High CPU load produced by USB (DW2)

[Marek Vasut]

On 02/19/2018 11:11 AM, Minas Harutyunyan wrote:
> On 2/19/2018 12:51 PM, Marek Vasut wrote:
>> On 02/19/2018 09:19 AM, Minas Harutyunyan wrote:
>>> On 2/17/2018 12:07 AM, Marek Vasut wrote:
>>>> On 02/16/2018 06:59 AM, Minas Harutyunyan wrote:
>>>>> On 2/15/2018 5:20 PM, Mirza Krak wrote:
>>>>>> On 14 February 2018 at 13:07, Minas Harutyunyan
>>>>>> <Minas.Harutyunyan at synopsys.com> wrote:
>>>>>>> On 2/14/2018 12:57 PM, Mirza Krak wrote:
>>>>>>>> On 8 February 2018 at 14:53, Minas Harutyunyan
>>>>>>>> <Minas.Harutyunyan at synopsys.com> wrote:
>>>>>>
>>>>>> < snip >
>>>>>>
>>>>>>>
>>>>>>> I reviewed your interrupt count log again. About 140,000 interrupts in 2
>>>>>>> seconds, obviously it's not SOF only interrupts. More probably, its NAK
>>>>>>> respond interrupts to SSPLIT/CSPLIT transactions. For this case I can
>>>>>>> recommend you to apply patch from Douglas Anderson: "[PATCH v2] usb:
>>>>>>> dwc2: host: Don't retry NAKed transactions right away" which already
>>>>>>> merged to 4.16-rc1.
>>>>>>>
>>>>>>> In your setups you see different behavior on different HUBs. Your HUBs
>>>>>>> have different "TT think time": 8 and 32. In USB2.0 spec "TT think time"
>>>>>>> described as follow "TT requires at most 8/32 FS bit times of inter
>>>>>>> transaction gap on a full-/low-speed downstream bus". So, your "worst"
>>>>>>> HUB with "TT think time"=8 sending more frequently SSPLIT/CSPLIT
>>>>>>> transactions which replied by NAK. As result you see about 4 time more
>>>>>>> interrupts comparing to "good" HUB. Could you please check interrupts
>>>>>>> count for "good" HUB and check "4 time" hypothesis.
>>>>>>
>>>>>> Did some further testing. The "good" HUB is actually as bad as the
>>>>>> "bad" HUB, and it was my setup that caused the different behavior. Did
>>>>>> not use the same devices etc.
>>>>>>
>>>>>> Once I made sure that the configuration and setup was the same on both
>>>>>> board I could see that the behaved similarly. And that is the
>>>>>> following interrupt load:
>>>>>>
>>>>>> - BT USB (FS) = ~80k interrupts / second
>>>>>> - Keyboard (FS) = ~80k interrupts / second
>>>>>> - WiFI USB (HS) = ~8k interrupts / second
>>>>>>
>>>>>> After applying the suggested patch [1], it is steady around 8k
>>>>>> interrupts / second no matter what device I connect (HS/FS,
>>>>>> HUB/NO-HUB). Which is acceptable and usable.
>>>>>
>>>>> Great!
>>>>
>>>> So 8k IRQs per second is what I should expect from this HW with HS
>>>> device attached, that's normal and cannot be reduced ?
>>>>
>>> If core acting as Host in Buffer DMA mode then 8k IRQs per second (SOF
>>> interrupts) is expected if connected device(s) has periodic endpoint.
>>
>> Is there a way to reduce that or is that the absolute minimum in HS mode?
>>
> We already discussed, in this email thread earlier, why SOF interrupts 
> required and unmasked.
> Only in case when connected device with CTRL+BLK EP's only (like flash 
> drive) and directly connected to cores root HUB, SOF's will be masked.

That's the setup I have on Altera SoCFPGA, yet the SOFs are still present.

-- 
Best regards,
Marek Vasut