Re: [PATCH v3 10/11] zram: use rw_semaphore for decompression streams

[Bo Zhang]

On Fri, 16 Oct 2026 17:54:25 +0900 Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx> wrote:
> A quick and dirty patch, as usual, just to play around with.  This
> attempts to improve the reader side, because the reader is side
> much more sensitive.  Tested only with zstd.

I applied the cross-CPU borrowing patch on-device and re-ran the same
zstd Android A/B, so it can be compared against the preempt-disable
numbers I posted earlier. Focusing on the read (exclusive decompression)
side, which is what this patch targets:

                        9+10+kmalloc   preempt-disable   cross-CPU borrow
  dexcl_contended       15,919         54                1,447
  dexcl_wait_ns_total   109.6 s        0.68 s            3.66 s
  dexcl_wait_ns_max     145 ms         34.6 ms           26.6 ms
  dexcl_avg_wait        6.88 ms        12.6 ms           2.53 ms
  dexcl_acquire         2.21M          2.75M             2.94M

Borrowing clearly works: dexcl contention drops 15,919 -> 1,447 (-91%),
and the residual 1,447 are the genuinely all-CPUs-busy cases that fall
through to wait_event(). More importantly for us, the worst case and the
average are the best of the three - max 145ms -> 26.6ms (-82%), avg down
to 2.53ms.

One note so the table isn't misread: the preempt-disable column also
disabled preemption on the *write* stream, which is why its
cwrite_contended was 48; this borrow-only build has no write-side change
(cwrite_contended back at ~16k, as before 9+10). That's expected - your
patch deliberately leaves the writer side for later, and the writer
isn't on the swap-in path, so it's lower priority for us too.

Scripts and raw numbers on request.

Thanks,
Bo