On 7/18/22 22:12, Sean Christopherson wrote:
On Mon, Jul 18, 2022, Paolo Bonzini wrote:
This needs to be fixed in the kernel because old QEMU/new KVM is supported.
I can't object to adding a quirk for this since KVM is breaking userspace, but
on
the KVM side we really need to stop "sanitizing" userspace inputs unless it puts
the host at risk, because inevitably it leads to needing a quirk.
The problem is not the sanitizing, it's that userspace literally cannot
know that this needs to be done because the feature bits are "backwards"
(1 = unavailable).
The right way to fix it is probably to use feature MSRs and, by default,
leave the features marked as unavailable. I'll think it through and
post a patch tomorrow for both KVM and QEMU (to enable PEBS).
But apart from that, where does Linux check MSR_IA32_MISC_ENABLE_BTS_UNAVAIL
and MSR_IA32_MISC_ENABLE_PEBS_UNAVAIL?
The kernel uses synthetic feature flags that are set by:
static void init_intel(struct cpuinfo_x86 *c)
if (boot_cpu_has(X86_FEATURE_DS)) {
unsigned int l1, l2;
rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
if (!(l1 & (1<<11)))
set_cpu_cap(c, X86_FEATURE_BTS);
if (!(l1 & (1<<12)))
set_cpu_cap(c, X86_FEATURE_PEBS);
}
Gah, shift constants are evil. I sent
https://lore.kernel.org/all/20220719174714.2410374-1-pbonz...@redhat.com/ to
clean this up.
Paolo
and consumed by:
void __init intel_ds_init(void)
/*
* No support for 32bit formats
*/
if (!boot_cpu_has(X86_FEATURE_DTES64))
return;
x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
