Re: BUG with linux 5.9.0 with dwc3 in gadget mode

From: Felipe Balbi <balbi@kernel.org>
To: Ferry Toth <fntoth@gmail.com>,
	Andy Shevchenko <andy.shevchenko@gmail.com>
Cc: "linux-usb@vger.kernel.org" <linux-usb@vger.kernel.org>,
	Thinh Nguyen <Thinh.Nguyen@synopsys.com>,
	Heikki Krogerus <heikki.krogerus@linux.intel.com>,
	Jack Pham <jackp@codeaurora.org>
Subject: Re: BUG with linux 5.9.0 with dwc3 in gadget mode
Date: Fri, 06 Nov 2020 08:38:07 +0200	[thread overview]
Message-ID: <87eel7q9bk.fsf@kernel.org> (raw)
In-Reply-To: <ec15a515-6dbc-1224-a457-d5ac4a4db280@gmail.com>

Hi,

Ferry Toth <fntoth@gmail.com> writes:
> Hi,
>
> Op 28-10-2020 om 10:18 schreef Felipe Balbi:
>> Hi,
>>
>> Ferry Toth <fntoth@gmail.com> writes:
>>> Op 27-10-2020 om 22:16 schreef Andy Shevchenko:
>>>> On Tue, Oct 27, 2020 at 10:13 PM Ferry Toth <fntoth@gmail.com> wrote:
>>>>> Op 22-10-2020 om 15:43 schreef Andy Shevchenko:
>>>>>> On Thu, Oct 22, 2020 at 4:21 PM Thinh Nguyen <Thinh.Nguyen@synopsys.com> wrote:
>>>>>>> Ferry Toth wrote:
>>>> ...
>>>>
>>>>>>> There are some fixes to dwc3 in kernel mainline. Is it possible to test
>>>>>>> this against linux-next?
>>>>>> I think the best is to wait for v5.10-rc1 and retest.
>>>> Can you give a try of v5.10-rc1?
>>> Yes, I just tried:
>>>
>>> I booted in host mode, then flip the switch. Gadget come up, go down
>>> once, then come up again and stay up.
>> please collect trace events. It's important to figure out why it's going
>> down, even if only once. Make sure to collect trace *and* dmesg so we
>> can correlate trace with the reenumeration that should show up in dmesg.
>>
>> thanks
>
> Sorry, I had to replace mobo. Now back on this.
>
> As is, on Edison I can record with something like "perf record -e 
> 'dwc3:dwc3_gadget*' -e 'gadget:*' -g -a".
> Then get the trace buffer with "perf script > gadget.txt". Then at each 
> trace point we get a stack trace like:
>
> file-storage   831 [001]  4445.240038: dwc3:dwc3_gadget_ep_cmd: [FAILED 
> TO PARSE] name=ep4in cmd=524295 param0=0 param1=0 param2=0 cmd_status=0
>      ffffffff9a35b7e7 __traceiter_dwc3_gadget_ep_cmd+0x37 
> ([kernel.kallsyms])
>      ffffffff9a35b7e7 __traceiter_dwc3_gadget_ep_cmd+0x37 
> ([kernel.kallsyms])
>      ffffffff9a35fa40 dwc3_send_gadget_ep_cmd+0x320 ([kernel.kallsyms])
>      ffffffff9a3606d0 __dwc3_gadget_kick_transfer+0x200 ([kernel.kallsyms])
>      ffffffff9a361114 dwc3_gadget_ep_queue+0xe4 ([kernel.kallsyms])
>      ffffffff9a3afc3a usb_ep_queue+0x2a ([kernel.kallsyms])
>      ffffffffc047c301 start_transfer.isra.0+0x21 ([kernel.kallsyms])
>      ffffffffc047c88a start_in_transfer.isra.0+0x3a ([kernel.kallsyms])
>      ffffffffc047c93d send_status+0x8d ([kernel.kallsyms])
>      ffffffffc047dd05 fsg_main_thread+0x3c5 ([kernel.kallsyms])
>      ffffffff99c853b9 kthread+0xf9 ([kernel.kallsyms])
>      ffffffff99c01a32 ret_from_fork+0x22 ([kernel.kallsyms])
>
> "perf list" shows the tracepoint events, the same as under 
> /sys/kernel/debug/tracing/events/
>
> Question is which points to trace (above command fills buffer to 35MB in 
> 10sec). Do you have suggestions?

don't enable any gadget event. Only dwc3 events. Also, enable *all* dwc3
events. Usually, I avoid perf when doing this and just go straight to
/sys/kernel/trace/.

-- 
balbi