[PATCH] drm/xe/vm: prevent UAF in rebind_work_func()

[Matthew Brost]

On Wed, Apr 17, 2024 at 06:01:04PM +0000, Matthew Brost wrote:
> On Wed, Apr 17, 2024 at 05:31:08PM +0100, Matthew Auld wrote:
> > We flush the rebind worker during the vm close phase, however in places
> > like preempt_fence_work_func() we seem to queue the rebind worker
> > without first checking if the vm has already been closed.  The concern
> > here is the vm being closed with the worker flushed, but then being
> > rearmed later, which looks like potential uaf, since there is no actual
> > refcounting to track the queued worker. To ensure this can't happen
> > prevent queueing the rebind worker once the vm has been closed.
> > 
> > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> > Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1591
> > Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1304
> > Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1249
> > Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> > Cc: Matthew Brost <matthew.brost at intel.com>
> > Cc: <stable at vger.kernel.org> # v6.8+
> > ---
> >  drivers/gpu/drm/xe/xe_pt.c |  2 +-
> >  drivers/gpu/drm/xe/xe_vm.h | 17 ++++++++++++++---
> >  2 files changed, 15 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
> > index 5b7930f46cf3..e21461be904f 100644
> > --- a/drivers/gpu/drm/xe/xe_pt.c
> > +++ b/drivers/gpu/drm/xe/xe_pt.c
> > @@ -1327,7 +1327,7 @@ __xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue
> >  		}
> >  		if (!rebind && last_munmap_rebind &&
> >  		    xe_vm_in_preempt_fence_mode(vm))
> > -			xe_vm_queue_rebind_worker(vm);
> > +			xe_vm_queue_rebind_worker_locked(vm);
> >  	} else {
> >  		kfree(rfence);
> >  		kfree(ifence);
> > diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> > index 306cd0934a19..8420fbf19f6d 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.h
> > +++ b/drivers/gpu/drm/xe/xe_vm.h
> > @@ -211,10 +211,20 @@ int xe_vm_rebind(struct xe_vm *vm, bool rebind_worker);
> >  
> >  int xe_vm_invalidate_vma(struct xe_vma *vma);
> >  
> > -static inline void xe_vm_queue_rebind_worker(struct xe_vm *vm)
> > +static inline void xe_vm_queue_rebind_worker_locked(struct xe_vm *vm)
> >  {
> >  	xe_assert(vm->xe, xe_vm_in_preempt_fence_mode(vm));
> > -	queue_work(vm->xe->ordered_wq, &vm->preempt.rebind_work);
> > +	lockdep_assert_held(&vm->lock);
> > +
> > +	if (!xe_vm_is_closed(vm))
> 
> xe_vm_is_closed_or_banned
> 
> Otherwise LGTM. With the above changed:
> Reviewed-by: Matthew Brost <matthew.brost at intel.com>

Revoking RB, should have looked at CI first.

This doesn't work as it deadlocks [1].

- VMA invalidate notifier waits on VM dma-resv (preempt fences)
- Preempt fences signaled via worker which queues rebind worked (VM read lock), order WQ too so serially
- Preempt rebind worker waits on VMA invalidate under VM write lock

Going to have to rethink this and aside from this new bug in actually
rather dangerous too as taking the VM lock any path that block a
signaling of a fence seemly can deadlock. Wondering if we can use
lockdep to catch bugs like this?

In the middle of few other things right now so haven't thought a ton
about this but open to ideas.

Matt

[1] https://intel-gfx-ci.01.org/tree/intel-xe/xe-pw-132571v1/bat-atsm-2/igt@xe_exec_threads@threads-mixed-userptr-invalidate.html

> 
> > +		queue_work(vm->xe->ordered_wq, &vm->preempt.rebind_work);
> > +}
> > +
> > +static inline void xe_vm_queue_rebind_worker(struct xe_vm *vm)
> > +{
> > +	down_read(&vm->lock);
> > +	xe_vm_queue_rebind_worker_locked(vm);
> > +	up_read(&vm->lock);
> >  }
> >  
> >  /**
> > @@ -225,12 +235,13 @@ static inline void xe_vm_queue_rebind_worker(struct xe_vm *vm)
> >   * If the rebind functionality on a compute vm was disabled due
> >   * to nothing to execute. Reactivate it and run the rebind worker.
> >   * This function should be called after submitting a batch to a compute vm.
> > + *
> >   */
> >  static inline void xe_vm_reactivate_rebind(struct xe_vm *vm)
> >  {
> >  	if (xe_vm_in_preempt_fence_mode(vm) && vm->preempt.rebind_deactivated) {
> >  		vm->preempt.rebind_deactivated = false;
> > -		xe_vm_queue_rebind_worker(vm);
> > +		xe_vm_queue_rebind_worker_locked(vm);
> >  	}
> >  }
> >  
> > -- 
> > 2.44.0
> >