Re: Virtual Swap Sync

From: Nhat Pham

Date: Sun Nov 22 2026 - 10:32:05 EST


On Thu, Nov 5, 2026 at 2:06 PM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
>
> Hi folks,
>

Hi folks - thanks everyone for attending. :)

I summarized the points we discussed below - please let me know if I
missed anything or misunderstood a point.

* Control interface: Google is interested in a knob to cease zswapping
(which already exists and will still work in vswap world). Kairui has
a use case for remapping all swapped-out pages.

Vswap currently bypassed when zswap is off; per-cgroup selection could
be added as concrete use case arises (for e.g migration/compaction).

David objected that a virtual swap abstraction should not carry a
swapoff interface per se. Folks also asked whether making vswap a
mandatory layer would make the implementation cleaner (Nhat will
explore this direction and follow up on this).

* Virtual space sizing: We discussed using vm.overcommit_ratio and
vm.overcommit_memory to limit the host's logical memory footprint. The
exact sizing formula is still unclear - Nhat will follow up with
Kairui on whether it is acceptable.

(Kairui, if possible, can you respond here for your exact use case? Do
you want to limit physical memory usage, or virtual? My understanding
is you want to gate physical + swap, which would be closer to the
latter).

* Priority and multiple swapfiles: should not be added for now.
No good use case for swap priority was identified.
Multiple swapfiles were used in the past to work around swapfile
lock contention, but it is unclear whether they are still needed with
Kairui's new cluster-based allocator. If contention remains a problem,
we should fix it in the kernel before putting that burden on
userspace.

* Proposed phase split and charging (from Nhat):

- Phase 1 - initial foundation:
- Decoupled zswap: Let zswap operate without reserving disk swap
for every virtual entry.
- Physical backing and writeback: Keep disk swap available as a
backend, including when zswap cannot store a page or an entry needs
writeback.
- Charging: Retain memsw charging for virtual entries; keep
virtual-slot usage out of memory.swap.max. Actual physical swap usage
keeps its resource charging.
- Usage counter: Expose vswap usage so users can monitor
virtual-space consumption separately from physical swap usage.
Essentially this gives you a method to estimate the logical footprint
(memory.current - swapcached + vswap).

- Subsequent phase(s) - further support and controls:
- VM_SPARSE: Add VM_SPARSE as an optimization after the
first-phase foundation.
- Runtime controls (added as use cases arise):
- Host-wide runtime disablement knob (and remapping of swapped out data).
- Per-cgroup vswap control (for e.g, if some cgroups want
migrations/physical swapfile compaction).

- Orthogonal use cases and concerns:
- Cgroup-level logical memory limits in cgroupv2 (memsw?) - swap
developers will follow up with memcg maintainers.
- Static sizing: Define the host-sizing formula and decide whether
another interface is needed.
- Swap tiering for cgroups: Explore backend selection when a
concrete workload needs it.

Folks, please comment on the summary and proposal.

Let us keep the communication line open and coordinate on the next
steps. I will schedule something again.

Also taking a step back - once we reach a consensus on vswap, maybe we
can repurpose this series of meeting for general swap alignment? There
are other swap/compressed swap developments right now (looking at the
orthogonal use cases list above for e.g), so maybe we can do this once
or twice a month. My experience in the past year or so has shown that
we clearly need better alignment :)

I will reach out to set up another round of meeting. I'm hoping we can
do it one more time in December, but I understand that holiday season
is coming :)

For folks in the U.S, I wish you all a merry and festive Thanksgiving.