Re: [PATCH] media: v4l: async: Fix NULL pointer when v4l2 flash subdev binding

[Sakari Ailus]

Hi Laurent,

On Fri, May 17, 2024 at 02:19:44PM +0300, Laurent Pinchart wrote:
> On Fri, May 17, 2024 at 08:00:32AM +0000, Sakari Ailus wrote:
> > Hi Chi Yuan,
> > 
> > On Fri, May 17, 2024 at 02:31:50PM +0800, ChiYuan Huang wrote:
> > > Hi, Sakari:
> > > 
> > > 	Thanks for your reply.
> > > If any misunderstanding, please correct me.
> > > 
> > > On Thu, May 16, 2024 at 10:42:05AM +0000, Sakari Ailus wrote:
> > > > Hi Chi Yuan,
> > > > 
> > > > On Wed, May 08, 2024 at 10:51:49AM +0800, cy_huang@xxxxxxxxxxx wrote:
> > > > > From: ChiYuan Huang <cy_huang@xxxxxxxxxxx>
> > > > > 
> > > > > In v4l2_async_create_ancillary_links(), if v4l2 async notifier is
> > > > > created from v4l2 device, the v4l2 flash subdev async binding will enter
> > > > > the logic to create media link. Due to the subdev of notifier is NULL,
> > > > > this will cause NULL pointer to access the subdev entity. Therefore, add
> > > > > the check to bypass it.
> > > > > 
> > > > > Fixes: aa4faf6eb271 ("media: v4l2-async: Create links during v4l2_async_match_notify()")
> > > > > Signed-off-by: ChiYuan Huang <cy_huang@xxxxxxxxxxx>
> > > > > ---
> > > > > Hi,
> > > > > 
> > > > >   I'm trying to bind the v4l2 subdev for flashlight testing. It seems
> > > > > some logic in v4l2 asynd binding is incorrect.
> > > > > 
> > > > > From the change, I modified vim2m as the test driver to bind mt6370 flashlight.
> > > > > 
> > > > > Here's the backtrace log.
> > > > > 
> > > > >  vim2m soc:vim2m: bound [white:flash-2]
> > > > >  Unable to handle kernel NULL pointer dereference at virtual address 0000000000000058
> > > > >  ......skipping
> > > > >  Call trace:
> > > > >   media_create_ancillary_link+0x48/0xd8 [mc]
> > > > >   v4l2_async_match_notify+0x17c/0x208 [v4l2_async]
> > > > >   v4l2_async_register_subdev+0xb8/0x1d0 [v4l2_async]
> > > > 
> > > > There's something wrong obviously somewhere but wherea?
> > >
> > > In vim2m driver, I added v4l2_async_nf_init -> v4l2_async_nf_add_fwnode_remote ->
> > > v4l2_async_nf_register.
> > > 
> > > From the async flow, in notifier complete ops to create v4l-subdevX node for the 
> > > specified subdev.
> > >
> > > > A sub-notifier does have a sub-device after the notifier initialisation.
> > > 
> > > Why? Are you saying to the notifier can only be used for subdev and subdev binding, 
> > > not v4l2 and subdev binding?
> > > 
> > > But to create v4l-subdevX, the key is only v4l2 device and its needed subdev.
> > > 
> > > > Maybe the initialisation does not happen in the right order?
> > >
> > > AFAIK, Async flow can solve the probe order and makes the user no need to care
> > > the probe order.
> > > 
> > > From the stacktrace, I'm pretty sure it's not the probe order issue.
> > >
> > > > >   __v4l2_flash_init.part.0+0x3b4/0x4b0 [v4l2_flash_led_class]
> > > > >   v4l2_flash_init+0x28/0x48 [v4l2_flash_led_class]
> > > > >   mt6370_led_probe+0x348/0x690 [leds_mt6370_flash]
> > > > > 
> > > > > After tracing the code, it will let the subdev labeled as F_LENS or
> > > > > F_FLASH function to create media link. To prevent the NULL pointer
> > > > > issue, the simplest way is add a check when 'n->sd' is NULL and bypass
> > > > > the later media link creataion.
> > > > > ---
> > > > >  drivers/media/v4l2-core/v4l2-async.c | 3 +++
> > > > >  1 file changed, 3 insertions(+)
> > > > > 
> > > > > diff --git a/drivers/media/v4l2-core/v4l2-async.c b/drivers/media/v4l2-core/v4l2-async.c
> > > > > index 3ec323bd528b..9d3161c51954 100644
> > > > > --- a/drivers/media/v4l2-core/v4l2-async.c
> > > > > +++ b/drivers/media/v4l2-core/v4l2-async.c
> > > > > @@ -324,6 +324,9 @@ static int v4l2_async_create_ancillary_links(struct v4l2_async_notifier *n,
> > > > >  	    sd->entity.function != MEDIA_ENT_F_FLASH)
> > > > >  		return 0;
> > > > >  
> > > > > +	if (!n->sd)
> > > > > +		return 0;
> > > > 
> > > > This isn't the right fix: the ancillary link won't be created as a result.
> > >
> > > Due to the notifier is created by v4l2 device not subdev, this 'n->sd' is NULL.
> > > The NULL 'n->sd' will be referenced by the next flow 'media_create_ancillary_link'.
> > 
> > Ah, right. I took a new look into the code and agree this is a problem.
> > This probably hasn't been hit previously as the root notifier driver tends
> > not to have any lens or flash devices.
> > 
> > I'd change the commit message slightly:
> > 
> > --------8<-------------
> > In v4l2_async_create_ancillary_links(), ancillary links are created for
> > lens and flash sub-devices. These are sub-device to sub-device links and if
> > the async notifier is related to a V4L2 device, the source sub-device of
> > the ancillary link is NULL, leading to a NULL pointer dereference. Check
> > the notifier's sd field is non-NULL in v4l2_async_create_ancillary_links().
> > --------8<-------------
> 
> What's the use case for including lens or flash devices in the root
> notifier ? Shouldn't lens and flash subdevices always be linked to
> something ? We should of course not crash, but it seems that simply
> ignoring the subdevs and not linking them isn't a great idea either.

Yes, I think triggering this does require a very peculiar setup if not a
driver bug. We could also print a warning if this happens.

Also using the notifier's sub-device to create ancillary links is somewhat
opportunistic. We seem to rely on it currently but it would only seem
meaningful for a sensor in practice.

> 
> > > Or is it caused by the wrong usage? 
> > > 
> > > > > +
> > > > >  	link = media_create_ancillary_link(&n->sd->entity, &sd->entity);
> > > > >  
> > > > >  #endif
> 

-- 
Kind regards,

Sakari Ailus