Re: [PATCH RFC 3/5] blk-mq: Facilitate a shared tags per tagset

[John Garry]

On 13/11/2019 14:06, Hannes Reinecke wrote:
> On 11/13/19 2:36 PM, John Garry wrote:
> > Some SCSI HBAs (such as HPSA, megaraid, mpt3sas, hisi_sas_v3 ..) support
> > multiple reply queues with single hostwide tags.
> >
> > In addition, these drivers want to use interrupt assignment in
> > pci_alloc_irq_vectors(PCI_IRQ_AFFINITY). However, as discussed in [0],
> > CPU hotplug may cause in-flight IO completion to not be serviced when an
> > interrupt is shutdown.
> >
> > To solve that problem, Ming's patchset to drain hctx's should ensure no
> > IOs are missed in-flight [1].
> >
> > However, to take advantage of that patchset, we need to map the HBA HW
> > queues to blk mq hctx's; to do that, we need to expose the HBA HW queues.
> >
> > In making that transition, the per-SCSI command request tags are no
> > longer unique per Scsi host - they are just unique per hctx. As such, the
> > HBA LLDD would have to generate this tag internally, which has a certain
> > performance overhead.
> >
> > However another problem is that blk mq assumes the host may accept
> > (Scsi_host.can_queue * #hw queue) commands. In [2], we removed the Scsi
> > host busy counter, which would stop the LLDD being sent more than
> > .can_queue commands; however, we should still ensure that the block layer
> > does not issue more than .can_queue commands to the Scsi host.
> >
> > To solve this problem, introduce a shared tags per blk_mq_tag_set, which
> > may be requested when allocating the tagset.
> >
> > New flag BLK_MQ_F_TAG_HCTX_SHARED should be set when requesting the
> > tagset.
> >
> > This is based on work originally from Ming Lei in [3].
> >
> > [0] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@xxxxxxxxxxxxxxxxxxxxxxx/
> > [1] https://lore.kernel.org/linux-block/20191014015043.25029-1-ming.lei@xxxxxxxxxx/
> > [2] https://lore.kernel.org/linux-scsi/20191025065855.6309-1-ming.lei@xxxxxxxxxx/
> > [3] https://lore.kernel.org/linux-block/20190531022801.10003-1-ming.lei@xxxxxxxxxx/
> >
> > Signed-off-by: John Garry <john.garry@xxxxxxxxxx>
> > ---
> >   block/blk-core.c       |  1 +
> >   block/blk-flush.c      |  2 +
> >   block/blk-mq-debugfs.c |  2 +-
> >   block/blk-mq-tag.c     | 85 ++++++++++++++++++++++++++++++++++++++++++
> >   block/blk-mq-tag.h     |  1 +
> >   block/blk-mq.c         | 61 +++++++++++++++++++++++++-----
> >   block/blk-mq.h         |  9 +++++
> >   include/linux/blk-mq.h |  3 ++
> >   include/linux/blkdev.h |  1 +
> >   9 files changed, 155 insertions(+), 10 deletions(-)
> [ .. ]
> > @@ -396,15 +398,17 @@ static struct request *blk_mq_get_request(struct request_queue *q,
> >   		blk_mq_tag_busy(data->hctx);
> >   	}
> >   
> > -	tag = blk_mq_get_tag(data);
> > -	if (tag == BLK_MQ_TAG_FAIL) {
> > -		if (clear_ctx_on_error)
> > -			data->ctx = NULL;
> > -		blk_queue_exit(q);
> > -		return NULL;
> > +	if (data->hctx->shared_tags) {
> > +		shared_tag = blk_mq_get_shared_tag(data);
> > +		if (shared_tag == BLK_MQ_TAG_FAIL)
> > +			goto err_shared_tag;
> >   	}
> >   
> > -	rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
> > +	tag = blk_mq_get_tag(data);
> > +	if (tag == BLK_MQ_TAG_FAIL)
> > +		goto err_tag;
> > +
> > +	rq = blk_mq_rq_ctx_init(data, tag, shared_tag, data->cmd_flags, alloc_time_ns);
> >   	if (!op_is_flush(data->cmd_flags)) {
> >   		rq->elv.icq = NULL;
> >   		if (e && e->type->ops.prepare_request) {

Hi Hannes,

> Why do you need to keep a parallel tag accounting between 'normal' and
> 'shared' tags here?
> Isn't is sufficient to get a shared tag only, and us that in lieo of the
> 'real' one?

In theory, yes. Just the 'shared' tag should be adequate.

A problem I see with this approach is that we lose the identity of which 
tags are allocated for each hctx. As an example for this, consider 
blk_mq_queue_tag_busy_iter(), which iterates the bits for each hctx. 
Now, if you're just using shared tags only, that wouldn't work.

Consider blk_mq_can_queue() as another example - this tells us if a hctx 
has any bits unset, while with only using shared tags it would tell if 
any bits unset over all queues, and this change in semantics could break 
things. At a glance, function __blk_mq_tag_idle() looks problematic also.

And this is where it becomes messy to implement.

> I would love to combine both,

Same here...

 as then we can easily do a reverse mapping
> by using the 'tag' value to lookup the command itself, and can possibly
> do the 'scsi_cmd_priv' trick of embedding the LLDD-specific parts within
> the command. With this split we'll be wasting quite some memory there,
> as the possible 'tag' values are actually nr_hw_queues * shared_tags.

Yeah, understood. That's just a trade-off I saw.

Thanks,
John