Re: [PATCH v4 00/12] Fine grained fault locking, threaded prefetch, storm cache

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

On Thu, Feb 26, 2026 at 02:43:29PM +0100, Thomas Hellström wrote:
> Hi, Matt.
> 
> On Wed, 2026-02-25 at 20:28 -0800, Matthew Brost wrote:
> > Fine-grained fault locking provides immediate benefits: it allows
> > page
> > faults from the same VM to be processed in parallel (unless they
> > target
> > the same range) and enables a sane multi-threaded prefetch
> > implementation. UMD prefetch benchmarks see 10% to 50% improvement in
> > prefetch performance on BMG depending on PCIe bus speed.
> > 
> > Once parallel fault processing is available, the pagefault queue can
> > be
> > unified into a single queue with multiple workers pulling faults to
> > process. A single queue then allows a sensible pagefault cache to be
> > implemented, so that multiple faults targeting the same region can be
> > batched together and acknowledged in, ideally, a single pass. This
> > saves
> > CPU cycles during pagefault handling and improves overall throughput
> > of
> > the fault handler.
> > 
> > Significant improvements in UMD pagefault benchmarks can be seen when
> > utilizing this caching.
> > 
> > v3:
> >  - Fix kunit build (CI)
> > v4:
> >  - Actually fix kunit build (CI)
> > 
> > Matt
> > 
> > Matthew Brost (12):
> >   drm/xe: Fine grained page fault locking
> >   drm/xe: Allow prefetch-only VM bind IOCTLs to use VM read lock
> >   drm/xe: Thread prefetch of SVM ranges
> >   drm/xe: Use a single page-fault queue with multiple workers
> >   drm/xe: Add num_pf_work modparam
> >   drm/xe: Engine class and instance into a u8
> >   drm/xe: Track pagefault worker runtime
> >   drm/xe: Chain page faults via queue-resident cache to avoid fault
> >     storms
> >   drm/xe: Add pagefault chaining stats
> >   drm/xe: Add debugfs pagefault_info
> >   drm/xe: batch CT pagefault acks with periodic flush
> >   drm/xe: Track parallel page fault activity in GT stats
> > 
> >  drivers/gpu/drm/drm_gpusvm.c            |   2 +-
> >  drivers/gpu/drm/xe/xe_debugfs.c         |  11 +
> >  drivers/gpu/drm/xe/xe_defaults.h        |   1 +
> >  drivers/gpu/drm/xe/xe_device.c          |  17 +-
> >  drivers/gpu/drm/xe/xe_device_types.h    |  17 +-
> >  drivers/gpu/drm/xe/xe_gt_stats.c        |   7 +
> >  drivers/gpu/drm/xe/xe_gt_stats_types.h  |   7 +
> >  drivers/gpu/drm/xe/xe_guc_ct.c          |  94 +++-
> >  drivers/gpu/drm/xe/xe_guc_ct.h          |  35 +-
> >  drivers/gpu/drm/xe/xe_guc_pagefault.c   |  35 +-
> >  drivers/gpu/drm/xe/xe_guc_types.h       |   6 +
> >  drivers/gpu/drm/xe/xe_module.c          |   4 +
> >  drivers/gpu/drm/xe/xe_module.h          |   1 +
> >  drivers/gpu/drm/xe/xe_pagefault.c       | 675 ++++++++++++++++++++--
> > --
> >  drivers/gpu/drm/xe/xe_pagefault.h       |  74 +++
> >  drivers/gpu/drm/xe/xe_pagefault_types.h | 109 +++-
> >  drivers/gpu/drm/xe/xe_svm.c             | 129 +++--
> >  drivers/gpu/drm/xe/xe_svm.h             |  59 ++-
> >  drivers/gpu/drm/xe/xe_userptr.c         |  20 +-
> >  drivers/gpu/drm/xe/xe_vm.c              | 215 ++++++--
> >  drivers/gpu/drm/xe/xe_vm_types.h        |  37 +-
> >  21 files changed, 1309 insertions(+), 246 deletions(-)
> 
> Before I get to reviewing this, some suggestions from Claude:
> 
>   Confirmed regressions (3 commits with issues):
> 
>   c664c1b91090 — Fine grained page fault locking
> 
>    - Reference leak in vm_bind_ioctl_ops_create (xe_vm.c). 
>   xe_svm_range_find_or_insert() was changed to take a reference, but
> two paths
>    don't put it: (1) when xe_svm_range_validate() returns true → goto 
>   check_next_range, and (2) when xa_alloc() fails → goto
> unwind_prefetch_ops. 
>   The validate path is on every prefetch of an already-populated range,
> so 
>   refcounts grow unbounded.
>

Indeed. Will fix.

>   80012f80c75f — Chain page faults
> 
>    - Commit message typos only: "samr ASID" → "same ASID", "IRQ pathd"
> → "IRQ 
>   paths". No code issues.
>

Yes.

>   569104fb76ed — batch CT pagefault acks with periodic flush
> 
>    - Off-by-one in flush period: guc_ack_fault_begin initialises 

This is pretty good, caught this one myself after posting.

I'm convinced everyone should use Claude as a spot check before posting.

>   pagefault_ack_counter to PERIOD - 2 = 14, but the comment says first
> flush 
>   should be at ack #2. With counter=14 the first flush fires at ack #3 
>   (counter hits 16, 16&15==0). Fix: = XE_GUC_PAGEFAULT_FLUSH_PERIOD -
> 1.
>    - Commit message typo: "Assistent-by" → "Assisted-by".

Yes.

Matt

> 
> /Thomas
> 
> 
>