Re: [PATCH 00/10] blk-mq: fix blk_mq_tags double free while nr_requests grown

[Yu Kuai]

Hi,

在 2025/08/15 16:30, Ming Lei 写道:
> On Fri, Aug 15, 2025 at 04:02:06PM +0800, Yu Kuai wrote:
> > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> >
> > In the case user trigger tags grow by queue sysfs attribute nr_requests,
> > hctx->sched_tags will be freed directly and replaced with a new
> > allocated tags, see blk_mq_tag_update_depth().
> >
> > The problem is that hctx->sched_tags is from elevator->et->tags, while
> > et->tags is still the freed tags, hence later elevator exist will try to
> > free the tags again, causing kernel panic.
> >
> > patch 1-6 are prep cleanup and refactor patches for updating nr_requests
> > patch 7,8 are the fix patches for the regression
> > patch 9 is cleanup patch after patch 8
> > patch 10 fix the stale nr_requests documentation
>
> Please do not mix bug(regression) fix with cleanup.
>
> The bug fix for updating nr_requests should have been simple enough in single
> or two patches, why do you make 10-patches for dealing with the regression?

Ok, in short, my solution is:

- serialize switching elevator with updating nr_requests
- check the case that nr_requests will grow and allocate elevator_tags
before freezing the queue.
- for the grow case, switch to new elevator_tags.

I do tried and I can't find a easy way to fix this without making
related code uncomfortable. Perhaps because I do the cleanups and
refactor first and I can't think outside the box...

> Not mention this way is really unfriendly for stable tree backport.

I checked the last time related code to queue_requests_store() was
changed is commit 3efe7571c3ae ("block: protect nr_requests update using
q->elevator_lock"), and I believe this is what the fixed patch relied
on, so I think backport will not have much conflicts.

Whatever stbale branch that f5a6604f7a44 ("block: fix lockdep warning
caused by lock dependency in elv_iosched_store") is backported, I can
make sure a proper fix is backported as well.

Thanks,
Kuai

> Thanks,
> Ming
>
>
> .