From: Gao Xiang <xiang@kernel.org>
To: Wenjie Qi <qwjhust@gmail.com>
Cc: xiang@kernel.org, chao@kernel.org, linux-erofs@lists.ozlabs.org,
linux-fsdevel@vger.kernel.org, shengyong1@xiaomi.com,
yangyongpeng@xiaomi.com, qiwenjie@xiaomi.com
Subject: Re: [PATCH] erofs: enable buffered RWF_DONTCACHE reads
Date: Thu, 19 Nov 2026 22:17:47 +0100 [thread overview]
Message-ID: <av9ne1tKVIBQgCI0@MacBookPro.fritz.box> (raw)
In-Reply-To: <b168501806b41e3f0be15efec4b54b371d8dd4e1.1795059802.git.qiwenjie@xiaomi.com>
On Thu, Nov 19, 2026 at 11:52:50AM +0800, Wenjie Qi wrote:
> From: Wenjie Qi <qiwenjie@xiaomi.com>
>
> EROFS uses filemap_read() for buffered reads, including the inode-sharing
> path, which preserves request flags when cloning the kiocb.
>
> Advertise FOP_DONTCACHE in both file operation tables so preadv2() callers
> can request that newly instantiated file-cache folios be dropped after
> consumption. This is useful for one-pass reads that need not retain data
> in the page cache.
>
> For compressed reads, retain trailing readmore and propagate dropbehind
> to the single-folio neighbor lookup. Readahead expansion already marks
> new folios through readahead_control. Suppress prefix readmore so it
> cannot refill pages already consumed and evicted. Ordinary readmore is
> unchanged.
>
> Tested sequential reads of 1-GiB files with 1-MiB requests on NVMe-backed
> LZ4 and Zstd EROFS images in x86-64 QEMU/KVM with two vCPUs and 2 GiB RAM.
> Each reader ran with memory.max=255852544 (244 MiB) and
> memory.swap.max=0. Normal and dontcache ran in ten paired rounds per
> layout, alternating their order and using a fresh guest for each run.
>
> Results are arithmetic means of ten runs. Memory values average the
> run-level time-weighted memory.current means. Percentage changes are
> relative to normal reads:
>
> MiB/s Mean memory, MiB
> Layout normal/dontcache change normal/dontcache change
> LZ4 2513.57/6350.61 +152.65% 169.53/17.39 -89.74%
> Zstd 2535.57/4066.49 +60.38% 179.01/18.17 -89.85%
>
> In this test, dontcache reduced average cgroup memory by about 90%,
> left no consumed file pages resident and recorded no memcg reclaim
> scans or steals. Both layouts were faster in all ten pairs. This
> offers a smaller cache footprint for one-pass reads on memory-constrained
> systems.
>
> Signed-off-by: Wenjie Qi <qiwenjie@xiaomi.com>
> Signed-off-by: Sheng Yong <shengyong1@xiaomi.com>
> Reviewed-by: Yongpeng Yang <yangyongpeng@xiaomi.com>
Yeah, that was in my own plan too but I never find a slot, thank you.
It would be useful for one-time load on large files like AI models
and these numbers are impressive:
Reviewed-by: Gao Xiang <xiang@kernel.org>
Thanks,
Gao Xiang
prev parent reply other threads:[~2026-11-19 21:17 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-19 3:52 [PATCH] erofs: enable buffered RWF_DONTCACHE reads Wenjie Qi
2026-11-19 21:17 ` Gao Xiang [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=av9ne1tKVIBQgCI0@MacBookPro.fritz.box \
--to=xiang@kernel.org \
--cc=chao@kernel.org \
--cc=linux-erofs@lists.ozlabs.org \
--cc=linux-fsdevel@vger.kernel.org \
--cc=qiwenjie@xiaomi.com \
--cc=qwjhust@gmail.com \
--cc=shengyong1@xiaomi.com \
--cc=yangyongpeng@xiaomi.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.