Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <CACsaVZ+mt3CfdXV0_yJh7d50tRcGcRZ12j3n6-hoX2cz3+njsg@mail.gmail.com>
 <20220219210354.GF59715@dread.disaster.area> <CACsaVZ+LZUebtsGuiKhNV_No8fNLTv5kJywFKOigieB1cZcKUw@mail.gmail.com>
 <YhN76/ONC9qgIKQc@silpixa00400314>
In-Reply-To: <YhN76/ONC9qgIKQc@silpixa00400314>
From:   Kyle Sanderson <kyle.leet@gmail.com>
Date:   Mon, 28 Feb 2022 00:18:33 -0800
Message-ID: <CACsaVZJFane88cXxG_E1VkcMcJm8YVN+GDqQ2+tRYNpCf+m8zA@mail.gmail.com>
Subject: Re: Intel QAT on A2SDi-8C-HLN4F causes massive data corruption with
 dm-crypt + xfs
To:     Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Cc:     herbert@gondor.apana.org.au, Dave Chinner <david@fromorbit.com>,
        qat-linux@intel.com, Linux-Kernal <linux-kernel@vger.kernel.org>,
        linux-xfs@vger.kernel.org, linux-crypto@vger.kernel.org,
        dm-devel@redhat.com,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Greg KH <gregkh@linuxfoundation.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

> 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.

Thanks Giovanni. Joel (from Intel) reached out to me out of band to
try and sell me further on QAT but wasn't able to follow-up on any
questions (like - how is the device actually used, how can I
personally help, etc).

> 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.

Makes sense - this kernel driver has been destroying users for many
years. I'm disappointed that this critical bricking failure isn't
searchable for others.

> 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.

That's nice to hear that the device itself isn't dying, but it's been
completely destroying systems for years which itself is a DoS.

> I'm working at patch that resolves this problem. In the meanwhile a workaround is to blacklist the qat_c3xxx.ko driver.

I'm not writing this facetiously, but this driver has caused
incredible harm over the past 5+ years and seems to continue to do so.
As there's no patch proposed yet, I'm looking for the driver to be
completely removed from the tree as it's presently a pure marketing
campaign that's caused significant harm. If the marketing benefits
(like accelerated crypto + hashing) aren't there when the accelerated
instruction set was pulled from these integrated chips - the driver
continues to serve no purpose for consumers beyond damage. Disabling
the core I/O bits in December 2020 to make this barely work continues
to promote this as a side project as it was never resolved in the
driver.

If I can test patches, or assist with the removal of this present
in-tree malware I'm happy to help.

Kyle.


On Mon, Feb 21, 2022 at 3:48 AM Giovanni Cabiddu
<giovanni.cabiddu@intel.com> wrote:
>
> 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@fromorbit.com> 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@fromorbit.com