+       tlist_for_each_rev(&history, i, list) {
+               switch (i->type) {
+               case FAILTEST_PIPE:
+                       /* From a pipe? */
+                       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 (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 (i->u.open.ret == fd) {
+                               assert((void *)i->cleanup
+                                      == (void *)cleanup_open);
+                               i->cleanup = NULL;
+                               goto out;
+                       }
+                       break;
+               default:
+                       break;
+               }
        }
 
+out:
        locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
        return close(fd);
 }
@@ -708,7 +1183,6 @@ 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;
@@ -717,6 +1191,7 @@ 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)
@@ -736,69 +1211,46 @@ int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
        return p->u.fcntl.ret;
 }
 
-void failtest_init(int argc, char *argv[])
+pid_t failtest_getpid(const char *file, unsigned line)
 {
-       if (argc == 2
-           && strncmp(argv[1], "--failpath=", strlen("--failpath=")) == 0) {
-               failpath = argv[1] + strlen("--failpath=");
-       }
+       /* You must call failtest_init first! */
+       assert(orig_pid);
+       return orig_pid;
 }
-
-/* Free up memory, so valgrind doesn't report leaks. */
-static void free_everything(void)
+       
+void failtest_init(int argc, char *argv[])
 {
        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);
+       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 = warnfd;
+                       failtest_timeout_ms = -1;
+               } else if (!strncmp(argv[i], "--debugpath=",
+                                   strlen("--debugpath="))) {
+                       debugpath = argv[i] + strlen("--debugpath=");
+               }
        }
-       free(history);
+       failtable_init(&failtable);
+       start = time_now();
 }
 
-void failtest_exit(int status)
+bool failtest_has_failed(void)
 {
-       unsigned int i;
-
-       if (control_fd == -1) {
-               free_everything();
-               exit(status);
-       }
+       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");
        }
 
-       /* 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);
 }