Re: [PATCH v2] KVM: kvm_io_bus_unregister_dev() should never fail

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Mar 23, 2017 at 06:24:19PM +0100, David Hildenbrand wrote:
> No caller currently checks the return value of
> kvm_io_bus_unregister_dev(). This is evil, as all callers silently go on
> freeing their device. A stale reference will remain in the io_bus,
> getting at least used again, when the iobus gets teared down on
> kvm_destroy_vm() - leading to use after free errors.
> 
> There is nothing the callers could do, except retrying over and over
> again.
> 
> So let's simply remove the bus altogether, print an error and make
> sure no one can access this broken bus again (returning -ENOMEM on any
> attempt to access it).
> 
> Fixes: e93f8a0f821e ("KVM: convert io_bus to SRCU")
> Cc: stable@vger.kernel.org # 3.4+
> Reported-by: Dmitry Vyukov <dvyukov@google.com>
> Signed-off-by: David Hildenbrand <david@redhat.com>
> ---
> 
> Based on kvm/queue, where we just got 2a108a4e7c1 ("KVM: x86: clear bus
> pointer when destroyed"), which added a check we need here.
> 
> v1 -> v2:
> - added a check in kvm_destroy_vm()
> - added a check in virt/kvm/eventfd.c
> 
> Using 'git grep -C 4 "kvm->buses"' should help to find all users. The other
> user in virt/kvm/eventfd.c should be fine.
> 
> ---
>  include/linux/kvm_host.h |  4 ++--
>  virt/kvm/eventfd.c       |  3 ++-
>  virt/kvm/kvm_main.c      | 42 +++++++++++++++++++++++++-----------------
>  3 files changed, 29 insertions(+), 20 deletions(-)
> 
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index 2c14ad9..d025074 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -162,8 +162,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
>  		    int len, void *val);
>  int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  			    int len, struct kvm_io_device *dev);
> -int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
> -			      struct kvm_io_device *dev);
> +void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
> +			       struct kvm_io_device *dev);
>  struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
>  					 gpa_t addr);
>  
> diff --git a/virt/kvm/eventfd.c b/virt/kvm/eventfd.c
> index a29786d..4d28a9d 100644
> --- a/virt/kvm/eventfd.c
> +++ b/virt/kvm/eventfd.c
> @@ -870,7 +870,8 @@ kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
>  			continue;
>  
>  		kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
> -		kvm->buses[bus_idx]->ioeventfd_count--;
> +		if (kvm->buses[bus_idx])
> +			kvm->buses[bus_idx]->ioeventfd_count--;
>  		ioeventfd_release(p);
>  		ret = 0;
>  		break;
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 7445566..ef1aa7f 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -728,7 +728,8 @@ static void kvm_destroy_vm(struct kvm *kvm)
>  	spin_unlock(&kvm_lock);
>  	kvm_free_irq_routing(kvm);
>  	for (i = 0; i < KVM_NR_BUSES; i++) {
> -		kvm_io_bus_destroy(kvm->buses[i]);
> +		if (kvm->buses[i])
> +			kvm_io_bus_destroy(kvm->buses[i]);
>  		kvm->buses[i] = NULL;
>  	}
>  	kvm_coalesced_mmio_free(kvm);
> @@ -3476,6 +3477,8 @@ int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
>  	};
>  
>  	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
> +	if (!bus)
> +		return -ENOMEM;
>  	r = __kvm_io_bus_write(vcpu, bus, &range, val);
>  	return r < 0 ? r : 0;
>  }
> @@ -3493,6 +3496,8 @@ int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
>  	};
>  
>  	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
> +	if (!bus)
> +		return -ENOMEM;
>  
>  	/* First try the device referenced by cookie. */
>  	if ((cookie >= 0) && (cookie < bus->dev_count) &&
> @@ -3543,6 +3548,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
>  	};
>  
>  	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
> +	if (!bus)
> +		return -ENOMEM;
>  	r = __kvm_io_bus_read(vcpu, bus, &range, val);
>  	return r < 0 ? r : 0;
>  }
> @@ -3555,6 +3562,9 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  	struct kvm_io_bus *new_bus, *bus;
>  
>  	bus = kvm->buses[bus_idx];
> +	if (!bus)
> +		return -ENOMEM;
> +
>  	/* exclude ioeventfd which is limited by maximum fd */
>  	if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
>  		return -ENOSPC;
> @@ -3574,45 +3584,41 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  }
>  
>  /* Caller must hold slots_lock. */
> -int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
> -			      struct kvm_io_device *dev)
> +void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
> +			       struct kvm_io_device *dev)
>  {
> -	int i, r;
> +	int i;
>  	struct kvm_io_bus *new_bus, *bus;
>  
>  	bus = kvm->buses[bus_idx];
> -
> -	/*
> -	 * It's possible the bus being released before hand. If so,
> -	 * we're done here.
> -	 */
>  	if (!bus)
> -		return 0;
> +		return;
>  
> -	r = -ENOENT;
>  	for (i = 0; i < bus->dev_count; i++)
>  		if (bus->range[i].dev == dev) {
> -			r = 0;
>  			break;
>  		}
>  
> -	if (r)
> -		return r;
> +	if (i == bus->dev_count)
> +		return;
>  
>  	new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
>  			  sizeof(struct kvm_io_range)), GFP_KERNEL);
> -	if (!new_bus)
> -		return -ENOMEM;
> +	if (!new_bus)  {
> +		pr_err("kvm: failed to shrink bus, removing it completely\n");
> +		goto broken;

The guest will fail in mysterious ways, if you do this (and
io_bus_unregister_dev can be called during runtime): in-kernel device
accesses will fail with unknown behaviour in the guest.

Can't you retry a handful of times with GFP_KERNEL before switching to GFP_ATOMIC? 
(which in case fails the machine is likely to be crashing soon).