Re: [PATCH] arm64: Fix compile error with KVM and !HARDEN_BRANCH_PREDICTOR

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On Fri, 29 Nov 2019 at 08:21, Marc Zyngier <maz@kernel.org> wrote:
>
> On Thu, 28 Nov 2019 17:20:20 +0000,
>
> [fixing Will's email address]
>
> Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote:
> >
> > (+ Marc)
> >
> >
> >
> > On Wed, 27 Nov 2019 at 19:10, <minyard@acm.org> wrote:
> > >
> > > From: Corey Minyard <cminyard@mvista.com>
> > >
> > > When compiling with KVM enabled and without HARDEN_BRANCH_PREDICTOR,
> > > the following compile error happens:
> > >
> > > arch/arm64/kernel/cpu_errata.c:92:23:
> > > error: '__bp_harden_hyp_vecs_start' undeclared (first use in this function);
> > > did you mean 'hyp_vecs_start'?
> > >   void *dst = lm_alias(__bp_harden_hyp_vecs_start + slot * SZ_2K);
> > >
> > > Some ifdefs were removed by 3e91f3eacc91d9 "arm64: Always enable
> > > spectre-v2 vulnerability detection" for CONFIG_HARDEN_BRANCH_PREDICTOR,
> > > but __bp_harden_hyp_vecs_start is only defined if that config is
> > > enabled.
> > >
> > > Add CONFIG_HARDEN_BRANCH_PREDICTOR to the #if that has CONFIG_KVM,
> > > It looks like you need both of those for that code to be valid.
> > >
> > > Fixes: 3e91f3eacc91d9 "arm64: Always enable spectre-v2 vulnerability detection"
> > > Cc: Andre Przywara <andre.przywara@arm.com>
> > > Cc: Catalin Marinas <catalin.marinas@arm.com>
> > > Cc: Stefan Wahren <stefan.wahren@i2se.com>
> > > Cc: Will Deacon <will.deacon@arm.com>
> > > Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
> > > Signed-off-by: Corey Minyard <cminyard@mvista.com>
> > > ---
> > > This is for 4.14, I'm not sure if it is needed for other kernels.
> > >
> > > It is not needed in master because a new config item was added,
> > > CONFIG_KVM_INDIRECT_VECTORS, that depends on KVM and
> > > HARDEN_BRANCH_PREDICTOR being configured.  I looked at pulling the
> > > patches that add the required changes, and they make a lot of
> > > changes.  This change is the simple fix, but I'm not sure if we want to
> > > pull all those other changes into 4.14 and whatever other kernels
> > > are required.
> > >
> >
> > I agree that backporting this cleanly is going to be problematic,
> > since it pulls in the entire EL2 VA randomization feature with all its
> > prerequisites.
> >
> > Backporting the following set could be done fairly cleanly, and fixes
> > the problem at build time, but unfortunately, it causes a boot time
> > crash (see below)
> >
> > 97eca4d2bfec (HEAD -> linux-4.14.y) arm64: Move the content of bpi.S
> > to hyp-entry.S
> > 514dd33114c6 arm64: Make BP hardening slot counter available
> > d7ddf3ae9470 arm64; insn: Add encoder for the EXTR instruction
> > c0b2c4e56e10 arm64: KVM: Introduce EL2 VA randomisation
> > 56ab0a87c737 arm64: KVM: Dynamically compute the HYP VA mask
> > d92c02628dfb arm64: KVM: Allow far branches from vector slots to the
> > main vectors
> > 7adec01c9674 arm64: cpufeatures: Drop the ARM64_HYP_OFFSET_LOW feature flag
> > 1095e4fc3134 arm64: KVM: Move stashing of x0/x1 into the vector code itself
> > bb2e1aceb423 arm64: KVM: Dynamically patch the kernel/hyp VA mask
> > 6f0ccfc451be arm64: KVM: Reserve 4 additional instructions in the BPI template
> > bf425ffee07a arm64: insn: Add encoder for bitwise operations using literals
> > 41dda94d1a9f arm64: insn: Add N immediate encoding
> > 3225668ebe00 arm64: KVM: Move BP hardening vectors into .hyp.text section
> >
> > Marc?
>
> You need at least these:
>
> 1bb32a44aea1 KVM: arm/arm64: Keep GICv2 HYP VAs in kvm_vgic_global_state
> 44a497abd621 KVM: arm/arm64: Do not use kern_hyp_va() with kvm_vgic_global_state
>

Yeah, I only did a fairly mechanical backport based on the actual
diffs depend on each other, but I did spot those two as possibly
related.

> But that's definitely not enough to fix the crash.
>
> >
> > [    0.062126] CPU: All CPU(s) started at EL1
> > [    0.063109] alternatives: patching kernel code
> > [    0.064228] random: get_random_u64 called from
> > compute_layout+0x94/0xe8 with crng_init=0
> > [    0.066313] aarch64_insn_gen_add_sub_imm: invalid immediate encoding 1904640
>
> OK, that one is really bizarre. This value (tag_val) is supposed to be
> a small value (only 12 significant bits out of 24 at any given time),
> and it is not (0x1D1000). So somehow compute_instruction() is not
> doing the right thing.
>
> Do you have a tree somewhere with this patches?
>

Sure, thanks for having a look.

https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm64-v4.14-backport%2b%2b

Another thing I found bizarre is that we actually run this code when
all CPUs boot at EL1. Or is that intended?

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel