Re: [PATCH] drm/xe: Add min and max context TLB invalidation sizes

["Summers, Stuart" <stuart.summers@intel.com>]

On Mon, 2026-03-23 at 10:23 -0700, Matthew Brost wrote:
> On Fri, Mar 20, 2026 at 08:46:30PM +0000, Stuart Summers wrote:
> > Allow platform-defined TLB invalidation min and max lengths.
> > 
> > This gives finer granular control to which invalidations we
> > decide to send to GuC. The min size is essentially a round
> > up. The max allows us to switch to a full invalidation.
> > 
> > The expectation here is that GuC will translate the full
> > invalidation in this instance into a series of per context
> > invalidaitons. These are then issued with no H2G or G2H
> > messages and therefore should be quicker than splitting
> > the invalidations from the KMD in max size chunks and sending
> > separately.
> > 
> > v2: Add proper defaults for min/max if not set in the device
> >     structures
> > v3: Add coverage for pow-of-2 out of bounds cases
> > 
> > Signed-off-by: Stuart Summers <stuart.summers@intel.com>
> > Reviewed-by: Jonathan Cavitt <jonathan.cavitt@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_device_types.h  |  4 +++
> >  drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 39 +++++++++++++++++------
> > ----
> >  drivers/gpu/drm/xe/xe_pci.c           |  3 +++
> >  drivers/gpu/drm/xe/xe_pci_types.h     |  2 ++
> >  4 files changed, 34 insertions(+), 14 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_device_types.h
> > b/drivers/gpu/drm/xe/xe_device_types.h
> > index 615218d775b1..0c4168fe2ffb 100644
> > --- a/drivers/gpu/drm/xe/xe_device_types.h
> > +++ b/drivers/gpu/drm/xe/xe_device_types.h
> > @@ -137,6 +137,10 @@ struct xe_device {
> >                 u8 vm_max_level;
> >                 /** @info.va_bits: Maximum bits of a virtual
> > address */
> >                 u8 va_bits;
> > +               /** @info.min_tlb_inval_size: Minimum size of
> > context based TLB invalidations */
> > +               u64 min_tlb_inval_size;
> > +               /** @info.max_tlb_inval_size: Maximum size of
> > context based TLB invalidations */
> > +               u64 max_tlb_inval_size;
> >  
> >                 /*
> >                  * Keep all flags below alphabetically sorted
> > diff --git a/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
> > b/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
> > index ced58f46f846..e9e0be94ceef 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
> > @@ -115,14 +115,23 @@ static int send_page_reclaim(struct xe_guc
> > *guc, u32 seqno,
> >                               G2H_LEN_DW_PAGE_RECLAMATION, 1);
> >  }
> >  
> > +/*
> > + * Ensure that roundup_pow_of_two(length) doesn't overflow.
> > + * Note that roundup_pow_of_two() operates on unsigned long,
> > + * not on u64.
> > + */
> > +#define MAX_RANGE_TLB_INVALIDATION_LENGTH
> > (rounddown_pow_of_two(ULONG_MAX))
> > +
> >  static u64 normalize_invalidation_range(struct xe_gt *gt, u64
> > *start, u64 *end)
> >  {
> > +       struct xe_device *xe = gt_to_xe(gt);
> >         u64 orig_start = *start;
> >         u64 length = *end - *start;
> >         u64 align;
> >  
> > -       if (length < SZ_4K)
> > -               length = SZ_4K;
> > +       xe_gt_assert(gt, length <=
> > MAX_RANGE_TLB_INVALIDATION_LENGTH);
> > +
> > +       length = max_t(u64, xe->info.min_tlb_inval_size, length);
> >  
> >         align = roundup_pow_of_two(length);
> >         *start = ALIGN_DOWN(*start, align);
> > @@ -147,13 +156,6 @@ static u64 normalize_invalidation_range(struct
> > xe_gt *gt, u64 *start, u64 *end)
> >         return length;
> >  }
> >  
> > -/*
> > - * Ensure that roundup_pow_of_two(length) doesn't overflow.
> > - * Note that roundup_pow_of_two() operates on unsigned long,
> > - * not on u64.
> > - */
> > -#define MAX_RANGE_TLB_INVALIDATION_LENGTH
> > (rounddown_pow_of_two(ULONG_MAX))
> > -
> >  static int send_tlb_inval_ppgtt(struct xe_guc *guc, u32 seqno, u64
> > start,
> >                                 u64 end, u32 id, u32 type,
> >                                 struct drm_suballoc *prl_sa)
> > @@ -162,8 +164,20 @@ static int send_tlb_inval_ppgtt(struct xe_guc
> > *guc, u32 seqno, u64 start,
> >         struct xe_gt *gt = guc_to_gt(guc);
> >         struct xe_device *xe = guc_to_xe(guc);
> >         u32 action[MAX_TLB_INVALIDATION_LEN];
> > -       u64 length = end - start;
> > +       u64 normalize_len, length = end - start;
> >         int len = 0, err;
> > +       bool do_full_inval = false;
> > +
> > +       if (!xe->info.has_range_tlb_inval ||
> > +           length > MAX_RANGE_TLB_INVALIDATION_LENGTH) {
> > +               do_full_inval = true;
> > +       } else {
> > +               normalize_len = normalize_invalidation_range(gt,
> > &start,
> > +                                                            &end);
> > +
> > +               if (normalize_len > xe->info.max_tlb_inval_size)
> > +                       do_full_inval = true;
> > +       }
> 
> I suggested this is the last rev, can this logic be moved to
> send_tlb_inval_asid_ppgtt / send_tlb_inval_ctx_ppgtt?
> 
> For send_tlb_inval_asid_ppgtt it doesn't really matter as
> send_tlb_inval_ppgtt is called once.
> 
> But consider send_tlb_inval_ctx_ppgtt where send_tlb_inval_ppgtt is
> called multiple times and each call fails the
> normalize_invalidation_range step (i.e., we set do_full_inval). We
> only
> need to issue one full invalidation, not multiple.

Yeah it's a good suggestion. I'll split this out in the next rev.

And I'll repond to those other comments in the earlier rev.

Thanks,
Stuart

> 
> So likely want to hook in early in existing if statement in
> send_tlb_inval_ctx_ppgtt.
> 
> 244 #define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8
> 245         if (vm->exec_queues.count[id] >=
> EXEC_QUEUE_COUNT_FULL_THRESHOLD) {
> 246                 u32 action[] = {
> 247                         XE_GUC_ACTION_TLB_INVALIDATION,
> 248                         seqno,
> 249                         MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL),
> 250                 };
> 251
> 252                 err = send_tlb_inval(guc, action,
> ARRAY_SIZE(action));
> 253                 goto err_unlock;
> 254         }
> 255 #undef EXEC_QUEUE_COUNT_FULL_THRESHOLD
> 
> Matt
> 
> >  
> >         xe_gt_assert(gt, (type == XE_GUC_TLB_INVAL_PAGE_SELECTIVE
> > &&
> >                           !xe->info.has_ctx_tlb_inval) ||
> > @@ -172,12 +186,9 @@ static int send_tlb_inval_ppgtt(struct xe_guc
> > *guc, u32 seqno, u64 start,
> >  
> >         action[len++] = XE_GUC_ACTION_TLB_INVALIDATION;
> >         action[len++] = !prl_sa ? seqno :
> > TLB_INVALIDATION_SEQNO_INVALID;
> > -       if (!gt_to_xe(gt)->info.has_range_tlb_inval ||
> > -           length > MAX_RANGE_TLB_INVALIDATION_LENGTH) {
> > +       if (do_full_inval) {
> >                 action[len++] =
> > MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL);
> >         } else {
> > -               u64 normalize_len =
> > normalize_invalidation_range(gt, &start,
> > -                                                               
> > &end);
> >                 bool need_flush = !prl_sa &&
> >                         seqno != TLB_INVALIDATION_SEQNO_INVALID;
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_pci.c
> > b/drivers/gpu/drm/xe/xe_pci.c
> > index 189e2a1c29f9..5e02f9ab625b 100644
> > --- a/drivers/gpu/drm/xe/xe_pci.c
> > +++ b/drivers/gpu/drm/xe/xe_pci.c
> > @@ -743,6 +743,9 @@ static int xe_info_init_early(struct xe_device
> > *xe,
> >         xe->info.vm_max_level = desc->vm_max_level;
> >         xe->info.vram_flags = desc->vram_flags;
> >  
> > +       xe->info.min_tlb_inval_size = desc->min_tlb_inval_size ?:
> > SZ_4K;
> > +       xe->info.max_tlb_inval_size = desc->max_tlb_inval_size ?:
> > SZ_1G;
> > +
> >         xe->info.is_dgfx = desc->is_dgfx;
> >         xe->info.has_cached_pt = desc->has_cached_pt;
> >         xe->info.has_fan_control = desc->has_fan_control;
> > diff --git a/drivers/gpu/drm/xe/xe_pci_types.h
> > b/drivers/gpu/drm/xe/xe_pci_types.h
> > index 8eee4fb1c57c..cd9d3ad96fe0 100644
> > --- a/drivers/gpu/drm/xe/xe_pci_types.h
> > +++ b/drivers/gpu/drm/xe/xe_pci_types.h
> > @@ -34,6 +34,8 @@ struct xe_device_desc {
> >         u8 va_bits;
> >         u8 vm_max_level;
> >         u8 vram_flags;
> > +       u64 min_tlb_inval_size;
> > +       u64 max_tlb_inval_size;
> >  
> >         u8 require_force_probe:1;
> >         u8 is_dgfx:1;
> > -- 
> > 2.43.0
> >