Re: [PATCH] usb: dwc3: gadget: Fix broken gadget on system suspend/resume

[Roger Quadros <rogerq@ti.com>]

On 11/12/2014 05:36 PM, Felipe Balbi wrote:
> Hi,
> 
> On Wed, Nov 12, 2014 at 05:32:00PM +0200, Roger Quadros wrote:
>> On 11/12/2014 05:08 PM, Felipe Balbi wrote:
>>> On Wed, Nov 12, 2014 at 04:58:16PM +0200, Roger Quadros wrote:
>>>> On TI SoCs (e.g. DRA7) we don't support the DWC3 hibernation feature.
>>>> We need to stop the gadget controller while system suspend
>>>> else it results in L3 Bus errors on resume with broken
>>>> USB gadget on J6-evm.
>>>>
>>>> [   55.718226] WARNING: CPU: 0 PID: 0 at drivers/bus/omap_l3_noc.c:147 l3_interrupt_handler+0x224/0x31c()
>>>> [   55.718232] 44000000.ocp:L3 Custom Error: MASTER USB3 TARGET GPMC (Idle): Data Access in User mode during Functional access
>>>> [   55.718263] Modules linked in: usb_f_ss_lb g_zero libcomposite configfs xhci_hcd btwilink bluetooth dwc3 6lowpan_iphc m25p80 dwc3_omap omap_remoteproc
>>>> [   55.718271] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.14.19-02011-g3ece3ca-dirty #658
>>>> [   55.718290] [<c0016090>] (unwind_backtrace) from [<c0012060>] (show_stack+0x10/0x14)
>>>> [   55.718302] [<c0012060>] (show_stack) from [<c062a2e4>] (dump_stack+0x78/0x94)
>>>> [   55.718315] [<c062a2e4>] (dump_stack) from [<c0043bc4>] (warn_slowpath_common+0x6c/0x90)
>>>> [   55.718325] [<c0043bc4>] (warn_slowpath_common) from [<c0043c7c>] (warn_slowpath_fmt+0x30/0x40)
>>>> [   55.718336] [<c0043c7c>] (warn_slowpath_fmt) from [<c030fb24>] (l3_interrupt_handler+0x224/0x31c)
>>>> [   55.718348] [<c030fb24>] (l3_interrupt_handler) from [<c0093010>] (handle_irq_event_percpu+0x5c/0x23c)
>>>> [   55.718358] [<c0093010>] (handle_irq_event_percpu) from [<c009322c>] (handle_irq_event+0x3c/0x5c)
>>>> [   55.718367] [<c009322c>] (handle_irq_event) from [<c0096160>] (handle_fasteoi_irq+0x98/0x158)
>>>> [   55.718377] [<c0096160>] (handle_fasteoi_irq) from [<c00929dc>] (generic_handle_irq+0x20/0x30)
>>>> [   55.718385] [<c00929dc>] (generic_handle_irq) from [<c000eea8>] (handle_IRQ+0x4c/0xb0)
>>>> [   55.718393] [<c000eea8>] (handle_IRQ) from [<c0008638>] (gic_handle_irq+0x28/0x5c)
>>>> [   55.718402] [<c0008638>] (gic_handle_irq) from [<c0630964>] (__irq_svc+0x44/0x5c)
>>>> [   55.718406] Exception stack(0xc09c1f68 to 0xc09c1fb0)
>>>> [   55.718412] 1f60:                   00000001 00000001 00000000 00000000 c09c0000 00000000
>>>> [   55.718418] 1f80: c09c0000 c09c0000 c0a7a1e4 c09c8938 c09c89b4 00000000 60000093 c09c1fb0
>>>> [   55.718423] 1fa0: c008a2e4 c00926cc 60000013 ffffffff
>>>> [   55.718433] [<c0630964>] (__irq_svc) from [<c00926cc>] (cpu_startup_entry+0x100/0x1f4)
>>>> [   55.718444] [<c00926cc>] (cpu_startup_entry) from [<c0901ac4>] (start_kernel+0x324/0x388)
>>>>
>>>> Signed-off-by: Roger Quadros <rogerq@ti.com>
>>>> ---
>>>>  drivers/usb/dwc3/gadget.c | 8 +++++++-
>>>>  1 file changed, 7 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
>>>> index 3818b26..bc15b54 100644
>>>> --- a/drivers/usb/dwc3/gadget.c
>>>> +++ b/drivers/usb/dwc3/gadget.c
>>>> @@ -2741,7 +2741,13 @@ int dwc3_gadget_prepare(struct dwc3 *dwc)
>>>
>>> there is no more dwc3_gadget_prepare(), please rebase on 'next'.
>>
>> right.
>>
>>>
>>>>  {
>>>>  	if (dwc->pullups_connected) {
>>>>  		dwc3_gadget_disable_irq(dwc);
>>>> -		dwc3_gadget_run_stop(dwc, true, true);
>>>> +		if (dwc->has_hibernation) {
>>>> +			dwc3_gadget_run_stop(dwc, true, true);
>>>> +		} else {
>>>> +			dwc3_gadget_run_stop(dwc, false, true);
>>>> +			/* remember to connect back on resume */
>>>> +			dwc->pullups_connected = true;
>>>> +		}
>>>
>>> Another thing you probably want to do is make sure there is nothing
>>> pending on our request list because if there is, then we need to wait
>>> for any in-flight transfers to complete before disconnecting.
>>
>> shouldn't that check be done in dwc3_gadget_run_stop()?
> 
> no, gadget drivers only disconnect pullups when they know there's
> nothing pending. And sometimes, we actually want to break such transfers
> to see how the thing would behave. Only suspend/resume needs to make
> transfers complete before hand.
> 
>>> I also wonder if current code isn't really fragile... I mean, when you
>>> clear run_stop bit, you should notify the gadget by means of
>>> ->disconnect() and so on... But since we can't test this with mainline,
>>> then it's difficult to tell.
>>>
>> that is true.
> 
> Not sure what to do with this patch :-p
> 
:). Maybe we can wait till we have suspend/resume working in mainline.

cheers,
-roger