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;
Message-ID: <de8d1830-9553-dd8f-8b12-308985c98b40@intel.com>
Date:   Tue, 28 Jun 2022 09:52:28 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.9.1
Subject: Re: [REGRESSION] Unable to unlock encrypted disk starting with kernel
 5.19-rc1+
Content-Language: en-US
To:     Borislav Petkov <bp@alien8.de>,
        Alexandre Messier <alex@me.ssier.org>
Cc:     linux-kernel@vger.kernel.org, tglx@linutronix.de,
        Andrew.Cooper3@citrix.com, mingo@redhat.com,
        dave.hansen@linux.intel.com, x86@kernel.org,
        regressions@lists.linux.dev
References: <6025678c-e94a-6966-e298-82fad658a889@me.ssier.org>
 <YrrHwxtD2dpts7PF@zn.tnic>
From:   Dave Hansen <dave.hansen@intel.com>
In-Reply-To: <YrrHwxtD2dpts7PF@zn.tnic>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

First of all, thank you for bisecting this!  I know those are a lot of work.

That XSAVEC patch modifies the AVX register save/restore code.  There is
a set of x86 AES acceleration instructions called AES-NI.  Those
instructions use the AVX registers.  So, it's at least a plausible
connection between that patch and your symptoms.  But, I don't think
anyone's been able to reproduce what you're seeing yet.

The kernel XSAVE buffer formats also differ slightly between AMD and
Intel.  That *should* be OK, but it might explain why I can't reproduce
this.

If you get a chance, could you apply this (ugly hackish) patch to the
userspace 'cryptsetup' utility and run it?

	https://sr71.net/~dave/intel/cryptsetup-memcmp.patch

On Ubuntu at least, it was as simple as:

	apt-get source cryptsetup
	apt-get build-dep cryptsetup
	cd cryptsetup-1.6.6
	./configure
	make

Then I could run:

	./src/cryptsetup benchmark --cipher=aes-xts --key-size=512
and
	./src/cryptsetup benchmark --cipher=aes-xts --key-size=256

With that patch applied, you should see some output like:

# ./src/cryptsetup benchmark --cipher=aes-xts --key-size=512
# Tests are approximate using memory only (no storage IO).
memcmp12: 0
memcmp23: 0
memcmp13: 0
memcmp12: -173
memcmp23: 173
memcmp13: 0
#  Algorithm | Key |  Encryption |  Decryption
     aes-xts   512b  4592.2 MiB/s  4192.0 MiB/s

The "memcmp13:" lines should both be 0.  That means that an encryption
and decryption cycle didn't change the data.  You *might* have to run
this in a loop if there's some kind of bad timing involved in triggering
the bug.

If you see a "memcmp13:" with something other than 0, that will narrow
things down and means we'll have a pretty quick reproducer that doesn't
involve luks which should speed things along.