Re: [bug report] shared tags causes IO hang and performance drop

[John Garry <john.garry@huawei.com>]

Hi Ming,

> 
> It is reported inside RH that CPU utilization is increased ~20% when
> running simple FIO test inside VM which disk is built on image stored
> on XFS/megaraid_sas.
> 
> When I try to investigate by reproducing the issue via scsi_debug, I found
> IO hang when running randread IO(8k, direct IO, libaio) on scsi_debug disk
> created by the following command:
> 
> 	modprobe scsi_debug host_max_queue=128 submit_queues=$NR_CPUS virtual_gb=256
> 

So I can recreate this hang for using mq-deadline IO sched for scsi 
debug, in that fio does not exit. I'm using v5.12-rc7.

Do you have any idea of what changed to cause this, as we would have 
tested this before? Or maybe only none IO sched on scsi_debug. And 
normally 4k block size and only rw=read (for me, anyway).

Note that host_max_queue=128 will cap submit queue depth at 128, while 
would be 192 by default.

Will check more...including CPU utilization.

Thanks,
John

> Looks it is caused by SCHED_RESTART because current RESTART is just done
> on current hctx, and we may need to restart all hctxs for shared tags, and the
> issue can be fixed by the append patch. However, IOPS drops more than 10% with
> the patch.
> 
> So any idea for this issue and the original performance drop?
> 
> diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
> index e1e997af89a0..45188f7aa789 100644
> --- a/block/blk-mq-sched.c
> +++ b/block/blk-mq-sched.c
> @@ -59,10 +59,18 @@ EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
>   
>   void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
>   {
> +	bool shared_tag = blk_mq_is_sbitmap_shared(hctx->flags);
> +
> +	if (shared_tag)
> +		blk_mq_run_hw_queues(hctx->queue, true);
> +
>   	if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
>   		return;
>   	clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
>   
> +	if (shared_tag)
> +		return;
> +
>   	/*
>   	 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
>   	 * in blk_mq_run_hw_queue(). Its pair is the barrier in
> 
> Thanks,
> Ming
> 
> .
>