On Thu, 7 May 2026 14:38:23 -0700
Rob Clark <[email protected]> wrote:

> On Thu, May 7, 2026 at 5:46 AM Boris Brezillon
> <[email protected]> wrote:
> >
> > On Wed, 06 May 2026 14:16:27 +0200
> > Boris Brezillon <[email protected]> wrote:
> >  
> > > The following race can currently happen:
> > >
> > > | Thread 0 in `drm_gem_lru_scan`               | Thread 1 in 
> > > `drm_gem_object_release` |
> > > | -                                            | -                        
> > >             |
> > > | move obj1 with refcount==0 to `still_in_lru` |                          
> > >             |
> > > | move obj2 with refcount!=0 to `still_in_lru` |                          
> > >             |
> > > | mutex_unlock                                 |                          
> > >             |
> > > | shrink obj2                                  |                          
> > >             |
> > > |                                              | lru = obj1->lru; // 
> > > `still_in_lru`   |
> > > | mutex_lock                                   |                          
> > >             |
> > > | move obj1 back to the original lru           |                          
> > >             |
> > > | mutex_unlock                                 |                          
> > >             |
> > > | return                                       |                          
> > >             |
> > > |                                              | dereference 
> > > `still_in_lru`           |
> > >
> > > Move the drm_gem_lru_move_tail_locked() after the
> > > kref_get_unless_zero() check so that we don't end up with a
> > > vanishing LRU when we hit drm_gem_object_release(). We also need to
> > > remove the skipped object from its LRU, otherwise we'll keep hitting
> > > it on subsequent loop iterations until it's actually removed from the
> > > list in the drm_gem_release().
> > >
> > > Fixes: e7c2af13f811 ("drm/gem: Add LRU/shrinker helper")
> > > Reported-by: Chia-I Wu <[email protected]>
> > > Closes: https://gitlab.freedesktop.org/panfrost/linux/-/work_items/86
> > > Signed-off-by: Boris Brezillon <[email protected]>
> > > Reviewed-by: Chia-I Wu <[email protected]>
> > > ---
> > >  drivers/gpu/drm/drm_gem.c | 14 +++++++++-----
> > >  1 file changed, 9 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> > > index fca42949eb2b..97cf63de0112 100644
> > > --- a/drivers/gpu/drm/drm_gem.c
> > > +++ b/drivers/gpu/drm/drm_gem.c
> > > @@ -1660,15 +1660,19 @@ drm_gem_lru_scan(struct drm_gem_lru *lru,
> > >               if (!obj)
> > >                       break;
> > >
> > > -             drm_gem_lru_move_tail_locked(&still_in_lru, obj);
> > > -
> > >               /*
> > >                * If it's in the process of being freed, gem_object->free()
> > > -              * may be blocked on lock waiting to remove it.  So just
> > > -              * skip it.
> > > +              * may be blocked on lock waiting to remove it.  So just 
> > > remove
> > > +              * it from its current LRU and skip it.
> > >                */
> > > -             if (!kref_get_unless_zero(&obj->refcount))
> > > +             if (!kref_get_unless_zero(&obj->refcount)) {
> > > +                     if (obj->lru)
> > > +                             drm_gem_lru_remove_locked(obj);
> > > +  
> >
> > Actually, this thing is still racy, because obj->lru is dereferenced
> > without the lru->lock held in drm_gem_object_release(). At this point
> > I'm wondering if we should expose a drm_gem_lru_remove() taking the LRU
> > lock as an argument as suggested by Steve, and delegate the
> > responsibility to call drm_gem_lru_remove() to the driver. Either that,
> > or we make it so the LRU lock is attached to the drm_device instead of
> > the GEM (both MSM and panthor assume a device-wide lock for LRU
> > manipulation).
> >
> > Rob, what's your take on this matter?  
> 
> I don't think there is a race, because of the kref_get_unless_zero().
> Other than lru_scan, there shouldn't be cases where someone is moving
> an obj between LRUs racing with drm_gem_object_release(), because that
> means they don't own a reference on the obj they are manipulating.

Yeah, but the race I'm talking about is drm_gem_object_release()
vs drm_gem_lru_scan(), so at this point refcount is zero, and this
patch only moves the needle, but doesn't fix the problem entirely:


| Thread 0 in `drm_gem_lru_scan`               | Thread 1 in 
`drm_gem_object_release` |
| -                                            | -                              
      |
|                                              | drm_gem_lru_remove()           
      |
|                                              |    lru = obj->lru              
      |
|                                              |    if (!lru) return;           
      |
| lock(still_in_lru.lock)                      |                                
      |
|    if (refcount == 0)                        |                                
      |
|       drm_gem_lru_remove_locked(obj)         |                                
      |
|         obj->lru = NULL                      |                                
      |
|    .....                                     |                                
      |
| unlock(still_in_lru.lock)                    |                                
      |
|                                              |    lock(lru->lock)             
      |
|                                              |       
drm_gem_lru_remove_locked(obj) |
|                                              |         obj->lru==NULL => NULL 
deref |
|                                              |    unlock(lru->lock)           
      |

We can of course add an extra

        if (!obj->lru) return;

in drm_gem_lru_remove_locked() to cover for this race, and add a
READ_ONCE in drm_gem_lru_remove() to make sure the compiler doesn't
do crazy things like dereferencing obj->lru twice instead of having
the LRU pointer stored in a register. That still assumes that the lru
we assigned to our local variable is valid even after the
drm_gem_lru_remove_locked(obj) call, which is true at least for MSM and
and panthor because they have their LRUs attached to the drm_device,
which outlives any GEMs attached to it. But it's not something the API
enforce or document as a requirement.

> 
> That said, I can't really think of a sensible thing to do with more
> than a single LRU lock per device.  And it does make things easier to
> reason about.

Okay, I'll give it a try then.

Reply via email to