block, bfq: honor already-setup queue merges

The function bfq_setup_merge prepares the merging between two
bfq_queues, say bfqq and new_bfqq. To this goal, it assigns
bfqq->new_bfqq = new_bfqq. Then, each time some I/O for bfqq arrives,
the process that generated that I/O is disassociated from bfqq and
associated with new_bfqq (merging is actually a redirection). In this
respect, bfq_setup_merge increases new_bfqq->ref in advance, adding
the number of processes that are expected to be associated with
new_bfqq.

Unfortunately, the stable-merging mechanism interferes with this
setup. After bfqq->new_bfqq has been set by bfq_setup_merge, and
before all the expected processes have been associated with
bfqq->new_bfqq, bfqq may happen to be stably merged with a different
queue than the current bfqq->new_bfqq. In this case, bfqq->new_bfqq
gets changed. So, some of the processes that have been already
accounted for in the ref counter of the previous new_bfqq will not be
associated with that queue.  This creates an unbalance, because those
references will never be decremented.

This commit fixes this issue by reestablishing the previous, natural
behaviour: once bfqq->new_bfqq has been set, it will not be changed
until all expected redirections have occurred.

Signed-off-by: Davide Zini <davidezini2@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20210802141352.74353-2-paolo.valente@linaro.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index 480e1a13485966104362f265d842887640775781..dd13c2bbc29c1164b6d01432b28c13545cad8a90 100644 (file)
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -2662,6 +2662,15 @@ bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
         * are likely to increase the throughput.
         */
        bfqq->new_bfqq = new_bfqq;
+       /*
+        * The above assignment schedules the following redirections:
+        * each time some I/O for bfqq arrives, the process that
+        * generated that I/O is disassociated from bfqq and
+        * associated with new_bfqq. Here we increases new_bfqq->ref
+        * in advance, adding the number of processes that are
+        * expected to be associated with new_bfqq as they happen to
+        * issue I/O.
+        */
        new_bfqq->ref += process_refs;
        return new_bfqq;
 }
@@ -2724,6 +2733,10 @@ bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 {
        struct bfq_queue *in_service_bfqq, *new_bfqq;
 
+       /* if a merge has already been setup, then proceed with that first */
+       if (bfqq->new_bfqq)
+               return bfqq->new_bfqq;
+
        /*
         * Check delayed stable merge for rotational or non-queueing
         * devs. For this branch to be executed, bfqq must not be
@@ -2825,9 +2838,6 @@ bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
        if (bfq_too_late_for_merging(bfqq))
                return NULL;
 
-       if (bfqq->new_bfqq)
-               return bfqq->new_bfqq;
-
        if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
                return NULL;