Re: [PATCH v2 2/2] btrfs: fix compressed write bio attribution

[Dennis Zhou <dennis@kernel.org>]

On Tue, Dec 17, 2019 at 04:05:48PM +0100, David Sterba wrote:
> On Fri, Dec 13, 2019 at 02:21:49PM -0800, Dennis Zhou wrote:
> > On Fri, Dec 13, 2019 at 01:24:01PM +0100, David Sterba wrote:
> > > On Thu, Dec 12, 2019 at 10:19:34AM -0800, Dennis Zhou wrote:
> > > > From a0569aebde08e31e994c92d0b70befb84f7f5563 Mon Sep 17 00:00:00 2001
> > > > From: Dennis Zhou <dennis@kernel.org>
> > > > Date: Wed, 11 Dec 2019 15:20:15 -0800
> > > > 
> > > > Bio attribution is handled at bio_set_dev() as once we have a device, we
> > > > have a corresponding request_queue and then can derive the current css.
> > > > In special cases, we want to attribute to bio to someone else. This can
> > > > be done by calling bio_associate_blkg_from_css() or
> > > > kthread_associate_blkcg() depending on the scenario. Btrfs does this for
> > > > compressed writeback as they are handled by kworkers, so the latter can
> > > > be done here.
> > > > 
> > > > Commit 1a41802701ec ("btrfs: drop bio_set_dev where not needed") removes
> > > > early bio_set_dev() calls prior to submit_stripe_bio(). This breaks the
> > > > above assumption that we'll have a request_queue when we are doing
> > > > association. To fix this, switch to using kthread_associate_blkcg().
> > > 
> > > Can be kthread_associate_blkcg used also for submit_extent_page that
> > > calls bio_associate_blkg_from_css indirectly when initializing wbc?
> > > 
> > > 2996                 bio_set_dev(bio, bdev);
> > > 2997                 wbc_init_bio(wbc, bio);
> > > 2998                 wbc_account_cgroup_owner(wbc, page, page_size);
> > > 
> > > wbc_init_bio:
> > > 
> > > 	if (wbc)
> > > 		bio_associate_blkg_from_css();
> > 
> > Correct me if I'm wrong, but I don't think submit_extent_page() is only
> > called from kthread contexts. So, we wouldn't be able to rely on
> > kthread_associate_blkcg().
> 
> Yeah, the kthread is not guaranteed here.
> 
> > I can think about how to make wbc better for association in general, but
> > it's a percpu decrement and increment so it shouldn't really be much in
> > overhead.
> 
> Performance is not my concern here, the addition of bios and blkcg
> association is new and there were some integration bugs where I
> independently removed early bdev association while the blkg relied on
> that. I'm looking for ways to make it less error prone and the kthread
> association looks exactly like that so I was curious if it's possible to
> use it everywhere. If not, the bdev needs to be found from other
> available data.

Yeah. At the time, going through bio_set_dev() was a way to guarantee we
weren't missing an association with a blk-cgroup. This simplified
auditing and prevented newer use cases from missing it. However, I do
agree it's quite error prone.. I'll put it on my list and see if I can
come up with something better.

Thanks,
Dennis