Hi Kyle,
The issue is that the implementations of aead and skcipher in the QAT
driver are not properly supporting requests with the
CRYPTO_TFM_REQ_MAY_BACKLOG flag set.
If the HW queue is full, the driver returns -EBUSY [1] but does not
enqueues the request as dm-crypt expects [2]. Dm-crypt ends up waiting
indefinitely for a completion to a request that was never submitted,
therefore the stall.
This is not related to QATE-7495 'An incorrectly formatted request to
QAT can hang the entire QAT endpoint' [3], which occurs when a malformed
request is sent to the device.
I'm working at patch that resolves this problem. In the meanwhile a
workaround is to blacklist the qat_c3xxx.ko driver.
Regarding avoiding this issue on stable kernels. The usage of QAT with
dm-crypt was already disabled in kernel 5.10 for a different issue
(the driver allocates memory in the datapath).
The following patches implement the change:
7bcb2c99f8ed crypto: algapi - use common mechanism for inheriting flags
2eb27c11937e crypto: algapi - add NEED_FALLBACK to INHERITED_FLAGS
fbb6cda44190 crypto: algapi - introduce the flag CRYPTO_ALG_ALLOCATES_MEMORY
b8aa7dc5c753 crypto: drivers - set the flag CRYPTO_ALG_ALLOCATES_MEMORY
cd74693870fb dm crypt: don't use drivers that have CRYPTO_ALG_ALLOCATES_MEMORY
An option would be to send the patches above to stable, another is to wait
for a patch that fixes the problems in the QAT driver and send that to
stable.
@Herbert, what is the preferred approach here?
Thanks,
[1] https://elixir.bootlin.com/linux/latest/source/drivers/crypto/qat/qat_common/qat_algs.c#L1022
[2] https://elixir.bootlin.com/linux/latest/source/drivers/md/dm-crypt.c#L1584
[3] https://01.org/sites/default/files/downloads//336211qatsoftwareforlinux-rn-hwversion1.7021.pdf - page 25
--
Giovanni
On Sat, Feb 19, 2022 at 03:00:51PM -0800, Kyle Sanderson wrote:
> hi Dave,
>
> > This really sounds like broken hardware, not a kernel problem.
>
> It is indeed a hardware issue, specifically the intel qat crypto
> driver that's in-tree - the hardware is fine (see below). The IQAT
> eratta documentation states that if a request is not submitted
> properly it can stall the entire device. The remediation guidance from
> 2020 was "don't do that" and "don't allow unprivileged users access to
> the device". The in-tree driver is not implemented properly either for
> this SoC or board - I'm thinking it's related to QATE-7495.
>
> https://01.org/sites/default/files/downloads//336211qatsoftwareforlinux-rn-hwversion1.7021.pdf
>
> > This implies a dmcrypt level problem - XFS can't make progress is dmcrypt is not completing IOs.
>
> That's the weird part about it. Some bio's are completing, others are
> completely dropped, with some stalling forever. I had to use
> xfs_repair to get the volumes operational again. I lost a good deal of
> files and had to recover from backup after toggling the device back on
> on a production system (silly, I know).
>
> > Where are the XFS corruption reports that the subject implies is occurring?
>
> I think you're right, it's dm-crypt that's broken here, with
> ultimately the crypto driver causing this corruption. XFS being the
> edge to the end-user is taking the brunt of it. There's reports going
> back to late 2017 of significant issues with this mainlined stable
> driver.
>
> https://bugzilla.redhat.com/show_bug.cgi?id=1522962
> https://serverfault.com/questions/1010108/luks-hangs-on-centos-running-on-atom-c3758-cpu
> https://www.phoronix.com/forums/forum/software/distributions/1172231-fedora-33-s-enterprise-linux-next-effort-approved-testbed-for-raising-cpu-requirements-etc?p=1174560#post1174560
>
> Any guidance would be appreciated.
> Kyle.
> On Sat, Feb 19, 2022 at 1:03 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Feb 18, 2022 at 09:02:28PM -0800, Kyle Sanderson wrote:
> > > A2SDi-8C-HLN4F has IQAT enabled by default, when this device is
> > > attempted to be used by xfs (through dm-crypt) the entire kernel
> > > thread stalls forever. Multiple users have hit this over the years
> > > (through sporadic reporting) - I ended up trying ZFS and encryption
> > > wasn't an issue there at all because I guess they don't use this
> > > device. Returning to sanity (xfs), I was able to provision a dm-crypt
> > > volume no problem on the disk, however when running mkfs.xfs on the
> > > volume is what triggers the cascading failure (each request kills a
> > > kthread).
> >
> > Can you provide the full stack traces for these errors so we can see
> > exactly what this cascading failure looks like, please? In reality,
> > the stall messages some time after this are not interesting - it's
> > the first errors that cause the stall that need to be investigated.
> >
> > A good idea would be to provide the full storage stack decription
> > and hardware in use, as per:
> >
> > https://xfs.org/index.php/XFS_FAQ#Q:_What_information_should_I_include_when_reporting_a_problem.3F
> >
> > > Disabling IQAT on the south bridge results in a working
> > > system, however this is not the default configuration for the
> > > distribution of choice (Ubuntu 20.04.3 LTS), nor the motherboard. I'm
> > > convinced this never worked properly based on the lack of popularity
> > > for kernel encryption (crypto), and the embedded nature that
> > > SuperMicro has integrated this device in collaboration with intel as
> > > it looks like the primary usage is through external accelerator cards.
> >
> > This really sounds like broken hardware, not a kernel problem.
> >
> > > Kernels tried were from RHEL8 over a year ago, and this impacts the
> > > entirety of the 5.4 series on Ubuntu.
> > > Please CC me on replies as I'm not subscribed to all lists. CPU is C3758.
> >
> > [snip stalled kcryptd worker threads]
> >
> > This implies a dmcrypt level problem - XFS can't make progress is
> > dmcrypt is not completing IOs.
> >
> > Where are the XFS corruption reports that the subject implies is
> > occurring?
> >
> > Cheers,
> >
> > Dave.
> > --
> > Dave Chinner
> > david@xxxxxxxxxxxxx
