From: Pavel Begunkov <asml.silence@xxxxxxxxx>
Commit 4bc6339a583cec650b05 ("block: move blk_stat_add() to
__blk_mq_end_request()") moved blk_stat_add() to reuse ktime_get_ns(),
so now it's called after blk_update_request(), which zeroes
rq->__data_len. Without length, blk_stat_add() can't calculate stat
bucket and returns error, effectively disabling hybrid polling.
v2: Hybrid polling needs pure io time for precision, but according to
the feedback from Omar Sandoval other components require end-to-end
time. So, it can't be reused, and should be sampled twice insted.
Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
---
block/blk-mq.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 32b8ad3d341b..907799282d57 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -537,11 +537,6 @@ inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
if (blk_mq_need_time_stamp(rq))
now = ktime_get_ns();
- if (rq->rq_flags & RQF_STATS) {
- blk_mq_poll_stats_start(rq->q);
- blk_stat_add(rq, now);
- }
-
if (rq->internal_tag != -1)
blk_mq_sched_completed_request(rq, now);
@@ -580,6 +575,11 @@ static void __blk_mq_complete_request(struct request *rq)
int cpu;
WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
+
+ if (rq->rq_flags & RQF_STATS) {
+ blk_mq_poll_stats_start(rq->q);
+ blk_stat_add(rq, ktime_get_ns());
+ }
/*
* Most of single queue controllers, there is only one irq vector
* for handling IO completion, and the only irq's affinity is set
--
2.21.0
