On 08.12.2017 13:06, Alexander Kappner wrote:
> Hi,
> 
>> I think we need to dig a bit deeper. It's good to check if spriv is
>> valid
>> but there are probably other reasons than kzalloc failing.
> 
> I agree -- this small allocation is  unlikely to fail in practice.
> Also, while my patch prevents the kernel oops, it also prevents the
> debugfs entries from being created.
> 
> I've been debugging this more trying to come up with a better
> solution, but I might need some guidance as I'm not too familiar with
> the USB subsystem. The immediate cause of the crash was usbmuxd
> sending a USBDEVFS_SETCONFIGURATION ioctl to a device, which _only if
> it fails_ calls usb_hcd_alloc_bandwidth to try and reset the device,
> which in turn calls xhci_debugfs_create_endpoint. The ioctl handler
> acquires a device-specific lock via usb_lock_device.
> 
> When the system resumed from hibernate, xhci_resume was called. This
> in turn called xhci_mem_cleanup to deallocate the device structures,
> which include setting the debugfs_private pointer to NULL  (via
> xhci_free_virt_devices_depth_first). It thus seems likely that the
> ioctl is somehow racing with the hibernate. The call to xhci_resume
> is protected by a host-controller specific lock (xhci->lock) but it
> doesn't attempt to take the usb_lock_device device-specific lock.
> 
> Now my suspicion is that xhci_resume freed and zeroed the device
> structures while racing with the ioctl handler. I can't seem to find
> any exclusion mechanism that would prevent xhci_resume from racing
> with the USBDEVFS_SETCONFIGURATION ioctl (or any other ioctl, for
> that matter). Am I missing something? If not, is there any reason why
> an ioctl might need to execute in parallel with the xhci_resume? If
> not, can we just do a busy wait in xhci_resume until all pending
> ioctls have returned?

I'm not sure, but if I recall correctly then power management is supposed
to make sure a driver doesn't access usb devices while the host controller
is still resuming.

The odd thing here is that
xhci_debugfs_remove_slot(xhci, slot_id), and
xhci_free_virt_device(xhci, slot_id) are called together when
xhci_mem_cleanup() calls xhci_free_virt_devices_depth_first()

That means both the xhci_virt_device *dev and dev->debugfs_private
should both be freed and xhci->devs[slot_id] set NULL for that virt_device.

so xhci_add_endpoint() should fail a lot earlier because the xhci->devs[slot_id]
should be a null pointer as well.

Allocation is also done together in xhci_alloc_dev()

Looking at it more closely there is actually the .free_dev callback that
first frees the dev->debugs_private but the virt_dev is only freed
conditionally later

Attached a patch that frees them together, can you try it out?

If it doesn't help we need to add some elaborate tracing

Thanks
-Mathias