Hi Hanjun,
On 23/11/2018 09:40, Hanjun Guo wrote:
Hi Marc,
On 2018/11/23 17:10, Marc Zyngier wrote:
On 23/11/2018 01:25, Hanjun Guo wrote:
On 2018/10/31 22:04, David Long wrote:
From: "David A. Long" dave.long@linaro.org
V4.4 backport of spectre patches from Russell M. King's spectre branch. Most KVM patches are excluded. Patches not yet in upstream are excluded.
I tested this patch set on top of stable 4.4 kernel, running on boards with A9 and A15 based Hisilicon SoCs, didn't see boot regression and other function regressions in our CI system,
Tested-by: Hanjun Guo hanjun.guo@linaro.org
Since this patch set didn't include PSCI based hardening for arm32, so bugfix 6282e916f774 ("ARM: 8809/1: proc-v7: fix Thumb annotation of cpu_v7_hvc_switch_mm") is not needed for this patch set and this patch set is in a good shape I think. So what's the plan for this patch set?
Well, not having these patches means that a 32bit kernel won't be get any Spectre-v2 mitigation when run as a guest on an arm64 platform. It turns out that this is a pretty common setup among people building large pieces of SW, such as distributions.
I almost miss this point, that makes sense to me :)
Not having KVM host mitigation on 32bit ARM is probably OK (let's face it, I'm the only user), but not mitigating it as a guest doesn't seem completely OK to me.
We are working on a patch set which is backported from mainline to fix ARM64 spectre-v1, spectre-v2 and SSBD for stable 4.4 kernel, and that patch set (almost done) has PSCI patches which is needed by 32bit ARM, so how about posting those ARM64 spectre fixes then backport all those kvm patches for 32bit ARM spectre fix as well?
I'm not sure I get what you mean by PSCI. PSCI is not involved in the Spectre-v2 mitigation, as we use a specially designed SMC call, relying on the SMCCC 1.1 infrastructure. Maybe it is what you're referring to here?
Again, I don't think it is worth the hassle backporting the KVM patches. What I'd like to see is the guest (and bare metal) support code that uses the ARCH_WORKAROUND_1 SMCCC 1.1 infrastructure.
I also don't think it is worth creating an artificial dependency between the two architectures. Yes, some patches are common (the SMCCC infrastructure), but that can be easily be solved at merge time. My vote would be for David to carry the relevant patches in this series.
Thanks,
M.