KVM: x86/xen: Fix potential deadlock in kvm_xen_update_runstate_guest()
authorDavid Woodhouse <dwmw@amazon.co.uk>
Wed, 11 Jan 2023 18:06:49 +0000 (18:06 +0000)
committerPaolo Bonzini <pbonzini@redhat.com>
Wed, 11 Jan 2023 18:32:21 +0000 (13:32 -0500)
The kvm_xen_update_runstate_guest() function can be called when the vCPU
is being scheduled out, from a preempt notifier. It *opportunistically*
updates the runstate area in the guest memory, if the gfn_to_pfn_cache
which caches the appropriate address is still valid.

If there is *contention* when it attempts to obtain gpc->lock, then
locking inside the priority inheritance checks may cause a deadlock.
Lockdep reports:

[13890.148997] Chain exists of:
                 &gpc->lock --> &p->pi_lock --> &rq->__lock

[13890.149002]  Possible unsafe locking scenario:

[13890.149003]        CPU0                    CPU1
[13890.149004]        ----                    ----
[13890.149005]   lock(&rq->__lock);
[13890.149007]                                lock(&p->pi_lock);
[13890.149009]                                lock(&rq->__lock);
[13890.149011]   lock(&gpc->lock);
[13890.149013]
                *** DEADLOCK ***

In the general case, if there's contention for a read lock on gpc->lock,
that's going to be because something else is either invalidating or
revalidating the cache. Either way, we've raced with seeing it in an
invalid state, in which case we would have aborted the opportunistic
update anyway.

So in the 'atomic' case when called from the preempt notifier, just
switch to using read_trylock() and avoid the PI handling altogether.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20230111180651.14394-2-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/xen.c

index bfa9809721b5b8275ca733bd6bae8274e1190a20..651f9c5b873ddcd11625a498c6fef1db516185a2 100644 (file)
@@ -271,7 +271,15 @@ static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
         * Attempt to obtain the GPC lock on *both* (if there are two)
         * gfn_to_pfn caches that cover the region.
         */
-       read_lock_irqsave(&gpc1->lock, flags);
+       if (atomic) {
+               local_irq_save(flags);
+               if (!read_trylock(&gpc1->lock)) {
+                       local_irq_restore(flags);
+                       return;
+               }
+       } else {
+               read_lock_irqsave(&gpc1->lock, flags);
+       }
        while (!kvm_gpc_check(gpc1, user_len1)) {
                read_unlock_irqrestore(&gpc1->lock, flags);
 
@@ -308,7 +316,14 @@ static void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, bool atomic)
                 * gpc1 lock to make lockdep shut up about it.
                 */
                lock_set_subclass(&gpc1->lock.dep_map, 1, _THIS_IP_);
-               read_lock(&gpc2->lock);
+               if (atomic) {
+                       if (!read_trylock(&gpc2->lock)) {
+                               read_unlock_irqrestore(&gpc1->lock, flags);
+                               return;
+                       }
+               } else {
+                       read_lock(&gpc2->lock);
+               }
 
                if (!kvm_gpc_check(gpc2, user_len2)) {
                        read_unlock(&gpc2->lock);