Re: [RFC PATCH 5/9] mm: zswap: remove zswap_same_filled_pages_enabled

[Yosry Ahmed <yosryahmed@google.com>]

On Fri, Mar 29, 2024 at 10:45 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
>
> On Fri, Mar 29, 2024 at 03:02:10PM +0100, Maciej S. Szmigiero wrote:
> > On 29.03.2024 03:14, Yosry Ahmed wrote:
> > > On Thu, Mar 28, 2024 at 1:06 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > >>
> > >> On Thu, Mar 28, 2024 at 12:11 PM Johannes Weiner <hannes@cmpxchg.org> wrote:
> > >>>
> > >>> On Mon, Mar 25, 2024 at 11:50:13PM +0000, Yosry Ahmed wrote:
> > >>>> There is no logical reason to refuse storing same-filled pages more
> > >>>> efficiently and opt for compression. Remove the userspace knob.
> > >>>>
> > >>>> Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
> > >>>
> > >>> Acked-by: Johannes Weiner <hannes@cmpxchg.org>
> > >>>
> > >>> I also think the non_same_filled_pages_enabled option should go
> > >>> away. Both of these tunables are pretty bizarre.
> > >>
> > >> Happy to remove both in the next version :)
> > >
> > > I thought non_same_filled_pages_enabled was introduced with the
> > > initial support for same-filled pages, but it was introduced
> > > separately (and much more recently):
> > > https://lore.kernel.org/all/7dbafa963e8bab43608189abbe2067f4b9287831.1641247624.git.maciej.szmigiero@oracle.com/
> > >
> > > I am CCing Maciej to hear more about the use case for this.
> >
> > Thanks for CCing me.
> >
> > I introduced "non_same_filled_pages_enabled" a few years ago to
> > enable using zswap in a lightweight mode where it is only used for
> > its ability to store same-filled pages effectively.
>
> But all the pages it rejects go to disk swap instead, which is much
> slower than compression...
>
> > As far as I remember, there were some interactions between full
> > zswap and the cgroup memory controller - like, it made it easier
> > for an aggressive workload to exceed its cgroup memory.high limits.
>
> Ok, that makes sense! A container fairness measure, rather than a
> performance optimization.
>
> Fair enough, but that's moot then with cgroup accounting of the
> backing memory, f4840ccfca25 ("zswap: memcg accounting").

Right, this should no longer be needed with the zswap charging.

Maciej, is this still being used on kernels with f4840ccfca25 (5.19+)?
Any objections to removing it now?