Date:   Thu, 7 Dec 2023 09:11:09 +0800
From:   Ming Lei <ming.lei@...hat.com>
To:     Yury Norov <yury.norov@...il.com>
Cc:     Thomas Gleixner <tglx@...utronix.de>,
        Andrew Morton <akpm@...ux-foundation.org>,
        linux-kernel@...r.kernel.org, Keith Busch <kbusch@...nel.org>,
        linux-nvme@...ts.infradead.org, linux-block@...r.kernel.org,
        Yi Zhang <yi.zhang@...hat.com>,
        Guangwu Zhang <guazhang@...hat.com>,
        Chengming Zhou <zhouchengming@...edance.com>,
        Jens Axboe <axboe@...nel.dk>
Subject: Re: [PATCH V4 resend] lib/group_cpus.c: avoid to acquire cpu hotplug
 lock in group_cpus_evenly

On Wed, Dec 06, 2023 at 04:41:44PM -0800, Yury Norov wrote:
> Hi Ming,
> 
> On Mon, Nov 20, 2023 at 04:35:59PM +0800, Ming Lei wrote:
> > group_cpus_evenly() could be part of storage driver's error handler,
> > such as nvme driver, when may happen during CPU hotplug, in which
> > storage queue has to drain its pending IOs because all CPUs associated
> > with the queue are offline and the queue is becoming inactive. And
> > handling IO needs error handler to provide forward progress.
> > 
> > Then dead lock is caused:
> > 
> > 1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's
> > handler is waiting for inflight IO
> > 
> > 2) error handler is waiting for CPU hotplug lock
> > 
> > 3) inflight IO can't be completed in blk-mq's CPU hotplug handler because
> > error handling can't provide forward progress.
> > 
> > Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(),
> > in which two stage spreads are taken: 1) the 1st stage is over all present
> > CPUs; 2) the end stage is over all other CPUs.
> > 
> > Turns out the two stage spread just needs consistent 'cpu_present_mask', and
> > remove the CPU hotplug lock by storing it into one local cache. This way
> > doesn't change correctness, because all CPUs are still covered.
> > 
> > Cc: Keith Busch <kbusch@...nel.org>
> > Cc: linux-nvme@...ts.infradead.org
> > Cc: linux-block@...r.kernel.org
> > Reported-by: Yi Zhang <yi.zhang@...hat.com>
> > Reported-by: Guangwu Zhang <guazhang@...hat.com>
> > Tested-by: Guangwu Zhang <guazhang@...hat.com>
> > Reviewed-by: Chengming Zhou <zhouchengming@...edance.com>
> > Reviewed-by: Jens Axboe <axboe@...nel.dk>
> > Signed-off-by: Ming Lei <ming.lei@...hat.com>
> > ---
> >  lib/group_cpus.c | 22 ++++++++++++++++------
> >  1 file changed, 16 insertions(+), 6 deletions(-)
> > 
> > diff --git a/lib/group_cpus.c b/lib/group_cpus.c
> > index aa3f6815bb12..ee272c4cefcc 100644
> > --- a/lib/group_cpus.c
> > +++ b/lib/group_cpus.c
> > @@ -366,13 +366,25 @@ struct cpumask *group_cpus_evenly(unsigned int numgrps)
> >  	if (!masks)
> >  		goto fail_node_to_cpumask;
> >  
> > -	/* Stabilize the cpumasks */
> > -	cpus_read_lock();
> >  	build_node_to_cpumask(node_to_cpumask);
> >  
> > +	/*
> > +	 * Make a local cache of 'cpu_present_mask', so the two stages
> > +	 * spread can observe consistent 'cpu_present_mask' without holding
> > +	 * cpu hotplug lock, then we can reduce deadlock risk with cpu
> > +	 * hotplug code.
> > +	 *
> > +	 * Here CPU hotplug may happen when reading `cpu_present_mask`, and
> > +	 * we can live with the case because it only affects that hotplug
> > +	 * CPU is handled in the 1st or 2nd stage, and either way is correct
> > +	 * from API user viewpoint since 2-stage spread is sort of
> > +	 * optimization.
> > +	 */
> > +	cpumask_copy(npresmsk, data_race(cpu_present_mask));
> 
> Now that you initialize the npresmsk explicitly, you can allocate it
> using alloc_cpumask_var().

Indeed, but this way is actually before this patch, and not related with
this fix.

> 
> The same actually holds for nmsk too, and even before this patch. Maybe
> fix it in a separate prepending patch?

Yeah, 'nmsk' is similar with 'npresmsk', and it is not fix, just one
optimization.

group_cpus_evenly() is only run in slow path, so this kind of
micro-optimization is not urgent and should be done in standalone
patch, and even we can live with it.

> 
> > +
> >  	/* grouping present CPUs first */
> >  	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
> > -				  cpu_present_mask, nmsk, masks);
> > +				  npresmsk, nmsk, masks);
> >  	if (ret < 0)
> >  		goto fail_build_affinity;
> >  	nr_present = ret;
> > @@ -387,15 +399,13 @@ struct cpumask *group_cpus_evenly(unsigned int numgrps)
> >  		curgrp = 0;
> >  	else
> >  		curgrp = nr_present;
> > -	cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
> > +	cpumask_andnot(npresmsk, cpu_possible_mask, npresmsk);
> >  	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
> >  				  npresmsk, nmsk, masks);
> 
> The first thing the helper does is checking if nprepmask is empty.
> cpumask_andnot() returns false in that case. So, assuming that present
> cpumask in the previous call can't be empty, we can save few cycles if
> drop corresponding check in the helper and do like this:
>         
> 	if (cpumask_andnot(npresmsk, cpu_possible_mask, npresmsk) == 0) {
>                 nr_others = 0;
>                 goto fail_build_affinity;
>         }
> 
>   	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
>   				  npresmsk, nmsk, masks);
> 
> Although, it's not related to this patch directly. So, if you fix
> zalloc_cpumask_var(), the patch looks good to me.

I'd rather not make things complicated, as mentioned this API is only
run in slow path.

> 
> Reviewed-by: Yury Norov <yury.norov@...il.com>

Thanks for the review!


Thanks, 
Ming

Powered by blists - more mailing lists