On Thu, May 23, 2019 at 05:21:00PM +1000, Alexey Kardashevskiy wrote: > > > On 23/05/2019 16:34, Paul Mackerras wrote: > > Recent reports of lockdep splats in the HV KVM code revealed that it > > was taking the kvm->lock mutex in several contexts where a vcpu mutex > > was already held. Lockdep has only started warning since I added code > > to take the vcpu mutexes in the XIVE device release functions, but > > since Documentation/virtual/kvm/locking.txt specifies that the vcpu > > mutexes nest inside kvm->lock, it seems that the new code is correct > > and it is most of the old uses of kvm->lock that are wrong. > > > > This series should fix the problems, by adding new mutexes that nest > > inside the vcpu mutexes and using them instead of kvm->lock. > > > I applied these 4, compiled, installed, rebooted, tried running a guest > (which failed because I also updated QEMU and its cli has changed), got > this. So VM was created and then destroyed without executing a single > instruction, if that matters. Looks like I need to remove the lockdep_assert_held(&kvm->arch.rtas_token_lock); in kvmppc_rtas_tokens_free(). We don't have the rtas_token_lock, but it doesn't matter because we are destroying the VM and nothing else has a reference to it by now. Paul.
