[RFC PATCH 0/3] mm/zsmalloc: reduce lock contention in zs_free()

[Wenchao Hao]

Swap freeing can be expensive when unmapping a VMA containing many swap
entries. This has been reported to significantly delay memory reclamation
during Android's low-memory killing, especially when multiple processes
are terminated to free memory, with slot_free() accounting for more than
80% of the total cost of freeing swap entries.

Lock contention in zs_free() is a major contributor to this cost:
  - pool->lock (rwlock) read-side atomic operations become expensive
    under multi-process concurrency due to cacheline bouncing
  - class->lock held across zspage page freeing causes zone->lock
    contention to propagate back

This series addresses both issues:

Patch 1: Encode class_idx in obj value
  On 64-bit systems, OBJ_INDEX_BITS is over-provisioned. We split it
  into class_idx + obj_idx so that zs_free() can determine the correct
  size_class from the obj value alone, without needing pool->lock.

Patch 2: Remove pool->lock from zs_free()
  With class_idx available from the obj encoding, zs_free() acquires
  only class->lock (re-reading obj for a stable PFN). This eliminates
  rwlock read-side contention between concurrent zs_free() calls and
  page migration/compaction.

Patch 3: Drop class->lock before freeing zspage pages
  Move the actual page release (free_zspage) outside class->lock. The
  bookkeeping is done under the lock, but buddy allocator interaction
  (zone->lock) no longer nests inside class->lock.

Performance results:

Test: each process independently mmap 256MB, write data, madvise
MADV_PAGEOUT to swap out via zram (lzo-rle), then concurrent munmap.

Raspberry Pi 4B (4-core ARM64 Cortex-A72):

  mode        Base       Patched     Speedup
  single      59.0ms     56.0ms      1.05x
  multi 2p    94.6ms     66.7ms      1.42x
  multi 4p    202.9ms    110.6ms     1.83x

x86 (20-core Intel i7-12700, 16 concurrent processes):

  mode        Base       Patched     Speedup
  single      11.7ms     9.8ms       1.19x
  multi 2p    24.1ms     17.2ms      1.40x
  multi 4p    63.0ms     45.3ms      1.39x

Single-process shows modest improvement. With multiple processes,
each read_lock/read_unlock atomically modifies the shared rwlock
reader count, and the cost of these atomic operations increases
with more CPUs accessing the same cacheline concurrently.
Eliminating pool->lock removes this overhead entirely.

Patch 1-2 only work on 64-bit systems (gated by ZS_OBJ_CLASS_IDX);
32-bit falls back to the original pool->lock path. Patch 3 benefits
all architectures.

Wenchao Hao (2):
  mm/zsmalloc: encode class index in obj value for lockless class lookup
  mm/zsmalloc: remove pool->lock from zs_free on 64-bit systems

Xueyuan Chen (1):
  mm/zsmalloc: drop class lock before freeing zspage

 mm/zsmalloc.c | 146 ++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 131 insertions(+), 15 deletions(-)

--
2.34.1