Re: [PATCH net-next v3 08/19] net: ravb: Move the IRQs get and request in the probe function

From: Sergey Shtylyov
Date: Sun Jan 14 2024 - 13:08:20 EST


On 1/10/24 2:55 PM, claudiu beznea wrote:

[...]

>>>>> From: Claudiu Beznea <claudiu.beznea.uj@xxxxxxxxxxxxxx>
>>>>>
>>>>> The runtime PM implementation will disable clocks at the end of
>>>>> ravb_probe(). As some IP variants switch to reset mode as a result of
>>>>> setting module standby through clock disable APIs, to implement runtime PM
>>>>> the resource parsing and requesting are moved in the probe function and IP
>>>>> settings are moved in the open function. This is done because at the end of
>>>>> the probe some IP variants will switch anyway to reset mode and the
>>>>> registers content is lost. Also keeping only register specific operations
>>>>> in the ravb_open()/ravb_close() functions will make them faster.
>>>>>
>>>>> Commit moves IRQ requests to ravb_probe() to have all the IRQs ready when
>>>>> the interface is open. As now IRQs gets and requests are in a single place
>>>>> there is no need to keep intermediary data (like ravb_rx_irqs[] and
>>>>> ravb_tx_irqs[] arrays or IRQs in struct ravb_private).
>>>>
>>>> There's one thing that you probably didn't take into account: after
>>>> you call request_irq(), you should be able to handle your IRQ as it's
>>>> automatically unmasked, unless you pass IRQF_NO_AUTOEN to request_irq().
>>>> Your device may be held i reset or even powered off but if you pass
>>>> IRQF_SHARED to request_irq() (you do in a single IRQ config), you must
>>>> be prepared to get your device's registers read (in order to ascertain
>>
>> And, at least on arm32, reading a powered off (or not clocked?) device's
>> register causes an imprecise external abort exception -- which results in a
>> kernel oops...
>>
>>>> whether it's your IRQ or not). And you can't even pass IRQF_NO_AUTOEN
>>>> along with IRQF_SHARED, according to my reading of the IRQ code...
>>>
>>> Good point!

Iff we can come up with a robust check whether the device is powered on,
we can overcome even the IRQF_SHARED issue though...
I'm seeing pm_runtime_active() API and wondering whether we can use it
from the IRQ context. Alternatively, we can add a is_opened flag, like
sh_eth.c does...

>>>>> This is a preparatory change to add runtime PM support for all IP variants.
>>>>
>>>> I don't readily see why this is necessary for the full-fledged RPM
>>>> support...
>>>
>>> I tried to speed up the ravb_open()/ravb_close() but missed the IRQF_SHARED
>>
>> I doubt that optimizing ravb_{open,close}() is worth pursuing, frankly...

OTOH, we'll get a simpler/cleaner code if we do this...
Previously, I was under an impression that it's common behavior of
the networking drivers to call request_irq() from their ndo_open() methods.
Apparently, it's not true anymore (probably with the introduction of the
managed device API) -- the newer drivers often call devm_request_irq()
from their probe() methods instead...

>>> IRQ. As there is only one IRQ requested w/ IRQF_SHARED, are you OK with
>>> still keeping the rest of IRQs handled as proposed by this patch?
>>
>> I'm not, as this doesn't really seem necessary for your main goal.
>> It's not clear in what state U-Boot leaves EtherAVB...

This still seems an issue though... My prior experience with the R-Car
MMC driver tells me that U-Boot may leave interrupts enabled... :-/

> Ok. One other reason I did this is, as commit message states, to keep
> resource parsing and allocation/freeing in probe/remove and hardware
> settings in open/close.
>
> Anyway, I'll revert all the changes IRQ related.

Now I've changed my mind -- let's retain your patch! It needs some work
though...

> Thank you,
> Claudiu Beznea

[...]

MBR, Sergey