On Thu, Jul 26, 2018 at 02:55:44PM +0100, Alex Bennée wrote:
>
> Dave Martin <Dave.Martin@xxxxxxx> writes:
>
> > On Wed, Jul 25, 2018 at 04:58:30PM +0100, Alex Bennée wrote:
> >>
> >> Dave Martin <Dave.Martin@xxxxxxx> writes:
> >>
> >> > The Arm SVE architecture defines registers that are up to 2048 bits
> >> > in size (with some possibility of further future expansion).
> >> >
> >> > In order to avoid the need for an excessively large number of
> >> > ioctls when saving and restoring a vcpu's registers, this patch
> >> > adds a #define to make support for individual 2048-bit registers
> >> > through the KVM_{GET,SET}_ONE_REG ioctl interface official. This
> >> > will allow each SVE register to be accessed in a single call.
> >> >
> >> > There are sufficient spare bits in the register id size field for
> >> > this change, so there is no ABI impact providing that
> >> > KVM_GET_REG_LIST does not enumerate any 2048-bit register unless
> >> > userspace explicitly opts in to the relevant architecture-specific
> >> > features.
> >>
> >> Does it? It's not in this patch and looking at the final tree:
> >>
> >> unsigned long kvm_arm_num_regs(struct kvm_vcpu *vcpu)
> >> {
> >> unsigned long res = 0;
> >>
> >> res += num_core_regs();
> >> res += num_sve_regs(vcpu);
> >> res += kvm_arm_num_sys_reg_descs(vcpu);
> >> res += kvm_arm_get_fw_num_regs(vcpu);
> >> res += NUM_TIMER_REGS;
> >>
> >> return res;
> >> }
> >>
> >>
> >> which leads to:
> >>
> >> static int enumerate_sve_regs(const struct kvm_vcpu *vcpu, u64 __user **uind)
> >> {
> >> unsigned int n, i;
> >> int err = 0;
> >> int total = 0;
> >> unsigned int slices;
> >>
> >> if (!vcpu_has_sve(&vcpu->arch))
> >> return 0;
> >>
> >> Which enumerates the SVE regs if vcpu_has_sve() which AFAICT is true if
> >> the host supports it, not if the user has requested it.
> >>
> >> I'll have to check what but given the indirection of kvm_one_reg I
> >> wonder if existing binaries might end up spamming a badly sized array
> >> when run on a new SVE supporting kernel?
> >
> > That shouldn't be the case: vcpu_has_sve() checks for the
> > KVM_ARM64_GUEST_HAS_SVE flag, which should only be set if userspace asks
> > for it.
> >
> > Give me a shout if this doesn't seem to be the case...
>
> Ahh I missed it the first time:
>
> if (system_supports_sve() &&
> test_bit(KVM_ARM_VCPU_SVE, vcpu->arch.features)) {
> vcpu->arch.flags |= KVM_ARM64_GUEST_HAS_SVE;
>
> And vcpu->arch.features is set by the user. However it will be set as
> unless you specify otherwise we use the results of probed:
>
> ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, init);
>
> so the user will get it by definition when they first run on SVE capable
> hardware.
AFAIK qemu and kvmtool don't currently propagate the feature flags from
KVM_ARM_PREFERRED_TARGET to KVM_VCPU_INIT. That would probably not be
the right thing to do, because if you don't know what a feature flag
means then you can't safely turn it on.
But in light of Andrew's comments I will need to review how this works.
The semantics of the feature bits are not well defined, so it may be
safer not to use them for enabling SVE. I was hoping we could choose
a meaning for them since they weren't previously used for anything,
but that may be optimistic.
Cheers
---Dave
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
