fio: fix aio trim completion latencies
[fio.git] / rate-submit.c
index a68a3dd3a7b26bf1e4a03c03e3583c28ebfc6faa..cf00d9bc75c56f057b41a337ae8bdc740aebaa66 100644 (file)
@@ -1,18 +1,66 @@
+/*
+ * Rated submission helpers
+ *
+ * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
+ *
+ */
 #include "fio.h"
-#include "ioengine.h"
+#include "ioengines.h"
 #include "lib/getrusage.h"
+#include "rate-submit.h"
 
-static void io_workqueue_fn(struct submit_worker *sw,
-                           struct workqueue_work *work)
+static void check_overlap(struct io_u *io_u)
+{
+       int i;
+       struct thread_data *td;
+       bool overlap = false;
+
+       do {
+               /*
+                * Allow only one thread to check for overlap at a
+                * time to prevent two threads from thinking the coast
+                * is clear and then submitting IOs that overlap with
+                * each other
+                *
+                * If an overlap is found, release the lock and
+                * re-acquire it before checking again to give other
+                * threads a chance to make progress
+                *
+                * If an overlap is not found, release the lock when the
+                * io_u's IO_U_F_FLIGHT flag is set so that this io_u
+                * can be checked by other threads as they assess overlap
+                */
+               pthread_mutex_lock(&overlap_check);
+               for_each_td(td, i) {
+                       if (td->runstate <= TD_SETTING_UP ||
+                               td->runstate >= TD_FINISHING ||
+                               !td->o.serialize_overlap ||
+                               td->o.io_submit_mode != IO_MODE_OFFLOAD)
+                               continue;
+
+                       overlap = in_flight_overlap(&td->io_u_all, io_u);
+                       if (overlap) {
+                               pthread_mutex_unlock(&overlap_check);
+                               break;
+                       }
+               }
+       } while (overlap);
+}
+
+static int io_workqueue_fn(struct submit_worker *sw,
+                          struct workqueue_work *work)
 {
        struct io_u *io_u = container_of(work, struct io_u, work);
        const enum fio_ddir ddir = io_u->ddir;
-       struct thread_data *td = sw->private;
-       int ret;
+       struct thread_data *td = sw->priv;
+       int ret, error;
+
+       if (td->o.serialize_overlap)
+               check_overlap(io_u);
 
        dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
 
-       io_u_set(io_u, IO_U_F_NO_FILE_PUT);
+       io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 
        td->cur_depth++;
 
@@ -23,12 +71,14 @@ static void io_workqueue_fn(struct submit_worker *sw,
                ret = io_u_queued_complete(td, 1);
                if (ret > 0)
                        td->cur_depth -= ret;
-               io_u_clear(io_u, IO_U_F_FLIGHT);
+               else if (ret < 0)
+                       break;
+               io_u_clear(td, io_u, IO_U_F_FLIGHT);
        } while (1);
 
        dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
 
-       io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
+       error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
 
        if (ret == FIO_Q_COMPLETED)
                td->cur_depth--;
@@ -43,17 +93,20 @@ static void io_workqueue_fn(struct submit_worker *sw,
                ret = io_u_queued_complete(td, min_evts);
                if (ret > 0)
                        td->cur_depth -= ret;
-       } else if (ret == FIO_Q_BUSY) {
-               ret = io_u_queued_complete(td, td->cur_depth);
-               if (ret > 0)
-                       td->cur_depth -= ret;
        }
+
+       if (error || td->error)
+               pthread_cond_signal(&td->parent->free_cond);
+
+       return 0;
 }
 
 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
 {
-       struct thread_data *td = sw->private;
+       struct thread_data *td = sw->priv;
 
+       if (td->error)
+               return false;
        if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
                return true;
 
@@ -62,7 +115,7 @@ static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
 
 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
 {
-       struct thread_data *td = sw->private;
+       struct thread_data *td = sw->priv;
        int ret;
 
        ret = io_u_quiesce(td);
@@ -75,21 +128,20 @@ static int io_workqueue_alloc_fn(struct submit_worker *sw)
        struct thread_data *td;
 
        td = calloc(1, sizeof(*td));
-       sw->private = td;
+       sw->priv = td;
        return 0;
 }
 
 static void io_workqueue_free_fn(struct submit_worker *sw)
 {
-       free(sw->private);
-       sw->private = NULL;
+       free(sw->priv);
+       sw->priv = NULL;
 }
 
 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
 {
        struct thread_data *parent = sw->wq->td;
-       struct thread_data *td = sw->private;
-       int fio_unused ret;
+       struct thread_data *td = sw->priv;
 
        memcpy(&td->o, &parent->o, sizeof(td->o));
        memcpy(&td->ts, &parent->ts, sizeof(td->ts));
@@ -101,28 +153,27 @@ static int io_workqueue_init_worker_fn(struct submit_worker *sw)
        if (ioengine_load(td))
                goto err;
 
-       if (td->o.odirect)
-               td->io_ops->flags |= FIO_RAWIO;
-
        td->pid = gettid();
 
        INIT_FLIST_HEAD(&td->io_log_list);
        INIT_FLIST_HEAD(&td->io_hist_list);
        INIT_FLIST_HEAD(&td->verify_list);
        INIT_FLIST_HEAD(&td->trim_list);
-       INIT_FLIST_HEAD(&td->next_rand_list);
        td->io_hist_tree = RB_ROOT;
 
        td->o.iodepth = 1;
        if (td_io_init(td))
                goto err_io_init;
 
-       fio_gettime(&td->epoch, NULL);
+       if (td->io_ops->post_init && td->io_ops->post_init(td))
+               goto err_io_init;
+
+       set_epoch_time(td, td->o.log_unix_epoch);
        fio_getrusage(&td->ru_start);
        clear_io_state(td, 1);
 
        td_set_runstate(td, TD_RUNNING);
-       td->flags |= TD_F_CHILD;
+       td->flags |= TD_F_CHILD | TD_F_NEED_LOCK;
        td->parent = parent;
        return 0;
 
@@ -136,7 +187,7 @@ err:
 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
                                        unsigned int *sum_cnt)
 {
-       struct thread_data *td = sw->private;
+       struct thread_data *td = sw->priv;
 
        (*sum_cnt)++;
        sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
@@ -204,7 +255,7 @@ static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 
 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
 {
-       struct thread_data *src = sw->private;
+       struct thread_data *src = sw->priv;
        struct thread_data *dst = sw->wq->td;
 
        if (td_read(src))
@@ -216,7 +267,7 @@ static void io_workqueue_update_acct_fn(struct submit_worker *sw)
 
 }
 
-struct workqueue_ops rated_wq_ops = {
+static struct workqueue_ops rated_wq_ops = {
        .fn                     = io_workqueue_fn,
        .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
        .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
@@ -226,3 +277,19 @@ struct workqueue_ops rated_wq_ops = {
        .init_worker_fn         = io_workqueue_init_worker_fn,
        .exit_worker_fn         = io_workqueue_exit_worker_fn,
 };
+
+int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
+{
+       if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
+               return 0;
+
+       return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
+}
+
+void rate_submit_exit(struct thread_data *td)
+{
+       if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
+               return;
+
+       workqueue_exit(&td->io_wq);
+}