On 13.08.2018 20:53, Roman Gushchin wrote:

On Mon, Aug 13, 2018 at 01:11:19PM -0400, Johannes Weiner wrote:

quoted

On Mon, Aug 13, 2018 at 06:48:42AM -0700, Tejun Heo wrote:

quoted

Hello, Konstantin.

On Mon, Aug 13, 2018 at 09:58:05AM +0300, Konstantin Khlebnikov wrote:

quoted

After removing cgroup subsystem state could leak or live in background
forever because it is pinned by some reference. For example memory cgroup
could be pinned by pages in cache or tmpfs.

This patch adds common debugfs interface for listing basic state for each
controller. Controller could define callback for dumping own attributes.

In file /sys/kernel/debug/cgroup/<controller> each line shows state in
format: <common_attr>=<value>... [-- <controller_attr>=<value>... ]

Seems pretty useful to me.  Roman, Johannes, what do you guys think?

Totally agree with the idea and was about to suggest something similar.

quoted

Generally I like the idea of having more introspection into offlined
cgroups, but I wonder if having only memory= and swap= could be a
little too terse to track down what exactly is pinning the groups.

Roman has more experience debugging these pileups, but it seems to me
that unless we add a breakdown off memory, and maybe make slabinfo
available for these groups, that in practice this might not provide
that much more insight than per-cgroup stat counters of dead children.

I agree here.

I don't think that we could cover all cases with single interface.

This debugfs just gives simple entry point for debugging:
- paths for guessing user
- pointers for looking with gdb via kcore
- inode numbers for page-types - see second and third patch

For slab: this could show one of remaining slab. Anyway each of them pins css.

It's hard to say in advance what numbers are useful, so let's export
these numbers, but also make the format more extendable, so we can
easily add new information later. Maybe, something like:

cgroup {
   path = ...
   ino = ...
   main css {
     refcnt = ...
     key = value
     ...
   }
   memcg css {
     refcnt = ...
     ...
   }
   some other controller css {
   }
   ...
}

Also, because we do batch charges, printing numbers without draining stocks
is not that useful. All stats are also per-cpu cached, what adds some
inaccuracy.

Seems too verbose. And one-line key=value format is more grep/awk friendly.

Anyway such extended debugging could be implemented as gdb plugin.
While simple list of pointers gives enough information for dumping structures
with gdb alone without extra plugins.

Thanks!