Re: [PATCH v2 0/3] mm: bypass swap readahead for zswap

[Alexandre Ghiti]

Hi everyone,

On Wed, Jul 22, 2026 at 6:35 PM Alexandre Ghiti <alex@xxxxxxxx> wrote:
>
> >
> First, the title is a bit misleading but this patchset is a follow-up of
> the zswap bypass readahead, so I kept it.
>
> Swap readahead walks a window around the faulting entry and reads every
> neighbour. Some of those reads are synchronous (a zswap-resident entry
> or an entry on an SWP_SYNCHRONOUS_IO device) with no asynchronous device I/O
> to overlap them with. Today they run inline during the window walk, so they
> are pure added latency in front of the read the fault is actually waiting on.
>
> This series splits the window walk into two passes and reorders it around that
> distinction:
>
> Patch 1 (overlap): issue the asynchronous device reads first, defer the
> synchronous neighbours, and read them while the device I/O is in flight. This
> overlaps the synchronous work with the disk reads instead of serialising it in
> front of them -- the win for mixed (zswap + disk) workloads.
>
> Patch 2 (bypass unless the target needs I/O): when the faulting target is
> itself synchronous there is no in-flight I/O to hide the second pass behind, so
> skip it. This is the zswap readahead bypass, gated on the target: it helps
> pure-zswap workloads and, above all, keeps synchronous-target page-fault
> latency low.
>
> Patch 3 (SWP_SYNCHRONOUS_IO): treat SWP_SYNCHRONOUS_IO devices (zram) as
> synchronous sources too, so a disk-target fault overlaps its zram neighbours'
> reads behind its own disk I/O -- the same benefit patch 1 gives zswap
> neighbours, now for mixed zram+disk workloads.
>

Let me know if I went too far with this new version or if you'd prefer
something simpler. Or anything else you need, I'll be happy to come up
with a new version.

Thanks,

Alex

> Changes since the RFC [1]:
> - patch 1: overlap the synchronous decompression with the in-flight disk I/O
> - patch 2: bypass the second pass when the target needs no disk I/O
> - patch 3: extend the same handling to SWP_SYNCHRONOUS_IO devices (zram)
>
> [1] https://lore.kernel.org/all/20260624075700.751467-1-alex@xxxxxxxx/
>
> Alexandre Ghiti (3):
>   mm: swap: overlap synchronous readahead decompression with disk I/O
>   mm: swap: bypass synchronous readahead unless the target needs I/O
>   mm: swap: treat SWP_SYNCHRONOUS_IO devices as synchronous in readahead
>
>  include/linux/swap.h  |  6 ++++
>  include/linux/zswap.h |  6 ++++
>  mm/swap_state.c       | 67 +++++++++++++++++++++++++++++++++++++++++--
>  mm/swapfile.c         | 11 +++++++
>  mm/zswap.c            | 17 +++++++++++
>  5 files changed, 105 insertions(+), 2 deletions(-)
>
>
> base-commit: 13a1e1a618858407fa12c391f664ea750651f6b2
> --
> 2.53.0-Meta
>
>