Re: [RFC PATCH 7/9] mm: zswap: store zero-filled pages without a zswap_entry

[Yosry Ahmed]

On Thu, Mar 28, 2024 at 12:38 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> On Mon, Mar 25, 2024 at 11:50:15PM +0000, Yosry Ahmed wrote:
> > After the rbtree to xarray conversion, and dropping zswap_entry.refcount
> > and zswap_entry.value, the only members of zswap_entry utilized by
> > zero-filled pages are zswap_entry.length (always 0) and
> > zswap_entry.objcg. Store the objcg pointer directly in the xarray as a
> > tagged pointer and avoid allocating a zswap_entry completely for
> > zero-filled pages.
> >
> > This simplifies the code as we no longer need to special case
> > zero-length cases. We are also able to further separate the zero-filled
> > pages handling logic and completely isolate them within store/load
> > helpers.  Handling tagged xarray pointers is handled in these two
> > helpers, as well as the newly introduced helper for freeing tree
> > elements, zswap_tree_free_element().
> >
> > There is also a small performance improvement observed over 50 runs of
> > kernel build test (kernbench) comparing the mean build time on a skylake
> > machine when building the kernel in a cgroup v1 container with a 3G
> > limit. This is on top of the improvement from dropping support for
> > non-zero same-filled pages:
> >
> >               base            patched         % diff
> > real            69.915          69.757                -0.229%
> > user            2956.147        2955.244      -0.031%
> > sys             2594.718        2575.747      -0.731%
> >
> > This probably comes from avoiding the zswap_entry allocation and
> > cleanup/freeing for zero-filled pages. Note that the percentage of
> > zero-filled pages during this test was only around 1.5% on average.
> > Practical workloads could have a larger proportion of such pages (e.g.
> > Johannes observed around 10% [1]), so the performance improvement should
> > be larger.
> >
> > This change also saves a small amount of memory due to less allocated
> > zswap_entry's. In the kernel build test above, we save around 2M of
> > slab usage when we swap out 3G to zswap.
> >
> > [1]https://lore.kernel.org/linux-mm/20240320210716.GH294822@xxxxxxxxxxx/
> >
> > Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > ---
> >  mm/zswap.c | 137 ++++++++++++++++++++++++++++++-----------------------
> >  1 file changed, 78 insertions(+), 59 deletions(-)
>
> Tbh, I think this makes the code worse overall. Instead of increasing
> type safety, it actually reduces it, and places that previously dealt
> with a struct zswap_entry now deal with a void *.
>
> If we go ahead with this series, I think it makes more sense to either
>
> a) do the explicit subtyping of struct zswap_entry I had proposed, or

I suspect we won't get the small performance improvements (and memory
saving) with this approach. Neither are too significant, but it'd be
nice if we could keep them.

>
> b) at least keep the specialness handling of the xarray entry as local
>    as possible, so that instead of a proliferating API that operates
>    on void *, you have explicit filtering only where the tree is
>    accessed, and then work on struct zswap_entry as much as possible.

I was trying to go for option (b) by isolating filtering and casting
to the correct type in a few functions (zswap_tree_free_element(),
zswap_store_zero_filled(), and zswap_load_zero_filled()). If we
open-code filtering it will be repeated in a few places.

What did you have in mind?