Re: [RFC v3 00/15] xswap: extendable (virtual) swap device backed by zswap

[Nhat Pham]

On Thu, Aug 13, 2026 at 6:49 AM Baoquan He <hebaoquan@xxxxxxxxxx> wrote:
>
> xswap is an extendable swap device with no backing storage.
> Swapped-out pages live only in zswap, so the device wastes no disk space
> and its size is independent of any physical device.
>
> xswap decouples PTE swap entries from physical backing storage.  The
> cluster_info array is backed by a sparse vmalloc (VM_SPARSE) area that is
> grown and shrunk on demand:
>
> - Grow: when cluster allocation runs out of free clusters and the device
>   is below its ceiling, more physical pages are mapped into the VM_SPARSE
>   area and their clusters are added to the free list.
>
> - Shrink: when contiguous free clusters accumulate at the tail of the
>   mapped range (tracked in O(1) via nr_free_tail), they are unmapped and
>   the backing pages freed.  Shrink is deferred to a workqueue to avoid
>   lock recursion.
>
> A per-device ceiling (nr_clusters) bounds growth and is adjustable at
> runtime via debugfs.
>
> Interface:
>
>   /sys/kernel/mm/xswap/create           write a percent of RAM (0 for
>                                         the default) to create a device
>   /sys/kernel/mm/xswap/destroy          write a swap type to tear down
>                                         a device
>   /sys/kernel/debug/xswap/type<N>_cluster_limit
>                                         read/write the per-device
>                                         cluster ceiling
>
> Since xswap has no backing, swapped-out pages are stored compressed in
> zswap: physical writeout is skipped, and zswap writeback is disabled when
> every swapfile in the system is an xswap device.  xswap requires zswap, so
> device creation is refused when zswap is unavailable.
>
> Naming:
> ======
> I'm going with "xswap" (the "x" for extendable/extension) rather than "vswap".
> Chris suggested this name, and this aligns with the "VFS-like swap layers"
> direction Chris Li described in the first swap abstraction LPC talk
> (co-hosted with Yosry) — the swap ops and the xswap extension interfaces in
> this series are moving toward exactly that. I don't have a strong preference
> between xswap and vswap, so if reviewers object to the name, please comment.
>
> Note:
> =====
> Most of the added lines come from the switch to a sysfs-based create/destroy
> interface. v2 created an xswap device by handing swapon() a "header-only"
> swap file (mkswap followed by dd of just the first  4K), and teardown reused
> the normal swapoff path. v3 makes xswap truly file-less:
> /sys/kernel/mm/xswap/{create,destroy} replace that file dance.
>
> This patchset only build the base. On top of this, I believe Nhat's core code
> of xswap writeback, rmap etc can be implemented simpler. E.g, we only need add
> one field in struct swap_cluster_info to let xs_table point to physical swap
> entry, or zswap entry etc. No need to introduce struct swap_cluster_info_dynamic.

It's functionally equivalent. I think it looks cleaner to separate
vswap-only bits into a vswap-only struct (technically saving 1 word
for the cluster that doesnt need it too).

How would writeback, rmap be simpler? The complexity comes from the
operations around it, not the actual vtable/xs_table operations
itself.

>
> --- a/mm/swap.h
> +++ b/mm/swap.h
> @@ -57,6 +57,9 @@ struct swap_cluster_info {
>         u8 order;
>         atomic_long_t __rcu *table;     /* Swap table entries, see mm/swap_table.h */
>         unsigned int *extend_table;     /* For large swap count, protected by ci->lock */
> +#ifdef CONFIG_XSWAP
> +       unsigned long *xs_table;
> +#endif
>
>
> Testing:
> ========
> 1. enable zswap
> # echo 1 > /sys/module/zswap/parameters/enabled
>
> 2. create xswap device
> ~# echo 0 > /sys/kernel/mm/xswap/create
> ~# swapon
> NAME   TYPE  SIZE USED PRIO
> xswap0 xswap 1.2G   0B   -1
> ~# echo 80 > /sys/kernel/mm/xswap/create
> ~# swapon
> NAME   TYPE  SIZE USED PRIO
> xswap0 xswap 1.2G   0B   -1
> xswap1 xswap 3.1G   0B   -1
>
> 3. tune the zswap size
> ~# cat /sys/kernel/debug/xswap/type0_cluster_limit
> 1179
> ~# cat /sys/kernel/debug/xswap/type1_cluster_limit
> 3145
> ~# echo 2048 > /sys/kernel/debug/xswap/type0_cluster_limit
> ~# echo 2048 > /sys/kernel/debug/xswap/type1_cluster_limit
> ~# swapon
> NAME   TYPE  SIZE   USED PRIO
> xswap0 xswap   2G 377.9M   -1
> xswap1 xswap   2G 376.2M   -1
>
> 4. add memory pressure
> stress-ng --vm 1 --vm-bytes 4G --vm-keep --timeout 120s &
>
> create/destroy and grow/shrink xswap device casually, all passed.

Sigh.

We have something that has performance benchmarks, including on a
real, production workload:

https://lore.kernel.org/all/20260825153238.2695446-1-nphamcs@xxxxxxxxx/

And somehow, we decide to implement a different data structures,
without concrete numbers to show that it is more efficient than xarray
(only intuitions)?

At the risk of sounding like a broken record - I have to ask: why
don't we land use case first, optimize later? Especially since, we
have data showing that performance is fairly close, AND users who do
not want to risk it still have the old code and can do A/B testing
before they decide to switch?