Re: [PATCH] usb: gadget: composite: split spinlock to avoid recursion

[Alan Stern <stern@rowland.harvard.edu>]

On Thu, 14 Nov 2019, Michael Olbrich wrote:

> On Wed, Nov 13, 2019 at 10:36:25AM -0500, Alan Stern wrote:
> > On Wed, 13 Nov 2019, Peter Chen wrote:
> > > On 19-11-12 10:33:18, Michael Olbrich wrote:
> > > > 'delayed_status' and 'deactivations' are used completely independent but
> > > > they share the same spinlock. This can result in spinlock recursion:
> > > > 
> > > > BUG: spinlock recursion on CPU#1, uvc-gadget/322
> > > >  lock: 0xffffffc0570364e0, .magic: dead4ead, .owner: uvc-gadget/322, .owner_cpu: 1
> > > > CPU: 1 PID: 322 Comm: uvc-gadget Tainted: G         C O      5.3.0-20190916-1+ #55
> > > > Hardware name: XXXXX (DT)
> > > > Call trace:
> > > >  dump_backtrace+0x0/0x178
> > > >  show_stack+0x24/0x30
> > > >  dump_stack+0xc0/0x104
> > > >  spin_dump+0x90/0xa0
> > > >  do_raw_spin_lock+0xd8/0x108
> > > >  _raw_spin_lock_irqsave+0x40/0x50
> > > >  composite_disconnect+0x2c/0x80
> > > >  usb_gadget_disconnect+0x84/0x150
> > > >  usb_gadget_deactivate+0x64/0x120
> > > >  usb_function_deactivate+0x70/0x80
> > > >  uvc_function_disconnect+0x20/0x58
> > > >  uvc_v4l2_release+0x34/0x90
> > > >  v4l2_release+0xbc/0xf0
> > > >  __fput+0xb0/0x218
> > > >  ____fput+0x20/0x30
> > > >  task_work_run+0xa0/0xd0
> > > >  do_notify_resume+0x2f4/0x340
> > > >  work_pending+0x8/0x14
> > > > 
> > > > Fix this by using separate spinlocks.
> > > 
> > > This issue may be introduced by 0a55187a1ec8c ("USB: gadget core: Issue
> > > ->disconnect() callback from usb_gadget_disconnect()"), which adds
> > > gadget's disconnect at usb_gadget_disconnect. Add Alan, if he is Ok
> > > with your patch, you may cc to stable tree.
> > 
> > I wasn't aware of the dual usage of that lock in the composite core 
> > (and 0a55187a1ec8c touches only the gadget core, not composite.c).
> > 
> > In any case, I don't have a good feel for how the locking is supposed 
> > to work in the composite core.  This is really something Felipe should 
> > look at.
> > 
> > Would a better fix be to change usb_function_deactivate() so that it
> > doesn't hold the lock while calling usb_gadget_deactivate()?  Maybe
> > increment cdev->deactivations unconditionally before dropping the lock
> > (for mutual exclusion) and then decrement it again if the call fails?
> 
> Hmm, I think, that would mean that usb_gadget_activate() may be called
> while usb_gadget_deactivate() is still running right? That's not
> acceptable, is it?

It's a little tricky.  The lock in question belongs to the composite 
core, not the UDC core, so it doesn't really apply to the 
usb_gadget_{de}activate() routines.

As for mutual exclusion of usb_gadget_activate() and
usb_gadget_deactivate(), I don't know that anyone has ever considered
the matter.

> Anyways. Something else is needed because executing usb_gadget_deactivate()
> under the spinlock has another problem. It's hard to reproduce, but we've
> seen this one:
> 
> BUG: scheduling while atomic: pipewire/260/0x00000002
> Modules linked in: allegro(C) regmap_mmio v4l2_mem2mem xlnx_vcu st1232 uio_pdrv_genirq
> Preemption disabled at: [<ffffff801061dc40>] usb_function_deactivate+0x30/0x80
> CPU: 1 PID: 260 Comm: pipewire Tainted: G         C O 5.3.0-20191112-1 #2
> Hardware name: Wolfvision ZynqMP PF4 (DT)
> Call trace:
>  dump_backtrace+0x0/0x178
>  show_stack+0x24/0x30
>  dump_stack+0xc0/0x104
>  __schedule_bug+0xb0/0xc0
>  __schedule+0x354/0x4d8
>  schedule+0x44/0xd8
>  schedule_timeout+0x1b4/0x380
>  wait_for_common+0xc0/0x188
>  wait_for_completion_timeout+0x2c/0x38
>  dwc3_gadget_pullup+0x90/0xb0
>  usb_gadget_disconnect+0x38/0x150
>  usb_gadget_deactivate+0x64/0x120
>  usb_function_deactivate+0x70/0x80
>  uvc_function_disconnect+0x20/0x58
>  uvc_v4l2_release+0x34/0x90
>  v4l2_release+0xbc/0xf0
>  __fput+0x90/0x208
>  ____fput+0x20/0x30
>  task_work_run+0x98/0xb8
>  do_notify_resume+0x2f4/0x340
>  work_pending+0x8/0x14
> dwc3 fe200000.usb: timed out waiting for SETUP phase
> 
> Or maybe it's incorrect for dwc3_gadget_pullup() to sleep?

It isn't documented, so there's no definitive answer.  My feeling is 
that the UDC driver pullup routines should not sleep, but that's not 
official.

Of course, Felipe should have the last word on this.

Alan Stern