[PATCH] irqchip: GICv3: ITS: don't assume 64K page size in its_alloc_tables

[Marc Zyngier]

On 18/05/15 16:33, Stuart Yoder wrote:
> 
> 
>> -----Original Message-----
>> From: Marc Zyngier [mailto:marc.zyngier at arm.com]
>> Sent: Monday, May 18, 2015 9:09 AM
>> To: Yoder Stuart-B08248; tglx at linutronix.de
>> Cc: linux-arm-kernel at lists.infradead.org
>> Subject: Re: [PATCH] irqchip: GICv3: ITS: don't assume 64K page size in its_alloc_tables
>>
>> On 18/05/15 14:38, Stuart Yoder wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Marc Zyngier [mailto:marc.zyngier at arm.com]
>>>> Sent: Monday, May 18, 2015 8:23 AM
>>>> To: Yoder Stuart-B08248; tglx at linutronix.de
>>>> Cc: linux-arm-kernel at lists.infradead.org
>>>> Subject: Re: [PATCH] irqchip: GICv3: ITS: don't assume 64K page size in its_alloc_tables
>>>>
>>>> Hi Stuart,
>>>>
>>>> On 15/05/15 00:02, Stuart Yoder wrote:
>>>>> its_alloc_tables() needs to account for page sizes other than
>>>>> 64KB.  Without this change, when PAGE_SIZE=4KB its_alloc_tables()
>>>>> gets stuck in an infinite loop.
>>>>>
>>>>> Signed-off-by: Stuart Yoder <stuart.yoder at freescale.com>
>>>>> ---
>>>>>
>>>>> think this should go into 4.1 if at all possible...without it I am
>>>>> unable to boot a 4.1 kernel on the LS2085 SoC
>>>>
>>>> What you are suggesting here is a effectively a revert of commit
>>>> 790b57a, which would break other implementations.
>>>>
>>>> Can you please explain the actual issue? I'm failing to see how you end
>>>> up in an infinite loop here (the system page size and the ITS base
>>>> granule should be completely unrelated...).
>>>
>>> Here is the problem line:
>>>
>>>       val |= (alloc_size / psz) - 1;
>>>
>>> In our case:
>>>    alloc_size=16K
>>>    psz=64K
>>>
>>> ...so (alloc_size / psz) = 0, and thus val becomes -1, and everything
>>> is screwed up.  We get stuck in a loop to retry_baser:
>>
>> If alloc_size is 16k, you have an order of 2, and I have to assume this
>> is an allocation for a device table (otherwise order would be 4). So
>> things fail because we've computed an alloc_size smaller than what we
>> want to allocate as a minimum.
>>
>> Isn't that exactly what Minghuan's patch fixes?
> 
> Yes.  I'll get with Minghuan and between the 2 of us we'll get a properly
> commented version of his patch sent out.

Thanks Stuart.

	M.
-- 
Jazz is not dead. It just smells funny...