Linux Audio

Check our new training course

Loading...
Note: File does not exist in v4.6.
   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 <drm/xe_drm.h>
  14
  15#include "generated/xe_wa_oob.h"
  16#include "instructions/xe_mi_commands.h"
  17#include "regs/xe_gpu_commands.h"
  18#include "tests/xe_test.h"
  19#include "xe_assert.h"
  20#include "xe_bb.h"
  21#include "xe_bo.h"
  22#include "xe_exec_queue.h"
  23#include "xe_ggtt.h"
  24#include "xe_gt.h"
  25#include "xe_hw_engine.h"
  26#include "xe_lrc.h"
  27#include "xe_map.h"
  28#include "xe_mocs.h"
  29#include "xe_pt.h"
  30#include "xe_res_cursor.h"
  31#include "xe_sched_job.h"
  32#include "xe_sync.h"
  33#include "xe_trace.h"
  34#include "xe_vm.h"
  35#include "xe_wa.h"
  36
  37/**
  38 * struct xe_migrate - migrate context.
  39 */
  40struct xe_migrate {
  41	/** @q: Default exec queue used for migration */
  42	struct xe_exec_queue *q;
  43	/** @tile: Backpointer to the tile this struct xe_migrate belongs to. */
  44	struct xe_tile *tile;
  45	/** @job_mutex: Timeline mutex for @eng. */
  46	struct mutex job_mutex;
  47	/** @pt_bo: Page-table buffer object. */
  48	struct xe_bo *pt_bo;
  49	/** @batch_base_ofs: VM offset of the migration batch buffer */
  50	u64 batch_base_ofs;
  51	/** @usm_batch_base_ofs: VM offset of the usm batch buffer */
  52	u64 usm_batch_base_ofs;
  53	/** @cleared_mem_ofs: VM offset of @cleared_bo. */
  54	u64 cleared_mem_ofs;
  55	/**
  56	 * @fence: dma-fence representing the last migration job batch.
  57	 * Protected by @job_mutex.
  58	 */
  59	struct dma_fence *fence;
  60	/**
  61	 * @vm_update_sa: For integrated, used to suballocate page-tables
  62	 * out of the pt_bo.
  63	 */
  64	struct drm_suballoc_manager vm_update_sa;
  65	/** @min_chunk_size: For dgfx, Minimum chunk size */
  66	u64 min_chunk_size;
  67};
  68
  69#define MAX_PREEMPTDISABLE_TRANSFER SZ_8M /* Around 1ms. */
  70#define MAX_CCS_LIMITED_TRANSFER SZ_4M /* XE_PAGE_SIZE * (FIELD_MAX(XE2_CCS_SIZE_MASK) + 1) */
  71#define NUM_KERNEL_PDE 17
  72#define NUM_PT_SLOTS 32
  73#define LEVEL0_PAGE_TABLE_ENCODE_SIZE SZ_2M
  74
  75/**
  76 * xe_tile_migrate_engine() - Get this tile's migrate engine.
  77 * @tile: The tile.
  78 *
  79 * Returns the default migrate engine of this tile.
  80 * TODO: Perhaps this function is slightly misplaced, and even unneeded?
  81 *
  82 * Return: The default migrate engine
  83 */
  84struct xe_exec_queue *xe_tile_migrate_engine(struct xe_tile *tile)
  85{
  86	return tile->migrate->q;
  87}
  88
  89static void xe_migrate_fini(struct drm_device *dev, void *arg)
  90{
  91	struct xe_migrate *m = arg;
  92
  93	xe_vm_lock(m->q->vm, false);
  94	xe_bo_unpin(m->pt_bo);
  95	xe_vm_unlock(m->q->vm);
  96
  97	dma_fence_put(m->fence);
  98	xe_bo_put(m->pt_bo);
  99	drm_suballoc_manager_fini(&m->vm_update_sa);
 100	mutex_destroy(&m->job_mutex);
 101	xe_vm_close_and_put(m->q->vm);
 102	xe_exec_queue_put(m->q);
 103}
 104
 105static u64 xe_migrate_vm_addr(u64 slot, u32 level)
 106{
 107	XE_WARN_ON(slot >= NUM_PT_SLOTS);
 108
 109	/* First slot is reserved for mapping of PT bo and bb, start from 1 */
 110	return (slot + 1ULL) << xe_pt_shift(level + 1);
 111}
 112
 113static u64 xe_migrate_vram_ofs(struct xe_device *xe, u64 addr)
 114{
 115	/*
 116	 * Remove the DPA to get a correct offset into identity table for the
 117	 * migrate offset
 118	 */
 119	addr -= xe->mem.vram.dpa_base;
 120	return addr + (256ULL << xe_pt_shift(2));
 121}
 122
 123static int xe_migrate_prepare_vm(struct xe_tile *tile, struct xe_migrate *m,
 124				 struct xe_vm *vm)
 125{
 126	struct xe_device *xe = tile_to_xe(tile);
 127	u16 pat_index = xe->pat.idx[XE_CACHE_WB];
 128	u8 id = tile->id;
 129	u32 num_entries = NUM_PT_SLOTS, num_level = vm->pt_root[id]->level;
 130	u32 map_ofs, level, i;
 131	struct xe_bo *bo, *batch = tile->mem.kernel_bb_pool->bo;
 132	u64 entry;
 133
 134	/* Can't bump NUM_PT_SLOTS too high */
 135	BUILD_BUG_ON(NUM_PT_SLOTS > SZ_2M/XE_PAGE_SIZE);
 136	/* Must be a multiple of 64K to support all platforms */
 137	BUILD_BUG_ON(NUM_PT_SLOTS * XE_PAGE_SIZE % SZ_64K);
 138	/* And one slot reserved for the 4KiB page table updates */
 139	BUILD_BUG_ON(!(NUM_KERNEL_PDE & 1));
 140
 141	/* Need to be sure everything fits in the first PT, or create more */
 142	xe_tile_assert(tile, m->batch_base_ofs + batch->size < SZ_2M);
 143
 144	bo = xe_bo_create_pin_map(vm->xe, tile, vm,
 145				  num_entries * XE_PAGE_SIZE,
 146				  ttm_bo_type_kernel,
 147				  XE_BO_CREATE_VRAM_IF_DGFX(tile) |
 148				  XE_BO_CREATE_PINNED_BIT);
 149	if (IS_ERR(bo))
 150		return PTR_ERR(bo);
 151
 152	entry = vm->pt_ops->pde_encode_bo(bo, bo->size - XE_PAGE_SIZE, pat_index);
 153	xe_pt_write(xe, &vm->pt_root[id]->bo->vmap, 0, entry);
 154
 155	map_ofs = (num_entries - num_level) * XE_PAGE_SIZE;
 156
 157	/* Map the entire BO in our level 0 pt */
 158	for (i = 0, level = 0; i < num_entries; level++) {
 159		entry = vm->pt_ops->pte_encode_bo(bo, i * XE_PAGE_SIZE,
 160						  pat_index, 0);
 161
 162		xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64, entry);
 163
 164		if (vm->flags & XE_VM_FLAG_64K)
 165			i += 16;
 166		else
 167			i += 1;
 168	}
 169
 170	if (!IS_DGFX(xe)) {
 171		/* Write out batch too */
 172		m->batch_base_ofs = NUM_PT_SLOTS * XE_PAGE_SIZE;
 173		for (i = 0; i < batch->size;
 174		     i += vm->flags & XE_VM_FLAG_64K ? XE_64K_PAGE_SIZE :
 175		     XE_PAGE_SIZE) {
 176			entry = vm->pt_ops->pte_encode_bo(batch, i,
 177							  pat_index, 0);
 178
 179			xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64,
 180				  entry);
 181			level++;
 182		}
 183		if (xe->info.has_usm) {
 184			xe_tile_assert(tile, batch->size == SZ_1M);
 185
 186			batch = tile->primary_gt->usm.bb_pool->bo;
 187			m->usm_batch_base_ofs = m->batch_base_ofs + SZ_1M;
 188			xe_tile_assert(tile, batch->size == SZ_512K);
 189
 190			for (i = 0; i < batch->size;
 191			     i += vm->flags & XE_VM_FLAG_64K ? XE_64K_PAGE_SIZE :
 192			     XE_PAGE_SIZE) {
 193				entry = vm->pt_ops->pte_encode_bo(batch, i,
 194								  pat_index, 0);
 195
 196				xe_map_wr(xe, &bo->vmap, map_ofs + level * 8, u64,
 197					  entry);
 198				level++;
 199			}
 200		}
 201	} else {
 202		u64 batch_addr = xe_bo_addr(batch, 0, XE_PAGE_SIZE);
 203
 204		m->batch_base_ofs = xe_migrate_vram_ofs(xe, batch_addr);
 205
 206		if (xe->info.has_usm) {
 207			batch = tile->primary_gt->usm.bb_pool->bo;
 208			batch_addr = xe_bo_addr(batch, 0, XE_PAGE_SIZE);
 209			m->usm_batch_base_ofs = xe_migrate_vram_ofs(xe, batch_addr);
 210		}
 211	}
 212
 213	for (level = 1; level < num_level; level++) {
 214		u32 flags = 0;
 215
 216		if (vm->flags & XE_VM_FLAG_64K && level == 1)
 217			flags = XE_PDE_64K;
 218
 219		entry = vm->pt_ops->pde_encode_bo(bo, map_ofs + (level - 1) *
 220						  XE_PAGE_SIZE, pat_index);
 221		xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE * level, u64,
 222			  entry | flags);
 223	}
 224
 225	/* Write PDE's that point to our BO. */
 226	for (i = 0; i < num_entries - num_level; i++) {
 227		entry = vm->pt_ops->pde_encode_bo(bo, i * XE_PAGE_SIZE,
 228						  pat_index);
 229
 230		xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE +
 231			  (i + 1) * 8, u64, entry);
 232	}
 233
 234	/* Set up a 1GiB NULL mapping at 255GiB offset. */
 235	level = 2;
 236	xe_map_wr(xe, &bo->vmap, map_ofs + XE_PAGE_SIZE * level + 255 * 8, u64,
 237		  vm->pt_ops->pte_encode_addr(xe, 0, pat_index, level, IS_DGFX(xe), 0)
 238		  | XE_PTE_NULL);
 239	m->cleared_mem_ofs = (255ULL << xe_pt_shift(level));
 240
 241	/* Identity map the entire vram at 256GiB offset */
 242	if (IS_DGFX(xe)) {
 243		u64 pos, ofs, flags;
 244
 245		level = 2;
 246		ofs = map_ofs + XE_PAGE_SIZE * level + 256 * 8;
 247		flags = vm->pt_ops->pte_encode_addr(xe, 0, pat_index, level,
 248						    true, 0);
 249
 250		/*
 251		 * Use 1GB pages, it shouldn't matter the physical amount of
 252		 * vram is less, when we don't access it.
 253		 */
 254		for (pos = xe->mem.vram.dpa_base;
 255		     pos < xe->mem.vram.actual_physical_size + xe->mem.vram.dpa_base;
 256		     pos += SZ_1G, ofs += 8)
 257			xe_map_wr(xe, &bo->vmap, ofs, u64, pos | flags);
 258	}
 259
 260	/*
 261	 * Example layout created above, with root level = 3:
 262	 * [PT0...PT7]: kernel PT's for copy/clear; 64 or 4KiB PTE's
 263	 * [PT8]: Kernel PT for VM_BIND, 4 KiB PTE's
 264	 * [PT9...PT28]: Userspace PT's for VM_BIND, 4 KiB PTE's
 265	 * [PT29 = PDE 0] [PT30 = PDE 1] [PT31 = PDE 2]
 266	 *
 267	 * This makes the lowest part of the VM point to the pagetables.
 268	 * Hence the lowest 2M in the vm should point to itself, with a few writes
 269	 * and flushes, other parts of the VM can be used either for copying and
 270	 * clearing.
 271	 *
 272	 * For performance, the kernel reserves PDE's, so about 20 are left
 273	 * for async VM updates.
 274	 *
 275	 * To make it easier to work, each scratch PT is put in slot (1 + PT #)
 276	 * everywhere, this allows lockless updates to scratch pages by using
 277	 * the different addresses in VM.
 278	 */
 279#define NUM_VMUSA_UNIT_PER_PAGE	32
 280#define VM_SA_UPDATE_UNIT_SIZE		(XE_PAGE_SIZE / NUM_VMUSA_UNIT_PER_PAGE)
 281#define NUM_VMUSA_WRITES_PER_UNIT	(VM_SA_UPDATE_UNIT_SIZE / sizeof(u64))
 282	drm_suballoc_manager_init(&m->vm_update_sa,
 283				  (map_ofs / XE_PAGE_SIZE - NUM_KERNEL_PDE) *
 284				  NUM_VMUSA_UNIT_PER_PAGE, 0);
 285
 286	m->pt_bo = bo;
 287	return 0;
 288}
 289
 290/*
 291 * Due to workaround 16017236439, odd instance hardware copy engines are
 292 * faster than even instance ones.
 293 * This function returns the mask involving all fast copy engines and the
 294 * reserved copy engine to be used as logical mask for migrate engine.
 295 * Including the reserved copy engine is required to avoid deadlocks due to
 296 * migrate jobs servicing the faults gets stuck behind the job that faulted.
 297 */
 298static u32 xe_migrate_usm_logical_mask(struct xe_gt *gt)
 299{
 300	u32 logical_mask = 0;
 301	struct xe_hw_engine *hwe;
 302	enum xe_hw_engine_id id;
 303
 304	for_each_hw_engine(hwe, gt, id) {
 305		if (hwe->class != XE_ENGINE_CLASS_COPY)
 306			continue;
 307
 308		if (!XE_WA(gt, 16017236439) ||
 309		    xe_gt_is_usm_hwe(gt, hwe) || hwe->instance & 1)
 310			logical_mask |= BIT(hwe->logical_instance);
 311	}
 312
 313	return logical_mask;
 314}
 315
 316/**
 317 * xe_migrate_init() - Initialize a migrate context
 318 * @tile: Back-pointer to the tile we're initializing for.
 319 *
 320 * Return: Pointer to a migrate context on success. Error pointer on error.
 321 */
 322struct xe_migrate *xe_migrate_init(struct xe_tile *tile)
 323{
 324	struct xe_device *xe = tile_to_xe(tile);
 325	struct xe_gt *primary_gt = tile->primary_gt;
 326	struct xe_migrate *m;
 327	struct xe_vm *vm;
 328	int err;
 329
 330	m = drmm_kzalloc(&xe->drm, sizeof(*m), GFP_KERNEL);
 331	if (!m)
 332		return ERR_PTR(-ENOMEM);
 333
 334	m->tile = tile;
 335
 336	/* Special layout, prepared below.. */
 337	vm = xe_vm_create(xe, XE_VM_FLAG_MIGRATION |
 338			  XE_VM_FLAG_SET_TILE_ID(tile));
 339	if (IS_ERR(vm))
 340		return ERR_CAST(vm);
 341
 342	xe_vm_lock(vm, false);
 343	err = xe_migrate_prepare_vm(tile, m, vm);
 344	xe_vm_unlock(vm);
 345	if (err) {
 346		xe_vm_close_and_put(vm);
 347		return ERR_PTR(err);
 348	}
 349
 350	if (xe->info.has_usm) {
 351		struct xe_hw_engine *hwe = xe_gt_hw_engine(primary_gt,
 352							   XE_ENGINE_CLASS_COPY,
 353							   primary_gt->usm.reserved_bcs_instance,
 354							   false);
 355		u32 logical_mask = xe_migrate_usm_logical_mask(primary_gt);
 356
 357		if (!hwe || !logical_mask)
 358			return ERR_PTR(-EINVAL);
 359
 360		m->q = xe_exec_queue_create(xe, vm, logical_mask, 1, hwe,
 361					    EXEC_QUEUE_FLAG_KERNEL |
 362					    EXEC_QUEUE_FLAG_PERMANENT |
 363					    EXEC_QUEUE_FLAG_HIGH_PRIORITY);
 364	} else {
 365		m->q = xe_exec_queue_create_class(xe, primary_gt, vm,
 366						  XE_ENGINE_CLASS_COPY,
 367						  EXEC_QUEUE_FLAG_KERNEL |
 368						  EXEC_QUEUE_FLAG_PERMANENT);
 369	}
 370	if (IS_ERR(m->q)) {
 371		xe_vm_close_and_put(vm);
 372		return ERR_CAST(m->q);
 373	}
 374
 375	mutex_init(&m->job_mutex);
 376
 377	err = drmm_add_action_or_reset(&xe->drm, xe_migrate_fini, m);
 378	if (err)
 379		return ERR_PTR(err);
 380
 381	if (IS_DGFX(xe)) {
 382		if (xe_device_has_flat_ccs(xe))
 383			/* min chunk size corresponds to 4K of CCS Metadata */
 384			m->min_chunk_size = SZ_4K * SZ_64K /
 385				xe_device_ccs_bytes(xe, SZ_64K);
 386		else
 387			/* Somewhat arbitrary to avoid a huge amount of blits */
 388			m->min_chunk_size = SZ_64K;
 389		m->min_chunk_size = roundup_pow_of_two(m->min_chunk_size);
 390		drm_dbg(&xe->drm, "Migrate min chunk size is 0x%08llx\n",
 391			(unsigned long long)m->min_chunk_size);
 392	}
 393
 394	return m;
 395}
 396
 397static u64 max_mem_transfer_per_pass(struct xe_device *xe)
 398{
 399	if (!IS_DGFX(xe) && xe_device_has_flat_ccs(xe))
 400		return MAX_CCS_LIMITED_TRANSFER;
 401
 402	return MAX_PREEMPTDISABLE_TRANSFER;
 403}
 404
 405static u64 xe_migrate_res_sizes(struct xe_migrate *m, struct xe_res_cursor *cur)
 406{
 407	struct xe_device *xe = tile_to_xe(m->tile);
 408	u64 size = min_t(u64, max_mem_transfer_per_pass(xe), cur->remaining);
 409
 410	if (mem_type_is_vram(cur->mem_type)) {
 411		/*
 412		 * VRAM we want to blit in chunks with sizes aligned to
 413		 * min_chunk_size in order for the offset to CCS metadata to be
 414		 * page-aligned. If it's the last chunk it may be smaller.
 415		 *
 416		 * Another constraint is that we need to limit the blit to
 417		 * the VRAM block size, unless size is smaller than
 418		 * min_chunk_size.
 419		 */
 420		u64 chunk = max_t(u64, cur->size, m->min_chunk_size);
 421
 422		size = min_t(u64, size, chunk);
 423		if (size > m->min_chunk_size)
 424			size = round_down(size, m->min_chunk_size);
 425	}
 426
 427	return size;
 428}
 429
 430static bool xe_migrate_allow_identity(u64 size, const struct xe_res_cursor *cur)
 431{
 432	/* If the chunk is not fragmented, allow identity map. */
 433	return cur->size >= size;
 434}
 435
 436static u32 pte_update_size(struct xe_migrate *m,
 437			   bool is_vram,
 438			   struct ttm_resource *res,
 439			   struct xe_res_cursor *cur,
 440			   u64 *L0, u64 *L0_ofs, u32 *L0_pt,
 441			   u32 cmd_size, u32 pt_ofs, u32 avail_pts)
 442{
 443	u32 cmds = 0;
 444
 445	*L0_pt = pt_ofs;
 446	if (is_vram && xe_migrate_allow_identity(*L0, cur)) {
 447		/* Offset into identity map. */
 448		*L0_ofs = xe_migrate_vram_ofs(tile_to_xe(m->tile),
 449					      cur->start + vram_region_gpu_offset(res));
 450		cmds += cmd_size;
 451	} else {
 452		/* Clip L0 to available size */
 453		u64 size = min(*L0, (u64)avail_pts * SZ_2M);
 454		u64 num_4k_pages = DIV_ROUND_UP(size, XE_PAGE_SIZE);
 455
 456		*L0 = size;
 457		*L0_ofs = xe_migrate_vm_addr(pt_ofs, 0);
 458
 459		/* MI_STORE_DATA_IMM */
 460		cmds += 3 * DIV_ROUND_UP(num_4k_pages, 0x1ff);
 461
 462		/* PDE qwords */
 463		cmds += num_4k_pages * 2;
 464
 465		/* Each chunk has a single blit command */
 466		cmds += cmd_size;
 467	}
 468
 469	return cmds;
 470}
 471
 472static void emit_pte(struct xe_migrate *m,
 473		     struct xe_bb *bb, u32 at_pt,
 474		     bool is_vram, bool is_comp_pte,
 475		     struct xe_res_cursor *cur,
 476		     u32 size, struct ttm_resource *res)
 477{
 478	struct xe_device *xe = tile_to_xe(m->tile);
 479	struct xe_vm *vm = m->q->vm;
 480	u16 pat_index;
 481	u32 ptes;
 482	u64 ofs = at_pt * XE_PAGE_SIZE;
 483	u64 cur_ofs;
 484
 485	/* Indirect access needs compression enabled uncached PAT index */
 486	if (GRAPHICS_VERx100(xe) >= 2000)
 487		pat_index = is_comp_pte ? xe->pat.idx[XE_CACHE_NONE_COMPRESSION] :
 488					  xe->pat.idx[XE_CACHE_WB];
 489	else
 490		pat_index = xe->pat.idx[XE_CACHE_WB];
 491
 492	ptes = DIV_ROUND_UP(size, XE_PAGE_SIZE);
 493
 494	while (ptes) {
 495		u32 chunk = min(0x1ffU, ptes);
 496
 497		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(chunk);
 498		bb->cs[bb->len++] = ofs;
 499		bb->cs[bb->len++] = 0;
 500
 501		cur_ofs = ofs;
 502		ofs += chunk * 8;
 503		ptes -= chunk;
 504
 505		while (chunk--) {
 506			u64 addr, flags = 0;
 507			bool devmem = false;
 508
 509			addr = xe_res_dma(cur) & PAGE_MASK;
 510			if (is_vram) {
 511				if (vm->flags & XE_VM_FLAG_64K) {
 512					u64 va = cur_ofs * XE_PAGE_SIZE / 8;
 513
 514					xe_assert(xe, (va & (SZ_64K - 1)) ==
 515						  (addr & (SZ_64K - 1)));
 516
 517					flags |= XE_PTE_PS64;
 518				}
 519
 520				addr += vram_region_gpu_offset(res);
 521				devmem = true;
 522			}
 523
 524			addr = vm->pt_ops->pte_encode_addr(m->tile->xe,
 525							   addr, pat_index,
 526							   0, devmem, flags);
 527			bb->cs[bb->len++] = lower_32_bits(addr);
 528			bb->cs[bb->len++] = upper_32_bits(addr);
 529
 530			xe_res_next(cur, min_t(u32, size, PAGE_SIZE));
 531			cur_ofs += 8;
 532		}
 533	}
 534}
 535
 536#define EMIT_COPY_CCS_DW 5
 537static void emit_copy_ccs(struct xe_gt *gt, struct xe_bb *bb,
 538			  u64 dst_ofs, bool dst_is_indirect,
 539			  u64 src_ofs, bool src_is_indirect,
 540			  u32 size)
 541{
 542	struct xe_device *xe = gt_to_xe(gt);
 543	u32 *cs = bb->cs + bb->len;
 544	u32 num_ccs_blks;
 545	u32 num_pages;
 546	u32 ccs_copy_size;
 547	u32 mocs;
 548
 549	if (GRAPHICS_VERx100(xe) >= 2000) {
 550		num_pages = DIV_ROUND_UP(size, XE_PAGE_SIZE);
 551		xe_gt_assert(gt, FIELD_FIT(XE2_CCS_SIZE_MASK, num_pages - 1));
 552
 553		ccs_copy_size = REG_FIELD_PREP(XE2_CCS_SIZE_MASK, num_pages - 1);
 554		mocs = FIELD_PREP(XE2_XY_CTRL_SURF_MOCS_INDEX_MASK, gt->mocs.uc_index);
 555
 556	} else {
 557		num_ccs_blks = DIV_ROUND_UP(xe_device_ccs_bytes(gt_to_xe(gt), size),
 558					    NUM_CCS_BYTES_PER_BLOCK);
 559		xe_gt_assert(gt, FIELD_FIT(CCS_SIZE_MASK, num_ccs_blks - 1));
 560
 561		ccs_copy_size = REG_FIELD_PREP(CCS_SIZE_MASK, num_ccs_blks - 1);
 562		mocs = FIELD_PREP(XY_CTRL_SURF_MOCS_MASK, gt->mocs.uc_index);
 563	}
 564
 565	*cs++ = XY_CTRL_SURF_COPY_BLT |
 566		(src_is_indirect ? 0x0 : 0x1) << SRC_ACCESS_TYPE_SHIFT |
 567		(dst_is_indirect ? 0x0 : 0x1) << DST_ACCESS_TYPE_SHIFT |
 568		ccs_copy_size;
 569	*cs++ = lower_32_bits(src_ofs);
 570	*cs++ = upper_32_bits(src_ofs) | mocs;
 571	*cs++ = lower_32_bits(dst_ofs);
 572	*cs++ = upper_32_bits(dst_ofs) | mocs;
 573
 574	bb->len = cs - bb->cs;
 575}
 576
 577#define EMIT_COPY_DW 10
 578static void emit_copy(struct xe_gt *gt, struct xe_bb *bb,
 579		      u64 src_ofs, u64 dst_ofs, unsigned int size,
 580		      unsigned int pitch)
 581{
 582	struct xe_device *xe = gt_to_xe(gt);
 583	u32 mocs = 0;
 584	u32 tile_y = 0;
 585
 586	xe_gt_assert(gt, size / pitch <= S16_MAX);
 587	xe_gt_assert(gt, pitch / 4 <= S16_MAX);
 588	xe_gt_assert(gt, pitch <= U16_MAX);
 589
 590	if (GRAPHICS_VER(xe) >= 20)
 591		mocs = FIELD_PREP(XE2_XY_FAST_COPY_BLT_MOCS_INDEX_MASK, gt->mocs.uc_index);
 592
 593	if (GRAPHICS_VERx100(xe) >= 1250)
 594		tile_y = XY_FAST_COPY_BLT_D1_SRC_TILE4 | XY_FAST_COPY_BLT_D1_DST_TILE4;
 595
 596	bb->cs[bb->len++] = XY_FAST_COPY_BLT_CMD | (10 - 2);
 597	bb->cs[bb->len++] = XY_FAST_COPY_BLT_DEPTH_32 | pitch | tile_y | mocs;
 598	bb->cs[bb->len++] = 0;
 599	bb->cs[bb->len++] = (size / pitch) << 16 | pitch / 4;
 600	bb->cs[bb->len++] = lower_32_bits(dst_ofs);
 601	bb->cs[bb->len++] = upper_32_bits(dst_ofs);
 602	bb->cs[bb->len++] = 0;
 603	bb->cs[bb->len++] = pitch | mocs;
 604	bb->cs[bb->len++] = lower_32_bits(src_ofs);
 605	bb->cs[bb->len++] = upper_32_bits(src_ofs);
 606}
 607
 608static int job_add_deps(struct xe_sched_job *job, struct dma_resv *resv,
 609			enum dma_resv_usage usage)
 610{
 611	return drm_sched_job_add_resv_dependencies(&job->drm, resv, usage);
 612}
 613
 614static u64 xe_migrate_batch_base(struct xe_migrate *m, bool usm)
 615{
 616	return usm ? m->usm_batch_base_ofs : m->batch_base_ofs;
 617}
 618
 619static u32 xe_migrate_ccs_copy(struct xe_migrate *m,
 620			       struct xe_bb *bb,
 621			       u64 src_ofs, bool src_is_indirect,
 622			       u64 dst_ofs, bool dst_is_indirect, u32 dst_size,
 623			       u64 ccs_ofs, bool copy_ccs)
 624{
 625	struct xe_gt *gt = m->tile->primary_gt;
 626	u32 flush_flags = 0;
 627
 628	if (xe_device_has_flat_ccs(gt_to_xe(gt)) && !copy_ccs && dst_is_indirect) {
 629		/*
 630		 * If the src is already in vram, then it should already
 631		 * have been cleared by us, or has been populated by the
 632		 * user. Make sure we copy the CCS aux state as-is.
 633		 *
 634		 * Otherwise if the bo doesn't have any CCS metadata attached,
 635		 * we still need to clear it for security reasons.
 636		 */
 637		u64 ccs_src_ofs =  src_is_indirect ? src_ofs : m->cleared_mem_ofs;
 638
 639		emit_copy_ccs(gt, bb,
 640			      dst_ofs, true,
 641			      ccs_src_ofs, src_is_indirect, dst_size);
 642
 643		flush_flags = MI_FLUSH_DW_CCS;
 644	} else if (copy_ccs) {
 645		if (!src_is_indirect)
 646			src_ofs = ccs_ofs;
 647		else if (!dst_is_indirect)
 648			dst_ofs = ccs_ofs;
 649
 650		xe_gt_assert(gt, src_is_indirect || dst_is_indirect);
 651
 652		emit_copy_ccs(gt, bb, dst_ofs, dst_is_indirect, src_ofs,
 653			      src_is_indirect, dst_size);
 654		if (dst_is_indirect)
 655			flush_flags = MI_FLUSH_DW_CCS;
 656	}
 657
 658	return flush_flags;
 659}
 660
 661/**
 662 * xe_migrate_copy() - Copy content of TTM resources.
 663 * @m: The migration context.
 664 * @src_bo: The buffer object @src is currently bound to.
 665 * @dst_bo: If copying between resources created for the same bo, set this to
 666 * the same value as @src_bo. If copying between buffer objects, set it to
 667 * the buffer object @dst is currently bound to.
 668 * @src: The source TTM resource.
 669 * @dst: The dst TTM resource.
 670 * @copy_only_ccs: If true copy only CCS metadata
 671 *
 672 * Copies the contents of @src to @dst: On flat CCS devices,
 673 * the CCS metadata is copied as well if needed, or if not present,
 674 * the CCS metadata of @dst is cleared for security reasons.
 675 *
 676 * Return: Pointer to a dma_fence representing the last copy batch, or
 677 * an error pointer on failure. If there is a failure, any copy operation
 678 * started by the function call has been synced.
 679 */
 680struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
 681				  struct xe_bo *src_bo,
 682				  struct xe_bo *dst_bo,
 683				  struct ttm_resource *src,
 684				  struct ttm_resource *dst,
 685				  bool copy_only_ccs)
 686{
 687	struct xe_gt *gt = m->tile->primary_gt;
 688	struct xe_device *xe = gt_to_xe(gt);
 689	struct dma_fence *fence = NULL;
 690	u64 size = src_bo->size;
 691	struct xe_res_cursor src_it, dst_it, ccs_it;
 692	u64 src_L0_ofs, dst_L0_ofs;
 693	u32 src_L0_pt, dst_L0_pt;
 694	u64 src_L0, dst_L0;
 695	int pass = 0;
 696	int err;
 697	bool src_is_pltt = src->mem_type == XE_PL_TT;
 698	bool dst_is_pltt = dst->mem_type == XE_PL_TT;
 699	bool src_is_vram = mem_type_is_vram(src->mem_type);
 700	bool dst_is_vram = mem_type_is_vram(dst->mem_type);
 701	bool copy_ccs = xe_device_has_flat_ccs(xe) &&
 702		xe_bo_needs_ccs_pages(src_bo) && xe_bo_needs_ccs_pages(dst_bo);
 703	bool copy_system_ccs = copy_ccs && (!src_is_vram || !dst_is_vram);
 704
 705	/* Copying CCS between two different BOs is not supported yet. */
 706	if (XE_WARN_ON(copy_ccs && src_bo != dst_bo))
 707		return ERR_PTR(-EINVAL);
 708
 709	if (src_bo != dst_bo && XE_WARN_ON(src_bo->size != dst_bo->size))
 710		return ERR_PTR(-EINVAL);
 711
 712	if (!src_is_vram)
 713		xe_res_first_sg(xe_bo_sg(src_bo), 0, size, &src_it);
 714	else
 715		xe_res_first(src, 0, size, &src_it);
 716	if (!dst_is_vram)
 717		xe_res_first_sg(xe_bo_sg(dst_bo), 0, size, &dst_it);
 718	else
 719		xe_res_first(dst, 0, size, &dst_it);
 720
 721	if (copy_system_ccs)
 722		xe_res_first_sg(xe_bo_sg(src_bo), xe_bo_ccs_pages_start(src_bo),
 723				PAGE_ALIGN(xe_device_ccs_bytes(xe, size)),
 724				&ccs_it);
 725
 726	while (size) {
 727		u32 batch_size = 2; /* arb_clear() + MI_BATCH_BUFFER_END */
 728		struct xe_sched_job *job;
 729		struct xe_bb *bb;
 730		u32 flush_flags;
 731		u32 update_idx;
 732		u64 ccs_ofs, ccs_size;
 733		u32 ccs_pt;
 734
 735		bool usm = xe->info.has_usm;
 736		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
 737
 738		src_L0 = xe_migrate_res_sizes(m, &src_it);
 739		dst_L0 = xe_migrate_res_sizes(m, &dst_it);
 740
 741		drm_dbg(&xe->drm, "Pass %u, sizes: %llu & %llu\n",
 742			pass++, src_L0, dst_L0);
 743
 744		src_L0 = min(src_L0, dst_L0);
 745
 746		batch_size += pte_update_size(m, src_is_vram, src, &src_it, &src_L0,
 747					      &src_L0_ofs, &src_L0_pt, 0, 0,
 748					      avail_pts);
 749
 750		batch_size += pte_update_size(m, dst_is_vram, dst, &dst_it, &src_L0,
 751					      &dst_L0_ofs, &dst_L0_pt, 0,
 752					      avail_pts, avail_pts);
 753
 754		if (copy_system_ccs) {
 755			ccs_size = xe_device_ccs_bytes(xe, src_L0);
 756			batch_size += pte_update_size(m, false, NULL, &ccs_it, &ccs_size,
 757						      &ccs_ofs, &ccs_pt, 0,
 758						      2 * avail_pts,
 759						      avail_pts);
 760			xe_assert(xe, IS_ALIGNED(ccs_it.start, PAGE_SIZE));
 761		}
 762
 763		/* Add copy commands size here */
 764		batch_size += ((copy_only_ccs) ? 0 : EMIT_COPY_DW) +
 765			((xe_device_has_flat_ccs(xe) ? EMIT_COPY_CCS_DW : 0));
 766
 767		bb = xe_bb_new(gt, batch_size, usm);
 768		if (IS_ERR(bb)) {
 769			err = PTR_ERR(bb);
 770			goto err_sync;
 771		}
 772
 773		if (src_is_vram && xe_migrate_allow_identity(src_L0, &src_it))
 774			xe_res_next(&src_it, src_L0);
 775		else
 776			emit_pte(m, bb, src_L0_pt, src_is_vram, copy_system_ccs,
 777				 &src_it, src_L0, src);
 778
 779		if (dst_is_vram && xe_migrate_allow_identity(src_L0, &dst_it))
 780			xe_res_next(&dst_it, src_L0);
 781		else
 782			emit_pte(m, bb, dst_L0_pt, dst_is_vram, copy_system_ccs,
 783				 &dst_it, src_L0, dst);
 784
 785		if (copy_system_ccs)
 786			emit_pte(m, bb, ccs_pt, false, false, &ccs_it, ccs_size, src);
 787
 788		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
 789		update_idx = bb->len;
 790
 791		if (!copy_only_ccs)
 792			emit_copy(gt, bb, src_L0_ofs, dst_L0_ofs, src_L0, XE_PAGE_SIZE);
 793
 794		flush_flags = xe_migrate_ccs_copy(m, bb, src_L0_ofs,
 795						  IS_DGFX(xe) ? src_is_vram : src_is_pltt,
 796						  dst_L0_ofs,
 797						  IS_DGFX(xe) ? dst_is_vram : dst_is_pltt,
 798						  src_L0, ccs_ofs, copy_ccs);
 799
 800		mutex_lock(&m->job_mutex);
 801		job = xe_bb_create_migration_job(m->q, bb,
 802						 xe_migrate_batch_base(m, usm),
 803						 update_idx);
 804		if (IS_ERR(job)) {
 805			err = PTR_ERR(job);
 806			goto err;
 807		}
 808
 809		xe_sched_job_add_migrate_flush(job, flush_flags);
 810		if (!fence) {
 811			err = job_add_deps(job, src_bo->ttm.base.resv,
 812					   DMA_RESV_USAGE_BOOKKEEP);
 813			if (!err && src_bo != dst_bo)
 814				err = job_add_deps(job, dst_bo->ttm.base.resv,
 815						   DMA_RESV_USAGE_BOOKKEEP);
 816			if (err)
 817				goto err_job;
 818		}
 819
 820		xe_sched_job_arm(job);
 821		dma_fence_put(fence);
 822		fence = dma_fence_get(&job->drm.s_fence->finished);
 823		xe_sched_job_push(job);
 824
 825		dma_fence_put(m->fence);
 826		m->fence = dma_fence_get(fence);
 827
 828		mutex_unlock(&m->job_mutex);
 829
 830		xe_bb_free(bb, fence);
 831		size -= src_L0;
 832		continue;
 833
 834err_job:
 835		xe_sched_job_put(job);
 836err:
 837		mutex_unlock(&m->job_mutex);
 838		xe_bb_free(bb, NULL);
 839
 840err_sync:
 841		/* Sync partial copy if any. FIXME: under job_mutex? */
 842		if (fence) {
 843			dma_fence_wait(fence, false);
 844			dma_fence_put(fence);
 845		}
 846
 847		return ERR_PTR(err);
 848	}
 849
 850	return fence;
 851}
 852
 853static void emit_clear_link_copy(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
 854				 u32 size, u32 pitch)
 855{
 856	struct xe_device *xe = gt_to_xe(gt);
 857	u32 *cs = bb->cs + bb->len;
 858	u32 len = PVC_MEM_SET_CMD_LEN_DW;
 859
 860	*cs++ = PVC_MEM_SET_CMD | PVC_MEM_SET_MATRIX | (len - 2);
 861	*cs++ = pitch - 1;
 862	*cs++ = (size / pitch) - 1;
 863	*cs++ = pitch - 1;
 864	*cs++ = lower_32_bits(src_ofs);
 865	*cs++ = upper_32_bits(src_ofs);
 866	if (GRAPHICS_VERx100(xe) >= 2000)
 867		*cs++ = FIELD_PREP(XE2_MEM_SET_MOCS_INDEX_MASK, gt->mocs.uc_index);
 868	else
 869		*cs++ = FIELD_PREP(PVC_MEM_SET_MOCS_INDEX_MASK, gt->mocs.uc_index);
 870
 871	xe_gt_assert(gt, cs - bb->cs == len + bb->len);
 872
 873	bb->len += len;
 874}
 875
 876static void emit_clear_main_copy(struct xe_gt *gt, struct xe_bb *bb,
 877				 u64 src_ofs, u32 size, u32 pitch, bool is_vram)
 878{
 879	struct xe_device *xe = gt_to_xe(gt);
 880	u32 *cs = bb->cs + bb->len;
 881	u32 len = XY_FAST_COLOR_BLT_DW;
 882
 883	if (GRAPHICS_VERx100(xe) < 1250)
 884		len = 11;
 885
 886	*cs++ = XY_FAST_COLOR_BLT_CMD | XY_FAST_COLOR_BLT_DEPTH_32 |
 887		(len - 2);
 888	if (GRAPHICS_VERx100(xe) >= 2000)
 889		*cs++ = FIELD_PREP(XE2_XY_FAST_COLOR_BLT_MOCS_INDEX_MASK, gt->mocs.uc_index) |
 890			(pitch - 1);
 891	else
 892		*cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, gt->mocs.uc_index) |
 893			(pitch - 1);
 894	*cs++ = 0;
 895	*cs++ = (size / pitch) << 16 | pitch / 4;
 896	*cs++ = lower_32_bits(src_ofs);
 897	*cs++ = upper_32_bits(src_ofs);
 898	*cs++ = (is_vram ? 0x0 : 0x1) <<  XY_FAST_COLOR_BLT_MEM_TYPE_SHIFT;
 899	*cs++ = 0;
 900	*cs++ = 0;
 901	*cs++ = 0;
 902	*cs++ = 0;
 903
 904	if (len > 11) {
 905		*cs++ = 0;
 906		*cs++ = 0;
 907		*cs++ = 0;
 908		*cs++ = 0;
 909		*cs++ = 0;
 910	}
 911
 912	xe_gt_assert(gt, cs - bb->cs == len + bb->len);
 913
 914	bb->len += len;
 915}
 916
 917static bool has_service_copy_support(struct xe_gt *gt)
 918{
 919	/*
 920	 * What we care about is whether the architecture was designed with
 921	 * service copy functionality (specifically the new MEM_SET / MEM_COPY
 922	 * instructions) so check the architectural engine list rather than the
 923	 * actual list since these instructions are usable on BCS0 even if
 924	 * all of the actual service copy engines (BCS1-BCS8) have been fused
 925	 * off.
 926	 */
 927	return gt->info.__engine_mask & GENMASK(XE_HW_ENGINE_BCS8,
 928						XE_HW_ENGINE_BCS1);
 929}
 930
 931static u32 emit_clear_cmd_len(struct xe_gt *gt)
 932{
 933	if (has_service_copy_support(gt))
 934		return PVC_MEM_SET_CMD_LEN_DW;
 935	else
 936		return XY_FAST_COLOR_BLT_DW;
 937}
 938
 939static void emit_clear(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
 940		       u32 size, u32 pitch, bool is_vram)
 941{
 942	if (has_service_copy_support(gt))
 943		emit_clear_link_copy(gt, bb, src_ofs, size, pitch);
 944	else
 945		emit_clear_main_copy(gt, bb, src_ofs, size, pitch,
 946				     is_vram);
 947}
 948
 949/**
 950 * xe_migrate_clear() - Copy content of TTM resources.
 951 * @m: The migration context.
 952 * @bo: The buffer object @dst is currently bound to.
 953 * @dst: The dst TTM resource to be cleared.
 954 *
 955 * Clear the contents of @dst to zero. On flat CCS devices,
 956 * the CCS metadata is cleared to zero as well on VRAM destinations.
 957 * TODO: Eliminate the @bo argument.
 958 *
 959 * Return: Pointer to a dma_fence representing the last clear batch, or
 960 * an error pointer on failure. If there is a failure, any clear operation
 961 * started by the function call has been synced.
 962 */
 963struct dma_fence *xe_migrate_clear(struct xe_migrate *m,
 964				   struct xe_bo *bo,
 965				   struct ttm_resource *dst)
 966{
 967	bool clear_vram = mem_type_is_vram(dst->mem_type);
 968	struct xe_gt *gt = m->tile->primary_gt;
 969	struct xe_device *xe = gt_to_xe(gt);
 970	bool clear_system_ccs = (xe_bo_needs_ccs_pages(bo) && !IS_DGFX(xe)) ? true : false;
 971	struct dma_fence *fence = NULL;
 972	u64 size = bo->size;
 973	struct xe_res_cursor src_it;
 974	struct ttm_resource *src = dst;
 975	int err;
 976	int pass = 0;
 977
 978	if (!clear_vram)
 979		xe_res_first_sg(xe_bo_sg(bo), 0, bo->size, &src_it);
 980	else
 981		xe_res_first(src, 0, bo->size, &src_it);
 982
 983	while (size) {
 984		u64 clear_L0_ofs;
 985		u32 clear_L0_pt;
 986		u32 flush_flags = 0;
 987		u64 clear_L0;
 988		struct xe_sched_job *job;
 989		struct xe_bb *bb;
 990		u32 batch_size, update_idx;
 991
 992		bool usm = xe->info.has_usm;
 993		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
 994
 995		clear_L0 = xe_migrate_res_sizes(m, &src_it);
 996
 997		drm_dbg(&xe->drm, "Pass %u, size: %llu\n", pass++, clear_L0);
 998
 999		/* Calculate final sizes and batch size.. */
1000		batch_size = 2 +
1001			pte_update_size(m, clear_vram, src, &src_it,
1002					&clear_L0, &clear_L0_ofs, &clear_L0_pt,
1003					clear_system_ccs ? 0 : emit_clear_cmd_len(gt), 0,
1004					avail_pts);
1005
1006		if (xe_device_has_flat_ccs(xe))
1007			batch_size += EMIT_COPY_CCS_DW;
1008
1009		/* Clear commands */
1010
1011		if (WARN_ON_ONCE(!clear_L0))
1012			break;
1013
1014		bb = xe_bb_new(gt, batch_size, usm);
1015		if (IS_ERR(bb)) {
1016			err = PTR_ERR(bb);
1017			goto err_sync;
1018		}
1019
1020		size -= clear_L0;
1021		/* Preemption is enabled again by the ring ops. */
1022		if (clear_vram && xe_migrate_allow_identity(clear_L0, &src_it))
1023			xe_res_next(&src_it, clear_L0);
1024		else
1025			emit_pte(m, bb, clear_L0_pt, clear_vram, clear_system_ccs,
1026				 &src_it, clear_L0, dst);
1027
1028		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1029		update_idx = bb->len;
1030
1031		if (!clear_system_ccs)
1032			emit_clear(gt, bb, clear_L0_ofs, clear_L0, XE_PAGE_SIZE, clear_vram);
1033
1034		if (xe_device_has_flat_ccs(xe)) {
1035			emit_copy_ccs(gt, bb, clear_L0_ofs, true,
1036				      m->cleared_mem_ofs, false, clear_L0);
1037			flush_flags = MI_FLUSH_DW_CCS;
1038		}
1039
1040		mutex_lock(&m->job_mutex);
1041		job = xe_bb_create_migration_job(m->q, bb,
1042						 xe_migrate_batch_base(m, usm),
1043						 update_idx);
1044		if (IS_ERR(job)) {
1045			err = PTR_ERR(job);
1046			goto err;
1047		}
1048
1049		xe_sched_job_add_migrate_flush(job, flush_flags);
1050		if (!fence) {
1051			/*
1052			 * There can't be anything userspace related at this
1053			 * point, so we just need to respect any potential move
1054			 * fences, which are always tracked as
1055			 * DMA_RESV_USAGE_KERNEL.
1056			 */
1057			err = job_add_deps(job, bo->ttm.base.resv,
1058					   DMA_RESV_USAGE_KERNEL);
1059			if (err)
1060				goto err_job;
1061		}
1062
1063		xe_sched_job_arm(job);
1064		dma_fence_put(fence);
1065		fence = dma_fence_get(&job->drm.s_fence->finished);
1066		xe_sched_job_push(job);
1067
1068		dma_fence_put(m->fence);
1069		m->fence = dma_fence_get(fence);
1070
1071		mutex_unlock(&m->job_mutex);
1072
1073		xe_bb_free(bb, fence);
1074		continue;
1075
1076err_job:
1077		xe_sched_job_put(job);
1078err:
1079		mutex_unlock(&m->job_mutex);
1080		xe_bb_free(bb, NULL);
1081err_sync:
1082		/* Sync partial copies if any. FIXME: job_mutex? */
1083		if (fence) {
1084			dma_fence_wait(m->fence, false);
1085			dma_fence_put(fence);
1086		}
1087
1088		return ERR_PTR(err);
1089	}
1090
1091	if (clear_system_ccs)
1092		bo->ccs_cleared = true;
1093
1094	return fence;
1095}
1096
1097static void write_pgtable(struct xe_tile *tile, struct xe_bb *bb, u64 ppgtt_ofs,
1098			  const struct xe_vm_pgtable_update *update,
1099			  struct xe_migrate_pt_update *pt_update)
1100{
1101	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1102	u32 chunk;
1103	u32 ofs = update->ofs, size = update->qwords;
1104
1105	/*
1106	 * If we have 512 entries (max), we would populate it ourselves,
1107	 * and update the PDE above it to the new pointer.
1108	 * The only time this can only happen if we have to update the top
1109	 * PDE. This requires a BO that is almost vm->size big.
1110	 *
1111	 * This shouldn't be possible in practice.. might change when 16K
1112	 * pages are used. Hence the assert.
1113	 */
1114	xe_tile_assert(tile, update->qwords <= 0x1ff);
1115	if (!ppgtt_ofs)
1116		ppgtt_ofs = xe_migrate_vram_ofs(tile_to_xe(tile),
1117						xe_bo_addr(update->pt_bo, 0,
1118							   XE_PAGE_SIZE));
1119
1120	do {
1121		u64 addr = ppgtt_ofs + ofs * 8;
1122
1123		chunk = min(update->qwords, 0x1ffU);
1124
1125		/* Ensure populatefn can do memset64 by aligning bb->cs */
1126		if (!(bb->len & 1))
1127			bb->cs[bb->len++] = MI_NOOP;
1128
1129		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(chunk);
1130		bb->cs[bb->len++] = lower_32_bits(addr);
1131		bb->cs[bb->len++] = upper_32_bits(addr);
1132		ops->populate(pt_update, tile, NULL, bb->cs + bb->len, ofs, chunk,
1133			      update);
1134
1135		bb->len += chunk * 2;
1136		ofs += chunk;
1137		size -= chunk;
1138	} while (size);
1139}
1140
1141struct xe_vm *xe_migrate_get_vm(struct xe_migrate *m)
1142{
1143	return xe_vm_get(m->q->vm);
1144}
1145
1146#if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
1147struct migrate_test_params {
1148	struct xe_test_priv base;
1149	bool force_gpu;
1150};
1151
1152#define to_migrate_test_params(_priv) \
1153	container_of(_priv, struct migrate_test_params, base)
1154#endif
1155
1156static struct dma_fence *
1157xe_migrate_update_pgtables_cpu(struct xe_migrate *m,
1158			       struct xe_vm *vm, struct xe_bo *bo,
1159			       const struct  xe_vm_pgtable_update *updates,
1160			       u32 num_updates, bool wait_vm,
1161			       struct xe_migrate_pt_update *pt_update)
1162{
1163	XE_TEST_DECLARE(struct migrate_test_params *test =
1164			to_migrate_test_params
1165			(xe_cur_kunit_priv(XE_TEST_LIVE_MIGRATE));)
1166	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1167	struct dma_fence *fence;
1168	int err;
1169	u32 i;
1170
1171	if (XE_TEST_ONLY(test && test->force_gpu))
1172		return ERR_PTR(-ETIME);
1173
1174	if (bo && !dma_resv_test_signaled(bo->ttm.base.resv,
1175					  DMA_RESV_USAGE_KERNEL))
1176		return ERR_PTR(-ETIME);
1177
1178	if (wait_vm && !dma_resv_test_signaled(xe_vm_resv(vm),
1179					       DMA_RESV_USAGE_BOOKKEEP))
1180		return ERR_PTR(-ETIME);
1181
1182	if (ops->pre_commit) {
1183		pt_update->job = NULL;
1184		err = ops->pre_commit(pt_update);
1185		if (err)
1186			return ERR_PTR(err);
1187	}
1188	for (i = 0; i < num_updates; i++) {
1189		const struct xe_vm_pgtable_update *update = &updates[i];
1190
1191		ops->populate(pt_update, m->tile, &update->pt_bo->vmap, NULL,
1192			      update->ofs, update->qwords, update);
1193	}
1194
1195	if (vm) {
1196		trace_xe_vm_cpu_bind(vm);
1197		xe_device_wmb(vm->xe);
1198	}
1199
1200	fence = dma_fence_get_stub();
1201
1202	return fence;
1203}
1204
1205static bool no_in_syncs(struct xe_vm *vm, struct xe_exec_queue *q,
1206			struct xe_sync_entry *syncs, u32 num_syncs)
1207{
1208	struct dma_fence *fence;
1209	int i;
1210
1211	for (i = 0; i < num_syncs; i++) {
1212		fence = syncs[i].fence;
1213
1214		if (fence && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1215				       &fence->flags))
1216			return false;
1217	}
1218	if (q) {
1219		fence = xe_exec_queue_last_fence_get(q, vm);
1220		if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
1221			dma_fence_put(fence);
1222			return false;
1223		}
1224		dma_fence_put(fence);
1225	}
1226
1227	return true;
1228}
1229
1230/**
1231 * xe_migrate_update_pgtables() - Pipelined page-table update
1232 * @m: The migrate context.
1233 * @vm: The vm we'll be updating.
1234 * @bo: The bo whose dma-resv we will await before updating, or NULL if userptr.
1235 * @q: The exec queue to be used for the update or NULL if the default
1236 * migration engine is to be used.
1237 * @updates: An array of update descriptors.
1238 * @num_updates: Number of descriptors in @updates.
1239 * @syncs: Array of xe_sync_entry to await before updating. Note that waits
1240 * will block the engine timeline.
1241 * @num_syncs: Number of entries in @syncs.
1242 * @pt_update: Pointer to a struct xe_migrate_pt_update, which contains
1243 * pointers to callback functions and, if subclassed, private arguments to
1244 * those.
1245 *
1246 * Perform a pipelined page-table update. The update descriptors are typically
1247 * built under the same lock critical section as a call to this function. If
1248 * using the default engine for the updates, they will be performed in the
1249 * order they grab the job_mutex. If different engines are used, external
1250 * synchronization is needed for overlapping updates to maintain page-table
1251 * consistency. Note that the meaing of "overlapping" is that the updates
1252 * touch the same page-table, which might be a higher-level page-directory.
1253 * If no pipelining is needed, then updates may be performed by the cpu.
1254 *
1255 * Return: A dma_fence that, when signaled, indicates the update completion.
1256 */
1257struct dma_fence *
1258xe_migrate_update_pgtables(struct xe_migrate *m,
1259			   struct xe_vm *vm,
1260			   struct xe_bo *bo,
1261			   struct xe_exec_queue *q,
1262			   const struct xe_vm_pgtable_update *updates,
1263			   u32 num_updates,
1264			   struct xe_sync_entry *syncs, u32 num_syncs,
1265			   struct xe_migrate_pt_update *pt_update)
1266{
1267	const struct xe_migrate_pt_update_ops *ops = pt_update->ops;
1268	struct xe_tile *tile = m->tile;
1269	struct xe_gt *gt = tile->primary_gt;
1270	struct xe_device *xe = tile_to_xe(tile);
1271	struct xe_sched_job *job;
1272	struct dma_fence *fence;
1273	struct drm_suballoc *sa_bo = NULL;
1274	struct xe_vma *vma = pt_update->vma;
1275	struct xe_bb *bb;
1276	u32 i, batch_size, ppgtt_ofs, update_idx, page_ofs = 0;
1277	u64 addr;
1278	int err = 0;
1279	bool usm = !q && xe->info.has_usm;
1280	bool first_munmap_rebind = vma &&
1281		vma->gpuva.flags & XE_VMA_FIRST_REBIND;
1282	struct xe_exec_queue *q_override = !q ? m->q : q;
1283	u16 pat_index = xe->pat.idx[XE_CACHE_WB];
1284
1285	/* Use the CPU if no in syncs and engine is idle */
1286	if (no_in_syncs(vm, q, syncs, num_syncs) && xe_exec_queue_is_idle(q_override)) {
1287		fence =  xe_migrate_update_pgtables_cpu(m, vm, bo, updates,
1288							num_updates,
1289							first_munmap_rebind,
1290							pt_update);
1291		if (!IS_ERR(fence) || fence == ERR_PTR(-EAGAIN))
1292			return fence;
1293	}
1294
1295	/* fixed + PTE entries */
1296	if (IS_DGFX(xe))
1297		batch_size = 2;
1298	else
1299		batch_size = 6 + num_updates * 2;
1300
1301	for (i = 0; i < num_updates; i++) {
1302		u32 num_cmds = DIV_ROUND_UP(updates[i].qwords, 0x1ff);
1303
1304		/* align noop + MI_STORE_DATA_IMM cmd prefix */
1305		batch_size += 4 * num_cmds + updates[i].qwords * 2;
1306	}
1307
1308	/*
1309	 * XXX: Create temp bo to copy from, if batch_size becomes too big?
1310	 *
1311	 * Worst case: Sum(2 * (each lower level page size) + (top level page size))
1312	 * Should be reasonably bound..
1313	 */
1314	xe_tile_assert(tile, batch_size < SZ_128K);
1315
1316	bb = xe_bb_new(gt, batch_size, !q && xe->info.has_usm);
1317	if (IS_ERR(bb))
1318		return ERR_CAST(bb);
1319
1320	/* For sysmem PTE's, need to map them in our hole.. */
1321	if (!IS_DGFX(xe)) {
1322		ppgtt_ofs = NUM_KERNEL_PDE - 1;
1323		if (q) {
1324			xe_tile_assert(tile, num_updates <= NUM_VMUSA_WRITES_PER_UNIT);
1325
1326			sa_bo = drm_suballoc_new(&m->vm_update_sa, 1,
1327						 GFP_KERNEL, true, 0);
1328			if (IS_ERR(sa_bo)) {
1329				err = PTR_ERR(sa_bo);
1330				goto err;
1331			}
1332
1333			ppgtt_ofs = NUM_KERNEL_PDE +
1334				(drm_suballoc_soffset(sa_bo) /
1335				 NUM_VMUSA_UNIT_PER_PAGE);
1336			page_ofs = (drm_suballoc_soffset(sa_bo) %
1337				    NUM_VMUSA_UNIT_PER_PAGE) *
1338				VM_SA_UPDATE_UNIT_SIZE;
1339		}
1340
1341		/* Map our PT's to gtt */
1342		bb->cs[bb->len++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(num_updates);
1343		bb->cs[bb->len++] = ppgtt_ofs * XE_PAGE_SIZE + page_ofs;
1344		bb->cs[bb->len++] = 0; /* upper_32_bits */
1345
1346		for (i = 0; i < num_updates; i++) {
1347			struct xe_bo *pt_bo = updates[i].pt_bo;
1348
1349			xe_tile_assert(tile, pt_bo->size == SZ_4K);
1350
1351			addr = vm->pt_ops->pte_encode_bo(pt_bo, 0, pat_index, 0);
1352			bb->cs[bb->len++] = lower_32_bits(addr);
1353			bb->cs[bb->len++] = upper_32_bits(addr);
1354		}
1355
1356		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1357		update_idx = bb->len;
1358
1359		addr = xe_migrate_vm_addr(ppgtt_ofs, 0) +
1360			(page_ofs / sizeof(u64)) * XE_PAGE_SIZE;
1361		for (i = 0; i < num_updates; i++)
1362			write_pgtable(tile, bb, addr + i * XE_PAGE_SIZE,
1363				      &updates[i], pt_update);
1364	} else {
1365		/* phys pages, no preamble required */
1366		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
1367		update_idx = bb->len;
1368
1369		for (i = 0; i < num_updates; i++)
1370			write_pgtable(tile, bb, 0, &updates[i], pt_update);
1371	}
1372
1373	if (!q)
1374		mutex_lock(&m->job_mutex);
1375
1376	job = xe_bb_create_migration_job(q ?: m->q, bb,
1377					 xe_migrate_batch_base(m, usm),
1378					 update_idx);
1379	if (IS_ERR(job)) {
1380		err = PTR_ERR(job);
1381		goto err_bb;
1382	}
1383
1384	/* Wait on BO move */
1385	if (bo) {
1386		err = job_add_deps(job, bo->ttm.base.resv,
1387				   DMA_RESV_USAGE_KERNEL);
1388		if (err)
1389			goto err_job;
1390	}
1391
1392	/*
1393	 * Munmap style VM unbind, need to wait for all jobs to be complete /
1394	 * trigger preempts before moving forward
1395	 */
1396	if (first_munmap_rebind) {
1397		err = job_add_deps(job, xe_vm_resv(vm),
1398				   DMA_RESV_USAGE_BOOKKEEP);
1399		if (err)
1400			goto err_job;
1401	}
1402
1403	err = xe_sched_job_last_fence_add_dep(job, vm);
1404	for (i = 0; !err && i < num_syncs; i++)
1405		err = xe_sync_entry_add_deps(&syncs[i], job);
1406
1407	if (err)
1408		goto err_job;
1409
1410	if (ops->pre_commit) {
1411		pt_update->job = job;
1412		err = ops->pre_commit(pt_update);
1413		if (err)
1414			goto err_job;
1415	}
1416	xe_sched_job_arm(job);
1417	fence = dma_fence_get(&job->drm.s_fence->finished);
1418	xe_sched_job_push(job);
1419
1420	if (!q)
1421		mutex_unlock(&m->job_mutex);
1422
1423	xe_bb_free(bb, fence);
1424	drm_suballoc_free(sa_bo, fence);
1425
1426	return fence;
1427
1428err_job:
1429	xe_sched_job_put(job);
1430err_bb:
1431	if (!q)
1432		mutex_unlock(&m->job_mutex);
1433	xe_bb_free(bb, NULL);
1434err:
1435	drm_suballoc_free(sa_bo, NULL);
1436	return ERR_PTR(err);
1437}
1438
1439/**
1440 * xe_migrate_wait() - Complete all operations using the xe_migrate context
1441 * @m: Migrate context to wait for.
1442 *
1443 * Waits until the GPU no longer uses the migrate context's default engine
1444 * or its page-table objects. FIXME: What about separate page-table update
1445 * engines?
1446 */
1447void xe_migrate_wait(struct xe_migrate *m)
1448{
1449	if (m->fence)
1450		dma_fence_wait(m->fence, false);
1451}
1452
1453#if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
1454#include "tests/xe_migrate.c"
1455#endif