[RFC PATCH 0/7] mm: store zswap entries in swap table (Pointer entry)

[Baoquan He]

Currently zswap uses a per-swap-device xarray to map swap offsets to
zswap_entry pointers. This introduces an additional tree walk on every
zswap entry access.

This series replaces the xarray with a "Pointer" entry type stored
directly in the swap cluster table. The new swp_tb_t variant embeds
a zswap_entry pointer plus a entry count, eliminating the xarray and
its lookup overhead entirely.

Design overview
---------------
A swap table slot can now be one of five types:

  NULL:    |---------------- 0 ---------------|
  Shadow:  |SWAP_COUNT|Z|---- SHADOW_VAL ---|1|
  PFN:     |SWAP_COUNT|Z|------ PFN -------|10|
  Pointer: |----------- Pointer ----------|100|
  Bad:     |------------- 1 -------------|1000|

When zswap compresses a page, it saves the original swap table entry
(PFN or Shadow) in zswap_entry->swp_tb_val, then writes the Pointer
entry into the swap table.  The swap count and flags remain accessible
via helper functions that read them from swp_tb_val.

The zswap_entry is threaded through the swapin call chain via a new
void **zentry parameter, so zswap_load() receives it directly instead
of doing an xarray lookup.

Performance
-----------
[~]# free -h
               total        used        free      shared  buff/cache   available
Mem:           3.8Gi       155Mi       3.4Gi       656Ki       315Mi       3.7Gi
Swap:          4.0Gi        54Mi       3.9Gi

Scalability benchmark: 'usemem -O -y -x -n 1 5G', 10 runs under cgroup
MemoryMax=2G MemorySwapMax=3G.  zswap enabled with zstd compressor.

  Kernel A: cfb8731f5396 (xarray-based lookup, baseline)
  Kernel B: (Pointer entry)

5G test â?? the gap widens significantly as swap pressure grows:

     Metric           xarray(A)       Pointer(B)
    --------------------------------------------------
    Avg throughput      709,556 KB/s    792,681 KB/s  (+11.7%)
    Avg free time       167,994 µs     132,896 µs   (-20.9%)

  Free time trend across 10 rounds at 5G:

    Pointer: 126â??124â??126â??126â??124â??129â??138â??145â??142â??148  (gradual, +17%)
    xarray:  136â??139â??136â??160â??162â??161â??159â??230â??219â??178  (collapse, +31%)

  xarray degrades sharply in the later rounds (last 3 avg free time
  209 ms vs Pointer's 145 ms, 1.44Ã? worse) as the radix tree
  insertion/deletion cost grows non-linearly under sustained swap
  churn.  Pointer entry avoids this entirely by looking up entries
  directly in the swap cluster table, which is already hot in cache.

Baoquan He (7):
  mm/swap_table: add Pointer type support for zswap entry storage
  mm/zswap: add swp_tb_val to zswap_entry and Pointer accessor helpers
  mm/swap_state: do zswap conversion in __swap_cache_do_del_folio
  mm/zswap, mm/swap_state: handle Pointer entries in add_folio path
  mm/swapfile: handle Pointer entries in count/dup/put paths
  mm/zswap, mm/swap_state: migrate zswap to use swap_table for entry
    indexing
  mm/zswap, mm/swap_state: replace xarray indexing with swap table
    lookups

 include/linux/zswap.h |   4 +-
 mm/page_io.c          |   4 +-
 mm/swap.h             |  52 +++++-
 mm/swap_state.c       | 117 +++++++++++---
 mm/swap_table.h       |  35 ++++-
 mm/swapfile.c         | 107 ++++++++++---
 mm/zswap.c            | 358 +++++++++++++++++++++++++++++-------------
 7 files changed, 508 insertions(+), 169 deletions(-)

-- 
2.54.0