1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "intel_context.h"
8 #include "intel_gpu_commands.h"
9 #include "intel_gt.h"
10 #include "intel_gtt.h"
11 #include "intel_migrate.h"
12 #include "intel_ring.h"
13 
14 struct insert_pte_data {
15 	u64 offset;
16 };
17 
18 #define CHUNK_SZ SZ_8M /* ~1ms at 8GiB/s preemption delay */
19 
20 #define GET_CCS_BYTES(i915, size)	(HAS_FLAT_CCS(i915) ? \
21 					 DIV_ROUND_UP(size, NUM_BYTES_PER_CCS_BYTE) : 0)
22 static bool engine_supports_migration(struct intel_engine_cs *engine)
23 {
24 	if (!engine)
25 		return false;
26 
27 	/*
28 	 * We need the ability to prevent aribtration (MI_ARB_ON_OFF),
29 	 * the ability to write PTE using inline data (MI_STORE_DATA)
30 	 * and of course the ability to do the block transfer (blits).
31 	 */
32 	GEM_BUG_ON(engine->class != COPY_ENGINE_CLASS);
33 
34 	return true;
35 }
36 
37 static void xehpsdv_toggle_pdes(struct i915_address_space *vm,
38 				struct i915_page_table *pt,
39 				void *data)
40 {
41 	struct insert_pte_data *d = data;
42 
43 	/*
44 	 * Insert a dummy PTE into every PT that will map to LMEM to ensure
45 	 * we have a correctly setup PDE structure for later use.
46 	 */
47 	vm->insert_page(vm, 0, d->offset, I915_CACHE_NONE, PTE_LM);
48 	GEM_BUG_ON(!pt->is_compact);
49 	d->offset += SZ_2M;
50 }
51 
52 static void xehpsdv_insert_pte(struct i915_address_space *vm,
53 			       struct i915_page_table *pt,
54 			       void *data)
55 {
56 	struct insert_pte_data *d = data;
57 
58 	/*
59 	 * We are playing tricks here, since the actual pt, from the hw
60 	 * pov, is only 256bytes with 32 entries, or 4096bytes with 512
61 	 * entries, but we are still guaranteed that the physical
62 	 * alignment is 64K underneath for the pt, and we are careful
63 	 * not to access the space in the void.
64 	 */
65 	vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE, PTE_LM);
66 	d->offset += SZ_64K;
67 }
68 
69 static void insert_pte(struct i915_address_space *vm,
70 		       struct i915_page_table *pt,
71 		       void *data)
72 {
73 	struct insert_pte_data *d = data;
74 
75 	vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE,
76 			i915_gem_object_is_lmem(pt->base) ? PTE_LM : 0);
77 	d->offset += PAGE_SIZE;
78 }
79 
80 static struct i915_address_space *migrate_vm(struct intel_gt *gt)
81 {
82 	struct i915_vm_pt_stash stash = {};
83 	struct i915_ppgtt *vm;
84 	int err;
85 	int i;
86 
87 	/*
88 	 * We construct a very special VM for use by all migration contexts,
89 	 * it is kept pinned so that it can be used at any time. As we need
90 	 * to pre-allocate the page directories for the migration VM, this
91 	 * limits us to only using a small number of prepared vma.
92 	 *
93 	 * To be able to pipeline and reschedule migration operations while
94 	 * avoiding unnecessary contention on the vm itself, the PTE updates
95 	 * are inline with the blits. All the blits use the same fixed
96 	 * addresses, with the backing store redirection being updated on the
97 	 * fly. Only 2 implicit vma are used for all migration operations.
98 	 *
99 	 * We lay the ppGTT out as:
100 	 *
101 	 *	[0, CHUNK_SZ) -> first object
102 	 *	[CHUNK_SZ, 2 * CHUNK_SZ) -> second object
103 	 *	[2 * CHUNK_SZ, 2 * CHUNK_SZ + 2 * CHUNK_SZ >> 9] -> PTE
104 	 *
105 	 * By exposing the dma addresses of the page directories themselves
106 	 * within the ppGTT, we are then able to rewrite the PTE prior to use.
107 	 * But the PTE update and subsequent migration operation must be atomic,
108 	 * i.e. within the same non-preemptible window so that we do not switch
109 	 * to another migration context that overwrites the PTE.
110 	 *
111 	 * This changes quite a bit on platforms with HAS_64K_PAGES support,
112 	 * where we instead have three windows, each CHUNK_SIZE in size. The
113 	 * first is reserved for mapping system-memory, and that just uses the
114 	 * 512 entry layout using 4K GTT pages. The other two windows just map
115 	 * lmem pages and must use the new compact 32 entry layout using 64K GTT
116 	 * pages, which ensures we can address any lmem object that the user
117 	 * throws at us. We then also use the xehpsdv_toggle_pdes as a way of
118 	 * just toggling the PDE bit(GEN12_PDE_64K) for us, to enable the
119 	 * compact layout for each of these page-tables, that fall within the
120 	 * [CHUNK_SIZE, 3 * CHUNK_SIZE) range.
121 	 *
122 	 * We lay the ppGTT out as:
123 	 *
124 	 * [0, CHUNK_SZ) -> first window/object, maps smem
125 	 * [CHUNK_SZ, 2 * CHUNK_SZ) -> second window/object, maps lmem src
126 	 * [2 * CHUNK_SZ, 3 * CHUNK_SZ) -> third window/object, maps lmem dst
127 	 *
128 	 * For the PTE window it's also quite different, since each PTE must
129 	 * point to some 64K page, one for each PT(since it's in lmem), and yet
130 	 * each is only <= 4096bytes, but since the unused space within that PTE
131 	 * range is never touched, this should be fine.
132 	 *
133 	 * So basically each PT now needs 64K of virtual memory, instead of 4K,
134 	 * which looks like:
135 	 *
136 	 * [3 * CHUNK_SZ, 3 * CHUNK_SZ + ((3 * CHUNK_SZ / SZ_2M) * SZ_64K)] -> PTE
137 	 */
138 
139 	vm = i915_ppgtt_create(gt, I915_BO_ALLOC_PM_EARLY);
140 	if (IS_ERR(vm))
141 		return ERR_CAST(vm);
142 
143 	if (!vm->vm.allocate_va_range || !vm->vm.foreach) {
144 		err = -ENODEV;
145 		goto err_vm;
146 	}
147 
148 	if (HAS_64K_PAGES(gt->i915))
149 		stash.pt_sz = I915_GTT_PAGE_SIZE_64K;
150 
151 	/*
152 	 * Each engine instance is assigned its own chunk in the VM, so
153 	 * that we can run multiple instances concurrently
154 	 */
155 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
156 		struct intel_engine_cs *engine;
157 		u64 base = (u64)i << 32;
158 		struct insert_pte_data d = {};
159 		struct i915_gem_ww_ctx ww;
160 		u64 sz;
161 
162 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
163 		if (!engine_supports_migration(engine))
164 			continue;
165 
166 		/*
167 		 * We copy in 8MiB chunks. Each PDE covers 2MiB, so we need
168 		 * 4x2 page directories for source/destination.
169 		 */
170 		if (HAS_64K_PAGES(gt->i915))
171 			sz = 3 * CHUNK_SZ;
172 		else
173 			sz = 2 * CHUNK_SZ;
174 		d.offset = base + sz;
175 
176 		/*
177 		 * We need another page directory setup so that we can write
178 		 * the 8x512 PTE in each chunk.
179 		 */
180 		if (HAS_64K_PAGES(gt->i915))
181 			sz += (sz / SZ_2M) * SZ_64K;
182 		else
183 			sz += (sz >> 12) * sizeof(u64);
184 
185 		err = i915_vm_alloc_pt_stash(&vm->vm, &stash, sz);
186 		if (err)
187 			goto err_vm;
188 
189 		for_i915_gem_ww(&ww, err, true) {
190 			err = i915_vm_lock_objects(&vm->vm, &ww);
191 			if (err)
192 				continue;
193 			err = i915_vm_map_pt_stash(&vm->vm, &stash);
194 			if (err)
195 				continue;
196 
197 			vm->vm.allocate_va_range(&vm->vm, &stash, base, sz);
198 		}
199 		i915_vm_free_pt_stash(&vm->vm, &stash);
200 		if (err)
201 			goto err_vm;
202 
203 		/* Now allow the GPU to rewrite the PTE via its own ppGTT */
204 		if (HAS_64K_PAGES(gt->i915)) {
205 			vm->vm.foreach(&vm->vm, base, d.offset - base,
206 				       xehpsdv_insert_pte, &d);
207 			d.offset = base + CHUNK_SZ;
208 			vm->vm.foreach(&vm->vm,
209 				       d.offset,
210 				       2 * CHUNK_SZ,
211 				       xehpsdv_toggle_pdes, &d);
212 		} else {
213 			vm->vm.foreach(&vm->vm, base, d.offset - base,
214 				       insert_pte, &d);
215 		}
216 	}
217 
218 	return &vm->vm;
219 
220 err_vm:
221 	i915_vm_put(&vm->vm);
222 	return ERR_PTR(err);
223 }
224 
225 static struct intel_engine_cs *first_copy_engine(struct intel_gt *gt)
226 {
227 	struct intel_engine_cs *engine;
228 	int i;
229 
230 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
231 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
232 		if (engine_supports_migration(engine))
233 			return engine;
234 	}
235 
236 	return NULL;
237 }
238 
239 static struct intel_context *pinned_context(struct intel_gt *gt)
240 {
241 	static struct lock_class_key key;
242 	struct intel_engine_cs *engine;
243 	struct i915_address_space *vm;
244 	struct intel_context *ce;
245 
246 	engine = first_copy_engine(gt);
247 	if (!engine)
248 		return ERR_PTR(-ENODEV);
249 
250 	vm = migrate_vm(gt);
251 	if (IS_ERR(vm))
252 		return ERR_CAST(vm);
253 
254 	ce = intel_engine_create_pinned_context(engine, vm, SZ_512K,
255 						I915_GEM_HWS_MIGRATE,
256 						&key, "migrate");
257 	i915_vm_put(vm);
258 	return ce;
259 }
260 
261 int intel_migrate_init(struct intel_migrate *m, struct intel_gt *gt)
262 {
263 	struct intel_context *ce;
264 
265 	memset(m, 0, sizeof(*m));
266 
267 	ce = pinned_context(gt);
268 	if (IS_ERR(ce))
269 		return PTR_ERR(ce);
270 
271 	m->context = ce;
272 	return 0;
273 }
274 
275 static int random_index(unsigned int max)
276 {
277 	return upper_32_bits(mul_u32_u32(get_random_u32(), max));
278 }
279 
280 static struct intel_context *__migrate_engines(struct intel_gt *gt)
281 {
282 	struct intel_engine_cs *engines[MAX_ENGINE_INSTANCE];
283 	struct intel_engine_cs *engine;
284 	unsigned int count, i;
285 
286 	count = 0;
287 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
288 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
289 		if (engine_supports_migration(engine))
290 			engines[count++] = engine;
291 	}
292 
293 	return intel_context_create(engines[random_index(count)]);
294 }
295 
296 struct intel_context *intel_migrate_create_context(struct intel_migrate *m)
297 {
298 	struct intel_context *ce;
299 
300 	/*
301 	 * We randomly distribute contexts across the engines upon constrction,
302 	 * as they all share the same pinned vm, and so in order to allow
303 	 * multiple blits to run in parallel, we must construct each blit
304 	 * to use a different range of the vm for its GTT. This has to be
305 	 * known at construction, so we can not use the late greedy load
306 	 * balancing of the virtual-engine.
307 	 */
308 	ce = __migrate_engines(m->context->engine->gt);
309 	if (IS_ERR(ce))
310 		return ce;
311 
312 	ce->ring = NULL;
313 	ce->ring_size = SZ_256K;
314 
315 	i915_vm_put(ce->vm);
316 	ce->vm = i915_vm_get(m->context->vm);
317 
318 	return ce;
319 }
320 
321 static inline struct sgt_dma sg_sgt(struct scatterlist *sg)
322 {
323 	dma_addr_t addr = sg_dma_address(sg);
324 
325 	return (struct sgt_dma){ sg, addr, addr + sg_dma_len(sg) };
326 }
327 
328 static int emit_no_arbitration(struct i915_request *rq)
329 {
330 	u32 *cs;
331 
332 	cs = intel_ring_begin(rq, 2);
333 	if (IS_ERR(cs))
334 		return PTR_ERR(cs);
335 
336 	/* Explicitly disable preemption for this request. */
337 	*cs++ = MI_ARB_ON_OFF;
338 	*cs++ = MI_NOOP;
339 	intel_ring_advance(rq, cs);
340 
341 	return 0;
342 }
343 
344 static int emit_pte(struct i915_request *rq,
345 		    struct sgt_dma *it,
346 		    enum i915_cache_level cache_level,
347 		    bool is_lmem,
348 		    u64 offset,
349 		    int length)
350 {
351 	bool has_64K_pages = HAS_64K_PAGES(rq->engine->i915);
352 	const u64 encode = rq->context->vm->pte_encode(0, cache_level,
353 						       is_lmem ? PTE_LM : 0);
354 	struct intel_ring *ring = rq->ring;
355 	int pkt, dword_length;
356 	u32 total = 0;
357 	u32 page_size;
358 	u32 *hdr, *cs;
359 
360 	GEM_BUG_ON(GRAPHICS_VER(rq->engine->i915) < 8);
361 
362 	page_size = I915_GTT_PAGE_SIZE;
363 	dword_length = 0x400;
364 
365 	/* Compute the page directory offset for the target address range */
366 	if (has_64K_pages) {
367 		GEM_BUG_ON(!IS_ALIGNED(offset, SZ_2M));
368 
369 		offset /= SZ_2M;
370 		offset *= SZ_64K;
371 		offset += 3 * CHUNK_SZ;
372 
373 		if (is_lmem) {
374 			page_size = I915_GTT_PAGE_SIZE_64K;
375 			dword_length = 0x40;
376 		}
377 	} else {
378 		offset >>= 12;
379 		offset *= sizeof(u64);
380 		offset += 2 * CHUNK_SZ;
381 	}
382 
383 	offset += (u64)rq->engine->instance << 32;
384 
385 	cs = intel_ring_begin(rq, 6);
386 	if (IS_ERR(cs))
387 		return PTR_ERR(cs);
388 
389 	/* Pack as many PTE updates as possible into a single MI command */
390 	pkt = min_t(int, dword_length, ring->space / sizeof(u32) + 5);
391 	pkt = min_t(int, pkt, (ring->size - ring->emit) / sizeof(u32) + 5);
392 
393 	hdr = cs;
394 	*cs++ = MI_STORE_DATA_IMM | REG_BIT(21); /* as qword elements */
395 	*cs++ = lower_32_bits(offset);
396 	*cs++ = upper_32_bits(offset);
397 
398 	do {
399 		if (cs - hdr >= pkt) {
400 			int dword_rem;
401 
402 			*hdr += cs - hdr - 2;
403 			*cs++ = MI_NOOP;
404 
405 			ring->emit = (void *)cs - ring->vaddr;
406 			intel_ring_advance(rq, cs);
407 			intel_ring_update_space(ring);
408 
409 			cs = intel_ring_begin(rq, 6);
410 			if (IS_ERR(cs))
411 				return PTR_ERR(cs);
412 
413 			dword_rem = dword_length;
414 			if (has_64K_pages) {
415 				if (IS_ALIGNED(total, SZ_2M)) {
416 					offset = round_up(offset, SZ_64K);
417 				} else {
418 					dword_rem = SZ_2M - (total & (SZ_2M - 1));
419 					dword_rem /= page_size;
420 					dword_rem *= 2;
421 				}
422 			}
423 
424 			pkt = min_t(int, dword_rem, ring->space / sizeof(u32) + 5);
425 			pkt = min_t(int, pkt, (ring->size - ring->emit) / sizeof(u32) + 5);
426 
427 			hdr = cs;
428 			*cs++ = MI_STORE_DATA_IMM | REG_BIT(21);
429 			*cs++ = lower_32_bits(offset);
430 			*cs++ = upper_32_bits(offset);
431 		}
432 
433 		GEM_BUG_ON(!IS_ALIGNED(it->dma, page_size));
434 
435 		*cs++ = lower_32_bits(encode | it->dma);
436 		*cs++ = upper_32_bits(encode | it->dma);
437 
438 		offset += 8;
439 		total += page_size;
440 
441 		it->dma += page_size;
442 		if (it->dma >= it->max) {
443 			it->sg = __sg_next(it->sg);
444 			if (!it->sg || sg_dma_len(it->sg) == 0)
445 				break;
446 
447 			it->dma = sg_dma_address(it->sg);
448 			it->max = it->dma + sg_dma_len(it->sg);
449 		}
450 	} while (total < length);
451 
452 	*hdr += cs - hdr - 2;
453 	*cs++ = MI_NOOP;
454 
455 	ring->emit = (void *)cs - ring->vaddr;
456 	intel_ring_advance(rq, cs);
457 	intel_ring_update_space(ring);
458 
459 	return total;
460 }
461 
462 static bool wa_1209644611_applies(int ver, u32 size)
463 {
464 	u32 height = size >> PAGE_SHIFT;
465 
466 	if (ver != 11)
467 		return false;
468 
469 	return height % 4 == 3 && height <= 8;
470 }
471 
472 /**
473  * DOC: Flat-CCS - Memory compression for Local memory
474  *
475  * On Xe-HP and later devices, we use dedicated compression control state (CCS)
476  * stored in local memory for each surface, to support the 3D and media
477  * compression formats.
478  *
479  * The memory required for the CCS of the entire local memory is 1/256 of the
480  * local memory size. So before the kernel boot, the required memory is reserved
481  * for the CCS data and a secure register will be programmed with the CCS base
482  * address.
483  *
484  * Flat CCS data needs to be cleared when a lmem object is allocated.
485  * And CCS data can be copied in and out of CCS region through
486  * XY_CTRL_SURF_COPY_BLT. CPU can't access the CCS data directly.
487  *
488  * I915 supports Flat-CCS on lmem only objects. When an objects has smem in
489  * its preference list, on memory pressure, i915 needs to migrate the lmem
490  * content into smem. If the lmem object is Flat-CCS compressed by userspace,
491  * then i915 needs to decompress it. But I915 lack the required information
492  * for such decompression. Hence I915 supports Flat-CCS only on lmem only objects.
493  *
494  * When we exhaust the lmem, Flat-CCS capable objects' lmem backing memory can
495  * be temporarily evicted to smem, along with the auxiliary CCS state, where
496  * it can be potentially swapped-out at a later point, if required.
497  * If userspace later touches the evicted pages, then we always move
498  * the backing memory back to lmem, which includes restoring the saved CCS state,
499  * and potentially performing any required swap-in.
500  *
501  * For the migration of the lmem objects with smem in placement list, such as
502  * {lmem, smem}, objects are treated as non Flat-CCS capable objects.
503  */
504 
505 static inline u32 *i915_flush_dw(u32 *cmd, u32 flags)
506 {
507 	*cmd++ = MI_FLUSH_DW | flags;
508 	*cmd++ = 0;
509 	*cmd++ = 0;
510 
511 	return cmd;
512 }
513 
514 static u32 calc_ctrl_surf_instr_size(struct drm_i915_private *i915, int size)
515 {
516 	u32 num_cmds, num_blks, total_size;
517 
518 	if (!GET_CCS_BYTES(i915, size))
519 		return 0;
520 
521 	/*
522 	 * XY_CTRL_SURF_COPY_BLT transfers CCS in 256 byte
523 	 * blocks. one XY_CTRL_SURF_COPY_BLT command can
524 	 * transfer upto 1024 blocks.
525 	 */
526 	num_blks = DIV_ROUND_UP(GET_CCS_BYTES(i915, size),
527 				NUM_CCS_BYTES_PER_BLOCK);
528 	num_cmds = DIV_ROUND_UP(num_blks, NUM_CCS_BLKS_PER_XFER);
529 	total_size = XY_CTRL_SURF_INSTR_SIZE * num_cmds;
530 
531 	/*
532 	 * Adding a flush before and after XY_CTRL_SURF_COPY_BLT
533 	 */
534 	total_size += 2 * MI_FLUSH_DW_SIZE;
535 
536 	return total_size;
537 }
538 
539 static int emit_copy_ccs(struct i915_request *rq,
540 			 u32 dst_offset, u8 dst_access,
541 			 u32 src_offset, u8 src_access, int size)
542 {
543 	struct drm_i915_private *i915 = rq->engine->i915;
544 	int mocs = rq->engine->gt->mocs.uc_index << 1;
545 	u32 num_ccs_blks, ccs_ring_size;
546 	u32 *cs;
547 
548 	ccs_ring_size = calc_ctrl_surf_instr_size(i915, size);
549 	WARN_ON(!ccs_ring_size);
550 
551 	cs = intel_ring_begin(rq, round_up(ccs_ring_size, 2));
552 	if (IS_ERR(cs))
553 		return PTR_ERR(cs);
554 
555 	num_ccs_blks = DIV_ROUND_UP(GET_CCS_BYTES(i915, size),
556 				    NUM_CCS_BYTES_PER_BLOCK);
557 	GEM_BUG_ON(num_ccs_blks > NUM_CCS_BLKS_PER_XFER);
558 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
559 
560 	/*
561 	 * The XY_CTRL_SURF_COPY_BLT instruction is used to copy the CCS
562 	 * data in and out of the CCS region.
563 	 *
564 	 * We can copy at most 1024 blocks of 256 bytes using one
565 	 * XY_CTRL_SURF_COPY_BLT instruction.
566 	 *
567 	 * In case we need to copy more than 1024 blocks, we need to add
568 	 * another instruction to the same batch buffer.
569 	 *
570 	 * 1024 blocks of 256 bytes of CCS represent a total 256KB of CCS.
571 	 *
572 	 * 256 KB of CCS represents 256 * 256 KB = 64 MB of LMEM.
573 	 */
574 	*cs++ = XY_CTRL_SURF_COPY_BLT |
575 		src_access << SRC_ACCESS_TYPE_SHIFT |
576 		dst_access << DST_ACCESS_TYPE_SHIFT |
577 		((num_ccs_blks - 1) & CCS_SIZE_MASK) << CCS_SIZE_SHIFT;
578 	*cs++ = src_offset;
579 	*cs++ = rq->engine->instance |
580 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
581 	*cs++ = dst_offset;
582 	*cs++ = rq->engine->instance |
583 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
584 
585 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
586 	if (ccs_ring_size & 1)
587 		*cs++ = MI_NOOP;
588 
589 	intel_ring_advance(rq, cs);
590 
591 	return 0;
592 }
593 
594 static int emit_copy(struct i915_request *rq,
595 		     u32 dst_offset, u32 src_offset, int size)
596 {
597 	const int ver = GRAPHICS_VER(rq->engine->i915);
598 	u32 instance = rq->engine->instance;
599 	u32 *cs;
600 
601 	cs = intel_ring_begin(rq, ver >= 8 ? 10 : 6);
602 	if (IS_ERR(cs))
603 		return PTR_ERR(cs);
604 
605 	if (ver >= 9 && !wa_1209644611_applies(ver, size)) {
606 		*cs++ = GEN9_XY_FAST_COPY_BLT_CMD | (10 - 2);
607 		*cs++ = BLT_DEPTH_32 | PAGE_SIZE;
608 		*cs++ = 0;
609 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
610 		*cs++ = dst_offset;
611 		*cs++ = instance;
612 		*cs++ = 0;
613 		*cs++ = PAGE_SIZE;
614 		*cs++ = src_offset;
615 		*cs++ = instance;
616 	} else if (ver >= 8) {
617 		*cs++ = XY_SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (10 - 2);
618 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
619 		*cs++ = 0;
620 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
621 		*cs++ = dst_offset;
622 		*cs++ = instance;
623 		*cs++ = 0;
624 		*cs++ = PAGE_SIZE;
625 		*cs++ = src_offset;
626 		*cs++ = instance;
627 	} else {
628 		GEM_BUG_ON(instance);
629 		*cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
630 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
631 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE;
632 		*cs++ = dst_offset;
633 		*cs++ = PAGE_SIZE;
634 		*cs++ = src_offset;
635 	}
636 
637 	intel_ring_advance(rq, cs);
638 	return 0;
639 }
640 
641 static int scatter_list_length(struct scatterlist *sg)
642 {
643 	int len = 0;
644 
645 	while (sg && sg_dma_len(sg)) {
646 		len += sg_dma_len(sg);
647 		sg = sg_next(sg);
648 	};
649 
650 	return len;
651 }
652 
653 static void
654 calculate_chunk_sz(struct drm_i915_private *i915, bool src_is_lmem,
655 		   int *src_sz, u32 bytes_to_cpy, u32 ccs_bytes_to_cpy)
656 {
657 	if (ccs_bytes_to_cpy) {
658 		if (!src_is_lmem)
659 			/*
660 			 * When CHUNK_SZ is passed all the pages upto CHUNK_SZ
661 			 * will be taken for the blt. in Flat-ccs supported
662 			 * platform Smem obj will have more pages than required
663 			 * for main meory hence limit it to the required size
664 			 * for main memory
665 			 */
666 			*src_sz = min_t(int, bytes_to_cpy, CHUNK_SZ);
667 	} else { /* ccs handling is not required */
668 		*src_sz = CHUNK_SZ;
669 	}
670 }
671 
672 static void get_ccs_sg_sgt(struct sgt_dma *it, u32 bytes_to_cpy)
673 {
674 	u32 len;
675 
676 	do {
677 		GEM_BUG_ON(!it->sg || !sg_dma_len(it->sg));
678 		len = it->max - it->dma;
679 		if (len > bytes_to_cpy) {
680 			it->dma += bytes_to_cpy;
681 			break;
682 		}
683 
684 		bytes_to_cpy -= len;
685 
686 		it->sg = __sg_next(it->sg);
687 		it->dma = sg_dma_address(it->sg);
688 		it->max = it->dma + sg_dma_len(it->sg);
689 	} while (bytes_to_cpy);
690 }
691 
692 int
693 intel_context_migrate_copy(struct intel_context *ce,
694 			   const struct i915_deps *deps,
695 			   struct scatterlist *src,
696 			   enum i915_cache_level src_cache_level,
697 			   bool src_is_lmem,
698 			   struct scatterlist *dst,
699 			   enum i915_cache_level dst_cache_level,
700 			   bool dst_is_lmem,
701 			   struct i915_request **out)
702 {
703 	struct sgt_dma it_src = sg_sgt(src), it_dst = sg_sgt(dst), it_ccs;
704 	struct drm_i915_private *i915 = ce->engine->i915;
705 	u32 ccs_bytes_to_cpy = 0, bytes_to_cpy;
706 	enum i915_cache_level ccs_cache_level;
707 	u32 src_offset, dst_offset;
708 	u8 src_access, dst_access;
709 	struct i915_request *rq;
710 	int src_sz, dst_sz;
711 	bool ccs_is_src, overwrite_ccs;
712 	int err;
713 
714 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
715 	GEM_BUG_ON(IS_DGFX(ce->engine->i915) && (!src_is_lmem && !dst_is_lmem));
716 	*out = NULL;
717 
718 	GEM_BUG_ON(ce->ring->size < SZ_64K);
719 
720 	src_sz = scatter_list_length(src);
721 	bytes_to_cpy = src_sz;
722 
723 	if (HAS_FLAT_CCS(i915) && src_is_lmem ^ dst_is_lmem) {
724 		src_access = !src_is_lmem && dst_is_lmem;
725 		dst_access = !src_access;
726 
727 		dst_sz = scatter_list_length(dst);
728 		if (src_is_lmem) {
729 			it_ccs = it_dst;
730 			ccs_cache_level = dst_cache_level;
731 			ccs_is_src = false;
732 		} else if (dst_is_lmem) {
733 			bytes_to_cpy = dst_sz;
734 			it_ccs = it_src;
735 			ccs_cache_level = src_cache_level;
736 			ccs_is_src = true;
737 		}
738 
739 		/*
740 		 * When there is a eviction of ccs needed smem will have the
741 		 * extra pages for the ccs data
742 		 *
743 		 * TO-DO: Want to move the size mismatch check to a WARN_ON,
744 		 * but still we have some requests of smem->lmem with same size.
745 		 * Need to fix it.
746 		 */
747 		ccs_bytes_to_cpy = src_sz != dst_sz ? GET_CCS_BYTES(i915, bytes_to_cpy) : 0;
748 		if (ccs_bytes_to_cpy)
749 			get_ccs_sg_sgt(&it_ccs, bytes_to_cpy);
750 	}
751 
752 	overwrite_ccs = HAS_FLAT_CCS(i915) && !ccs_bytes_to_cpy && dst_is_lmem;
753 
754 	src_offset = 0;
755 	dst_offset = CHUNK_SZ;
756 	if (HAS_64K_PAGES(ce->engine->i915)) {
757 		src_offset = 0;
758 		dst_offset = 0;
759 		if (src_is_lmem)
760 			src_offset = CHUNK_SZ;
761 		if (dst_is_lmem)
762 			dst_offset = 2 * CHUNK_SZ;
763 	}
764 
765 	do {
766 		int len;
767 
768 		rq = i915_request_create(ce);
769 		if (IS_ERR(rq)) {
770 			err = PTR_ERR(rq);
771 			goto out_ce;
772 		}
773 
774 		if (deps) {
775 			err = i915_request_await_deps(rq, deps);
776 			if (err)
777 				goto out_rq;
778 
779 			if (rq->engine->emit_init_breadcrumb) {
780 				err = rq->engine->emit_init_breadcrumb(rq);
781 				if (err)
782 					goto out_rq;
783 			}
784 
785 			deps = NULL;
786 		}
787 
788 		/* The PTE updates + copy must not be interrupted. */
789 		err = emit_no_arbitration(rq);
790 		if (err)
791 			goto out_rq;
792 
793 		calculate_chunk_sz(i915, src_is_lmem, &src_sz,
794 				   bytes_to_cpy, ccs_bytes_to_cpy);
795 
796 		len = emit_pte(rq, &it_src, src_cache_level, src_is_lmem,
797 			       src_offset, src_sz);
798 		if (!len) {
799 			err = -EINVAL;
800 			goto out_rq;
801 		}
802 		if (len < 0) {
803 			err = len;
804 			goto out_rq;
805 		}
806 
807 		err = emit_pte(rq, &it_dst, dst_cache_level, dst_is_lmem,
808 			       dst_offset, len);
809 		if (err < 0)
810 			goto out_rq;
811 		if (err < len) {
812 			err = -EINVAL;
813 			goto out_rq;
814 		}
815 
816 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
817 		if (err)
818 			goto out_rq;
819 
820 		err = emit_copy(rq, dst_offset,	src_offset, len);
821 		if (err)
822 			goto out_rq;
823 
824 		bytes_to_cpy -= len;
825 
826 		if (ccs_bytes_to_cpy) {
827 			int ccs_sz;
828 
829 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
830 			if (err)
831 				goto out_rq;
832 
833 			ccs_sz = GET_CCS_BYTES(i915, len);
834 			err = emit_pte(rq, &it_ccs, ccs_cache_level, false,
835 				       ccs_is_src ? src_offset : dst_offset,
836 				       ccs_sz);
837 			if (err < 0)
838 				goto out_rq;
839 			if (err < ccs_sz) {
840 				err = -EINVAL;
841 				goto out_rq;
842 			}
843 
844 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
845 			if (err)
846 				goto out_rq;
847 
848 			err = emit_copy_ccs(rq, dst_offset, dst_access,
849 					    src_offset, src_access, len);
850 			if (err)
851 				goto out_rq;
852 
853 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
854 			if (err)
855 				goto out_rq;
856 			ccs_bytes_to_cpy -= ccs_sz;
857 		} else if (overwrite_ccs) {
858 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
859 			if (err)
860 				goto out_rq;
861 
862 			/*
863 			 * While we can't always restore/manage the CCS state,
864 			 * we still need to ensure we don't leak the CCS state
865 			 * from the previous user, so make sure we overwrite it
866 			 * with something.
867 			 */
868 			err = emit_copy_ccs(rq, dst_offset, INDIRECT_ACCESS,
869 					    dst_offset, DIRECT_ACCESS, len);
870 			if (err)
871 				goto out_rq;
872 
873 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
874 			if (err)
875 				goto out_rq;
876 		}
877 
878 		/* Arbitration is re-enabled between requests. */
879 out_rq:
880 		if (*out)
881 			i915_request_put(*out);
882 		*out = i915_request_get(rq);
883 		i915_request_add(rq);
884 
885 		if (err)
886 			break;
887 
888 		if (!bytes_to_cpy && !ccs_bytes_to_cpy) {
889 			if (src_is_lmem)
890 				WARN_ON(it_src.sg && sg_dma_len(it_src.sg));
891 			else
892 				WARN_ON(it_dst.sg && sg_dma_len(it_dst.sg));
893 			break;
894 		}
895 
896 		if (WARN_ON(!it_src.sg || !sg_dma_len(it_src.sg) ||
897 			    !it_dst.sg || !sg_dma_len(it_dst.sg) ||
898 			    (ccs_bytes_to_cpy && (!it_ccs.sg ||
899 						  !sg_dma_len(it_ccs.sg))))) {
900 			err = -EINVAL;
901 			break;
902 		}
903 
904 		cond_resched();
905 	} while (1);
906 
907 out_ce:
908 	return err;
909 }
910 
911 static int emit_clear(struct i915_request *rq, u32 offset, int size,
912 		      u32 value, bool is_lmem)
913 {
914 	struct drm_i915_private *i915 = rq->engine->i915;
915 	int mocs = rq->engine->gt->mocs.uc_index << 1;
916 	const int ver = GRAPHICS_VER(i915);
917 	int ring_sz;
918 	u32 *cs;
919 
920 	GEM_BUG_ON(size >> PAGE_SHIFT > S16_MAX);
921 
922 	if (HAS_FLAT_CCS(i915) && ver >= 12)
923 		ring_sz = XY_FAST_COLOR_BLT_DW;
924 	else if (ver >= 8)
925 		ring_sz = 8;
926 	else
927 		ring_sz = 6;
928 
929 	cs = intel_ring_begin(rq, ring_sz);
930 	if (IS_ERR(cs))
931 		return PTR_ERR(cs);
932 
933 	if (HAS_FLAT_CCS(i915) && ver >= 12) {
934 		*cs++ = XY_FAST_COLOR_BLT_CMD | XY_FAST_COLOR_BLT_DEPTH_32 |
935 			(XY_FAST_COLOR_BLT_DW - 2);
936 		*cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, mocs) |
937 			(PAGE_SIZE - 1);
938 		*cs++ = 0;
939 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
940 		*cs++ = offset;
941 		*cs++ = rq->engine->instance;
942 		*cs++ = !is_lmem << XY_FAST_COLOR_BLT_MEM_TYPE_SHIFT;
943 		/* BG7 */
944 		*cs++ = value;
945 		*cs++ = 0;
946 		*cs++ = 0;
947 		*cs++ = 0;
948 		/* BG11 */
949 		*cs++ = 0;
950 		*cs++ = 0;
951 		/* BG13 */
952 		*cs++ = 0;
953 		*cs++ = 0;
954 		*cs++ = 0;
955 	} else if (ver >= 8) {
956 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (7 - 2);
957 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
958 		*cs++ = 0;
959 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
960 		*cs++ = offset;
961 		*cs++ = rq->engine->instance;
962 		*cs++ = value;
963 		*cs++ = MI_NOOP;
964 	} else {
965 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
966 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
967 		*cs++ = 0;
968 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
969 		*cs++ = offset;
970 		*cs++ = value;
971 	}
972 
973 	intel_ring_advance(rq, cs);
974 	return 0;
975 }
976 
977 int
978 intel_context_migrate_clear(struct intel_context *ce,
979 			    const struct i915_deps *deps,
980 			    struct scatterlist *sg,
981 			    enum i915_cache_level cache_level,
982 			    bool is_lmem,
983 			    u32 value,
984 			    struct i915_request **out)
985 {
986 	struct drm_i915_private *i915 = ce->engine->i915;
987 	struct sgt_dma it = sg_sgt(sg);
988 	struct i915_request *rq;
989 	u32 offset;
990 	int err;
991 
992 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
993 	*out = NULL;
994 
995 	GEM_BUG_ON(ce->ring->size < SZ_64K);
996 
997 	offset = 0;
998 	if (HAS_64K_PAGES(i915) && is_lmem)
999 		offset = CHUNK_SZ;
1000 
1001 	do {
1002 		int len;
1003 
1004 		rq = i915_request_create(ce);
1005 		if (IS_ERR(rq)) {
1006 			err = PTR_ERR(rq);
1007 			goto out_ce;
1008 		}
1009 
1010 		if (deps) {
1011 			err = i915_request_await_deps(rq, deps);
1012 			if (err)
1013 				goto out_rq;
1014 
1015 			if (rq->engine->emit_init_breadcrumb) {
1016 				err = rq->engine->emit_init_breadcrumb(rq);
1017 				if (err)
1018 					goto out_rq;
1019 			}
1020 
1021 			deps = NULL;
1022 		}
1023 
1024 		/* The PTE updates + clear must not be interrupted. */
1025 		err = emit_no_arbitration(rq);
1026 		if (err)
1027 			goto out_rq;
1028 
1029 		len = emit_pte(rq, &it, cache_level, is_lmem, offset, CHUNK_SZ);
1030 		if (len <= 0) {
1031 			err = len;
1032 			goto out_rq;
1033 		}
1034 
1035 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1036 		if (err)
1037 			goto out_rq;
1038 
1039 		err = emit_clear(rq, offset, len, value, is_lmem);
1040 		if (err)
1041 			goto out_rq;
1042 
1043 		if (HAS_FLAT_CCS(i915) && is_lmem && !value) {
1044 			/*
1045 			 * copy the content of memory into corresponding
1046 			 * ccs surface
1047 			 */
1048 			err = emit_copy_ccs(rq, offset, INDIRECT_ACCESS, offset,
1049 					    DIRECT_ACCESS, len);
1050 			if (err)
1051 				goto out_rq;
1052 		}
1053 
1054 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1055 
1056 		/* Arbitration is re-enabled between requests. */
1057 out_rq:
1058 		if (*out)
1059 			i915_request_put(*out);
1060 		*out = i915_request_get(rq);
1061 		i915_request_add(rq);
1062 		if (err || !it.sg || !sg_dma_len(it.sg))
1063 			break;
1064 
1065 		cond_resched();
1066 	} while (1);
1067 
1068 out_ce:
1069 	return err;
1070 }
1071 
1072 int intel_migrate_copy(struct intel_migrate *m,
1073 		       struct i915_gem_ww_ctx *ww,
1074 		       const struct i915_deps *deps,
1075 		       struct scatterlist *src,
1076 		       enum i915_cache_level src_cache_level,
1077 		       bool src_is_lmem,
1078 		       struct scatterlist *dst,
1079 		       enum i915_cache_level dst_cache_level,
1080 		       bool dst_is_lmem,
1081 		       struct i915_request **out)
1082 {
1083 	struct intel_context *ce;
1084 	int err;
1085 
1086 	*out = NULL;
1087 	if (!m->context)
1088 		return -ENODEV;
1089 
1090 	ce = intel_migrate_create_context(m);
1091 	if (IS_ERR(ce))
1092 		ce = intel_context_get(m->context);
1093 	GEM_BUG_ON(IS_ERR(ce));
1094 
1095 	err = intel_context_pin_ww(ce, ww);
1096 	if (err)
1097 		goto out;
1098 
1099 	err = intel_context_migrate_copy(ce, deps,
1100 					 src, src_cache_level, src_is_lmem,
1101 					 dst, dst_cache_level, dst_is_lmem,
1102 					 out);
1103 
1104 	intel_context_unpin(ce);
1105 out:
1106 	intel_context_put(ce);
1107 	return err;
1108 }
1109 
1110 int
1111 intel_migrate_clear(struct intel_migrate *m,
1112 		    struct i915_gem_ww_ctx *ww,
1113 		    const struct i915_deps *deps,
1114 		    struct scatterlist *sg,
1115 		    enum i915_cache_level cache_level,
1116 		    bool is_lmem,
1117 		    u32 value,
1118 		    struct i915_request **out)
1119 {
1120 	struct intel_context *ce;
1121 	int err;
1122 
1123 	*out = NULL;
1124 	if (!m->context)
1125 		return -ENODEV;
1126 
1127 	ce = intel_migrate_create_context(m);
1128 	if (IS_ERR(ce))
1129 		ce = intel_context_get(m->context);
1130 	GEM_BUG_ON(IS_ERR(ce));
1131 
1132 	err = intel_context_pin_ww(ce, ww);
1133 	if (err)
1134 		goto out;
1135 
1136 	err = intel_context_migrate_clear(ce, deps, sg, cache_level,
1137 					  is_lmem, value, out);
1138 
1139 	intel_context_unpin(ce);
1140 out:
1141 	intel_context_put(ce);
1142 	return err;
1143 }
1144 
1145 void intel_migrate_fini(struct intel_migrate *m)
1146 {
1147 	struct intel_context *ce;
1148 
1149 	ce = fetch_and_zero(&m->context);
1150 	if (!ce)
1151 		return;
1152 
1153 	intel_engine_destroy_pinned_context(ce);
1154 }
1155 
1156 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1157 #include "selftest_migrate.c"
1158 #endif
1159