[Intel-xe] [PATCH 03/26] drm/xe: Add backpointer from gt to tile
Matt Roper
matthew.d.roper at intel.com
Fri May 12 17:00:26 UTC 2023
On Fri, May 12, 2023 at 09:31:27AM -0700, Matt Atwood wrote:
> On Fri, May 12, 2023 at 09:20:02AM -0700, Matt Roper wrote:
> > On Thu, May 11, 2023 at 05:07:46PM -0700, Lucas De Marchi wrote:
> > > On Wed, May 10, 2023 at 08:46:59PM -0700, Matt Roper wrote:
> > > > Rather than a backpointer to the xe_device, a GT should have a
> > > > backpointer to its tile (which can then be used to lookup the device if
> > > > necessary).
> > > >
> > > > The gt_to_xe() helper macro (which moves from xe_gt.h to xe_gt_types.h)
> > > > can and should still be used to jump directly from an xe_gt to
> > > > xe_device.
> > > >
> > > > Signed-off-by: Matt Roper <matthew.d.roper at intel.com>
> > > > ---
> > > > drivers/gpu/drm/xe/xe_bb.c | 2 +-
> > > > drivers/gpu/drm/xe/xe_gt.h | 5 -----
> > > > drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c | 4 ++--
> > > > drivers/gpu/drm/xe/xe_gt_types.h | 14 ++++++++++++--
> > > > drivers/gpu/drm/xe/xe_mocs.c | 14 +++++++-------
> > > > drivers/gpu/drm/xe/xe_pci.c | 11 +++++++----
> > > > drivers/gpu/drm/xe/xe_pt.c | 2 +-
> > > > 7 files changed, 30 insertions(+), 22 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/xe/xe_bb.c b/drivers/gpu/drm/xe/xe_bb.c
> > > > index 3deb2d55f421..bf7c94b769d7 100644
> > > > --- a/drivers/gpu/drm/xe/xe_bb.c
> > > > +++ b/drivers/gpu/drm/xe/xe_bb.c
> > > > @@ -16,7 +16,7 @@
> > > >
> > > > static int bb_prefetch(struct xe_gt *gt)
> > > > {
> > > > - struct xe_device *xe = gt->xe;
> > > > + struct xe_device *xe = gt_to_xe(gt);
> > > >
> > > > if (GRAPHICS_VERx100(xe) >= 1250 && !xe_gt_is_media_type(gt))
> > > > /*
> > > > diff --git a/drivers/gpu/drm/xe/xe_gt.h b/drivers/gpu/drm/xe/xe_gt.h
> > > > index 086369f7ee6d..f4e98f499b36 100644
> > > > --- a/drivers/gpu/drm/xe/xe_gt.h
> > > > +++ b/drivers/gpu/drm/xe/xe_gt.h
> > > > @@ -49,11 +49,6 @@ static inline bool xe_gt_is_media_type(struct xe_gt *gt)
> > > > return gt->info.type == XE_GT_TYPE_MEDIA;
> > > > }
> > > >
> > > > -#define gt_to_xe(gt__) \
> > > > - _Generic(gt__, \
> > > > - const struct xe_gt *: (const struct xe_device *)((gt__)->xe), \
> > > > - struct xe_gt *: (gt__)->xe)
> > > > -
> > > > static inline bool xe_gt_is_usm_hwe(struct xe_gt *gt, struct xe_hw_engine *hwe)
> > > > {
> > > > struct xe_device *xe = gt_to_xe(gt);
> > > > diff --git a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > > > index c815a42e2cdb..c9e8825c02aa 100644
> > > > --- a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > > > +++ b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
> > > > @@ -322,8 +322,8 @@ int xe_guc_tlb_invalidation_done_handler(struct xe_guc *guc, u32 *msg, u32 len)
> > > > TLB_INVALIDATION_SEQNO_MAX;
> > > > if (!expected_seqno)
> > > > expected_seqno = 1;
> > > > - if (drm_WARN_ON(>->xe->drm, expected_seqno != msg[0])) {
> > > > - drm_err(>->xe->drm, "TLB expected_seqno(%d) != msg(%u)\n",
> > > > + if (drm_WARN_ON(>_to_xe(gt)->drm, expected_seqno != msg[0])) {
> > > > + drm_err(>_to_xe(gt)->drm, "TLB expected_seqno(%d) != msg(%u)\n",
> > > > expected_seqno, msg[0]);
> > > > }
> > > >
> > > > diff --git a/drivers/gpu/drm/xe/xe_gt_types.h b/drivers/gpu/drm/xe/xe_gt_types.h
> > > > index e0ed4508269b..c4376d50786b 100644
> > > > --- a/drivers/gpu/drm/xe/xe_gt_types.h
> > > > +++ b/drivers/gpu/drm/xe/xe_gt_types.h
> > > > @@ -76,6 +76,16 @@ enum xe_steering_type {
> > > > NUM_STEERING_TYPES
> > > > };
> > > >
> > > > +#define gt_to_tile(gt__) \
> > > > + _Generic(gt__, \
> > > > + const struct xe_gt *: (const struct xe_tile *)((gt__)->tile), \
> > > > + struct xe_gt *: (gt__)->tile)
> > > > +
> > > > +#define gt_to_xe(gt__) \
> > > > + _Generic(gt__, \
> > > > + const struct xe_gt *: (const struct xe_device *)(gt_to_tile(gt__)->xe), \
> > > > + struct xe_gt *: gt_to_tile(gt__)->xe)
> > > > +
> > > > /**
> > > > * struct xe_gt - A "Graphics Technology" unit of the GPU
> > > > *
> > > > @@ -90,8 +100,8 @@ enum xe_steering_type {
> > > > * within a tile.
> > > > */
> > > > struct xe_gt {
> > > > - /** @xe: backpointer to XE device */
> > > > - struct xe_device *xe;
> > > > + /** @tile: Backpointer to GT's tile */
> > > > + struct xe_tile *tile;
> > > >
> > > > /** @info: GT info */
> > > > struct {
> > > > diff --git a/drivers/gpu/drm/xe/xe_mocs.c b/drivers/gpu/drm/xe/xe_mocs.c
> > > > index 817afd301d52..d57fbf16a3ef 100644
> > > > --- a/drivers/gpu/drm/xe/xe_mocs.c
> > > > +++ b/drivers/gpu/drm/xe/xe_mocs.c
> > > > @@ -471,7 +471,7 @@ static void __init_mocs_table(struct xe_gt *gt,
> > > > unsigned int i;
> > > > u32 mocs;
> > > >
> > > > - mocs_dbg(>->xe->drm, "entries:%d\n", info->n_entries);
> > > > + mocs_dbg(>_to_xe(gt)->drm, "entries:%d\n", info->n_entries);
> > > > drm_WARN_ONCE(&xe->drm, !info->unused_entries_index,
> > > > "Unused entries index should have been defined\n");
> > > > for (i = 0;
> > > > @@ -479,7 +479,7 @@ static void __init_mocs_table(struct xe_gt *gt,
> > > > i++) {
> > > > struct xe_reg reg = XE_REG(addr + i * 4);
> > > >
> > > > - mocs_dbg(>->xe->drm, "%d 0x%x 0x%x\n", i, reg.addr, mocs);
> > > > + mocs_dbg(>_to_xe(gt)->drm, "%d 0x%x 0x%x\n", i, reg.addr, mocs);
> > > > xe_mmio_write32(gt, reg, mocs);
> > > > }
> > > > }
> > > > @@ -508,13 +508,13 @@ static void init_l3cc_table(struct xe_gt *gt,
> > > > unsigned int i;
> > > > u32 l3cc;
> > > >
> > > > - mocs_dbg(>->xe->drm, "entries:%d\n", info->n_entries);
> > > > + mocs_dbg(>_to_xe(gt)->drm, "entries:%d\n", info->n_entries);
> > > > for (i = 0;
> > > > i < (info->n_entries + 1) / 2 ?
> > > > (l3cc = l3cc_combine(get_entry_l3cc(info, 2 * i),
> > > > get_entry_l3cc(info, 2 * i + 1))), 1 : 0;
> > > > i++) {
> > > > - mocs_dbg(>->xe->drm, "%d 0x%x 0x%x\n", i, LNCFCMOCS(i).addr,
> > > > + mocs_dbg(>_to_xe(gt)->drm, "%d 0x%x 0x%x\n", i, LNCFCMOCS(i).addr,
> > > > l3cc);
> > > > xe_mmio_write32(gt, LNCFCMOCS(i), l3cc);
> > > > }
> > > > @@ -524,7 +524,7 @@ void xe_mocs_init_early(struct xe_gt *gt)
> > > > {
> > > > struct xe_mocs_info table;
> > > >
> > > > - get_mocs_settings(gt->xe, &table);
> > > > + get_mocs_settings(gt_to_xe(gt), &table);
> > > > gt->mocs.uc_index = table.uc_index;
> > > > gt->mocs.wb_index = table.wb_index;
> > > > }
> > > > @@ -537,8 +537,8 @@ void xe_mocs_init(struct xe_gt *gt)
> > > > /*
> > > > * LLC and eDRAM control values are not applicable to dgfx
> > > > */
> > > > - flags = get_mocs_settings(gt->xe, &table);
> > > > - mocs_dbg(>->xe->drm, "flag:0x%x\n", flags);
> > > > + flags = get_mocs_settings(gt_to_xe(gt), &table);
> > > > + mocs_dbg(>_to_xe(gt)->drm, "flag:0x%x\n", flags);
> > > >
> > > > if (flags & HAS_GLOBAL_MOCS)
> > > > __init_mocs_table(gt, &table, GLOBAL_MOCS(0).addr);
> > > > diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c
> > > > index e79b16d8bf7f..87c328106aca 100644
> > > > --- a/drivers/gpu/drm/xe/xe_pci.c
> > > > +++ b/drivers/gpu/drm/xe/xe_pci.c
> > > > @@ -471,6 +471,7 @@ static int xe_info_init(struct xe_device *xe,
> > > > {
> > > > const struct xe_graphics_desc *graphics_desc = NULL;
> > > > const struct xe_media_desc *media_desc = NULL;
> > > > + struct xe_tile *tile;
> > > > struct xe_gt *gt;
> > > > u8 id;
> > > >
> > > > @@ -525,13 +526,15 @@ static int xe_info_init(struct xe_device *xe,
> > > > xe->info.step = xe_step_get(xe);
> > > >
> > > > for (id = 0; id < xe->info.tile_count; ++id) {
> > > > - xe->tiles[id].xe = xe;
> > > > - xe->tiles[id].id = id;
> > > > + tile = &xe->tiles[id];
> > > > + tile->xe = xe;
> > > > + tile->id = id;
> > > >
> > > > - gt = &xe->tiles[id].primary_gt;
> > > > + gt = &tile->primary_gt;
> > >
> > > these seem to have been squash in the wrong patch, changing the style
> > > from using xe->tiles[i].* to declaring a local var. Maybe squash it in
> > > the previous one to avoid the back and forth here?
> > >
> > > otherwise,
> > >
> > >
> > > Reviewed-by: Lucas De Marchi <lucas.demarchi at intel.com>
> > >
> > > > gt->info.id = id;
> > > > - gt->xe = xe;
> > > > + gt->tile = tile;
> > > >
> > > > + gt->info.id = id;
> > >
> > > the gt than has the id of the tile? Because right now there can only be
> > > primary_gt? Looks odd, but it's a consequence of removing the only
> > > platform with 1 tile and multiple gts.
> >
> > Yeah, deciding how we want to number GTs is a bit of an open question.
> > Right now at the end of this series, the numbering is
> >
> > * PVC: 0 = root tile primary, 1 = remote tile primary
> > * MTL: 0 = root tile primary, 1 = root tile media
> > * everything else: only has GT0
> >
> > Some day we may have a platform with multiple tiles and separate
> > graphics/media GTs within each tile. In that case would all graphics
> > GTs get even numbers and all media GTs get odd? Or would all primary
> > GTs get gt_id = tile_id and all media GTs get gt_id = tile_id +
> > tile_count?
> >
> > The bigger open question is whether uapi should still expose GTs (like
> > it does today) or whether it should be converted to just expose tiles
> > and keep the graphics vs media separation an internal implementation
> > detail. If uapi changes to only expose tiles, then we probably don't
> > even need IDs for the GTs...
> I recall doing the uapi changes for topology that UMD seemed really
> bound to the idea of interacting with "GTs". Likely only exposing tiles
> would cause a need for alot of changes in implementation and methodolgy in the UMD.
> MattA
On i915, neither tiles nor GTs are exposed to userspace. Exposing GTs
as a first-class concepts is a large design change on the Xe driver.
I think there's even been talk about splitting this out at the /dev/dri
level, with separate render nodes being exposed per tile. So there may
be even larger design changes coming before we settle on exactly how Xe
should work.
Matt
> >
> >
> > Matt
> >
> > >
> > > Lucas De Marchi
> > >
> > > > if (id == 0) {
> > > > gt->info.type = XE_GT_TYPE_MAIN;
> > > > gt->info.vram_id = id;
> > > > diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
> > > > index f15282996c3b..61126cefe0b5 100644
> > > > --- a/drivers/gpu/drm/xe/xe_pt.c
> > > > +++ b/drivers/gpu/drm/xe/xe_pt.c
> > > > @@ -695,7 +695,7 @@ xe_pt_stage_bind_entry(struct xe_ptw *parent, pgoff_t offset,
> > > > * TODO: Suballocate the pt bo to avoid wasting a lot of
> > > > * memory.
> > > > */
> > > > - if (GRAPHICS_VERx100(xe_walk->gt->xe) >= 1250 && level == 1 &&
> > > > + if (GRAPHICS_VERx100(gt_to_xe(xe_walk->gt)) >= 1250 && level == 1 &&
> > > > covers && xe_pt_scan_64K(addr, next, xe_walk)) {
> > > > walk->shifts = xe_compact_pt_shifts;
> > > > flags |= XE_PDE_64K;
> > > > --
> > > > 2.40.0
> > > >
> >
> > --
> > Matt Roper
> > Graphics Software Engineer
> > Linux GPU Platform Enablement
> > Intel Corporation
--
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
More information about the Intel-xe
mailing list