Guard writes to the global osvw_status and osvw_len variables with a
spinlock to ensure enabling virtualization on multiple CPUs in parallel
doesn't effectively drop any writes due to writing back stale data. Don't
bother taking the lock when the boot CPU doesn't support the feature, as
that check is constant for all CPUs, i.e. racing writes will always write
the same value (zero).
Note, the bug was inadvertently "fixed" by commit
9a798b1337af ("KVM:
Register cpuhp and syscore callbacks when enabling hardware"), which
effectively serialized calls to enable virtualization due to how the cpuhp
framework "brings up" CPU. But KVM shouldn't rely on the mechanics of
cphup to provide serialization.
Link: https://patch.msgid.link/20251113231420.1695919-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
* are published and we know what the new status bits are
*/
static uint64_t osvw_len = 4, osvw_status;
+static DEFINE_SPINLOCK(osvw_lock);
static DEFINE_PER_CPU(u64, current_tsc_ratio);
if (!err)
err = native_read_msr_safe(MSR_AMD64_OSVW_STATUS, &status);
- if (err)
+ guard(spinlock)(&osvw_lock);
+
+ if (err) {
osvw_status = osvw_len = 0;
- else {
+ } else {
if (len < osvw_len)
osvw_len = len;
osvw_status |= status;
osvw_status &= (1ULL << osvw_len) - 1;
}
- } else
+ } else {
osvw_status = osvw_len = 0;
+ }
svm_init_erratum_383();