On Wed 01-07-15 21:10:56, Tejun Heo wrote:

Hello, Jan.

On Tue, Jun 30, 2015 at 11:37:51AM +0200, Jan Kara wrote:

quoted

Hum, you later changed this to use a per-sb flag instead of a per-fs-type
flag, right? We could do it as well here but OK.

The commits were already in stable branch at that point and landed in
mainline during this merge window, so I'm afraid the review points
will have to be addressed as additional patches.

Yeah, I know but I just didn't get to the series earlier. Anyway, I didn't
find fundamental issues so it's easy to change things in followup patches.

quoted

One more question - what does prevent us from supporting CGROUP_WRITEBACK
for all bdis capable of writeback? I guess the reason is that currently
blkcgs are bound to request_queue and we have to have blkcg(s) for
CGROUP_WRITEBACK to work, am I right? But in principle tracking writeback
state and doing writeback per memcg doesn't seem to be bound to any device
properties so we could do that right?

The main issue is that cgroup should somehow know how the processes
are mapped to the underlying IO layer - the IO domain should somehow
be defined.  We can introduce an intermediate abstraction which maps
to blkcg and whatever other cgroup controllers which may define cgroup
IO domains but given that such cases would be fairly niche, I think
we'd be better off making those corner cases represent themselves
using blkcg rather than introducing an additional layer.

Well, unless there is some specific mapping for the device, we could just
fall back to attributing everything to the root cgroup. We would still
account dirty pages in memcg, throttle writers in memcg when there are too
many dirty pages, issue writeback for inodes in memcg with enough dirty
pages etc. Just all IO from different memcgs would be equal so no
separation would be there. But it would still seem better that just
ignoring the split of dirty pages among memcgs as we do now... Thoughts?

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR