Add support for queuing > 1 command at the time
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 78dca9a09610e3140ae5eaabc7518012bff40abe..858d6b81e6c5df199e17ae33f09d802105b1011b 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -27,6 +27,7 @@
 #include <signal.h>
 #include <time.h>
 #include <locale.h>
+#include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
@@ -162,20 +163,14 @@ static struct fio_file *get_next_file(struct thread_data *td)
  */
 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) {
-               icd.nr = r;
-               ios_completed(td, &icd);
-       }
+       io_u_queued_complete(td, 0, NULL);
 
        /*
         * now cancel remaining active events
@@ -190,13 +185,8 @@ 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) {
-                       icd.nr = r;
-                       ios_completed(td, &icd);
-               }
-       }
+       if (td->cur_depth)
+               io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -206,7 +196,6 @@ 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)
@@ -220,24 +209,26 @@ 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) {
+       if (ret < 0) {
                td_verror(td, io_u->error);
                put_io_u(td, io_u);
                return 1;
-       }
-
-       ret = td_io_getevents(td, 1, td->cur_depth, NULL);
-       if (ret < 0) {
-               td_verror(td, ret);
-               return 1;
-       }
+       } else if (ret == FIO_Q_QUEUED) {
+               if (io_u_queued_complete(td, 1, NULL))
+                       return 1;
+       } else if (ret == FIO_Q_COMPLETED) {
+               if (io_u->error) {
+                       td_verror(td, io_u->error);
+                       return 1;
+               }
 
-       icd.nr = ret;
-       ios_completed(td, &icd);
-       if (icd.error) {
-               td_verror(td, icd.error);
-               return 1;
+               io_u_sync_complete(td, io_u, NULL);
+       } else if (ret == FIO_Q_BUSY) {
+               if (td_io_commit(td))
+                       return 1;
+               goto requeue;
        }
 
        return 0;
@@ -249,10 +240,9 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
  */
 static void do_verify(struct thread_data *td)
 {
-       struct io_u *io_u, *v_io_u = NULL;
-       struct io_completion_data icd;
        struct fio_file *f;
-       int ret, i;
+       struct io_u *io_u;
+       int ret, i, min_events;
 
        /*
         * sync io first and invalidate cache, to make sure we really
@@ -265,78 +255,72 @@ static void do_verify(struct thread_data *td)
 
        td_set_runstate(td, TD_VERIFYING);
 
-       do {
-               if (td->terminate)
-                       break;
-
+       io_u = NULL;
+       while (!td->terminate) {
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
-               if (runtime_exceeded(td, &io_u->start_time)) {
-                       put_io_u(td, io_u);
+               if (runtime_exceeded(td, &io_u->start_time))
                        break;
-               }
 
-               if (get_next_verify(td, io_u)) {
-                       put_io_u(td, io_u);
+               if (get_next_verify(td, io_u))
                        break;
-               }
 
-               f = get_next_file(td);
-               if (!f)
+               if (td_io_prep(td, io_u))
                        break;
 
-               io_u->file = f;
+requeue:
+               ret = td_io_queue(td, io_u);
 
-               if (td_io_prep(td, io_u)) {
-                       put_io_u(td, io_u);
-                       break;
-               }
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (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;
 
-               ret = td_io_queue(td, io_u);
-               if (ret) {
-                       td_verror(td, io_u->error);
-                       put_io_u(td, io_u);
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       ret = io_u_sync_complete(td, io_u, verify_io_u);
+                       if (ret)
+                               break;
+                       continue;
+               case FIO_Q_QUEUED:
                        break;
-               }
-
-               /*
-                * we have one pending to verify, do that while
-                * we are doing io on the next one
-                */
-               if (do_io_u_verify(td, &v_io_u))
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
                        break;
-
-               ret = td_io_getevents(td, 1, 1, NULL);
-               if (ret != 1) {
-                       if (ret < 0)
-                               td_verror(td, ret);
+               default:
+                       assert(ret < 0);
+                       td_verror(td, -ret);
                        break;
                }
 
-               v_io_u = td->io_ops->event(td, 0);
-               icd.nr = 1;
-               icd.error = 0;
-               fio_gettime(&icd.time, NULL);
-               io_completed(td, v_io_u, &icd);
-
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       put_io_u(td, v_io_u);
-                       v_io_u = NULL;
+               if (ret < 0 || td->error)
                        break;
-               }
 
                /*
-                * if we can't submit more io, we need to verify now
+                * if we can queue more, do so. but check if there are
+                * completed io_u's first.
                 */
-               if (queue_full(td) && do_io_u_verify(td, &v_io_u))
-                       break;
+               min_events = 0;
+               if (queue_full(td) || ret == FIO_Q_BUSY)
+                       min_events = 1;
 
-       } while (1);
+               /*
+                * 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))
+                       break;
+       }
 
-       do_io_u_verify(td, &v_io_u);
+       if (io_u)
+               put_io_u(td, io_u);
 
        if (td->cur_depth)
                cleanup_pending_aio(td);
@@ -375,7 +359,6 @@ 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;
@@ -384,7 +367,8 @@ static void do_io(struct thread_data *td)
        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;
 
@@ -400,74 +384,97 @@ static void do_io(struct thread_data *td)
                        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);
-               if (ret) {
-                       if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
-                           io_u->resid) {
-                               /*
-                                * short read/write. requeue.
-                                */
+
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (io_u->error) {
+                               ret = io_u->error;
+                               break;
+                       }
+                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
+
                                io_u->xfer_buflen = io_u->resid;
-                               io_u->xfer_buf += ret;
+                               io_u->xfer_buf += bytes;
                                goto requeue;
-                       } else {
-                               put_io_u(td, io_u);
-                               break;
                        }
+                       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:
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       put_io_u(td, io_u);
+                       break;
                }
 
-               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
+               if (ret < 0 || td->error)
+                       break;
 
-               if (td->cur_depth < td->iodepth) {
-                       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+               if (io_u)
+                       add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 
-                       timeout = &ts;
+               /*
+                * See if we need to complete some commands
+                */
+               if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
                        min_evts = 0;
-               } else {
-                       timeout = NULL;
-                       min_evts = 1;
+                       if (queue_full(td) || ret == FIO_Q_BUSY)
+                               min_evts = 1;
+
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
+                       if (bytes_done < 0)
+                               break;
                }
 
-               ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-               if (ret < 0) {
-                       td_verror(td, ret);
-                       break;
-               } else if (!ret)
+               if (!bytes_done)
                        continue;
 
-               icd.nr = ret;
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       break;
-               }
-
                /*
                 * 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;
 
                        b = td->io_blocks[0] + td->io_blocks[1];
-                       if (!(b % td->thinktime_blocks))
-                               usec_sleep(td, td->thinktime);
+                       if (!(b % td->thinktime_blocks)) {
+                               int left;
+
+                               if (td->thinktime_spin)
+                                       __usec_sleep(td->thinktime_spin);
+
+                               left = td->thinktime - td->thinktime_spin;
+                               if (left)
+                                       usec_sleep(td, left);
+                       }
                }
        }
 
@@ -640,6 +647,7 @@ 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);