Message-ID: <20250704181748.63181-1-ryncsn@gmail.com>
Date: Sat,  5 Jul 2025 02:17:39 +0800
From: Kairui Song <ryncsn@...il.com>
To: linux-mm@...ck.org
Cc: Andrew Morton <akpm@...ux-foundation.org>,
	Hugh Dickins <hughd@...gle.com>,
	Baolin Wang <baolin.wang@...ux.alibaba.com>,
	Matthew Wilcox <willy@...radead.org>,
	Kemeng Shi <shikemeng@...weicloud.com>,
	Chris Li <chrisl@...nel.org>,
	Nhat Pham <nphamcs@...il.com>,
	Baoquan He <bhe@...hat.com>,
	Barry Song <baohua@...nel.org>,
	linux-kernel@...r.kernel.org,
	Kairui Song <kasong@...cent.com>
Subject: [PATCH v4 0/9] mm/shmem, swap: bugfix and improvement of mTHP swap-in

From: Kairui Song <kasong@...cent.com>

The current mTHP swapin path have several problems. It may potentially
hang, may cause redundant faults due to false positive swap cache lookup,
and it will involve at least 4 Xarray tree walks (get order, get order
again, confirm swap, insert folio). And for !CONFIG_TRANSPARENT_HUGEPAGE
builds, it will performs some mTHP related checks.

This series fixes all of the mentioned issues, and the code should be more
robust and prepared for the swap table series. Now tree walks is reduced
to twice (get order & confirm, insert folio), !CONFIG_TRANSPARENT_HUGEPAGE
build overhead is also minimized, and comes with a sanity check now.

The performance is slightly better after this series, sequential swap in of
24G data from ZRAM, using transparent_hugepage_tmpfs=always (24 samples each):

Before:        11.02,  stddev: 0.06
After patch 1: 10.74,  stddev: 0.03
After patch 2: 10.72,  stddev: 0.01
After patch 3: 10.73,  stddev: 0.04
After patch 4: 10.72,  stddev: 0.02
After patch 5: 10.74,  stddev: 0.01
After patch 6: 10.13,  stddev: 0.09
After patch 7: 9.95,   stddev: 0.02
After patch 8: 9.88,   stddev: 0.04

Each patch improves the performance by a little, which is about ~10%
faster in total.

Build kernel test showed very slightly improvement, testing with make -j24
with defconfig in a 256M memcg also using ZRAM as swap, and
transparent_hugepage_tmpfs=always (6 test runs):

Before:      system time avg: 3911.80s
After:       system time avg: 3863.76s

--

V3: https://lore.kernel.org/linux-mm/20250627062020.534-1-ryncsn@gmail.com/
Updates:
- Split and reorganize a few intermediate patches [ Baolin Wang ]
- Fix a duplicated fault issue that may occur in one intermediate patch
  [ Baolin Wang ]
- Improve variable naming [ Baolin Wang ]
- Fix a wrong error value problem, return proper error value when
  swapin failed.
- No major code change from V3.

V2: https://lore.kernel.org/linux-mm/20250619175538.15799-1-ryncsn@gmail.com/
Updates:
- Split the clean up patch into 3 individual patches [ Baolin Wang ]
- Fix a code error in the first patch [ Baolin Wang ]
- I found there are some other remaining issue that can be fixed after
  the clean up so includes these too: fix major fault counter, and clean
  up the goto labels.

V1: https://lore.kernel.org/linux-mm/20250617183503.10527-1-ryncsn@gmail.com/
Updates:
- Improve of funtion name and variable names, also commit message [ Kemeng Shi,
  Dev Jain ]
- Correct Fixes: tag [ Andrew Morton ]
- Collect Reviewed-by.

Two of the patches in this series comes from the swap table series [1],
and worth noting that the performance gain of this series is independent
to the swap table series, we'll see another bigger performance gain and
reduce of memory usage after the swap table series.

I found these issues while trying to split the shmem changes out of the
swap table series for easier reviewing, and found several more issues
while doing stress tests for performance comparision. Barry also mentioned
that CONFIG_TRANSPARENT_HUGEPAGE may have redundant checks [2] and I
managed to clean them up properly too.

No issue is found with a few days of stress testing.

Link: https://lore.kernel.org/linux-mm/20250514201729.48420-1-ryncsn@gmail.com/ [1]
Link: https://lore.kernel.org/linux-mm/CAMgjq7AsKFz7UN+seR5atznE_RBTDC9qjDmwN5saMe+KL3b1mQ@mail.gmail.com/ [2]
Kairui Song (9):
  mm/shmem, swap: improve cached mTHP handling and fix potential hung
  mm/shmem, swap: avoid redundant Xarray lookup during swapin
  mm/shmem, swap: tidy up THP swapin checks
  mm/shmem, swap: tidy up swap entry splitting
  mm/shmem, swap: avoid false positive swap cache lookup
  mm/shmem, swap: never use swap cache and readahead for
    SWP_SYNCHRONOUS_IO
  mm/shmem, swap: simplify swapin path and result handling
  mm/shmem, swap: simplify swap entry and index calculation of large
    swapin
  mm/shmem, swap: fix major fault counting

 mm/shmem.c | 273 ++++++++++++++++++++++++++++-------------------------
 1 file changed, 147 insertions(+), 126 deletions(-)

-- 
2.50.0

Powered by blists - more mailing lists