Re: [PATCH 5/6] scsi: core: don't limit per-LUN queue depth for SSD when HBA needs

From: Ming Lei <ming.lei@redhat.com>
To: "Martin K. Petersen" <martin.petersen@oracle.com>
Cc: James Bottomley <James.Bottomley@hansenpartnership.com>,
	linux-scsi@vger.kernel.org, linux-block@vger.kernel.org,
	Jens Axboe <axboe@kernel.dk>,
	Sathya Prakash <sathya.prakash@broadcom.com>,
	Chaitra P B <chaitra.basappa@broadcom.com>,
	Suganath Prabu Subramani  <suganath-prabu.subramani@broadcom.com>,
	Kashyap Desai <kashyap.desai@broadcom.com>,
	Sumit Saxena <sumit.saxena@broadcom.com>,
	Shivasharan S <shivasharan.srikanteshwara@broadcom.com>,
	"Ewan D . Milne" <emilne@redhat.com>,
	Christoph Hellwig <hch@lst.de>, Hannes Reinecke <hare@suse.de>,
	Bart Van Assche <bart.vanassche@wdc.com>
Subject: Re: [PATCH 5/6] scsi: core: don't limit per-LUN queue depth for SSD when HBA needs
Date: Thu, 23 Jan 2020 10:54:29 +0800	[thread overview]
Message-ID: <20200123025429.GA5191@ming.t460p> (raw)
In-Reply-To: <yq1y2u1if7t.fsf@oracle.com>

Hi Martin,

On Mon, Jan 20, 2020 at 11:52:06PM -0500, Martin K. Petersen wrote:
> 
> Ming,
> 
> > NVMe doesn't have such per-request-queue(namespace) queue depth, so it
> > is reasonable to ignore the limit for SCSI SSD too.
> 
> It is really not. A free host tag does not guarantee that the target
> device can queue the command.

Yeah, I agree.

> 
> The assumption that SSDs are somehow special because they are "fast" is
> not valid. Given the common hardware queue depth for a SAS device of
> ~128 it is often trivial to drive a device into a congestion
> scenario. We see it all the time for non-rotational devices, SSDs and
> arrays alike. The SSD heuristic is simply not going to fly.

In reality, the channel number of SSD is very limited, it should be
dozens of in enterprise grade SSD, so the device itself can be
saturated by limited in-flight IOs.

However, it depends on if the target device returns the congestion
to host. From my observation, looks there isn't such feedback
from NVMe target.

If SSD target device doesn't provide such kind of congestion feedback,
tracking in-flight per-LUN IO via .device_busy doesn't make any
difference.

Even if there was such SSD target which provides such congestion
feedback, bypassing .device_busy won't cause big effect too since
blk-mq's SCHED_RESTART will retry this IO returning STS_RESOURCE
only after another in-flight one is completed.

> 
> Don't get me wrong, I am very sympathetic to obliterating device_busy in
> the hot path. I just don't think it is as easy as just ignoring the
> counter and hope for the best. Dynamic queue depth management is an
> integral part of the SCSI protocol, not something we can just decide to
> bypass because a device claims to be of a certain media type or speed.

At least, Broadcom guys tests this patch on megaraid raid and the results
shows that big improvement was got, that is why the flag is only set
on megaraid host.

> 
> I would prefer not to touch drivers that rely on cmd_per_lun / untagged
> operation and focus exclusively on the ones that use .track_queue_depth.
> For those we could consider an adaptive queue depth management scheme.
> Something like not maintaining device_busy until we actually get a
> QUEUE_FULL condition. And then rely on the existing queue depth ramp up
> heuristics to determine when to disable the busy counter again. Maybe
> with an additional watermark or time limit to avoid flip-flopping.
> 
> If that approach turns out to work, we should convert all remaining
> non-legacy drivers to .track_queue_depth so we only have two driver
> queuing flavors to worry about.

In theory, .track_queue_depth may only improve sequential IO's
performance for HDD., not very effective for SSD. Or just save a bit
CPU cycles in case of SSD.

I will study the related drivers/device a bit more wrt. track_queue_depth().

Thanks,
Ming