Re: [PATCH 4.4 119/268] xen/pirq: fix error path cleanup when binding MSIs

Re: [PATCH 4.4 119/268] xen/pirq: fix error path cleanup when binding MSIs

[Ben Hutchings]

On Wed, 2018-02-28 at 09:19 +0000, Roger Pau Monne wrote:
> From: Roger Pau Monne <roger.pau@citrix.com>
> 
> [ Upstream commit 910f8befdf5bccf25287d9f1743e3e546bcb7ce0 ]
> 
> Current cleanup in the error path of xen_bind_pirq_msi_to_irq is
> wrong. First of all there's an off-by-one in the cleanup loop, which
> can lead to unbinding wrong IRQs.
> 
> Secondly IRQs not bound won't be freed, thus leaking IRQ numbers.
> 
> Note that there's no need to differentiate between bound and unbound
> IRQs when freeing them, __unbind_from_irq will deal with both of them
> correctly.

It appears to me that it is safe to call __unbind_from_irq() after
xen_irq_info_common_setup() fails, but *not* if the latter hasn't been
called at all.  In that case the IRQ type will still be set to
IRQT_UNBOUND and this will trigger the BUG_ON() in __unbind_from_irq().

[...]
> --- a/drivers/xen/events/events_base.c
> +++ b/drivers/xen/events/events_base.c
> @@ -764,8 +764,8 @@ out:
>  	mutex_unlock(&irq_mapping_update_lock);
>  	return irq;
>  error_irq:
> -	for (; i >= 0; i--)
> -		__unbind_from_irq(irq + i);
> +	while (nvec--)
> +		__unbind_from_irq(irq + nvec);

If nvec > 1, and xen_irq_info_pirq_setup() fails for i != nvec - 1,
then we reach here without having called xen_irq_info_common_setup()
for all these IRQs.

In that case, I think we will still want to call xen_free_irq() for all
IRQs.  So maybe the fix would be to remove the BUG_ON() in
__unbind_from_irq()?

Ben.

>  	mutex_unlock(&irq_mapping_update_lock);
>  	return ret;
>  }

-- 
Ben Hutchings, Software Developer                         Codethink Ltd
https://www.codethink.co.uk/                 Dale House, 35 Dale Street
                                     Manchester, M1 2HF, United Kingdom

Re: [PATCH 4.4 119/268] xen/pirq: fix error path cleanup when binding MSIs

[Roger Pau Monné]

On Wed, Jun 13, 2018 at 07:48:50PM +0100, Ben Hutchings wrote:
> On Wed, 2018-02-28 at 09:19 +0000, Roger Pau Monne wrote:
> > From: Roger Pau Monne <roger.pau@citrix.com>
> > 
> > [ Upstream commit 910f8befdf5bccf25287d9f1743e3e546bcb7ce0 ]
> > 
> > Current cleanup in the error path of xen_bind_pirq_msi_to_irq is
> > wrong. First of all there's an off-by-one in the cleanup loop, which
> > can lead to unbinding wrong IRQs.
> > 
> > Secondly IRQs not bound won't be freed, thus leaking IRQ numbers.
> > 
> > Note that there's no need to differentiate between bound and unbound
> > IRQs when freeing them, __unbind_from_irq will deal with both of them
> > correctly.
> 
> It appears to me that it is safe to call __unbind_from_irq() after
> xen_irq_info_common_setup() fails, but *not* if the latter hasn't been
> called at all.  In that case the IRQ type will still be set to
> IRQT_UNBOUND and this will trigger the BUG_ON() in __unbind_from_irq().
> 
> [...]
> > --- a/drivers/xen/events/events_base.c
> > +++ b/drivers/xen/events/events_base.c
> > @@ -764,8 +764,8 @@ out:
> >  	mutex_unlock(&irq_mapping_update_lock);
> >  	return irq;
> >  error_irq:
> > -	for (; i >= 0; i--)
> > -		__unbind_from_irq(irq + i);
> > +	while (nvec--)
> > +		__unbind_from_irq(irq + nvec);
> 
> If nvec > 1, and xen_irq_info_pirq_setup() fails for i != nvec - 1,
> then we reach here without having called xen_irq_info_common_setup()
> for all these IRQs.
> 
> In that case, I think we will still want to call xen_free_irq() for all
> IRQs.  So maybe the fix would be to remove the BUG_ON() in
> __unbind_from_irq()?

I think your analysis is right, and I agree that removing the BUG_ON
from __unbind_from_irq seems like the right solution.

I can't see any issues from calling xen_free_irq with type ==
IRQT_UNBOUND, but I've already attempted to fix this once and failed,
so I would like to get second opinions. Also I'm not sure of the
reason behind that BUG_ON.

Roger.

Re: [PATCH 4.4 119/268] xen/pirq: fix error path cleanup when binding MSIs

[Boris Ostrovsky]

On 06/14/2018 04:21 AM, Roger Pau Monné wrote:
> On Wed, Jun 13, 2018 at 07:48:50PM +0100, Ben Hutchings wrote:
>> On Wed, 2018-02-28 at 09:19 +0000, Roger Pau Monne wrote:
>>> From: Roger Pau Monne <roger.pau@citrix.com>
>>>
>>> [ Upstream commit 910f8befdf5bccf25287d9f1743e3e546bcb7ce0 ]
>>>
>>> Current cleanup in the error path of xen_bind_pirq_msi_to_irq is
>>> wrong. First of all there's an off-by-one in the cleanup loop, which
>>> can lead to unbinding wrong IRQs.
>>>
>>> Secondly IRQs not bound won't be freed, thus leaking IRQ numbers.
>>>
>>> Note that there's no need to differentiate between bound and unbound
>>> IRQs when freeing them, __unbind_from_irq will deal with both of them
>>> correctly.
>> It appears to me that it is safe to call __unbind_from_irq() after
>> xen_irq_info_common_setup() fails, but *not* if the latter hasn't been
>> called at all.  In that case the IRQ type will still be set to
>> IRQT_UNBOUND and this will trigger the BUG_ON() in __unbind_from_irq().
>>
>> [...]
>>> --- a/drivers/xen/events/events_base.c
>>> +++ b/drivers/xen/events/events_base.c
>>> @@ -764,8 +764,8 @@ out:
>>>  	mutex_unlock(&irq_mapping_update_lock);
>>>  	return irq;
>>>  error_irq:
>>> -	for (; i >= 0; i--)
>>> -		__unbind_from_irq(irq + i);
>>> +	while (nvec--)
>>> +		__unbind_from_irq(irq + nvec);
>> If nvec > 1, and xen_irq_info_pirq_setup() fails for i != nvec - 1,
>> then we reach here without having called xen_irq_info_common_setup()
>> for all these IRQs.
>>
>> In that case, I think we will still want to call xen_free_irq() for all
>> IRQs.  So maybe the fix would be to remove the BUG_ON() in
>> __unbind_from_irq()?
> I think your analysis is right, and I agree that removing the BUG_ON
> from __unbind_from_irq seems like the right solution.
>
> I can't see any issues from calling xen_free_irq with type ==
> IRQT_UNBOUND, but I've already attempted to fix this once and failed,
> so I would like to get second opinions. Also I'm not sure of the
> reason behind that BUG_ON.

I don't see a reason for the BUG_ON either.

-boris