cfq: improve fsync performance for small files
Corrado Zoccolo [Mon, 20 Sep 2010 13:24:50 +0000 (15:24 +0200)]
Fsync performance for small files achieved by cfq on high-end disks is
lower than what deadline can achieve, due to idling introduced between
the sync write happening in process context and the journal commit.

Moreover, when competing with a sequential reader, a process writing
small files and fsync-ing them is starved.

This patch fixes the two problems by:
- marking journal commits as WRITE_SYNC, so that they get the REQ_NOIDLE
  flag set,
- force all queues that have REQ_NOIDLE requests to be put in the noidle
  tree.

Having the queue associated to the fsync-ing process and the one associated
 to journal commits in the noidle tree allows:
- switching between them without idling,
- fairness vs. competing idling queues, since they will be serviced only
  after the noidle tree expires its slice.

Acked-by: Vivek Goyal <vgoyal@redhat.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Tested-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

block/cfq-iosched.c
fs/jbd/commit.c
fs/jbd2/commit.c

index b9f8619..6845926 100644 (file)
@@ -216,7 +216,6 @@ struct cfq_data {
        enum wl_type_t serving_type;
        unsigned long workload_expires;
        struct cfq_group *serving_group;
-       bool noidle_tree_requires_idle;
 
        /*
         * Each priority tree is sorted by next_request position.  These
@@ -2126,7 +2125,6 @@ static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
        slice = max_t(unsigned, slice, CFQ_MIN_TT);
        cfq_log(cfqd, "workload slice:%d", slice);
        cfqd->workload_expires = jiffies + slice;
-       cfqd->noidle_tree_requires_idle = false;
 }
 
 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
@@ -3108,7 +3106,9 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        if (cfqq->queued[0] + cfqq->queued[1] >= 4)
                cfq_mark_cfqq_deep(cfqq);
 
-       if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
+       if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
+               enable_idle = 0;
+       else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
            (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
                enable_idle = 0;
        else if (sample_valid(cic->ttime_samples)) {
@@ -3421,17 +3421,7 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
                        cfq_slice_expired(cfqd, 1);
                else if (sync && cfqq_empty &&
                         !cfq_close_cooperator(cfqd, cfqq)) {
-                       cfqd->noidle_tree_requires_idle |=
-                               !(rq->cmd_flags & REQ_NOIDLE);
-                       /*
-                        * Idling is enabled for SYNC_WORKLOAD.
-                        * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
-                        * only if we processed at least one !REQ_NOIDLE request
-                        */
-                       if (cfqd->serving_type == SYNC_WORKLOAD
-                           || cfqd->noidle_tree_requires_idle
-                           || cfqq->cfqg->nr_cfqq == 1)
-                               cfq_arm_slice_timer(cfqd);
+                       cfq_arm_slice_timer(cfqd);
                }
        }
 
index 95d8c11..3f030e9 100644 (file)
@@ -318,7 +318,7 @@ void journal_commit_transaction(journal_t *journal)
        int first_tag = 0;
        int tag_flag;
        int i;
-       int write_op = WRITE;
+       int write_op = WRITE_SYNC;
 
        /*
         * First job: lock down the current transaction and wait for
index 7c068c1..80910f5 100644 (file)
@@ -360,7 +360,7 @@ void jbd2_journal_commit_transaction(journal_t *journal)
        int tag_bytes = journal_tag_bytes(journal);
        struct buffer_head *cbh = NULL; /* For transactional checksums */
        __u32 crc32_sum = ~0;
-       int write_op = WRITE;
+       int write_op = WRITE_SYNC;
 
        /*
         * First job: lock down the current transaction and wait for