Hi Kashyap,
I am using <mq-deadline> which is MQ version of SQ deadline.
For now we can just consider case without scheduler.
I have some more findings. May be someone from upstream community can
connect the dots.
#1. hctx_may_queue() throttle the IO at hctx level. This is eventually per
sdev throttling for megaraid_sas driver because It creates only one
context - hctx0 for each scsi device.
If driver is using only one h/w queue, active_queues will be always steady.
In my test it was 64 thread, so active_queues=64.
So I figure that 64 threads comes from 64 having disks.
Even though <fio> thread is shared among allowed cpumask
(cpus_allowed_policy=shared option in fio), active_queues will be always 64
because we have only one h/w queue.
All the logical cpus are mapped to one h/w queue. It means, thread moving
from one cpu to another cpu will not change active_queues per hctx.
In case of this RFC, active_queues are now per hctx and there are multiple
hctx, but tags are shared.
Right, so we need a policy to divide up the shared tags across request
queues, based on principle of fairness.
This can create unwanted throttling and
eventually more lock contention in sbitmap.
I added below patch and things improved a bit, but not a full proof.
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 586c9d6..c708fbc 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -60,10 +60,12 @@ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
* For shared tag users, we track the number of currently active users
* and attempt to provide a fair share of the tag depth for each of them.
*/
-static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
+static inline bool hctx_may_queue(struct request_queue *q,
+ struct blk_mq_hw_ctx *hctx,
struct sbitmap_queue *bt)
{
- unsigned int depth, users;
+ unsigned int depth, users, i, outstanding = 0;
+ struct blk_mq_hw_ctx *hctx_iter;
if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
return true;
@@ -84,14 +86,18 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx
*hctx,
* Allow at least some tags
*/
depth = max((bt->sb.depth + users - 1) / users, 4U);
- return atomic_read(&hctx->nr_active) < depth;
+
+ queue_for_each_hw_ctx(q, hctx_iter, i)
+ outstanding += atomic_read(&hctx_iter->nr_active);
+
OK, I think that we need to find a cleaner way to do this.
+ return outstanding < depth;
}
static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
struct sbitmap_queue *bt)
{
if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
- !hctx_may_queue(data->hctx, bt))
#2 - In completion path, scsi module call blk_mq_run_hw_queues() upon IO
completion. I am not sure if it is good to run all the h/w queue or just
h/w queue of current reference is good enough.
Below patch helped to reduce contention in hcxt_lock().
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 610ee41..f72de2a 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -572,6 +572,7 @@ static bool scsi_end_request(struct request *req,
blk_status_t error,
struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
struct scsi_device *sdev = cmd->device;
struct request_queue *q = sdev->request_queue;
+ struct blk_mq_hw_ctx *mq_hctx = req->mq_hctx;
if (blk_update_request(req, error, bytes))
return true;
@@ -613,7 +614,7 @@ static bool scsi_end_request(struct request *req,
blk_status_t error,
!list_empty(&sdev->host->starved_list))
kblockd_schedule_work(&sdev->requeue_work);
else
- blk_mq_run_hw_queues(q, true);
+ blk_mq_run_hw_queue(mq_hctx, true);
Not sure on this. But, indeed, I found running all queues did add lots
of extra load for when enabling the deadline scheduler.
percpu_ref_put(&q->q_usage_counter);
return false;
#3 - __blk_mq_tag_idle() calls blk_mq_tag_wakeup_all which may not be
optimal for shared queue.
There is a penalty if we are calling __sbq_wake_up() frequently. In case of
nr_hw_queue = 1, things are better because one hctx and hctx->state will
avoid multiple calls.
If blk_mq_tag_wakeup_all is called from hctx0 context, there is no need to
call from hctx1, hctx2 etc.
I have added below patch in my testing.
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 586c9d6..5b331e5 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -53,7 +53,9 @@ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
atomic_dec(&tags->active_queues);
- blk_mq_tag_wakeup_all(tags, false);
+ /* TBD - Do this only for hctx->flags & BLK_MQ_F_TAG_HCTX_SHARED */
+ if (hctx->queue_num == 0)
+ blk_mq_tag_wakeup_all(tags, false);
ok, I see. But, again, I think we need to find a cleaner way to do this.
}
/*
With all above mentioned changes, I see performance improved from 2.2M IOPS
to 2.7M on same workload and profile.
But still short of nr_hw_queue = 1, which got 3.1M IOPS, as below, right?
Thanks,
John
Old Driver which has nr_hw_queue = 1 and I issue IOs from <fio> queue
depth = 128. We get 3.1M IOPS in this config. This eventually exhaust
host can_queue.
So I think I need to find a point where we start to get throttled.
Note - Very low contention in sbitmap_get()
- 23.58% 0.25% fio [kernel.vmlinux] [k]
blk_mq_make_request
- 23.33% blk_mq_make_request
- 21.68% blk_mq_get_request
- 20.19% blk_mq_get_tag
+ 10.08% prepare_to_wait_exclusive
+ 4.51% io_schedule
- 3.59% __sbitmap_queue_get
- 2.82% sbitmap_get
0.86% __sbitmap_get_word
0.75% _raw_spin_lock_irqsave
0.55% _raw_spin_unlock_irqrestore
Driver with RFC which has nr_hw_queue = N and I issue IOs from <fio>
queue depth = 128. We get 2.3 M IOPS in this config. This eventually
exhaust host can_queue.
Note - Very high contention in sbitmap_get()
- 42.39% 0.12% fio [kernel.vmlinux] [k]
generic_make_request
- 42.27% generic_make_request
- 41.00% blk_mq_make_request
- 38.28% blk_mq_get_request
- 33.76% blk_mq_get_tag
- 30.25% __sbitmap_queue_get
- 29.90% sbitmap_get
+ 9.06% _raw_spin_lock_irqsave
+ 7.94% _raw_spin_unlock_irqrestore
+ 3.86% __sbitmap_get_word
+ 1.78% call_function_single_interrupt
+ 0.67% ret_from_intr
+ 1.69% io_schedule
0.59% prepare_to_wait_exclusive
0.55% __blk_mq_get_tag
In this particular case, I observed alloc_hint = zeros which means,
sbitmap_get is not able to find free tags from hint. That may lead to
contention.
This condition is not happening with nr_hw_queue=1 (without RFC) driver.
alloc_hint=
{663, 2425, 3060, 54, 3149, 4319, 4175, 4867, 543, 2481, 0, 4779, 377,
***0***, 2010, 0, 909, 3350, 1546, 2179, 2875, 659, 3902, 2224, 3212,
836, 1892, 1669, 2420, 3415, 1904, 512, 3027, 4810, 2845, 4690, 712,
3105, 0, 0, 0, 3268, 4915, 3897, 1349, 547, 4, 733, 1765, 2068, 979,
51, 880, 0, 370, 3520, 2877, 4097, 418, 4501, 3717, 2893, 604, 508,
759, 3329, 4038, 4829, 715, 842, 1443, 556}
Driver with RFC which has nr_hw_queue = N and I issue IOs from <fio>
queue depth = 32. We get 3.1M IOPS in this config. This workload does
*not* exhaust host can_queue.
Please ensure .host_tagset is set for whenever nr_hw_queue = N. This is as
per
RFC, and I don't think you modified from the RFC for your test.
But I just wanted to mention that to be crystal clear.
Yes I have two separate driver copy. One with RFC change and another without
RFC.
- 5.07% 0.14% fio [kernel.vmlinux] [k]
generic_make_request
- 4.93% generic_make_request
- 3.61% blk_mq_make_request
- 2.04% blk_mq_get_request
- 1.08% blk_mq_get_tag
- 0.70% __sbitmap_queue_get
0.67% sbitmap_get
In summary, RFC has some performance bottleneck in sbitmap_get () if
outstanding per shost is about to exhaust. Without this RFC also
driver works in nr_hw_queue = 1, but that case is managed very well.
I am not sure why it happens only with shared host tag ? Theoretically
all the hctx is sharing the same bitmaptag which is same as
nr_hw_queue=1, so why contention is only visible in shared host tag
case.
Let me check this.
If you want to reproduce this issue, may be you have to reduce the
can_queue in hisi_sas driver.
Thanks,
John
