Hi John,
John Youn <John.Youn@xxxxxxxxxxxx> writes:
>> John Youn <John.Youn@xxxxxxxxxxxx> writes:
>>>>> John Youn <John.Youn@xxxxxxxxxxxx> writes:
>>>>>> On 8/31/2016 2:38 AM, Felipe Balbi wrote:
>>>>>>> If we don't know what are the actual U1/U2 exit
>>>>>>> latencies from the UDC, we're better off using
>>>>>>> maximum latencies as default. This should avoid
>>>>>>> any problems with too frequent U1/U2 entry.
>>>>>>>
>>>>>>> Signed-off-by: Felipe Balbi <felipe.balbi@xxxxxxxxxxxxxxx>
>>>>>>> ---
>>>>>>> include/linux/usb/gadget.h | 12 ++++++++++--
>>>>>>> 1 file changed, 10 insertions(+), 2 deletions(-)
>>>>>>>
>>>>>>> diff --git a/include/linux/usb/gadget.h b/include/linux/usb/gadget.h
>>>>>>> index 3667d667cab1..20cb590c658e 100644
>>>>>>> --- a/include/linux/usb/gadget.h
>>>>>>> +++ b/include/linux/usb/gadget.h
>>>>>>> @@ -276,11 +276,19 @@ static inline void usb_ep_fifo_flush(struct usb_ep *ep)
>>>>>>>
>>>>>>> /*-------------------------------------------------------------------------*/
>>>>>>>
>>>>>>> +/**
>>>>>>> + * struct usb_dcd_config_params - Default U1/U2 exit latencies
>>>>>>> + * @bU1DevExitLat: U1 Exit Latency in us
>>>>>>> + * @bU2DevExitLat: U2 Exit Latency in us
>>>>>>> + *
>>>>>>> + * Note that we will be setting U1/U2 exit latencies to their maximum
>>>>>>> + * by default if no value is passed by the UDC Driver.
>>>>>>> + */
>>>>>>> struct usb_dcd_config_params {
>>>>>>> __u8 bU1DevExitLat; /* U1 Device exit Latency */
>>>>>>> -#define USB_DEFAULT_U1_DEV_EXIT_LAT 0x01 /* Less then 1 microsec */
>>>>>>> +#define USB_DEFAULT_U1_DEV_EXIT_LAT 10 /* us */
>>>>>>> __le16 wU2DevExitLat; /* U2 Device exit Latency */
>>>>>>> -#define USB_DEFAULT_U2_DEV_EXIT_LAT 0x1F4 /* Less then 500 microsec */
>>>>>>> +#define USB_DEFAULT_U2_DEV_EXIT_LAT 2047 /* us */
>>>>>>> };
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>> Hi Felipe,
>>>>>>
>>>>>> Speaking of this, how would you feel about adding module parameters in
>>>>>> the dwc3-pci to set these? And we also have several more settings in
>>>>>> our internal tree that we need for IP validation and debugging.
>>>>>>
>>>>>> As you know the IP is highly configurable, and we do much of the
>>>>>> testing against these configurations via software settings. And our
>>>>>> validation platform is not a typical platform. Basically we have a
>>>>>> single platform, but the instantiated IP and PHY could be configured
>>>>>> in any way so it could need different values passed in for
>>>>>> testing. Like the U1/U2 exit latencies, LPM NYET threshold, HIRD,
>>>>>> USB2/3 SUSPHY, LPM sleep mode, GFLADJ, etc.
>>>>>>
>>>>>> And some things that are automatically detected we need to restrict or
>>>>>> disable to get full coverage. Such as disabling U1/U2 and LPM, maximum
>>>>>> speed, dr_mode, NUMP, RX thresholding, RX thresholding packet count.
>>>>>>
>>>>>> I know module parameters are typically frowned upon so do you
>>>>>> recommend another approach?
>>>>>
>>>>> I completely understand the situation. Module parameters are, well,
>>>>> rather unsophisticated. FPGA validation is, however, a 'peculiar' use
>>>>> case.
>>>>>
>>>>> I want to avoid module parameters as much as possible, but apart from
>>>>> module parameters, the only thing I can think of is a specific
>>>>> sub-directory on debugfs which only gets compile if EXPERT &&
>>>>> DWC3_VALIDATION_MODE or whatever.
>>>>>
>>>>> Would debugfs work for you? The only problem is for things which get
>>>>> discovered during probe()
>>>>
>>>> Yes that's the problem, otherwise I'd be fine with debugfs. Almost
>>>> everything we need is initialized on probe. I thought of maybe
>>>> refactoring the dwc3 code so that this doesn't have to happen on probe
>>>> and we can trigger a "module reset" or something. But that is not
>>>> exactly clean either.
>>>>
>>>> Regards,
>>>> John
>>>>
>>>
>>> Does it seem reasonable to add module params to the PCI driver for
>>> this use case? At least until/if there is some better solution. We can
>>> guard it with a config option if necessary.
>>
>> module parameters are user-space ABI. Once added, we can't easily
>> remove. I've been considering if kprobes could be used to change stuff
>> like this.
>>
>> module parameters also feel like a big, big hammer to hit a tiny nail
>> head. I don't want to add any module parameters for stuff like this. And
>> since you've been pushing for them for a while, it only shows that
>> you're only concerned about your use case ;-)
>
> Maybe so, but module params are the easiest, workable solution. It
It might be easy to implement but it becomes a pain as time
passes. I can give you a concrete example (using device properties to
illustrate):
We introduced way back a property for platform code to tell us that
$this dwc3 instance needed the driver to resize FIFOs. The only reason
for this was because OMAP5 ES1.0 had default FIFO sizes which were less
than a full bulk USB 3.0 packet (< 1024 bytes) so we couldn't receive
any USB 3.0 packets.
Documentation was clear that this property was only needed if default
FIFO sizes were bad and yet several bindings started using it
blindly. Granted, this is one occasion where it really didn't cause
problems to resize FIFO, but now imagine what happens when we introduce
several module parameters and people start using it without really
knowing what they're for. We will start getting "bug reports" because
someone passed a e.g. "number_of_endpoints=0" parameter to dwc3 and the
driver didn't allocate any EP structure, or something silly like that.
> doesn't affect any other modules other than dwc3-pci, and they will
> only touch certain already-defined DT bindings. So in terms of
> maintainability, the code is all in one place, in one module, and it
> should be stable since the bindings are already defined as part of the
> ABI of dwc3.ko.
dwc3-pci is also used by non-FPGA platforms :-) I really don't want to
introduce module parameters and I really think debugfs can be used here
for your case. Just expose all parameters you want to dwc3-pci's debugfs
(needs to be created) and blacklist dwc3.ko so it doesn't load
automatically.
Then, load dwc3-pci, mount debugfs, set all parameters you need and
manually modprobe dwc3.ko. No module parameters, debugfs is usually not
shipping in products, and we can still wrap dwc3-pci's debugfs creation
with a #ifdef DWC3_FPGA_PROTOTYPE_EXTRAS or something along those lines.
>> Remember that module parameters are 'global', every instance of
>> dwc3-pci.ko/dwc3.ko will behave with the same set of module parameters.
>>
>> Sorry, but I have to say 'no' again. module parameters are not an
>> option; and I'd strongly suggest you don't do it for dwc2 either because
>> it'll make maintaining the driver a lot more difficult.
>>
>
> Well you never said 'no' in the first place. The thread just died
> before any decision was made either way. Hence why I bring it up
> again.
>
> As I mentioned previously, one other thing I can think of is to rework
> the platform probe to take out the core and parameter initialization
> to be called from other contexts. What do you think about this?
this will be even worse :-(
> In any case, whatever the ultimate solution, I will follow the same in
> dwc2.
fair enough, I'll try to cook something up, just a sec.
--
balbi
Attachment:
signature.asc
Description: PGP signature
