[PATCH v4 3/4] drm/xe: Hold a PM ref when GT TLB invalidations are inflight

[Matthew Brost]

On Fri, Jul 19, 2024 at 08:50:02AM +0100, Matthew Auld wrote:
> On 18/07/2024 19:55, Matthew Brost wrote:
> > Avoid GT TLB invalidation timeouts by holding a PM ref when
> > invalidations are inflight.
> > 
> > v2:
> >   - Drop PM ref before signaling fence (CI)
> > 
> > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> > Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > Cc: Nirmoy Das <nirmoy.das at intel.com>
> > Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> > Reviewed-by: Nirmoy Das <nirmoy.das at intel.com>
> > ---
> >   drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c   | 62 ++++++++++++-------
> >   drivers/gpu/drm/xe/xe_gt_tlb_invalidation.h   |  1 +
> >   .../gpu/drm/xe/xe_gt_tlb_invalidation_types.h |  4 ++
> >   drivers/gpu/drm/xe/xe_vm.c                    |  4 +-
> >   4 files changed, 47 insertions(+), 24 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > index 668c1a3f06ac..481d83d07367 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > @@ -13,6 +13,7 @@
> >   #include "xe_guc.h"
> >   #include "xe_guc_ct.h"
> >   #include "xe_mmio.h"
> > +#include "xe_pm.h"
> >   #include "xe_sriov.h"
> >   #include "xe_trace.h"
> >   #include "regs/xe_guc_regs.h"
> > @@ -35,6 +36,24 @@ static long tlb_timeout_jiffies(struct xe_gt *gt)
> >   	return hw_tlb_timeout + 2 * delay;
> >   }
> > +static void
> > +__invalidation_fence_signal(struct xe_device *xe, struct xe_gt_tlb_invalidation_fence *fence)
> > +{
> > +	bool stack = test_bit(FENCE_STACK_BIT, &fence->base.flags);
> > +
> > +	trace_xe_gt_tlb_invalidation_fence_signal(xe, fence);
> > +	xe_gt_tlb_invalidation_fence_fini(fence);
> > +	dma_fence_signal(&fence->base);
> > +	if (!stack)
> > +		dma_fence_put(&fence->base);
> > +}
> > +
> > +static void
> > +invalidation_fence_signal(struct xe_device *xe, struct xe_gt_tlb_invalidation_fence *fence)
> > +{
> > +	list_del(&fence->link);
> > +	__invalidation_fence_signal(xe, fence);
> > +}
> >   static void xe_gt_tlb_fence_timeout(struct work_struct *work)
> >   {
> > @@ -56,10 +75,8 @@ static void xe_gt_tlb_fence_timeout(struct work_struct *work)
> >   		xe_gt_err(gt, "TLB invalidation fence timeout, seqno=%d recv=%d",
> >   			  fence->seqno, gt->tlb_invalidation.seqno_recv);
> > -		list_del(&fence->link);
> >   		fence->base.error = -ETIME;
> > -		dma_fence_signal(&fence->base);
> > -		dma_fence_put(&fence->base);
> > +		invalidation_fence_signal(xe, fence);
> 
> Should this be moved to the previous patch, which introduces the fence on
> stack?

Yes, will fix when merging.

Matt