]> git.ozlabs.org Git - ccan/blobdiff - ccan/failtest/failtest.c
failtest: use ccan/time
[ccan] / ccan / failtest / failtest.c
index 5d7a54f187f1f627ed6c950e9c7f06dfef82b1e1..bac0aa350ac24378e8af2b949cb0781a7699906a 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 <fcntl.h>
+#include <sys/time.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>
 
-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,
+       RELEASE_LOCKS,
        FAILURE,
        SUCCESS,
        UNEXPECTED
 };
 
-struct write_info_hdr {
-       size_t len;
-       off_t offset;
-       int fd;
-};
-
-struct fd_orig {
+struct lock_info {
        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;
+       /* end is inclusive: you can't have a 0-byte lock. */
+       off_t start, end;
+       int type;
 };
 
 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
@@ -54,17 +49,19 @@ bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
 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_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 pid_t orig_pid;
 
-static const char info_to_arg[] = "mceoprw";
+static const char info_to_arg[] = "mceoxprwf";
 
 /* Dummy call used for failtest_undo wrappers. */
 static struct failtest_call unrecorded_call;
@@ -83,6 +80,7 @@ static struct failtest_call *add_history_(enum failtest_call_type type,
        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++];
 }
@@ -90,58 +88,42 @@ 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)))
 
-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;
-}
+/* 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))
 
 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;
+       char *ret = malloc(history_num + 1);
 
-       printf("To reproduce: --failpath=");
        for (i = 0; i < history_num; i++) {
+               ret[i] = info_to_arg[history[i].type];
                if (history[i].fail)
-                       printf("%c", toupper(info_to_arg[history[i].type]));
-               else
-                       printf("%c", info_to_arg[history[i].type]);
+                       ret[i] = toupper(ret[i]);
        }
-       printf("\n");
+       ret[i] = '\0';
+       return ret;
 }
 
 static void tell_parent(enum info_type type)
@@ -153,22 +135,231 @@ 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)
+{
+       /* Locks were never acquired/reacquired? */
+       if (lock_owner == 0)
+               return;
+
+       /* We own them?  Release them all. */
+       if (lock_owner == getpid()) {
+               unsigned int i;
+               struct flock fl;
+               fl.l_type = F_UNLCK;
+               fl.l_whence = SEEK_SET;
+               fl.l_start = 0;
+               fl.l_len = 0;
+
+               for (i = 0; i < lock_num; i++)
+                       fcntl(locks[i].fd, F_SETLK, &fl);
+       } else {
+               /* Our parent must have them; pass request up. */
+               enum info_type type = RELEASE_LOCKS;
+               assert(control_fd != -1);
+               write_all(control_fd, &type, sizeof(type));
+       }
+       lock_owner = 0;
+}
+
+/* off_t is a signed type.  Getting its max is non-trivial. */
+static off_t off_max(void)
+{
+       BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
+       if (sizeof(off_t) == 4)
+               return (off_t)0x7FFFFFF;
+       else
+               return (off_t)0x7FFFFFFFFFFFFFFULL;
+}
+
+static void get_locks(void)
+{
+       unsigned int i;
+       struct flock fl;
+
+       if (lock_owner == getpid())
+               return;
+
+       if (lock_owner != 0) {
+               enum info_type type = RELEASE_LOCKS;
+               assert(control_fd != -1);
+               write_all(control_fd, &type, sizeof(type));
+       }
+
+       fl.l_whence = SEEK_SET;
+
+       for (i = 0; i < lock_num; i++) {
+               fl.l_type = locks[i].type;
+               fl.l_start = locks[i].start;
+               if (locks[i].end == off_max())
+                       fl.l_len = 0;
+               else
+                       fl.l_len = locks[i].end - locks[i].start + 1;
+
+               if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
+                       abort();
+       }
+       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;
+       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);
+               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;
+       }
+}
+
+/* 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)
@@ -178,22 +369,82 @@ static bool should_fail(struct failtest_call *call)
        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 && control_fd != -1)
+               failtest_cleanup(true, 0);
 
        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 '%c' got '%c'\n",
+                                    info_to_arg[call->type], *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();
+
        /* We're going to fail in the child. */
        call->fail = true;
        if (pipe(control) != 0 || pipe(output) != 0)
@@ -206,6 +457,22 @@ static bool should_fail(struct failtest_call *call)
                err(1, "forking failed");
 
        if (child == 0) {
+               if (tracefd != -1) {
+                       struct timeval diff;
+                       const char *p;
+
+                       diff = time_sub(time_now(), start);
+                       p = failpath_string();
+                       trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
+                             (int)diff.tv_sec, (int)diff.tv_usec / 10000, p);
+                       free((char *)p);
+                       p = strrchr(history[history_num-1].file, '/');
+                       if (p)
+                               trace("%s", p+1);
+                       else
+                               trace("%s", history[history_num-1].file);
+                       trace(":%u)\n", history[history_num-1].line);
+               }
                close(control[0]);
                close(output[0]);
                dup2(output[1], STDOUT_FILENO);
@@ -213,6 +480,8 @@ static bool should_fail(struct failtest_call *call)
                if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
                        close(output[1]);
                control_fd = control[1];
+               /* Valgrind spots the leak if we don't free these. */
+               free_files(files);
                return true;
        }
 
@@ -237,8 +506,13 @@ 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;
@@ -251,6 +525,9 @@ static bool should_fail(struct failtest_call *call)
                                if (type == WRITE) {
                                        if (!read_write_info(control[0]))
                                                break;
+                               } else if (type == RELEASE_LOCKS) {
+                                       release_locks();
+                                       /* FIXME: Tell them we're done... */
                                }
                        }
                } else if (pfd[0].revents & POLLHUP) {
@@ -261,9 +538,13 @@ 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);
@@ -277,11 +558,18 @@ static bool should_fail(struct failtest_call *call)
        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)
 {
@@ -296,11 +584,17 @@ 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;
@@ -313,11 +607,49 @@ void *failtest_malloc(size_t size, const char *file, unsigned line)
                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;
@@ -330,41 +662,68 @@ 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);
                p->u.realloc.ret = realloc(ptr, size);
+               set_cleanup(p, cleanup_realloc, struct realloc_call);
        }
        errno = p->error;
        return p->u.realloc.ret;
 }
 
-int failtest_open(const char *pathname, int flags,
+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, ...)
 {
        struct failtest_call *p;
        struct open_call call;
+       va_list ap;
 
        call.pathname = strdup(pathname);
-       call.flags = flags;
-       if (flags & O_CREAT) {
-               va_list ap;
-               va_start(ap, line);
-               call.mode = va_arg(ap, mode_t);
+       va_start(ap, line);
+       call.flags = va_arg(ap, int);
+       if (call.flags & O_CREAT) {
+               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 (!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, 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;
@@ -377,6 +736,8 @@ 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));
@@ -384,73 +745,48 @@ 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. */
@@ -459,124 +795,288 @@ 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;
 }
 
-/* We only trap this so we can dup fds in case we need to restore. */
-int failtest_close(int fd)
+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)
 {
        unsigned int i;
-       int newfd = -1;
+       struct lock_info *l;
+
+       for (i = 0; i < lock_num; i++) {
+               l = &locks[i];
 
-       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;
+               if (l->fd != fd)
+                       continue;
+               /* Four cases we care about:
+                * Start overlap:
+                *      l =    |      |
+                *      new = |   |
+                * Mid overlap:
+                *      l =    |      |
+                *      new =    |  |
+                * End overlap:
+                *      l =    |      |
+                *      new =      |    |
+                * Total overlap:
+                *      l =    |      |
+                *      new = |         |
+                */
+               if (start > l->start && end < l->end) {
+                       /* Mid overlap: trim entry, add new one. */
+                       off_t new_start, new_end;
+                       new_start = end + 1;
+                       new_end = l->end;
+                       l->end = start - 1;
+                       locks = add_lock(locks,
+                                        fd, new_start, new_end, l->type);
+                       l = &locks[i];
+               } else if (start <= l->start && end >= l->end) {
+                       /* Total overlap: eliminate entry. */
+                       l->end = 0;
+                       l->start = 1;
+               } else if (end >= l->start && end < l->end) {
+                       /* Start overlap: trim entry. */
+                       l->start = end + 1;
+               } else if (start > l->start && start <= l->end) {
+                       /* End overlap: trim entry. */
+                       l->end = start-1;
+               }
+               /* Nothing left?  Remove it. */
+               if (l->end < l->start) {
+                       memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
+                       i--;
                }
        }
 
-       for (i = 0; i < writes_num; i++) {
-               if (writes[i].hdr.fd == fd)
-                       writes[i].hdr.fd = newfd;
+       if (type != F_UNLCK) {
+               locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
+               l = &locks[lock_num++];
+               l->fd = fd;
+               l->start = start;
+               l->end = end;
+               l->type = type;
        }
-       return close(fd);
+       return locks;
 }
 
-void failtest_init(int argc, char *argv[])
+/* We trap this so we can record it: we don't fail it. */
+int failtest_close(int fd, const char *file, unsigned line)
 {
-       if (argc == 2
-           && strncmp(argv[1], "--failpath=", strlen("--failpath=")) == 0) {
-               failpath = argv[1] + strlen("--failpath=");
+       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);
+
+       /* 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);
 }
 
-/* Free up memory, so valgrind doesn't report leaks. */
-static void free_everything(void)
+/* Zero length means "to end of file" */
+static off_t end_of(off_t start, off_t len)
 {
-       unsigned int i;
+       if (len == 0)
+               return off_max();
+       return start + len - 1;
+}
+
+/* FIXME: This only handles locks, really. */
+int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
+{
+       struct failtest_call *p;
+       struct fcntl_call call;
+       va_list ap;
 
-       for (i = 0; i < writes_num; i++) {
-               free(writes[i].data);
-               if (writes[i].hdr.offset != (off_t)-1)
-                       free(writes[i].olddata);
+       call.fd = fd;
+       call.cmd = cmd;
+
+       /* Argument extraction. */
+       switch (cmd) {
+       case F_SETFL:
+       case F_SETFD:
+               va_start(ap, cmd);
+               call.arg.l = va_arg(ap, long);
+               va_end(ap);
+               return fcntl(fd, cmd, call.arg.l);
+       case F_GETFD:
+       case F_GETFL:
+               return fcntl(fd, cmd);
+       case F_GETLK:
+               get_locks();
+               va_start(ap, cmd);
+               call.arg.fl = *va_arg(ap, struct flock *);
+               va_end(ap);
+               return fcntl(fd, cmd, &call.arg.fl);
+       case F_SETLK:
+       case F_SETLKW:
+               va_start(ap, cmd);
+               call.arg.fl = *va_arg(ap, struct flock *);
+               va_end(ap);
+               break;
+       default:
+               /* This means you need to implement it here. */
+               err(1, "failtest: unknown fcntl %u", cmd);
        }
-       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);
+
+       p = add_history(FAILTEST_FCNTL, file, line, &call);
+
+       if (should_fail(p)) {
+               p->u.fcntl.ret = -1;
+               if (p->u.fcntl.cmd == F_SETLK)
+                       p->error = EAGAIN;
+               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)
+                       p->error = errno;
+               else {
+                       /* We don't handle anything else yet. */
+                       assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
+                       locks = add_lock(locks,
+                                        p->u.fcntl.fd,
+                                        p->u.fcntl.arg.fl.l_start,
+                                        end_of(p->u.fcntl.arg.fl.l_start,
+                                               p->u.fcntl.arg.fl.l_len),
+                                        p->u.fcntl.arg.fl.l_type);
+               }
        }
-       free(history);
+       errno = p->error;
+       return p->u.fcntl.ret;
 }
 
-void failtest_exit(int status)
+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;
 
-       if (control_fd == -1) {
-               free_everything();
-               exit(status);
+       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=");
+               }
        }
+       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))
                        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);
 }