Improve submission latency calculation

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

diff --git a/engines/libaio.c b/engines/libaio.c
index 510ecab417b43bf369bb3062d2238435b03e1a4f..a659ba9bc141890b57ff0b29180eb446bf6dc17f 100644 (file)
--- a/engines/libaio.c
+++ b/engines/libaio.c
@@ -19,6 +19,7 @@ struct libaio_data {
        io_context_t aio_ctx;
        struct io_event *aio_events;
        struct iocb **iocbs;
+       struct io_u **io_us;
        int iocbs_nr;
 };
 
@@ -90,27 +91,49 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
        }
 
        ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
+       ld->io_us[ld->iocbs_nr] = io_u;
        ld->iocbs_nr++;
        return FIO_Q_QUEUED;
 }
 
+static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
+                             unsigned int nr)
+{
+       struct timeval now;
+       unsigned int i;
+
+       fio_gettime(&now, NULL);
+
+       for (i = 0; i < nr; i++) {
+               struct io_u *io_u = io_us[i];
+
+               memcpy(&io_u->issue_time, &now, sizeof(now));
+               io_u_queued(td, io_u);
+       }
+}
+
 static int fio_libaio_commit(struct thread_data *td)
 {
        struct libaio_data *ld = td->io_ops->data;
        struct iocb **iocbs;
+       struct io_u **io_us;
        int ret, iocbs_nr;
 
        if (!ld->iocbs_nr)
                return 0;
 
        iocbs_nr = ld->iocbs_nr;
+       io_us = ld->io_us;
        iocbs = ld->iocbs;
        do {
                ret = io_submit(ld->aio_ctx, iocbs_nr, iocbs);
                if (ret == iocbs_nr) {
+                       fio_libaio_queued(td, io_us, ret);
                        ret = 0;
                        break;
                } else if (ret > 0) {
+                       fio_libaio_queued(td, io_us, ret);
+                       io_us += ret;
                        iocbs += ret;
                        iocbs_nr -= ret;
                        continue;
@@ -141,11 +164,9 @@ static void fio_libaio_cleanup(struct thread_data *td)
 
        if (ld) {
                io_destroy(ld->aio_ctx);
-               if (ld->aio_events)
-                       free(ld->aio_events);
-               if (ld->iocbs)
-                       free(ld->iocbs);
-
+               free(ld->aio_events);
+               free(ld->iocbs);
+               free(ld->io_us);
                free(ld);
                td->io_ops->data = NULL;
        }
@@ -166,6 +187,8 @@ static int fio_libaio_init(struct thread_data *td)
        memset(ld->aio_events, 0, td->iodepth * sizeof(struct io_event));
        ld->iocbs = malloc(td->iodepth * sizeof(struct iocb *));
        memset(ld->iocbs, 0, sizeof(struct iocb *));
+       ld->io_us = malloc(td->iodepth * sizeof(struct io_u *));
+       memset(ld->io_us, 0, td->iodepth * sizeof(struct io_u *));
        ld->iocbs_nr = 0;
 
        td->io_ops->data = ld;

diff --git a/fio.c b/fio.c
index 6176b77a22d2cf54147e03da51684bdb431d2d30..4596544429c9212e6cabc60b15ff9f09f7ef6291 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -385,6 +385,13 @@ 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);
@@ -399,9 +406,6 @@ 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
                 */

diff --git a/fio.h b/fio.h
index 12afa0651ae6e6800d7d316d09735a13178d88bf..259166fca1b8e842a2c001bc1bdaabfb54cb9be3 100644 (file)
--- a/fio.h
+++ b/fio.h
@@ -614,6 +614,7 @@ extern void put_io_u(struct thread_data *, struct io_u *);
 extern void requeue_io_u(struct thread_data *, struct io_u **);
 extern long io_u_sync_complete(struct thread_data *, struct io_u *, endio_handler *);
 extern long io_u_queued_complete(struct thread_data *, int, endio_handler *);
+extern void io_u_queued(struct thread_data *, struct io_u *);
 
 /*
  * io engine entry points

diff --git a/io_u.c b/io_u.c
index ab46cbe7aadb163578517578a7b41b962c6711c4..6234c42b3a91b32305110f96e9e216472a9f537c 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -555,3 +555,14 @@ long io_u_queued_complete(struct thread_data *td, int min_events,
 
        return -1;
 }
+
+/*
+ * Call when io_u is really queued, to update the submission latency.
+ */
+void io_u_queued(struct thread_data *td, struct io_u *io_u)
+{
+       unsigned long slat_time;
+
+       slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
+       add_slat_sample(td, io_u->ddir, slat_time);
+}

diff --git a/ioengines.c b/ioengines.c
index 16ea928fe727c094f651e75922f52308eb36c70a..2a11ed322ecd49fba5751cf958397da09097c4af 100644 (file)
--- a/ioengines.c
+++ b/ioengines.c
@@ -187,12 +187,15 @@ int td_io_getevents(struct thread_data *td, int min, int max,
 
 int td_io_queue(struct thread_data *td, struct io_u *io_u)
 {
-       fio_gettime(&io_u->issue_time, NULL);
+       int ret;
+
 
        if (io_u->ddir != DDIR_SYNC)
                td->io_issues[io_u->ddir]++;
 
-       return td->io_ops->queue(td, io_u);
+       ret = td->io_ops->queue(td, io_u);
+       fio_gettime(&io_u->issue_time, NULL);
+       return ret;
 }
 
 int td_io_init(struct thread_data *td)