On 2/14/20 12:22 PM, Laurent Pinchart wrote:
> Hi Hans,
>
> On Fri, Feb 14, 2020 at 11:06:36AM +0100, Hans Verkuil wrote:
>> On 1/24/20 3:13 PM, Dan Carpenter wrote:
>>> Syzbot triggered a use after free in v5.5-rc6:
>>>
>>> BUG: KASAN: use-after-free in v4l2_release+0x2f1/0x390 drivers/media/v4l2-core/v4l2-dev.c:459
>>>
>>> Allocated by task 94:
>>> usbvision_alloc drivers/media/usb/usbvision/usbvision-video.c:1315 [inline]
>>> usbvision_probe.cold+0x5c5/0x1f21 drivers/media/usb/usbvision/usbvision-video.c:1469
>>>
>>> Freed by task 1913:
>>> kfree+0xd5/0x300 mm/slub.c:3957
>>> usbvision_release+0x181/0x1c0 drivers/media/usb/usbvision/usbvision-video.c:1364
>>> usbvision_radio_close.cold+0x2b/0x74 drivers/media/usb/usbvision/usbvision-video.c:1130
>>> v4l2_release+0x2e7/0x390 drivers/media/v4l2-core/v4l2-dev.c:455
>>>
>>> The problem is that the v4l2_release() calls usbvision_release() which
>>> frees "usbvision" but v4l2_release() still wants to use
>>> "usbvision->vdev". One solution is to make this devm_ allocated memory
>>> so the memory isn't freed until later.
>>
>> devm_ allocated memory is freed after disconnect, so I doubt this will help, or at
>> best it will just move the problem elsewhere.
>
> Yes, devm_*alloc is evil :-( It has spread to many drivers and is used
> incorrectly in most cases.
>
>> The right approach would be to use the release() callback from struct v4l2_device:
>> that's called when the very last open filehandle is closed.
>
> Hillf Danton has sent a patch to do so in the "Re: KASAN: use-after-free
> Read in v4l2_release (3)" thread. Have you seen it ?
Ah, that was never mailed to linux-media, so never ended up in patchwork. And
if it ain't in patchwork, then I don't know about it :-)
Hillf, if you want your patch to be merged, then make sure it is CC-ed to
linux-media as well.
Regards,
Hans
>
>> But I'm not sure if it is worth the effort. The usbvision driver is a mess and
>> personally I think it should be deprecated.
>
> I agree.
>
>>> Reported-by: syzbot+75287f75e2fedd69d680@xxxxxxxxxxxxxxxxxxxxxxxxx
>>> Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
>>> ---
>>> I copied this idea from a different driver, but I haven't tested it.
>>> I wanted to try the #syz fix command to see if it works.
>>>
>>> drivers/media/usb/usbvision/usbvision-video.c | 4 +---
>>> 1 file changed, 1 insertion(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/media/usb/usbvision/usbvision-video.c b/drivers/media/usb/usbvision/usbvision-video.c
>>> index 93d36aab824f..07b4763062c4 100644
>>> --- a/drivers/media/usb/usbvision/usbvision-video.c
>>> +++ b/drivers/media/usb/usbvision/usbvision-video.c
>>> @@ -1312,7 +1312,7 @@ static struct usb_usbvision *usbvision_alloc(struct usb_device *dev,
>>> {
>>> struct usb_usbvision *usbvision;
>>>
>>> - usbvision = kzalloc(sizeof(*usbvision), GFP_KERNEL);
>>> + usbvision = devm_kzalloc(&dev->dev, sizeof(*usbvision), GFP_KERNEL);
>>> if (!usbvision)
>>> return NULL;
>>>
>>> @@ -1336,7 +1336,6 @@ static struct usb_usbvision *usbvision_alloc(struct usb_device *dev,
>>> v4l2_ctrl_handler_free(&usbvision->hdl);
>>> v4l2_device_unregister(&usbvision->v4l2_dev);
>>> err_free:
>>> - kfree(usbvision);
>>> return NULL;
>>> }
>>>
>>> @@ -1361,7 +1360,6 @@ static void usbvision_release(struct usb_usbvision *usbvision)
>>>
>>> v4l2_ctrl_handler_free(&usbvision->hdl);
>>> v4l2_device_unregister(&usbvision->v4l2_dev);
>>> - kfree(usbvision);
>>>
>>> PDEBUG(DBG_PROBE, "success");
>>> }
>
