Re: [PATCH v4 10/10] mm: zswap: Compress batching with Intel IAA in zswap_batch_store() of large folios.

[Johannes Weiner <hannes@cmpxchg.org>]

On Mon, Nov 25, 2024 at 12:20:01PM -0800, Yosry Ahmed wrote:
> On Fri, Nov 22, 2024 at 11:01 PM Kanchana P Sridhar
> <kanchana.p.sridhar@intel.com> wrote:
> >
> > This patch adds two new zswap API:
> >
> >  1) bool zswap_can_batch(void);
> >  2) void zswap_batch_store(struct folio_batch *batch, int *errors);
> >
> > Higher level mm code, for instance, swap_writepage(), can query if the
> > current zswap pool supports batching, by calling zswap_can_batch(). If so
> > it can invoke zswap_batch_store() to swapout a large folio much more
> > efficiently to zswap, instead of calling zswap_store().
> >
> > Hence, on systems with Intel IAA hardware compress/decompress accelerators,
> > swap_writepage() will invoke zswap_batch_store() for large folios.
> >
> > zswap_batch_store() will call crypto_acomp_batch_compress() to compress up
> > to SWAP_CRYPTO_BATCH_SIZE (i.e. 8) pages in large folios in parallel using
> > the multiple compress engines available in IAA.
> >
> > On platforms with multiple IAA devices per package, compress jobs from all
> > cores in a package will be distributed among all IAA devices in the package
> > by the iaa_crypto driver.
> >
> > The newly added zswap_batch_store() follows the general structure of
> > zswap_store(). Some amount of restructuring and optimization is done to
> > minimize failure points for a batch, fail early and maximize the zswap
> > store pipeline occupancy with SWAP_CRYPTO_BATCH_SIZE pages, potentially
> > from multiple folios in future. This is intended to maximize reclaim
> > throughput with the IAA hardware parallel compressions.
> >
> > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > Suggested-by: Yosry Ahmed <yosryahmed@google.com>
> 
> This is definitely not what I suggested :)
> 
> I won't speak for Johannes here but I suspect it's not quite what he
> wanted either.

It is not.

I suggested having an integrated code path where "legacy" stores of
single pages is just the batch_size=1 case.

https://lore.kernel.org/linux-mm/20241107185340.GG1172372@cmpxchg.org/

> What we really need to do (and I suppose what Johannes meant, but
> please correct me if I am wrong), is to make the existing flow work
> with batches.
> 
> For example, most of zswap_store() should remain the same. It is still
> getting a folio to compress, the only difference is that we will
> parallelize the page compressions. zswap_store_page() is where some
> changes need to be made. Instead of a single function that handles the
> storage of each page, we need a vectorized function that handles the
> storage of N pages in a folio (allocate zswap_entry's, do xarray
> insertions, etc). This should be refactoring in a separate patch.
> 
> Once we have that, the logic introduced by this patch should really be
> mostly limited to zswap_compress(), where the acomp interfacing would
> be different based on whether batching is supported or not. This could
> be changes in zswap_compress() itself, or maybe at this point we can
> have a completely different path (e.g. zswap_compress_batch()). But
> outside of that, I don't see why we should have a completely different
> store path for the batching.

+1