Merge branch 'doc-patches' of https://github.com/vincentkfu/fio

[fio.git] / engines / libaio.c

diff --git a/engines/libaio.c b/engines/libaio.c
index bcc296ffdbc2c1046e9d91774381c34748b4d1ea..e0d7cbbafb780551f35c62c83bb9dd3b26b4b164 100644 (file)
--- a/engines/libaio.c
+++ b/engines/libaio.c
@@ -12,6 +12,10 @@
 #include <libaio.h>
 
 #include "../fio.h"
+#include "../lib/pow2.h"
+#include "../optgroup.h"
+
+static int fio_libaio_commit(struct thread_data *td);
 
 struct libaio_data {
        io_context_t aio_ctx;
@@ -35,7 +39,7 @@ struct libaio_data {
 };
 
 struct libaio_options {
-       struct thread_data *td;
+       void *pad;
        unsigned int userspace_reap;
 };
 
@@ -79,7 +83,7 @@ static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 
 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct io_event *ev;
        struct io_u *io_u;
 
@@ -139,13 +143,19 @@ static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
 }
 
 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
-                               unsigned int max, struct timespec *t)
+                               unsigned int max, const struct timespec *t)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct libaio_options *o = td->eo;
-       unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
+       unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
+       struct timespec __lt, *lt = NULL;
        int r, events = 0;
 
+       if (t) {
+               __lt = *t;
+               lt = &__lt;
+       }
+
        do {
                if (o->userspace_reap == 1
                    && actual_min == 0
@@ -155,12 +165,15 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
                                ld->aio_events + events);
                } else {
                        r = io_getevents(ld->aio_ctx, actual_min,
-                               max, ld->aio_events + events, t);
+                               max, ld->aio_events + events, lt);
                }
-               if (r >= 0)
+               if (r > 0)
                        events += r;
-               else if (r == -EAGAIN)
+               else if ((min && r == 0) || r == -EAGAIN) {
+                       fio_libaio_commit(td);
                        usleep(100);
+               } else if (r != -EINTR)
+                       break;
        } while (events < min);
 
        return r < 0 ? r : events;
@@ -168,7 +181,7 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
 
 static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        fio_ro_check(td, io_u);
 
@@ -207,7 +220,7 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
                              unsigned int nr)
 {
-       struct timeval now;
+       struct timespec now;
        unsigned int i;
 
        if (!fio_fill_issue_time(td))
@@ -225,10 +238,11 @@ static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
 
 static int fio_libaio_commit(struct thread_data *td)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct iocb **iocbs;
        struct io_u **io_us;
-       int ret;
+       struct timespec ts;
+       int ret, wait_start = 0;
 
        if (!ld->queued)
                return 0;
@@ -248,38 +262,70 @@ static int fio_libaio_commit(struct thread_data *td)
                        ld->queued -= ret;
                        ring_inc(ld, &ld->tail, ret);
                        ret = 0;
+                       wait_start = 0;
                } else if (ret == -EINTR || !ret) {
                        if (!ret)
                                io_u_mark_submit(td, ret);
+                       wait_start = 0;
                        continue;
                } else if (ret == -EAGAIN) {
                        /*
                         * If we get EAGAIN, we should break out without
                         * error and let the upper layer reap some
-                        * events for us.
+                        * events for us. If we have no queued IO, we
+                        * must loop here. If we loop for more than 30s,
+                        * just error out, something must be buggy in the
+                        * IO path.
                         */
-                       ret = 0;
+                       if (ld->queued) {
+                               ret = 0;
+                               break;
+                       }
+                       if (!wait_start) {
+                               fio_gettime(&ts, NULL);
+                               wait_start = 1;
+                       } else if (mtime_since_now(&ts) > 30000) {
+                               log_err("fio: aio appears to be stalled, giving up\n");
+                               break;
+                       }
+                       usleep(1);
+                       continue;
+               } else if (ret == -ENOMEM) {
+                       /*
+                        * If we get -ENOMEM, reap events if we can. If
+                        * we cannot, treat it as a fatal event since there's
+                        * nothing we can do about it.
+                        */
+                       if (ld->queued)
+                               ret = 0;
                        break;
                } else
                        break;
-       } while (ld->head != ld->tail);
+       } while (ld->queued);
 
        return ret;
 }
 
 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
 }
 
 static void fio_libaio_cleanup(struct thread_data *td)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        if (ld) {
-               io_destroy(ld->aio_ctx);
+               /*
+                * Work-around to avoid huge RCU stalls at exit time. If we
+                * don't do this here, then it'll be torn down by exit_aio().
+                * But for that case we can parallellize the freeing, thus
+                * speeding it up a lot.
+                */
+               if (!(td->flags & TD_F_CHILD))
+                       io_destroy(ld->aio_ctx);
                free(ld->aio_events);
                free(ld->iocbs);
                free(ld->io_us);
@@ -317,7 +363,7 @@ static int fio_libaio_init(struct thread_data *td)
        ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
        ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
 
-       td->io_ops->data = ld;
+       td->io_ops_data = ld;
        return 0;
 }