[PATCH 02/27] arm64: KVM: Hide unsupported AArch64 CPU features from guests

[Dave Martin]

On Thu, Aug 17, 2017 at 09:45:51AM +0100, Marc Zyngier wrote:
> On 16/08/17 21:32, Dave Martin wrote:
> > On Wed, Aug 16, 2017 at 12:10:38PM +0100, Marc Zyngier wrote:
> >> On 09/08/17 13:05, Dave Martin wrote:
> >>> Currently, a guest kernel sees the true CPU feature registers
> >>> (ID_*_EL1) when it reads them using MRS instructions.  This means
> >>> that the guest will observe features that are present in the
> >>> hardware but the host doesn't understand or doesn't provide support
> >>> for.  A guest may legimitately try to use such a feature as per the
> >>> architecture, but use of the feature may trap instead of working
> >>> normally, triggering undef injection into the guest.
> >>>
> >>> This is not a problem for the host, but the guest may go wrong when
> >>> running on newer hardware than the host knows about.
> >>>
> >>> This patch hides from guest VMs any AArch64-specific CPU features
> >>> that the host doesn't support, by exposing to the guest the
> >>> sanitised versions of the registers computed by the cpufeatures
> >>> framework, instead of the true hardware registers.  To achieve
> >>> this, HCR_EL2.TID3 is now set for AArch64 guests, and emulation
> >>> code is added to KVM to report the sanitised versions of the
> >>> affected registers in response to MRS and register reads from
> >>> userspace.
> >>>
> >>> The affected registers are removed from invariant_sys_regs[] (since
> >>> the invariant_sys_regs handling is no longer quite correct for
> >>> them) and added to sys_reg_desgs[], with appropriate access(),
> >>> get_user() and set_user() methods.  No runtime vcpu storage is
> >>> allocated for the registers: instead, they are read on demand from
> >>> the cpufeatures framework.  This may need modification in the
> >>> future if there is a need for userspace to customise the features
> >>> visible to the guest.
> >>>
> >>> Attempts by userspace to write the registers are handled similarly
> >>> to the current invariant_sys_regs handling: writes are permitted,
> >>> but only if they don't attempt to change the value.  This is
> >>> sufficient to support VM snapshot/restore from userspace.
> >>>
> >>> Because of the additional registers, restoring a VM on an older
> >>> kernel may not work unless userspace knows how to handle the extra
> >>> VM registers exposed to the KVM user ABI by this patch.
> >>>
> >>> Under the principle of least damage, this patch makes no attempt to
> >>> handle any of the other registers currently in
> >>> invariant_sys_regs[], or to emulate registers for AArch32: however,
> >>> these could be handled in a similar way in future, as necessary.
> >>>
> >>> Signed-off-by: Dave Martin <Dave.Martin at arm.com>
> >>> ---
> >>>  arch/arm64/kvm/hyp/switch.c |   6 ++
> >>>  arch/arm64/kvm/sys_regs.c   | 224 +++++++++++++++++++++++++++++++++++---------
> >>>  2 files changed, 185 insertions(+), 45 deletions(-)

[...]

> >>> +static bool __access_id_reg(struct kvm_vcpu *vcpu,
> >>> +			    struct sys_reg_params *p,
> >>> +			    const struct sys_reg_desc const *r,
> >>> +			    bool raz)
> >>> +{
> >>> +	if (p->is_write) {
> >>> +		kvm_inject_undefined(vcpu);
> >>> +		return false;
> >>> +	}
> >>
> >> I don't think this is supposed to happen (should have UNDEF-ed at EL1).
> >> You can call write_to_read_only() in that case, which will spit out a
> >> warning and inject the exception.
> > 
> > I'll check this -- sounds about right.
> > 
> > If is should never happen, should I just delete that code or BUG()?  I
> > notice a BUG_ON() for a similar situation in access_vm_reg() for example.
> > 
> > Or do we not quite trust hardware not to get this wrong?
> > (It feels like the kind of thing that could slip through validation
> > and/or would be considered not worth a respin, but it seems wrong to
> > work around a theoretical hardware bug before it's confirmed to exist,
> > unless we think for some reason that it's really likely.)
> 
> That's the way we handle this for the rest of the accessors. We used to
> have a BUG_ON(), but it is pretty silly to kill the whole system for
> such a small deviation from the architecture. And maybe it is useless,
> but it doesn't hurt either.

OK, that makes sense -- I'll follow the precedent here and call
write_to_read_only() if this happens.

> >>> +
> >>> +	p->regval = read_id_reg(r, raz);
> >>> +	return true;
> >>> +}
> > 
> > [...]
> > 
> >>> @@ -944,6 +1073,32 @@ static const struct sys_reg_desc sys_reg_descs[] = {
> >>>  	{ SYS_DESC(SYS_DBGVCR32_EL2), NULL, reset_val, DBGVCR32_EL2, 0 },
> >>>  
> >>>  	{ SYS_DESC(SYS_MPIDR_EL1), NULL, reset_mpidr, MPIDR_EL1 },
> >>> +
> >>> +	/*
> >>> +	 * All non-RAZ feature registers listed here must also be
> >>> +	 * present in arm64_ftr_regs[].
> >>> +	 */
> >>> +
> >>> +	/* AArch64 mappings of the AArch32 ID registers */
> >>> +	/* ID_AFR0_EL1 not exposed to guests for now */
> >>> +	ID(PFR0),	ID(PFR1),	ID(DFR0),	_ID_RAZ(1,3),
> >>> +	ID(MMFR0),	ID(MMFR1),	ID(MMFR2),	ID(MMFR3),
> >>> +	ID(ISAR0),	ID(ISAR1),	ID(ISAR2),	ID(ISAR3),
> >>> +	ID(ISAR4),	ID(ISAR5),	ID(MMFR4),	_ID_RAZ(2,7),
> >>> +	_ID(MVFR0),	_ID(MVFR1),	_ID(MVFR2),	_ID_RAZ(3,3),
> >>> +	_ID_RAZ(3,4),	_ID_RAZ(3,5),	_ID_RAZ(3,6),	_ID_RAZ(3,7),
> >>
> >> #bikeshed:
> >>
> >> OK, this is giving me a headache. Too many variants with similar names.
> >> ID and _ID
> >> I'm also slightly perplexed with the amalgamation of RAZ because the
> >> register is not defined yet in the architecture, and RAZ because we
> >> don't expose it (like ID_AFR0_EL1). Yes, there is a number of comments
> > 
> > This "raz" overloading already seems present in other places, such as the
> > cpufeatures code.  (Which is not necessarily a good reason for adding
> > more of it...)
> > 
> >> to document that, but the code should aim to be be self-documenting. How
> >> about IDRAZ() for those we want to "hide", and IDRSV for encodings that
> >> are not allocated yet? It would look like this:
> >>
> >> 	IDREG(ID_PFR0),		IDREG(ID_PFR1),		IDREG(ID_DFR0),
> >> 	IDRAZ(ID_AFR0),		IDREG(ID_MMFR0),	IDREG(ID_MMFR1),
> >> 	IDREG(ID_MMFR2),	IDREG(ID_MMFR3),	IDREG(ID_ISAR0),
> >> 	IDREG(ID_ISAR1),	IDREG(ID_ISAR2),	IDREG(ID_ISAR3),
> >> 	IDREG(ID_ISAR4),	IDREG(ID_ISAR5),	IDREG(ID_MMFR4),
> >> 	IDRSV(2,7),		IDREG(MVFR0),		IDREG(MVFR1),
> >> 	IDREG(MVFR2),		IDRSV(3,3),		IDRSV(3,4),	
> >> 	IDRSV(3,5),		IDRSV(3,6),		IDRSV(3,7),
> >>
> >> Yes, only 3 a line. Lines are cheap. And yes, they also have similar
> >> names, but I said #bikeshed.
> > 
> > So, point taken, but the main reason for making this a table was to make
> > it easy to see by eye how the entries map to the encoding while hacking
> > this up, which helped me to make sure no entries were missed or in the
> > wrong place etc.
> > 
> > With 3 entries per line that visual map is lost, and with 2 entries per
> > line it's debatable whether it's worth having multiple entries per line
> > at all.
> 
> Let's be clear. I don't care at all about the number of entries per
> line. I can widen my editor to 200 columns if I need to. If you think 4
> is the way, keep it to 4.
> 
> My point is about the readability of both the macros and the
> identifiers, and your initial proposal did seem to lack on both counts.

Agreed, I was just trying to explain why it ended up that way in the
first place, and I'm happy to change it.

> > So now that the table exists maybe we should just have one entry per
> > line like everything else -- it really depends on which option you think
> > is best for ongoing maintenance.
> > 
> > 
> > Having one per line allows much less cryptic names, allowing the
> > temptingly short but ambiguous "RAZ" to be avoided:
> > 
> > 	ID_SANITISED(ID_ISAR5),
> > 	ID_RAZ_FOR_GUEST(ID_AFR0),
> > 	ID_UNALLOCATED(crm, op2)
> > 
> > With a whole line and different lengths, it's easier to pick out
> > the different cases by eye, so they don't all look like IDRXX (and are a
> > more tasteful colour perhaps).
> > 
> > Blank lines and/or comments can split the list into sensible blocks for
> > readability if needed.
> > 
> > If you're happy with naming along those broad lines then I'm happy to
> > see what it looks like.
> 
> Sure. If you're happy with that, so am I.
> 
> >>> +
> >>> +	/* AArch64 ID registers */
> >>> +	ID(AA64PFR0),	ID(AA64PFR1),	_ID_RAZ(4,2),	_ID_RAZ(4,3),
> >>> +	_ID_RAZ(4,4),	_ID_RAZ(4,5),	_ID_RAZ(4,6),	_ID_RAZ(4,7),
> >>> +	ID(AA64DFR0),	ID(AA64DFR1),	_ID_RAZ(5,2),	_ID_RAZ(5,3),
> >>> +	/* ID_AA64AFR0_EL1 and ID_AA64AFR0_EL1 not exposed to guests for now */
> > 
> > There are no sysreg definitions for IA_AA64AFR{0,1}_EL1 yet.
> > 
> > If we want to macroise those rather than just commenting, I guess
> > they'll need adding in sysreg.h.  I'd prefer not to imply these are
> > "unallocated" or similar when the architecture does define them.
> > 
> > Can I take it there's no problem with zombie entries in sysreg.h so long
> > as they're at least referenced somewhere?  (Arguably they wouldn't be
> > zombies then, but hopefully you see what I mean.)
> 
> That'd be the right thing to do. The register exists, and KVM handles it
> by returning 0 when a guest reads it. So I'd argue that it *must* be
> defined in sysreg.h, and given its full visibility in that table.

OK, sounds good -- I'll reroll with that change.

Cheers
---Dave