Re: [PATCH v4 4/4] mm/page_owner: Record task command name

[Waiman Long]

On 2/3/22 07:10, Vlastimil Babka wrote:
> On 2/2/22 17:53, Waiman Long wrote:
> > On 2/1/22 10:28, Michal Hocko wrote:
> > > Cc Vlastimil
> > >
> > > On Mon 31-01-22 17:03:28, Waiman Long wrote:
> > > > The page_owner information currently includes the pid of the calling
> > > > task. That is useful as long as the task is still running. Otherwise,
> > > > the number is meaningless. To have more information about the allocating
> > > > tasks that had exited by the time the page_owner information is
> > > > retrieved, we need to store the command name of the task.
> > > >
> > > > Add a new comm field into page_owner structure to store the command name
> > > > and display it when the page_owner information is retrieved.
> > > I completely agree that pid is effectivelly useless (if not misleading)
> > > but is comm really telling all that much to compensate for the
> > > additional storage required for _each_ page in the system?
> > Yes, it does add an extra 16 bytes per page overhead. The command name can
> > be useful if one want to find out which userspace command is responsible for
> > a problematic page allocation. Maybe we can remove pid from page_owner to
> > save 8 bytes as you also agree that this number is not that useful.
> Pid could be used to correlate command instances (not perfectly if reuse
> happens), but command name could have a higher chance to be useful. In my
> experience the most useful were the stacktraces and gfp/order etc. anyway.
> So I wouldn't be opposed replacing pid with comm. The mild size increase
> should be acceptable, this is an opt-in feature for debugging sessions with
> known tradeoff for memory and cpu overhead for the extra info.

Thanks for the information.

I floated around dropping pid just as a possible way to reduce overall 
memory overhead. I did not do that in my patch and I am not planning to 
post any patch unless everybody agree.

Cheer,
Longman