xref: /linux/drivers/gpu/drm/xe/xe_migrate.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include "xe_migrate.h"
7 
8 #include <linux/bitfield.h>
9 #include <linux/sizes.h>
10 
11 #include <drm/drm_managed.h>
12 #include <drm/ttm/ttm_tt.h>
13 #include <uapi/drm/xe_drm.h>
14 
15 #include <generated/xe_wa_oob.h>
16 
17 #include "instructions/xe_gpu_commands.h"
18 #include "instructions/xe_mi_commands.h"
19 #include "regs/xe_gtt_defs.h"
20 #include "tests/xe_test.h"
21 #include "xe_assert.h"
22 #include "xe_bb.h"
23 #include "xe_bo.h"
24 #include "xe_exec_queue.h"
25 #include "xe_ggtt.h"
26 #include "xe_gt.h"
27 #include "xe_hw_engine.h"
28 #include "xe_lrc.h"
29 #include "xe_map.h"
30 #include "xe_mocs.h"
31 #include "xe_pt.h"
32 #include "xe_res_cursor.h"
33 #include "xe_sched_job.h"
34 #include "xe_sync.h"
35 #include "xe_trace_bo.h"
36 #include "xe_vm.h"
37 
38 /**
39  * struct xe_migrate - migrate context.
40  */
41 struct xe_migrate {
42 	/** @q: Default exec queue used for migration */
43 	struct xe_exec_queue *q;
44 	/** @tile: Backpointer to the tile this struct xe_migrate belongs to. */
45 	struct xe_tile *tile;
46 	/** @job_mutex: Timeline mutex for @eng. */
47 	struct mutex job_mutex;
48 	/** @pt_bo: Page-table buffer object. */
49 	struct xe_bo *pt_bo;
50 	/** @batch_base_ofs: VM offset of the migration batch buffer */
51 	u64 batch_base_ofs;
52 	/** @usm_batch_base_ofs: VM offset of the usm batch buffer */
53 	u64 usm_batch_base_ofs;
54 	/** @cleared_mem_ofs: VM offset of @cleared_bo. */
55 	u64 cleared_mem_ofs;
56 	/**
57 	 * @fence: dma-fence representing the last migration job batch.
58 	 * Protected by @job_mutex.
59 	 */
60 	struct dma_fence *fence;
61 	/**
62 	 * @vm_update_sa: For integrated, used to suballocate page-tables
63 	 * out of the pt_bo.
64 	 */
65 	struct drm_suballoc_manager vm_update_sa;
66 	/** @min_chunk_size: For dgfx, Minimum chunk size */
67 	u64 min_chunk_size;
68 };
69 
70 #define MAX_PREEMPTDISABLE_TRANSFER SZ_8M /* Around 1ms. */
71 #define MAX_CCS_LIMITED_TRANSFER SZ_4M /* XE_PAGE_SIZE * (FIELD_MAX(XE2_CCS_SIZE_MASK) + 1) */
72 #define NUM_KERNEL_PDE 15
73 #define NUM_PT_SLOTS 32
74 #define LEVEL0_PAGE_TABLE_ENCODE_SIZE SZ_2M
75 #define MAX_NUM_PTE 512
76 #define IDENTITY_OFFSET 256ULL
77 
78 /*
79  * Although MI_STORE_DATA_IMM's "length" field is 10-bits, 0x3FE is the largest
80  * legal value accepted.  Since that instruction field is always stored in
81  * (val-2) format, this translates to 0x400 dwords for the true maximum length
82  * of the instruction.  Subtracting the instruction header (1 dword) and
83  * address (2 dwords), that leaves 0x3FD dwords (0x1FE qwords) for PTE values.
84  */
85 #define MAX_PTE_PER_SDI 0x1FE
86 
87 /**
88  * xe_tile_migrate_exec_queue() - Get this tile's migrate exec queue.
89  * @tile: The tile.
90  *
91  * Returns the default migrate exec queue of this tile.
92  *
93  * Return: The default migrate exec queue
94  */
95 struct xe_exec_queue *xe_tile_migrate_exec_queue(struct xe_tile *tile)
96 {
97 	return tile->migrate->q;
98 }
99 
100 static void xe_migrate_fini(struct drm_device *dev, void *arg)
101 {
102 	struct xe_migrate *m = arg;
103 
104 	xe_vm_lock(m->q->vm, false);
105 	xe_bo_unpin(m->pt_bo);
106 	xe_vm_unlock(m->q->vm);
107 
108 	dma_fence_put(m->fence);
109 	xe_bo_put(m->pt_bo);
110 	drm_suballoc_manager_fini(&m->vm_update_sa);
111 	mutex_destroy(&m->job_mutex);
112 	xe_vm_close_and_put(m->q->vm);
113 	xe_exec_queue_put(m->q);
114 }
115 
116 static u64 xe_migrate_vm_addr(u64 slot, u32 level)
117 {
118 	XE_WARN_ON(slot >= NUM_PT_SLOTS);
119 
120 	/* First slot is reserved for mapping of PT bo and bb, start from 1 */
121 	return (slot + 1ULL) << xe_pt_shift(level + 1);
122 }
123 
124 static u64 xe_migrate_vram_ofs(struct xe_device *xe, u64 addr, bool is_comp_pte)
125 {
126 	/*
127 	 * Remove the DPA to get a correct offset into identity table for the
128 	 * migrate offset
129 	 */
130 	u64 identity_offset = IDENTITY_OFFSET;
131 
132 	if (GRAPHICS_VER(xe) >= 20 && is_comp_pte)
133 		identity_offset += DIV_ROUND_UP_ULL(xe->mem.vram.actual_physical_size, SZ_1G);
134 
135 	addr -= xe->mem.vram.dpa_base;
136 	return addr + (identity_offset << xe_pt_shift(2));
137 }
138 
139 static void xe_migrate_program_identity(struct xe_device *xe, struct xe_vm *vm, struct xe_bo *bo,
140 					u64 map_ofs, u64 vram_offset, u16 pat_index, u64 pt_2m_ofs)
141 {
142 	u64 pos, ofs, flags;
143 	u64 entry;
144 	/* XXX: Unclear if this should be usable_size? */
145 	u64 vram_limit =  xe->mem.vram.actual_physical_size +
146 		xe->mem.vram.dpa_base;
147 	u32 level = 2;
148 
149 	ofs = map_ofs + XE_PAGE_SIZE * level + vram_offset * 8;
150 	flags = vm->pt_ops->pte_encode_addr(xe, 0, pat_index, level,
151 					    true, 0);
152 
153 	xe_assert(xe, IS_ALIGNED(xe->mem.vram.usable_size, SZ_2M));
154 
155 	/*
156 	 * Use 1GB pages when possible, last chunk always use 2M
157 	 * pages as mixing reserved memory (stolen, WOCPM) with a single
158 	 * mapping is not allowed on certain platforms.
159 	 */
160 	for (pos = xe->mem.vram.dpa_base; pos < vram_limit;
161 	     pos += SZ_1G, ofs += 8) {
162 		if (pos + SZ_1G >= vram_limit) {
163 			entry = vm->pt_ops->pde_encode_bo(bo, pt_2m_ofs,
164 							  pat_index);
165 			xe_map_wr(xe, &bo->vmap, ofs, u64, entry);
166 
167 			flags = vm->pt_ops->pte_encode_addr(xe, 0,
168 							    pat_index,
169 							    level - 1,
170 							    true, 0);
171 
172 			for (ofs = pt_2m_ofs; pos < vram_limit;
173 			     pos += SZ_2M, ofs += 8)
174 				xe_map_wr(xe, &bo->vmap, ofs, u64, pos | flags);
175 			break;	/* Ensure pos == vram_limit assert correct */
176 		}
177 
178 		xe_map_wr(xe, &bo->vmap, ofs, u64, pos | flags);
179 	}
180 
181 	xe_assert(xe, pos == vram_limit);
182 }
183 
184 static int xe_migrate_prepare_vm(struct xe_tile *tile, struct xe_migrate *m,
185 				 struct xe_vm *vm)
186 {
187 	struct xe_device *xe = tile_to_xe(tile);
188 	u16 pat_index = xe->pat.idx[XE_CACHE_WB];
189 	u8 id = tile->id;
190 	u32 num_entries = NUM_PT_SLOTS, num_level = vm->pt_root[id]->level;
191 #define VRAM_IDENTITY_MAP_COUNT	2
192 	u32 num_setup = num_level + VRAM_IDENTITY_MAP_COUNT;
193 #undef VRAM_IDENTITY_MAP_COUNT
194 	u32 map_ofs, level, i;
195 	struct xe_bo *bo, *batch = tile->mem.kernel_bb_pool->bo;
196 	u64 entry, pt29_ofs;
197 
198 	/* Can't bump NUM_PT_SLOTS too high */
199 	BUILD_BUG_ON(NUM_PT_SLOTS > SZ_2M/XE_PAGE_SIZE);
200 	/* Must be a multiple of 64K to support all platforms */
201 	BUILD_BUG_ON(NUM_PT_SLOTS * XE_PAGE_SIZE % SZ_64K);
202 	/* And one slot reserved for the 4KiB page table updates */
203 	BUILD_BUG_ON(!(NUM_KERNEL_PDE & 1));
204 
205 	/* Need to be sure everything fits in the first PT, or create more */
206 	xe_tile_assert(tile, m->batch_base_ofs + batch->size < SZ_2M);
207 
208 	bo = xe_bo_create_pin_map(vm->xe, tile, vm,
209 				  num_entries * XE_PAGE_SIZE,
210 				  ttm_bo_type_kernel,
211 				  XE_BO_FLAG_VRAM_IF_DGFX(tile) |
212 				  XE_BO_FLAG_PINNED |
213 				  XE_BO_FLAG_PAGETABLE);
214 	if (IS_ERR(bo))
215 		return PTR_ERR(bo);
216 
217 	/* PT30 & PT31 reserved for 2M identity map */
218 	pt29_ofs = bo->size - 3 * XE_PAGE_SIZE;
219 	entry = vm->pt_ops->pde_encode_bo(bo, pt29_ofs, pat_index);
220 	xe_pt_write(xe, &vm->pt_root[id]->bo->vmap, 0, entry);
221 
222 	map_ofs = (num_entries - num_setup) * XE_PAGE_SIZE;
223 
224 	/* Map the entire BO in our level 0 pt */
225 	for (i = 0, level = 0; i < num_entries; level++) {
226 		entry = vm->pt_ops->pte_encode_bo(bo, i * XE_PAGE_SIZE,
227 						  pat_index, 0);
228 
229 		xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64, entry);
230 
231 		if (vm->flags & XE_VM_FLAG_64K)
232 			i += 16;
233 		else
234 			i += 1;
235 	}
236 
237 	if (!IS_DGFX(xe)) {
238 		/* Write out batch too */
239 		m->batch_base_ofs = NUM_PT_SLOTS * XE_PAGE_SIZE;
240 		for (i = 0; i < batch->size;
241 		     i += vm->flags & XE_VM_FLAG_64K ? XE_64K_PAGE_SIZE :
242 		     XE_PAGE_SIZE) {
243 			entry = vm->pt_ops->pte_encode_bo(batch, i,
244 							  pat_index, 0);
245 
246 			xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64,
247 				  entry);
248 			level++;
249 		}
250 		if (xe->info.has_usm) {
251 			xe_tile_assert(tile, batch->size == SZ_1M);
252 
253 			batch = tile->primary_gt->usm.bb_pool->bo;
254 			m->usm_batch_base_ofs = m->batch_base_ofs + SZ_1M;
255 			xe_tile_assert(tile, batch->size == SZ_512K);
256 
257 			for (i = 0; i < batch->size;
258 			     i += vm->flags & XE_VM_FLAG_64K ? XE_64K_PAGE_SIZE :
259 			     XE_PAGE_SIZE) {
260 				entry = vm->pt_ops->pte_encode_bo(batch, i,
261 								  pat_index, 0);
262 
263 				xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64,
264 					  entry);
265 				level++;
266 			}
267 		}
268 	} else {
269 		u64 batch_addr = xe_bo_addr(batch, 0, XE_PAGE_SIZE);
270 
271 		m->batch_base_ofs = xe_migrate_vram_ofs(xe, batch_addr, false);
272 
273 		if (xe->info.has_usm) {
274 			batch = tile->primary_gt->usm.bb_pool->bo;
275 			batch_addr = xe_bo_addr(batch, 0, XE_PAGE_SIZE);
276 			m->usm_batch_base_ofs = xe_migrate_vram_ofs(xe, batch_addr, false);
277 		}
278 	}
279 
280 	for (level = 1; level < num_level; level++) {
281 		u32 flags = 0;
282 
283 		if (vm->flags & XE_VM_FLAG_64K && level == 1)
284 			flags = XE_PDE_64K;
285 
286 		entry = vm->pt_ops->pde_encode_bo(bo, map_ofs + (u64)(level - 1) *
287 						  XE_PAGE_SIZE, pat_index);
288 		xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE * level, u64,
289 			  entry | flags);
290 	}
291 
292 	/* Write PDE's that point to our BO. */
293 	for (i = 0; i < map_ofs / PAGE_SIZE; i++) {
294 		entry = vm->pt_ops->pde_encode_bo(bo, (u64)i * XE_PAGE_SIZE,
295 						  pat_index);
296 
297 		xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE +
298 			  (i + 1) * 8, u64, entry);
299 	}
300 
301 	/* Set up a 1GiB NULL mapping at 255GiB offset. */
302 	level = 2;
303 	xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE * level + 255 * 8, u64,
304 		  vm->pt_ops->pte_encode_addr(xe, 0, pat_index, level, IS_DGFX(xe), 0)
305 		  | XE_PTE_NULL);
306 	m->cleared_mem_ofs = (255ULL << xe_pt_shift(level));
307 
308 	/* Identity map the entire vram at 256GiB offset */
309 	if (IS_DGFX(xe)) {
310 		u64 pt30_ofs = bo->size - 2 * XE_PAGE_SIZE;
311 
312 		xe_migrate_program_identity(xe, vm, bo, map_ofs, IDENTITY_OFFSET,
313 					    pat_index, pt30_ofs);
314 		xe_assert(xe, xe->mem.vram.actual_physical_size <=
315 					(MAX_NUM_PTE - IDENTITY_OFFSET) * SZ_1G);
316 
317 		/*
318 		 * Identity map the entire vram for compressed pat_index for xe2+
319 		 * if flat ccs is enabled.
320 		 */
321 		if (GRAPHICS_VER(xe) >= 20 && xe_device_has_flat_ccs(xe)) {
322 			u16 comp_pat_index = xe->pat.idx[XE_CACHE_NONE_COMPRESSION];
323 			u64 vram_offset = IDENTITY_OFFSET +
324 				DIV_ROUND_UP_ULL(xe->mem.vram.actual_physical_size, SZ_1G);
325 			u64 pt31_ofs = bo->size - XE_PAGE_SIZE;
326 
327 			xe_assert(xe, xe->mem.vram.actual_physical_size <= (MAX_NUM_PTE -
328 						IDENTITY_OFFSET - IDENTITY_OFFSET / 2) * SZ_1G);
329 			xe_migrate_program_identity(xe, vm, bo, map_ofs, vram_offset,
330 						    comp_pat_index, pt31_ofs);
331 		}
332 	}
333 
334 	/*
335 	 * Example layout created above, with root level = 3:
336 	 * [PT0...PT7]: kernel PT's for copy/clear; 64 or 4KiB PTE's
337 	 * [PT8]: Kernel PT for VM_BIND, 4 KiB PTE's
338 	 * [PT9...PT26]: Userspace PT's for VM_BIND, 4 KiB PTE's
339 	 * [PT27 = PDE 0] [PT28 = PDE 1] [PT29 = PDE 2] [PT30 & PT31 = 2M vram identity map]
340 	 *
341 	 * This makes the lowest part of the VM point to the pagetables.
342 	 * Hence the lowest 2M in the vm should point to itself, with a few writes
343 	 * and flushes, other parts of the VM can be used either for copying and
344 	 * clearing.
345 	 *
346 	 * For performance, the kernel reserves PDE's, so about 20 are left
347 	 * for async VM updates.
348 	 *
349 	 * To make it easier to work, each scratch PT is put in slot (1 + PT #)
350 	 * everywhere, this allows lockless updates to scratch pages by using
351 	 * the different addresses in VM.
352 	 */
353 #define NUM_VMUSA_UNIT_PER_PAGE	32
354 #define VM_SA_UPDATE_UNIT_SIZE		(XE_PAGE_SIZE / NUM_VMUSA_UNIT_PER_PAGE)
355 #define NUM_VMUSA_WRITES_PER_UNIT	(VM_SA_UPDATE_UNIT_SIZE / sizeof(u64))
356 	drm_suballoc_manager_init(&m->vm_update_sa,
357 				  (size_t)(map_ofs / XE_PAGE_SIZE - NUM_KERNEL_PDE) *
358 				  NUM_VMUSA_UNIT_PER_PAGE, 0);
359 
360 	m->pt_bo = bo;
361 	return 0;
362 }
363 
364 /*
365  * Including the reserved copy engine is required to avoid deadlocks due to
366  * migrate jobs servicing the faults gets stuck behind the job that faulted.
367  */
368 static u32 xe_migrate_usm_logical_mask(struct xe_gt *gt)
369 {
370 	u32 logical_mask = 0;
371 	struct xe_hw_engine *hwe;
372 	enum xe_hw_engine_id id;
373 
374 	for_each_hw_engine(hwe, gt, id) {
375 		if (hwe->class != XE_ENGINE_CLASS_COPY)
376 			continue;
377 
378 		if (xe_gt_is_usm_hwe(gt, hwe))
379 			logical_mask |= BIT(hwe->logical_instance);
380 	}
381 
382 	return logical_mask;
383 }
384 
385 static bool xe_migrate_needs_ccs_emit(struct xe_device *xe)
386 {
387 	return xe_device_has_flat_ccs(xe) && !(GRAPHICS_VER(xe) >= 20 && IS_DGFX(xe));
388 }
389 
390 /**
391  * xe_migrate_init() - Initialize a migrate context
392  * @tile: Back-pointer to the tile we're initializing for.
393  *
394  * Return: Pointer to a migrate context on success. Error pointer on error.
395  */
396 struct xe_migrate *xe_migrate_init(struct xe_tile *tile)
397 {
398 	struct xe_device *xe = tile_to_xe(tile);
399 	struct xe_gt *primary_gt = tile->primary_gt;
400 	struct xe_migrate *m;
401 	struct xe_vm *vm;
402 	int err;
403 
404 	m = drmm_kzalloc(&xe->drm, sizeof(*m), GFP_KERNEL);
405 	if (!m)
406 		return ERR_PTR(-ENOMEM);
407 
408 	m->tile = tile;
409 
410 	/* Special layout, prepared below.. */
411 	vm = xe_vm_create(xe, XE_VM_FLAG_MIGRATION |
412 			  XE_VM_FLAG_SET_TILE_ID(tile));
413 	if (IS_ERR(vm))
414 		return ERR_CAST(vm);
415 
416 	xe_vm_lock(vm, false);
417 	err = xe_migrate_prepare_vm(tile, m, vm);
418 	xe_vm_unlock(vm);
419 	if (err) {
420 		xe_vm_close_and_put(vm);
421 		return ERR_PTR(err);
422 	}
423 
424 	if (xe->info.has_usm) {
425 		struct xe_hw_engine *hwe = xe_gt_hw_engine(primary_gt,
426 							   XE_ENGINE_CLASS_COPY,
427 							   primary_gt->usm.reserved_bcs_instance,
428 							   false);
429 		u32 logical_mask = xe_migrate_usm_logical_mask(primary_gt);
430 
431 		if (!hwe || !logical_mask)
432 			return ERR_PTR(-EINVAL);
433 
434 		/*
435 		 * XXX: Currently only reserving 1 (likely slow) BCS instance on
436 		 * PVC, may want to revisit if performance is needed.
437 		 */
438 		m->q = xe_exec_queue_create(xe, vm, logical_mask, 1, hwe,
439 					    EXEC_QUEUE_FLAG_KERNEL |
440 					    EXEC_QUEUE_FLAG_PERMANENT |
441 					    EXEC_QUEUE_FLAG_HIGH_PRIORITY, 0);
442 	} else {
443 		m->q = xe_exec_queue_create_class(xe, primary_gt, vm,
444 						  XE_ENGINE_CLASS_COPY,
445 						  EXEC_QUEUE_FLAG_KERNEL |
446 						  EXEC_QUEUE_FLAG_PERMANENT, 0);
447 	}
448 	if (IS_ERR(m->q)) {
449 		xe_vm_close_and_put(vm);
450 		return ERR_CAST(m->q);
451 	}
452 
453 	mutex_init(&m->job_mutex);
454 	fs_reclaim_acquire(GFP_KERNEL);
455 	might_lock(&m->job_mutex);
456 	fs_reclaim_release(GFP_KERNEL);
457 
458 	err = drmm_add_action_or_reset(&xe->drm, xe_migrate_fini, m);
459 	if (err)
460 		return ERR_PTR(err);
461 
462 	if (IS_DGFX(xe)) {
463 		if (xe_migrate_needs_ccs_emit(xe))
464 			/* min chunk size corresponds to 4K of CCS Metadata */
465 			m->min_chunk_size = SZ_4K * SZ_64K /
466 				xe_device_ccs_bytes(xe, SZ_64K);
467 		else
468 			/* Somewhat arbitrary to avoid a huge amount of blits */
469 			m->min_chunk_size = SZ_64K;
470 		m->min_chunk_size = roundup_pow_of_two(m->min_chunk_size);
471 		drm_dbg(&xe->drm, "Migrate min chunk size is 0x%08llx\n",
472 			(unsigned long long)m->min_chunk_size);
473 	}
474 
475 	return m;
476 }
477 
478 static u64 max_mem_transfer_per_pass(struct xe_device *xe)
479 {
480 	if (!IS_DGFX(xe) && xe_device_has_flat_ccs(xe))
481 		return MAX_CCS_LIMITED_TRANSFER;
482 
483 	return MAX_PREEMPTDISABLE_TRANSFER;
484 }
485 
486 static u64 xe_migrate_res_sizes(struct xe_migrate *m, struct xe_res_cursor *cur)
487 {
488 	struct xe_device *xe = tile_to_xe(m->tile);
489 	u64 size = min_t(u64, max_mem_transfer_per_pass(xe), cur->remaining);
490 
491 	if (mem_type_is_vram(cur->mem_type)) {
492 		/*
493 		 * VRAM we want to blit in chunks with sizes aligned to
494 		 * min_chunk_size in order for the offset to CCS metadata to be
495 		 * page-aligned. If it's the last chunk it may be smaller.
496 		 *
497 		 * Another constraint is that we need to limit the blit to
498 		 * the VRAM block size, unless size is smaller than
499 		 * min_chunk_size.
500 		 */
501 		u64 chunk = max_t(u64, cur->size, m->min_chunk_size);
502 
503 		size = min_t(u64, size, chunk);
504 		if (size > m->min_chunk_size)
505 			size = round_down(size, m->min_chunk_size);
506 	}
507 
508 	return size;
509 }
510 
511 static bool xe_migrate_allow_identity(u64 size, const struct xe_res_cursor *cur)
512 {
513 	/* If the chunk is not fragmented, allow identity map. */
514 	return cur->size >= size;
515 }
516 
517 #define PTE_UPDATE_FLAG_IS_VRAM		BIT(0)
518 #define PTE_UPDATE_FLAG_IS_COMP_PTE	BIT(1)
519 
520 static u32 pte_update_size(struct xe_migrate *m,
521 			   u32 flags,
522 			   struct ttm_resource *res,
523 			   struct xe_res_cursor *cur,
524 			   u64 *L0, u64 *L0_ofs, u32 *L0_pt,
525 			   u32 cmd_size, u32 pt_ofs, u32 avail_pts)
526 {
527 	u32 cmds = 0;
528 	bool is_vram = PTE_UPDATE_FLAG_IS_VRAM & flags;
529 	bool is_comp_pte = PTE_UPDATE_FLAG_IS_COMP_PTE & flags;
530 
531 	*L0_pt = pt_ofs;
532 	if (is_vram && xe_migrate_allow_identity(*L0, cur)) {
533 		/* Offset into identity map. */
534 		*L0_ofs = xe_migrate_vram_ofs(tile_to_xe(m->tile),
535 					      cur->start + vram_region_gpu_offset(res),
536 					      is_comp_pte);
537 		cmds += cmd_size;
538 	} else {
539 		/* Clip L0 to available size */
540 		u64 size = min(*L0, (u64)avail_pts * SZ_2M);
541 		u32 num_4k_pages = (size + XE_PAGE_SIZE - 1) >> XE_PTE_SHIFT;
542 
543 		*L0 = size;
544 		*L0_ofs = xe_migrate_vm_addr(pt_ofs, 0);
545 
546 		/* MI_STORE_DATA_IMM */
547 		cmds += 3 * DIV_ROUND_UP(num_4k_pages, MAX_PTE_PER_SDI);
548 
549 		/* PDE qwords */
550 		cmds += num_4k_pages * 2;
551 
552 		/* Each chunk has a single blit command */
553 		cmds += cmd_size;
554 	}
555 
556 	return cmds;
557 }
558 
559 static void emit_pte(struct xe_migrate *m,
560 		     struct xe_bb *bb, u32 at_pt,
561 		     bool is_vram, bool is_comp_pte,
562 		     struct xe_res_cursor *cur,
563 		     u32 size, struct ttm_resource *res)
564 {
565 	struct xe_device *xe = tile_to_xe(m->tile);
566 	struct xe_vm *vm = m->q->vm;
567 	u16 pat_index;
568 	u32 ptes;
569 	u64 ofs = (u64)at_pt * XE_PAGE_SIZE;
570 	u64 cur_ofs;
571 
572 	/* Indirect access needs compression enabled uncached PAT index */
573 	if (GRAPHICS_VERx100(xe) >= 2000)
574 		pat_index = is_comp_pte ? xe->pat.idx[XE_CACHE_NONE_COMPRESSION] :
575 					  xe->pat.idx[XE_CACHE_WB];
576 	else
577 		pat_index = xe->pat.idx[XE_CACHE_WB];
578 
579 	ptes = DIV_ROUND_UP(size, XE_PAGE_SIZE);
580 
581 	while (ptes) {
582 		u32 chunk = min(MAX_PTE_PER_SDI, ptes);
583 
584 		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(chunk);
585 		bb->cs[bb->len++] = ofs;
586 		bb->cs[bb->len++] = 0;
587 
588 		cur_ofs = ofs;
589 		ofs += chunk * 8;
590 		ptes -= chunk;
591 
592 		while (chunk--) {
593 			u64 addr, flags = 0;
594 			bool devmem = false;
595 
596 			addr = xe_res_dma(cur) & PAGE_MASK;
597 			if (is_vram) {
598 				if (vm->flags & XE_VM_FLAG_64K) {
599 					u64 va = cur_ofs * XE_PAGE_SIZE / 8;
600 
601 					xe_assert(xe, (va & (SZ_64K - 1)) ==
602 						  (addr & (SZ_64K - 1)));
603 
604 					flags |= XE_PTE_PS64;
605 				}
606 
607 				addr += vram_region_gpu_offset(res);
608 				devmem = true;
609 			}
610 
611 			addr = vm->pt_ops->pte_encode_addr(m->tile->xe,
612 							   addr, pat_index,
613 							   0, devmem, flags);
614 			bb->cs[bb->len++] = lower_32_bits(addr);
615 			bb->cs[bb->len++] = upper_32_bits(addr);
616 
617 			xe_res_next(cur, min_t(u32, size, PAGE_SIZE));
618 			cur_ofs += 8;
619 		}
620 	}
621 }
622 
623 #define EMIT_COPY_CCS_DW 5
624 static void emit_copy_ccs(struct xe_gt *gt, struct xe_bb *bb,
625 			  u64 dst_ofs, bool dst_is_indirect,
626 			  u64 src_ofs, bool src_is_indirect,
627 			  u32 size)
628 {
629 	struct xe_device *xe = gt_to_xe(gt);
630 	u32 *cs = bb->cs + bb->len;
631 	u32 num_ccs_blks;
632 	u32 num_pages;
633 	u32 ccs_copy_size;
634 	u32 mocs;
635 
636 	if (GRAPHICS_VERx100(xe) >= 2000) {
637 		num_pages = DIV_ROUND_UP(size, XE_PAGE_SIZE);
638 		xe_gt_assert(gt, FIELD_FIT(XE2_CCS_SIZE_MASK, num_pages - 1));
639 
640 		ccs_copy_size = REG_FIELD_PREP(XE2_CCS_SIZE_MASK, num_pages - 1);
641 		mocs = FIELD_PREP(XE2_XY_CTRL_SURF_MOCS_INDEX_MASK, gt->mocs.uc_index);
642 
643 	} else {
644 		num_ccs_blks = DIV_ROUND_UP(xe_device_ccs_bytes(gt_to_xe(gt), size),
645 					    NUM_CCS_BYTES_PER_BLOCK);
646 		xe_gt_assert(gt, FIELD_FIT(CCS_SIZE_MASK, num_ccs_blks - 1));
647 
648 		ccs_copy_size = REG_FIELD_PREP(CCS_SIZE_MASK, num_ccs_blks - 1);
649 		mocs = FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, gt->mocs.uc_index);
650 	}
651 
652 	*cs++ = XY_CTRL_SURF_COPY_BLT |
653 		(src_is_indirect ? 0x0 : 0x1) << SRC_ACCESS_TYPE_SHIFT |
654 		(dst_is_indirect ? 0x0 : 0x1) << DST_ACCESS_TYPE_SHIFT |
655 		ccs_copy_size;
656 	*cs++ = lower_32_bits(src_ofs);
657 	*cs++ = upper_32_bits(src_ofs) | mocs;
658 	*cs++ = lower_32_bits(dst_ofs);
659 	*cs++ = upper_32_bits(dst_ofs) | mocs;
660 
661 	bb->len = cs - bb->cs;
662 }
663 
664 #define EMIT_COPY_DW 10
665 static void emit_copy(struct xe_gt *gt, struct xe_bb *bb,
666 		      u64 src_ofs, u64 dst_ofs, unsigned int size,
667 		      unsigned int pitch)
668 {
669 	struct xe_device *xe = gt_to_xe(gt);
670 	u32 mocs = 0;
671 	u32 tile_y = 0;
672 
673 	xe_gt_assert(gt, size / pitch <= S16_MAX);
674 	xe_gt_assert(gt, pitch / 4 <= S16_MAX);
675 	xe_gt_assert(gt, pitch <= U16_MAX);
676 
677 	if (GRAPHICS_VER(xe) >= 20)
678 		mocs = FIELD_PREP(XE2_XY_FAST_COPY_BLT_MOCS_INDEX_MASK, gt->mocs.uc_index);
679 
680 	if (GRAPHICS_VERx100(xe) >= 1250)
681 		tile_y = XY_FAST_COPY_BLT_D1_SRC_TILE4 | XY_FAST_COPY_BLT_D1_DST_TILE4;
682 
683 	bb->cs[bb->len++] = XY_FAST_COPY_BLT_CMD | (10 - 2);
684 	bb->cs[bb->len++] = XY_FAST_COPY_BLT_DEPTH_32 | pitch | tile_y | mocs;
685 	bb->cs[bb->len++] = 0;
686 	bb->cs[bb->len++] = (size / pitch) << 16 | pitch / 4;
687 	bb->cs[bb->len++] = lower_32_bits(dst_ofs);
688 	bb->cs[bb->len++] = upper_32_bits(dst_ofs);
689 	bb->cs[bb->len++] = 0;
690 	bb->cs[bb->len++] = pitch | mocs;
691 	bb->cs[bb->len++] = lower_32_bits(src_ofs);
692 	bb->cs[bb->len++] = upper_32_bits(src_ofs);
693 }
694 
695 static u64 xe_migrate_batch_base(struct xe_migrate *m, bool usm)
696 {
697 	return usm ? m->usm_batch_base_ofs : m->batch_base_ofs;
698 }
699 
700 static u32 xe_migrate_ccs_copy(struct xe_migrate *m,
701 			       struct xe_bb *bb,
702 			       u64 src_ofs, bool src_is_indirect,
703 			       u64 dst_ofs, bool dst_is_indirect, u32 dst_size,
704 			       u64 ccs_ofs, bool copy_ccs)
705 {
706 	struct xe_gt *gt = m->tile->primary_gt;
707 	u32 flush_flags = 0;
708 
709 	if (!copy_ccs && dst_is_indirect) {
710 		/*
711 		 * If the src is already in vram, then it should already
712 		 * have been cleared by us, or has been populated by the
713 		 * user. Make sure we copy the CCS aux state as-is.
714 		 *
715 		 * Otherwise if the bo doesn't have any CCS metadata attached,
716 		 * we still need to clear it for security reasons.
717 		 */
718 		u64 ccs_src_ofs =  src_is_indirect ? src_ofs : m->cleared_mem_ofs;
719 
720 		emit_copy_ccs(gt, bb,
721 			      dst_ofs, true,
722 			      ccs_src_ofs, src_is_indirect, dst_size);
723 
724 		flush_flags = MI_FLUSH_DW_CCS;
725 	} else if (copy_ccs) {
726 		if (!src_is_indirect)
727 			src_ofs = ccs_ofs;
728 		else if (!dst_is_indirect)
729 			dst_ofs = ccs_ofs;
730 
731 		xe_gt_assert(gt, src_is_indirect || dst_is_indirect);
732 
733 		emit_copy_ccs(gt, bb, dst_ofs, dst_is_indirect, src_ofs,
734 			      src_is_indirect, dst_size);
735 		if (dst_is_indirect)
736 			flush_flags = MI_FLUSH_DW_CCS;
737 	}
738 
739 	return flush_flags;
740 }
741 
742 /**
743  * xe_migrate_copy() - Copy content of TTM resources.
744  * @m: The migration context.
745  * @src_bo: The buffer object @src is currently bound to.
746  * @dst_bo: If copying between resources created for the same bo, set this to
747  * the same value as @src_bo. If copying between buffer objects, set it to
748  * the buffer object @dst is currently bound to.
749  * @src: The source TTM resource.
750  * @dst: The dst TTM resource.
751  * @copy_only_ccs: If true copy only CCS metadata
752  *
753  * Copies the contents of @src to @dst: On flat CCS devices,
754  * the CCS metadata is copied as well if needed, or if not present,
755  * the CCS metadata of @dst is cleared for security reasons.
756  *
757  * Return: Pointer to a dma_fence representing the last copy batch, or
758  * an error pointer on failure. If there is a failure, any copy operation
759  * started by the function call has been synced.
760  */
761 struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
762 				  struct xe_bo *src_bo,
763 				  struct xe_bo *dst_bo,
764 				  struct ttm_resource *src,
765 				  struct ttm_resource *dst,
766 				  bool copy_only_ccs)
767 {
768 	struct xe_gt *gt = m->tile->primary_gt;
769 	struct xe_device *xe = gt_to_xe(gt);
770 	struct dma_fence *fence = NULL;
771 	u64 size = src_bo->size;
772 	struct xe_res_cursor src_it, dst_it, ccs_it;
773 	u64 src_L0_ofs, dst_L0_ofs;
774 	u32 src_L0_pt, dst_L0_pt;
775 	u64 src_L0, dst_L0;
776 	int pass = 0;
777 	int err;
778 	bool src_is_pltt = src->mem_type == XE_PL_TT;
779 	bool dst_is_pltt = dst->mem_type == XE_PL_TT;
780 	bool src_is_vram = mem_type_is_vram(src->mem_type);
781 	bool dst_is_vram = mem_type_is_vram(dst->mem_type);
782 	bool copy_ccs = xe_device_has_flat_ccs(xe) &&
783 		xe_bo_needs_ccs_pages(src_bo) && xe_bo_needs_ccs_pages(dst_bo);
784 	bool copy_system_ccs = copy_ccs && (!src_is_vram || !dst_is_vram);
785 	bool use_comp_pat = xe_device_has_flat_ccs(xe) &&
786 		GRAPHICS_VER(xe) >= 20 && src_is_vram && !dst_is_vram;
787 
788 	/* Copying CCS between two different BOs is not supported yet. */
789 	if (XE_WARN_ON(copy_ccs && src_bo != dst_bo))
790 		return ERR_PTR(-EINVAL);
791 
792 	if (src_bo != dst_bo && XE_WARN_ON(src_bo->size != dst_bo->size))
793 		return ERR_PTR(-EINVAL);
794 
795 	if (!src_is_vram)
796 		xe_res_first_sg(xe_bo_sg(src_bo), 0, size, &src_it);
797 	else
798 		xe_res_first(src, 0, size, &src_it);
799 	if (!dst_is_vram)
800 		xe_res_first_sg(xe_bo_sg(dst_bo), 0, size, &dst_it);
801 	else
802 		xe_res_first(dst, 0, size, &dst_it);
803 
804 	if (copy_system_ccs)
805 		xe_res_first_sg(xe_bo_sg(src_bo), xe_bo_ccs_pages_start(src_bo),
806 				PAGE_ALIGN(xe_device_ccs_bytes(xe, size)),
807 				&ccs_it);
808 
809 	while (size) {
810 		u32 batch_size = 2; /* arb_clear() + MI_BATCH_BUFFER_END */
811 		struct xe_sched_job *job;
812 		struct xe_bb *bb;
813 		u32 flush_flags = 0;
814 		u32 update_idx;
815 		u64 ccs_ofs, ccs_size;
816 		u32 ccs_pt;
817 		u32 pte_flags;
818 
819 		bool usm = xe->info.has_usm;
820 		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
821 
822 		src_L0 = xe_migrate_res_sizes(m, &src_it);
823 		dst_L0 = xe_migrate_res_sizes(m, &dst_it);
824 
825 		drm_dbg(&xe->drm, "Pass %u, sizes: %llu & %llu\n",
826 			pass++, src_L0, dst_L0);
827 
828 		src_L0 = min(src_L0, dst_L0);
829 
830 		pte_flags = src_is_vram ? PTE_UPDATE_FLAG_IS_VRAM : 0;
831 		pte_flags |= use_comp_pat ? PTE_UPDATE_FLAG_IS_COMP_PTE : 0;
832 		batch_size += pte_update_size(m, pte_flags, src, &src_it, &src_L0,
833 					      &src_L0_ofs, &src_L0_pt, 0, 0,
834 					      avail_pts);
835 
836 		pte_flags = dst_is_vram ? PTE_UPDATE_FLAG_IS_VRAM : 0;
837 		batch_size += pte_update_size(m, pte_flags, dst, &dst_it, &src_L0,
838 					      &dst_L0_ofs, &dst_L0_pt, 0,
839 					      avail_pts, avail_pts);
840 
841 		if (copy_system_ccs) {
842 			ccs_size = xe_device_ccs_bytes(xe, src_L0);
843 			batch_size += pte_update_size(m, 0, NULL, &ccs_it, &ccs_size,
844 						      &ccs_ofs, &ccs_pt, 0,
845 						      2 * avail_pts,
846 						      avail_pts);
847 			xe_assert(xe, IS_ALIGNED(ccs_it.start, PAGE_SIZE));
848 		}
849 
850 		/* Add copy commands size here */
851 		batch_size += ((copy_only_ccs) ? 0 : EMIT_COPY_DW) +
852 			((xe_migrate_needs_ccs_emit(xe) ? EMIT_COPY_CCS_DW : 0));
853 
854 		bb = xe_bb_new(gt, batch_size, usm);
855 		if (IS_ERR(bb)) {
856 			err = PTR_ERR(bb);
857 			goto err_sync;
858 		}
859 
860 		if (src_is_vram && xe_migrate_allow_identity(src_L0, &src_it))
861 			xe_res_next(&src_it, src_L0);
862 		else
863 			emit_pte(m, bb, src_L0_pt, src_is_vram, copy_system_ccs,
864 				 &src_it, src_L0, src);
865 
866 		if (dst_is_vram && xe_migrate_allow_identity(src_L0, &dst_it))
867 			xe_res_next(&dst_it, src_L0);
868 		else
869 			emit_pte(m, bb, dst_L0_pt, dst_is_vram, copy_system_ccs,
870 				 &dst_it, src_L0, dst);
871 
872 		if (copy_system_ccs)
873 			emit_pte(m, bb, ccs_pt, false, false, &ccs_it, ccs_size, src);
874 
875 		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
876 		update_idx = bb->len;
877 
878 		if (!copy_only_ccs)
879 			emit_copy(gt, bb, src_L0_ofs, dst_L0_ofs, src_L0, XE_PAGE_SIZE);
880 
881 		if (xe_migrate_needs_ccs_emit(xe))
882 			flush_flags = xe_migrate_ccs_copy(m, bb, src_L0_ofs,
883 							  IS_DGFX(xe) ? src_is_vram : src_is_pltt,
884 							  dst_L0_ofs,
885 							  IS_DGFX(xe) ? dst_is_vram : dst_is_pltt,
886 							  src_L0, ccs_ofs, copy_ccs);
887 
888 		job = xe_bb_create_migration_job(m->q, bb,
889 						 xe_migrate_batch_base(m, usm),
890 						 update_idx);
891 		if (IS_ERR(job)) {
892 			err = PTR_ERR(job);
893 			goto err;
894 		}
895 
896 		xe_sched_job_add_migrate_flush(job, flush_flags);
897 		if (!fence) {
898 			err = xe_sched_job_add_deps(job, src_bo->ttm.base.resv,
899 						    DMA_RESV_USAGE_BOOKKEEP);
900 			if (!err && src_bo != dst_bo)
901 				err = xe_sched_job_add_deps(job, dst_bo->ttm.base.resv,
902 							    DMA_RESV_USAGE_BOOKKEEP);
903 			if (err)
904 				goto err_job;
905 		}
906 
907 		mutex_lock(&m->job_mutex);
908 		xe_sched_job_arm(job);
909 		dma_fence_put(fence);
910 		fence = dma_fence_get(&job->drm.s_fence->finished);
911 		xe_sched_job_push(job);
912 
913 		dma_fence_put(m->fence);
914 		m->fence = dma_fence_get(fence);
915 
916 		mutex_unlock(&m->job_mutex);
917 
918 		xe_bb_free(bb, fence);
919 		size -= src_L0;
920 		continue;
921 
922 err_job:
923 		xe_sched_job_put(job);
924 err:
925 		xe_bb_free(bb, NULL);
926 
927 err_sync:
928 		/* Sync partial copy if any. FIXME: under job_mutex? */
929 		if (fence) {
930 			dma_fence_wait(fence, false);
931 			dma_fence_put(fence);
932 		}
933 
934 		return ERR_PTR(err);
935 	}
936 
937 	return fence;
938 }
939 
940 static void emit_clear_link_copy(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
941 				 u32 size, u32 pitch)
942 {
943 	struct xe_device *xe = gt_to_xe(gt);
944 	u32 *cs = bb->cs + bb->len;
945 	u32 len = PVC_MEM_SET_CMD_LEN_DW;
946 
947 	*cs++ = PVC_MEM_SET_CMD | PVC_MEM_SET_MATRIX | (len - 2);
948 	*cs++ = pitch - 1;
949 	*cs++ = (size / pitch) - 1;
950 	*cs++ = pitch - 1;
951 	*cs++ = lower_32_bits(src_ofs);
952 	*cs++ = upper_32_bits(src_ofs);
953 	if (GRAPHICS_VERx100(xe) >= 2000)
954 		*cs++ = FIELD_PREP(XE2_MEM_SET_MOCS_INDEX_MASK, gt->mocs.uc_index);
955 	else
956 		*cs++ = FIELD_PREP(PVC_MEM_SET_MOCS_INDEX_MASK, gt->mocs.uc_index);
957 
958 	xe_gt_assert(gt, cs - bb->cs == len + bb->len);
959 
960 	bb->len += len;
961 }
962 
963 static void emit_clear_main_copy(struct xe_gt *gt, struct xe_bb *bb,
964 				 u64 src_ofs, u32 size, u32 pitch, bool is_vram)
965 {
966 	struct xe_device *xe = gt_to_xe(gt);
967 	u32 *cs = bb->cs + bb->len;
968 	u32 len = XY_FAST_COLOR_BLT_DW;
969 
970 	if (GRAPHICS_VERx100(xe) < 1250)
971 		len = 11;
972 
973 	*cs++ = XY_FAST_COLOR_BLT_CMD | XY_FAST_COLOR_BLT_DEPTH_32 |
974 		(len - 2);
975 	if (GRAPHICS_VERx100(xe) >= 2000)
976 		*cs++ = FIELD_PREP(XE2_XY_FAST_COLOR_BLT_MOCS_INDEX_MASK, gt->mocs.uc_index) |
977 			(pitch - 1);
978 	else
979 		*cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, gt->mocs.uc_index) |
980 			(pitch - 1);
981 	*cs++ = 0;
982 	*cs++ = (size / pitch) << 16 | pitch / 4;
983 	*cs++ = lower_32_bits(src_ofs);
984 	*cs++ = upper_32_bits(src_ofs);
985 	*cs++ = (is_vram ? 0x0 : 0x1) <<  XY_FAST_COLOR_BLT_MEM_TYPE_SHIFT;
986 	*cs++ = 0;
987 	*cs++ = 0;
988 	*cs++ = 0;
989 	*cs++ = 0;
990 
991 	if (len > 11) {
992 		*cs++ = 0;
993 		*cs++ = 0;
994 		*cs++ = 0;
995 		*cs++ = 0;
996 		*cs++ = 0;
997 	}
998 
999 	xe_gt_assert(gt, cs - bb->cs == len + bb->len);
1000 
1001 	bb->len += len;
1002 }
1003 
1004 static bool has_service_copy_support(struct xe_gt *gt)
1005 {
1006 	/*
1007 	 * What we care about is whether the architecture was designed with
1008 	 * service copy functionality (specifically the new MEM_SET / MEM_COPY
1009 	 * instructions) so check the architectural engine list rather than the
1010 	 * actual list since these instructions are usable on BCS0 even if
1011 	 * all of the actual service copy engines (BCS1-BCS8) have been fused
1012 	 * off.
1013 	 */
1014 	return gt->info.engine_mask & GENMASK(XE_HW_ENGINE_BCS8,
1015 					      XE_HW_ENGINE_BCS1);
1016 }
1017 
1018 static u32 emit_clear_cmd_len(struct xe_gt *gt)
1019 {
1020 	if (has_service_copy_support(gt))
1021 		return PVC_MEM_SET_CMD_LEN_DW;
1022 	else
1023 		return XY_FAST_COLOR_BLT_DW;
1024 }
1025 
1026 static void emit_clear(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
1027 		       u32 size, u32 pitch, bool is_vram)
1028 {
1029 	if (has_service_copy_support(gt))
1030 		emit_clear_link_copy(gt, bb, src_ofs, size, pitch);
1031 	else
1032 		emit_clear_main_copy(gt, bb, src_ofs, size, pitch,
1033 				     is_vram);
1034 }
1035 
1036 /**
1037  * xe_migrate_clear() - Copy content of TTM resources.
1038  * @m: The migration context.
1039  * @bo: The buffer object @dst is currently bound to.
1040  * @dst: The dst TTM resource to be cleared.
1041  * @clear_flags: flags to specify which data to clear: CCS, BO, or both.
1042  *
1043  * Clear the contents of @dst to zero when XE_MIGRATE_CLEAR_FLAG_BO_DATA is set.
1044  * On flat CCS devices, the CCS metadata is cleared to zero with XE_MIGRATE_CLEAR_FLAG_CCS_DATA.
1045  * Set XE_MIGRATE_CLEAR_FLAG_FULL to clear bo as well as CCS metadata.
1046  * TODO: Eliminate the @bo argument.
1047  *
1048  * Return: Pointer to a dma_fence representing the last clear batch, or
1049  * an error pointer on failure. If there is a failure, any clear operation
1050  * started by the function call has been synced.
1051  */
1052 struct dma_fence *xe_migrate_clear(struct xe_migrate *m,
1053 				   struct xe_bo *bo,
1054 				   struct ttm_resource *dst,
1055 				   u32 clear_flags)
1056 {
1057 	bool clear_vram = mem_type_is_vram(dst->mem_type);
1058 	bool clear_bo_data = XE_MIGRATE_CLEAR_FLAG_BO_DATA & clear_flags;
1059 	bool clear_ccs = XE_MIGRATE_CLEAR_FLAG_CCS_DATA & clear_flags;
1060 	struct xe_gt *gt = m->tile->primary_gt;
1061 	struct xe_device *xe = gt_to_xe(gt);
1062 	bool clear_only_system_ccs = false;
1063 	struct dma_fence *fence = NULL;
1064 	u64 size = bo->size;
1065 	struct xe_res_cursor src_it;
1066 	struct ttm_resource *src = dst;
1067 	int err;
1068 
1069 	if (WARN_ON(!clear_bo_data && !clear_ccs))
1070 		return NULL;
1071 
1072 	if (!clear_bo_data && clear_ccs && !IS_DGFX(xe))
1073 		clear_only_system_ccs = true;
1074 
1075 	if (!clear_vram)
1076 		xe_res_first_sg(xe_bo_sg(bo), 0, bo->size, &src_it);
1077 	else
1078 		xe_res_first(src, 0, bo->size, &src_it);
1079 
1080 	while (size) {
1081 		u64 clear_L0_ofs;
1082 		u32 clear_L0_pt;
1083 		u32 flush_flags = 0;
1084 		u64 clear_L0;
1085 		struct xe_sched_job *job;
1086 		struct xe_bb *bb;
1087 		u32 batch_size, update_idx;
1088 		u32 pte_flags;
1089 
1090 		bool usm = xe->info.has_usm;
1091 		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
1092 
1093 		clear_L0 = xe_migrate_res_sizes(m, &src_it);
1094 
1095 		/* Calculate final sizes and batch size.. */
1096 		pte_flags = clear_vram ? PTE_UPDATE_FLAG_IS_VRAM : 0;
1097 		batch_size = 2 +
1098 			pte_update_size(m, pte_flags, src, &src_it,
1099 					&clear_L0, &clear_L0_ofs, &clear_L0_pt,
1100 					clear_bo_data ? emit_clear_cmd_len(gt) : 0, 0,
1101 					avail_pts);
1102 
1103 		if (xe_migrate_needs_ccs_emit(xe))
1104 			batch_size += EMIT_COPY_CCS_DW;
1105 
1106 		/* Clear commands */
1107 
1108 		if (WARN_ON_ONCE(!clear_L0))
1109 			break;
1110 
1111 		bb = xe_bb_new(gt, batch_size, usm);
1112 		if (IS_ERR(bb)) {
1113 			err = PTR_ERR(bb);
1114 			goto err_sync;
1115 		}
1116 
1117 		size -= clear_L0;
1118 		/* Preemption is enabled again by the ring ops. */
1119 		if (clear_vram && xe_migrate_allow_identity(clear_L0, &src_it))
1120 			xe_res_next(&src_it, clear_L0);
1121 		else
1122 			emit_pte(m, bb, clear_L0_pt, clear_vram, clear_only_system_ccs,
1123 				 &src_it, clear_L0, dst);
1124 
1125 		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1126 		update_idx = bb->len;
1127 
1128 		if (clear_bo_data)
1129 			emit_clear(gt, bb, clear_L0_ofs, clear_L0, XE_PAGE_SIZE, clear_vram);
1130 
1131 		if (xe_migrate_needs_ccs_emit(xe)) {
1132 			emit_copy_ccs(gt, bb, clear_L0_ofs, true,
1133 				      m->cleared_mem_ofs, false, clear_L0);
1134 			flush_flags = MI_FLUSH_DW_CCS;
1135 		}
1136 
1137 		job = xe_bb_create_migration_job(m->q, bb,
1138 						 xe_migrate_batch_base(m, usm),
1139 						 update_idx);
1140 		if (IS_ERR(job)) {
1141 			err = PTR_ERR(job);
1142 			goto err;
1143 		}
1144 
1145 		xe_sched_job_add_migrate_flush(job, flush_flags);
1146 		if (!fence) {
1147 			/*
1148 			 * There can't be anything userspace related at this
1149 			 * point, so we just need to respect any potential move
1150 			 * fences, which are always tracked as
1151 			 * DMA_RESV_USAGE_KERNEL.
1152 			 */
1153 			err = xe_sched_job_add_deps(job, bo->ttm.base.resv,
1154 						    DMA_RESV_USAGE_KERNEL);
1155 			if (err)
1156 				goto err_job;
1157 		}
1158 
1159 		mutex_lock(&m->job_mutex);
1160 		xe_sched_job_arm(job);
1161 		dma_fence_put(fence);
1162 		fence = dma_fence_get(&job->drm.s_fence->finished);
1163 		xe_sched_job_push(job);
1164 
1165 		dma_fence_put(m->fence);
1166 		m->fence = dma_fence_get(fence);
1167 
1168 		mutex_unlock(&m->job_mutex);
1169 
1170 		xe_bb_free(bb, fence);
1171 		continue;
1172 
1173 err_job:
1174 		xe_sched_job_put(job);
1175 err:
1176 		xe_bb_free(bb, NULL);
1177 err_sync:
1178 		/* Sync partial copies if any. FIXME: job_mutex? */
1179 		if (fence) {
1180 			dma_fence_wait(m->fence, false);
1181 			dma_fence_put(fence);
1182 		}
1183 
1184 		return ERR_PTR(err);
1185 	}
1186 
1187 	if (clear_ccs)
1188 		bo->ccs_cleared = true;
1189 
1190 	return fence;
1191 }
1192 
1193 static void write_pgtable(struct xe_tile *tile, struct xe_bb *bb, u64 ppgtt_ofs,
1194 			  const struct xe_vm_pgtable_update_op *pt_op,
1195 			  const struct xe_vm_pgtable_update *update,
1196 			  struct xe_migrate_pt_update *pt_update)
1197 {
1198 	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1199 	u32 chunk;
1200 	u32 ofs = update->ofs, size = update->qwords;
1201 
1202 	/*
1203 	 * If we have 512 entries (max), we would populate it ourselves,
1204 	 * and update the PDE above it to the new pointer.
1205 	 * The only time this can only happen if we have to update the top
1206 	 * PDE. This requires a BO that is almost vm->size big.
1207 	 *
1208 	 * This shouldn't be possible in practice.. might change when 16K
1209 	 * pages are used. Hence the assert.
1210 	 */
1211 	xe_tile_assert(tile, update->qwords < MAX_NUM_PTE);
1212 	if (!ppgtt_ofs)
1213 		ppgtt_ofs = xe_migrate_vram_ofs(tile_to_xe(tile),
1214 						xe_bo_addr(update->pt_bo, 0,
1215 							   XE_PAGE_SIZE), false);
1216 
1217 	do {
1218 		u64 addr = ppgtt_ofs + ofs * 8;
1219 
1220 		chunk = min(size, MAX_PTE_PER_SDI);
1221 
1222 		/* Ensure populatefn can do memset64 by aligning bb->cs */
1223 		if (!(bb->len & 1))
1224 			bb->cs[bb->len++] = MI_NOOP;
1225 
1226 		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(chunk);
1227 		bb->cs[bb->len++] = lower_32_bits(addr);
1228 		bb->cs[bb->len++] = upper_32_bits(addr);
1229 		if (pt_op->bind)
1230 			ops->populate(pt_update, tile, NULL, bb->cs + bb->len,
1231 				      ofs, chunk, update);
1232 		else
1233 			ops->clear(pt_update, tile, NULL, bb->cs + bb->len,
1234 				   ofs, chunk, update);
1235 
1236 		bb->len += chunk * 2;
1237 		ofs += chunk;
1238 		size -= chunk;
1239 	} while (size);
1240 }
1241 
1242 struct xe_vm *xe_migrate_get_vm(struct xe_migrate *m)
1243 {
1244 	return xe_vm_get(m->q->vm);
1245 }
1246 
1247 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
1248 struct migrate_test_params {
1249 	struct xe_test_priv base;
1250 	bool force_gpu;
1251 };
1252 
1253 #define to_migrate_test_params(_priv) \
1254 	container_of(_priv, struct migrate_test_params, base)
1255 #endif
1256 
1257 static struct dma_fence *
1258 xe_migrate_update_pgtables_cpu(struct xe_migrate *m,
1259 			       struct xe_migrate_pt_update *pt_update)
1260 {
1261 	XE_TEST_DECLARE(struct migrate_test_params *test =
1262 			to_migrate_test_params
1263 			(xe_cur_kunit_priv(XE_TEST_LIVE_MIGRATE));)
1264 	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1265 	struct xe_vm *vm = pt_update->vops->vm;
1266 	struct xe_vm_pgtable_update_ops *pt_update_ops =
1267 		&pt_update->vops->pt_update_ops[pt_update->tile_id];
1268 	int err;
1269 	u32 i, j;
1270 
1271 	if (XE_TEST_ONLY(test && test->force_gpu))
1272 		return ERR_PTR(-ETIME);
1273 
1274 	if (ops->pre_commit) {
1275 		pt_update->job = NULL;
1276 		err = ops->pre_commit(pt_update);
1277 		if (err)
1278 			return ERR_PTR(err);
1279 	}
1280 
1281 	for (i = 0; i < pt_update_ops->num_ops; ++i) {
1282 		const struct xe_vm_pgtable_update_op *pt_op =
1283 			&pt_update_ops->ops[i];
1284 
1285 		for (j = 0; j < pt_op->num_entries; j++) {
1286 			const struct xe_vm_pgtable_update *update =
1287 				&pt_op->entries[j];
1288 
1289 			if (pt_op->bind)
1290 				ops->populate(pt_update, m->tile,
1291 					      &update->pt_bo->vmap, NULL,
1292 					      update->ofs, update->qwords,
1293 					      update);
1294 			else
1295 				ops->clear(pt_update, m->tile,
1296 					   &update->pt_bo->vmap, NULL,
1297 					   update->ofs, update->qwords, update);
1298 		}
1299 	}
1300 
1301 	trace_xe_vm_cpu_bind(vm);
1302 	xe_device_wmb(vm->xe);
1303 
1304 	return dma_fence_get_stub();
1305 }
1306 
1307 static struct dma_fence *
1308 __xe_migrate_update_pgtables(struct xe_migrate *m,
1309 			     struct xe_migrate_pt_update *pt_update,
1310 			     struct xe_vm_pgtable_update_ops *pt_update_ops)
1311 {
1312 	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1313 	struct xe_tile *tile = m->tile;
1314 	struct xe_gt *gt = tile->primary_gt;
1315 	struct xe_device *xe = tile_to_xe(tile);
1316 	struct xe_sched_job *job;
1317 	struct dma_fence *fence;
1318 	struct drm_suballoc *sa_bo = NULL;
1319 	struct xe_bb *bb;
1320 	u32 i, j, batch_size = 0, ppgtt_ofs, update_idx, page_ofs = 0;
1321 	u32 num_updates = 0, current_update = 0;
1322 	u64 addr;
1323 	int err = 0;
1324 	bool is_migrate = pt_update_ops->q == m->q;
1325 	bool usm = is_migrate && xe->info.has_usm;
1326 
1327 	for (i = 0; i < pt_update_ops->num_ops; ++i) {
1328 		struct xe_vm_pgtable_update_op *pt_op = &pt_update_ops->ops[i];
1329 		struct xe_vm_pgtable_update *updates = pt_op->entries;
1330 
1331 		num_updates += pt_op->num_entries;
1332 		for (j = 0; j < pt_op->num_entries; ++j) {
1333 			u32 num_cmds = DIV_ROUND_UP(updates[j].qwords,
1334 						    MAX_PTE_PER_SDI);
1335 
1336 			/* align noop + MI_STORE_DATA_IMM cmd prefix */
1337 			batch_size += 4 * num_cmds + updates[j].qwords * 2;
1338 		}
1339 	}
1340 
1341 	/* fixed + PTE entries */
1342 	if (IS_DGFX(xe))
1343 		batch_size += 2;
1344 	else
1345 		batch_size += 6 * (num_updates / MAX_PTE_PER_SDI + 1) +
1346 			num_updates * 2;
1347 
1348 	bb = xe_bb_new(gt, batch_size, usm);
1349 	if (IS_ERR(bb))
1350 		return ERR_CAST(bb);
1351 
1352 	/* For sysmem PTE's, need to map them in our hole.. */
1353 	if (!IS_DGFX(xe)) {
1354 		u16 pat_index = xe->pat.idx[XE_CACHE_WB];
1355 		u32 ptes, ofs;
1356 
1357 		ppgtt_ofs = NUM_KERNEL_PDE - 1;
1358 		if (!is_migrate) {
1359 			u32 num_units = DIV_ROUND_UP(num_updates,
1360 						     NUM_VMUSA_WRITES_PER_UNIT);
1361 
1362 			if (num_units > m->vm_update_sa.size) {
1363 				err = -ENOBUFS;
1364 				goto err_bb;
1365 			}
1366 			sa_bo = drm_suballoc_new(&m->vm_update_sa, num_units,
1367 						 GFP_KERNEL, true, 0);
1368 			if (IS_ERR(sa_bo)) {
1369 				err = PTR_ERR(sa_bo);
1370 				goto err_bb;
1371 			}
1372 
1373 			ppgtt_ofs = NUM_KERNEL_PDE +
1374 				(drm_suballoc_soffset(sa_bo) /
1375 				 NUM_VMUSA_UNIT_PER_PAGE);
1376 			page_ofs = (drm_suballoc_soffset(sa_bo) %
1377 				    NUM_VMUSA_UNIT_PER_PAGE) *
1378 				VM_SA_UPDATE_UNIT_SIZE;
1379 		}
1380 
1381 		/* Map our PT's to gtt */
1382 		i = 0;
1383 		j = 0;
1384 		ptes = num_updates;
1385 		ofs = ppgtt_ofs * XE_PAGE_SIZE + page_ofs;
1386 		while (ptes) {
1387 			u32 chunk = min(MAX_PTE_PER_SDI, ptes);
1388 			u32 idx = 0;
1389 
1390 			bb->cs[bb->len++] = MI_STORE_DATA_IMM |
1391 				MI_SDI_NUM_QW(chunk);
1392 			bb->cs[bb->len++] = ofs;
1393 			bb->cs[bb->len++] = 0; /* upper_32_bits */
1394 
1395 			for (; i < pt_update_ops->num_ops; ++i) {
1396 				struct xe_vm_pgtable_update_op *pt_op =
1397 					&pt_update_ops->ops[i];
1398 				struct xe_vm_pgtable_update *updates = pt_op->entries;
1399 
1400 				for (; j < pt_op->num_entries; ++j, ++current_update, ++idx) {
1401 					struct xe_vm *vm = pt_update->vops->vm;
1402 					struct xe_bo *pt_bo = updates[j].pt_bo;
1403 
1404 					if (idx == chunk)
1405 						goto next_cmd;
1406 
1407 					xe_tile_assert(tile, pt_bo->size == SZ_4K);
1408 
1409 					/* Map a PT at most once */
1410 					if (pt_bo->update_index < 0)
1411 						pt_bo->update_index = current_update;
1412 
1413 					addr = vm->pt_ops->pte_encode_bo(pt_bo, 0,
1414 									 pat_index, 0);
1415 					bb->cs[bb->len++] = lower_32_bits(addr);
1416 					bb->cs[bb->len++] = upper_32_bits(addr);
1417 				}
1418 
1419 				j = 0;
1420 			}
1421 
1422 next_cmd:
1423 			ptes -= chunk;
1424 			ofs += chunk * sizeof(u64);
1425 		}
1426 
1427 		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1428 		update_idx = bb->len;
1429 
1430 		addr = xe_migrate_vm_addr(ppgtt_ofs, 0) +
1431 			(page_ofs / sizeof(u64)) * XE_PAGE_SIZE;
1432 		for (i = 0; i < pt_update_ops->num_ops; ++i) {
1433 			struct xe_vm_pgtable_update_op *pt_op =
1434 				&pt_update_ops->ops[i];
1435 			struct xe_vm_pgtable_update *updates = pt_op->entries;
1436 
1437 			for (j = 0; j < pt_op->num_entries; ++j) {
1438 				struct xe_bo *pt_bo = updates[j].pt_bo;
1439 
1440 				write_pgtable(tile, bb, addr +
1441 					      pt_bo->update_index * XE_PAGE_SIZE,
1442 					      pt_op, &updates[j], pt_update);
1443 			}
1444 		}
1445 	} else {
1446 		/* phys pages, no preamble required */
1447 		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1448 		update_idx = bb->len;
1449 
1450 		for (i = 0; i < pt_update_ops->num_ops; ++i) {
1451 			struct xe_vm_pgtable_update_op *pt_op =
1452 				&pt_update_ops->ops[i];
1453 			struct xe_vm_pgtable_update *updates = pt_op->entries;
1454 
1455 			for (j = 0; j < pt_op->num_entries; ++j)
1456 				write_pgtable(tile, bb, 0, pt_op, &updates[j],
1457 					      pt_update);
1458 		}
1459 	}
1460 
1461 	job = xe_bb_create_migration_job(pt_update_ops->q, bb,
1462 					 xe_migrate_batch_base(m, usm),
1463 					 update_idx);
1464 	if (IS_ERR(job)) {
1465 		err = PTR_ERR(job);
1466 		goto err_sa;
1467 	}
1468 
1469 	if (ops->pre_commit) {
1470 		pt_update->job = job;
1471 		err = ops->pre_commit(pt_update);
1472 		if (err)
1473 			goto err_job;
1474 	}
1475 	if (is_migrate)
1476 		mutex_lock(&m->job_mutex);
1477 
1478 	xe_sched_job_arm(job);
1479 	fence = dma_fence_get(&job->drm.s_fence->finished);
1480 	xe_sched_job_push(job);
1481 
1482 	if (is_migrate)
1483 		mutex_unlock(&m->job_mutex);
1484 
1485 	xe_bb_free(bb, fence);
1486 	drm_suballoc_free(sa_bo, fence);
1487 
1488 	return fence;
1489 
1490 err_job:
1491 	xe_sched_job_put(job);
1492 err_sa:
1493 	drm_suballoc_free(sa_bo, NULL);
1494 err_bb:
1495 	xe_bb_free(bb, NULL);
1496 	return ERR_PTR(err);
1497 }
1498 
1499 /**
1500  * xe_migrate_update_pgtables() - Pipelined page-table update
1501  * @m: The migrate context.
1502  * @pt_update: PT update arguments
1503  *
1504  * Perform a pipelined page-table update. The update descriptors are typically
1505  * built under the same lock critical section as a call to this function. If
1506  * using the default engine for the updates, they will be performed in the
1507  * order they grab the job_mutex. If different engines are used, external
1508  * synchronization is needed for overlapping updates to maintain page-table
1509  * consistency. Note that the meaing of "overlapping" is that the updates
1510  * touch the same page-table, which might be a higher-level page-directory.
1511  * If no pipelining is needed, then updates may be performed by the cpu.
1512  *
1513  * Return: A dma_fence that, when signaled, indicates the update completion.
1514  */
1515 struct dma_fence *
1516 xe_migrate_update_pgtables(struct xe_migrate *m,
1517 			   struct xe_migrate_pt_update *pt_update)
1518 
1519 {
1520 	struct xe_vm_pgtable_update_ops *pt_update_ops =
1521 		&pt_update->vops->pt_update_ops[pt_update->tile_id];
1522 	struct dma_fence *fence;
1523 
1524 	fence =  xe_migrate_update_pgtables_cpu(m, pt_update);
1525 
1526 	/* -ETIME indicates a job is needed, anything else is legit error */
1527 	if (!IS_ERR(fence) || PTR_ERR(fence) != -ETIME)
1528 		return fence;
1529 
1530 	return __xe_migrate_update_pgtables(m, pt_update, pt_update_ops);
1531 }
1532 
1533 /**
1534  * xe_migrate_wait() - Complete all operations using the xe_migrate context
1535  * @m: Migrate context to wait for.
1536  *
1537  * Waits until the GPU no longer uses the migrate context's default engine
1538  * or its page-table objects. FIXME: What about separate page-table update
1539  * engines?
1540  */
1541 void xe_migrate_wait(struct xe_migrate *m)
1542 {
1543 	if (m->fence)
1544 		dma_fence_wait(m->fence, false);
1545 }
1546 
1547 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
1548 #include "tests/xe_migrate.c"
1549 #endif
1550