On 30/04/2021 16:52, Roger Pau Monne wrote:
> @@ -822,3 +825,28 @@ int xc_cpu_policy_serialise(xc_interface *xch, const
> xc_cpu_policy_t p,
> errno = 0;
> return 0;
> }
> +
> +int xc_cpu_policy_get_cpuid(xc_interface *xch, const xc_cpu_policy_t policy,
> + uint32_t leaf, uint32_t subleaf,
> + xen_cpuid_leaf_t *out)
> +{
> + unsigned int nr_leaves = ARRAY_SIZE(policy->leaves);
> + xen_cpuid_leaf_t *tmp;
> + int rc;
> +
> + rc = xc_cpu_policy_serialise(xch, policy, policy->leaves, &nr_leaves,
> + NULL, 0);
> + if ( rc )
> + return rc;
Sorry for not spotting this last time.
You don't need to serialise. You can look up leaf/subleaf in O(1) time
from cpuid_policy, which was a design goal of the structure originally.
It is probably best to adapt most of the first switch statement in
guest_cpuid() to be a libx86 function. The asserts aren't massively
interesting to keep, and instead of messing around with nospec, just
have the function return a pointer into the cpuid_policy (or NULL), and
have a single block_speculation() in Xen. We'll also want a unit test
to go with this new function to check that out-of-range leaves don't
result in out-of-bounds reads.
~Andrew
