Re: [PATCH 09/15] irq: arm: perform irqentry in entry code

From: Vladimir Murzin
Date: Tue Nov 30 2021 - 03:49:30 EST


On 10/23/21 2:36 PM, Vladimir Murzin wrote:
> On 10/23/21 2:18 PM, Marc Zyngier wrote:
>> On Sat, 23 Oct 2021 13:06:25 +0100,
>> Vladimir Murzin <vladimir.murzin@xxxxxxx> wrote:
>>>
>>> On 10/22/21 7:43 PM, Marc Zyngier wrote:
>>>> On Fri, 22 Oct 2021 18:58:54 +0100,
>>>> Mark Rutland <mark.rutland@xxxxxxx> wrote:
>>>>>
>>>>> On Fri, Oct 22, 2021 at 05:34:20PM +0100, Vladimir Murzin wrote:
>>
>> [...]
>>
>>>>>> As for TODO, is [1] look something you have been thinking of? IIUC,
>>>>>> the show stopper is that hwirq is being passed from exception entry
>>>>>> which retrieved via xPSR (IPSR to be precise). OTOH hwirq also available
>>>>>> via Interrupt Controller Status Register (ICSR) thus can be used in
>>>>>> driver itself... I gave [1] a go and it runs fine, yet I admit I might
>>>>>> be missing something...
>>>>>
>>>>> I hadn't thought about it in much detail, but that looks good!
>>>>>
>>>>> I was wondering if we needed something like a
>>>>> handle_arch_vectored_irq(), but if we can rely on the ICSR that seems
>>>>> simpler overall. I'm not at all familiar with M-class, so I'm not sure
>>>>> if there are pitfalls in this area.
>>>>
>>>> Why can't we just use IPSR instead from the C code? It has the
>>>> potential of being of lower latency then a MMIO read (though I have no
>>>> idea whether it makes a material difference on M-class) and from what
>>>> I can see in the arch spec, they are strictly equivalent.
>>>
>>> Hmmm, less arch specific asm(s) in driver code, no?
>>
>> Well, it isn't like this driver is going to be useful on anything
>> else, is it?
>>
>
> Well, with some work to unwire it from arch/arm it can be COMPILE_TEST :)
>
>> If there is no overhead in reading from MMIO compared to the
>> architected register, then I agree that ICSR is the way to
>> go. Is there any chance you could measure it on a HW platform? Or
>> maybe in emulation?
>
> My MPS{2,3} boards left in office and I'm on holiday next week... OTOH, I
> have no strong opinion on ICSR vs IPSR, I just wanted to check how much
> work it'd be to close TODO per my (quite limited) understanding :)

One month and a week later...

I observe that in terms of performance

MRS r0, ipsr

is equivalent to readl_relaxed(BASEADDR_V7M_SCB + V7M_SCB_ICSR)

MOV.W r3, #3758153728
LDR.W r0, [r3, #3332]

Old compilers can produce less performant sequence like

LDR r3,0xbcc0
ADD.W r3,r3,#0xaf00
LDR r0,[r3,#0]

So, what would be your preference?

Cheers
Vladimir

>
> Cheers
> Vladimir
>
>
>>
>> Thanks,
>>
>> M.
>>
>