Re: [GIT PULL FOR v4.18] R-Car VSP1 TLB optimisation

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Mauro,

On Saturday, 26 May 2018 14:28:18 EEST Mauro Carvalho Chehab wrote:
> Em Sat, 26 May 2018 03:24:00 +0300 Laurent Pinchart escreveu:

[snip]

> > I've reproduced the issue and created a minimal test case.
> > 
> >  1. struct vsp1_pipeline;
> >  2.
> >  3. struct vsp1_entity {
> >  4.         struct vsp1_pipeline *pipe;
> >  5.         struct vsp1_entity *sink;
> >  6.         unsigned int source_pad;
> >  7. };
> >  8.
> >  9. struct vsp1_pipeline {
> > 10.         struct vsp1_entity *brx;
> > 11. };
> > 12.
> > 13. struct vsp1_brx {
> > 14.         struct vsp1_entity entity;
> > 15. };
> > 16.
> > 17. struct vsp1_device {
> > 18.         struct vsp1_brx *bru;
> > 19.         struct vsp1_brx *brs;
> > 20. };
> > 21.
> > 22. unsigned int frob(struct vsp1_device *vsp1, struct vsp1_pipeline
> > *pipe)
> > 23. {
> > 24.         struct vsp1_entity *brx;
> > 25.
> > 26.         if (pipe->brx)
> > 27.                 brx = pipe->brx;
> > 28.         else if (!vsp1->bru->entity.pipe)
> > 29.                 brx = &vsp1->bru->entity;
> > 30.         else
> > 31.                 brx = &vsp1->brs->entity;
> > 32.
> > 33.         if (brx != pipe->brx)
> > 34.                 pipe->brx = brx;
> > 35.
> > 36.         return pipe->brx->source_pad;
> > 37. }
> > 
> > The reason why smatch complains is that it has no guarantee that vsp1->brs
> > is not NULL. It's quite tricky:
> > 
> > - On line 26, smatch assumes that pipe->brx can be NULL
> > - On line 27, brx is assigned a non-NULL value (as pipe->brx is not NULL
> > due to line 26)
> > - On line 28, smatch assumes that vsp1->bru is not NULL
> > - On line 29, brx is assigned a non-NULL value (as vsp1->bru is not NULL
> > due to line 28)
> > - On line 31, brx is assigned a possibly NULL value (as there's no
> > information regarding vsp1->brs)
> > - On line 34, pipe->brx is not assigned a non-NULL value if brx is NULL
> > - On line 36 pipe->brx is dereferenced
> > 
> > The problem comes from the fact that smatch assumes that vsp1->brs isn't
> > NULL. Adding a "(void)vsp1->brs->entity;" statement on line 25 makes the
> > warning disappear.
> > 
> > So how do we know that vsp1->brs isn't NULL in the original code ?
> > 
> >         if (pipe->num_inputs > 2)
> >                 brx = &vsp1->bru->entity;
> >         else if (pipe->brx && !drm_pipe->force_brx_release)
> >                 brx = pipe->brx;
> >         else if (!vsp1->bru->entity.pipe)
> >                 brx = &vsp1->bru->entity;
> >         else
> >                 brx = &vsp1->brs->entity;
> > 
> > A VSP1 instance can have no brs, so in general vsp1->brs can be NULL.
> > However, when that's the case, the following conditions are fulfilled.
> > 
> > - drm_pipe->force_brx_release will be false
> > - either pipe->brx will be non-NULL, or vsp1->bru->entity.pipe will be
> > NULL
> > 
> > The fourth branch should thus never be taken.
> 
> I don't think that adding a forth branch there would solve.
> 
> The thing is that Smatch knows that pipe->brx can be NULL, as the function
> explicly checks if pipe->brx != NULL.
> 
> When Smatch handles this if:
> 
> 	if (brx != pipe->brx) {
> 
> It wrongly assumes that this could be false if pipe->brx is NULL.
> I don't know why, as Smatch should know that brx can't be NULL.

brx can be NULL here if an only if vsp1->brs is NULL (as the entity field is 
first in the vsp1->brs structure, so &vsp1->brs->entity has the same address 
as vsp1->brs).

vsp1->brs can be NULL on some devices, but in that case we have the following 
guarantees:

- drm_pipe->force_brx_release will always be FALSE
- either pipe->brx will be non-NULL or vsp1->bru->entity.pipe will be NULL

So the fourth branch is never taken.

The above conditions come from outside this function, and smatch can't know 
about them. However, I don't know whether the problems comes from smatch 
assuming that vsp1->brs can be NULL, or from somewhere else.

> On such case, the next code to be executed would be:
> 
> 	format.pad = pipe->brx->source_pad;
> 
> With would be trying to de-ref a NULL pointer.
> 
> There are two ways to fix it:
> 
> 1) with my patch.
> 
> It is based to the fact that, if pipe->brx is null, then brx won't be
> NULL. So, the logic that "Switch BRx if needed." will always be called:
> 
> diff --git a/drivers/media/platform/vsp1/vsp1_drm.c
> b/drivers/media/platform/vsp1/vsp1_drm.c index 095dc48aa25a..cb6b60843400
> 100644
> --- a/drivers/media/platform/vsp1/vsp1_drm.c
> +++ b/drivers/media/platform/vsp1/vsp1_drm.c
> @@ -185,7 +185,7 @@ static int vsp1_du_pipeline_setup_brx(struct vsp1_device
> *vsp1, brx = &vsp1->brs->entity;
> 
>  	/* Switch BRx if needed. */
> -	if (brx != pipe->brx) {
> +	if (brx != pipe->brx || !pipe->brx) {
>  		struct vsp1_entity *released_brx = NULL;
> 
>  		/* Release our BRx if we have one. */
> 
> The code with switches BRx ensures that pipe->brx won't be null, as
> in the end, it sets:
> 
> 	pipe->brx = brx;
> 
> And brx can't be NULL.

The reason I don't like this is because the problem originally comes from the 
fact that smatch assumes that vsp1->brs can be NULL when it can't. I'd rather 
modify the code in a way that explicitly tests for vsp1->brs. However, smatch 
won't accept that happily :-/ I tried

        if (pipe->num_inputs > 2)
                brx = &vsp1->bru->entity;
        else if (pipe->brx && !drm_pipe->force_brx_release)
                brx = pipe->brx;
        else if (!vsp1->bru->entity.pipe)
                brx = &vsp1->bru->entity;
        else if (vsp1->brs)
                brx = &vsp1->brs->entity;
        else
                return -EINVAL;

and I still get the same warning. I had to write the following (which is 
obviously not correct) to silence the warning.

        if (pipe->num_inputs > 2)
                brx = &vsp1->bru->entity;
        else if (pipe->brx)
                brx = pipe->brx;
        else if (!vsp1->bru->entity.pipe)
                brx = &vsp1->bru->entity;
        else { 
                (void)vsp1->brs->entity;
                brx = &vsp1->brs->entity;
        }

Both the (void)vsp1->brs->entity and the removal of the !drm_pipe-
>force_brx_release were needed, any of those on its own didn't fix the 
problem.

> From my PoV, this patch has the advantage of explicitly showing
> to humans that the code inside the if statement will always be
> executed when pipe->brx is NULL.
> 
> -
> 
> Another way to solve would be to explicitly check if pipe->brx is still
> null before de-referencing:
> 
> diff --git a/drivers/media/platform/vsp1/vsp1_drm.c
> b/drivers/media/platform/vsp1/vsp1_drm.c index edb35a5c57ea..9fe063d6df31
> 100644
> --- a/drivers/media/platform/vsp1/vsp1_drm.c
> +++ b/drivers/media/platform/vsp1/vsp1_drm.c
> @@ -327,6 +327,9 @@ static int vsp1_du_pipeline_setup_brx(struct vsp1_device
> *vsp1, list_add_tail(&pipe->brx->list_pipe, &pipe->entities);
>  	}
> 
> +	if (!pipe->brx)
> +		return -EINVAL;
> +
>  	/*
>  	 * Configure the format on the BRx source and verify that it matches the
>  	 * requested format. We don't set the media bus code as it is configured
> 
> The right fix would be, instead, to fix Smatch to handle the:
> 
> 	if (brx != pipe->brx)
> 
> for the cases where one var can be NULL while the other can't be NULL,
> but, as I said before, I suspect that this can be a way more complex.

I'm not sure smatch is faulty here, or at least not when it interprets the brx 
!= pipe->brx check. The problem seems to come from the fact that is believes 
brx can be NULL.

-- 
Regards,

Laurent Pinchart