[PATCH v4 0/3] DG2 accelerated migration/clearing support

[PATCH v4 0/3] DG2 accelerated migration/clearing support

[Matthew Auld]

Enable accelerated moves and clearing on DG2. On such HW we have minimum page
size restrictions when accessing LMEM from the GTT, where we now have to use 64K
GTT pages or larger. With the ppGTT the page-table also has a slightly different
layout from past generations when using the 64K GTT mode(which is still enabled
on via some PDE bit), where it is now compacted down to 32 qword entries. Note
that on discrete the paging structures must also be placed in LMEM, and we need
to able to modify them via the GTT itself(see patch 3), which is one of the
complications here.

v2: Add missing cover letter
v3:
  - Add some r-b tags
  - Drop the GTT_MAPPABLE approach. We can instead simply pass along the
    required size/alignment using alloc_pt().
v4:
  - Drop already merged patches, and add some r-b tags
  - Add some better docs to patch 3

Matthew Auld (3):
  drm/i915/gtt: allow overriding the pt alignment
  drm/i915/gtt: add xehpsdv_ppgtt_insert_entry
  drm/i915/migrate: add acceleration support for DG2

 drivers/gpu/drm/i915/gt/gen8_ppgtt.c    |  50 +++++-
 drivers/gpu/drm/i915/gt/intel_gtt.h     |  10 +-
 drivers/gpu/drm/i915/gt/intel_migrate.c | 209 +++++++++++++++++++-----
 drivers/gpu/drm/i915/gt/intel_ppgtt.c   |  16 +-
 4 files changed, 239 insertions(+), 46 deletions(-)

-- 
2.31.1

[PATCH v4 1/3] drm/i915/gtt: allow overriding the pt alignment

[Matthew Auld]

On some platforms we have alignment restrictions when accessing LMEM
from the GTT. In the next few patches we need to be able to modify the
page-tables directly via the GTT itself.

Suggested-by: Ramalingam C <ramalingam.c@intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Reviewed-by: Ramalingam C <ramalingam.c@intel.com>
---
 drivers/gpu/drm/i915/gt/intel_gtt.h   | 10 +++++++++-
 drivers/gpu/drm/i915/gt/intel_ppgtt.c | 16 ++++++++++++----
 2 files changed, 21 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h
index 01e9a98846fb..5ca5caa667b8 100644
--- a/drivers/gpu/drm/i915/gt/intel_gtt.h
+++ b/drivers/gpu/drm/i915/gt/intel_gtt.h
@@ -199,6 +199,14 @@ void *__px_vaddr(struct drm_i915_gem_object *p);
 struct i915_vm_pt_stash {
 	/* preallocated chains of page tables/directories */
 	struct i915_page_table *pt[2];
+	/*
+	 * Optionally override the alignment/size of the physical page that
+	 * contains each PT. If not set defaults back to the usual
+	 * I915_GTT_PAGE_SIZE_4K. This does not influence the other paging
+	 * structures. MUST be a power-of-two. ONLY applicable on discrete
+	 * platforms.
+	 */
+	int pt_sz;
 };
 
 struct i915_vma_ops {
@@ -586,7 +594,7 @@ void free_scratch(struct i915_address_space *vm);
 
 struct drm_i915_gem_object *alloc_pt_dma(struct i915_address_space *vm, int sz);
 struct drm_i915_gem_object *alloc_pt_lmem(struct i915_address_space *vm, int sz);
-struct i915_page_table *alloc_pt(struct i915_address_space *vm);
+struct i915_page_table *alloc_pt(struct i915_address_space *vm, int sz);
 struct i915_page_directory *alloc_pd(struct i915_address_space *vm);
 struct i915_page_directory *__alloc_pd(int npde);
 
diff --git a/drivers/gpu/drm/i915/gt/intel_ppgtt.c b/drivers/gpu/drm/i915/gt/intel_ppgtt.c
index b8238f5bc8b1..3c90aea25072 100644
--- a/drivers/gpu/drm/i915/gt/intel_ppgtt.c
+++ b/drivers/gpu/drm/i915/gt/intel_ppgtt.c
@@ -12,7 +12,7 @@
 #include "gen6_ppgtt.h"
 #include "gen8_ppgtt.h"
 
-struct i915_page_table *alloc_pt(struct i915_address_space *vm)
+struct i915_page_table *alloc_pt(struct i915_address_space *vm, int sz)
 {
 	struct i915_page_table *pt;
 
@@ -20,7 +20,7 @@ struct i915_page_table *alloc_pt(struct i915_address_space *vm)
 	if (unlikely(!pt))
 		return ERR_PTR(-ENOMEM);
 
-	pt->base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
+	pt->base = vm->alloc_pt_dma(vm, sz);
 	if (IS_ERR(pt->base)) {
 		kfree(pt);
 		return ERR_PTR(-ENOMEM);
@@ -219,17 +219,25 @@ int i915_vm_alloc_pt_stash(struct i915_address_space *vm,
 			   u64 size)
 {
 	unsigned long count;
-	int shift, n;
+	int shift, n, pt_sz;
 
 	shift = vm->pd_shift;
 	if (!shift)
 		return 0;
 
+	pt_sz = stash->pt_sz;
+	if (!pt_sz)
+		pt_sz = I915_GTT_PAGE_SIZE_4K;
+	else
+		GEM_BUG_ON(!IS_DGFX(vm->i915));
+
+	GEM_BUG_ON(!is_power_of_2(pt_sz));
+
 	count = pd_count(size, shift);
 	while (count--) {
 		struct i915_page_table *pt;
 
-		pt = alloc_pt(vm);
+		pt = alloc_pt(vm, pt_sz);
 		if (IS_ERR(pt)) {
 			i915_vm_free_pt_stash(vm, stash);
 			return PTR_ERR(pt);
-- 
2.31.1

[PATCH v4 2/3] drm/i915/gtt: add xehpsdv_ppgtt_insert_entry

[Matthew Auld]

If this is LMEM then we get a 32 entry PT, with each PTE pointing to
some 64K block of memory, otherwise it's just the usual 512 entry PT.
This very much assumes the caller knows what they are doing.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Reviewed-by: Ramalingam C <ramalingam.c@intel.com>
---
 drivers/gpu/drm/i915/gt/gen8_ppgtt.c | 50 ++++++++++++++++++++++++++--
 1 file changed, 48 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/gen8_ppgtt.c b/drivers/gpu/drm/i915/gt/gen8_ppgtt.c
index 5db11d8f7c7a..b6e9bfecb023 100644
--- a/drivers/gpu/drm/i915/gt/gen8_ppgtt.c
+++ b/drivers/gpu/drm/i915/gt/gen8_ppgtt.c
@@ -728,13 +728,56 @@ static void gen8_ppgtt_insert_entry(struct i915_address_space *vm,
 		gen8_pdp_for_page_index(vm, idx);
 	struct i915_page_directory *pd =
 		i915_pd_entry(pdp, gen8_pd_index(idx, 2));
+	struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1));
 	gen8_pte_t *vaddr;
 
-	vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
+	GEM_BUG_ON(pt->is_compact);
+
+	vaddr = px_vaddr(pt);
 	vaddr[gen8_pd_index(idx, 0)] = gen8_pte_encode(addr, level, flags);
 	clflush_cache_range(&vaddr[gen8_pd_index(idx, 0)], sizeof(*vaddr));
 }
 
+static void __xehpsdv_ppgtt_insert_entry_lm(struct i915_address_space *vm,
+					    dma_addr_t addr,
+					    u64 offset,
+					    enum i915_cache_level level,
+					    u32 flags)
+{
+	u64 idx = offset >> GEN8_PTE_SHIFT;
+	struct i915_page_directory * const pdp =
+		gen8_pdp_for_page_index(vm, idx);
+	struct i915_page_directory *pd =
+		i915_pd_entry(pdp, gen8_pd_index(idx, 2));
+	struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1));
+	gen8_pte_t *vaddr;
+
+	GEM_BUG_ON(!IS_ALIGNED(addr, SZ_64K));
+	GEM_BUG_ON(!IS_ALIGNED(offset, SZ_64K));
+
+	if (!pt->is_compact) {
+		vaddr = px_vaddr(pd);
+		vaddr[gen8_pd_index(idx, 1)] |= GEN12_PDE_64K;
+		pt->is_compact = true;
+	}
+
+	vaddr = px_vaddr(pt);
+	vaddr[gen8_pd_index(idx, 0) / 16] = gen8_pte_encode(addr, level, flags);
+}
+
+static void xehpsdv_ppgtt_insert_entry(struct i915_address_space *vm,
+				       dma_addr_t addr,
+				       u64 offset,
+				       enum i915_cache_level level,
+				       u32 flags)
+{
+	if (flags & PTE_LM)
+		return __xehpsdv_ppgtt_insert_entry_lm(vm, addr, offset,
+						       level, flags);
+
+	return gen8_ppgtt_insert_entry(vm, addr, offset, level, flags);
+}
+
 static int gen8_init_scratch(struct i915_address_space *vm)
 {
 	u32 pte_flags;
@@ -937,7 +980,10 @@ struct i915_ppgtt *gen8_ppgtt_create(struct intel_gt *gt,
 
 	ppgtt->vm.bind_async_flags = I915_VMA_LOCAL_BIND;
 	ppgtt->vm.insert_entries = gen8_ppgtt_insert;
-	ppgtt->vm.insert_page = gen8_ppgtt_insert_entry;
+	if (HAS_64K_PAGES(gt->i915))
+		ppgtt->vm.insert_page = xehpsdv_ppgtt_insert_entry;
+	else
+		ppgtt->vm.insert_page = gen8_ppgtt_insert_entry;
 	ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc;
 	ppgtt->vm.clear_range = gen8_ppgtt_clear;
 	ppgtt->vm.foreach = gen8_ppgtt_foreach;
-- 
2.31.1

[PATCH v4 3/3] drm/i915/migrate: add acceleration support for DG2

[Matthew Auld]

This is all kinds of awkward since we now have to contend with using 64K
GTT pages when mapping anything in LMEM(including the page-tables
themselves).

v2(Ram)
  - Document the ppGTT layout and add a better description for the
    different windows.

Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Reviewed-by: Ramalingam C <ramalingam.c@intel.com>
---
 drivers/gpu/drm/i915/gt/intel_migrate.c | 209 +++++++++++++++++++-----
 1 file changed, 170 insertions(+), 39 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_migrate.c b/drivers/gpu/drm/i915/gt/intel_migrate.c
index 64ffaacac1e0..c8f564b976e3 100644
--- a/drivers/gpu/drm/i915/gt/intel_migrate.c
+++ b/drivers/gpu/drm/i915/gt/intel_migrate.c
@@ -33,6 +33,38 @@ static bool engine_supports_migration(struct intel_engine_cs *engine)
 	return true;
 }
 
+static void xehpsdv_toggle_pdes(struct i915_address_space *vm,
+				struct i915_page_table *pt,
+				void *data)
+{
+	struct insert_pte_data *d = data;
+
+	/*
+	 * Insert a dummy PTE into every PT that will map to LMEM to ensure
+	 * we have a correctly setup PDE structure for later use.
+	 */
+	vm->insert_page(vm, 0, d->offset, I915_CACHE_NONE, PTE_LM);
+	GEM_BUG_ON(!pt->is_compact);
+	d->offset += SZ_2M;
+}
+
+static void xehpsdv_insert_pte(struct i915_address_space *vm,
+			       struct i915_page_table *pt,
+			       void *data)
+{
+	struct insert_pte_data *d = data;
+
+	/*
+	 * We are playing tricks here, since the actual pt, from the hw
+	 * pov, is only 256bytes with 32 entries, or 4096bytes with 512
+	 * entries, but we are still guaranteed that the physical
+	 * alignment is 64K underneath for the pt, and we are careful
+	 * not to access the space in the void.
+	 */
+	vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE, PTE_LM);
+	d->offset += SZ_64K;
+}
+
 static void insert_pte(struct i915_address_space *vm,
 		       struct i915_page_table *pt,
 		       void *data)
@@ -75,7 +107,32 @@ static struct i915_address_space *migrate_vm(struct intel_gt *gt)
 	 * i.e. within the same non-preemptible window so that we do not switch
 	 * to another migration context that overwrites the PTE.
 	 *
-	 * TODO: Add support for huge LMEM PTEs
+	 * This changes quite a bit on platforms with HAS_64K_PAGES support,
+	 * where we instead have three windows, each CHUNK_SIZE in size. The
+	 * first is reserved for mapping system-memory, and that just uses the
+	 * 512 entry layout using 4K GTT pages. The other two windows just map
+	 * lmem pages and must use the new compact 32 entry layout using 64K GTT
+	 * pages, which ensures we can address any lmem object that the user
+	 * throws at us. We then also use the xehpsdv_toggle_pdes as a way of
+	 * just toggling the PDE bit(GEN12_PDE_64K) for us, to enable the
+	 * compact layout for each of these page-tables, that fall within the
+	 * [CHUNK_SIZE, 3 * CHUNK_SIZE) range.
+	 *
+	 * We lay the ppGTT out as:
+	 *
+	 * [0, CHUNK_SZ) -> first window/object, maps smem
+	 * [CHUNK_SZ, 2 * CHUNK_SZ) -> second window/object, maps lmem src
+	 * [2 * CHUNK_SZ, 3 * CHUNK_SZ) -> third window/object, maps lmem dst
+	 *
+	 * For the PTE window it's also quite different, since each PTE must
+	 * point to some 64K page, one for each PT(since it's in lmem), and yet
+	 * each is only <= 4096bytes, but since the unused space within that PTE
+	 * range is never touched, this should be fine.
+	 *
+	 * So basically each PT now needs 64K of virtual memory, instead of 4K,
+	 * which looks like:
+	 *
+	 * [3 * CHUNK_SZ, 3 * CHUNK_SZ + ((3 * CHUNK_SZ / SZ_2M) * SZ_64K)] -> PTE
 	 */
 
 	vm = i915_ppgtt_create(gt, I915_BO_ALLOC_PM_EARLY);
@@ -87,6 +144,9 @@ static struct i915_address_space *migrate_vm(struct intel_gt *gt)
 		goto err_vm;
 	}
 
+	if (HAS_64K_PAGES(gt->i915))
+		stash.pt_sz = I915_GTT_PAGE_SIZE_64K;
+
 	/*
 	 * Each engine instance is assigned its own chunk in the VM, so
 	 * that we can run multiple instances concurrently
@@ -106,14 +166,20 @@ static struct i915_address_space *migrate_vm(struct intel_gt *gt)
 		 * We copy in 8MiB chunks. Each PDE covers 2MiB, so we need
 		 * 4x2 page directories for source/destination.
 		 */
-		sz = 2 * CHUNK_SZ;
+		if (HAS_64K_PAGES(gt->i915))
+			sz = 3 * CHUNK_SZ;
+		else
+			sz = 2 * CHUNK_SZ;
 		d.offset = base + sz;
 
 		/*
 		 * We need another page directory setup so that we can write
 		 * the 8x512 PTE in each chunk.
 		 */
-		sz += (sz >> 12) * sizeof(u64);
+		if (HAS_64K_PAGES(gt->i915))
+			sz += (sz / SZ_2M) * SZ_64K;
+		else
+			sz += (sz >> 12) * sizeof(u64);
 
 		err = i915_vm_alloc_pt_stash(&vm->vm, &stash, sz);
 		if (err)
@@ -134,7 +200,18 @@ static struct i915_address_space *migrate_vm(struct intel_gt *gt)
 			goto err_vm;
 
 		/* Now allow the GPU to rewrite the PTE via its own ppGTT */
-		vm->vm.foreach(&vm->vm, base, d.offset - base, insert_pte, &d);
+		if (HAS_64K_PAGES(gt->i915)) {
+			vm->vm.foreach(&vm->vm, base, d.offset - base,
+				       xehpsdv_insert_pte, &d);
+			d.offset = base + CHUNK_SZ;
+			vm->vm.foreach(&vm->vm,
+				       d.offset,
+				       2 * CHUNK_SZ,
+				       xehpsdv_toggle_pdes, &d);
+		} else {
+			vm->vm.foreach(&vm->vm, base, d.offset - base,
+				       insert_pte, &d);
+		}
 	}
 
 	return &vm->vm;
@@ -270,19 +347,38 @@ static int emit_pte(struct i915_request *rq,
 		    u64 offset,
 		    int length)
 {
+	bool has_64K_pages = HAS_64K_PAGES(rq->engine->i915);
 	const u64 encode = rq->context->vm->pte_encode(0, cache_level,
 						       is_lmem ? PTE_LM : 0);
 	struct intel_ring *ring = rq->ring;
-	int total = 0;
+	int pkt, dword_length;
+	u32 total = 0;
+	u32 page_size;
 	u32 *hdr, *cs;
-	int pkt;
 
 	GEM_BUG_ON(GRAPHICS_VER(rq->engine->i915) < 8);
 
+	page_size = I915_GTT_PAGE_SIZE;
+	dword_length = 0x400;
+
 	/* Compute the page directory offset for the target address range */
-	offset >>= 12;
-	offset *= sizeof(u64);
-	offset += 2 * CHUNK_SZ;
+	if (has_64K_pages) {
+		GEM_BUG_ON(!IS_ALIGNED(offset, SZ_2M));
+
+		offset /= SZ_2M;
+		offset *= SZ_64K;
+		offset += 3 * CHUNK_SZ;
+
+		if (is_lmem) {
+			page_size = I915_GTT_PAGE_SIZE_64K;
+			dword_length = 0x40;
+		}
+	} else {
+		offset >>= 12;
+		offset *= sizeof(u64);
+		offset += 2 * CHUNK_SZ;
+	}
+
 	offset += (u64)rq->engine->instance << 32;
 
 	cs = intel_ring_begin(rq, 6);
@@ -290,7 +386,7 @@ static int emit_pte(struct i915_request *rq,
 		return PTR_ERR(cs);
 
 	/* Pack as many PTE updates as possible into a single MI command */
-	pkt = min_t(int, 0x400, ring->space / sizeof(u32) + 5);
+	pkt = min_t(int, dword_length, ring->space / sizeof(u32) + 5);
 	pkt = min_t(int, pkt, (ring->size - ring->emit) / sizeof(u32) + 5);
 
 	hdr = cs;
@@ -300,6 +396,8 @@ static int emit_pte(struct i915_request *rq,
 
 	do {
 		if (cs - hdr >= pkt) {
+			int dword_rem;
+
 			*hdr += cs - hdr - 2;
 			*cs++ = MI_NOOP;
 
@@ -311,7 +409,18 @@ static int emit_pte(struct i915_request *rq,
 			if (IS_ERR(cs))
 				return PTR_ERR(cs);
 
-			pkt = min_t(int, 0x400, ring->space / sizeof(u32) + 5);
+			dword_rem = dword_length;
+			if (has_64K_pages) {
+				if (IS_ALIGNED(total, SZ_2M)) {
+					offset = round_up(offset, SZ_64K);
+				} else {
+					dword_rem = SZ_2M - (total & (SZ_2M - 1));
+					dword_rem /= page_size;
+					dword_rem *= 2;
+				}
+			}
+
+			pkt = min_t(int, dword_rem, ring->space / sizeof(u32) + 5);
 			pkt = min_t(int, pkt, (ring->size - ring->emit) / sizeof(u32) + 5);
 
 			hdr = cs;
@@ -320,13 +429,15 @@ static int emit_pte(struct i915_request *rq,
 			*cs++ = upper_32_bits(offset);
 		}
 
+		GEM_BUG_ON(!IS_ALIGNED(it->dma, page_size));
+
 		*cs++ = lower_32_bits(encode | it->dma);
 		*cs++ = upper_32_bits(encode | it->dma);
 
 		offset += 8;
-		total += I915_GTT_PAGE_SIZE;
+		total += page_size;
 
-		it->dma += I915_GTT_PAGE_SIZE;
+		it->dma += page_size;
 		if (it->dma >= it->max) {
 			it->sg = __sg_next(it->sg);
 			if (!it->sg || sg_dma_len(it->sg) == 0)
@@ -357,7 +468,8 @@ static bool wa_1209644611_applies(int ver, u32 size)
 	return height % 4 == 3 && height <= 8;
 }
 
-static int emit_copy(struct i915_request *rq, int size)
+static int emit_copy(struct i915_request *rq,
+		     u32 dst_offset, u32 src_offset, int size)
 {
 	const int ver = GRAPHICS_VER(rq->engine->i915);
 	u32 instance = rq->engine->instance;
@@ -372,31 +484,31 @@ static int emit_copy(struct i915_request *rq, int size)
 		*cs++ = BLT_DEPTH_32 | PAGE_SIZE;
 		*cs++ = 0;
 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
-		*cs++ = CHUNK_SZ; /* dst offset */
+		*cs++ = dst_offset;
 		*cs++ = instance;
 		*cs++ = 0;
 		*cs++ = PAGE_SIZE;
-		*cs++ = 0; /* src offset */
+		*cs++ = src_offset;
 		*cs++ = instance;
 	} else if (ver >= 8) {
 		*cs++ = XY_SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (10 - 2);
 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
 		*cs++ = 0;
 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
-		*cs++ = CHUNK_SZ; /* dst offset */
+		*cs++ = dst_offset;
 		*cs++ = instance;
 		*cs++ = 0;
 		*cs++ = PAGE_SIZE;
-		*cs++ = 0; /* src offset */
+		*cs++ = src_offset;
 		*cs++ = instance;
 	} else {
 		GEM_BUG_ON(instance);
 		*cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE;
-		*cs++ = CHUNK_SZ; /* dst offset */
+		*cs++ = dst_offset;
 		*cs++ = PAGE_SIZE;
-		*cs++ = 0; /* src offset */
+		*cs++ = src_offset;
 	}
 
 	intel_ring_advance(rq, cs);
@@ -424,6 +536,7 @@ intel_context_migrate_copy(struct intel_context *ce,
 	GEM_BUG_ON(ce->ring->size < SZ_64K);
 
 	do {
+		u32 src_offset, dst_offset;
 		int len;
 
 		rq = i915_request_create(ce);
@@ -451,15 +564,28 @@ intel_context_migrate_copy(struct intel_context *ce,
 		if (err)
 			goto out_rq;
 
-		len = emit_pte(rq, &it_src, src_cache_level, src_is_lmem, 0,
-			       CHUNK_SZ);
+		src_offset = 0;
+		dst_offset = CHUNK_SZ;
+		if (HAS_64K_PAGES(ce->engine->i915)) {
+			GEM_BUG_ON(!src_is_lmem && !dst_is_lmem);
+
+			src_offset = 0;
+			dst_offset = 0;
+			if (src_is_lmem)
+				src_offset = CHUNK_SZ;
+			if (dst_is_lmem)
+				dst_offset = 2 * CHUNK_SZ;
+		}
+
+		len = emit_pte(rq, &it_src, src_cache_level, src_is_lmem,
+			       src_offset, CHUNK_SZ);
 		if (len <= 0) {
 			err = len;
 			goto out_rq;
 		}
 
 		err = emit_pte(rq, &it_dst, dst_cache_level, dst_is_lmem,
-			       CHUNK_SZ, len);
+			       dst_offset, len);
 		if (err < 0)
 			goto out_rq;
 		if (err < len) {
@@ -471,7 +597,7 @@ intel_context_migrate_copy(struct intel_context *ce,
 		if (err)
 			goto out_rq;
 
-		err = emit_copy(rq, len);
+		err = emit_copy(rq, dst_offset, src_offset, len);
 
 		/* Arbitration is re-enabled between requests. */
 out_rq:
@@ -569,18 +695,20 @@ static u32 *_i915_ctrl_surf_copy_blt(u32 *cmd, u64 src_addr, u64 dst_addr,
 }
 
 static int emit_clear(struct i915_request *rq,
+		      u64 offset,
 		      int size,
 		      u32 value,
 		      bool is_lmem)
 {
-	const int ver = GRAPHICS_VER(rq->engine->i915);
-	u32 instance = rq->engine->instance;
-	u32 *cs;
 	struct drm_i915_private *i915 = rq->engine->i915;
+	const int ver = GRAPHICS_VER(rq->engine->i915);
 	u32 num_ccs_blks, ccs_ring_size;
+	u32 *cs;
 
 	GEM_BUG_ON(size >> PAGE_SHIFT > S16_MAX);
 
+	offset += (u64)rq->engine->instance << 32;
+
 	/* Clear flat css only when value is 0 */
 	ccs_ring_size = (is_lmem && !value) ?
 			 calc_ctrl_surf_instr_size(i915, size)
@@ -595,17 +723,17 @@ static int emit_clear(struct i915_request *rq,
 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
 		*cs++ = 0;
 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
-		*cs++ = 0; /* offset */
-		*cs++ = instance;
+		*cs++ = lower_32_bits(offset);
+		*cs++ = upper_32_bits(offset);
 		*cs++ = value;
 		*cs++ = MI_NOOP;
 	} else {
-		GEM_BUG_ON(instance);
+		GEM_BUG_ON(upper_32_bits(offset));
 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
 		*cs++ = 0;
 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
-		*cs++ = 0;
+		*cs++ = lower_32_bits(offset);
 		*cs++ = value;
 	}
 
@@ -621,17 +749,15 @@ static int emit_clear(struct i915_request *rq,
 		 * and use it as a source.
 		 */
 
-		cs = i915_flush_dw(cs, (u64)instance << 32,
-				   MI_FLUSH_LLC | MI_FLUSH_CCS);
+		cs = i915_flush_dw(cs, offset, MI_FLUSH_LLC | MI_FLUSH_CCS);
 		cs = _i915_ctrl_surf_copy_blt(cs,
-					      (u64)instance << 32,
-					      (u64)instance << 32,
+					      offset,
+					      offset,
 					      DIRECT_ACCESS,
 					      INDIRECT_ACCESS,
 					      1, 1,
 					      num_ccs_blks);
-		cs = i915_flush_dw(cs, (u64)instance << 32,
-				   MI_FLUSH_LLC | MI_FLUSH_CCS);
+		cs = i915_flush_dw(cs, offset, MI_FLUSH_LLC | MI_FLUSH_CCS);
 	}
 	intel_ring_advance(rq, cs);
 	return 0;
@@ -656,6 +782,7 @@ intel_context_migrate_clear(struct intel_context *ce,
 	GEM_BUG_ON(ce->ring->size < SZ_64K);
 
 	do {
+		u32 offset;
 		int len;
 
 		rq = i915_request_create(ce);
@@ -683,7 +810,11 @@ intel_context_migrate_clear(struct intel_context *ce,
 		if (err)
 			goto out_rq;
 
-		len = emit_pte(rq, &it, cache_level, is_lmem, 0, CHUNK_SZ);
+		offset = 0;
+		if (HAS_64K_PAGES(ce->engine->i915) && is_lmem)
+			offset = CHUNK_SZ;
+
+		len = emit_pte(rq, &it, cache_level, is_lmem, offset, CHUNK_SZ);
 		if (len <= 0) {
 			err = len;
 			goto out_rq;
@@ -693,7 +824,7 @@ intel_context_migrate_clear(struct intel_context *ce,
 		if (err)
 			goto out_rq;
 
-		err = emit_clear(rq, len, value, is_lmem);
+		err = emit_clear(rq, offset, len, value, is_lmem);
 
 		/* Arbitration is re-enabled between requests. */
 out_rq:
-- 
2.31.1