[PATCH 2/2] drm/xe: RFC Deny unbinds if uapi ufence pending

Matthew Brost matthew.brost at intel.com
Thu Feb 15 06:33:33 UTC 2024


On Wed, Feb 14, 2024 at 04:15:58PM +0200, Mika Kuoppala wrote:
> Mika Kuoppala <mika.kuoppala at linux.intel.com> writes:
> 
> > If user fence was provided for MAP in vm_bind_ioctl
> > and it has still not been signalled, deny UNMAP of said
> > vma with EBUSY as long as unsignalled fence exists.
> >
> > This guarantees that MAP vs UNMAP sequences won't
> > escape under the radar if we ever want to track the
> > client's state wrt to completed and accessible MAPs.
> > By means of intercepting the ufence release signalling.
> >
> > Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1159
> > Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> > Cc: Matthew Brost <matthew.brost at intel.com>
> > Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > Signed-off-by: Mika Kuoppala <mika.kuoppala at linux.intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_vm.c       | 22 ++++++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_vm_types.h |  7 +++++++
> >  2 files changed, 29 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > index 836a6e849cda8..c26297568e697 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -897,6 +897,11 @@ static void xe_vma_destroy_late(struct xe_vma *vma)
> >  	struct xe_device *xe = vm->xe;
> >  	bool read_only = xe_vma_read_only(vma);
> >  
> > +	if (vma->ufence) {
> > +		xe_sync_ufence_put(vma->ufence);
> > +		vma->ufence = NULL;
> > +	}
> > +
> >  	if (xe_vma_is_userptr(vma)) {
> >  		struct xe_userptr *userptr = &to_userptr_vma(vma)->userptr;
> >  
> > @@ -1608,6 +1613,16 @@ xe_vm_unbind_vma(struct xe_vma *vma, struct xe_exec_queue *q,
> >  
> >  	trace_xe_vma_unbind(vma);
> >  
> > +	if (vma->ufence) {
> > +		struct xe_user_fence * const f = vma->ufence;
> > +
> > +		if (!xe_sync_ufence_get_status(f))
> > +			return ERR_PTR(-EBUSY);
> > +
> > +		vma->ufence = NULL;
> > +		xe_sync_ufence_put(f);
> > +	}
> > +
> >  	if (number_tiles > 1) {
> >  		fences = kmalloc_array(number_tiles, sizeof(*fences),
> >  				       GFP_KERNEL);
> > @@ -1751,6 +1766,13 @@ static int __xe_vm_bind(struct xe_vm *vm, struct xe_vma *vma,
> >  
> >  	xe_vm_assert_held(vm);
> >  
> > +	if (num_syncs == 1 && xe_sync_is_ufence(&syncs[0])) {

Maybe a helper here? Also I think it is possible for num_syncs > 1 and
have a ufence too.

> > +		if (XE_WARN_ON(vma->ufence))
> 
> This will get triggered with xe_exec_threads: threads-cm-userptr.
> 
> Should we just silently swap into a newer provided ufence
> and only complain if the existing fence in here is still non signalled?
> 

CI caught a failure too [1].

I think the failure in threads-cm-userptr must be a userptr invalidation
triggering a rebind? In this case the ufence will not be signaled but
vma->ufence == xe_sync_is_ufence(&syncs[0]).

For [1], I think it is possible that ufence could not be signaled too.

With that I believe is ok just to swap without an error.

Matt

[1] https://intel-gfx-ci.01.org/tree/intel-xe/xe-pw-129900v1/bat-pvc-2/igt@xe_exec_fault_mode@twice-basic-prefetch.html

> -Mika
> 
> > +			xe_sync_ufence_put(vma->ufence);
> > +
> > +		vma->ufence = xe_sync_ufence_get(&syncs[0]);
> > +	}
> > +
> >  	if (immediate) {
> >  		fence = xe_vm_bind_vma(vma, q, syncs, num_syncs, first_op,
> >  				       last_op);
> > diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
> > index 5ac9c5bebabc3..4a06420b941ea 100644
> > --- a/drivers/gpu/drm/xe/xe_vm_types.h
> > +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> > @@ -19,6 +19,7 @@
> >  
> >  struct xe_bo;
> >  struct xe_sync_entry;
> > +struct xe_user_fence;
> >  struct xe_vm;
> >  
> >  #define XE_VMA_READ_ONLY	DRM_GPUVA_USERBITS
> > @@ -102,6 +103,12 @@ struct xe_vma {
> >  	 * @pat_index: The pat index to use when encoding the PTEs for this vma.
> >  	 */
> >  	u16 pat_index;
> > +
> > +	/**
> > +	 * @ufence: The user fence that was provided with MAP.
> > +	 * Needs to be signalled before UNMAP can be processed.
> > +	 */
> > +	struct xe_user_fence *ufence;
> >  };
> >  
> >  /**
> > -- 
> > 2.34.1


More information about the Intel-xe mailing list