Re: [PATCH V2 5/5] blk-mq: Wait for for hctx inflight requests on CPU unplug

[Ming Lei <ming.lei@redhat.com>]

On Thu, May 30, 2019 at 10:31:34AM +0100, John Garry wrote:
> Hi Ming,
> 
> > > 
> > > > Thinking of this issue further, so far, one doable solution is to
> > > > expose reply queues
> > > > as blk-mq hw queues, as done by the following patchset:
> > > > 
> > > > https://lore.kernel.org/linux-block/20180205152035.15016-1-ming.lei@redhat.com/
> > > 
> > > I thought that this patchset had fundamental issues, in terms of working for
> > > all types of hosts. FYI, I did the backport of latest hisi_sas_v3 to v4.15
> > 
> > Could you explain it a bit about the fundamental issues for all types of
> > host?
> > 
> 
> *As I understand*, splitting the tagset has issues with dual-mode HBAs - as
> in supporting NVMe and SCSI, as some HBAs do.

The patchset I mentioned doesn't split tagset. The patch just
converts SCSI's reply queue into blk_mq hw queue, and all hw queues
share the host-wide tags. You can get unique tag too.

This way isn't very different with the current single hw queue(tags),
that is why I think the performance shouldn't be bad compared with
the current single hw queue. Meantime, drivers can get simplified.

> 
> > It is just for hosts with multiple reply queues, such as hisi_sas v3,
> > megaraid_sas, mpt3sas and hpsa.
> > 
> > > with this patchset (as you may have noticed in my git send mistake), but we
> > > have not got to test it yet.
> > > 
> > > On a related topic, we did test exposing reply queues as blk-mq hw queues
> > > and generating the host-wide tag internally in the LLDD with sbitmap, and
> > > unfortunately we were experiencing a significant performance hit, like 2300K
> > > -> 1800K IOPs for 4K read.
> > > 
> > > We need to test this further. I don't understand why we get such a big hit.
> > 
> > The performance regression shouldn't have been introduced in theory, and it is
> > because blk_mq_queue_tag_busy_iter() iterates over the same duplicated tags multiple
> > times, which can be fixed easily.
> > 
> 
> We are testing further, and I will tentatively say that we're getting better
> results (than previously) after fixing something in the LLDD. TBC.
> 
> > > 
> > > > 
> > > > In which global host-wide tags are shared for all blk-mq hw queues.
> > > > 
> > > > Also we can remove all the reply_map stuff in drivers, then solve the problem of
> > > > draining in-flight requests during unplugging CPU in a generic approach.
> > > 
> > > So you're saying that removing this reply queue stuff can make the solution
> > > to the problem more generic, but do you have an idea of the overall
> > > solution?
> > 
> > 1) convert reply queue into blk-mq hw queue first
> > 
> > 2) then all drivers are in same position wrt. handling requests vs.
> > unplugging CPU (shutdown managed IRQ)
> > 
> > The current handling in blk_mq_hctx_notify_dead() is actually wrong,
> 
> Yeah, the comment reads that it's going away, but it's actually gone.
> 
> > at that time, all CPUs on the hctx are dead, blk_mq_run_hw_queue()
> > still dispatches requests on driver's hw queue, and driver is invisible
> > to DEAD CPUs mapped to this hctx, and finally interrupt for these
> > requests on the hctx are lost.
> > 
> > Frankly speaking, the above 2nd problem is still hard to solve.
> > 
> > 1) take_cpu_down() shutdown managed IRQ first, then run teardown callback
> > for states in [CPUHP_AP_ONLINE, CPUHP_AP_OFFLINE) on the to-be-offline
> > CPU
> > 
> > 2) However, all runnable tasks are removed from the CPU in the teardown
> > callback for CPUHP_AP_SCHED_STARTING, which is run after managed IRQs
> > are shutdown. That said it is hard to avoid new request queued to
> > the hctx with all DEAD CPUs.
> > 
> > 3) we don't support to freeze queue for specific hctx yet, or that way
> > may not be accepted because of extra cost in fast path
> > 
> > 4) once request is allocated, it should be submitted to driver no matter
> > if CPU hotplug happens or not. Or free it and re-allocate new request
> > on proper sw/hw queue?
> > 
> > > 
> > > > 
> > > > Last time, it was reported that the patchset causes performance regression,
> > > > which is actually caused by duplicated io accounting in
> > > > blk_mq_queue_tag_busy_iter(),
> > > > which should be fixed easily.
> > > > 
> > > > What do you think of this approach?
> > > 
> > > It would still be good to have a forward port of this patchset for testing,
> > > if we're serious about it. Or at least this bug you mention fixed.
> > 
> > I plan to make this patchset workable on 5.2-rc for your test first.
> > 
> 
> ok, thanks. I assume that we're still open to not adding support for global
> tags in blk-mq, but rather the LLDD generating the unique tag with sbitmap.

Actually it is global tags.

Thanks,
Ming