Re: mm: mkfs.ext4 invoked oom-killer on i386 - pagecache_get_page

[Johannes Weiner]

On Thu, May 21, 2020 at 01:06:28PM -0700, Hugh Dickins wrote:
> On Thu, 21 May 2020, Johannes Weiner wrote:
> > do_memsw_account() used to be automatically false when the cgroup
> > controller was disabled. Now that it's replaced by
> > cgroup_memory_noswap, for which this isn't true, make the
> > mem_cgroup_disabled() checks explicit in the swap control API.
> > 
> > [hannes@xxxxxxxxxxx: use mem_cgroup_disabled() in all API functions]
> > Reported-by: Naresh Kamboju <naresh.kamboju@xxxxxxxxxx>
> > Debugged-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > Debugged-by: Michal Hocko <mhocko@xxxxxxxxxx>
> > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> > ---
> >  mm/memcontrol.c | 47 +++++++++++++++++++++++++++++++++++++++++------
> >  1 file changed, 41 insertions(+), 6 deletions(-)
> 
> I'm certainly not against a mem_cgroup_disabled() check in the only
> place that's been observed to need it, as a fixup to merge into your
> original patch; but this seems rather an over-reaction - and I'm a
> little surprised that setting mem_cgroup_disabled() doesn't just
> force cgroup_memory_noswap, saving repetitious checks elsewhere
> (perhaps there's a difficulty in that, I haven't looked).

Fair enough, I changed it to set the flag at initialization time if
mem_cgroup_disabled(). I was never a fan of the old flags, where it
was never clear what was commandline, and what was internal runtime
state - do_swap_account? really_do_swap_account? But I think it's
straight-forward in this case now.

> Historically, I think we've added mem_cgroup_disabled() checks
> (accessing a cacheline we'd rather avoid) where they're necessary,
> rather than at every "interface".

To me that always seemed like bugs waiting to happen. Like this one!

It's a jump label nowadays, so I've been liberal with these to avoid
subtle bugs.

> And you seem to be in a very "goto out" mood today - we all have
> our "goto out" days, alternating with our "return 0" days :)

:-)

But I agree, best to keep this fixup self-contained and defer anything
else to separate cleanup patches.

How about the below? It survives a swaptest with cgroup_disable=memory
for me.

Hugh, I started with your patch, which is why I kept you as the
author, but as the patch now (and arguably the previous one) is
sufficiently different, I dropped that now. I hope that's okay.

---