Fix out-of-band deref of (potentially) gone threads structure
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 6a72ce773820d0a19cad1497dd41e0a64ae31128..37a425e85ee6fde71ec7aa537b0f6947687cc128 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -49,6 +49,7 @@ int nr_process = 0;
 int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
+unsigned long done_secs = 0;
 
 static struct fio_mutex *startup_mutex;
 static volatile int fio_abort;
@@ -74,10 +75,12 @@ static void terminate_threads(int group_id)
        struct thread_data *td;
        int i;
 
+       dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
+
        for_each_td(td, i) {
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
-                       dprint(FD_PROCESS, "setting terminate on %d\n",td->pid);
-
+                       dprint(FD_PROCESS, "setting terminate on %s/%d\n",
+                                                       td->o.name, td->pid);
                        td->terminate = 1;
                        td->o.start_delay = 0;
 
@@ -98,17 +101,20 @@ static void terminate_threads(int group_id)
 
 static void sig_handler(int sig)
 {
+       if (!threads)
+               return;
+
        switch (sig) {
-               case SIGALRM:
-                       update_io_ticks();
-                       disk_util_timer_arm();
-                       print_thread_status();
-                       break;
-               default:
-                       printf("\nfio: terminating on signal %d\n", sig);
-                       fflush(stdout);
-                       terminate_threads(TERMINATE_ALL);
-                       break;
+       case SIGALRM:
+               update_io_ticks();
+               disk_util_timer_arm();
+               print_thread_status();
+               break;
+       default:
+               printf("\nfio: terminating on signal %d\n", sig);
+               fflush(stdout);
+               terminate_threads(TERMINATE_ALL);
+               break;
        }
 }
 
@@ -156,12 +162,16 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                         * check bandwidth specified rate
                         */
                        if (bytes < td->rate_bytes) {
-                               log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
+                               log_err("%s: min rate %u not met\n", td->o.name,
+                                                               td->o.ratemin);
                                return 1;
                        } else {
                                rate = (bytes - td->rate_bytes) / spent;
-                               if (rate < td->o.ratemin || bytes < td->rate_bytes) {
-                                       log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
+                               if (rate < td->o.ratemin ||
+                                   bytes < td->rate_bytes) {
+                                       log_err("%s: min rate %u not met, got"
+                                               " %luKiB/sec\n", td->o.name,
+                                                       td->o.ratemin, rate);
                                        return 1;
                                }
                        }
@@ -170,12 +180,17 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                         * checks iops specified rate
                         */
                        if (iops < td->o.rate_iops) {
-                               log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
+                               log_err("%s: min iops rate %u not met\n",
+                                               td->o.name, td->o.rate_iops);
                                return 1;
                        } else {
                                rate = (iops - td->rate_blocks) / spent;
-                               if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
-                                       log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
+                               if (rate < td->o.rate_iops_min ||
+                                   iops < td->rate_blocks) {
+                                       log_err("%s: min iops rate %u not met,"
+                                               " got %lu\n", td->o.name,
+                                                       td->o.rate_iops_min,
+                                                       rate);
                                }
                        }
                }
@@ -400,10 +415,10 @@ sync_done:
                 */
                min_events = 0;
                if (queue_full(td) || ret == FIO_Q_BUSY) {
-                       min_events = 1;
-
-                       if (td->cur_depth > td->o.iodepth_low)
+                       if (td->cur_depth >= td->o.iodepth_low)
                                min_events = td->cur_depth - td->o.iodepth_low;
+                       if (!min_events)
+                               min_events = 1;
                }
 
                /*
@@ -464,7 +479,7 @@ static void do_io(struct thread_data *td)
                 * Add verification end_io handler, if asked to verify
                 * a previously written file.
                 */
-               if (td->o.verify != VERIFY_NONE) {
+               if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
                        io_u->end_io = verify_io_u;
                        td_set_runstate(td, TD_VERIFYING);
                } else
@@ -533,15 +548,12 @@ sync_done:
                /*
                 * See if we need to complete some commands
                 */
-               if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
+               if (queue_full(td) || ret == FIO_Q_BUSY) {
                        min_evts = 0;
-                       if (queue_full(td) || ret == FIO_Q_BUSY) {
+                       if (td->cur_depth >= td->o.iodepth_low)
+                               min_evts = td->cur_depth - td->o.iodepth_low;
+                       if (!min_evts)
                                min_evts = 1;
-
-                               if (td->cur_depth > td->o.iodepth_low)
-                                       min_evts = td->cur_depth - td->o.iodepth_low;
-                       }
-
                        fio_gettime(&comp_time, NULL);
                        bytes_done = io_u_queued_complete(td, min_evts);
                        if (bytes_done < 0)
@@ -629,22 +641,6 @@ static void cleanup_io_u(struct thread_data *td)
        free_io_mem(td);
 }
 
-/*
- * "randomly" fill the buffer contents
- */
-static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
-{
-       long *ptr = io_u->buf;
-
-       if (!td->o.zero_buffers) {
-               while ((void *) ptr - io_u->buf < max_bs) {
-                       *ptr = rand() * GOLDEN_RATIO_PRIME;
-                       ptr++;
-               }
-       } else
-               memset(ptr, 0, max_bs);
-}
-
 static int init_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
@@ -654,10 +650,15 @@ static int init_io_u(struct thread_data *td)
 
        max_units = td->o.iodepth;
        max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
-       td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
+       td->orig_buffer_size = (unsigned long long) max_bs
+                                       * (unsigned long long) max_units;
 
-       if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
-               td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
+       if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
+               unsigned long bs;
+
+               bs = td->orig_buffer_size + td->o.hugepage_size - 1;
+               td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
+       }
 
        if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
                log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
@@ -682,8 +683,8 @@ static int init_io_u(struct thread_data *td)
                if (!(td->io_ops->flags & FIO_NOIO)) {
                        io_u->buf = p + max_bs * i;
 
-                       if (td_write(td))
-                               fill_io_buf(td, io_u, max_bs);
+                       if (td_write(td) && !td->o.refill_buffers)
+                               io_u_fill_buffer(td, io_u, max_bs);
                }
 
                io_u->index = i;
@@ -710,7 +711,8 @@ static int switch_ioscheduler(struct thread_data *td)
        f = fopen(tmp, "r+");
        if (!f) {
                if (errno == ENOENT) {
-                       log_err("fio: os or kernel doesn't support IO scheduler switching\n");
+                       log_err("fio: os or kernel doesn't support IO scheduler"
+                               " switching\n");
                        return 0;
                }
                td_verror(td, errno, "fopen iosched");
@@ -764,7 +766,8 @@ static int keep_running(struct thread_data *td)
                return 1;
        }
 
-       io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
+       io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
+                       + td->io_skip_bytes;
        if (io_done < td->o.size)
                return 1;
 
@@ -897,7 +900,8 @@ static void *thread_main(void *data)
                memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->o.ratemin)
-                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
+                       memcpy(&td->lastrate, &td->ts.stat_sample_time,
+                                                       sizeof(td->lastrate));
 
                if (clear_state && clear_io_state(td))
                        break;
@@ -910,7 +914,8 @@ static void *thread_main(void *data)
 
                if (td_read(td) && td->io_bytes[DDIR_READ]) {
                        if (td->rw_end_set[DDIR_READ])
-                               elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
+                               elapsed = utime_since(&td->start,
+                                                     &td->rw_end[DDIR_READ]);
                        else
                                elapsed = utime_since_now(&td->start);
 
@@ -918,13 +923,14 @@ static void *thread_main(void *data)
                }
                if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
                        if (td->rw_end_set[DDIR_WRITE])
-                               elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
+                               elapsed = utime_since(&td->start,
+                                                     &td->rw_end[DDIR_WRITE]);
                        else
                                elapsed = utime_since_now(&td->start);
 
                        runtime[DDIR_WRITE] += elapsed;
                }
-               
+
                if (td->error || td->terminate)
                        break;
 
@@ -969,7 +975,8 @@ static void *thread_main(void *data)
 
 err:
        if (td->error)
-               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
+               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
+                                                       td->verror);
        close_and_free_files(td);
        close_ioengine(td);
        cleanup_io_u(td);
@@ -1056,7 +1063,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                ret = waitpid(td->pid, &status, flags);
                if (ret < 0) {
                        if (errno == ECHILD) {
-                               log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
+                               log_err("fio: pid=%d disappeared %d\n", td->pid,
+                                                               td->runstate);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1066,7 +1074,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                                int sig = WTERMSIG(status);
 
                                if (sig != SIGQUIT)
-                                       log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+                                       log_err("fio: pid=%d, got signal=%d\n",
+                                                               td->pid, sig);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1085,23 +1094,16 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                pending++;
                continue;
 reaped:
-               if (td->o.use_thread) {
-                       long ret;
-
-                       dprint(FD_PROCESS, "joining tread %d\n", td->pid);
-                       if (pthread_join(td->thread, (void *) &ret)) {
-                               dprint(FD_PROCESS, "join failed %ld\n", ret);
-                               perror("pthread_join");
-                       }
-               }
-
                (*nr_running)--;
                (*m_rate) -= td->o.ratemin;
                (*t_rate) -= td->o.rate;
-               pending--;
+               if (!td->pid)
+                       pending--;
 
                if (td->error)
                        exit_value++;
+
+               done_secs += mtime_since_now(&td->epoch) / 1000;
        }
 
        if (*nr_running == cputhreads && !pending && realthreads)
@@ -1123,11 +1125,13 @@ static void run_threads(void)
        if (!terse_output) {
                printf("Starting ");
                if (nr_thread)
-                       printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
+                       printf("%d thread%s", nr_thread,
+                                               nr_thread > 1 ? "s" : "");
                if (nr_process) {
                        if (nr_thread)
                                printf(" and ");
-                       printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
+                       printf("%d process%s", nr_process,
+                                               nr_process > 1 ? "es" : "");
                }
                printf("\n");
                fflush(stdout);
@@ -1157,7 +1161,8 @@ static void run_threads(void)
                if (setup_files(td)) {
                        exit_value++;
                        if (td->error)
-                               log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
+                               log_err("fio: pid=%d, err=%d/%s\n", td->pid,
+                                                       td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                } else {
@@ -1171,7 +1176,7 @@ static void run_threads(void)
                         */
                        for_each_file(td, f, i)
                                td_io_close_file(td, f);
-               }
+               }
 
                init_disk_util(td);
        }
@@ -1206,8 +1211,11 @@ static void run_threads(void)
                                        continue;
                        }
 
-                       if (td->o.stonewall && (nr_started || nr_running))
+                       if (td->o.stonewall && (nr_started || nr_running)) {
+                               dprint(FD_PROCESS, "%s: stonewall wait\n",
+                                                       td->o.name);
                                break;
+                       }
 
                        /*
                         * Set state to created. Thread will transition
@@ -1219,18 +1227,24 @@ static void run_threads(void)
 
                        if (td->o.use_thread) {
                                dprint(FD_PROCESS, "will pthread_create\n");
-                               if (pthread_create(&td->thread, NULL, thread_main, td)) {
-                                       perror("thread_create");
+                               if (pthread_create(&td->thread, NULL,
+                                                  thread_main, td)) {
+                                       perror("pthread_create");
                                        nr_started--;
                                        break;
                                }
+                               if (pthread_detach(td->thread) < 0)
+                                       perror("pthread_detach");
                        } else {
+                               pid_t pid;
                                dprint(FD_PROCESS, "will fork\n");
-                               if (!fork()) {
+                               pid = fork();
+                               if (!pid) {
                                        int ret = fork_main(shm_id, i);
 
-                                       exit(ret);
-                               }
+                                       _exit(ret);
+                               } else if (i == fio_debug_jobno)
+                                       *fio_debug_jobp = pid;
                        }
                        fio_mutex_down(startup_mutex);
                }
@@ -1349,8 +1363,9 @@ int main(int argc, char *argv[])
        if (!fio_abort) {
                show_run_stats();
                if (write_bw_log) {
-                       __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
-                       __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
+                       __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
+                       __finish_log(agg_io_log[DDIR_WRITE],
+                                       "agg-write_bw.log");
                }
        }