Fwd: Possible io_uring regression with QEMU on Ubuntu's kernel

Thu Jul 1 08:13:47 UTC 2021

This came in as PM, so fwd'ing for all to be aware.

-Stefan

-------- Forwarded Message --------
Subject: Possible io_uring regression with QEMU on Ubuntu's kernel
Date: Wed, 30 Jun 2021 17:47:06 +0900
From: Juhyung Park <qkrwngud825 at gmail.com>
To: Kamal Mostafa <kamal at canonical.com>, Stefan Bader 
<stefan.bader at canonical.com>, io-uring <io-uring at vger.kernel.org>
CC: Jens Axboe <axboe at kernel.dk>, Stefano Garzarella <sgarzare at redhat.com>, 
qemu-devel at nongnu.org

Hi everyone.

With the latest Ubuntu 20.04's HWE kernel 5.8.0-59, I'm noticing some
weirdness when using QEMU/libvirt with the following storage
configuration:

<disk type="block" device="disk">
   <driver name="qemu" type="raw" cache="none" io="io_uring"
discard="unmap" detect_zeroes="unmap"/>
   <source dev="/dev/disk/by-id/md-uuid-df271a1e:9dfb7edb:8dc4fbb8:c43e652f-part1"
index="1"/>
   <backingStore/>
   <target dev="vda" bus="virtio"/>
   <alias name="virtio-disk0"/>
   <address type="pci" domain="0x0000" bus="0x07" slot="0x00" function="0x0"/>
</disk>

QEMU version is 5.2+dfsg-9ubuntu3 and libvirt version is 7.0.0-2ubuntu2.

The guest VM is unable to handle I/O properly with io_uring, and
nuking io="io_uring" fixes the issue.
On one machine (EPYC 7742), the partition table cannot be read and on
another (Ryzen 9 3950X), ext4 detects weirdness with journaling and
ultimately remounts the guest disk to R/O:

[    2.712321] virtio_blk virtio5: [vda] 3906519775 512-byte logical
blocks (2.00 TB/1.82 TiB)
[    2.714054] vda: detected capacity change from 0 to 2000138124800
[    2.963671] blk_update_request: I/O error, dev vda, sector 0 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.964909] Buffer I/O error on dev vda, logical block 0, async page read
[    2.966021] blk_update_request: I/O error, dev vda, sector 1 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.967177] Buffer I/O error on dev vda, logical block 1, async page read
[    2.968330] blk_update_request: I/O error, dev vda, sector 2 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.969504] Buffer I/O error on dev vda, logical block 2, async page read
[    2.970767] blk_update_request: I/O error, dev vda, sector 3 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.971624] Buffer I/O error on dev vda, logical block 3, async page read
[    2.972170] blk_update_request: I/O error, dev vda, sector 4 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.972728] Buffer I/O error on dev vda, logical block 4, async page read
[    2.973308] blk_update_request: I/O error, dev vda, sector 5 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.973920] Buffer I/O error on dev vda, logical block 5, async page read
[    2.974496] blk_update_request: I/O error, dev vda, sector 6 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.975093] Buffer I/O error on dev vda, logical block 6, async page read
[    2.975685] blk_update_request: I/O error, dev vda, sector 7 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.976295] Buffer I/O error on dev vda, logical block 7, async page read
[    2.980074] blk_update_request: I/O error, dev vda, sector 0 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.981104] Buffer I/O error on dev vda, logical block 0, async page read
[    2.981786] blk_update_request: I/O error, dev vda, sector 1 op
0x0:(READ) flags 0x0 phys_seg 1 prio class 0
[    2.982083] ixgbe 0000:06:00.0: Multiqueue Enabled: Rx Queue count
= 63, Tx Queue count = 63 XDP Queue count = 0
[    2.982442] Buffer I/O error on dev vda, logical block 1, async page read
[    2.983642] ldm_validate_partition_table(): Disk read failed.

Kernel 5.8.0-55 is fine, and the only io_uring-related change between
5.8.0-55 and 5.8.0-59 is the commit 4b982bd0f383 ("io_uring: don't
mark S_ISBLK async work as unbounded").

The weird thing is that this commit was first introduced with v5.12,
but neither the mainline v5.12.0 or v5.13.0 is affected by this issue.

I guess one of these commits following the backported commit from
v5.12 fixes the issue, but that's just a guess. It might be another
earlier commit:
c7d95613c7d6 io_uring: fix early sqd_list removal sqpoll hangs
9728463737db io_uring: fix rw req completion
6ad7f2332e84 io_uring: clear F_REISSUE right after getting it
e82ad4853948 io_uring: fix !CONFIG_BLOCK compilation failure
230d50d448ac io_uring: move reissue into regular IO path
07204f21577a io_uring: fix EIOCBQUEUED iter revert
696ee88a7c50 io_uring/io-wq: protect against sprintf overflow

It would be much appreciated if Jens could give pointers to Canonical
developers on how to fix the issue, and hopefully a suggestion to
prevent this from happening again.

Thanks,
Regards

-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <https://lists.ubuntu.com/archives/kernel-team/attachments/20210701/6bd06b1d/attachment.sig>