| From: |
| Chi Zhiling <chizhiling-AT-163.com> |
| To: |
| linux-fsdevel-AT-vger.kernel.org, linux-mm-AT-kvack.org, linux-kernel-AT-vger.kernel.org |
| Subject: |
| [RFC PATCH 0/4] mm/shmem: optimize read performance with folio batching |
| Date: |
| Fri, 15 May 2026 17:46:58 +0800 |
| Message-ID: |
| <[email protected]> |
| Cc: |
| Hugh Dickins <hughd-AT-google.com>, Baolin Wang <baolin.wang-AT-linux.alibaba.com>, Andrew Morton <akpm-AT-linux-foundation.org>, David Hildenbrand <david-AT-kernel.org>, Lorenzo Stoakes <ljs-AT-kernel.org>, Zi Yan <ziy-AT-nvidia.com>, "Liam R. Howlett" <liam-AT-infradead.org>, Nico Pache <npache-AT-redhat.com>, Ryan Roberts <ryan.roberts-AT-arm.com>, Dev Jain <dev.jain-AT-arm.com>, Barry Song <baohua-AT-kernel.org>, Lance Yang <lance.yang-AT-linux.dev>, "Matthew Wilcox (Oracle)" <willy-AT-infradead.org>, Jan Kara <jack-AT-suse.cz>, Chi Zhiling <chizhiling-AT-kylinos.cn> |
| Archive-link: |
| Article |
From: Chi Zhiling <[email protected]>
This series optimizes shmem read performance by implementing folio
batching in the read path and eliminating unnecessary lock operations.
Performance testing with fio:
(--ioengine=sync --rw=read --size=1G --runtime=120)
shmem (THP disabled):
bs=1M: 11.4 GiB/s
bs=64k: 11.1 GiB/s
bs=4k: 3814 MiB/s
shmem (THP disabled) + fbatch:
bs=1M: 12.8 GiB/s (+12%)
bs=64k: 12.3 GiB/s (+11%)
bs=4k: 3783 MiB/s (-0.8%)
shmem (THP enabled):
bs=1M: 13.8 GiB/s
bs=64k: 13.1 GiB/s
bs=4k: 3851 MiB/s
shmem (THP enabled) + fbatch:
bs=1M: 14.0 GiB/s (+1%)
bs=64k: 13.4 GiB/s (+2%)
bs=4k: 3811 MiB/s (-1%)
shmem preallocated via fallocate (THP disabled):
bs=1M: 24.0 GiB/s
bs=64k: 22.5 GiB/s
bs=4k: 4670 MiB/s
shmem preallocated via fallocate (THP disabled) + fbatch:
bs=1M: 29.3 GiB/s (+22%)
bs=64k: 26.7 GiB/s (+19%)
bs=4k: 4654 MiB/s (-0.3%)
shmem preallocated via fallocate (THP enabled):
bs=1M: 24.0 GiB/s
bs=64k: 22.9 GiB/s
bs=4k: 4698 MiB/s
shmem preallocated via fallocate (THP enabled) + fbatch:
bs=1M: 34.3 GiB/s (+43%)
bs=64k: 31.5 GiB/s (+38%)
bs=4k: 4689 MiB/s (-0.2%)
Chi Zhiling (4):
mm/shmem: add SGP_GET to get unlocked folio
mm/shmem: use SGP_GET in read operations
mm/shmem: optimize file read with folio batching
mm/shmem: make SGP_NOALLOC succeed on hole like SGP_READ
include/linux/shmem_fs.h | 5 +-
mm/khugepaged.c | 2 +-
mm/shmem.c | 132 ++++++++++++++++++++++++++++++++-------
3 files changed, 112 insertions(+), 27 deletions(-)
--
2.43.0
