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 static int emit_pte(struct i915_request *rq,
356 		    struct sgt_dma *it,
357 		    enum i915_cache_level cache_level,
358 		    bool is_lmem,
359 		    u64 offset,
360 		    int length)
361 {
362 	bool has_64K_pages = HAS_64K_PAGES(rq->engine->i915);
363 	const u64 encode = rq->context->vm->pte_encode(0, cache_level,
364 						       is_lmem ? PTE_LM : 0);
365 	struct intel_ring *ring = rq->ring;
366 	int pkt, dword_length;
367 	u32 total = 0;
368 	u32 page_size;
369 	u32 *hdr, *cs;
370 
371 	GEM_BUG_ON(GRAPHICS_VER(rq->engine->i915) < 8);
372 
373 	page_size = I915_GTT_PAGE_SIZE;
374 	dword_length = 0x400;
375 
376 	/* Compute the page directory offset for the target address range */
377 	if (has_64K_pages) {
378 		GEM_BUG_ON(!IS_ALIGNED(offset, SZ_2M));
379 
380 		offset /= SZ_2M;
381 		offset *= SZ_64K;
382 		offset += 3 * CHUNK_SZ;
383 
384 		if (is_lmem) {
385 			page_size = I915_GTT_PAGE_SIZE_64K;
386 			dword_length = 0x40;
387 		}
388 	} else {
389 		offset >>= 12;
390 		offset *= sizeof(u64);
391 		offset += 2 * CHUNK_SZ;
392 	}
393 
394 	offset += (u64)rq->engine->instance << 32;
395 
396 	cs = intel_ring_begin(rq, 6);
397 	if (IS_ERR(cs))
398 		return PTR_ERR(cs);
399 
400 	/* Pack as many PTE updates as possible into a single MI command */
401 	pkt = max_pte_pkt_size(rq, dword_length);
402 
403 	hdr = cs;
404 	*cs++ = MI_STORE_DATA_IMM | REG_BIT(21); /* as qword elements */
405 	*cs++ = lower_32_bits(offset);
406 	*cs++ = upper_32_bits(offset);
407 
408 	do {
409 		if (cs - hdr >= pkt) {
410 			int dword_rem;
411 
412 			*hdr += cs - hdr - 2;
413 			*cs++ = MI_NOOP;
414 
415 			ring->emit = (void *)cs - ring->vaddr;
416 			intel_ring_advance(rq, cs);
417 			intel_ring_update_space(ring);
418 
419 			cs = intel_ring_begin(rq, 6);
420 			if (IS_ERR(cs))
421 				return PTR_ERR(cs);
422 
423 			dword_rem = dword_length;
424 			if (has_64K_pages) {
425 				if (IS_ALIGNED(total, SZ_2M)) {
426 					offset = round_up(offset, SZ_64K);
427 				} else {
428 					dword_rem = SZ_2M - (total & (SZ_2M - 1));
429 					dword_rem /= page_size;
430 					dword_rem *= 2;
431 				}
432 			}
433 
434 			pkt = max_pte_pkt_size(rq, dword_rem);
435 
436 			hdr = cs;
437 			*cs++ = MI_STORE_DATA_IMM | REG_BIT(21);
438 			*cs++ = lower_32_bits(offset);
439 			*cs++ = upper_32_bits(offset);
440 		}
441 
442 		GEM_BUG_ON(!IS_ALIGNED(it->dma, page_size));
443 
444 		*cs++ = lower_32_bits(encode | it->dma);
445 		*cs++ = upper_32_bits(encode | it->dma);
446 
447 		offset += 8;
448 		total += page_size;
449 
450 		it->dma += page_size;
451 		if (it->dma >= it->max) {
452 			it->sg = __sg_next(it->sg);
453 			if (!it->sg || sg_dma_len(it->sg) == 0)
454 				break;
455 
456 			it->dma = sg_dma_address(it->sg);
457 			it->max = it->dma + sg_dma_len(it->sg);
458 		}
459 	} while (total < length);
460 
461 	*hdr += cs - hdr - 2;
462 	*cs++ = MI_NOOP;
463 
464 	ring->emit = (void *)cs - ring->vaddr;
465 	intel_ring_advance(rq, cs);
466 	intel_ring_update_space(ring);
467 
468 	return total;
469 }
470 
471 static bool wa_1209644611_applies(int ver, u32 size)
472 {
473 	u32 height = size >> PAGE_SHIFT;
474 
475 	if (ver != 11)
476 		return false;
477 
478 	return height % 4 == 3 && height <= 8;
479 }
480 
481 /**
482  * DOC: Flat-CCS - Memory compression for Local memory
483  *
484  * On Xe-HP and later devices, we use dedicated compression control state (CCS)
485  * stored in local memory for each surface, to support the 3D and media
486  * compression formats.
487  *
488  * The memory required for the CCS of the entire local memory is 1/256 of the
489  * local memory size. So before the kernel boot, the required memory is reserved
490  * for the CCS data and a secure register will be programmed with the CCS base
491  * address.
492  *
493  * Flat CCS data needs to be cleared when a lmem object is allocated.
494  * And CCS data can be copied in and out of CCS region through
495  * XY_CTRL_SURF_COPY_BLT. CPU can't access the CCS data directly.
496  *
497  * I915 supports Flat-CCS on lmem only objects. When an objects has smem in
498  * its preference list, on memory pressure, i915 needs to migrate the lmem
499  * content into smem. If the lmem object is Flat-CCS compressed by userspace,
500  * then i915 needs to decompress it. But I915 lack the required information
501  * for such decompression. Hence I915 supports Flat-CCS only on lmem only objects.
502  *
503  * When we exhaust the lmem, Flat-CCS capable objects' lmem backing memory can
504  * be temporarily evicted to smem, along with the auxiliary CCS state, where
505  * it can be potentially swapped-out at a later point, if required.
506  * If userspace later touches the evicted pages, then we always move
507  * the backing memory back to lmem, which includes restoring the saved CCS state,
508  * and potentially performing any required swap-in.
509  *
510  * For the migration of the lmem objects with smem in placement list, such as
511  * {lmem, smem}, objects are treated as non Flat-CCS capable objects.
512  */
513 
514 static inline u32 *i915_flush_dw(u32 *cmd, u32 flags)
515 {
516 	*cmd++ = MI_FLUSH_DW | flags;
517 	*cmd++ = 0;
518 	*cmd++ = 0;
519 
520 	return cmd;
521 }
522 
523 static int emit_copy_ccs(struct i915_request *rq,
524 			 u32 dst_offset, u8 dst_access,
525 			 u32 src_offset, u8 src_access, int size)
526 {
527 	struct drm_i915_private *i915 = rq->engine->i915;
528 	int mocs = rq->engine->gt->mocs.uc_index << 1;
529 	u32 num_ccs_blks;
530 	u32 *cs;
531 
532 	cs = intel_ring_begin(rq, 12);
533 	if (IS_ERR(cs))
534 		return PTR_ERR(cs);
535 
536 	num_ccs_blks = DIV_ROUND_UP(GET_CCS_BYTES(i915, size),
537 				    NUM_CCS_BYTES_PER_BLOCK);
538 	GEM_BUG_ON(num_ccs_blks > NUM_CCS_BLKS_PER_XFER);
539 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
540 
541 	/*
542 	 * The XY_CTRL_SURF_COPY_BLT instruction is used to copy the CCS
543 	 * data in and out of the CCS region.
544 	 *
545 	 * We can copy at most 1024 blocks of 256 bytes using one
546 	 * XY_CTRL_SURF_COPY_BLT instruction.
547 	 *
548 	 * In case we need to copy more than 1024 blocks, we need to add
549 	 * another instruction to the same batch buffer.
550 	 *
551 	 * 1024 blocks of 256 bytes of CCS represent a total 256KB of CCS.
552 	 *
553 	 * 256 KB of CCS represents 256 * 256 KB = 64 MB of LMEM.
554 	 */
555 	*cs++ = XY_CTRL_SURF_COPY_BLT |
556 		src_access << SRC_ACCESS_TYPE_SHIFT |
557 		dst_access << DST_ACCESS_TYPE_SHIFT |
558 		((num_ccs_blks - 1) & CCS_SIZE_MASK) << CCS_SIZE_SHIFT;
559 	*cs++ = src_offset;
560 	*cs++ = rq->engine->instance |
561 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
562 	*cs++ = dst_offset;
563 	*cs++ = rq->engine->instance |
564 		FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, mocs);
565 
566 	cs = i915_flush_dw(cs, MI_FLUSH_DW_LLC | MI_FLUSH_DW_CCS);
567 	*cs++ = MI_NOOP;
568 
569 	intel_ring_advance(rq, cs);
570 
571 	return 0;
572 }
573 
574 static int emit_copy(struct i915_request *rq,
575 		     u32 dst_offset, u32 src_offset, int size)
576 {
577 	const int ver = GRAPHICS_VER(rq->engine->i915);
578 	u32 instance = rq->engine->instance;
579 	u32 *cs;
580 
581 	cs = intel_ring_begin(rq, ver >= 8 ? 10 : 6);
582 	if (IS_ERR(cs))
583 		return PTR_ERR(cs);
584 
585 	if (ver >= 9 && !wa_1209644611_applies(ver, size)) {
586 		*cs++ = GEN9_XY_FAST_COPY_BLT_CMD | (10 - 2);
587 		*cs++ = BLT_DEPTH_32 | PAGE_SIZE;
588 		*cs++ = 0;
589 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
590 		*cs++ = dst_offset;
591 		*cs++ = instance;
592 		*cs++ = 0;
593 		*cs++ = PAGE_SIZE;
594 		*cs++ = src_offset;
595 		*cs++ = instance;
596 	} else if (ver >= 8) {
597 		*cs++ = XY_SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (10 - 2);
598 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
599 		*cs++ = 0;
600 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
601 		*cs++ = dst_offset;
602 		*cs++ = instance;
603 		*cs++ = 0;
604 		*cs++ = PAGE_SIZE;
605 		*cs++ = src_offset;
606 		*cs++ = instance;
607 	} else {
608 		GEM_BUG_ON(instance);
609 		*cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
610 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | PAGE_SIZE;
611 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE;
612 		*cs++ = dst_offset;
613 		*cs++ = PAGE_SIZE;
614 		*cs++ = src_offset;
615 	}
616 
617 	intel_ring_advance(rq, cs);
618 	return 0;
619 }
620 
621 static u64 scatter_list_length(struct scatterlist *sg)
622 {
623 	u64 len = 0;
624 
625 	while (sg && sg_dma_len(sg)) {
626 		len += sg_dma_len(sg);
627 		sg = sg_next(sg);
628 	}
629 
630 	return len;
631 }
632 
633 static int
634 calculate_chunk_sz(struct drm_i915_private *i915, bool src_is_lmem,
635 		   u64 bytes_to_cpy, u64 ccs_bytes_to_cpy)
636 {
637 	if (ccs_bytes_to_cpy && !src_is_lmem)
638 		/*
639 		 * When CHUNK_SZ is passed all the pages upto CHUNK_SZ
640 		 * will be taken for the blt. in Flat-ccs supported
641 		 * platform Smem obj will have more pages than required
642 		 * for main meory hence limit it to the required size
643 		 * for main memory
644 		 */
645 		return min_t(u64, bytes_to_cpy, CHUNK_SZ);
646 	else
647 		return CHUNK_SZ;
648 }
649 
650 static void get_ccs_sg_sgt(struct sgt_dma *it, u64 bytes_to_cpy)
651 {
652 	u64 len;
653 
654 	do {
655 		GEM_BUG_ON(!it->sg || !sg_dma_len(it->sg));
656 		len = it->max - it->dma;
657 		if (len > bytes_to_cpy) {
658 			it->dma += bytes_to_cpy;
659 			break;
660 		}
661 
662 		bytes_to_cpy -= len;
663 
664 		it->sg = __sg_next(it->sg);
665 		it->dma = sg_dma_address(it->sg);
666 		it->max = it->dma + sg_dma_len(it->sg);
667 	} while (bytes_to_cpy);
668 }
669 
670 int
671 intel_context_migrate_copy(struct intel_context *ce,
672 			   const struct i915_deps *deps,
673 			   struct scatterlist *src,
674 			   enum i915_cache_level src_cache_level,
675 			   bool src_is_lmem,
676 			   struct scatterlist *dst,
677 			   enum i915_cache_level dst_cache_level,
678 			   bool dst_is_lmem,
679 			   struct i915_request **out)
680 {
681 	struct sgt_dma it_src = sg_sgt(src), it_dst = sg_sgt(dst), it_ccs;
682 	struct drm_i915_private *i915 = ce->engine->i915;
683 	u64 ccs_bytes_to_cpy = 0, bytes_to_cpy;
684 	enum i915_cache_level ccs_cache_level;
685 	u32 src_offset, dst_offset;
686 	u8 src_access, dst_access;
687 	struct i915_request *rq;
688 	u64 src_sz, dst_sz;
689 	bool ccs_is_src, overwrite_ccs;
690 	int err;
691 
692 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
693 	GEM_BUG_ON(IS_DGFX(ce->engine->i915) && (!src_is_lmem && !dst_is_lmem));
694 	*out = NULL;
695 
696 	GEM_BUG_ON(ce->ring->size < SZ_64K);
697 
698 	src_sz = scatter_list_length(src);
699 	bytes_to_cpy = src_sz;
700 
701 	if (HAS_FLAT_CCS(i915) && src_is_lmem ^ dst_is_lmem) {
702 		src_access = !src_is_lmem && dst_is_lmem;
703 		dst_access = !src_access;
704 
705 		dst_sz = scatter_list_length(dst);
706 		if (src_is_lmem) {
707 			it_ccs = it_dst;
708 			ccs_cache_level = dst_cache_level;
709 			ccs_is_src = false;
710 		} else if (dst_is_lmem) {
711 			bytes_to_cpy = dst_sz;
712 			it_ccs = it_src;
713 			ccs_cache_level = src_cache_level;
714 			ccs_is_src = true;
715 		}
716 
717 		/*
718 		 * When there is a eviction of ccs needed smem will have the
719 		 * extra pages for the ccs data
720 		 *
721 		 * TO-DO: Want to move the size mismatch check to a WARN_ON,
722 		 * but still we have some requests of smem->lmem with same size.
723 		 * Need to fix it.
724 		 */
725 		ccs_bytes_to_cpy = src_sz != dst_sz ? GET_CCS_BYTES(i915, bytes_to_cpy) : 0;
726 		if (ccs_bytes_to_cpy)
727 			get_ccs_sg_sgt(&it_ccs, bytes_to_cpy);
728 	}
729 
730 	overwrite_ccs = HAS_FLAT_CCS(i915) && !ccs_bytes_to_cpy && dst_is_lmem;
731 
732 	src_offset = 0;
733 	dst_offset = CHUNK_SZ;
734 	if (HAS_64K_PAGES(ce->engine->i915)) {
735 		src_offset = 0;
736 		dst_offset = 0;
737 		if (src_is_lmem)
738 			src_offset = CHUNK_SZ;
739 		if (dst_is_lmem)
740 			dst_offset = 2 * CHUNK_SZ;
741 	}
742 
743 	do {
744 		int len;
745 
746 		rq = i915_request_create(ce);
747 		if (IS_ERR(rq)) {
748 			err = PTR_ERR(rq);
749 			goto out_ce;
750 		}
751 
752 		if (deps) {
753 			err = i915_request_await_deps(rq, deps);
754 			if (err)
755 				goto out_rq;
756 
757 			if (rq->engine->emit_init_breadcrumb) {
758 				err = rq->engine->emit_init_breadcrumb(rq);
759 				if (err)
760 					goto out_rq;
761 			}
762 
763 			deps = NULL;
764 		}
765 
766 		/* The PTE updates + copy must not be interrupted. */
767 		err = emit_no_arbitration(rq);
768 		if (err)
769 			goto out_rq;
770 
771 		src_sz = calculate_chunk_sz(i915, src_is_lmem,
772 					    bytes_to_cpy, ccs_bytes_to_cpy);
773 
774 		len = emit_pte(rq, &it_src, src_cache_level, src_is_lmem,
775 			       src_offset, src_sz);
776 		if (!len) {
777 			err = -EINVAL;
778 			goto out_rq;
779 		}
780 		if (len < 0) {
781 			err = len;
782 			goto out_rq;
783 		}
784 
785 		err = emit_pte(rq, &it_dst, dst_cache_level, dst_is_lmem,
786 			       dst_offset, len);
787 		if (err < 0)
788 			goto out_rq;
789 		if (err < len) {
790 			err = -EINVAL;
791 			goto out_rq;
792 		}
793 
794 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
795 		if (err)
796 			goto out_rq;
797 
798 		err = emit_copy(rq, dst_offset,	src_offset, len);
799 		if (err)
800 			goto out_rq;
801 
802 		bytes_to_cpy -= len;
803 
804 		if (ccs_bytes_to_cpy) {
805 			int ccs_sz;
806 
807 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
808 			if (err)
809 				goto out_rq;
810 
811 			ccs_sz = GET_CCS_BYTES(i915, len);
812 			err = emit_pte(rq, &it_ccs, ccs_cache_level, false,
813 				       ccs_is_src ? src_offset : dst_offset,
814 				       ccs_sz);
815 			if (err < 0)
816 				goto out_rq;
817 			if (err < ccs_sz) {
818 				err = -EINVAL;
819 				goto out_rq;
820 			}
821 
822 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
823 			if (err)
824 				goto out_rq;
825 
826 			err = emit_copy_ccs(rq, dst_offset, dst_access,
827 					    src_offset, src_access, len);
828 			if (err)
829 				goto out_rq;
830 
831 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
832 			if (err)
833 				goto out_rq;
834 			ccs_bytes_to_cpy -= ccs_sz;
835 		} else if (overwrite_ccs) {
836 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
837 			if (err)
838 				goto out_rq;
839 
840 			if (src_is_lmem) {
841 				/*
842 				 * If the src is already in lmem, then we must
843 				 * be doing an lmem -> lmem transfer, and so
844 				 * should be safe to directly copy the CCS
845 				 * state. In this case we have either
846 				 * initialised the CCS aux state when first
847 				 * clearing the pages (since it is already
848 				 * allocated in lmem), or the user has
849 				 * potentially populated it, in which case we
850 				 * need to copy the CCS state as-is.
851 				 */
852 				err = emit_copy_ccs(rq,
853 						    dst_offset, INDIRECT_ACCESS,
854 						    src_offset, INDIRECT_ACCESS,
855 						    len);
856 			} else {
857 				/*
858 				 * While we can't always restore/manage the CCS
859 				 * state, we still need to ensure we don't leak
860 				 * the CCS state from the previous user, so make
861 				 * sure we overwrite it with something.
862 				 */
863 				err = emit_copy_ccs(rq,
864 						    dst_offset, INDIRECT_ACCESS,
865 						    dst_offset, DIRECT_ACCESS,
866 						    len);
867 			}
868 
869 			if (err)
870 				goto out_rq;
871 
872 			err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
873 			if (err)
874 				goto out_rq;
875 		}
876 
877 		/* Arbitration is re-enabled between requests. */
878 out_rq:
879 		if (*out)
880 			i915_request_put(*out);
881 		*out = i915_request_get(rq);
882 		i915_request_add(rq);
883 
884 		if (err)
885 			break;
886 
887 		if (!bytes_to_cpy && !ccs_bytes_to_cpy) {
888 			if (src_is_lmem)
889 				WARN_ON(it_src.sg && sg_dma_len(it_src.sg));
890 			else
891 				WARN_ON(it_dst.sg && sg_dma_len(it_dst.sg));
892 			break;
893 		}
894 
895 		if (WARN_ON(!it_src.sg || !sg_dma_len(it_src.sg) ||
896 			    !it_dst.sg || !sg_dma_len(it_dst.sg) ||
897 			    (ccs_bytes_to_cpy && (!it_ccs.sg ||
898 						  !sg_dma_len(it_ccs.sg))))) {
899 			err = -EINVAL;
900 			break;
901 		}
902 
903 		cond_resched();
904 	} while (1);
905 
906 out_ce:
907 	return err;
908 }
909 
910 static int emit_clear(struct i915_request *rq, u32 offset, int size,
911 		      u32 value, bool is_lmem)
912 {
913 	struct drm_i915_private *i915 = rq->engine->i915;
914 	int mocs = rq->engine->gt->mocs.uc_index << 1;
915 	const int ver = GRAPHICS_VER(i915);
916 	int ring_sz;
917 	u32 *cs;
918 
919 	GEM_BUG_ON(size >> PAGE_SHIFT > S16_MAX);
920 
921 	if (HAS_FLAT_CCS(i915) && ver >= 12)
922 		ring_sz = XY_FAST_COLOR_BLT_DW;
923 	else if (ver >= 8)
924 		ring_sz = 8;
925 	else
926 		ring_sz = 6;
927 
928 	cs = intel_ring_begin(rq, ring_sz);
929 	if (IS_ERR(cs))
930 		return PTR_ERR(cs);
931 
932 	if (HAS_FLAT_CCS(i915) && ver >= 12) {
933 		*cs++ = XY_FAST_COLOR_BLT_CMD | XY_FAST_COLOR_BLT_DEPTH_32 |
934 			(XY_FAST_COLOR_BLT_DW - 2);
935 		*cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, mocs) |
936 			(PAGE_SIZE - 1);
937 		*cs++ = 0;
938 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
939 		*cs++ = offset;
940 		*cs++ = rq->engine->instance;
941 		*cs++ = !is_lmem << XY_FAST_COLOR_BLT_MEM_TYPE_SHIFT;
942 		/* BG7 */
943 		*cs++ = value;
944 		*cs++ = 0;
945 		*cs++ = 0;
946 		*cs++ = 0;
947 		/* BG11 */
948 		*cs++ = 0;
949 		*cs++ = 0;
950 		/* BG13 */
951 		*cs++ = 0;
952 		*cs++ = 0;
953 		*cs++ = 0;
954 	} else if (ver >= 8) {
955 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (7 - 2);
956 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
957 		*cs++ = 0;
958 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
959 		*cs++ = offset;
960 		*cs++ = rq->engine->instance;
961 		*cs++ = value;
962 		*cs++ = MI_NOOP;
963 	} else {
964 		*cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
965 		*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
966 		*cs++ = 0;
967 		*cs++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
968 		*cs++ = offset;
969 		*cs++ = value;
970 	}
971 
972 	intel_ring_advance(rq, cs);
973 	return 0;
974 }
975 
976 int
977 intel_context_migrate_clear(struct intel_context *ce,
978 			    const struct i915_deps *deps,
979 			    struct scatterlist *sg,
980 			    enum i915_cache_level cache_level,
981 			    bool is_lmem,
982 			    u32 value,
983 			    struct i915_request **out)
984 {
985 	struct drm_i915_private *i915 = ce->engine->i915;
986 	struct sgt_dma it = sg_sgt(sg);
987 	struct i915_request *rq;
988 	u32 offset;
989 	int err;
990 
991 	GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
992 	*out = NULL;
993 
994 	GEM_BUG_ON(ce->ring->size < SZ_64K);
995 
996 	offset = 0;
997 	if (HAS_64K_PAGES(i915) && is_lmem)
998 		offset = CHUNK_SZ;
999 
1000 	do {
1001 		int len;
1002 
1003 		rq = i915_request_create(ce);
1004 		if (IS_ERR(rq)) {
1005 			err = PTR_ERR(rq);
1006 			goto out_ce;
1007 		}
1008 
1009 		if (deps) {
1010 			err = i915_request_await_deps(rq, deps);
1011 			if (err)
1012 				goto out_rq;
1013 
1014 			if (rq->engine->emit_init_breadcrumb) {
1015 				err = rq->engine->emit_init_breadcrumb(rq);
1016 				if (err)
1017 					goto out_rq;
1018 			}
1019 
1020 			deps = NULL;
1021 		}
1022 
1023 		/* The PTE updates + clear must not be interrupted. */
1024 		err = emit_no_arbitration(rq);
1025 		if (err)
1026 			goto out_rq;
1027 
1028 		len = emit_pte(rq, &it, cache_level, is_lmem, offset, CHUNK_SZ);
1029 		if (len <= 0) {
1030 			err = len;
1031 			goto out_rq;
1032 		}
1033 
1034 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1035 		if (err)
1036 			goto out_rq;
1037 
1038 		err = emit_clear(rq, offset, len, value, is_lmem);
1039 		if (err)
1040 			goto out_rq;
1041 
1042 		if (HAS_FLAT_CCS(i915) && is_lmem && !value) {
1043 			/*
1044 			 * copy the content of memory into corresponding
1045 			 * ccs surface
1046 			 */
1047 			err = emit_copy_ccs(rq, offset, INDIRECT_ACCESS, offset,
1048 					    DIRECT_ACCESS, len);
1049 			if (err)
1050 				goto out_rq;
1051 		}
1052 
1053 		err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1054 
1055 		/* Arbitration is re-enabled between requests. */
1056 out_rq:
1057 		if (*out)
1058 			i915_request_put(*out);
1059 		*out = i915_request_get(rq);
1060 		i915_request_add(rq);
1061 		if (err || !it.sg || !sg_dma_len(it.sg))
1062 			break;
1063 
1064 		cond_resched();
1065 	} while (1);
1066 
1067 out_ce:
1068 	return err;
1069 }
1070 
1071 int intel_migrate_copy(struct intel_migrate *m,
1072 		       struct i915_gem_ww_ctx *ww,
1073 		       const struct i915_deps *deps,
1074 		       struct scatterlist *src,
1075 		       enum i915_cache_level src_cache_level,
1076 		       bool src_is_lmem,
1077 		       struct scatterlist *dst,
1078 		       enum i915_cache_level dst_cache_level,
1079 		       bool dst_is_lmem,
1080 		       struct i915_request **out)
1081 {
1082 	struct intel_context *ce;
1083 	int err;
1084 
1085 	*out = NULL;
1086 	if (!m->context)
1087 		return -ENODEV;
1088 
1089 	ce = intel_migrate_create_context(m);
1090 	if (IS_ERR(ce))
1091 		ce = intel_context_get(m->context);
1092 	GEM_BUG_ON(IS_ERR(ce));
1093 
1094 	err = intel_context_pin_ww(ce, ww);
1095 	if (err)
1096 		goto out;
1097 
1098 	err = intel_context_migrate_copy(ce, deps,
1099 					 src, src_cache_level, src_is_lmem,
1100 					 dst, dst_cache_level, dst_is_lmem,
1101 					 out);
1102 
1103 	intel_context_unpin(ce);
1104 out:
1105 	intel_context_put(ce);
1106 	return err;
1107 }
1108 
1109 int
1110 intel_migrate_clear(struct intel_migrate *m,
1111 		    struct i915_gem_ww_ctx *ww,
1112 		    const struct i915_deps *deps,
1113 		    struct scatterlist *sg,
1114 		    enum i915_cache_level cache_level,
1115 		    bool is_lmem,
1116 		    u32 value,
1117 		    struct i915_request **out)
1118 {
1119 	struct intel_context *ce;
1120 	int err;
1121 
1122 	*out = NULL;
1123 	if (!m->context)
1124 		return -ENODEV;
1125 
1126 	ce = intel_migrate_create_context(m);
1127 	if (IS_ERR(ce))
1128 		ce = intel_context_get(m->context);
1129 	GEM_BUG_ON(IS_ERR(ce));
1130 
1131 	err = intel_context_pin_ww(ce, ww);
1132 	if (err)
1133 		goto out;
1134 
1135 	err = intel_context_migrate_clear(ce, deps, sg, cache_level,
1136 					  is_lmem, value, out);
1137 
1138 	intel_context_unpin(ce);
1139 out:
1140 	intel_context_put(ce);
1141 	return err;
1142 }
1143 
1144 void intel_migrate_fini(struct intel_migrate *m)
1145 {
1146 	struct intel_context *ce;
1147 
1148 	ce = fetch_and_zero(&m->context);
1149 	if (!ce)
1150 		return;
1151 
1152 	intel_engine_destroy_pinned_context(ce);
1153 }
1154 
1155 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1156 #include "selftest_migrate.c"
1157 #endif
1158