On 08.07.21 08:37, Jan Beulich wrote:
> On 08.07.2021 07:47, Juergen Gross wrote:
>> On 17.05.21 17:33, Jan Beulich wrote:
>>> On 17.05.2021 17:22, Juergen Gross wrote:
>>>> On 17.05.21 17:12, Jan Beulich wrote:
>>>>> On 17.05.2021 16:23, Juergen Gross wrote:
>>>>>> On 17.05.21 16:11, Jan Beulich wrote:
>>>>>>> On 13.05.2021 12:02, Juergen Gross wrote:
>>>>>>>> @@ -1574,10 +1580,16 @@ static irqreturn_t blkif_interrupt(int irq, void *dev_id)
>>>>>>>>      	spin_lock_irqsave(&rinfo->ring_lock, flags);
>>>>>>>>       again:
>>>>>>>>      	rp = rinfo->ring.sring->rsp_prod;
>>>>>>>> +	if (RING_RESPONSE_PROD_OVERFLOW(&rinfo->ring, rp)) {
>>>>>>>> +		pr_alert("%s: illegal number of responses %u\n",
>>>>>>>> +			 info->gd->disk_name, rp - rinfo->ring.rsp_cons);
>>>>>>>> +		goto err;
>>>>>>>> +	}
>>>>>>>>      	rmb(); /* Ensure we see queued responses up to 'rp'. */
>>>>>>>
>>>>>>> I think you want to insert after the barrier.
>>>>>>
>>>>>> Why? The relevant variable which is checked is "rp". The result of the
>>>>>> check is in no way depending on the responses themselves. And any change
>>>>>> of rsp_cons is protected by ring_lock, so there is no possibility of
>>>>>> reading an old value here.
>>>>>
>>>>> But this is a standard double read situation: You might check a value
>>>>> and then (via a separate read) use a different one past the barrier.
>>>>
>>>> Yes and no.
>>>>
>>>> rsp_cons should never be written by the other side, and additionally
>>>> it would be read multiple times anyway.
>>>
>>> But I'm talking about rsp_prod, as that's what rp gets loaded from.
>>
>> Oh, now I get your problem.
>>
>> But shouldn't that better be solved by using READ_ONCE() for reading rp
>> instead?
> 
> Not sure - the rmb() is needed anyway aiui, and hence you could as well
> move your code addition.

Sure.

My question was rather: does the rmb() really eliminate the possibility
of a double read introduced by the compiler? If yes, moving the code is
the correct solution.


Juergen