v3 syslet engine support
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index eb857afc690f4d794dfcbfa74f42b61d2986e0b9..9594f80582266ce8a8a5794ea9c52fb02a6adb1d 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -136,47 +136,22 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
-{
-       unsigned int old_next_file = td->next_file;
-       struct fio_file *f;
-
-       do {
-               f = &td->files[td->next_file];
-
-               td->next_file++;
-               if (td->next_file >= td->nr_files)
-                       td->next_file = 0;
-
-               if (f->fd != -1)
-                       break;
-
-               f = NULL;
-       } while (td->next_file != old_next_file);
-
-       return f;
-}
-
 /*
  * When job exits, we can cancel the in-flight IO if we are using async
  * io. Attempt to do so.
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
        struct list_head *entry, *n;
-       struct io_completion_data icd;
        struct io_u *io_u;
        int r;
 
        /*
         * get immediately available events, if any
         */
-       r = td_io_getevents(td, 0, td->cur_depth, &ts);
-       if (r > 0) {
-               init_icd(&icd, NULL, r);
-               ios_completed(td, &icd);
-       }
+       r = io_u_queued_complete(td, 0, NULL);
+       if (r < 0)
+               return;
 
        /*
         * now cancel remaining active events
@@ -191,13 +166,8 @@ static void cleanup_pending_aio(struct thread_data *td)
                }
        }
 
-       if (td->cur_depth) {
-               r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
-               if (r > 0) {
-                       init_icd(&icd, NULL, r);
-                       ios_completed(td, &icd);
-               }
-       }
+       if (td->cur_depth)
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -207,7 +177,6 @@ static void cleanup_pending_aio(struct thread_data *td)
 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
        struct io_u *io_u = __get_io_u(td);
-       struct io_completion_data icd;
        int ret;
 
        if (!io_u)
@@ -221,33 +190,27 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
                return 1;
        }
 
+requeue:
        ret = td_io_queue(td, io_u);
        if (ret < 0) {
                td_verror(td, io_u->error);
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
-               ret = td_io_getevents(td, 1, td->cur_depth, NULL);
-               if (ret < 0) {
-                       td_verror(td, ret);
+               if (io_u_queued_complete(td, 1, NULL) < 0)
                        return 1;
-               }
-
-               init_icd(&icd, NULL, ret);
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       return 1;
-               }
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
                        td_verror(td, io_u->error);
                        return 1;
                }
 
-               init_icd(&icd, NULL, 1);
-               io_completed(td, io_u, &icd);
-               put_io_u(td, io_u);
+               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;
+               goto requeue;
        }
 
        return 0;
@@ -268,37 +231,44 @@ 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;
        while (!td->terminate) {
-               struct io_completion_data icd;
-               struct timespec *timeout;
-
                io_u = __get_io_u(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);
 
                switch (ret) {
                case FIO_Q_COMPLETED:
                        if (io_u->error)
-                               ret = io_u->error;
+                               ret = -io_u->error;
                        if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
 
@@ -306,63 +276,45 @@ requeue:
                                io_u->xfer_buf += bytes;
                                goto requeue;
                        }
-                       init_icd(&icd, verify_io_u, 1);
-                       io_completed(td, io_u, &icd);
-                       if (icd.error) {
-                               ret = icd.error;
+                       ret = io_u_sync_complete(td, io_u, verify_io_u);
+                       if (ret < 0)
                                break;
-                       }
-                       put_io_u(td, io_u);
                        continue;
                case FIO_Q_QUEUED:
                        break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
                default:
                        assert(ret < 0);
-                       td_verror(td, ret);
+                       td_verror(td, -ret);
                        break;
                }
 
-               if (ret < 0)
+               if (ret < 0 || td->error)
                        break;
 
                /*
                 * if we can queue more, do so. but check if there are
                 * completed io_u's first.
                 */
-               if (queue_full(td)) {
-                       timeout = NULL;
+               min_events = 0;
+               if (queue_full(td) || ret == FIO_Q_BUSY) {
                        min_events = 1;
-               } else {
-                       struct timespec ts;
 
-                       ts.tv_sec = 0;
-                       ts.tv_nsec = 0;
-                       timeout = &ts;
-                       min_events = 0;
+                       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
                 */
-               ret = td_io_getevents(td, min_events, td->cur_depth, timeout);
-               if (ret < 0)
-                       break;
-               else if (!ret)
-                       continue;
-
-               init_icd(&icd, verify_io_u, ret);
-               ios_completed(td, &icd);
-
-               if (icd.error) {
-                       td_verror(td, icd.error);
+               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);
 
@@ -400,31 +352,31 @@ static void do_cpuio(struct thread_data *td)
  */
 static void do_io(struct thread_data *td)
 {
-       struct io_completion_data icd;
        struct timeval s;
        unsigned long usec;
-       struct fio_file *f;
        int i, ret = 0;
 
        td_set_runstate(td, TD_RUNNING);
 
        while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
-               struct timespec *timeout;
+               struct timeval comp_time;
+               long bytes_done = 0;
                int min_evts = 0;
                struct io_u *io_u;
 
                if (td->terminate)
                        break;
 
-               f = get_next_file(td);
-               if (!f)
-                       break;
-
-               io_u = get_io_u(td, f);
+               io_u = get_io_u(td);
                if (!io_u)
                        break;
 
                memcpy(&s, &io_u->start_time, sizeof(s));
+
+               if (runtime_exceeded(td, &s)) {
+                       put_io_u(td, io_u);
+                       break;
+               }
 requeue:
                ret = td_io_queue(td, io_u);
 
@@ -441,11 +393,23 @@ requeue:
                                io_u->xfer_buf += bytes;
                                goto requeue;
                        }
-                       init_icd(&icd, NULL, 1);
-                       io_completed(td, io_u, &icd);
-                       put_io_u(td, io_u);
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_sync_complete(td, io_u, NULL);
+                       if (bytes_done < 0)
+                               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);
+                       ret = td_io_commit(td);
                        break;
                default:
                        assert(ret < 0);
@@ -453,58 +417,46 @@ requeue:
                        break;
                }
 
-               if (ret < 0)
+               if (ret < 0 || td->error)
                        break;
 
-               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
-               if (ret == FIO_Q_QUEUED) {
-                       if (td->cur_depth < td->iodepth) {
-                               struct timespec ts;
-
-                               ts.tv_sec = 0;
-                               ts.tv_nsec = 0;
-                               timeout = &ts;
-                               min_evts = 0;
-                       } else {
-                               timeout = NULL;
+               /*
+                * 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) {
                                min_evts = 1;
-                       }
 
-                       ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-                       if (ret < 0) {
-                               td_verror(td, ret);
-                               break;
-                       } else if (!ret)
-                               continue;
+                               if (td->cur_depth > td->iodepth_low)
+                                       min_evts = td->cur_depth - td->iodepth_low;
+                       }
 
-                       init_icd(&icd, NULL, ret);
-                       ios_completed(td, &icd);
-                       if (icd.error) {
-                               td_verror(td, icd.error);
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
+                       if (bytes_done < 0)
                                break;
-                       }
                }
 
+               if (!bytes_done)
+                       continue;
+
                /*
                 * the rate is batched for now, it should work for batches
                 * of completions except the very first one which may look
                 * a little bursty
                 */
-               usec = utime_since(&s, &icd.time);
+               usec = utime_since(&s, &comp_time);
 
-               rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
+               rate_throttle(td, usec, bytes_done, td->ddir);
 
-               if (check_min_rate(td, &icd.time)) {
+               if (check_min_rate(td, &comp_time)) {
                        if (exitall_on_terminate)
                                terminate_threads(td->groupid, 0);
                        td_verror(td, ENODATA);
                        break;
                }
 
-               if (runtime_exceeded(td, &icd.time))
-                       break;
-
                if (td->thinktime) {
                        unsigned long long b;
 
@@ -523,6 +475,8 @@ requeue:
        }
 
        if (!td->error) {
+               struct fio_file *f;
+
                if (td->cur_depth)
                        cleanup_pending_aio(td);
 
@@ -661,11 +615,15 @@ 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;
 
+       td->last_was_sync = 0;
+
        for_each_file(td, f, i) {
+               f->last_completed_pos = 0;
+
                f->last_pos = 0;
                if (td->io_ops->flags & FIO_SYNCIO)
                        lseek(f->fd, SEEK_SET, 0);
@@ -691,6 +649,7 @@ static void *thread_main(void *data)
 
        INIT_LIST_HEAD(&td->io_u_freelist);
        INIT_LIST_HEAD(&td->io_u_busylist);
+       INIT_LIST_HEAD(&td->io_u_requeues);
        INIT_LIST_HEAD(&td->io_hist_list);
        INIT_LIST_HEAD(&td->io_log_list);
 
@@ -745,15 +704,15 @@ 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);
@@ -789,12 +748,12 @@ 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) {
@@ -812,7 +771,7 @@ err:
        close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return (void *) td->error;
+       return (void *) (unsigned long) td->error;
 }
 
 /*
@@ -835,7 +794,7 @@ 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;
 }
 
 /*