[PATCH v4 0/7] Use DRM scheduler for delayed GT TLB invalidations

[Matthew Brost]

On Mon, Aug 11, 2025 at 09:31:08AM +0100, Tvrtko Ursulin wrote:
> 
> On 24/07/2025 20:12, Matthew Brost wrote:
> > MIME-Version: 1.0
> > Content-Type: text/plain; charset=UTF-8
> > Content-Transfer-Encoding: 8bit
> > 
> >        Use the DRM scheduler for delayed GT TLB invalidations, which properly
> > fixes the issue raised in [1]. GT TLB fences have their own dma-fence
> > context, so even if the invalidations are ordered, the dma-resv/DRM
> > scheduler cannot squash the fences. This results in O(M*N*N) complexity
> > in the garbage collector, where M is the number of ranges in the garbage
> > collector and N is the number of pending GT TLB invalidations. After
> > this change, the resulting complexity in O(M*C) where C is the number of
> > TLB invalidation contexts (i.e., number of exec queue, GTs tuples) with
> > an invalidation in flight.
> 
> Does this series improve the performance of the TLB invalidations in
> general?
> 
> I am asking because we are currently investigating a problem with Google
> Chrome which apparently suffers a lot from latencies caused by TLB
> invalidations piling up after a lot of tab switching causes a lot of vm bind
> ioctl activity.

It doesn't improve performance of the actual invalidation rather the
time to create a VM unbind job as it avoids O(M*N*N) complexity
explosion. I originally saw this in a case of many SVM unbinds (i.e. a
free / unmap is called on a large piece of memory) but many VM unbinds
in short period of time could also cause this complexity explosion too.

By making all invalidations from the same (queue, GT) share a dma_fence
context, dma-resv and drm_sched can coalesce them into a single
dependenct per (queue, GT). That keeps the dependency count bounded and
eliminated the SVM panics I was seeing under heavy unbind stress.

> 
> Also, is it related to
> https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2162 ?
> 

Maybe. In this case, a flood of TLB invalidations overwhelmed the GuC,
leaving it no time to do anything else (e.g., schedule contexts). It’s a
multi-process test, so the driver-side complexity explosion likely
doesn’t apply as this explosion is seen within a single VM. What we need
is to hold TLB invalidations in the driver and coalesce them to issue
fewer operations. I have patches for this, but I haven’t cleaned them up
for posting yet—we’re in the middle of other TLB-invalidation refactors.

Matt

> Regards,
> 
> Tvrtko
> 
> > Admittedly, it's quite a lot of code, but the series includes extensive
> > kernel documentation and clear code comments. It introduces a generic
> > dependency scheduler that can be reused in the future and is logically
> > much cleaner than the previous open-coded solution for delaying GT TLB
> > invalidations until a bind job completes.
> > 
> > v2:
> >   - Various cleanup covered in detail in change logs
> >   - Use a per-GT ordered workqueue as DRM scheduler workqueue
> >   - Remove unused ftrace points
> > v3:
> >   - Address Stuarts feedback
> >   - Fix kernel doc, minor cleanups
> > v4:
> >   - Rebase
> > 
> > Matt
> > 
> > [1] https://patchwork.freedesktop.org/patch/658370/?series=150188&rev=1
> > 
> > Matthew Brost (7):
> >    drm/xe: Explicitly mark migration queues with flag
> >    drm/xe: Add generic dependecy jobs / scheduler
> >    drm/xe: Create ordered workqueue for GT TLB invalidation jobs
> >    drm/xe: Add dependency scheduler for GT TLB invalidations to bind
> >      queues
> >    drm/xe: Add GT TLB invalidation jobs
> >    drm/xe: Use GT TLB invalidation jobs in PT layer
> >    drm/xe: Remove unused GT TLB invalidation trace points
> > 
> >   drivers/gpu/drm/xe/Makefile                 |   2 +
> >   drivers/gpu/drm/xe/xe_dep_job_types.h       |  29 +++
> >   drivers/gpu/drm/xe/xe_dep_scheduler.c       | 143 ++++++++++
> >   drivers/gpu/drm/xe/xe_dep_scheduler.h       |  21 ++
> >   drivers/gpu/drm/xe/xe_exec_queue.c          |  48 ++++
> >   drivers/gpu/drm/xe/xe_exec_queue_types.h    |  15 ++
> >   drivers/gpu/drm/xe/xe_gt_tlb_inval_job.c    | 274 ++++++++++++++++++++
> >   drivers/gpu/drm/xe/xe_gt_tlb_inval_job.h    |  34 +++
> >   drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c |   8 +
> >   drivers/gpu/drm/xe/xe_gt_types.h            |   2 +
> >   drivers/gpu/drm/xe/xe_migrate.c             |  42 ++-
> >   drivers/gpu/drm/xe/xe_migrate.h             |  13 +
> >   drivers/gpu/drm/xe/xe_pt.c                  | 178 +++++--------
> >   drivers/gpu/drm/xe/xe_trace.h               |  16 --
> >   14 files changed, 700 insertions(+), 125 deletions(-)
> >   create mode 100644 drivers/gpu/drm/xe/xe_dep_job_types.h
> >   create mode 100644 drivers/gpu/drm/xe/xe_dep_scheduler.c
> >   create mode 100644 drivers/gpu/drm/xe/xe_dep_scheduler.h
> >   create mode 100644 drivers/gpu/drm/xe/xe_gt_tlb_inval_job.c
> >   create mode 100644 drivers/gpu/drm/xe/xe_gt_tlb_inval_job.h
> > 
>