Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
drm/xe: Use proper vram offset
In xe_migrate functions, use proper vram io offset of the
tiles while calculating addresses.
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
authored by
Niranjana Vishwanathapura
and committed by
Rodrigo Vivi
c33a7219
370997d1
+15
-11
+6
-6
drivers/gpu/drm/xe/xe_bo.c
+6
-6
drivers/gpu/drm/xe/xe_bo.c
···
1206
1206
* XXX: This is in the VM bind data path, likely should calculate this once and
1207
1207
* store, with a recalculation if the BO is moved.
1208
1208
*/
1209
-
static uint64_t vram_region_io_offset(struct xe_bo *bo)
1209
+
uint64_t vram_region_io_offset(struct ttm_resource *res)
1210
1210
{
1211
-
struct xe_device *xe = xe_bo_device(bo);
1212
-
struct xe_gt *gt = mem_type_to_gt(xe, bo->ttm.resource->mem_type);
1211
+
struct xe_device *xe = ttm_to_xe_device(res->bo->bdev);
1212
+
struct xe_gt *gt = mem_type_to_gt(xe, res->mem_type);
1213
1213
1214
-
if (bo->ttm.resource->mem_type == XE_PL_STOLEN)
1214
+
if (res->mem_type == XE_PL_STOLEN)
1215
1215
return xe_ttm_stolen_gpu_offset(xe);
1216
1216
1217
1217
return gt->mem.vram.io_start - xe->mem.vram.io_start;
···
1298
1298
XE_BUG_ON(!(place->flags & TTM_PL_FLAG_CONTIGUOUS));
1299
1299
1300
1300
place->fpfn = (xe_bo_addr(bo, 0, PAGE_SIZE, &vram) -
1301
-
vram_region_io_offset(bo)) >> PAGE_SHIFT;
1301
+
vram_region_io_offset(bo->ttm.resource)) >> PAGE_SHIFT;
1302
1302
place->lpfn = place->fpfn + (bo->size >> PAGE_SHIFT);
1303
1303
1304
1304
spin_lock(&xe->pinned.lock);
···
1442
1442
1443
1443
xe_res_first(bo->ttm.resource, page << PAGE_SHIFT,
1444
1444
page_size, &cur);
1445
-
return cur.start + offset + vram_region_io_offset(bo);
1445
+
return cur.start + offset + vram_region_io_offset(bo->ttm.resource);
1446
1446
}
1447
1447
}
1448
1448
+1
drivers/gpu/drm/xe/xe_bo.h
+1
drivers/gpu/drm/xe/xe_bo.h
+8
-5
drivers/gpu/drm/xe/xe_migrate.c
+8
-5
drivers/gpu/drm/xe/xe_migrate.c
···
392
392
393
393
static u32 pte_update_size(struct xe_migrate *m,
394
394
bool is_vram,
395
+
struct ttm_resource *res,
395
396
struct xe_res_cursor *cur,
396
397
u64 *L0, u64 *L0_ofs, u32 *L0_pt,
397
398
u32 cmd_size, u32 pt_ofs, u32 avail_pts)
···
418
417
cmds += cmd_size;
419
418
} else {
420
419
/* Offset into identity map. */
421
-
*L0_ofs = xe_migrate_vram_ofs(cur->start);
420
+
*L0_ofs = xe_migrate_vram_ofs(cur->start +
421
+
vram_region_io_offset(res));
422
422
cmds += cmd_size;
423
423
}
424
424
···
469
467
addr |= GEN12_PTE_PS64;
470
468
}
471
469
470
+
addr += vram_region_io_offset(bo->ttm.resource);
472
471
addr |= GEN12_PPGTT_PTE_LM;
473
472
}
474
473
addr |= PPAT_CACHED | GEN8_PAGE_PRESENT | GEN8_PAGE_RW;
···
649
646
650
647
src_L0 = min(src_L0, dst_L0);
651
648
652
-
batch_size += pte_update_size(m, src_is_vram, &src_it, &src_L0,
649
+
batch_size += pte_update_size(m, src_is_vram, src, &src_it, &src_L0,
653
650
&src_L0_ofs, &src_L0_pt, 0, 0,
654
651
NUM_PT_PER_BLIT);
655
652
656
-
batch_size += pte_update_size(m, dst_is_vram, &dst_it, &src_L0,
653
+
batch_size += pte_update_size(m, dst_is_vram, dst, &dst_it, &src_L0,
657
654
&dst_L0_ofs, &dst_L0_pt, 0,
658
655
NUM_PT_PER_BLIT, NUM_PT_PER_BLIT);
659
656
660
657
if (copy_system_ccs) {
661
658
ccs_size = xe_device_ccs_bytes(xe, src_L0);
662
-
batch_size += pte_update_size(m, false, &ccs_it, &ccs_size,
659
+
batch_size += pte_update_size(m, false, NULL, &ccs_it, &ccs_size,
663
660
&ccs_ofs, &ccs_pt, 0,
664
661
2 * NUM_PT_PER_BLIT,
665
662
NUM_PT_PER_BLIT);
···
882
879
883
880
/* Calculate final sizes and batch size.. */
884
881
batch_size = 2 +
885
-
pte_update_size(m, clear_vram, &src_it,
882
+
pte_update_size(m, clear_vram, src, &src_it,
886
883
&clear_L0, &clear_L0_ofs, &clear_L0_pt,
887
884
emit_clear_cmd_len(xe), 0,
888
885
NUM_PT_PER_BLIT);