[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <ZNW0WF/K8HS8AIu5@casper.infradead.org>
Date: Fri, 11 Aug 2023 05:08:56 +0100
From: Matthew Wilcox <willy@...radead.org>
To: Hui Zhu <teawaterz@...ux.alibaba.com>
Cc: viro@...iv.linux.org.uk, brauner@...nel.org, tytso@....edu,
adilger.kernel@...ger.ca, akpm@...ux-foundation.org, jack@...e.cz,
yi.zhang@...wei.com, hare@...e.de, p.raghav@...sung.com,
ritesh.list@...il.com, mpatocka@...hat.com,
linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
linux-ext4@...r.kernel.org, teawater@...group.com,
teawater@...il.com
Subject: Re: [PATCH] ext4_sb_breadahead_unmovable: Change to be no-blocking
On Fri, Aug 11, 2023 at 03:57:05AM +0000, Hui Zhu wrote:
> Encountered an issue where a large number of filesystem reads and writes
> occurred suddenly within a container. At the same time, other tasks on
> the same host that were performing filesystem read and write operations
> became blocked. It was observed that many of the blocked tasks were
> blocked on the ext4 journal lock. For example:
> PID: 171453 TASK: ffff926566c9440 CPU: 54 COMMAND: "Thread"
>
> Meanwhile, it was observed that the task holding the ext4 journal lock
> was blocked for an extended period of time on "shrink_page_list" due to
> "ext4_sb_breadahead_unmovable".
>
> The function "grow_dev_page" increased the gfp mask with "__GFP_NOFAIL",
> causing longer blocking times.
> /*
> * XXX: __getblk_slow() can not really deal with failure and
> * will endlessly loop on improvised global reclaim. Prefer
> * looping in the allocator rather than here, at least that
> * code knows what it's doing.
> */
> gfp_mask |= __GFP_NOFAIL;
> However, "ext4_sb_breadahead_unmovable" is a prefetch function and
> failures are acceptable.
That's a really good point.
> Therefore, this commit changes "ext4_sb_breadahead_unmovable" to be
> non-blocking, removing "__GFP_DIRECT_RECLAIM" from the gfp mask in the
> "grow_dev_page" function if caller is ext4_sb_breadahead_unmovable to
> alleviate memory-related blocking issues.
Uh, not like this though. Fix the gfp flags in the callers instead of
working this new "bool" flag through the buffer head layers.
Powered by blists - more mailing lists