Re: [PATCH v2] usb:xhci fix panic in xhci_free_virt_devices_depth_first

[Guenter Roeck <groeck@google.com>]

On Mon, Nov 6, 2017 at 6:00 AM, Mathias Nyman
<mathias.nyman@linux.intel.com> wrote:
> On 06.11.2017 14:36, Chen Yu wrote:
>>
>>
>>
>> On 2017/11/6 19:32, Greg KH wrote:
>>>>
>>>> A simple process is as below:
>>>>         xhci_plat_probe()
>>>>                 |
>>>>         usb_add_hcd()
>>>> xhci_plat_remove()
>>>>                |                                                |
>>>>         find some device                                usb_remove_hcd()
>>>>                |                                                |
>>>>         hub_port_connect() -> usb_alloc_dev()           usb_disconnect()
>>>>                |                                                |
>>>>         before hub_enable_device()                      xhci_stop()
>>>>                                                                 |
>>>>
>>>> xhci_mem_cleanup()
>>>>                                                                 |
>>>>
>>>> xhci_free_virt_devices_depth_first()
>>>>                                                                 |
>>>>                                                         real_port is 0
>>>> access xhci->rh_bw[vdev->real_port-1]
>>>>
>>>> The problem came from https://bugs.96boards.org/show_bug.cgi?id=535
>>>> Also look at crbug.com/700041
>>>
>>>
>>> Then the bug needs to be fixed, throwing a huge kernel trace message
>>> into the kernel log is not "fixing" the problem at all, right?
>>>
>>> thanks,
>>>
>>> greg k-h
>>>
>>> .
>>>
>>
>> You are right, the way that xhci_plat_remove() to be called needs to be
>> fixed.
>> But there is still possibility for this crash.
>> What do you think if just add an "xhci_warn" instead of "WARN_ON"?
>> +       if (!vdev->real_port) {
>> +               xhci_warn(xhci, "Bad vdev->real_port\n");
>> +               goto out;
>> +       }
>> +
>>
>
> This patch solves the issue, just drop all the error messages.
>
> vdev->real_port is not set until the the device enable/address
> stage, and we know it won't have any children yet then, so no need to
> worry about a child having tt pointers to this device.
>
> The "goto out" to xhci_free_virt_device() you do is fine here.
>
> xhci_plat_remove() is the .remove callback for the xhci platform driver.
> It might get called before a device is properly enabled/addressed.
> Not really a error. A unlikely but possible situation.
>

Agreed. I can reproduce the problem with a well timed bind/unbind loop.

Guenter

> xhci_free_tt_info() already has a similar check
>
> Thanks
> -Mathias
>