Re: [RFC][PATCH] dwc3: gadget: Fix losing list items in dwc3_gadget_ep_cleanup_completed_requests()

[Thinh Nguyen]

Wesley Cheng wrote:
> Hi Thinh,
> 
> On 8/9/2021 3:57 PM, Thinh Nguyen wrote:
>> John Stultz wrote:
>>> On Mon, Aug 9, 2021 at 3:44 PM Thinh Nguyen <Thinh.Nguyen@xxxxxxxxxxxx> wrote:
>>>>
>>>> John Stultz wrote:
>>>>> In commit d25d85061bd8 ("usb: dwc3: gadget: Use
>>>>> list_replace_init() before traversing lists"), a local list_head
>>>>> was introduced to process the started_list items to avoid races.
>>>>>
>>>>> However, in dwc3_gadget_ep_cleanup_completed_requests() if
>>>>> dwc3_gadget_ep_cleanup_completed_request() fails, we break early,
>>>>> causing the items on the local list_head to be lost.
>>>>>
>>>>> This issue showed up as problems on the db845c/RB3 board, where
>>>>> adb connetions would fail, showing the device as "offline".
>>>>>
>>>>> This patch tries to fix the issue by if we are returning early
>>>>> we splice in the local list head back into the started_list
>>>>> and return (avoiding an infinite loop, as the started_list is
>>>>> now non-null).
>>>>>
>>>>> Not sure if this is fully correct, but seems to work for me so I
>>>>> wanted to share for feedback.
>>>>>
>>>>> Cc: Wesley Cheng <wcheng@xxxxxxxxxxxxxx>
>>>>> Cc: Felipe Balbi <balbi@xxxxxxxxxx>
>>>>> Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
>>>>> Cc: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
>>>>> Cc: Jack Pham <jackp@xxxxxxxxxxxxxx>
>>>>> Cc: Thinh Nguyen <thinh.nguyen@xxxxxxxxxxxx>
>>>>> Cc: Todd Kjos <tkjos@xxxxxxxxxx>
>>>>> Cc: Amit Pundir <amit.pundir@xxxxxxxxxx>
>>>>> Cc: YongQin Liu <yongqin.liu@xxxxxxxxxx>
>>>>> Cc: Sumit Semwal <sumit.semwal@xxxxxxxxxx>
>>>>> Cc: Petri Gynther <pgynther@xxxxxxxxxx>
>>>>> Cc: linux-usb@xxxxxxxxxxxxxxx
>>>>> Fixes: d25d85061bd8 ("usb: dwc3: gadget: Use list_replace_init() before traversing lists")
>>>>> Signed-off-by: John Stultz <john.stultz@xxxxxxxxxx>
>>>>> ---
>>>>>  drivers/usb/dwc3/gadget.c | 6 ++++++
>>>>>  1 file changed, 6 insertions(+)
>>>>>
>>>>> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
>>>>> index b8d4b2d327b23..a73ebe8e75024 100644
>>>>> --- a/drivers/usb/dwc3/gadget.c
>>>>> +++ b/drivers/usb/dwc3/gadget.c
>>>>> @@ -2990,6 +2990,12 @@ static void dwc3_gadget_ep_cleanup_completed_requests(struct dwc3_ep *dep,
>>>>>                       break;
>>>>>       }
>>>>>
>>>>> +     if (!list_empty(&local)) {
>>>>> +             list_splice_tail(&local, &dep->started_list);
>>>>> +             /* Return so we don't hit the restart case and loop forever */
>>>>> +             return;
>>>>> +     }
>>>>> +
>>>>>       if (!list_empty(&dep->started_list))
>>>>>               goto restart;
>>>>>  }
>>>>>
>>>>
>>>> No, we should revert the change for
>>>> dwc3_gadget_ep_cleaup_completed_requests(). As I mentioned previously,
>>>> we don't cleanup the entire started_list. If the original problem is due
>>>> to disconnection in the middle of request completion, then we can just
>>>> check for pullup_connected and exit the loop and let the
>>>> dwc3_remove_requests() do the cleanup.
>>>
>>> Ok, sorry, I didn't read your mail in depth until I had this patch
>>> sent out. If a revert of d25d85061bd8 is the better fix, I'm fine with
>>> that too.
>>>
>>> thanks
>>> -john
>>>
>>
>> IMO, we should revert this patch for now since it will cause regression.
>> We can review and test a proper fix at a later time.
>>
>> Thanks,
>> Thinh
>>
> 
> Another suggestion would just be to replace the loop with a while() loop
> and using list_entry() instead.  That was what was discussed in the
> earlier patch series which also addresses the problem as well.  Issue
> here is the tmp variable still carries a stale request after the dwc3
> giveback is called.  We can avoid that by always fetching the
> list_entry() instead of relying on list_for_each_safe()
> 
> https://urldefense.com/v3/__https://lore.kernel.org/linux-usb/1620716636-12422-1-git-send-email-wcheng@xxxxxxxxxxxxxx/__;!!A4F2R9G_pg!P0E1pv3C0PStDepKyy8iqKgUaOhDy0ZDhYdz-_cZwnJRQjNjvw0MdJQCdU6Xwnt3YAs_$ 
> 

This should work, but the awkward thing is 2 loops from 2 separate
threads competing to remove/giveback the requests and may report mix status.

BR,
Thinh