On Thu, Aug 19, 2021, Peter Gonda wrote:

quoted

quoted

+static int svm_sev_lock_for_migration(struct kvm *kvm)
+{
+       struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
+       int ret;
+
+       /*
+        * Bail if this VM is already involved in a migration to avoid deadlock
+        * between two VMs trying to migrate to/from each other.
+        */
+       spin_lock(&sev->migration_lock);
+       if (sev->migration_in_progress)
+               ret = -EBUSY;
+       else {
+               /*
+                * Otherwise indicate VM is migrating and take the KVM lock.
+                */
+               sev->migration_in_progress = true;
+               mutex_lock(&kvm->lock);

Deadlock aside, mutex_lock() can sleep, which is not allowed while holding a
spinlock, i.e. this patch does not work.  That's my suggestion did the crazy
dance of "acquiring" a flag.

What I don't know is why on earth I suggested a global spinlock, a simple atomic
should work, e.g.

		if (atomic_cmpxchg_acquire(&sev->migration_in_progress, 0, 1))
			return -EBUSY;

		mutex_lock(&kvm->lock);

and on the backend...

		mutex_unlock(&kvm->lock); 

		atomic_set_release(&sev->migration_in_progress, 0);

quoted

quoted

+               ret = 0;
+       }
+       spin_unlock(&sev->migration_lock);
+
+       return ret;
+}
+
+static void svm_unlock_after_migration(struct kvm *kvm)
+{
+       struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
+
+       mutex_unlock(&kvm->lock);
+       WRITE_ONCE(sev->migration_in_progress, false);
+}
+

This entire locking scheme seems over-complicated to me. Can we simply
rely on `migration_lock` and get rid of `migration_in_progress`? I was
chatting about these patches with Peter, while he worked on this new
version. But he mentioned that this locking scheme had been suggested
by Sean in a previous review. Sean: what do you think? My rationale
was that this is called via a VM-level ioctl. So serializing the
entire code path on `migration_lock` seems fine. But maybe I'm missing
something?

Marc I think that only having the spin lock could result in
deadlocking. If userspace double migrated 2 VMs, A and B for
discussion, A could grab VM_A.spin_lock then VM_A.kvm_mutex. Meanwhile
B could grab VM_B.spin_lock and VM_B.kvm_mutex. Then A attempts to
grab VM_B.spin_lock and we have a deadlock. If the same happens with
the proposed scheme when A attempts to lock B, VM_B.spin_lock will be
open but the bool will mark the VM under migration so A will unlock
and bail. Sean originally proposed a global spin lock but I thought a
per kvm_sev_info struct would also be safe.

 
Close.  The issue is taking kvm->lock from both VM_A and VM_B.  If userspace
double migrates we'll end up with lock ordering A->B and B-A, so we need a way
to guarantee one of those wins.  My proposed solution is to use a flag as a sort
of one-off "try lock" to detect a mean userspace.