Re: [patch] zswap: fix zswap_frontswap_load() vs zsmalloc::map/unmap() might_sleep() splat

[Vitaly Wool]

On Sat, 19 Dec 2020, 11:27 Mike Galbraith, <efault@xxxxxx> wrote:
>
> On Sat, 2020-12-19 at 11:20 +0100, Vitaly Wool wrote:
> > Hi Mike,
> >
> > On Sat, Dec 19, 2020 at 11:12 AM Mike Galbraith <efault@xxxxxx> wrote:
> > >
> > > (mailer partially munged formatting? resend)
> > >
> > > mm/zswap: fix zswap_frontswap_load() vs zsmalloc::map/unmap() might_sleep() splat
> > >
> > > zsmalloc map/unmap methods use preemption disabling bit spinlocks.  Take the
> > > mutex outside of pool map/unmap methods in zswap_frontswap_load() as is done
> > > in zswap_frontswap_store().
> >
> > oh wait... So is zsmalloc taking a spin lock in its map callback and
> > releasing it only in unmap? In this case, I would rather keep zswap as
> > is, mark zsmalloc as RT unsafe and have zsmalloc maintainer fix it.
>
> The kernel that generated that splat was NOT an RT kernel, it was plain
> master.today with a PREEMPT config.


I see, thanks. I don't think it makes things better for zsmalloc
though. From what I can see, the offending code is this:

>        /* From now on, migration cannot move the object */
>        pin_tag(handle);

Bit spinlock is taken in pin_tag(). I find the comment above somewhat
misleading, why is it necessary to take a spinlock to prevent
migration? I would guess an atomic flag should normally be enough.

zswap is not broken here, it is zsmalloc that needs to be fixed.

Best regards,
   Vitaly