[PATCH v6 0/9] memcg: per cgroup dirty page accounting

[Curt Wohlgemuth]

On Thu, Mar 17, 2011 at 7:53 AM, Jan Kara <jack at suse.cz> wrote:
> On Thu 17-03-11 13:43:50, Johannes Weiner wrote:
>> > - mem_cgroup_balance_dirty_pages(): if memcg dirty memory usage if above
>> >   background limit, then add memcg to global memcg_over_bg_limit list and use
>> >   memcg's set of memcg_bdi to wakeup each(?) corresponding bdi flusher.  If over
>> >   fg limit, then use IO-less style foreground throttling with per-memcg per-bdi
>> >   (aka memcg_bdi) accounting structure.
>>
>> I wonder if we could just schedule a for_background work manually in
>> the memcg case that writes back the corresponding memcg_bdi set (and
>> e.g. having it continue until either the memcg is below bg thresh OR
>> the global bg thresh is exceeded OR there is other work scheduled)?
>> Then we would get away without the extra list, and it doesn't sound
>> overly complex to implement.
>  But then when you stop background writeback because of other work, you
> have to know you should restart it after that other work is done. For this
> you basically need the list. With this approach of one-work-per-memcg
> you also get into problems that one cgroup can livelock the flusher thread
> and thus other memcgs won't get writeback. So you have to switch between
> memcgs once in a while.

In pre-2.6.38 kernels (when background writeback enqueued work items,
and we didn't break the loop in wb_writeback() with for_background for
other work items), we experimented with this issue.  One solution we
came up with was enqueuing a background work item for a given memory
cgroup, but limiting nr_pages to something like 2048 instead of
LONG_MAX, to avoid livelock.  Writeback would only operate on inodes
with dirty pages from this memory cgroup.

If BG writeback takes place for all memcgs that are over their BG
limts, it seems that simply asking if each inode is "related" somehow
to the a of dirty memcgs is the simplest way to go.  Assuming of
course that efficient data structures are built to answer this
question.

Thanks,
Curt

> We've tried several approaches with global background writeback before we
> arrived at what we have now and what seems to work at least reasonably...
>
>                                                                Honza
> --
> Jan Kara <jack at suse.cz>
> SUSE Labs, CR
>