usb: gadget: ffs: Fix BUG when userland exits with submitted AIO transfers

[Vincent Pelletier <plr.vincent@gmail.com>]

Hello,

On Thu, 2 Aug 2018 14:23:51 +0000, Vincent Pelletier
<plr.vincent@gmail.com> wrote:
> On Thu, 2 Aug 2018 00:45:14 +0000, "He, Bo" <bo.he@intel.com> wrote:
> > Your patch fix the issue BUG: scheduling while atomic:  
> 
> Yes, although from my understanding of Felipe's answer, the actual bug
> is the "scheduling" part (sleeping in dwc3 UDC) rather than the
> "atomic" part.
> 
> So my patch addresses, still if my understanding is correct, the wrong
> half of the problem, and even introduced the regression you identified.
> Hence my uncertainty...

I notice that neither He's patch, nor a dwc3 change to prevent it from
scheduling inside usb_ep_dequeue are in Linus' master. Please correct
me if I missed something.

Just in case my previous emails were not clear:
- I have no objection to He's patch on its own (and I do not know the
  code nearly enough to provide a meaningful reviewed-by).
- I do not intend to work on making changes to dwc3 gadget to stop it
  from scheduling in usb_ep_dequeue in the foreseeable future.

So if there is no ongoing work on dwc3 cancel behaviour (Felipe ?),
please do resume/carry on with reviewing and integrating He's patch.

It is only *if* dwc3 cancel stops scheduling that I believe my patch
should be reverted (here is the hash as of Linus' master):
  commit d52e4d0c0c428bf2ba35074a7495cdb28e2efbae
  Author: Vincent Pelletier <plr.vincent@gmail.com>
  Date:   Wed Jun 13 11:05:06 2018 +0000

      usb: gadget: ffs: Fix BUG when userland exits with submitted AIO transfers

      This bug happens only when the UDC needs to sleep during usb_ep_dequeue,
      as is the case for (at least) dwc3.
and, in my understanding, a consequence is that He's fix would not
be needed anymore - the bug my patch introduced disappearing in the
revert.

Regards,