[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <YKbaqfTwdxkLaSUD@slm.duckdns.org>
Date: Thu, 20 May 2021 17:54:49 -0400
From: Tejun Heo <tj@...nel.org>
To: Dmitry Monakhov <dmtrmonakhov@...dex-team.ru>
Cc: linux-kernel@...r.kernel.org, axboe@...nel.dk
Subject: Re: [PATCH v2] blk-throttle: fix race between submitter and
throttler thread
On Thu, May 20, 2021 at 07:44:13PM +0000, Dmitry Monakhov wrote:
> Changes since v1:
> - Improve comments
>
> Currently we call bio_set_flag(bio, BIO_THROTTLED) unconditionally
> at the end of blk_throtl_bio w/o queue_lock. But once we drop queue_lock,
> bio may already be processed by thottler thread, so both threads
> may update bio->flags concurently
>
> Dipite that race window is tiny, it happens in real life under heavy load.
> It looks like follows:
>
> SUBMITTER_THREAD (CPU1) THROTTLER_THREAD (CPU2)
> ->blk_throtl_bio
> ->throtl_add_bio_tg
> (1) bio_set_flag(bio, BIO_THROTTLED);
> spin_unlock_irq(q->queue_lock);
> ->blk_throtl_dispatch_work_fn
> (2)spin_lock_irq(q->queue_lock);
> ->generic_make_request
> ->blk_queue_split
> (3)bio_set_flag(bio, BIO_CHAINED)
>
> (4) bio_set_flag(bio, BIO_THROTTLED);
>
> Since bio->bi_flags is not atomic it will be cached on each CPU
> CPU1 will cache it at the step (1), and changes from step(3) is not visiable,
> so BIO_CHAINED flag will be lost and rewritten at step(4).
> This result in ->bi_end_io() will be called multiple times once for each
> chained bio and once for parent bio.
>
> Bug#2: submit_bio_checks() call blkcg_bio_issue_init() for throttled bio,
> but at this moment bio may be already be completed and freed by throttler thread
>
> In order to fix both issues we should modify throttled bio under queue_lock only.
>
> Fixes: 111be88398174 ("block-throttle: avoid double charge")
> Signed-off-by: Dmitry Monakhov <dmtrmonakhov@...dex-team.ru>
Acked-by: Tejun Heo <tj@...nel.org>
Thanks.
--
tejun
Powered by blists - more mailing lists