Re: [Patch v4 1/2] freezer: check OOM kill while being frozen

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Monday, September 08, 2014 01:58:30 PM Cong Wang wrote:
> On Mon, Sep 8, 2014 at 1:54 PM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> > On Monday, September 08, 2014 10:40:17 AM Cong Wang wrote:
> >>
> >> We asked you to comment on either if this patch is safe for PM freeze
> >> if we don't have the cgroup_freezing() check, or if it is not safe why (so that
> >> I can put it in the comment).
> >
> > OK, which version of it?  Anything that has been posted as a complete patch you
> > have a link to?  Or am I supposed to go through the whole thread and figure out
> > how the patch *might* have looked had it been posted and then comment?
> 
> Sorry, I meant this version, the one we are discussing about.
> In case you missed it too, let me show the related code for you:
> 
> +       /* It might not be safe to check TIF_MEMDIE for pm freeze. */
> +       if (cgroup_freezing(current) && test_thread_flag(TIF_MEMDIE))
> +               return true;
> +
> 
> (return true means it will break the loop of freezing, that is thaw itself.)
> 
> So our question is if we really need the cgroup_freezing() check here?

No, I don't think so.

> IOW, is it safe to check TIF_MEMDIE and then thaw itself even if the
> freeze request is from PM? Hope it is clear now.

Yes, it is clear, thanks.

Why would the cgroup_freezing() test matter here at all?

> >> OK, maybe just one or two sentences. Let me know if the following
> >> comment is okay for you:
> >>
> >> /* OOM killer might decide to kill this process after it is frozen,
> >>   in this case it should thaw and die. */
> >
> > I don't think it's sufficient.  In particular, what does "thaw and die" mean
> > exactly?  It should be thawed immediately and then die or it should die right
> > after it's thawed (at one point in the future)?
> >
> 
> OK, let me try again:
> 
> /* OOM killer may decide to kill this process after it is frozen,
>   in this case SIGKILL can never be handled, so we should check
>  TIF_MEMDIE and if it is set, thaw and let SIGKILL kill it. */
> 
> (Again, more detailed explanation is in the changelog, comment is just
> for a quick reference, this seems what we all agree on.)

OK

I think you can simply add the test_thread_flag(TIF_MEMDIE) check to
__refrigerator() and it should be OK without the comment.

But please tell me this: If TIF_MEMDIE is set and we thaw the process,
can it do *anything* before dying or will it die immediately?

Rafael