Track io_u state (free or in-flight)
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 6176b77a22d2cf54147e03da51684bdb431d2d30..0ccd5d857e0002abac481e72e5934fcde1d34f74 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -149,7 +149,9 @@ static void cleanup_pending_aio(struct thread_data *td)
        /*
         * get immediately available events, if any
         */
-       io_u_queued_complete(td, 0, NULL);
+       r = io_u_queued_complete(td, 0, NULL);
+       if (r < 0)
+               return;
 
        /*
         * now cancel remaining active events
@@ -158,14 +160,24 @@ static void cleanup_pending_aio(struct thread_data *td)
                list_for_each_safe(entry, n, &td->io_u_busylist) {
                        io_u = list_entry(entry, struct io_u, list);
 
-                       r = td->io_ops->cancel(td, io_u);
-                       if (!r)
+                       /*
+                        * if the io_u isn't in flight, then that generally
+                        * means someone leaked an io_u. complain but fix
+                        * it up, so we don't stall here.
+                        */
+                       if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+                               log_err("fio: non-busy IO on busy list\n");
                                put_io_u(td, io_u);
+                       } else {
+                               r = td->io_ops->cancel(td, io_u);
+                               if (!r)
+                                       put_io_u(td, io_u);
+                       }
                }
        }
 
        if (td->cur_depth)
-               io_u_queued_complete(td, td->cur_depth, NULL);
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -195,7 +207,7 @@ requeue:
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
-               if (io_u_queued_complete(td, 1, NULL))
+               if (io_u_queued_complete(td, 1, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
@@ -203,7 +215,8 @@ requeue:
                        return 1;
                }
 
-               io_u_sync_complete(td, io_u, NULL);
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
+                       return 1;
        } else if (ret == FIO_Q_BUSY) {
                if (td_io_commit(td))
                        return 1;
@@ -228,10 +241,15 @@ static void do_verify(struct thread_data *td)
         * read from disk.
         */
        for_each_file(td, f, i) {
-               fio_io_sync(td, f);
-               file_invalidate_cache(td, f);
+               if (fio_io_sync(td, f))
+                       break;
+               if (file_invalidate_cache(td, f))
+                       break;
        }
 
+       if (td->error)
+               return;
+
        td_set_runstate(td, TD_VERIFYING);
 
        io_u = NULL;
@@ -240,15 +258,20 @@ static void do_verify(struct thread_data *td)
                if (!io_u)
                        break;
 
-               if (runtime_exceeded(td, &io_u->start_time))
+               if (runtime_exceeded(td, &io_u->start_time)) {
+                       put_io_u(td, io_u);
                        break;
+               }
 
-               if (get_next_verify(td, io_u))
+               if (get_next_verify(td, io_u)) {
+                       put_io_u(td, io_u);
                        break;
+               }
 
-               if (td_io_prep(td, io_u))
+               if (td_io_prep(td, io_u)) {
+                       put_io_u(td, io_u);
                        break;
-
+               }
 requeue:
                ret = td_io_queue(td, io_u);
 
@@ -264,7 +287,7 @@ requeue:
                                goto requeue;
                        }
                        ret = io_u_sync_complete(td, io_u, verify_io_u);
-                       if (ret)
+                       if (ret < 0)
                                break;
                        continue;
                case FIO_Q_QUEUED:
@@ -287,20 +310,21 @@ requeue:
                 * completed io_u's first.
                 */
                min_events = 0;
-               if (queue_full(td) || ret == FIO_Q_BUSY)
+               if (queue_full(td) || ret == FIO_Q_BUSY) {
                        min_events = 1;
 
+                       if (td->cur_depth > td->iodepth_low)
+                               min_events = td->cur_depth - td->iodepth_low;
+               }
+
                /*
                 * Reap required number of io units, if any, and do the
                 * verification on them through the callback handler
                 */
-               if (io_u_queued_complete(td, min_events, verify_io_u))
+               if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
                        break;
        }
 
-       if (io_u)
-               put_io_u(td, io_u);
-
        if (td->cur_depth)
                cleanup_pending_aio(td);
 
@@ -385,6 +409,13 @@ requeue:
                                ret = bytes_done;
                        break;
                case FIO_Q_QUEUED:
+                       /*
+                        * if the engine doesn't have a commit hook,
+                        * the io_u is really queued. if it does have such
+                        * a hook, it has to call io_u_queued() itself.
+                        */
+                       if (td->io_ops->commit == NULL)
+                               io_u_queued(td, io_u);
                        break;
                case FIO_Q_BUSY:
                        requeue_io_u(td, &io_u);
@@ -399,17 +430,18 @@ requeue:
                if (ret < 0 || td->error)
                        break;
 
-               if (io_u)
-                       add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
                /*
                 * See if we need to complete some commands
                 */
                if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
                        min_evts = 0;
-                       if (queue_full(td) || ret == FIO_Q_BUSY)
+                       if (queue_full(td) || ret == FIO_Q_BUSY) {
                                min_evts = 1;
 
+                               if (td->cur_depth > td->iodepth_low)
+                                       min_evts = td->cur_depth - td->iodepth_low;
+                       }
+
                        fio_gettime(&comp_time, NULL);
                        bytes_done = io_u_queued_complete(td, min_evts, NULL);
                        if (bytes_done < 0)
@@ -531,6 +563,7 @@ static int init_io_u(struct thread_data *td)
                        fill_rand_buf(io_u, max_bs);
 
                io_u->index = i;
+               io_u->flags = IO_U_F_FREE;
                list_add(&io_u->list, &td->io_u_freelist);
        }
 
@@ -593,7 +626,7 @@ static void clear_io_state(struct thread_data *td)
        struct fio_file *f;
        int i;
 
-       td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+       td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
 
@@ -682,15 +715,15 @@ static void *thread_main(void *data)
        }
 
        fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ru_start);
+       getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
        runtime[0] = runtime[1] = 0;
        while (td->loops--) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
-                       memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
                clear_io_state(td);
                prune_io_piece_log(td);
@@ -726,12 +759,12 @@ static void *thread_main(void *data)
        td->runtime[0] = runtime[0] / 1000;
        td->runtime[1] = runtime[1] / 1000;
 
-       if (td->bw_log)
-               finish_log(td, td->bw_log, "bw");
-       if (td->slat_log)
-               finish_log(td, td->slat_log, "slat");
-       if (td->clat_log)
-               finish_log(td, td->clat_log, "clat");
+       if (td->ts.bw_log)
+               finish_log(td, td->ts.bw_log, "bw");
+       if (td->ts.slat_log)
+               finish_log(td, td->ts.slat_log, "slat");
+       if (td->ts.clat_log)
+               finish_log(td, td->ts.clat_log, "clat");
        if (td->write_iolog_file)
                write_iolog_close(td);
        if (td->exec_postrun) {
@@ -749,7 +782,7 @@ err:
        close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return (void *) td->error;
+       return (void *) (unsigned long) td->error;
 }
 
 /*
@@ -772,7 +805,7 @@ static int fork_main(int shmid, int offset)
        td = data + offset * sizeof(struct thread_data);
        ret = thread_main(td);
        shmdt(data);
-       return (int) ret;
+       return (int) (unsigned long) ret;
 }
 
 /*