Re: [PATCH] ipc/mqueue: Avoid relying on a stack reference past its expiry

[Varad Gautam]

On 5/5/21 5:11 PM, Davidlohr Bueso wrote:
> On 2021-05-05 00:49, Varad Gautam wrote:
>> The race here really is about the lifetime of __pipelined_op's `this`
>> argument only
>> being guaranteed for some duration of the call (ie, until someone sets
>> ->state = STATE_READY). It is not about when wake_q addition happens,
>> as long as it is
>> being fed a valid task_struct.
> 
> Again, it's all about ensuring that the READY_STATE is set last, the blocked
> task has no business returning in the first place, making both races (exit and
> the one reported here) similar by ending up using bogus memory.
> 
> ...
> 
>> I considered that initially, but given that the race isn't connected
>> with wakeup, I
>> preferred the current approach which makes this fact explicit by showing what's
>> valid/invalid during __pipelined_op.
> 
> I understand your point, but this is why I updated the ordering comments. Furthermore
> there is no reason to decouple the task's refcount with the wake_q_add operation, it
> just makes the code weird and harder to follow.
> 

Different tastes I guess - I'd avoid an additional case to account for in the
MQ_BARRIER comment block and have the __pipelined_op code describe itself,
which avoids some back and forth while reading.

If you're still unconvinced, I'll send out a v2 w/ wake_q_add called before
smp_store_release.

Thanks,
Varad

> Thanks,
> Davidlohr
> 

-- 
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5
90409 Nürnberg
Germany

HRB 36809, AG Nürnberg
Geschäftsführer: Felix Imendörffer