Re: [RFC] kvm: reverse call order of kvm_arch_destroy_vm() and kvm_destroy_devices()

From: Anthony Krowiak
Date: Thu Aug 11 2022 - 10:41:00 EST


On 8/1/22 7:53 AM, Halil Pasic wrote:
On Wed, 27 Jul 2022 15:00:02 -0400
Anthony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:

Any Takers??????

On 7/5/22 2:54 PM, Tony Krowiak wrote:
There is a new requirement for s390 secure execution guests that the
hypervisor ensures all AP queues are reset and disassociated from the
KVM guest before the secure configuration is torn down. It is the
responsibility of the vfio_ap device driver to handle this.

Prior to commit ("vfio: remove VFIO_GROUP_NOTIFY_SET_KVM"),
the driver reset all AP queues passed through to a KVM guest when notified
that the KVM pointer was being set to NULL. Subsequently, the AP queues
are only reset when the fd for the mediated device used to pass the queues
through to the guest is closed (the vfio_ap_mdev_close_device() callback).
This is not a problem when userspace is well-behaved and uses the
KVM_DEV_VFIO_GROUP_DEL attribute to remove the VFIO group; however, if
userspace for some reason does not close the mdev fd, a secure execution
guest will tear down its configuration before the AP queues are
reset because the teardown is done in the kvm_arch_destroy_vm function
which is invoked prior to kvm_destroy_devices.
As Matt has pointed out: we did not have the guarantee we need prior
that commit. Please for the next version drop the digression about
the old behavior.

This patch proposes a simple solution; rather than introducing a new
notifier into vfio or callback into KVM, what aoubt reversing the order
in which the kvm_arch_destroy_vm and kvm_destroy_devices are called. In
some very limited testing (i.e., the automated regression tests for
the vfio_ap device driver) this did not seem to cause any problems.

The question remains, is there a good technical reason why the VM
is destroyed before the devices it is using? This is not intuitive, so
this is a request for comments on this proposed patch. The assumption
here is that the medev fd will get closed when the devices are destroyed.
I did some digging! The function and the corresponding mechanism was
introduced by 07f0a7bdec5c ("kvm: destroy emulated devices on VM
exit"). Before that patch we used to have ref-counting, and the refcound
got decremented in kvmppc_mpic_disconnect_vcpu() which in turn was
called by kvm_arch_vcpu_free(). So this was basically arch specific
stuff. For power (the patch came form power) the refcount was decremented
before calling kvmppc_core_vcpu_free(). So I conclude the old scheme
would have worked for us.

Since the patch does not state any technical reasons, my guess is, that
the choice was made somewhat arbitrarily under the assumption, that
there is no requirements or dependency with regards to the destruction
of devices or with regards towards severing the connection between
the devices and the VM. Under these assumptions the placement of
the invocation of kvm_destroy_devices after kvm_arch_destroy_vm()
did made sense, because if something that is destroyed in destroy_vm()
did hold a live reference to the device, this reference will be cleaned
up before kvm_destroy_devices() is invoked. So basically unless the
devices hold references to each other, things look good. If the
positions of kvm_arch_destroy_vm() and kvm_destroy_devices() are
changed, then we basically need to assume that nothing that is destroyed
in kvm_arch_destoy_vm() may logically hold a live reference (remember
the refcount is gone, but pointers may still exist) to a kvm device.
Does that hold? @Antony, maybe you can answer this question for us...


I do not have an answer for this without doing a deep dive into the code. I am not very familiar with the VM lifecycle. My hope was that someone who knows this area would respond to this RFC. I am copying the Signed-off-by email addresses for the patch (07f0a7bdec) you mentioned above; maybe they can provide some insight as to for their choice in ordering of the kvm_arch_destroy_vm() and kvm_destroy_devices() functions.



Otherwise I will continue the digging from here, eventually.

Also I have concerns about the following comments:

static void kvm_destroy_devices(struct kvm *kvm)
{
struct kvm_device *dev, *tmp;
/*
* We do not need to take the kvm->lock here, because nobody else
* has a reference to the struct kvm at this point and therefore
* cannot access the devices list anyhow.
[..]

Would this till hold when the order is changed?

struct kvm_device_ops {
[..]
/*
* Destroy is responsible for freeing dev.
*
* Destroy may be called before or after destructors are called
* on emulated I/O regions, depending on whether a reference is
* held by a vcpu or other kvm component that gets destroyed
* after the emulated I/O.
*/
void (*destroy)(struct kvm_device *dev);

This seems to document the order of things as is.

Btw I would like to understand more about the lifecycle of these
emulated I/O regions....

@Paolo: I believe this is ultimately your truff. I'm just digging
through the code, and the history to try to help along with this. We
definitely need a solution for our problem. We would very much appreciate
having your opinion!

Regards,
Halil

Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxx>
---
virt/kvm/kvm_main.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index a49df8988cd6..edaf2918be9b 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -1248,8 +1248,8 @@ static void kvm_destroy_vm(struct kvm *kvm)
#else
kvm_flush_shadow_all(kvm);
#endif
- kvm_arch_destroy_vm(kvm);
kvm_destroy_devices(kvm);
+ kvm_arch_destroy_vm(kvm);
for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
kvm_free_memslots(kvm, &kvm->__memslots[i][0]);
kvm_free_memslots(kvm, &kvm->__memslots[i][1]);