Messages in this thread Patch in this message | | | Subject | Re: [PATCH v5 00/14] blk-mq: Reduce static requests memory footprint for shared sbitmap | From | John Garry <> | Date | Fri, 8 Oct 2021 11:17:35 +0100 |
| |
On 07/10/2021 21:31, Kashyap Desai wrote: > Perf top data indicates lock contention in "blk_mq_find_and_get_req" call. > > 1.31% 1.31% kworker/57:1H-k [kernel.vmlinux] > native_queued_spin_lock_slowpath > ret_from_fork > kthread > worker_thread > process_one_work > blk_mq_timeout_work > blk_mq_queue_tag_busy_iter > bt_iter > blk_mq_find_and_get_req > _raw_spin_lock_irqsave > native_queued_spin_lock_slowpath > > > Kernel v5.14 Data - > > %Node1 : 8.4 us, 31.2 sy, 0.0 ni, 43.7 id, 0.0 wa, 0.0 hi, 16.8 si, 0.0 > st > 4.46% [kernel] [k] complete_cmd_fusion > 3.69% [kernel] [k] megasas_build_and_issue_cmd_fusion > 2.97% [kernel] [k] blk_mq_find_and_get_req > 2.81% [kernel] [k] megasas_build_ldio_fusion > 2.62% [kernel] [k] syscall_return_via_sysret > 2.17% [kernel] [k] __entry_text_start > 2.01% [kernel] [k] io_submit_one > 1.87% [kernel] [k] scsi_queue_rq > 1.77% [kernel] [k] native_queued_spin_lock_slowpath > 1.76% [kernel] [k] scsi_complete > 1.66% [kernel] [k] llist_reverse_order > 1.63% [kernel] [k] _raw_spin_lock_irqsave > 1.61% [kernel] [k] llist_add_batch > 1.39% [kernel] [k] aio_complete_rw > 1.37% [kernel] [k] read_tsc > 1.07% [kernel] [k] blk_complete_reqs > 1.07% [kernel] [k] native_irq_return_iret > 1.04% [kernel] [k] __x86_indirect_thunk_rax > 1.03% fio [.] __fio_gettime > 1.00% [kernel] [k] flush_smp_call_function_queue > > > Test #2: Three VDs (each VD consist of 8 SAS SSDs). > (numactl -N 1 fio > 3vd.fio --rw=randread --bs=4k --iodepth=32 --numjobs=8 > --ioscheduler=none/mq-deadline) > > There is a performance regression but it is not due to this patch set. > Kernel v5.11 gives 2.1M IOPs on mq-deadline but 5.15 (without this patchset) > gives 1.8M IOPs. > In this test I did not noticed CPU issue as mentioned in Test-1. > > In general, I noticed host_busy is incorrect once I apply this patchset. It > should not be more than can_queue, but sysfs host_busy value is very high > when IOs are running. This issue is only after applying this patchset. > > Is this patch set only change the behavior of <shared_host_tag> enabled > driver ? Will there be any impact on mpi3mr driver ? I can test that as > well.
I can see where the high value of host_busy is coming from in this series - we incorrectly re-iter the tags by #hw queues times in blk_mq_tagset_busy_iter() - d'oh.
Please try the below patch. I have looked at other places where we may have similar problems in looping the hw queue count for tagset->tags[], and they look ok. But I will double-check. I think that blk_mq_queue_tag_busy_iter() should be fine - Ming?
--->8----
From e6ecaa6d624ebb903fa773ca2a2035300b4c55c5 Mon Sep 17 00:00:00 2001 From: John Garry <john.garry@huawei.com> Date: Fri, 8 Oct 2021 10:55:11 +0100 Subject: [PATCH] blk-mq: Fix blk_mq_tagset_busy_iter() for shared tags
Since it is now possible for a tagset to share a single set of tags, the iter function should not re-iter the tags for the count of hw queues in that case. Rather it should just iter once.
Signed-off-by: John Garry <john.garry@huawei.com>
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c index 72a2724a4eee..ef888aab81b3 100644 --- a/block/blk-mq-tag.c +++ b/block/blk-mq-tag.c @@ -378,9 +378,15 @@ void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn, void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, busy_tag_iter_fn *fn, void *priv) { + int nr_hw_queues; int i;
- for (i = 0; i < tagset->nr_hw_queues; i++) { + if (blk_mq_is_shared_tags(tagset->flags)) + nr_hw_queues = 1; + else + nr_hw_queues = tagset->nr_hw_queues; + + for (i = 0; i < nr_hw_queues; i++) { if (tagset->tags && tagset->tags[i]) __blk_mq_all_tag_iter(tagset->tags[i], fn, priv, BT_TAG_ITER_STARTED); ----8<---- Thanks, john
| |