Re: [PATCH] drm/panthor: fix for dma-fence safe access rules

[Boris Brezillon <boris.brezillon@collabora.com>]

On Thu, 4 Dec 2025 09:42:37 -0800
Chia-I Wu <olvaffe@gmail.com> wrote:

> On Thu, Dec 4, 2025 at 1:27 AM Tvrtko Ursulin <tvrtko.ursulin@igalia.com> wrote:
> >
> >
> > On 04/12/2025 01:50, Chia-I Wu wrote:  
> > > Commit 506aa8b02a8d6 ("dma-fence: Add safe access helpers and document
> > > the rules") details the dma-fence safe access rules. The most common
> > > culprit is that drm_sched_fence_get_timeline_name may race with
> > > group_free_queue.
> > >
> > > Fixes: d2624d90a0b77 ("drm/panthor: assign unique names to queues")
> > > Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> > > ---
> > >   drivers/gpu/drm/panthor/panthor_sched.c | 4 ++++
> > >   1 file changed, 4 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> > > index 33b9ef537e359..a8b1347e4da71 100644
> > > --- a/drivers/gpu/drm/panthor/panthor_sched.c
> > > +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> > > @@ -23,6 +23,7 @@
> > >   #include <linux/module.h>
> > >   #include <linux/platform_device.h>
> > >   #include <linux/pm_runtime.h>
> > > +#include <linux/rcupdate.h>
> > >
> > >   #include "panthor_devfreq.h"
> > >   #include "panthor_device.h"
> > > @@ -923,6 +924,9 @@ static void group_release_work(struct work_struct *work)
> > >                                                  release_work);
> > >       u32 i;
> > >
> > > +     /* dma-fences may still be accessing group->queues under rcu lock. */
> > > +     synchronize_rcu();
> > > +
> > >       for (i = 0; i < group->queue_count; i++)
> > >               group_free_queue(group, group->queues[i]);
> > >  
> >
> > This handles the shared queue->fence_ctx.lock as well (which is also
> > unsafe until Christian lands the inline lock, etc patch series) so it
> > looks good to me as well.  
> Yeah, I will send v2 to drop the misleading "Fixes:" tag.
> 
> FWIW, the UAF I saw was from accessing the string returned by
> 
> static const char *drm_sched_fence_get_timeline_name(struct dma_fence *f)
> {
>         struct drm_sched_fence *fence = to_drm_sched_fence(f);
>         return (const char *)fence->sched->name;
> }

IIRC, the only place calling this callback is some debugsfs knob
dumping fences attached to dma_buf resvs, and we're not supposed to
expose our driver fences to the outside world (we use the
drm_sched_fence proxy for that), so I'm curious where the access was
coming from.

> 
> I thought it was "name" and added the "Fixes:" tag. But actually
> "sched" was also freed by group_release_work.
> 
> >
> > Just to mention an alternative could be to simply switch release_work to
> > INIT_RCU_WORK/queue_rcu_work, but I am not sure if that has an advantage.
> >
> > Regards,
> >
> > Tvrtko
> >