Re: [PATCH 0/2] Workaround for Cavium ThunderX2 erratum 219

From: Will Deacon <will@kernel.org>
To: Jayachandran Chandrasekharan Nair <jnair@marvell.com>
Cc: Tomasz Nowicki <tnowicki@marvell.com>,
	Catalin Marinas <catalin.marinas@arm.com>,
	Robert Richter <rrichter@marvell.com>,
	Marc Zyngier <maz@kernel.org>,
	"kvmarm@lists.cs.columbia.edu" <kvmarm@lists.cs.columbia.edu>,
	"linux-arm-kernel@lists.infradead.org"
	<linux-arm-kernel@lists.infradead.org>
Subject: Re: [PATCH 0/2] Workaround for Cavium ThunderX2 erratum 219
Date: Fri, 11 Oct 2019 11:44:55 +0100	[thread overview]
Message-ID: <20191011104454.d7dplgyjcnpfi5p3@willie-the-truck> (raw)
In-Reply-To: <1570790105-31829-1-git-send-email-jnair@marvell.com>

Hi JC,

Thanks for posting this.

On Fri, Oct 11, 2019 at 10:35:21AM +0000, Jayachandran Chandrasekharan Nair wrote:
> These two patches are based on the work by Marc Zyngier and addresses
> Cavium ThunderX2 erratum 219.
> 
> This erratum (originally reported by ARM folks) is from an interesting
> use of the prefetch instruction in the KPTI patchset. The prefetch
> was done between a TTBR change and the corresponding ISB, and this
> occasionally caused a crash on ThunderX2.
> 
> The first patch removes the troublesome prefetch for ThunderX2.
> The second patch addresses the case where the issue can be triggered
> from a guest kernel. The workaround in this case is to trap TTBR
> accesses by setting HCR_EL2.TVM for guests and doing the system
> register update from EL2 in a fast path.

FWIW, I was already planning to send the following to Linus:

https://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git/log/?h=errata/tx2-219

so please base any changes on top of that branch.

> Due to the nature of the erratum, the trap-and-emulate is only
> needed when SMT is enabled.
> 
> The overhead of trap-and-emulate is expected to be negligible on most
> workloads. A command line option kvm-arm.vm_msr_trap has been
> provided to override trapping on guest TTBR updates.  This is to
> address a very limited case where a user wants to run SMT enabled,
> with a trustworthy guest kernel, and wants to avoid the performance
> overhead associated with emulating the address translation register
> changes.

Do you have any performance data to show the impact of the workaround on
non-kpti guests? I don't think we can justify the inclusion of a cmdline
option for this without figures showing that it's really necessary.
Otherwise, the "very limited case" really is a niche scenario where the
CONFIG option can simply be disabled.

Cheers,

Will
_______________________________________________
kvmarm mailing list
kvmarm@lists.cs.columbia.edu
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm