[PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[Chen Gang]

evtchn_make_refcounted() may return failure, so need process the failure
case. In failure case, it need call unbind_from_irqhandler() just like
evtchn_unbind_from_user() has done.

irq_from_evtchn() must be OK when bind_evtchn_to_irqhandler() succeed,
so need not check it again.

Also still need remain the closing port code, because when the failure
occurs, unbind_from_irqhandler() will not close port internally.


Signed-off-by: Chen Gang <gang.chen.5i5j@gmail.com>
---
 drivers/xen/evtchn.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/xen/evtchn.c b/drivers/xen/evtchn.c
index 00f40f0..8a629d6 100644
--- a/drivers/xen/evtchn.c
+++ b/drivers/xen/evtchn.c
@@ -311,6 +311,11 @@ static int evtchn_bind_to_user(struct per_user_data *u, int port)
 		goto err;
 
 	rc = evtchn_make_refcounted(port);
+	if (rc < 0) {
+		unbind_from_irqhandler(irq_from_evtchn(port), evtchn);
+		goto err;
+	}
+
 	return rc;
 
 err:
-- 
1.7.11.7

Re: [PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[David Vrabel]

On 28/08/14 16:13, Chen Gang wrote:
> evtchn_make_refcounted() may return failure, so need process the failure
> case. In failure case, it need call unbind_from_irqhandler() just like
> evtchn_unbind_from_user() has done.
> 
> irq_from_evtchn() must be OK when bind_evtchn_to_irqhandler() succeed,
> so need not check it again.
> 
> Also still need remain the closing port code, because when the failure
> occurs, unbind_from_irqhandler() will not close port internally.

None of the evtchn_make_refcounted() failures can occur since we know we
have a valid irq and info at the single call site.

I would like to see this fixed by refactoring the code to remove
evtchn_make_refcounted() by making all irqs for event channels reference
counted.  These no need for this refcounting to be optional.

David

Re: [PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[Chen Gang]

On 8/28/14 23:49, David Vrabel wrote:
> On 28/08/14 16:13, Chen Gang wrote:
>> evtchn_make_refcounted() may return failure, so need process the failure
>> case. In failure case, it need call unbind_from_irqhandler() just like
>> evtchn_unbind_from_user() has done.
>>
>> irq_from_evtchn() must be OK when bind_evtchn_to_irqhandler() succeed,
>> so need not check it again.
>>
>> Also still need remain the closing port code, because when the failure
>> occurs, unbind_from_irqhandler() will not close port internally.
> 
> None of the evtchn_make_refcounted() failures can occur since we know we
> have a valid irq and info at the single call site.
> 

OK, thanks. I guess what you said is correct.

But only according to the code, for me, I am not quite sure about 'info'
must be always valid. If bind_evtchn_to_irqhandler() succeeds, I can not
find any related code to prove 'info' must be valid.

 - for a new irq, it will allocate 'info' for it.

 - but for an existing irq, the code assumes it may has no 'info'.
   (so several areas check 'info' whether valid, although irq is OK).

So could you give some additional related proofs for it? And if 'info'
must be always OK, can we remove all the related check about 'info'?


> I would like to see this fixed by refactoring the code to remove
> evtchn_make_refcounted() by making all irqs for event channels reference
> counted.  These no need for this refcounting to be optional.
> 

If 'info' really must always be valid, what you said sounds reasonable
to me.


Thanks
-- 
Chen Gang

Open, share, and attitude like air, water, and life which God blessed

Re: [PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[David Vrabel]

On 29/08/14 14:34, Chen Gang wrote:
> 
> 
> On 8/28/14 23:49, David Vrabel wrote:
>> On 28/08/14 16:13, Chen Gang wrote:
>>> evtchn_make_refcounted() may return failure, so need process the failure
>>> case. In failure case, it need call unbind_from_irqhandler() just like
>>> evtchn_unbind_from_user() has done.
>>>
>>> irq_from_evtchn() must be OK when bind_evtchn_to_irqhandler() succeed,
>>> so need not check it again.
>>>
>>> Also still need remain the closing port code, because when the failure
>>> occurs, unbind_from_irqhandler() will not close port internally.
>>
>> None of the evtchn_make_refcounted() failures can occur since we know we
>> have a valid irq and info at the single call site.
>>
> 
> OK, thanks. I guess what you said is correct.
> 
> But only according to the code, for me, I am not quite sure about 'info'
> must be always valid. If bind_evtchn_to_irqhandler() succeeds, I can not
> find any related code to prove 'info' must be valid.
> 
>  - for a new irq, it will allocate 'info' for it.
> 
>  - but for an existing irq, the code assumes it may has no 'info'.
>    (so several areas check 'info' whether valid, although irq is OK).
> 
> So could you give some additional related proofs for it? And if 'info'
> must be always OK, can we remove all the related check about 'info'?

I'm not sure what you mean by an existing irq.  If it's an irq for an
event channel it will have had info set when it was allocated.  the
irq_mapping_update_lock protects against seeing partially setup irqs.

So, the checks for !info can be removed, yes.

David

Re: [PATCH] drivers/xen/evtchn.c: Check failure for evtchn_make_refcounted()

[Chen Gang]

On 8/29/14 21:43, David Vrabel wrote:
> On 29/08/14 14:34, Chen Gang wrote:
>>
>>
>> On 8/28/14 23:49, David Vrabel wrote:
>>> On 28/08/14 16:13, Chen Gang wrote:
>>>> evtchn_make_refcounted() may return failure, so need process the failure
>>>> case. In failure case, it need call unbind_from_irqhandler() just like
>>>> evtchn_unbind_from_user() has done.
>>>>
>>>> irq_from_evtchn() must be OK when bind_evtchn_to_irqhandler() succeed,
>>>> so need not check it again.
>>>>
>>>> Also still need remain the closing port code, because when the failure
>>>> occurs, unbind_from_irqhandler() will not close port internally.
>>>
>>> None of the evtchn_make_refcounted() failures can occur since we know we
>>> have a valid irq and info at the single call site.
>>>
>>
>> OK, thanks. I guess what you said is correct.
>>
>> But only according to the code, for me, I am not quite sure about 'info'
>> must be always valid. If bind_evtchn_to_irqhandler() succeeds, I can not
>> find any related code to prove 'info' must be valid.
>>
>>  - for a new irq, it will allocate 'info' for it.
>>
>>  - but for an existing irq, the code assumes it may has no 'info'.
>>    (so several areas check 'info' whether valid, although irq is OK).
>>
>> So could you give some additional related proofs for it? And if 'info'
>> must be always OK, can we remove all the related check about 'info'?
> 
> I'm not sure what you mean by an existing irq.  If it's an irq for an
> event channel it will have had info set when it was allocated.  the
> irq_mapping_update_lock protects against seeing partially setup irqs.
> 

After check the code details again. I guess, really no "existing irq",
just like you said.

But in honest, only based on the code, for me, it is not quite clear (
I guess it is OK, but I am not sure it must be OK -- still worry about
it).


> So, the checks for !info can be removed, yes.
> 

So for me, for safety (also easy understanding) reason, I still prefer
my original patch, although it is not the best one.

If you are sure about it (I guess you are sure), please help send patch
for it (skip checking "!info"). If necessary, may mark Cc or Reported-by
to me.


Thanks.
-- 
Chen Gang

Open, share, and attitude like air, water, and life which God blessed