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

[Bo Zhang <zhangbo0325@gmail.com>]

Following up on the RFC thread [1]: Sergey pointed me at this series as
the reworked version of the by-direction split I had tested, so I applied
patches 9 and 10 and re-ran the same A/B. Big improvement.

[1] https://lore.kernel.org/lkml/20261012123328.1138059-1-zhangbo56@xiaomi.com

Setup as before: real Android device (MemTotal ~7.7GB, zram ~8GB), cycle
20 apps through fg/bg to force swap-out/in, 8 rounds, reboot before each.
Same trylock instrumentation on the stream locks across all builds. For
patch 10, "read contended" = dpin_contended + dexcl_contended. Decompress
("read") is the swap-in path, which is what matters for us.

I use decompress acquire count (swapped-in pages) as the comparability anchor
instead of swap volume, which stays within ~20% across builds.

                      baseline   preempt-disable   9 + 10
  read contended      24,749     406               1,959
  read wait_ns total  239.8 s    3.78 s            ~12 s
  read wait_ns max    757 ms     51.7 ms           72-96 ms
  read acquire        3,948,150  4,305,078         3,510,610

Decompression on 9+10 splits into:
  shared (stateless backend, single-page): 76.7% -> down_read never
      blocked (dpin_contended = 0)
  exclusive (cross-page object): 23.3% -> excl_lock, ~1,959 contended

- 9+10: read contention -92%, total wait ~-95%, worst case 757ms -> ~90ms.
  The rwsem does the heavy lifting: ~77% of swap-in decompressions now run
  concurrently with zero pin contention - exactly the reader-vs-reader
  queuing we were hitting. Clear win over the two-mutex split.
- Residual contention is all on the exclusive path, mostly cross-page
  objects (which need the ->local_copy bounce), not a specific backend -
  lzo-rle single-page objects all go shared.
- preempt-disable still shows the lowest read contention (406). It covers
  the other half of the root cause (holder preempted/migrated while
  holding the stream), so the two are complementary rather than
  competing: 9+10 removes reader-vs-reader queuing structurally, and the
  ~90ms tail that remains on 9+10 is the preempt/migrate effect.

Happy to share scripts and raw numbers on request.

Thanks,
Bo