Re: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in isochronous mode

[John Youn]

On 9/10/2015 10:56 AM, Roman Bacik wrote:
>> -----Original Message-----
>> From: Roman Bacik
>> Sent: September-09-15 7:59 PM
>> To: 'John Youn'; Scott Branden; 'Greg Kroah-Hartman'; 'linux-
>> usb@xxxxxxxxxxxxxxx'
>> Cc: 'linux-kernel@xxxxxxxxxxxxxxx'; bcm-kernel-feedback-list
>> Subject: RE: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in isochronous mode
>>
>>> -----Original Message-----
>>> From: Roman Bacik
>>> Sent: September-09-15 7:36 PM
>>> To: 'John Youn'; Scott Branden; Greg Kroah-Hartman; linux-
>>> usb@xxxxxxxxxxxxxxx
>>> Cc: linux-kernel@xxxxxxxxxxxxxxx; bcm-kernel-feedback-list
>>> Subject: RE: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in
>>> isochronous mode
>>>
>>>> -----Original Message-----
>>>> From: John Youn [mailto:John.Youn@xxxxxxxxxxxx]
>>>> Sent: September-09-15 7:25 PM
>>>> To: Roman Bacik; John Youn; Scott Branden; Greg Kroah-Hartman;
>>>> linux- usb@xxxxxxxxxxxxxxx
>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx; bcm-kernel-feedback-list
>>>> Subject: Re: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in
>>>> isochronous mode
>>>>
>>>> On 9/9/2015 7:16 PM, Roman Bacik wrote:
>>>>>> -----Original Message-----
>>>>>> From: John Youn [mailto:John.Youn@xxxxxxxxxxxx]
>>>>>> Sent: September-09-15 7:11 PM
>>>>>> To: Roman Bacik; John Youn; Scott Branden; Greg Kroah-Hartman;
>>>>>> linux- usb@xxxxxxxxxxxxxxx
>>>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx; bcm-kernel-feedback-list
>>>>>> Subject: Re: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in
>>>>>> isochronous mode
>>>>>>
>>>>>> On 9/9/2015 11:16 AM, Roman Bacik wrote:
>>>>>>>> -----Original Message-----
>>>>>>>> From: John Youn [mailto:John.Youn@xxxxxxxxxxxx]
>>>>>>>> Sent: September-03-15 11:53 PM
>>>>>>>> To: Scott Branden; John Youn; Greg Kroah-Hartman; linux-
>>>>>>>> usb@xxxxxxxxxxxxxxx; Roman Bacik
>>>>>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx; bcm-kernel-feedback-list
>>>>>>>> Subject: Re: [PATCH v2 1/1] usb: dwc2: gadget: parity fix in
>>>>>>>> isochronous mode
>>>>>>>>
>>>>>>>> On 8/31/2015 9:17 AM, Scott Branden wrote:
>>>>>>>>> From: Roman Bacik <rbacik@xxxxxxxxxxxx>
>>>>>>>>>
>>>>>>>>> USB OTG driver in isochronous mode has to set the parity of
>>>>>>>>> the receiving microframe. The parity is set to even by
>>>>>>>>> default. This causes problems for an audio gadget, if the host
>>>>>>>>> starts transmitting on odd
>>>>>>>> microframes.
>>>>>>>>>
>>>>>>>>> This fix uses Incomplete Periodic Transfer interrupt to toggle
>>>>>>>>> between even and odd parity until the Transfer Complete
>>>>>>>>> interrupt is
>>>>>> received.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Roman Bacik <rbacik@xxxxxxxxxxxx>
>>>>>>>>> Reviewed-by: Abhinav Ratna <aratna@xxxxxxxxxxxx>
>>>>>>>>> Reviewed-by: Srinath Mannam
>> <srinath.mannam@xxxxxxxxxxxx>
>>>>>>>>> Signed-off-by: Scott Branden <sbranden@xxxxxxxxxxxx>
>>>>>>>>> ---
>>>>>>>>>  drivers/usb/dwc2/core.h   |  1 +
>>>>>>>>>  drivers/usb/dwc2/gadget.c | 51
>>>>>>>> ++++++++++++++++++++++++++++++++++++++++++++++-
>>>>>>>>>  drivers/usb/dwc2/hw.h     |  1 +
>>>>>>>>>  3 files changed, 52 insertions(+), 1 deletion(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/usb/dwc2/core.h b/drivers/usb/dwc2/core.h
>>>>>>>>> index 0ed87620..a5634fd 100644
>>>>>>>>> --- a/drivers/usb/dwc2/core.h
>>>>>>>>> +++ b/drivers/usb/dwc2/core.h
>>>>>>>>> @@ -150,6 +150,7 @@ struct s3c_hsotg_ep {
>>>>>>>>>  	unsigned int            periodic:1;
>>>>>>>>>  	unsigned int            isochronous:1;
>>>>>>>>>  	unsigned int            send_zlp:1;
>>>>>>>>> +	unsigned int            has_correct_parity:1;
>>>>>>>>>
>>>>>>>>>  	char                    name[10];
>>>>>>>>>  };
>>>>>>>>> diff --git a/drivers/usb/dwc2/gadget.c
>>>>>>>>> b/drivers/usb/dwc2/gadget.c index 4d47b7c..fac3e2f 100644
>>>>>>>>> --- a/drivers/usb/dwc2/gadget.c
>>>>>>>>> +++ b/drivers/usb/dwc2/gadget.c
>>>>>>>>> @@ -1954,6 +1954,7 @@ static void s3c_hsotg_epint(struct
>>>>>>>>> dwc2_hsotg
>>>>>>>> *hsotg, unsigned int idx,
>>>>>>>>>  		ints &= ~DXEPINT_XFERCOMPL;
>>>>>>>>>
>>>>>>>>>  	if (ints & DXEPINT_XFERCOMPL) {
>>>>>>>>> +		hs_ep->has_correct_parity = 1;
>>>>>>>>>  		if (hs_ep->isochronous && hs_ep->interval == 1) {
>>>>>>>>>  			if (ctrl & DXEPCTL_EOFRNUM)
>>>>>>>>>  				ctrl |= DXEPCTL_SETEVENFR; @@ -
>> 2316,7 +2317,8 @@ void
>>>> s3c_hsotg_core_init_disconnected(struct
>>>>>>>> dwc2_hsotg *hsotg,
>>>>>>>>>  		GINTSTS_CONIDSTSCHNG | GINTSTS_USBRST |
>>>>>>>>>  		GINTSTS_RESETDET | GINTSTS_ENUMDONE |
>>>>>>>>>  		GINTSTS_OTGINT | GINTSTS_USBSUSP |
>>>>>>>>> -		GINTSTS_WKUPINT,
>>>>>>>>> +		GINTSTS_WKUPINT |
>>>>>>>>> +		GINTSTS_INCOMPL_SOIN |
>>>> GINTSTS_INCOMPL_SOOUT,
>>>>>>>>>  		hsotg->regs + GINTMSK);
>>>>>>>>>
>>>>>>>>>  	if (using_dma(hsotg))
>>>>>>>>> @@ -2581,6 +2583,52 @@ irq_retry:
>>>>>>>>>  		s3c_hsotg_dump(hsotg);
>>>>>>>>>  	}
>>>>>>>>>
>>>>>>>>> +	if (gintsts & GINTSTS_INCOMPL_SOIN) {
>>>>>>>>> +		u32 idx, epctl_reg, ctrl;
>>>>>>>>> +		struct s3c_hsotg_ep *hs_ep;
>>>>>>>>> +
>>>>>>>>> +		dev_dbg(hsotg->dev, "%s:
>>>> GINTSTS_INCOMPL_SOIN\n",
>>>>>>>> __func__);
>>>>>>>>> +		for (idx = 1; idx < MAX_EPS_CHANNELS; idx++) {
>>>>>>>>
>>>>>>>> Valid endpoints are only up to hsotg->num_of_eps so this might
>>>>>>>> crash on certain configurations.
>>>>>>>>
>>>>>>>> Also, have you tried to find the endpoint which caused the
>>>>>>>> incomplete interrupt by reading the control registers as
>>>>>>>> described in
>>>> the databook?
>>>>>>>>
>>>>>>>
>>>>>>> We are using procedure based on description from this source:
>>>>>>>
>>>>>>> Synopsys, Inc. SolvNet 527
>>>>>>> DesignWare.com
>>>>>>> 3.00a
>>>>>>> April 2012
>>>>>>> USB 2.0 Hi-Speed On-The-Go (OTG) Databook Isochronous Endpoints
>>> in
>>>>>>> DWC_otg Slave Mode
>>>>>>>
>>>>>>> Synopsys databook is not in a public domain to quote the exact
>>>>>>> paragraph
>>>>>> here. You can find it in Chapter E, pp 526-527. There is no
>>>>>> register in this databook, which would provide information about
>>>>>> the source endpoint of the incomplete interrupt, as you have
>> described.
>>>>>>>
>>>>>>> Please, provide an exact reference and possibly enough
>>>>>>> information that
>>>>>> we can turn it into a working code.
>>>>>>>
>>>>>>
>>>>>> You can check the source of the interrupt by reading the DSTS and
>>>>>> ep control registers as described in the databook page 526, last
>>>>>> bullet, and 527 first bullet.
>>>>>>
>>>>>> John
>>>>>>
>>>>>
>>>>> Yes, that is exactly what we are doing what is described in the
>>>>> last bullet of
>>>> 526 and the first two bullets of page 527. So what is the problem?
>>>>>
>>>>
>>>> Maybe I missed it. Where exactly are you doing it like it says in
>>>> these two paragraphs?
>>>>
>>>> Quote:
>>>>
>>>> GINTSTS.incompISOIN is a global interrupt. Therefore, when more than
>>>> one isochronous endpoints is active, then the application must
>>>> determine which isochronous IN endpoints have incomplete data.
>>>>
>>>> To accomplish this, read DSTS and DIEPCTLn for all isochronous
>>>> endpoints. If DSTS.SOFFN[0] =DIEPCTLn.EO_FrNum (Bit 16),then that
>>>> endpoint has an incomplete transfer
>>>>
>>>> John
>>>>
>>>>
>>>
> 
> We have replaced the original code with the following code for both IN and OUT endpoints and retested:
> 
> 	if (gintsts & GINTSTS_INCOMPL_SOIN) {
> 		u32 idx, epctl_reg, ctrl, dsts;
> 		struct s3c_hsotg_ep *hs_ep;
> 
> +		dsts = readl(hsotg->regs + DSTS);
> 		dev_dbg(hsotg->dev, "%s: GINTSTS_INCOMPL_SOIN\n", __func__);
> 		for (idx = 1; idx < MAX_EPS_CHANNELS; idx++) {
> 			hs_ep = hsotg->eps_in[idx];
> 
> 			if (!hs_ep->isochronous || hs_ep->has_correct_parity)
> 				continue;
> 
> 			epctl_reg = DIEPCTL(idx);
> 			ctrl = readl(hsotg->regs + epctl_reg);
> 
> +			if (((dsts >> DSTS_SOFFN_SHIFT) & 1) !=
> +				((ctrl >> DXEPCTL_EOFRNUM) & 1))
> +				continue;
> 
> 			if (ctrl & DXEPCTL_EOFRNUM)
> 				ctrl |= DXEPCTL_SETEVENFR;
> 			else
> 				ctrl |= DXEPCTL_SETODDFR;
> 			writel(ctrl, hsotg->regs + epctl_reg);
> 			break;
> 		}
> 		writel(GINTSTS_INCOMPL_SOIN, hsotg->regs + GINTSTS);
> 	}
> 

Yes that is what I was asking about, checking for the source of
the interrupt.

But I'm convinced that the original code should be ok even with
multiple iso endpoints given that all you are doing is toggling
to find the correct parity the first time.

However a full implementation of this interrupt would need to
know the exact source because, even with parity set, you could
get this interrupt and will need to handle it properly. For
example if you are slow reading or writing the fifo.


> It still works. However, based on our experience with the audio gadget, USB host would either always use odd microframes or even microframes. It does not change the parity and hence gadget either always works or never works without this patch. So we can call hosts either odd or even.
> 
> When a gadget connects to even host, since it has even parity by default, it would receive transfer complete packets. So it recognizes it has correct parity and stop flipping. 
> 
> When a gadget connects to an odd host, it would flip all parities to the correct one on all endpoints after receiving an incomplete interrupt from any of the endpoints connected to this odd host using the original code. On the other hand it would require an interrupt for each connected endpoint in order to get the parity right using the new code above. 
> 

You would still get this interrupt just once per microframe. And
you would determine all the endpoints that missed that microframe
at once in the loop. The result should be the same as before, the
only overhead being the register read/check per iteration.


> Even when connected to multiple PCs with different parities, the worst case number of swaps (and hence lost packets) would be the number of connected hosts when using the original code, and the total number of endpoints, which is bigger than the number of connected hosts, when using the new code.
> 
> Please, let us know if you would prefer the new code anyway.

The original code is fine. But we do plan to update the gadget
code for isochronous so we may change how this is handled in the
future.

If you resubmit with my other feedback I will review and test on
our platforms.

Regards,
John



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/