Re: [PATCH] drm/xe: Always invalidate TLBs on userptr invalidation

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

On Mon, Mar 23, 2026 at 09:17:42PM +0000, Stuart Summers wrote:
> Right now we are only invalidating TLBs when we are
> running in fault mode. For non-fault mode based userptr
> VMs, we then rely on context switches after an MMAP has
> happened to ensure the TLB is clean for the next submission.
> 
> With context based TLB invalidation, we can no longer rely
> on the implicit invalidation happening during context switch,
> so remove the fault mode limiter and simply always perform
> that invalidation.
> 
> I was able to see this behavior using the following test:
> xe_exec_compute_mode --r twice-userptr-invalidate
>

Hmm, this implies that preempt fences don't issue a TLB invalidation...
I thought we landed on preempt fences do a TLB invalidation in offline
discussions.

We may have more work to do here wrt context based TLB invalidations.

I think BO move path likely also needs to be updated then too.
 
> Signed-off-by: Stuart Summers <stuart.summers@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_userptr.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_userptr.c b/drivers/gpu/drm/xe/xe_userptr.c
> index 6761005c0b90..dfd679dd98d9 100644
> --- a/drivers/gpu/drm/xe/xe_userptr.c
> +++ b/drivers/gpu/drm/xe/xe_userptr.c
> @@ -102,7 +102,7 @@ xe_vma_userptr_do_inval(struct xe_vm *vm, struct xe_userptr_vma *uvma, bool is_d
>  				    false, MAX_SCHEDULE_TIMEOUT);
>  	XE_WARN_ON(err <= 0);
>  
> -	if (xe_vm_in_fault_mode(vm) && userptr->initial_bind) {
> +	if (userptr->initial_bind) {

Should be change this if statement based on hardware support?

Matt

>  		if (!userptr->finish_inuse) {
>  			/*
>  			 * Defer the TLB wait to an extra pass so the caller
> -- 
> 2.43.0
>