Re: [PATCH] mm: mmap_lock: replace get_memcg_path_buf() with on-stack buffer

[Axel Rasmussen <axelrasmussen@google.com>]

On Fri, Jun 21, 2024 at 10:57 PM Tetsuo Handa
<penguin-kernel@i-love.sakura.ne.jp> wrote:
>
> On 2024/06/22 8:03, Axel Rasmussen wrote:
> > No objections. Looking back all the way to the first version [1] the
> > buffers were already percpu, instead of on the stack like this. IOW,
> > there was no on-list discussion about why this shouldn't go on the
> > stack. It has been a while, but if memory serves I opted to do it that
> > way just out of paranoia around putting large buffers on the stack.
> > But, I agree 256 bytes isn't all that large.
> >
> > That v1 patch wasn't all that complex, but then again it didn't deal
> > with various edge cases properly :) so it has grown significantly more
> > complex over time. Reconsidering the approach seems reasonable now,
> > given how much code this removes.
> >
> > This change looks straightforwardly correct to me. You can take:
> >
> > Reviewed-by: Axel Rasmussen <axelrasmussen@google.com>
>
> Thank you. One question. CONTEXT_COUNT was defined as below.
>
> >> -/*
> >> - * How many contexts our trace events might be called in: normal, softirq, irq,
> >> - * and NMI.
> >> - */
> >> -#define CONTEXT_COUNT 4
>
> Is there possibility that this function (or in general, trace events) is called from NMI
> context? If yes, I worry that functions called from get_mm_memcg_path() are not NMI-safe.
> Original change at
> https://lkml.kernel.org/r/3e9b2a54-73d4-48cb-a510-d17984c97a45@I-love.SAKURA.ne.jp was
> posted due to worrying about NMI safety.

I think it's unlikely, but I'm not certain. Although trace events *in
general* can be called from NMI context.

This code was added based on the discussion here [1] from v4 of this
patchset. At the time I don't think we had any certainty that this
*was* called from NMI context, but rather just that if CONTEXT_COUNT
was 4 when it only had to be 3 in practice, it wasn't a huge deal, but
getting it wrong the other way would be much worse.

[1]: https://lore.kernel.org/r/20201020184746.300555-2-axelrasmussen@google.com

>