[SRU][N][O][PATCH 0/3] Introduce and use sendpages_ok() instead of sendpage_ok() in nvme-tcp and drbg

Thu Jan 30 01:26:40 UTC 2025

BugLink: https://bugs.launchpad.net/bugs/2093871

[Impact]

Currently the nvme-tcp and drbg subsystems try to enable the MSG_SPLICE_PAGES
flag on pages to be written, and when MSG_SPLICE_PAGES is set, eventually it
calls skb_splice_from_iter(), which then checks all pages with sendpage_ok()
to see if all the pages are sendable.

At the moment, both subsystems only check the first page in a potentially
contiguous block of pages, if they are sendpage_ok(), and if the first page is,
then it just assumes all the rest are sendpage_ok() too, and sends the I/O off
to eventually be found out by skb_splice_from_iter(). If one or more of the
pages in the contiguous block is not sendpage_ok(), then we get a warn printed,
data transfer is aborted. In the nvme-tcp case, IO then hangs.

This patchset introduces sendpages_ok() which iterates over each page in a
contiguous block, checks if it is sendpage_ok(), and only returns true if all
of them are.

This resolves the whole MSG_SPLICE_PAGES flag situation, since you can now
depend on the result of sendpages_ok(), instead of just assuming everything is
okay.

This issue is what caused bug 2075110 [0] to be discovered in the first place,
since it was responsible for contigious blocks of pages where the first was
sendpage_ok(), but pages further into the block were not.

[0] https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2075110

Even with "md/md-bitmap: fix writing non bitmap pages" applied, the issue can
still happen, e.g. with merged IO pages, so this fix is still needed to
eliminate the issue.

[Fix]

The fixes landed in mainline 6.12-rc1:

commit 23a55f4492fcf868d068da31a2cd30c15f46207d
Author: Ofir Gal <ofir.gal at volumez.com>
Date:   Thu Jul 18 11:45:12 2024 +0300
Subject: net: introduce helper sendpages_ok()
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=23a55f4492fcf868d068da31a2cd30c15f46207d

commit 6af7331a70b4888df43ec1d7e1803ae2c43b6981
Author: Ofir Gal <ofir.gal at volumez.com>
Date:   Thu Jul 18 11:45:13 2024 +0300
Subject: nvme-tcp: use sendpages_ok() instead of sendpage_ok()
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=6af7331a70b4888df43ec1d7e1803ae2c43b6981

commit 7960af373ade3b39e10106ef415e43a1d2aa48c6
Author: Ofir Gal <ofir.gal at volumez.com>
Date:  Thu Jul 18 11:45:14 2024 +0300
Subject: drbd: use sendpages_ok() instead of sendpage_ok()
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7960af373ade3b39e10106ef415e43a1d2aa48c6

They are needed for noble and oracular.

[Testcase]

This is the same testcase as the original bug 2075110 [0], as the fix is
designed to prevent it or similar other bugs from happening again.

[0] https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2075110

Because of this, the fix:

commit ab99a87542f194f28e2364a42afbf9fb48b1c724
Author: Ofir Gal <ofir.gal at volumez.com>
Date: Fri Jun 7 10:27:44 2024 +0300
Subject: md/md-bitmap: fix writing non bitmap pages
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ab99a87542f194f28e2364a42afbf9fb48b1c724

needs to be reverted during your test runs, or you won't see the issue
reproduce.

You can use this ppa for updated kernels with the revert to trigger the issue:

https://launchpad.net/~mruffell/+archive/ubuntu/sf404844-revert

This can be reproduced by running blktests md/001 [1], which the author of the fix created to act as a regression test for this issue.

[1] https://github.com/osandov/blktests/commit/a24a7b462816fbad7dc6c175e53fcc764ad0a822

Deploy a fresh Noble VM, that has a scratch NVME disk.

$ sudo apt install build-essential fio
$ git clone https://github.com/osandov/blktests.git
$ cd blktests
$ make
$ echo "TEST_DEVS=(/dev/nvme0n1)" > config
$ sudo ./check md/001

The md/001 test will hang an affected system, and the above oops message will be visible in dmesg.

A test kernel is available in the following ppa:

https://launchpad.net/~mruffell/+archive/ubuntu/sf404844-test

This has both the fixes for this bug, and also bug 2075110. The issue will not
reproduce.

There is also a test kernel available with the fix for this bug present, and the
fix for bug 2075110 reverted, so you can see the impact of these patches only:

https://launchpad.net/~mruffell/+archive/ubuntu/sf404844-repro

This will also not reproduce the issue anymore.

[Where problems could occur]

What we are changing is rather simple. Instead of checking the first page and
assuming all the rest in the contiguous block are sendpage_ok(), we now
check each page in the contiguous block to see if all of them are sendpage_ok().

If any aren't, then we abort the write to the driver, and try again later. This
saves us time.

However, it does take longer to call sendpage_ok() on each of the pages in the
contiguous block, so there will be a minor performance hit.

Small performance hit for correctness should be okay.

Currently we are only applying to nvme-tcp and drbg subsystems. If a regression
were to occur, it would affect users of those subsystems only.

[Other info]

Upstream mailing list:
https://lore.kernel.org/all/20240718084515.3833733-1-ofir.gal@volumez.com/T/#u

Ofir Gal (3):
  net: introduce helper sendpages_ok()
  nvme-tcp: use sendpages_ok() instead of sendpage_ok()
  drbd: use sendpages_ok() instead of sendpage_ok()

 drivers/block/drbd/drbd_main.c |  2 +-
 drivers/nvme/host/tcp.c        |  2 +-
 include/linux/net.h            | 19 +++++++++++++++++++
 3 files changed, 21 insertions(+), 2 deletions(-)

-- 
2.45.2