On Thu, 7 May 2026 15:40:51 +0100 Liviu Dudau <[email protected]> wrote:
> On Thu, May 07, 2026 at 02:10:27PM +0200, Boris Brezillon wrote: > > On Thu, 7 May 2026 11:01:25 +0100 > > Liviu Dudau <[email protected]> wrote: > > > > > On Wed, May 06, 2026 at 02:16:26PM +0200, Boris Brezillon wrote: > > > > drm_gem_lru_remove() dereference stores drm_gem_object::lru in a local > > > > variable that's then dereferenced to acquire the LRU lock. Because this > > > > assignment in done without the LRU lock held, it can race with > > > > drm_gem_lru_scan() where drm_gem_object::lru is temporarily assigned > > > > a stack-allcated LRU that goes away when leaving the function. By > > > > the time we dereference this local lru variable, the object might > > > > already > > > > be gone. > > > > > > > > It feels like drm_gem_lru_move_tail() was never meant to be used this > > > > way, because there's no easy way we can avoid this race unless we defer > > > > the locking to the caller. Let's add an explicit LRU for unreclaimable > > > > BOs instead, and have all BOs added to this LRU at creation time. > > > > > > I would argue that drm_gem_lru_scan() is broken by design. If you're going > > > to release the LRU lock in the middle of a loop you can expect that > > > someone > > > will get hold of your stack-allocated LRU and end up picking the pieces. > > > > I think it's fine as long as you always use the drm_gem_lru helpers to > > manipulate the lru field, which is true of a lot of kernel constructs. > > I think drm_gem_lru_scan() should never set an object's lru field to > still_in_lru. > It should set it to NULL when the object's node is removed from its lru and > add > it into still_in_lru without making the drm_gem_object->lru to point back to > it. > At the very end when we splice back the still_in_lru list back into lru's > list we > can then update obj->lru. Then you run into another race between drm_gem_lru_scan() and drm_gem_object_release(), where the LRU removal in _release() is skipped because obj->lru is NULL, and all of a sudden, the still_in_lru list has an element that's freed. Honestly, I don't think obj->lru pointing to a stack allocated object is a problem as long as we don't let gem users play freely with obj->lru (which we shouldn't do anyway). > > > > > > This patch is fine in itself by trying to avoid stepping into the fight, > > > but I think we should also add a warning in drm_gem_lru_scan() for future > > > users to be aware of the dangers. > > > > Warning the user about what? There's nothing they can do about it, and > > I don't even think it's unsafe per-se, unless someone goes off and > > stores the drm_gem_object::lru value somewhere else while their shrink() > > callback is called, and accesses it later, outside the shrinker path. > > Given drm_gem_lru is not refcounted, there's no way one could safely > > hold on the LRU they saw in the shrink() callback anyway, so I don't > > think that's fair to blame the drm_gem_lru API for this kind of misuse. > > Yeah, that would be the warning: don't store the object's lru as you might > get a temporary one that will become invalid after the shrinker has run. Oh, you mean a comment explaining this should be avoided, not an actual drm_warn(). Then yes, I think it's fine to document the expectations in the drm_gem_object::lru doc.
