Re: [PATCH] x86/pass-through: avoid double IRQ unbind during domain cleanup

[Jan Beulich <jbeulich@suse.com>]

On 29.04.2020 10:26, Roger Pau Monné wrote:
> On Wed, Apr 29, 2020 at 09:37:11AM +0200, Jan Beulich wrote:
>> On 28.04.2020 18:14, Roger Pau Monné wrote:
>>> On Tue, Apr 28, 2020 at 02:21:48PM +0200, Jan Beulich wrote:
>>>> XEN_DOMCTL_destroydomain creates a continuation if domain_kill -ERESTARTs.
>>>> In that scenario, it is possible to receive multiple _pirq_guest_unbind
>>>> calls for the same pirq from domain_kill, if the pirq has not yet been
>>>> removed from the domain's pirq_tree, as:
>>>>   domain_kill()
>>>>     -> domain_relinquish_resources()
>>>>       -> pci_release_devices()
>>>>         -> pci_clean_dpci_irq()
>>>>           -> pirq_guest_unbind()
>>>>             -> __pirq_guest_unbind()
>>>>
>>>> Avoid recurring invocations of pirq_guest_unbind() by removing the pIRQ
>>>> from the tree being iterated after the first call there. In case such a
>>>> removed entry still has a softirq outstanding, record it and re-check
>>>> upon re-invocation.
>>>>
>>>> Reported-by: Varad Gautam <vrd@amazon.de>
>>>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
>>>> Tested-by: Varad Gautam <vrd@amazon.de>
>>>>
>>>> --- a/xen/arch/x86/irq.c
>>>> +++ b/xen/arch/x86/irq.c
>>>> @@ -1323,7 +1323,7 @@ void (pirq_cleanup_check)(struct pirq *p
>>>>      }
>>>>  
>>>>      if ( radix_tree_delete(&d->pirq_tree, pirq->pirq) != pirq )
>>>> -        BUG();
>>>> +        BUG_ON(!d->is_dying);
>>>
>>> I think to keep the previous behavior this should be:
>>>
>>> BUG_ON(!is_hvm_domain(d) || !d->is_dying);
>>>
>>> Since the pirqs will only be removed elsewhere if the domain is HVM?
>>
>> pirq_cleanup_check() is a generic hook, and hence I consider it more
>> correct to not have it behave differently in this regard for different
>> types of guests. IOW while it _may_ (didn't check) not be the case
>> today that this can be called multiple times even for PV guests, I'd
>> view this as legitimate behavior.
> 
> Previous to this patch pirq_cleanup_check couldn't be called multiple
> times, as it would result in the BUG triggering, that was true for
> both PV and HVM. Now that the removal of PIRQs from the tree is done
> elsewhere for HVM when the domain is dying the check needs to be
> relaxed for HVM at least. I would prefer if it was kept as-is for PV
> (since there's been no change in behavior for PV that could allow for
> multiple calls to pirq_cleanup_check), or else a small comment in the
> commit message would help clarify this is done on purpose.

I've added

"Note that pirq_cleanup_check() gets relaxed beyond what's strictly
 needed here, to avoid introducing an asymmetry there between HVM and PV
 guests."

Does this sound suitable to you?

Jan