Re: [PATCH] drm/xe/gt: Abort GT init on LMTT initialization failure

From: Piotr Piórkowski

Date: Thu Nov 19 2026 - 04:50:15 EST


Summers, Stuart <stuart.summers@xxxxxxxxx> wrote on śro [2026-lis-18 21:03:09 +0100]:
> On Wed, 2026-11-18 at 16:54 +0800, yaolu@xxxxxxxxxx wrote:
> > From: Lu Yao <yaolu@xxxxxxxxxx>
> >
> > xe_lmtt_init() can fail, but gt_init_with_gt_forcewake() ignores its
> > return value. On a PF the probe then continues with a half-
> > initialized
> > LMTT: lmtt->pd is never set (and lmtt->ops is cleared), so
> > xe_lmtt_init_hw() returns early and the VF local memory is left
> > without
> > a translation directory while the driver still reports success.
> >
> > Return the xe_lmtt_init() error and abort GT init instead, as
> > required
> > by the documented LMTT contract ("Any LMTT initialization failure
> > should block VFs enabling").
> >
> > Fixes: b1d204058218 ("drm/xe/pf: Introduce Local Memory Translation
> > Table")
> > Signed-off-by: Lu Yao <yaolu@xxxxxxxxxx>
>
> I'm not an expert in this code, but it does look like if we tried to
> allocate memory here and that failed for some reason, causing the LMTT
> init to fail, it would cause downstream issues later in the PF's LMTT
> management. But the change as-is doesn't seem like the right approach
> here IMO. I don't think we want to kill the whole driver load because
> of this, it seems better to just log the error and not enumerate the
> VFs.
>
> But it would be good to get feedback from Piotr/Satya here.

I agree with Stuart that, since LMTT is only necessary for creating VFs, we could simply
block the creation of VFs (and all operations related to SR-IOV).
It seems that switching back to native mode at this point is already too late.
Interrupting the driver's probe is definitely too radical an approach.

We could do something like this:

--- a/drivers/gpu/drm/xe/xe_lmtt.h
+++ b/drivers/gpu/drm/xe/xe_lmtt.h
@@
#ifdef CONFIG_PCI_IOV
int xe_lmtt_init(struct xe_lmtt *lmtt);
+bool xe_lmtt_is_ready(struct xe_lmtt *lmtt);
void xe_lmtt_init_hw(struct xe_gt *gt);
--- a/drivers/gpu/drm/xe/xe_lmtt.c
+++ b/drivers/gpu/drm/xe/xe_lmtt.c
@@
+/**
+ * xe_lmtt_is_ready() - Check if LMTT is ready to be used.
+ * @lmtt: the &xe_lmtt to check
+ *
+ * This function shall be called only by PF.
+ *
+ * Return: true if LMTT is not needed on this device or was successfully
+ * initialized by xe_lmtt_init(), false otherwise.
+ */
+bool xe_lmtt_is_ready(struct xe_lmtt *lmtt)
+{
+ lmtt_assert(lmtt, IS_SRIOV_PF(lmtt_to_xe(lmtt)));
+
+ return !xe_device_has_lmtt(lmtt_to_xe(lmtt)) || lmtt->pd;
+}
--- a/drivers/gpu/drm/xe/xe_gt.c
+++ b/drivers/gpu/drm/xe/xe_gt.c
@@ static int gt_init_with_gt_forcewake(struct xe_gt *gt)
if (xe_gt_is_main_type(gt)) {
err = xe_ggtt_init(gt_to_tile(gt)->mem.ggtt);
if (err)
return err;
- if (IS_SRIOV_PF(gt_to_xe(gt)))
- xe_lmtt_init(&gt_to_tile(gt)->sriov.pf.lmtt);
+ if (IS_SRIOV_PF(gt_to_xe(gt))) {
+ err = xe_lmtt_init(&gt_to_tile(gt)->sriov.pf.lmtt);
+ if (err)
+ xe_gt_sriov_err(gt, "Failed to initialize LMTT (%pe), VFs will be unavailable\n",
+ ERR_PTR(err));
+ }
}
--- a/drivers/gpu/drm/xe/xe_sriov_pf.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf.c
@@
+static int pf_check_lmtt(struct xe_device *xe)
+{
+ struct xe_tile *tile;
+ unsigned int id;
+
+ for_each_tile(tile, xe, id)
+ if (!xe_lmtt_is_ready(&tile->sriov.pf.lmtt))
+ return -ENODEV;
+
+ return 0;
+}
+
int xe_sriov_pf_wait_ready(struct xe_device *xe)
{
struct xe_gt *gt;
unsigned int id;
int err;

if (xe_wedged(xe))
return -ECANCELED;

+ err = pf_check_lmtt(xe);
+ if (err)
+ return err;
+
for_each_gt(gt, xe, id) {

Thanks
Piotr

>
> Thanks,
> Stuart
>
> > ---
> >  drivers/gpu/drm/xe/xe_gt.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
> > index 286aee5775c0..1239cbb7c84c 100644
> > --- a/drivers/gpu/drm/xe/xe_gt.c
> > +++ b/drivers/gpu/drm/xe/xe_gt.c
> > @@ -582,10 +582,10 @@ static int gt_init_with_gt_forcewake(struct
> > xe_gt *gt)
> >  
> >         if (xe_gt_is_main_type(gt)) {
> >                 err = xe_ggtt_init(gt_to_tile(gt)->mem.ggtt);
> > +               if (!err && IS_SRIOV_PF(gt_to_xe(gt)))
> > +                       err = xe_lmtt_init(&gt_to_tile(gt)-
> > >sriov.pf.lmtt);
> >                 if (err)
> >                         return err;
> > -               if (IS_SRIOV_PF(gt_to_xe(gt)))
> > -                       xe_lmtt_init(&gt_to_tile(gt)->sriov.pf.lmtt);
> >         }
> >  
> >         /* Enable per hw engine IRQs */
>

--