From: Jens Axboe <axboe@kernel.dk>
To: Damien Le Moal <Damien.LeMoal@wdc.com>,
"linux-block@vger.kernel.org" <linux-block@vger.kernel.org>
Cc: Masato Suzuki <masato.suzuki@wdc.com>,
Naohiro Aota <Naohiro.Aota@wdc.com>
Subject: Re: [PATCH] block: Fix __blkdev_direct_IO()
Date: Thu, 1 Aug 2019 21:57:39 -0600 [thread overview]
Message-ID: <0bf921d3-162b-8e51-63fa-5da757ccbbd2@kernel.dk> (raw)
In-Reply-To: <BYAPR04MB581699C381CE35C34DE45EBCE7D90@BYAPR04MB5816.namprd04.prod.outlook.com>
On 8/1/19 7:05 PM, Damien Le Moal wrote:
> On 2019/08/02 4:51, Jens Axboe wrote:
>> On 8/1/19 4:21 AM, Damien Le Moal wrote:
>>> The recent fix to properly handle IOCB_NOWAIT for async O_DIRECT IO
>>> (patch 6a43074e2f46) introduced two problems with BIO fragment handling
>>> for direct IOs:
>>> 1) The dio size processed is claculated by incrementing the ret variable
>>> by the size of the bio fragment issued for the dio. However, this size
>>> is obtained directly from bio->bi_iter.bi_size AFTER the bio submission
>>> which may result in referencing the bi_size value after the bio
>>> completed, resulting in an incorrect value use.
>>> 2) The ret variable is not incremented by the size of the last bio
>>> fragment issued for the bio, leading to an invalid IO size being
>>> returned to the user.
>>>
>>> Fix both problem by using dio->size (which is incremented before the bio
>>> submission) to update the value of ret after bio submissions, including
>>> for the last bio fragment issued.
>>
>> Thanks, applied. Do you have a test case? I ran this through the usual
>> xfstests and block bits, but didn't catch anything.
>>
>
> The problem was detected with our weekly RC test runs for zoned block
> devices. RC1 last week was OK, but failures happen on RC2 Monday. We
> never hit a oops for the BIO reference after submission but we were
> getting a lot of unaligned write errors for all types of zoned drive
> tested (real SMR disks, tcmu-runner ZBC handler disks and nullblk in
> zoned mode) using various applications (fio, dd, ...)
>
> Masato isolated the problem to very large direct writes and could
> reliably recreate the problem with a dd doing a single 8MB write to a
> sequential zone. With this case, blktrace showed that the 8MB write
> was split into multiple BIOs (expected) and the whole 8MB being
> cleanly written sequentially. But this was followed by a stream of
> small 4K writes starting around the end of the 8MB chunk, but within
> it, causing unaligned write errors (overwrite in sequential zones not
> being possible).
>
> dd if=/dev/zero of=/dev/nullb0 bs=8M oflag=direct count=1
>
> On a nullblk disk in zoned mode should recreate the problem, and
> blktrace revealing that dd sees a short write for the 8M IO and issues
> remaining as 4K requests.
>
> Using a regular disk, this however does not generate any error at all,
> which may explain why you did not see any problem. I think we can
> create a blktest case for this using nullblk in zoned mode. Would you
> like us to send one ?
Thanks for the detailed explanation. I think we should absolutely have a
blktests case for this, even if O_DIRECT isn't strictly block level,
it's definitely in a grey zone that I think we should cover.
--
Jens Axboe
next prev parent reply other threads:[~2019-08-02 3:57 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-08-01 10:21 [PATCH] block: Fix __blkdev_direct_IO() Damien Le Moal
2019-08-01 19:51 ` Jens Axboe
2019-08-02 1:05 ` Damien Le Moal
2019-08-02 3:57 ` Jens Axboe [this message]
2019-08-02 4:11 ` Damien Le Moal
2019-08-05 17:11 ` Bart Van Assche
2019-08-06 19:53 ` Jens Axboe
2019-08-09 13:01 ` Damien Le Moal
2019-08-09 15:52 ` Jens Axboe
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=0bf921d3-162b-8e51-63fa-5da757ccbbd2@kernel.dk \
--to=axboe@kernel.dk \
--cc=Damien.LeMoal@wdc.com \
--cc=Naohiro.Aota@wdc.com \
--cc=linux-block@vger.kernel.org \
--cc=masato.suzuki@wdc.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 a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).