[PATCH v2 1/3] drm/xe/userptr: restore invalidation list on error

[Matthew Brost]

On Mon, Feb 17, 2025 at 07:58:11PM -0800, Matthew Brost wrote:
> On Mon, Feb 17, 2025 at 09:38:26AM +0000, Matthew Auld wrote:
> > On 15/02/2025 01:28, Matthew Brost wrote:
> > > On Fri, Feb 14, 2025 at 05:05:28PM +0000, Matthew Auld wrote:
> > > > On error restore anything still on the pin_list back to the invalidation
> > > > list on error. For the actual pin, so long as the vma is tracked on
> > > > either list it should get picked up on the next pin, however it looks
> > > > possible for the vma to get nuked but still be present on this per vm
> > > > pin_list leading to corruption. An alternative might be then to instead
> > > > just remove the link when destroying the vma.
> > > > 
> > > > Fixes: ed2bdf3b264d ("drm/xe/vm: Subclass userptr vmas")
> > > > Suggested-by: Matthew Brost <matthew.brost at intel.com>
> > > > Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> > > > Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> > > > Cc: <stable at vger.kernel.org> # v6.8+
> > > > ---
> > > >   drivers/gpu/drm/xe/xe_vm.c | 26 +++++++++++++++++++-------
> > > >   1 file changed, 19 insertions(+), 7 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > > > index d664f2e418b2..668b0bde7822 100644
> > > > --- a/drivers/gpu/drm/xe/xe_vm.c
> > > > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > > > @@ -670,12 +670,12 @@ int xe_vm_userptr_pin(struct xe_vm *vm)
> > > >   	list_for_each_entry_safe(uvma, next, &vm->userptr.invalidated,
> > > >   				 userptr.invalidate_link) {
> > > >   		list_del_init(&uvma->userptr.invalidate_link);
> > > > -		list_move_tail(&uvma->userptr.repin_link,
> > > > -			       &vm->userptr.repin_list);
> > > > +		list_add_tail(&uvma->userptr.repin_link,
> > > > +			      &vm->userptr.repin_list);
> > > 
> > > Why this change?
> > 
> > Just that with this patch the repin_link should now always be empty at this
> > point, I think. add should complain if that is not the case.
> > 
> 
> If it is always expected to be empty, then yea maybe add a xe_assert for
> this as the list management is pretty tricky. 
> 
> > > 
> > > >   	}
> > > >   	spin_unlock(&vm->userptr.invalidated_lock);
> > > > -	/* Pin and move to temporary list */
> > > > +	/* Pin and move to bind list */
> > > >   	list_for_each_entry_safe(uvma, next, &vm->userptr.repin_list,
> > > >   				 userptr.repin_link) {
> > > >   		err = xe_vma_userptr_pin_pages(uvma);
> > > > @@ -691,10 +691,10 @@ int xe_vm_userptr_pin(struct xe_vm *vm)
> > > >   			err = xe_vm_invalidate_vma(&uvma->vma);
> > > >   			xe_vm_unlock(vm);
> > > >   			if (err)
> > > > -				return err;
> > > > +				break;
> > > >   		} else {
> > > > -			if (err < 0)
> > > > -				return err;
> > > > +			if (err)
> > > > +				break;
> > > >   			list_del_init(&uvma->userptr.repin_link);
> > > >   			list_move_tail(&uvma->vma.combined_links.rebind,
> > > > @@ -702,7 +702,19 @@ int xe_vm_userptr_pin(struct xe_vm *vm)
> > > >   		}
> > > >   	}
> > > > -	return 0;
> > > > +	if (err) {
> > > > +		down_write(&vm->userptr.notifier_lock);
> > > 
> > > Can you explain why you take the notifier lock here? I don't think this
> > > required unless I'm missing something.
> > 
> > For the invalidated list, the docs say:
> > 
> > "Removing items from the list additionally requires @lock in write mode, and
> > adding items to the list requires the @userptr.notifer_lock in write mode."
> > 
> > Not sure if the docs needs to be updated here?
> > 
> 
> Oh. I believe the part of comment for 'adding items to the list
> requires the @userptr.notifer_lock in write mode' really means something
> like this:
> 
> 'When adding to @vm->userptr.invalidated in the notifier the
> @userptr.notifer_lock in write mode protects against concurrent VM binds
> from setting up newly invalidated pages.'
> 
> So with above and since this code path is in the VM bind path (i.e. we
> are not racing with other binds) I think the
> vm->userptr.invalidated_lock is sufficient. Maybe ask Thomas if he
> agrees here.
> 

After some discussion with Thomas, removing notifier lock here is safe.

However, for adding is either userptr.notifer_lock || vm->lock to also
avoid races between binds, execs, and rebind worker.

I'd like update the documentation and add a helper like this:

void xe_vma_userptr_add_invalidated(struct xe_userptr_vma *uvma)
{
       struct xe_vm *vm = xe_vma_vm(&uvma->vma);

       lockdep_assert(lock_is_held_type(&vm->lock.dep_map, 1) ||
                      lock_is_held_type(&vm->userptr.notifier_lock.dep_map, 1));

       spin_lock(&vm->userptr.invalidated_lock);
       list_move_tail(&uvma->userptr.invalidate_link,
                      &vm->userptr.invalidated);
       spin_unlock(&vm->userptr.invalidated_lock);
}

However, let's delay the helper until this series and recently post
series of mine [1] merge as both are fixes series and hoping for a clean
backport.

Matt

[1] https://patchwork.freedesktop.org/series/145198/

> Matt
> 
> > > 
> > > Matt
> > > 
> > > > +		spin_lock(&vm->userptr.invalidated_lock);
> > > > +		list_for_each_entry_safe(uvma, next, &vm->userptr.repin_list,
> > > > +					 userptr.repin_link) {
> > > > +			list_del_init(&uvma->userptr.repin_link);
> > > > +			list_move_tail(&uvma->userptr.invalidate_link,
> > > > +				       &vm->userptr.invalidated);
> > > > +		}
> > > > +		spin_unlock(&vm->userptr.invalidated_lock);
> > > > +		up_write(&vm->userptr.notifier_lock);
> > > > +	}
> > > > +	return err;
> > > >   }
> > > >   /**
> > > > -- 
> > > > 2.48.1
> > > > 
> >