From: Dongliang Mu <[email protected]>
flush_work will throw one WARN if worker->func is NULL. So we should always
initialize one worker before calling flush_work. When hdpvr_probe does not
initialize its worker, the hdpvr_disconnect will encounter one WARN. The
stack trace is in the following:
hdpvr_disconnect+0xb8/0xf2 drivers/media/usb/hdpvr/hdpvr-core.c:425
usb_unbind_interface+0xbf/0x3a0 drivers/usb/core/driver.c:458
__device_release_driver drivers/base/dd.c:1206 [inline]
device_release_driver_internal+0x22a/0x230 drivers/base/dd.c:1237
bus_remove_device+0x108/0x160 drivers/base/bus.c:529
device_del+0x1fe/0x510 drivers/base/core.c:3592
usb_disable_device+0xd1/0x1d0 drivers/usb/core/message.c:1419
usb_disconnect+0x109/0x330 drivers/usb/core/hub.c:2228
Fix this by adding a sanity check of the worker before flush_work.
Reported-by: syzkaller <[email protected]>
Signed-off-by: Dongliang Mu <[email protected]>
---
drivers/media/usb/hdpvr/hdpvr-core.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/media/usb/hdpvr/hdpvr-core.c b/drivers/media/usb/hdpvr/hdpvr-core.c
index 52e05a69c46e..d102b459d45d 100644
--- a/drivers/media/usb/hdpvr/hdpvr-core.c
+++ b/drivers/media/usb/hdpvr/hdpvr-core.c
@@ -422,7 +422,8 @@ static void hdpvr_disconnect(struct usb_interface *interface)
mutex_unlock(&dev->io_mutex);
v4l2_device_disconnect(&dev->v4l2_dev);
msleep(100);
- flush_work(&dev->worker);
+ if (dev->worker.func)
+ flush_work(&dev->worker);
mutex_lock(&dev->io_mutex);
hdpvr_cancel_queue(dev);
mutex_unlock(&dev->io_mutex);
--
2.25.1
On 17/02/2022 05:39, Dongliang Mu wrote:
> From: Dongliang Mu <[email protected]>
>
> flush_work will throw one WARN if worker->func is NULL. So we should always
> initialize one worker before calling flush_work. When hdpvr_probe does not
> initialize its worker, the hdpvr_disconnect will encounter one WARN. The
> stack trace is in the following:
>
> hdpvr_disconnect+0xb8/0xf2 drivers/media/usb/hdpvr/hdpvr-core.c:425
> usb_unbind_interface+0xbf/0x3a0 drivers/usb/core/driver.c:458
> __device_release_driver drivers/base/dd.c:1206 [inline]
> device_release_driver_internal+0x22a/0x230 drivers/base/dd.c:1237
> bus_remove_device+0x108/0x160 drivers/base/bus.c:529
> device_del+0x1fe/0x510 drivers/base/core.c:3592
> usb_disable_device+0xd1/0x1d0 drivers/usb/core/message.c:1419
> usb_disconnect+0x109/0x330 drivers/usb/core/hub.c:2228
>
> Fix this by adding a sanity check of the worker before flush_work.
>
> Reported-by: syzkaller <[email protected]>
> Signed-off-by: Dongliang Mu <[email protected]>
> ---
> drivers/media/usb/hdpvr/hdpvr-core.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/media/usb/hdpvr/hdpvr-core.c b/drivers/media/usb/hdpvr/hdpvr-core.c
> index 52e05a69c46e..d102b459d45d 100644
> --- a/drivers/media/usb/hdpvr/hdpvr-core.c
> +++ b/drivers/media/usb/hdpvr/hdpvr-core.c
> @@ -422,7 +422,8 @@ static void hdpvr_disconnect(struct usb_interface *interface)
> mutex_unlock(&dev->io_mutex);
> v4l2_device_disconnect(&dev->v4l2_dev);
> msleep(100);
> - flush_work(&dev->worker);
> + if (dev->worker.func)
> + flush_work(&dev->worker);
I don't think this is the right fix. Instead, move the INIT_WORK line from
hdpvr_start_streaming() to hdpvr_register_videodev(). That should initialize
the worker struct from the start instead of only when you start streaming,
as is the case today.
Can you try that?
Regards,
Hans
> mutex_lock(&dev->io_mutex);
> hdpvr_cancel_queue(dev);
> mutex_unlock(&dev->io_mutex);
On Thu, Feb 17, 2022 at 8:05 PM Hans Verkuil <[email protected]> wrote:
>
> On 17/02/2022 05:39, Dongliang Mu wrote:
> > From: Dongliang Mu <[email protected]>
> >
> > flush_work will throw one WARN if worker->func is NULL. So we should always
> > initialize one worker before calling flush_work. When hdpvr_probe does not
> > initialize its worker, the hdpvr_disconnect will encounter one WARN. The
> > stack trace is in the following:
> >
> > hdpvr_disconnect+0xb8/0xf2 drivers/media/usb/hdpvr/hdpvr-core.c:425
> > usb_unbind_interface+0xbf/0x3a0 drivers/usb/core/driver.c:458
> > __device_release_driver drivers/base/dd.c:1206 [inline]
> > device_release_driver_internal+0x22a/0x230 drivers/base/dd.c:1237
> > bus_remove_device+0x108/0x160 drivers/base/bus.c:529
> > device_del+0x1fe/0x510 drivers/base/core.c:3592
> > usb_disable_device+0xd1/0x1d0 drivers/usb/core/message.c:1419
> > usb_disconnect+0x109/0x330 drivers/usb/core/hub.c:2228
> >
> > Fix this by adding a sanity check of the worker before flush_work.
> >
> > Reported-by: syzkaller <[email protected]>
> > Signed-off-by: Dongliang Mu <[email protected]>
> > ---
> > drivers/media/usb/hdpvr/hdpvr-core.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/media/usb/hdpvr/hdpvr-core.c b/drivers/media/usb/hdpvr/hdpvr-core.c
> > index 52e05a69c46e..d102b459d45d 100644
> > --- a/drivers/media/usb/hdpvr/hdpvr-core.c
> > +++ b/drivers/media/usb/hdpvr/hdpvr-core.c
> > @@ -422,7 +422,8 @@ static void hdpvr_disconnect(struct usb_interface *interface)
> > mutex_unlock(&dev->io_mutex);
> > v4l2_device_disconnect(&dev->v4l2_dev);
> > msleep(100);
> > - flush_work(&dev->worker);
> > + if (dev->worker.func)
> > + flush_work(&dev->worker);
>
> I don't think this is the right fix. Instead, move the INIT_WORK line from
> hdpvr_start_streaming() to hdpvr_register_videodev(). That should initialize
> the worker struct from the start instead of only when you start streaming,
> as is the case today.
I see your point.
One small question: if we initialize worker at the beginning of
hdpvr_register_videodev, but without schedule_work, will flush_work at
hdpvr_disconnect lead to some issues?
Or we need to verify if the work is pending or running at hdpvr_disconnect?
>
> Can you try that?
>
> Regards,
>
> Hans
>
> > mutex_lock(&dev->io_mutex);
> > hdpvr_cancel_queue(dev);
> > mutex_unlock(&dev->io_mutex);
>
On 18/02/2022 06:58, Dongliang Mu wrote:
> On Thu, Feb 17, 2022 at 8:05 PM Hans Verkuil <[email protected]> wrote:
>>
>> On 17/02/2022 05:39, Dongliang Mu wrote:
>>> From: Dongliang Mu <[email protected]>
>>>
>>> flush_work will throw one WARN if worker->func is NULL. So we should always
>>> initialize one worker before calling flush_work. When hdpvr_probe does not
>>> initialize its worker, the hdpvr_disconnect will encounter one WARN. The
>>> stack trace is in the following:
>>>
>>> hdpvr_disconnect+0xb8/0xf2 drivers/media/usb/hdpvr/hdpvr-core.c:425
>>> usb_unbind_interface+0xbf/0x3a0 drivers/usb/core/driver.c:458
>>> __device_release_driver drivers/base/dd.c:1206 [inline]
>>> device_release_driver_internal+0x22a/0x230 drivers/base/dd.c:1237
>>> bus_remove_device+0x108/0x160 drivers/base/bus.c:529
>>> device_del+0x1fe/0x510 drivers/base/core.c:3592
>>> usb_disable_device+0xd1/0x1d0 drivers/usb/core/message.c:1419
>>> usb_disconnect+0x109/0x330 drivers/usb/core/hub.c:2228
>>>
>>> Fix this by adding a sanity check of the worker before flush_work.
>>>
>>> Reported-by: syzkaller <[email protected]>
>>> Signed-off-by: Dongliang Mu <[email protected]>
>>> ---
>>> drivers/media/usb/hdpvr/hdpvr-core.c | 3 ++-
>>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/media/usb/hdpvr/hdpvr-core.c b/drivers/media/usb/hdpvr/hdpvr-core.c
>>> index 52e05a69c46e..d102b459d45d 100644
>>> --- a/drivers/media/usb/hdpvr/hdpvr-core.c
>>> +++ b/drivers/media/usb/hdpvr/hdpvr-core.c
>>> @@ -422,7 +422,8 @@ static void hdpvr_disconnect(struct usb_interface *interface)
>>> mutex_unlock(&dev->io_mutex);
>>> v4l2_device_disconnect(&dev->v4l2_dev);
>>> msleep(100);
>>> - flush_work(&dev->worker);
>>> + if (dev->worker.func)
>>> + flush_work(&dev->worker);
>>
>> I don't think this is the right fix. Instead, move the INIT_WORK line from
>> hdpvr_start_streaming() to hdpvr_register_videodev(). That should initialize
>> the worker struct from the start instead of only when you start streaming,
>> as is the case today.
>
> I see your point.
>
> One small question: if we initialize worker at the beginning of
> hdpvr_register_videodev, but without schedule_work, will flush_work at
> hdpvr_disconnect lead to some issues?
>
> Or we need to verify if the work is pending or running at hdpvr_disconnect?
No, flush_work already checks if there is anything to do. If nothing was
scheduled, it will just return.
Regards,
Hans
>
>>
>> Can you try that?
>
>
>>
>> Regards,
>>
>> Hans
>>
>>> mutex_lock(&dev->io_mutex);
>>> hdpvr_cancel_queue(dev);
>>> mutex_unlock(&dev->io_mutex);
>>