On Fri, Dec 14, 2018 at 12:23:48AM -0500, David Long wrote:
> On 11/23/18 6:09 AM, Marc Zyngier wrote:
> >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@xxxxxxxxxx>
> >>>>>
> >>>>>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@xxxxxxxxxx>
> >>>>
> >>>>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.
> >
>
> Marc,
>
> Sorry to be slow in getting back to you on this.
>
> As I've been looking at the six or so virtualization-related patches I
> excluded from the backports for less ancient release streams, for the v4.4
> stream, I'm having a hard time believing you want the "KVM" patches left
> out. Just their subject lines sure make them sound like they would have the
> guest impact you are worried about. Here's the ones that worry me from the
> v4.9 backport:
>
> [PATCH 4.9 11/24] ARM: KVM: invalidate BTB on guest exit for Cortex-A12/A17
> [PATCH 4.9 12/24] ARM: KVM: invalidate icache on guest exit for Cortex-A15
> [PATCH 4.9 13/24] ARM: spectre-v2: KVM: invalidate icache on guest exit for
> Brahma B15
>
> Are these really not interesting for v4.4, or am I misunderstanding which
> patches you meant?
I'm getting the impression that this has completely de-railed the
backporting effort.
I'm also wondering if this is actually a good idea. The 4.4 and 4.9
32-bit kernels implement a particularly simple hypervisor, where the
hypervisor is nothing more than:
__hyp_stub_do_trap:
cmp r0, #-1
mrceq p15, 4, r0, c12, c0, 0 @ get HVBAR
mcrne p15, 4, r0, c12, c0, 0 @ set HVBAR
__ERET
Would it not be sane to assume that a v4.4 host kernel would support
a v4.4 guest kernel under virtualisation? Since this same code is in
v4.9, the same is true there (it got changed in v4.12).
If we backport the SMCCC_ARCH_WORKAROUND_1 bits, we end up with an
incompatibility between the hypervisor code in v4.4 and the guest
kernel code - the guest kernel will attempt to make a hypervisor call
with r0=SMCCC_ARCH_WORKAROUND_1, which is 0x80008000. This will end
up setting the HVBAR to that value, which is clearly not intended,
which will end up pointing the hypervisor vectors to that address.
Of course, the same will be true if we run a pre-4.12 host kernel
under a post-4.12 kernel with the SMCCC_ARCH_WORKAROUND_1 hack.
I forget whether the stable kernels picked up on the changes for this
hypervisor or not - if not, it isn't a trivial "just make the guest
use the SMCCC_ARCH_WORKAROUND_1 call."
I think Marc's of the opinion that he's the only one who runs kernels
under a 32-bit host kernel - how sure can we be that no one out there
other than Marc does this?
Apart from that, I'm getting very concerned about the amount of time
the backporting is taking - there are about 40 patches in all, and I
believe only around half that have so far been backported to any of
the stable kernels. We seem to be hung-up on dealing with v4.4 that
other stable kernels aren't getting the other fixes backported.
We've seen the stable people attempt to pick up patches from the series
that make no sense on their own, because the real Spectre fixes don't
apply because of previous Spectre patches that are missing.
All the time, we have people using the stable kernels without Spectre
mitigation in place - and Spectre will have been known about for a
year next month.
--
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up
