KVM: Acquire SCRU lock outside of irqfds.lock during assignment

Acquire SRCU outside of irqfds.lock so that the locking is symmetrical,
and add a comment explaining why on earth KVM holds SRCU for so long.

Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250522235223.3178519-3-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>

diff --git a/virt/kvm/eventfd.c b/virt/kvm/eventfd.c
index 0b655376734e444b5b1ec3e2855eb2aec8b18290..614f81cd37c11bb71e8d03f54c721790427b6581 100644 (file)
--- a/virt/kvm/eventfd.c
+++ b/virt/kvm/eventfd.c
@@ -394,6 +394,18 @@ kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
         */
        init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 
+       /*
+        * Set the irqfd routing and add it to KVM's list before registering
+        * the irqfd with the eventfd, so that the routing information is valid
+        * and stays valid, e.g. if there are GSI routing changes, prior to
+        * making the irqfd visible, i.e. before it might be signaled.
+        *
+        * Note, holding SRCU ensures a stable read of routing information, and
+        * also prevents irqfd_shutdown() from freeing the irqfd before it's
+        * fully initialized.
+        */
+       idx = srcu_read_lock(&kvm->irq_srcu);
+
        spin_lock_irq(&kvm->irqfds.lock);
 
        ret = 0;
@@ -402,11 +414,9 @@ kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
                        continue;
                /* This fd is used for another irq already. */
                ret = -EBUSY;
-               spin_unlock_irq(&kvm->irqfds.lock);
-               goto fail;
+               goto fail_duplicate;
        }
 
-       idx = srcu_read_lock(&kvm->irq_srcu);
        irqfd_update(kvm, irqfd);
 
        list_add_tail(&irqfd->list, &kvm->irqfds.items);
@@ -441,6 +451,9 @@ kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
        srcu_read_unlock(&kvm->irq_srcu, idx);
        return 0;
 
+fail_duplicate:
+       spin_unlock_irq(&kvm->irqfds.lock);
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 fail:
        if (irqfd->resampler)
                irqfd_resampler_shutdown(irqfd);