Re: [PATCH] drm/xe/vm: nuke PTs only after unlinking contested VMAs

[Matthew Brost <matthew.brost@intel.com>]

On Fri, Sep 18, 2026 at 04:35:43PM +0200, Thomas Hellström wrote:
> On Fri, 2026-09-18 at 14:10 +0100, Matthew Auld wrote:
> > In xe_vm_close_and_put(), external-BO VMAs are queued on the
> > contested
> > list for deferred destruction via xe_vma_destroy_unlocked(). However,
> > xe_vm_pt_destroy() was previously invoked before processing contested
> > VMAs, destroying vm->pt_root while those VMAs were still linked to
> > their
> > respective buffer objects (vm_bo->list.gpuva).
> > 
> > If a concurrent thread evicts one of those shared buffer objects,
> > xe_bo_trigger_rebind() holding only bo->resv walks the BO's VMAs and,
> > in
> > fault mode, calls xe_vm_invalidate_vma() -> xe_pt_zap_ptes(). Because
> > vm->pt_root[tile->id] is already NULL, dereferencing pt->level causes
> > a
> > NULL ptr deref.
> > 
> > Fix this by deferring xe_vm_free_scratch() and xe_vm_pt_destroy()
> > until
> > after all contested VMAs have been unlinked and destroyed.
> > 
> > User is reporting hitting a NULL ptr deref in xe_pt_zap_ptes(), which
> > could be explained by this race.
> > 
> > Assisted-by: LLM
> > Fixes: b06d47be7c83 ("drm/xe: Port Xe to GPUVA")
> > Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/9290
> > Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > Cc: Matthew Brost <matthew.brost@intel.com>
> > Cc: <stable@vger.kernel.org> # v6.12+
> 
> Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> 

I actually had a similar fix here [1] but I believe I missed the scratch
part.

Glad to see this fix getting in:
Reviewed-by: Matthew Brost <matthew.brost@intel.com>

[1] https://patchwork.freedesktop.org/patch/738642/?series=170045&rev=1

> 
> > ---
> >  drivers/gpu/drm/xe/xe_vm.c | 21 +++++++++------------
> >  1 file changed, 9 insertions(+), 12 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > index 17dc4debe7c1..390da884c727 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -1979,21 +1979,13 @@ void xe_vm_close_and_put(struct xe_vm *vm)
> >  		vma->gpuva.flags |= XE_VMA_DESTROYED;
> >  	}
> >  
> > -	/*
> > -	 * All vm operations will add shared fences to resv.
> > -	 * The only exception is eviction for a shared object,
> > -	 * but even so, the unbind when evicted would still
> > -	 * install a fence to resv. Hence it's safe to
> > -	 * destroy the pagetables immediately.
> > -	 */
> > -	xe_vm_free_scratch(vm);
> > -	xe_vm_pt_destroy(vm);
> >  	xe_vm_unlock(vm);
> >  
> >  	/*
> > -	 * VM is now dead, cannot re-add nodes to vm->vmas if it's
> > NULL
> > -	 * Since we hold a refcount to the bo, we can remove and
> > free
> > -	 * the members safely without locking.
> > +	 * Unlink and destroy all contested external-BO VMAs before
> > destroying
> > +	 * the page tables. Otherwise, concurrent eviction holding
> > only bo->resv
> > +	 * can walk the BO's VMAs and attempt to invalidate/zap page
> > tables that
> > +	 * have already been freed.
> >  	 */
> >  	list_for_each_entry_safe(vma, next_vma, &contested,
> >  				 combined_links.destroy) {
> > @@ -2001,6 +1993,11 @@ void xe_vm_close_and_put(struct xe_vm *vm)
> >  		xe_vma_destroy_unlocked(vma);
> >  	}
> >  
> > +	xe_vm_lock(vm, false);
> > +	xe_vm_free_scratch(vm);
> > +	xe_vm_pt_destroy(vm);
> > +	xe_vm_unlock(vm);
> > +
> >  	xe_svm_fini(vm);
> >  
> >  	up_write(&vm->lock);