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 #include "gem/i915_gem_lmem.h"
14 
15 struct insert_pte_data {
16 	u64 offset;
17 };
18 
19 #define CHUNK_SZ SZ_8M /* ~1ms at 8GiB/s preemption delay */
20 
21 #define GET_CCS_BYTES(i915, size)	(HAS_FLAT_CCS(i915) ? \
22 					 DIV_ROUND_UP(size, NUM_BYTES_PER_CCS_BYTE) : 0)
23 static bool engine_supports_migration(struct intel_engine_cs *engine)
24 {
25 	if (!engine)
26 		return false;
27 
28 	/*
29 	 * We need the ability to prevent aribtration (MI_ARB_ON_OFF),
30 	 * the ability to write PTE using inline data (MI_STORE_DATA)
31 	 * and of course the ability to do the block transfer (blits).
32 	 */
33 	GEM_BUG_ON(engine->class != COPY_ENGINE_CLASS);
34 
35 	return true;
36 }
37 
38 static void xehpsdv_toggle_pdes(struct i915_address_space *vm,
39 				struct i915_page_table *pt,
40 				void *data)
41 {
42 	struct insert_pte_data *d = data;
43 
44 	/*
45 	 * Insert a dummy PTE into every PT that will map to LMEM to ensure
46 	 * we have a correctly setup PDE structure for later use.
47 	 */
48 	vm->insert_page(vm, 0, d->offset, I915_CACHE_NONE, PTE_LM);
49 	GEM_BUG_ON(!pt->is_compact);
50 	d->offset += SZ_2M;
51 }
52 
53 static void xehpsdv_insert_pte(struct i915_address_space *vm,
54 			       struct i915_page_table *pt,
55 			       void *data)
56 {
57 	struct insert_pte_data *d = data;
58 
59 	/*
60 	 * We are playing tricks here, since the actual pt, from the hw
61 	 * pov, is only 256bytes with 32 entries, or 4096bytes with 512
62 	 * entries, but we are still guaranteed that the physical
63 	 * alignment is 64K underneath for the pt, and we are careful
64 	 * not to access the space in the void.
65 	 */
66 	vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE, PTE_LM);
67 	d->offset += SZ_64K;
68 }
69 
70 static void insert_pte(struct i915_address_space *vm,
71 		       struct i915_page_table *pt,
72 		       void *data)
73 {
74 	struct insert_pte_data *d = data;
75 
76 	vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE,
77 			i915_gem_object_is_lmem(pt->base) ? PTE_LM : 0);
78 	d->offset += PAGE_SIZE;
79 }
80 
81 static struct i915_address_space *migrate_vm(struct intel_gt *gt)
82 {
83 	struct i915_vm_pt_stash stash = {};
84 	struct i915_ppgtt *vm;
85 	int err;
86 	int i;
87 
88 	/*
89 	 * We construct a very special VM for use by all migration contexts,
90 	 * it is kept pinned so that it can be used at any time. As we need
91 	 * to pre-allocate the page directories for the migration VM, this
92 	 * limits us to only using a small number of prepared vma.
93 	 *
94 	 * To be able to pipeline and reschedule migration operations while
95 	 * avoiding unnecessary contention on the vm itself, the PTE updates
96 	 * are inline with the blits. All the blits use the same fixed
97 	 * addresses, with the backing store redirection being updated on the
98 	 * fly. Only 2 implicit vma are used for all migration operations.
99 	 *
100 	 * We lay the ppGTT out as:
101 	 *
102 	 *	[0, CHUNK_SZ) -> first object
103 	 *	[CHUNK_SZ, 2 * CHUNK_SZ) -> second object
104 	 *	[2 * CHUNK_SZ, 2 * CHUNK_SZ + 2 * CHUNK_SZ >> 9] -> PTE
105 	 *
106 	 * By exposing the dma addresses of the page directories themselves
107 	 * within the ppGTT, we are then able to rewrite the PTE prior to use.
108 	 * But the PTE update and subsequent migration operation must be atomic,
109 	 * i.e. within the same non-preemptible window so that we do not switch
110 	 * to another migration context that overwrites the PTE.
111 	 *
112 	 * This changes quite a bit on platforms with HAS_64K_PAGES support,
113 	 * where we instead have three windows, each CHUNK_SIZE in size. The
114 	 * first is reserved for mapping system-memory, and that just uses the
115 	 * 512 entry layout using 4K GTT pages. The other two windows just map
116 	 * lmem pages and must use the new compact 32 entry layout using 64K GTT
117 	 * pages, which ensures we can address any lmem object that the user
118 	 * throws at us. We then also use the xehpsdv_toggle_pdes as a way of
119 	 * just toggling the PDE bit(GEN12_PDE_64K) for us, to enable the
120 	 * compact layout for each of these page-tables, that fall within the
121 	 * [CHUNK_SIZE, 3 * CHUNK_SIZE) range.
122 	 *
123 	 * We lay the ppGTT out as:
124 	 *
125 	 * [0, CHUNK_SZ) -> first window/object, maps smem
126 	 * [CHUNK_SZ, 2 * CHUNK_SZ) -> second window/object, maps lmem src
127 	 * [2 * CHUNK_SZ, 3 * CHUNK_SZ) -> third window/object, maps lmem dst
128 	 *
129 	 * For the PTE window it's also quite different, since each PTE must
130 	 * point to some 64K page, one for each PT(since it's in lmem), and yet
131 	 * each is only <= 4096bytes, but since the unused space within that PTE
132 	 * range is never touched, this should be fine.
133 	 *
134 	 * So basically each PT now needs 64K of virtual memory, instead of 4K,
135 	 * which looks like:
136 	 *
137 	 * [3 * CHUNK_SZ, 3 * CHUNK_SZ + ((3 * CHUNK_SZ / SZ_2M) * SZ_64K)] -> PTE
138 	 */
139 
140 	vm = i915_ppgtt_create(gt, I915_BO_ALLOC_PM_EARLY);
141 	if (IS_ERR(vm))
142 		return ERR_CAST(vm);
143 
144 	if (!vm->vm.allocate_va_range || !vm->vm.foreach) {
145 		err = -ENODEV;
146 		goto err_vm;
147 	}
148 
149 	if (HAS_64K_PAGES(gt->i915))
150 		stash.pt_sz = I915_GTT_PAGE_SIZE_64K;
151 
152 	/*
153 	 * Each engine instance is assigned its own chunk in the VM, so
154 	 * that we can run multiple instances concurrently
155 	 */
156 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
157 		struct intel_engine_cs *engine;
158 		u64 base = (u64)i << 32;
159 		struct insert_pte_data d = {};
160 		struct i915_gem_ww_ctx ww;
161 		u64 sz;
162 
163 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
164 		if (!engine_supports_migration(engine))
165 			continue;
166 
167 		/*
168 		 * We copy in 8MiB chunks. Each PDE covers 2MiB, so we need
169 		 * 4x2 page directories for source/destination.
170 		 */
171 		if (HAS_64K_PAGES(gt->i915))
172 			sz = 3 * CHUNK_SZ;
173 		else
174 			sz = 2 * CHUNK_SZ;
175 		d.offset = base + sz;
176 
177 		/*
178 		 * We need another page directory setup so that we can write
179 		 * the 8x512 PTE in each chunk.
180 		 */
181 		if (HAS_64K_PAGES(gt->i915))
182 			sz += (sz / SZ_2M) * SZ_64K;
183 		else
184 			sz += (sz >> 12) * sizeof(u64);
185 
186 		err = i915_vm_alloc_pt_stash(&vm->vm, &stash, sz);
187 		if (err)
188 			goto err_vm;
189 
190 		for_i915_gem_ww(&ww, err, true) {
191 			err = i915_vm_lock_objects(&vm->vm, &ww);
192 			if (err)
193 				continue;
194 			err = i915_vm_map_pt_stash(&vm->vm, &stash);
195 			if (err)
196 				continue;
197 
198 			vm->vm.allocate_va_range(&vm->vm, &stash, base, sz);
199 		}
200 		i915_vm_free_pt_stash(&vm->vm, &stash);
201 		if (err)
202 			goto err_vm;
203 
204 		/* Now allow the GPU to rewrite the PTE via its own ppGTT */
205 		if (HAS_64K_PAGES(gt->i915)) {
206 			vm->vm.foreach(&vm->vm, base, d.offset - base,
207 				       xehpsdv_insert_pte, &d);
208 			d.offset = base + CHUNK_SZ;
209 			vm->vm.foreach(&vm->vm,
210 				       d.offset,
211 				       2 * CHUNK_SZ,
212 				       xehpsdv_toggle_pdes, &d);
213 		} else {
214 			vm->vm.foreach(&vm->vm, base, d.offset - base,
215 				       insert_pte, &d);
216 		}
217 	}
218 
219 	return &vm->vm;
220 
221 err_vm:
222 	i915_vm_put(&vm->vm);
223 	return ERR_PTR(err);
224 }
225 
226 static struct intel_engine_cs *first_copy_engine(struct intel_gt *gt)
227 {
228 	struct intel_engine_cs *engine;
229 	int i;
230 
231 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
232 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
233 		if (engine_supports_migration(engine))
234 			return engine;
235 	}
236 
237 	return NULL;
238 }
239 
240 static struct intel_context *pinned_context(struct intel_gt *gt)
241 {
242 	static struct lock_class_key key;
243 	struct intel_engine_cs *engine;
244 	struct i915_address_space *vm;
245 	struct intel_context *ce;
246 
247 	engine = first_copy_engine(gt);
248 	if (!engine)
249 		return ERR_PTR(-ENODEV);
250 
251 	vm = migrate_vm(gt);
252 	if (IS_ERR(vm))
253 		return ERR_CAST(vm);
254 
255 	ce = intel_engine_create_pinned_context(engine, vm, SZ_512K,
256 						I915_GEM_HWS_MIGRATE,
257 						&key, "migrate");
258 	i915_vm_put(vm);
259 	return ce;
260 }
261 
262 int intel_migrate_init(struct intel_migrate *m, struct intel_gt *gt)
263 {
264 	struct intel_context *ce;
265 
266 	memset(m, 0, sizeof(*m));
267 
268 	ce = pinned_context(gt);
269 	if (IS_ERR(ce))
270 		return PTR_ERR(ce);
271 
272 	m->context = ce;
273 	return 0;
274 }
275 
276 static int random_index(unsigned int max)
277 {
278 	return upper_32_bits(mul_u32_u32(get_random_u32(), max));
279 }
280 
281 static struct intel_context *__migrate_engines(struct intel_gt *gt)
282 {
283 	struct intel_engine_cs *engines[MAX_ENGINE_INSTANCE];
284 	struct intel_engine_cs *engine;
285 	unsigned int count, i;
286 
287 	count = 0;
288 	for (i = 0; i < ARRAY_SIZE(gt->engine_class[COPY_ENGINE_CLASS]); i++) {
289 		engine = gt->engine_class[COPY_ENGINE_CLASS][i];
290 		if (engine_supports_migration(engine))
291 			engines[count++] = engine;
292 	}
293 
294 	return intel_context_create(engines[random_index(count)]);
295 }
296 
297 struct intel_context *intel_migrate_create_context(struct intel_migrate *m)
298 {
299 	struct intel_context *ce;
300 
301 	/*
302 	 * We randomly distribute contexts across the engines upon constrction,
303 	 * as they all share the same pinned vm, and so in order to allow
304 	 * multiple blits to run in parallel, we must construct each blit
305 	 * to use a different range of the vm for its GTT. This has to be
306 	 * known at construction, so we can not use the late greedy load
307 	 * balancing of the virtual-engine.
308 	 */
309 	ce = __migrate_engines(m->context->engine->gt);
310 	if (IS_ERR(ce))
311 		return ce;
312 
313 	ce->ring = NULL;
314 	ce->ring_size = SZ_256K;
315 
316 	i915_vm_put(ce->vm);
317 	ce->vm = i915_vm_get(m->context->vm);
318 
319 	return ce;
320 }
321 
322 static inline struct sgt_dma sg_sgt(struct scatterlist *sg)
323 {
324 	dma_addr_t addr = sg_dma_address(sg);
325 
326 	return (struct sgt_dma){ sg, addr, addr + sg_dma_len(sg) };
327 }
328 
329 static int emit_no_arbitration(struct i915_request *rq)
330 {
331 	u32 *cs;
332 
333 	cs = intel_ring_begin(rq, 2);
334 	if (IS_ERR(cs))
335 		return PTR_ERR(cs);
336 
337 	/* Explicitly disable preemption for this request. */
338 	*cs++ = MI_ARB_ON_OFF;
339 	*cs++ = MI_NOOP;
340 	intel_ring_advance(rq, cs);
341 
342 	return 0;
343 }
344 
345 static int max_pte_pkt_size(struct i915_request *rq, int pkt)
346 {
347 	struct intel_ring *ring = rq->ring;
348 
349 	pkt = min_t(int, pkt, (ring->space - rq->reserved_space) / sizeof(u32) + 5);
350 	pkt = min_t(int, pkt, (ring->size - ring->emit) / sizeof(u32) + 5);
351 
352 	return pkt;
353 }
354 
355 #define I915_EMIT_PTE_NUM_DWORDS 6
356 
357 static int emit_pte(struct i915_request *rq,
358 		    struct sgt_dma *it,
359 		    enum i915_cache_level cache_level,
360 		    bool is_lmem,
361 		    u64 offset,
362 		    int length)
363 {
364 	bool has_64K_pages = HAS_64K_PAGES(rq->engine->i915);
365 	const u64 encode = rq->context->vm->pte_encode(0, cache_level,
366 						       is_lmem ? PTE_LM : 0);
367 	struct intel_ring *ring = rq->ring;
368 	int pkt, dword_length;
369 	u32 total = 0;
370 	u32 page_size;
371 	u32 *hdr, *cs;
372 
373 	GEM_BUG_ON(GRAPHICS_VER(rq->engine->i915) < 8);
374 
375 	page_size = I915_GTT_PAGE_SIZE;
376 	dword_length = 0x400;
377 
378 	/* Compute the page directory offset for the target address range */
379 	if (has_64K_pages) {
380 		GEM_BUG_ON(!IS_ALIGNED(offset, SZ_2M));
381 
382 		offset /= SZ_2M;
383 		offset *= SZ_64K;
384 		offset += 3 * CHUNK_SZ;
385 
386 		if (is_lmem) {
387 			page_size = I915_GTT_PAGE_SIZE_64K;
388 			dword_length = 0x40;
389 		}
390 	} else {
391 		offset >>= 12;
392 		offset *= sizeof(u64);
393 		offset += 2 * CHUNK_SZ;
394 	}
395 
396 	offset += (u64)rq->engine->instance << 32;
397 
398 	cs = intel_ring_begin(rq, I915_EMIT_PTE_NUM_DWORDS);
399 	if (IS_ERR(cs))
400 		return PTR_ERR(cs);
401 
402 	/* Pack as many PTE updates as possible into a single MI command */
403 	pkt = max_pte_pkt_size(rq, dword_length);
404 
405 	hdr = cs;
406 	*cs++ = MI_STORE_DATA_IMM | REG_BIT(21); /* as qword elements */
407 	*cs++ = lower_32_bits(offset);
408 	*cs++ = upper_32_bits(offset);
409 
410 	do {
411 		if (cs - hdr >= pkt) {
412 			int dword_rem;
413 
414 			*hdr += cs - hdr - 2;
415 			*cs++ = MI_NOOP;
416 
417 			ring->emit = (void *)cs - ring->vaddr;
418 			intel_ring_advance(rq, cs);
419 			intel_ring_update_space(ring);
420 
421 			cs = intel_ring_begin(rq, I915_EMIT_PTE_NUM_DWORDS);
422 			if (IS_ERR(cs))
423 				return PTR_ERR(cs);
424 
425 			dword_rem = dword_length;
426 			if (has_64K_pages) {
427 				if (IS_ALIGNED(total, SZ_2M)) {
428 					offset = round_up(offset, SZ_64K);
429 				} else {
430 					dword_rem = SZ_2M - (total & (SZ_2M - 1));
431 					dword_rem /= page_size;
432 					dword_rem *= 2;
433 				}
434 			}
435 
436 			pkt = max_pte_pkt_size(rq, dword_rem);
437 
438 			hdr = cs;
439 			*cs++ = MI_STORE_DATA_IMM | REG_BIT(21);
440 			*cs++ = lower_32_bits(offset);
441 			*cs++ = upper_32_bits(offset);
442 		}
443 
444 		GEM_BUG_ON(!IS_ALIGNED(it->dma, page_size));
445 
446 		*cs++ = lower_32_bits(encode | it->dma);
447 		*cs++ = upper_32_bits(encode | it->dma);
448 
449 		offset += 8;
450 		total += page_size;
451 
452 		it->dma += page_size;
453 		if (it->dma >= it->max) {
454 			it->sg = __sg_next(it->sg);
455 			if (!it->sg || sg_dma_len(it->sg) == 0)
456 				break;
457 
458 			it->dma = sg_dma_address(it->sg);
459 			it->max = it->dma + sg_dma_len(it->sg);
460 		}
461 	} while (total < length);
462 
463 	*hdr += cs - hdr - 2;
464 	*cs++ = MI_NOOP;
465 
466 	ring->emit = (void *)cs - ring->vaddr;
467 	intel_ring_advance(rq, cs);
468 	intel_ring_update_space(ring);
469 
470 	return total;
471 }
472 
473 static bool wa_1209644611_applies(int ver, u32 size)
474 {
475 	u32 height = size >> PAGE_SHIFT;
476 
477 	if (ver != 11)
478 		return false;
479 
480 	return height % 4 == 3 && height <= 8;
481 }
482 
483 /**
484  * DOC: Flat-CCS - Memory compression for Local memory
485  *
486  * On Xe-HP and later devices, we use dedicated compression control state (CCS)
487  * stored in local memory for each surface, to support the 3D and media
488  * compression formats.
489  *
490  * The memory required for the CCS of the entire local memory is 1/256 of the
491  * local memory size. So before the kernel boot, the required memory is reserved
492  * for the CCS data and a secure register will be programmed with the CCS base
493  * address.
494  *
495  * Flat CCS data needs to be cleared when a lmem object is allocated.
496  * And CCS data can be copied in and out of CCS region through
497  * XY_CTRL_SURF_COPY_BLT. CPU can't access the CCS data directly.
498  *
499  * I915 supports Flat-CCS on lmem only objects. When an objects has smem in
500  * its preference list, on memory pressure, i915 needs to migrate the lmem
501  * content into smem. If the lmem object is Flat-CCS compressed by userspace,
502  * then i915 needs to decompress it. But I915 lack the required information
503  * for such decompression. Hence I915 supports Flat-CCS only on lmem only objects.
504  *
505  * When we exhaust the lmem, Flat-CCS capable objects' lmem backing memory can
506  * be temporarily evicted to smem, along with the auxiliary CCS state, where
507  * it can be potentially swapped-out at a later point, if required.
508  * If userspace later touches the evicted pages, then we always move
509  * the backing memory back to lmem, which includes restoring the saved CCS state,
510  * and potentially performing any required swap-in.
511  *
512  * For the migration of the lmem objects with smem in placement list, such as
513  * {lmem, smem}, objects are treated as non Flat-CCS capable objects.
514  */
515 
516 static inline u32 *i915_flush_dw(u32 *cmd, u32 flags)
517 {
518 	*cmd++ = MI_FLUSH_DW | flags;
519 	*cmd++ = 0;
520 	*cmd++ = 0;
521 
522 	return cmd;
523 }
524 
525 static int emit_copy_ccs(struct i915_request *rq,
526 			 u32 dst_offset, u8 dst_access,
527 			 u32 src_offset, u8 src_access, int size)
528 {
529 	struct drm_i915_private *i915 = rq->engine->i915;
530 	int mocs = rq->engine->gt->mocs.uc_index << 1;
531 	u32 num_ccs_blks;
532 	u32 *cs;
533 
534 	cs = intel_ring_begin(rq, 12);
535 	if (IS_ERR(cs))
536 		return PTR_ERR(cs);
537 
538 	num_ccs_blks = DIV_ROUND_UP(GET_CCS_BYTES(i915, size),
539 				    NUM_CCS_BYTES_PER_BLOCK);
540 	GEM_BUG_ON(num_ccs_blks > NUM_CCS_BLKS_PER_XFER);
541 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
542 
543 	/*
544 	 * The XY_CTRL_SURF_COPY_BLT instruction is used to copy the CCS
545 	 * data in and out of the CCS region.
546 	 *
547 	 * We can copy at most 1024 blocks of 256 bytes using one
548 	 * XY_CTRL_SURF_COPY_BLT instruction.
549 	 *
550 	 * In case we need to copy more than 1024 blocks, we need to add
551 	 * another instruction to the same batch buffer.
552 	 *
553 	 * 1024 blocks of 256 bytes of CCS represent a total 256KB of CCS.
554 	 *
555 	 * 256 KB of CCS represents 256 * 256 KB = 64 MB of LMEM.
556 	 */
557 	*cs++ = XY_CTRL_SURF_COPY_BLT |
558 		src_access << SRC_ACCESS_TYPE_SHIFT |
559 		dst_access << DST_ACCESS_TYPE_SHIFT |
560 		((num_ccs_blks - 1) & CCS_SIZE_MASK) << CCS_SIZE_SHIFT;
561 	*cs++ = src_offset;
562 	*cs++ = rq->engine->instance |
563 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
564 	*cs++ = dst_offset;
565 	*cs++ = rq->engine->instance |
566 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
567 
568 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
569 	*cs++ = MI_NOOP;
570 
571 	intel_ring_advance(rq, cs);
572 
573 	return 0;
574 }
575 
576 static int emit_copy(struct i915_request *rq,
577 		     u32 dst_offset, u32 src_offset, int size)
578 {
579 	const int ver = GRAPHICS_VER(rq->engine->i915);
580 	u32 instance = rq->engine->instance;
581 	u32 *cs;
582 
583 	cs = intel_ring_begin(rq, ver >= 8 ? 10 : 6);
584 	if (IS_ERR(cs))
585 		return PTR_ERR(cs);
586 
587 	if (ver >= 9 && !wa_1209644611_applies(ver, size)) {
588 		*cs++ = GEN9_XY_FAST_COPY_BLT_CMD | (10 - 2);
589 		*cs++ = BLT_DEPTH_32 | PAGE_SIZE;
590 		*cs++ = 0;
591 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
592 		*cs++ = dst_offset;
593 		*cs++ = instance;
594 		*cs++ = 0;
595 		*cs++ = PAGE_SIZE;
596 		*cs++ = src_offset;
597 		*cs++ = instance;
598 	} else if (ver >= 8) {
599 		*cs++ = XY_SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (10 - 2);
600 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
601 		*cs++ = 0;
602 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
603 		*cs++ = dst_offset;
604 		*cs++ = instance;
605 		*cs++ = 0;
606 		*cs++ = PAGE_SIZE;
607 		*cs++ = src_offset;
608 		*cs++ = instance;
609 	} else {
610 		GEM_BUG_ON(instance);
611 		*cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
612 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
613 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE;
614 		*cs++ = dst_offset;
615 		*cs++ = PAGE_SIZE;
616 		*cs++ = src_offset;
617 	}
618 
619 	intel_ring_advance(rq, cs);
620 	return 0;
621 }
622 
623 static u64 scatter_list_length(struct scatterlist *sg)
624 {
625 	u64 len = 0;
626 
627 	while (sg && sg_dma_len(sg)) {
628 		len += sg_dma_len(sg);
629 		sg = sg_next(sg);
630 	}
631 
632 	return len;
633 }
634 
635 static int
636 calculate_chunk_sz(struct drm_i915_private *i915, bool src_is_lmem,
637 		   u64 bytes_to_cpy, u64 ccs_bytes_to_cpy)
638 {
639 	if (ccs_bytes_to_cpy && !src_is_lmem)
640 		/*
641 		 * When CHUNK_SZ is passed all the pages upto CHUNK_SZ
642 		 * will be taken for the blt. in Flat-ccs supported
643 		 * platform Smem obj will have more pages than required
644 		 * for main meory hence limit it to the required size
645 		 * for main memory
646 		 */
647 		return min_t(u64, bytes_to_cpy, CHUNK_SZ);
648 	else
649 		return CHUNK_SZ;
650 }
651 
652 static void get_ccs_sg_sgt(struct sgt_dma *it, u64 bytes_to_cpy)
653 {
654 	u64 len;
655 
656 	do {
657 		GEM_BUG_ON(!it->sg || !sg_dma_len(it->sg));
658 		len = it->max - it->dma;
659 		if (len > bytes_to_cpy) {
660 			it->dma += bytes_to_cpy;
661 			break;
662 		}
663 
664 		bytes_to_cpy -= len;
665 
666 		it->sg = __sg_next(it->sg);
667 		it->dma = sg_dma_address(it->sg);
668 		it->max = it->dma + sg_dma_len(it->sg);
669 	} while (bytes_to_cpy);
670 }
671 
672 int
673 intel_context_migrate_copy(struct intel_context *ce,
674 			   const struct i915_deps *deps,
675 			   struct scatterlist *src,
676 			   enum i915_cache_level src_cache_level,
677 			   bool src_is_lmem,
678 			   struct scatterlist *dst,
679 			   enum i915_cache_level dst_cache_level,
680 			   bool dst_is_lmem,
681 			   struct i915_request **out)
682 {
683 	struct sgt_dma it_src = sg_sgt(src), it_dst = sg_sgt(dst), it_ccs;
684 	struct drm_i915_private *i915 = ce->engine->i915;
685 	u64 ccs_bytes_to_cpy = 0, bytes_to_cpy;
686 	enum i915_cache_level ccs_cache_level;
687 	u32 src_offset, dst_offset;
688 	u8 src_access, dst_access;
689 	struct i915_request *rq;
690 	u64 src_sz, dst_sz;
691 	bool ccs_is_src, overwrite_ccs;
692 	int err;
693 
694 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
695 	GEM_BUG_ON(IS_DGFX(ce->engine->i915) && (!src_is_lmem && !dst_is_lmem));
696 	*out = NULL;
697 
698 	GEM_BUG_ON(ce->ring->size < SZ_64K);
699 
700 	src_sz = scatter_list_length(src);
701 	bytes_to_cpy = src_sz;
702 
703 	if (HAS_FLAT_CCS(i915) && src_is_lmem ^ dst_is_lmem) {
704 		src_access = !src_is_lmem && dst_is_lmem;
705 		dst_access = !src_access;
706 
707 		dst_sz = scatter_list_length(dst);
708 		if (src_is_lmem) {
709 			it_ccs = it_dst;
710 			ccs_cache_level = dst_cache_level;
711 			ccs_is_src = false;
712 		} else if (dst_is_lmem) {
713 			bytes_to_cpy = dst_sz;
714 			it_ccs = it_src;
715 			ccs_cache_level = src_cache_level;
716 			ccs_is_src = true;
717 		}
718 
719 		/*
720 		 * When there is a eviction of ccs needed smem will have the
721 		 * extra pages for the ccs data
722 		 *
723 		 * TO-DO: Want to move the size mismatch check to a WARN_ON,
724 		 * but still we have some requests of smem->lmem with same size.
725 		 * Need to fix it.
726 		 */
727 		ccs_bytes_to_cpy = src_sz != dst_sz ? GET_CCS_BYTES(i915, bytes_to_cpy) : 0;
728 		if (ccs_bytes_to_cpy)
729 			get_ccs_sg_sgt(&it_ccs, bytes_to_cpy);
730 	}
731 
732 	overwrite_ccs = HAS_FLAT_CCS(i915) && !ccs_bytes_to_cpy && dst_is_lmem;
733 
734 	src_offset = 0;
735 	dst_offset = CHUNK_SZ;
736 	if (HAS_64K_PAGES(ce->engine->i915)) {
737 		src_offset = 0;
738 		dst_offset = 0;
739 		if (src_is_lmem)
740 			src_offset = CHUNK_SZ;
741 		if (dst_is_lmem)
742 			dst_offset = 2 * CHUNK_SZ;
743 	}
744 
745 	do {
746 		int len;
747 
748 		rq = i915_request_create(ce);
749 		if (IS_ERR(rq)) {
750 			err = PTR_ERR(rq);
751 			goto out_ce;
752 		}
753 
754 		if (deps) {
755 			err = i915_request_await_deps(rq, deps);
756 			if (err)
757 				goto out_rq;
758 
759 			if (rq->engine->emit_init_breadcrumb) {
760 				err = rq->engine->emit_init_breadcrumb(rq);
761 				if (err)
762 					goto out_rq;
763 			}
764 
765 			deps = NULL;
766 		}
767 
768 		/* The PTE updates + copy must not be interrupted. */
769 		err = emit_no_arbitration(rq);
770 		if (err)
771 			goto out_rq;
772 
773 		src_sz = calculate_chunk_sz(i915, src_is_lmem,
774 					    bytes_to_cpy, ccs_bytes_to_cpy);
775 
776 		len = emit_pte(rq, &it_src, src_cache_level, src_is_lmem,
777 			       src_offset, src_sz);
778 		if (!len) {
779 			err = -EINVAL;
780 			goto out_rq;
781 		}
782 		if (len < 0) {
783 			err = len;
784 			goto out_rq;
785 		}
786 
787 		err = emit_pte(rq, &it_dst, dst_cache_level, dst_is_lmem,
788 			       dst_offset, len);
789 		if (err < 0)
790 			goto out_rq;
791 		if (err < len) {
792 			err = -EINVAL;
793 			goto out_rq;
794 		}
795 
796 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
797 		if (err)
798 			goto out_rq;
799 
800 		err = emit_copy(rq, dst_offset,	src_offset, len);
801 		if (err)
802 			goto out_rq;
803 
804 		bytes_to_cpy -= len;
805 
806 		if (ccs_bytes_to_cpy) {
807 			int ccs_sz;
808 
809 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
810 			if (err)
811 				goto out_rq;
812 
813 			ccs_sz = GET_CCS_BYTES(i915, len);
814 			err = emit_pte(rq, &it_ccs, ccs_cache_level, false,
815 				       ccs_is_src ? src_offset : dst_offset,
816 				       ccs_sz);
817 			if (err < 0)
818 				goto out_rq;
819 			if (err < ccs_sz) {
820 				err = -EINVAL;
821 				goto out_rq;
822 			}
823 
824 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
825 			if (err)
826 				goto out_rq;
827 
828 			err = emit_copy_ccs(rq, dst_offset, dst_access,
829 					    src_offset, src_access, len);
830 			if (err)
831 				goto out_rq;
832 
833 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
834 			if (err)
835 				goto out_rq;
836 			ccs_bytes_to_cpy -= ccs_sz;
837 		} else if (overwrite_ccs) {
838 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
839 			if (err)
840 				goto out_rq;
841 
842 			if (src_is_lmem) {
843 				/*
844 				 * If the src is already in lmem, then we must
845 				 * be doing an lmem -> lmem transfer, and so
846 				 * should be safe to directly copy the CCS
847 				 * state. In this case we have either
848 				 * initialised the CCS aux state when first
849 				 * clearing the pages (since it is already
850 				 * allocated in lmem), or the user has
851 				 * potentially populated it, in which case we
852 				 * need to copy the CCS state as-is.
853 				 */
854 				err = emit_copy_ccs(rq,
855 						    dst_offset, INDIRECT_ACCESS,
856 						    src_offset, INDIRECT_ACCESS,
857 						    len);
858 			} else {
859 				/*
860 				 * While we can't always restore/manage the CCS
861 				 * state, we still need to ensure we don't leak
862 				 * the CCS state from the previous user, so make
863 				 * sure we overwrite it with something.
864 				 */
865 				err = emit_copy_ccs(rq,
866 						    dst_offset, INDIRECT_ACCESS,
867 						    dst_offset, DIRECT_ACCESS,
868 						    len);
869 			}
870 
871 			if (err)
872 				goto out_rq;
873 
874 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
875 			if (err)
876 				goto out_rq;
877 		}
878 
879 		/* Arbitration is re-enabled between requests. */
880 out_rq:
881 		if (*out)
882 			i915_request_put(*out);
883 		*out = i915_request_get(rq);
884 		i915_request_add(rq);
885 
886 		if (err)
887 			break;
888 
889 		if (!bytes_to_cpy && !ccs_bytes_to_cpy) {
890 			if (src_is_lmem)
891 				WARN_ON(it_src.sg && sg_dma_len(it_src.sg));
892 			else
893 				WARN_ON(it_dst.sg && sg_dma_len(it_dst.sg));
894 			break;
895 		}
896 
897 		if (WARN_ON(!it_src.sg || !sg_dma_len(it_src.sg) ||
898 			    !it_dst.sg || !sg_dma_len(it_dst.sg) ||
899 			    (ccs_bytes_to_cpy && (!it_ccs.sg ||
900 						  !sg_dma_len(it_ccs.sg))))) {
901 			err = -EINVAL;
902 			break;
903 		}
904 
905 		cond_resched();
906 	} while (1);
907 
908 out_ce:
909 	return err;
910 }
911 
912 static int emit_clear(struct i915_request *rq, u32 offset, int size,
913 		      u32 value, bool is_lmem)
914 {
915 	struct drm_i915_private *i915 = rq->engine->i915;
916 	int mocs = rq->engine->gt->mocs.uc_index << 1;
917 	const int ver = GRAPHICS_VER(i915);
918 	int ring_sz;
919 	u32 *cs;
920 
921 	GEM_BUG_ON(size >> PAGE_SHIFT > S16_MAX);
922 
923 	if (HAS_FLAT_CCS(i915) && ver >= 12)
924 		ring_sz = XY_FAST_COLOR_BLT_DW;
925 	else if (ver >= 8)
926 		ring_sz = 8;
927 	else
928 		ring_sz = 6;
929 
930 	cs = intel_ring_begin(rq, ring_sz);
931 	if (IS_ERR(cs))
932 		return PTR_ERR(cs);
933 
934 	if (HAS_FLAT_CCS(i915) && ver >= 12) {
935 		*cs++ = XY_FAST_COLOR_BLT_CMD | XY_FAST_COLOR_BLT_DEPTH_32 |
936 			(XY_FAST_COLOR_BLT_DW - 2);
937 		*cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, mocs) |
938 			(PAGE_SIZE - 1);
939 		*cs++ = 0;
940 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
941 		*cs++ = offset;
942 		*cs++ = rq->engine->instance;
943 		*cs++ = !is_lmem << XY_FAST_COLOR_BLT_MEM_TYPE_SHIFT;
944 		/* BG7 */
945 		*cs++ = value;
946 		*cs++ = 0;
947 		*cs++ = 0;
948 		*cs++ = 0;
949 		/* BG11 */
950 		*cs++ = 0;
951 		*cs++ = 0;
952 		/* BG13 */
953 		*cs++ = 0;
954 		*cs++ = 0;
955 		*cs++ = 0;
956 	} else if (ver >= 8) {
957 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (7 - 2);
958 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
959 		*cs++ = 0;
960 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
961 		*cs++ = offset;
962 		*cs++ = rq->engine->instance;
963 		*cs++ = value;
964 		*cs++ = MI_NOOP;
965 	} else {
966 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
967 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
968 		*cs++ = 0;
969 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
970 		*cs++ = offset;
971 		*cs++ = value;
972 	}
973 
974 	intel_ring_advance(rq, cs);
975 	return 0;
976 }
977 
978 int
979 intel_context_migrate_clear(struct intel_context *ce,
980 			    const struct i915_deps *deps,
981 			    struct scatterlist *sg,
982 			    enum i915_cache_level cache_level,
983 			    bool is_lmem,
984 			    u32 value,
985 			    struct i915_request **out)
986 {
987 	struct drm_i915_private *i915 = ce->engine->i915;
988 	struct sgt_dma it = sg_sgt(sg);
989 	struct i915_request *rq;
990 	u32 offset;
991 	int err;
992 
993 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
994 	*out = NULL;
995 
996 	GEM_BUG_ON(ce->ring->size < SZ_64K);
997 
998 	offset = 0;
999 	if (HAS_64K_PAGES(i915) && is_lmem)
1000 		offset = CHUNK_SZ;
1001 
1002 	do {
1003 		int len;
1004 
1005 		rq = i915_request_create(ce);
1006 		if (IS_ERR(rq)) {
1007 			err = PTR_ERR(rq);
1008 			goto out_ce;
1009 		}
1010 
1011 		if (deps) {
1012 			err = i915_request_await_deps(rq, deps);
1013 			if (err)
1014 				goto out_rq;
1015 
1016 			if (rq->engine->emit_init_breadcrumb) {
1017 				err = rq->engine->emit_init_breadcrumb(rq);
1018 				if (err)
1019 					goto out_rq;
1020 			}
1021 
1022 			deps = NULL;
1023 		}
1024 
1025 		/* The PTE updates + clear must not be interrupted. */
1026 		err = emit_no_arbitration(rq);
1027 		if (err)
1028 			goto out_rq;
1029 
1030 		len = emit_pte(rq, &it, cache_level, is_lmem, offset, CHUNK_SZ);
1031 		if (len <= 0) {
1032 			err = len;
1033 			goto out_rq;
1034 		}
1035 
1036 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1037 		if (err)
1038 			goto out_rq;
1039 
1040 		err = emit_clear(rq, offset, len, value, is_lmem);
1041 		if (err)
1042 			goto out_rq;
1043 
1044 		if (HAS_FLAT_CCS(i915) && is_lmem && !value) {
1045 			/*
1046 			 * copy the content of memory into corresponding
1047 			 * ccs surface
1048 			 */
1049 			err = emit_copy_ccs(rq, offset, INDIRECT_ACCESS, offset,
1050 					    DIRECT_ACCESS, len);
1051 			if (err)
1052 				goto out_rq;
1053 		}
1054 
1055 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1056 
1057 		/* Arbitration is re-enabled between requests. */
1058 out_rq:
1059 		if (*out)
1060 			i915_request_put(*out);
1061 		*out = i915_request_get(rq);
1062 		i915_request_add(rq);
1063 		if (err || !it.sg || !sg_dma_len(it.sg))
1064 			break;
1065 
1066 		cond_resched();
1067 	} while (1);
1068 
1069 out_ce:
1070 	return err;
1071 }
1072 
1073 int intel_migrate_copy(struct intel_migrate *m,
1074 		       struct i915_gem_ww_ctx *ww,
1075 		       const struct i915_deps *deps,
1076 		       struct scatterlist *src,
1077 		       enum i915_cache_level src_cache_level,
1078 		       bool src_is_lmem,
1079 		       struct scatterlist *dst,
1080 		       enum i915_cache_level dst_cache_level,
1081 		       bool dst_is_lmem,
1082 		       struct i915_request **out)
1083 {
1084 	struct intel_context *ce;
1085 	int err;
1086 
1087 	*out = NULL;
1088 	if (!m->context)
1089 		return -ENODEV;
1090 
1091 	ce = intel_migrate_create_context(m);
1092 	if (IS_ERR(ce))
1093 		ce = intel_context_get(m->context);
1094 	GEM_BUG_ON(IS_ERR(ce));
1095 
1096 	err = intel_context_pin_ww(ce, ww);
1097 	if (err)
1098 		goto out;
1099 
1100 	err = intel_context_migrate_copy(ce, deps,
1101 					 src, src_cache_level, src_is_lmem,
1102 					 dst, dst_cache_level, dst_is_lmem,
1103 					 out);
1104 
1105 	intel_context_unpin(ce);
1106 out:
1107 	intel_context_put(ce);
1108 	return err;
1109 }
1110 
1111 int
1112 intel_migrate_clear(struct intel_migrate *m,
1113 		    struct i915_gem_ww_ctx *ww,
1114 		    const struct i915_deps *deps,
1115 		    struct scatterlist *sg,
1116 		    enum i915_cache_level cache_level,
1117 		    bool is_lmem,
1118 		    u32 value,
1119 		    struct i915_request **out)
1120 {
1121 	struct intel_context *ce;
1122 	int err;
1123 
1124 	*out = NULL;
1125 	if (!m->context)
1126 		return -ENODEV;
1127 
1128 	ce = intel_migrate_create_context(m);
1129 	if (IS_ERR(ce))
1130 		ce = intel_context_get(m->context);
1131 	GEM_BUG_ON(IS_ERR(ce));
1132 
1133 	err = intel_context_pin_ww(ce, ww);
1134 	if (err)
1135 		goto out;
1136 
1137 	err = intel_context_migrate_clear(ce, deps, sg, cache_level,
1138 					  is_lmem, value, out);
1139 
1140 	intel_context_unpin(ce);
1141 out:
1142 	intel_context_put(ce);
1143 	return err;
1144 }
1145 
1146 void intel_migrate_fini(struct intel_migrate *m)
1147 {
1148 	struct intel_context *ce;
1149 
1150 	ce = fetch_and_zero(&m->context);
1151 	if (!ce)
1152 		return;
1153 
1154 	intel_engine_destroy_pinned_context(ce);
1155 }
1156 
1157 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1158 #include "selftest_migrate.c"
1159 #endif
1160