Linux Audio

Check our new training course

Loading...
Note: File does not exist in v4.6.
   1// SPDX-License-Identifier: MIT
   2/*
   3 * Copyright © 2022 Intel Corporation
   4 */
   5
   6#include "xe_pt.h"
   7
   8#include "xe_bo.h"
   9#include "xe_device.h"
  10#include "xe_drm_client.h"
  11#include "xe_gt.h"
  12#include "xe_gt_tlb_invalidation.h"
  13#include "xe_migrate.h"
  14#include "xe_pt_types.h"
  15#include "xe_pt_walk.h"
  16#include "xe_res_cursor.h"
  17#include "xe_trace.h"
  18#include "xe_ttm_stolen_mgr.h"
  19#include "xe_vm.h"
  20
  21struct xe_pt_dir {
  22	struct xe_pt pt;
  23	/** @children: Array of page-table child nodes */
  24	struct xe_ptw *children[XE_PDES];
  25};
  26
  27#if IS_ENABLED(CONFIG_DRM_XE_DEBUG_VM)
  28#define xe_pt_set_addr(__xe_pt, __addr) ((__xe_pt)->addr = (__addr))
  29#define xe_pt_addr(__xe_pt) ((__xe_pt)->addr)
  30#else
  31#define xe_pt_set_addr(__xe_pt, __addr)
  32#define xe_pt_addr(__xe_pt) 0ull
  33#endif
  34
  35static const u64 xe_normal_pt_shifts[] = {12, 21, 30, 39, 48};
  36static const u64 xe_compact_pt_shifts[] = {16, 21, 30, 39, 48};
  37
  38#define XE_PT_HIGHEST_LEVEL (ARRAY_SIZE(xe_normal_pt_shifts) - 1)
  39
  40static struct xe_pt_dir *as_xe_pt_dir(struct xe_pt *pt)
  41{
  42	return container_of(pt, struct xe_pt_dir, pt);
  43}
  44
  45static struct xe_pt *xe_pt_entry(struct xe_pt_dir *pt_dir, unsigned int index)
  46{
  47	return container_of(pt_dir->children[index], struct xe_pt, base);
  48}
  49
  50static u64 __xe_pt_empty_pte(struct xe_tile *tile, struct xe_vm *vm,
  51			     unsigned int level)
  52{
  53	struct xe_device *xe = tile_to_xe(tile);
  54	u16 pat_index = xe->pat.idx[XE_CACHE_WB];
  55	u8 id = tile->id;
  56
  57	if (!xe_vm_has_scratch(vm))
  58		return 0;
  59
  60	if (level > MAX_HUGEPTE_LEVEL)
  61		return vm->pt_ops->pde_encode_bo(vm->scratch_pt[id][level - 1]->bo,
  62						 0, pat_index);
  63
  64	return vm->pt_ops->pte_encode_addr(xe, 0, pat_index, level, IS_DGFX(xe), 0) |
  65		XE_PTE_NULL;
  66}
  67
  68static void xe_pt_free(struct xe_pt *pt)
  69{
  70	if (pt->level)
  71		kfree(as_xe_pt_dir(pt));
  72	else
  73		kfree(pt);
  74}
  75
  76/**
  77 * xe_pt_create() - Create a page-table.
  78 * @vm: The vm to create for.
  79 * @tile: The tile to create for.
  80 * @level: The page-table level.
  81 *
  82 * Allocate and initialize a single struct xe_pt metadata structure. Also
  83 * create the corresponding page-table bo, but don't initialize it. If the
  84 * level is grater than zero, then it's assumed to be a directory page-
  85 * table and the directory structure is also allocated and initialized to
  86 * NULL pointers.
  87 *
  88 * Return: A valid struct xe_pt pointer on success, Pointer error code on
  89 * error.
  90 */
  91struct xe_pt *xe_pt_create(struct xe_vm *vm, struct xe_tile *tile,
  92			   unsigned int level)
  93{
  94	struct xe_pt *pt;
  95	struct xe_bo *bo;
  96	int err;
  97
  98	if (level) {
  99		struct xe_pt_dir *dir = kzalloc(sizeof(*dir), GFP_KERNEL);
 100
 101		pt = (dir) ? &dir->pt : NULL;
 102	} else {
 103		pt = kzalloc(sizeof(*pt), GFP_KERNEL);
 104	}
 105	if (!pt)
 106		return ERR_PTR(-ENOMEM);
 107
 108	pt->level = level;
 109	bo = xe_bo_create_pin_map(vm->xe, tile, vm, SZ_4K,
 110				  ttm_bo_type_kernel,
 111				  XE_BO_CREATE_VRAM_IF_DGFX(tile) |
 112				  XE_BO_CREATE_IGNORE_MIN_PAGE_SIZE_BIT |
 113				  XE_BO_CREATE_PINNED_BIT |
 114				  XE_BO_CREATE_NO_RESV_EVICT |
 115				  XE_BO_PAGETABLE);
 116	if (IS_ERR(bo)) {
 117		err = PTR_ERR(bo);
 118		goto err_kfree;
 119	}
 120	pt->bo = bo;
 121	pt->base.children = level ? as_xe_pt_dir(pt)->children : NULL;
 122
 123	if (vm->xef)
 124		xe_drm_client_add_bo(vm->xef->client, pt->bo);
 125	xe_tile_assert(tile, level <= XE_VM_MAX_LEVEL);
 126
 127	return pt;
 128
 129err_kfree:
 130	xe_pt_free(pt);
 131	return ERR_PTR(err);
 132}
 133
 134/**
 135 * xe_pt_populate_empty() - Populate a page-table bo with scratch- or zero
 136 * entries.
 137 * @tile: The tile the scratch pagetable of which to use.
 138 * @vm: The vm we populate for.
 139 * @pt: The pagetable the bo of which to initialize.
 140 *
 141 * Populate the page-table bo of @pt with entries pointing into the tile's
 142 * scratch page-table tree if any. Otherwise populate with zeros.
 143 */
 144void xe_pt_populate_empty(struct xe_tile *tile, struct xe_vm *vm,
 145			  struct xe_pt *pt)
 146{
 147	struct iosys_map *map = &pt->bo->vmap;
 148	u64 empty;
 149	int i;
 150
 151	if (!xe_vm_has_scratch(vm)) {
 152		/*
 153		 * FIXME: Some memory is allocated already allocated to zero?
 154		 * Find out which memory that is and avoid this memset...
 155		 */
 156		xe_map_memset(vm->xe, map, 0, 0, SZ_4K);
 157	} else {
 158		empty = __xe_pt_empty_pte(tile, vm, pt->level);
 159		for (i = 0; i < XE_PDES; i++)
 160			xe_pt_write(vm->xe, map, i, empty);
 161	}
 162}
 163
 164/**
 165 * xe_pt_shift() - Return the ilog2 value of the size of the address range of
 166 * a page-table at a certain level.
 167 * @level: The level.
 168 *
 169 * Return: The ilog2 value of the size of the address range of a page-table
 170 * at level @level.
 171 */
 172unsigned int xe_pt_shift(unsigned int level)
 173{
 174	return XE_PTE_SHIFT + XE_PDE_SHIFT * level;
 175}
 176
 177/**
 178 * xe_pt_destroy() - Destroy a page-table tree.
 179 * @pt: The root of the page-table tree to destroy.
 180 * @flags: vm flags. Currently unused.
 181 * @deferred: List head of lockless list for deferred putting. NULL for
 182 *            immediate putting.
 183 *
 184 * Puts the page-table bo, recursively calls xe_pt_destroy on all children
 185 * and finally frees @pt. TODO: Can we remove the @flags argument?
 186 */
 187void xe_pt_destroy(struct xe_pt *pt, u32 flags, struct llist_head *deferred)
 188{
 189	int i;
 190
 191	if (!pt)
 192		return;
 193
 194	XE_WARN_ON(!list_empty(&pt->bo->ttm.base.gpuva.list));
 195	xe_bo_unpin(pt->bo);
 196	xe_bo_put_deferred(pt->bo, deferred);
 197
 198	if (pt->level > 0 && pt->num_live) {
 199		struct xe_pt_dir *pt_dir = as_xe_pt_dir(pt);
 200
 201		for (i = 0; i < XE_PDES; i++) {
 202			if (xe_pt_entry(pt_dir, i))
 203				xe_pt_destroy(xe_pt_entry(pt_dir, i), flags,
 204					      deferred);
 205		}
 206	}
 207	xe_pt_free(pt);
 208}
 209
 210/**
 211 * DOC: Pagetable building
 212 *
 213 * Below we use the term "page-table" for both page-directories, containing
 214 * pointers to lower level page-directories or page-tables, and level 0
 215 * page-tables that contain only page-table-entries pointing to memory pages.
 216 *
 217 * When inserting an address range in an already existing page-table tree
 218 * there will typically be a set of page-tables that are shared with other
 219 * address ranges, and a set that are private to this address range.
 220 * The set of shared page-tables can be at most two per level,
 221 * and those can't be updated immediately because the entries of those
 222 * page-tables may still be in use by the gpu for other mappings. Therefore
 223 * when inserting entries into those, we instead stage those insertions by
 224 * adding insertion data into struct xe_vm_pgtable_update structures. This
 225 * data, (subtrees for the cpu and page-table-entries for the gpu) is then
 226 * added in a separate commit step. CPU-data is committed while still under the
 227 * vm lock, the object lock and for userptr, the notifier lock in read mode.
 228 * The GPU async data is committed either by the GPU or CPU after fulfilling
 229 * relevant dependencies.
 230 * For non-shared page-tables (and, in fact, for shared ones that aren't
 231 * existing at the time of staging), we add the data in-place without the
 232 * special update structures. This private part of the page-table tree will
 233 * remain disconnected from the vm page-table tree until data is committed to
 234 * the shared page tables of the vm tree in the commit phase.
 235 */
 236
 237struct xe_pt_update {
 238	/** @update: The update structure we're building for this parent. */
 239	struct xe_vm_pgtable_update *update;
 240	/** @parent: The parent. Used to detect a parent change. */
 241	struct xe_pt *parent;
 242	/** @preexisting: Whether the parent was pre-existing or allocated */
 243	bool preexisting;
 244};
 245
 246struct xe_pt_stage_bind_walk {
 247	/** base: The base class. */
 248	struct xe_pt_walk base;
 249
 250	/* Input parameters for the walk */
 251	/** @vm: The vm we're building for. */
 252	struct xe_vm *vm;
 253	/** @tile: The tile we're building for. */
 254	struct xe_tile *tile;
 255	/** @default_pte: PTE flag only template. No address is associated */
 256	u64 default_pte;
 257	/** @dma_offset: DMA offset to add to the PTE. */
 258	u64 dma_offset;
 259	/**
 260	 * @needs_64k: This address range enforces 64K alignment and
 261	 * granularity.
 262	 */
 263	bool needs_64K;
 264	/**
 265	 * @vma: VMA being mapped
 266	 */
 267	struct xe_vma *vma;
 268
 269	/* Also input, but is updated during the walk*/
 270	/** @curs: The DMA address cursor. */
 271	struct xe_res_cursor *curs;
 272	/** @va_curs_start: The Virtual address coresponding to @curs->start */
 273	u64 va_curs_start;
 274
 275	/* Output */
 276	struct xe_walk_update {
 277		/** @wupd.entries: Caller provided storage. */
 278		struct xe_vm_pgtable_update *entries;
 279		/** @wupd.num_used_entries: Number of update @entries used. */
 280		unsigned int num_used_entries;
 281		/** @wupd.updates: Tracks the update entry at a given level */
 282		struct xe_pt_update updates[XE_VM_MAX_LEVEL + 1];
 283	} wupd;
 284
 285	/* Walk state */
 286	/**
 287	 * @l0_end_addr: The end address of the current l0 leaf. Used for
 288	 * 64K granularity detection.
 289	 */
 290	u64 l0_end_addr;
 291	/** @addr_64K: The start address of the current 64K chunk. */
 292	u64 addr_64K;
 293	/** @found_64: Whether @add_64K actually points to a 64K chunk. */
 294	bool found_64K;
 295};
 296
 297static int
 298xe_pt_new_shared(struct xe_walk_update *wupd, struct xe_pt *parent,
 299		 pgoff_t offset, bool alloc_entries)
 300{
 301	struct xe_pt_update *upd = &wupd->updates[parent->level];
 302	struct xe_vm_pgtable_update *entry;
 303
 304	/*
 305	 * For *each level*, we could only have one active
 306	 * struct xt_pt_update at any one time. Once we move on to a
 307	 * new parent and page-directory, the old one is complete, and
 308	 * updates are either already stored in the build tree or in
 309	 * @wupd->entries
 310	 */
 311	if (likely(upd->parent == parent))
 312		return 0;
 313
 314	upd->parent = parent;
 315	upd->preexisting = true;
 316
 317	if (wupd->num_used_entries == XE_VM_MAX_LEVEL * 2 + 1)
 318		return -EINVAL;
 319
 320	entry = wupd->entries + wupd->num_used_entries++;
 321	upd->update = entry;
 322	entry->ofs = offset;
 323	entry->pt_bo = parent->bo;
 324	entry->pt = parent;
 325	entry->flags = 0;
 326	entry->qwords = 0;
 327
 328	if (alloc_entries) {
 329		entry->pt_entries = kmalloc_array(XE_PDES,
 330						  sizeof(*entry->pt_entries),
 331						  GFP_KERNEL);
 332		if (!entry->pt_entries)
 333			return -ENOMEM;
 334	}
 335
 336	return 0;
 337}
 338
 339/*
 340 * NOTE: This is a very frequently called function so we allow ourselves
 341 * to annotate (using branch prediction hints) the fastpath of updating a
 342 * non-pre-existing pagetable with leaf ptes.
 343 */
 344static int
 345xe_pt_insert_entry(struct xe_pt_stage_bind_walk *xe_walk, struct xe_pt *parent,
 346		   pgoff_t offset, struct xe_pt *xe_child, u64 pte)
 347{
 348	struct xe_pt_update *upd = &xe_walk->wupd.updates[parent->level];
 349	struct xe_pt_update *child_upd = xe_child ?
 350		&xe_walk->wupd.updates[xe_child->level] : NULL;
 351	int ret;
 352
 353	ret = xe_pt_new_shared(&xe_walk->wupd, parent, offset, true);
 354	if (unlikely(ret))
 355		return ret;
 356
 357	/*
 358	 * Register this new pagetable so that it won't be recognized as
 359	 * a shared pagetable by a subsequent insertion.
 360	 */
 361	if (unlikely(child_upd)) {
 362		child_upd->update = NULL;
 363		child_upd->parent = xe_child;
 364		child_upd->preexisting = false;
 365	}
 366
 367	if (likely(!upd->preexisting)) {
 368		/* Continue building a non-connected subtree. */
 369		struct iosys_map *map = &parent->bo->vmap;
 370
 371		if (unlikely(xe_child))
 372			parent->base.children[offset] = &xe_child->base;
 373
 374		xe_pt_write(xe_walk->vm->xe, map, offset, pte);
 375		parent->num_live++;
 376	} else {
 377		/* Shared pt. Stage update. */
 378		unsigned int idx;
 379		struct xe_vm_pgtable_update *entry = upd->update;
 380
 381		idx = offset - entry->ofs;
 382		entry->pt_entries[idx].pt = xe_child;
 383		entry->pt_entries[idx].pte = pte;
 384		entry->qwords++;
 385	}
 386
 387	return 0;
 388}
 389
 390static bool xe_pt_hugepte_possible(u64 addr, u64 next, unsigned int level,
 391				   struct xe_pt_stage_bind_walk *xe_walk)
 392{
 393	u64 size, dma;
 394
 395	if (level > MAX_HUGEPTE_LEVEL)
 396		return false;
 397
 398	/* Does the virtual range requested cover a huge pte? */
 399	if (!xe_pt_covers(addr, next, level, &xe_walk->base))
 400		return false;
 401
 402	/* Does the DMA segment cover the whole pte? */
 403	if (next - xe_walk->va_curs_start > xe_walk->curs->size)
 404		return false;
 405
 406	/* null VMA's do not have dma addresses */
 407	if (xe_vma_is_null(xe_walk->vma))
 408		return true;
 409
 410	/* Is the DMA address huge PTE size aligned? */
 411	size = next - addr;
 412	dma = addr - xe_walk->va_curs_start + xe_res_dma(xe_walk->curs);
 413
 414	return IS_ALIGNED(dma, size);
 415}
 416
 417/*
 418 * Scan the requested mapping to check whether it can be done entirely
 419 * with 64K PTEs.
 420 */
 421static bool
 422xe_pt_scan_64K(u64 addr, u64 next, struct xe_pt_stage_bind_walk *xe_walk)
 423{
 424	struct xe_res_cursor curs = *xe_walk->curs;
 425
 426	if (!IS_ALIGNED(addr, SZ_64K))
 427		return false;
 428
 429	if (next > xe_walk->l0_end_addr)
 430		return false;
 431
 432	/* null VMA's do not have dma addresses */
 433	if (xe_vma_is_null(xe_walk->vma))
 434		return true;
 435
 436	xe_res_next(&curs, addr - xe_walk->va_curs_start);
 437	for (; addr < next; addr += SZ_64K) {
 438		if (!IS_ALIGNED(xe_res_dma(&curs), SZ_64K) || curs.size < SZ_64K)
 439			return false;
 440
 441		xe_res_next(&curs, SZ_64K);
 442	}
 443
 444	return addr == next;
 445}
 446
 447/*
 448 * For non-compact "normal" 4K level-0 pagetables, we want to try to group
 449 * addresses together in 64K-contigous regions to add a 64K TLB hint for the
 450 * device to the PTE.
 451 * This function determines whether the address is part of such a
 452 * segment. For VRAM in normal pagetables, this is strictly necessary on
 453 * some devices.
 454 */
 455static bool
 456xe_pt_is_pte_ps64K(u64 addr, u64 next, struct xe_pt_stage_bind_walk *xe_walk)
 457{
 458	/* Address is within an already found 64k region */
 459	if (xe_walk->found_64K && addr - xe_walk->addr_64K < SZ_64K)
 460		return true;
 461
 462	xe_walk->found_64K = xe_pt_scan_64K(addr, addr + SZ_64K, xe_walk);
 463	xe_walk->addr_64K = addr;
 464
 465	return xe_walk->found_64K;
 466}
 467
 468static int
 469xe_pt_stage_bind_entry(struct xe_ptw *parent, pgoff_t offset,
 470		       unsigned int level, u64 addr, u64 next,
 471		       struct xe_ptw **child,
 472		       enum page_walk_action *action,
 473		       struct xe_pt_walk *walk)
 474{
 475	struct xe_pt_stage_bind_walk *xe_walk =
 476		container_of(walk, typeof(*xe_walk), base);
 477	u16 pat_index = xe_walk->vma->pat_index;
 478	struct xe_pt *xe_parent = container_of(parent, typeof(*xe_parent), base);
 479	struct xe_vm *vm = xe_walk->vm;
 480	struct xe_pt *xe_child;
 481	bool covers;
 482	int ret = 0;
 483	u64 pte;
 484
 485	/* Is this a leaf entry ?*/
 486	if (level == 0 || xe_pt_hugepte_possible(addr, next, level, xe_walk)) {
 487		struct xe_res_cursor *curs = xe_walk->curs;
 488		bool is_null = xe_vma_is_null(xe_walk->vma);
 489
 490		XE_WARN_ON(xe_walk->va_curs_start != addr);
 491
 492		pte = vm->pt_ops->pte_encode_vma(is_null ? 0 :
 493						 xe_res_dma(curs) + xe_walk->dma_offset,
 494						 xe_walk->vma, pat_index, level);
 495		pte |= xe_walk->default_pte;
 496
 497		/*
 498		 * Set the XE_PTE_PS64 hint if possible, otherwise if
 499		 * this device *requires* 64K PTE size for VRAM, fail.
 500		 */
 501		if (level == 0 && !xe_parent->is_compact) {
 502			if (xe_pt_is_pte_ps64K(addr, next, xe_walk)) {
 503				xe_walk->vma->gpuva.flags |= XE_VMA_PTE_64K;
 504				pte |= XE_PTE_PS64;
 505			} else if (XE_WARN_ON(xe_walk->needs_64K)) {
 506				return -EINVAL;
 507			}
 508		}
 509
 510		ret = xe_pt_insert_entry(xe_walk, xe_parent, offset, NULL, pte);
 511		if (unlikely(ret))
 512			return ret;
 513
 514		if (!is_null)
 515			xe_res_next(curs, next - addr);
 516		xe_walk->va_curs_start = next;
 517		xe_walk->vma->gpuva.flags |= (XE_VMA_PTE_4K << level);
 518		*action = ACTION_CONTINUE;
 519
 520		return ret;
 521	}
 522
 523	/*
 524	 * Descending to lower level. Determine if we need to allocate a
 525	 * new page table or -directory, which we do if there is no
 526	 * previous one or there is one we can completely replace.
 527	 */
 528	if (level == 1) {
 529		walk->shifts = xe_normal_pt_shifts;
 530		xe_walk->l0_end_addr = next;
 531	}
 532
 533	covers = xe_pt_covers(addr, next, level, &xe_walk->base);
 534	if (covers || !*child) {
 535		u64 flags = 0;
 536
 537		xe_child = xe_pt_create(xe_walk->vm, xe_walk->tile, level - 1);
 538		if (IS_ERR(xe_child))
 539			return PTR_ERR(xe_child);
 540
 541		xe_pt_set_addr(xe_child,
 542			       round_down(addr, 1ull << walk->shifts[level]));
 543
 544		if (!covers)
 545			xe_pt_populate_empty(xe_walk->tile, xe_walk->vm, xe_child);
 546
 547		*child = &xe_child->base;
 548
 549		/*
 550		 * Prefer the compact pagetable layout for L0 if possible. Only
 551		 * possible if VMA covers entire 2MB region as compact 64k and
 552		 * 4k pages cannot be mixed within a 2MB region.
 553		 * TODO: Suballocate the pt bo to avoid wasting a lot of
 554		 * memory.
 555		 */
 556		if (GRAPHICS_VERx100(tile_to_xe(xe_walk->tile)) >= 1250 && level == 1 &&
 557		    covers && xe_pt_scan_64K(addr, next, xe_walk)) {
 558			walk->shifts = xe_compact_pt_shifts;
 559			xe_walk->vma->gpuva.flags |= XE_VMA_PTE_COMPACT;
 560			flags |= XE_PDE_64K;
 561			xe_child->is_compact = true;
 562		}
 563
 564		pte = vm->pt_ops->pde_encode_bo(xe_child->bo, 0, pat_index) | flags;
 565		ret = xe_pt_insert_entry(xe_walk, xe_parent, offset, xe_child,
 566					 pte);
 567	}
 568
 569	*action = ACTION_SUBTREE;
 570	return ret;
 571}
 572
 573static const struct xe_pt_walk_ops xe_pt_stage_bind_ops = {
 574	.pt_entry = xe_pt_stage_bind_entry,
 575};
 576
 577/**
 578 * xe_pt_stage_bind() - Build a disconnected page-table tree for a given address
 579 * range.
 580 * @tile: The tile we're building for.
 581 * @vma: The vma indicating the address range.
 582 * @entries: Storage for the update entries used for connecting the tree to
 583 * the main tree at commit time.
 584 * @num_entries: On output contains the number of @entries used.
 585 *
 586 * This function builds a disconnected page-table tree for a given address
 587 * range. The tree is connected to the main vm tree for the gpu using
 588 * xe_migrate_update_pgtables() and for the cpu using xe_pt_commit_bind().
 589 * The function builds xe_vm_pgtable_update structures for already existing
 590 * shared page-tables, and non-existing shared and non-shared page-tables
 591 * are built and populated directly.
 592 *
 593 * Return 0 on success, negative error code on error.
 594 */
 595static int
 596xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
 597		 struct xe_vm_pgtable_update *entries, u32 *num_entries)
 598{
 599	struct xe_device *xe = tile_to_xe(tile);
 600	struct xe_bo *bo = xe_vma_bo(vma);
 601	bool is_devmem = !xe_vma_is_userptr(vma) && bo &&
 602		(xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo));
 603	struct xe_res_cursor curs;
 604	struct xe_pt_stage_bind_walk xe_walk = {
 605		.base = {
 606			.ops = &xe_pt_stage_bind_ops,
 607			.shifts = xe_normal_pt_shifts,
 608			.max_level = XE_PT_HIGHEST_LEVEL,
 609		},
 610		.vm = xe_vma_vm(vma),
 611		.tile = tile,
 612		.curs = &curs,
 613		.va_curs_start = xe_vma_start(vma),
 614		.vma = vma,
 615		.wupd.entries = entries,
 616		.needs_64K = (xe_vma_vm(vma)->flags & XE_VM_FLAG_64K) && is_devmem,
 617	};
 618	struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
 619	int ret;
 620
 621	if (vma && (vma->gpuva.flags & XE_VMA_ATOMIC_PTE_BIT) &&
 622	    (is_devmem || !IS_DGFX(xe)))
 623		xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
 624
 625	if (is_devmem) {
 626		xe_walk.default_pte |= XE_PPGTT_PTE_DM;
 627		xe_walk.dma_offset = vram_region_gpu_offset(bo->ttm.resource);
 628	}
 629
 630	if (!xe_vma_has_no_bo(vma) && xe_bo_is_stolen(bo))
 631		xe_walk.dma_offset = xe_ttm_stolen_gpu_offset(xe_bo_device(bo));
 632
 633	xe_bo_assert_held(bo);
 634
 635	if (!xe_vma_is_null(vma)) {
 636		if (xe_vma_is_userptr(vma))
 637			xe_res_first_sg(to_userptr_vma(vma)->userptr.sg, 0,
 638					xe_vma_size(vma), &curs);
 639		else if (xe_bo_is_vram(bo) || xe_bo_is_stolen(bo))
 640			xe_res_first(bo->ttm.resource, xe_vma_bo_offset(vma),
 641				     xe_vma_size(vma), &curs);
 642		else
 643			xe_res_first_sg(xe_bo_sg(bo), xe_vma_bo_offset(vma),
 644					xe_vma_size(vma), &curs);
 645	} else {
 646		curs.size = xe_vma_size(vma);
 647	}
 648
 649	ret = xe_pt_walk_range(&pt->base, pt->level, xe_vma_start(vma),
 650			       xe_vma_end(vma), &xe_walk.base);
 651
 652	*num_entries = xe_walk.wupd.num_used_entries;
 653	return ret;
 654}
 655
 656/**
 657 * xe_pt_nonshared_offsets() - Determine the non-shared entry offsets of a
 658 * shared pagetable.
 659 * @addr: The start address within the non-shared pagetable.
 660 * @end: The end address within the non-shared pagetable.
 661 * @level: The level of the non-shared pagetable.
 662 * @walk: Walk info. The function adjusts the walk action.
 663 * @action: next action to perform (see enum page_walk_action)
 664 * @offset: Ignored on input, First non-shared entry on output.
 665 * @end_offset: Ignored on input, Last non-shared entry + 1 on output.
 666 *
 667 * A non-shared page-table has some entries that belong to the address range
 668 * and others that don't. This function determines the entries that belong
 669 * fully to the address range. Depending on level, some entries may
 670 * partially belong to the address range (that can't happen at level 0).
 671 * The function detects that and adjust those offsets to not include those
 672 * partial entries. Iff it does detect partial entries, we know that there must
 673 * be shared page tables also at lower levels, so it adjusts the walk action
 674 * accordingly.
 675 *
 676 * Return: true if there were non-shared entries, false otherwise.
 677 */
 678static bool xe_pt_nonshared_offsets(u64 addr, u64 end, unsigned int level,
 679				    struct xe_pt_walk *walk,
 680				    enum page_walk_action *action,
 681				    pgoff_t *offset, pgoff_t *end_offset)
 682{
 683	u64 size = 1ull << walk->shifts[level];
 684
 685	*offset = xe_pt_offset(addr, level, walk);
 686	*end_offset = xe_pt_num_entries(addr, end, level, walk) + *offset;
 687
 688	if (!level)
 689		return true;
 690
 691	/*
 692	 * If addr or next are not size aligned, there are shared pts at lower
 693	 * level, so in that case traverse down the subtree
 694	 */
 695	*action = ACTION_CONTINUE;
 696	if (!IS_ALIGNED(addr, size)) {
 697		*action = ACTION_SUBTREE;
 698		(*offset)++;
 699	}
 700
 701	if (!IS_ALIGNED(end, size)) {
 702		*action = ACTION_SUBTREE;
 703		(*end_offset)--;
 704	}
 705
 706	return *end_offset > *offset;
 707}
 708
 709struct xe_pt_zap_ptes_walk {
 710	/** @base: The walk base-class */
 711	struct xe_pt_walk base;
 712
 713	/* Input parameters for the walk */
 714	/** @tile: The tile we're building for */
 715	struct xe_tile *tile;
 716
 717	/* Output */
 718	/** @needs_invalidate: Whether we need to invalidate TLB*/
 719	bool needs_invalidate;
 720};
 721
 722static int xe_pt_zap_ptes_entry(struct xe_ptw *parent, pgoff_t offset,
 723				unsigned int level, u64 addr, u64 next,
 724				struct xe_ptw **child,
 725				enum page_walk_action *action,
 726				struct xe_pt_walk *walk)
 727{
 728	struct xe_pt_zap_ptes_walk *xe_walk =
 729		container_of(walk, typeof(*xe_walk), base);
 730	struct xe_pt *xe_child = container_of(*child, typeof(*xe_child), base);
 731	pgoff_t end_offset;
 732
 733	XE_WARN_ON(!*child);
 734	XE_WARN_ON(!level && xe_child->is_compact);
 735
 736	/*
 737	 * Note that we're called from an entry callback, and we're dealing
 738	 * with the child of that entry rather than the parent, so need to
 739	 * adjust level down.
 740	 */
 741	if (xe_pt_nonshared_offsets(addr, next, --level, walk, action, &offset,
 742				    &end_offset)) {
 743		xe_map_memset(tile_to_xe(xe_walk->tile), &xe_child->bo->vmap,
 744			      offset * sizeof(u64), 0,
 745			      (end_offset - offset) * sizeof(u64));
 746		xe_walk->needs_invalidate = true;
 747	}
 748
 749	return 0;
 750}
 751
 752static const struct xe_pt_walk_ops xe_pt_zap_ptes_ops = {
 753	.pt_entry = xe_pt_zap_ptes_entry,
 754};
 755
 756/**
 757 * xe_pt_zap_ptes() - Zap (zero) gpu ptes of an address range
 758 * @tile: The tile we're zapping for.
 759 * @vma: GPU VMA detailing address range.
 760 *
 761 * Eviction and Userptr invalidation needs to be able to zap the
 762 * gpu ptes of a given address range in pagefaulting mode.
 763 * In order to be able to do that, that function needs access to the shared
 764 * page-table entrieaso it can either clear the leaf PTEs or
 765 * clear the pointers to lower-level page-tables. The caller is required
 766 * to hold the necessary locks to ensure neither the page-table connectivity
 767 * nor the page-table entries of the range is updated from under us.
 768 *
 769 * Return: Whether ptes were actually updated and a TLB invalidation is
 770 * required.
 771 */
 772bool xe_pt_zap_ptes(struct xe_tile *tile, struct xe_vma *vma)
 773{
 774	struct xe_pt_zap_ptes_walk xe_walk = {
 775		.base = {
 776			.ops = &xe_pt_zap_ptes_ops,
 777			.shifts = xe_normal_pt_shifts,
 778			.max_level = XE_PT_HIGHEST_LEVEL,
 779		},
 780		.tile = tile,
 781	};
 782	struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
 783
 784	if (!(vma->tile_present & BIT(tile->id)))
 785		return false;
 786
 787	(void)xe_pt_walk_shared(&pt->base, pt->level, xe_vma_start(vma),
 788				xe_vma_end(vma), &xe_walk.base);
 789
 790	return xe_walk.needs_invalidate;
 791}
 792
 793static void
 794xe_vm_populate_pgtable(struct xe_migrate_pt_update *pt_update, struct xe_tile *tile,
 795		       struct iosys_map *map, void *data,
 796		       u32 qword_ofs, u32 num_qwords,
 797		       const struct xe_vm_pgtable_update *update)
 798{
 799	struct xe_pt_entry *ptes = update->pt_entries;
 800	u64 *ptr = data;
 801	u32 i;
 802
 803	for (i = 0; i < num_qwords; i++) {
 804		if (map)
 805			xe_map_wr(tile_to_xe(tile), map, (qword_ofs + i) *
 806				  sizeof(u64), u64, ptes[i].pte);
 807		else
 808			ptr[i] = ptes[i].pte;
 809	}
 810}
 811
 812static void xe_pt_abort_bind(struct xe_vma *vma,
 813			     struct xe_vm_pgtable_update *entries,
 814			     u32 num_entries)
 815{
 816	u32 i, j;
 817
 818	for (i = 0; i < num_entries; i++) {
 819		if (!entries[i].pt_entries)
 820			continue;
 821
 822		for (j = 0; j < entries[i].qwords; j++)
 823			xe_pt_destroy(entries[i].pt_entries[j].pt, xe_vma_vm(vma)->flags, NULL);
 824		kfree(entries[i].pt_entries);
 825	}
 826}
 827
 828static void xe_pt_commit_locks_assert(struct xe_vma *vma)
 829{
 830	struct xe_vm *vm = xe_vma_vm(vma);
 831
 832	lockdep_assert_held(&vm->lock);
 833
 834	if (xe_vma_is_userptr(vma))
 835		lockdep_assert_held_read(&vm->userptr.notifier_lock);
 836	else if (!xe_vma_is_null(vma))
 837		dma_resv_assert_held(xe_vma_bo(vma)->ttm.base.resv);
 838
 839	xe_vm_assert_held(vm);
 840}
 841
 842static void xe_pt_commit_bind(struct xe_vma *vma,
 843			      struct xe_vm_pgtable_update *entries,
 844			      u32 num_entries, bool rebind,
 845			      struct llist_head *deferred)
 846{
 847	u32 i, j;
 848
 849	xe_pt_commit_locks_assert(vma);
 850
 851	for (i = 0; i < num_entries; i++) {
 852		struct xe_pt *pt = entries[i].pt;
 853		struct xe_pt_dir *pt_dir;
 854
 855		if (!rebind)
 856			pt->num_live += entries[i].qwords;
 857
 858		if (!pt->level) {
 859			kfree(entries[i].pt_entries);
 860			continue;
 861		}
 862
 863		pt_dir = as_xe_pt_dir(pt);
 864		for (j = 0; j < entries[i].qwords; j++) {
 865			u32 j_ = j + entries[i].ofs;
 866			struct xe_pt *newpte = entries[i].pt_entries[j].pt;
 867
 868			if (xe_pt_entry(pt_dir, j_))
 869				xe_pt_destroy(xe_pt_entry(pt_dir, j_),
 870					      xe_vma_vm(vma)->flags, deferred);
 871
 872			pt_dir->children[j_] = &newpte->base;
 873		}
 874		kfree(entries[i].pt_entries);
 875	}
 876}
 877
 878static int
 879xe_pt_prepare_bind(struct xe_tile *tile, struct xe_vma *vma,
 880		   struct xe_vm_pgtable_update *entries, u32 *num_entries,
 881		   bool rebind)
 882{
 883	int err;
 884
 885	*num_entries = 0;
 886	err = xe_pt_stage_bind(tile, vma, entries, num_entries);
 887	if (!err)
 888		xe_tile_assert(tile, *num_entries);
 889	else /* abort! */
 890		xe_pt_abort_bind(vma, entries, *num_entries);
 891
 892	return err;
 893}
 894
 895static void xe_vm_dbg_print_entries(struct xe_device *xe,
 896				    const struct xe_vm_pgtable_update *entries,
 897				    unsigned int num_entries)
 898#if (IS_ENABLED(CONFIG_DRM_XE_DEBUG_VM))
 899{
 900	unsigned int i;
 901
 902	vm_dbg(&xe->drm, "%u entries to update\n", num_entries);
 903	for (i = 0; i < num_entries; i++) {
 904		const struct xe_vm_pgtable_update *entry = &entries[i];
 905		struct xe_pt *xe_pt = entry->pt;
 906		u64 page_size = 1ull << xe_pt_shift(xe_pt->level);
 907		u64 end;
 908		u64 start;
 909
 910		xe_assert(xe, !entry->pt->is_compact);
 911		start = entry->ofs * page_size;
 912		end = start + page_size * entry->qwords;
 913		vm_dbg(&xe->drm,
 914		       "\t%u: Update level %u at (%u + %u) [%llx...%llx) f:%x\n",
 915		       i, xe_pt->level, entry->ofs, entry->qwords,
 916		       xe_pt_addr(xe_pt) + start, xe_pt_addr(xe_pt) + end, 0);
 917	}
 918}
 919#else
 920{}
 921#endif
 922
 923#ifdef CONFIG_DRM_XE_USERPTR_INVAL_INJECT
 924
 925static int xe_pt_userptr_inject_eagain(struct xe_userptr_vma *uvma)
 926{
 927	u32 divisor = uvma->userptr.divisor ? uvma->userptr.divisor : 2;
 928	static u32 count;
 929
 930	if (count++ % divisor == divisor - 1) {
 931		struct xe_vm *vm = xe_vma_vm(&uvma->vma);
 932
 933		uvma->userptr.divisor = divisor << 1;
 934		spin_lock(&vm->userptr.invalidated_lock);
 935		list_move_tail(&uvma->userptr.invalidate_link,
 936			       &vm->userptr.invalidated);
 937		spin_unlock(&vm->userptr.invalidated_lock);
 938		return true;
 939	}
 940
 941	return false;
 942}
 943
 944#else
 945
 946static bool xe_pt_userptr_inject_eagain(struct xe_userptr_vma *uvma)
 947{
 948	return false;
 949}
 950
 951#endif
 952
 953/**
 954 * struct xe_pt_migrate_pt_update - Callback argument for pre-commit callbacks
 955 * @base: Base we derive from.
 956 * @bind: Whether this is a bind or an unbind operation. A bind operation
 957 *        makes the pre-commit callback error with -EAGAIN if it detects a
 958 *        pending invalidation.
 959 * @locked: Whether the pre-commit callback locked the userptr notifier lock
 960 *          and it needs unlocking.
 961 */
 962struct xe_pt_migrate_pt_update {
 963	struct xe_migrate_pt_update base;
 964	bool bind;
 965	bool locked;
 966};
 967
 968/*
 969 * This function adds the needed dependencies to a page-table update job
 970 * to make sure racing jobs for separate bind engines don't race writing
 971 * to the same page-table range, wreaking havoc. Initially use a single
 972 * fence for the entire VM. An optimization would use smaller granularity.
 973 */
 974static int xe_pt_vm_dependencies(struct xe_sched_job *job,
 975				 struct xe_range_fence_tree *rftree,
 976				 u64 start, u64 last)
 977{
 978	struct xe_range_fence *rtfence;
 979	struct dma_fence *fence;
 980	int err;
 981
 982	rtfence = xe_range_fence_tree_first(rftree, start, last);
 983	while (rtfence) {
 984		fence = rtfence->fence;
 985
 986		if (!dma_fence_is_signaled(fence)) {
 987			/*
 988			 * Is this a CPU update? GPU is busy updating, so return
 989			 * an error
 990			 */
 991			if (!job)
 992				return -ETIME;
 993
 994			dma_fence_get(fence);
 995			err = drm_sched_job_add_dependency(&job->drm, fence);
 996			if (err)
 997				return err;
 998		}
 999
1000		rtfence = xe_range_fence_tree_next(rtfence, start, last);
1001	}
1002
1003	return 0;
1004}
1005
1006static int xe_pt_pre_commit(struct xe_migrate_pt_update *pt_update)
1007{
1008	struct xe_range_fence_tree *rftree =
1009		&xe_vma_vm(pt_update->vma)->rftree[pt_update->tile_id];
1010
1011	return xe_pt_vm_dependencies(pt_update->job, rftree,
1012				     pt_update->start, pt_update->last);
1013}
1014
1015static int xe_pt_userptr_pre_commit(struct xe_migrate_pt_update *pt_update)
1016{
1017	struct xe_pt_migrate_pt_update *userptr_update =
1018		container_of(pt_update, typeof(*userptr_update), base);
1019	struct xe_userptr_vma *uvma = to_userptr_vma(pt_update->vma);
1020	unsigned long notifier_seq = uvma->userptr.notifier_seq;
1021	struct xe_vm *vm = xe_vma_vm(&uvma->vma);
1022	int err = xe_pt_vm_dependencies(pt_update->job,
1023					&vm->rftree[pt_update->tile_id],
1024					pt_update->start,
1025					pt_update->last);
1026
1027	if (err)
1028		return err;
1029
1030	userptr_update->locked = false;
1031
1032	/*
1033	 * Wait until nobody is running the invalidation notifier, and
1034	 * since we're exiting the loop holding the notifier lock,
1035	 * nobody can proceed invalidating either.
1036	 *
1037	 * Note that we don't update the vma->userptr.notifier_seq since
1038	 * we don't update the userptr pages.
1039	 */
1040	do {
1041		down_read(&vm->userptr.notifier_lock);
1042		if (!mmu_interval_read_retry(&uvma->userptr.notifier,
1043					     notifier_seq))
1044			break;
1045
1046		up_read(&vm->userptr.notifier_lock);
1047
1048		if (userptr_update->bind)
1049			return -EAGAIN;
1050
1051		notifier_seq = mmu_interval_read_begin(&uvma->userptr.notifier);
1052	} while (true);
1053
1054	/* Inject errors to test_whether they are handled correctly */
1055	if (userptr_update->bind && xe_pt_userptr_inject_eagain(uvma)) {
1056		up_read(&vm->userptr.notifier_lock);
1057		return -EAGAIN;
1058	}
1059
1060	userptr_update->locked = true;
1061
1062	return 0;
1063}
1064
1065static const struct xe_migrate_pt_update_ops bind_ops = {
1066	.populate = xe_vm_populate_pgtable,
1067	.pre_commit = xe_pt_pre_commit,
1068};
1069
1070static const struct xe_migrate_pt_update_ops userptr_bind_ops = {
1071	.populate = xe_vm_populate_pgtable,
1072	.pre_commit = xe_pt_userptr_pre_commit,
1073};
1074
1075struct invalidation_fence {
1076	struct xe_gt_tlb_invalidation_fence base;
1077	struct xe_gt *gt;
1078	struct xe_vma *vma;
1079	struct dma_fence *fence;
1080	struct dma_fence_cb cb;
1081	struct work_struct work;
1082};
1083
1084static const char *
1085invalidation_fence_get_driver_name(struct dma_fence *dma_fence)
1086{
1087	return "xe";
1088}
1089
1090static const char *
1091invalidation_fence_get_timeline_name(struct dma_fence *dma_fence)
1092{
1093	return "invalidation_fence";
1094}
1095
1096static const struct dma_fence_ops invalidation_fence_ops = {
1097	.get_driver_name = invalidation_fence_get_driver_name,
1098	.get_timeline_name = invalidation_fence_get_timeline_name,
1099};
1100
1101static void invalidation_fence_cb(struct dma_fence *fence,
1102				  struct dma_fence_cb *cb)
1103{
1104	struct invalidation_fence *ifence =
1105		container_of(cb, struct invalidation_fence, cb);
1106
1107	trace_xe_gt_tlb_invalidation_fence_cb(&ifence->base);
1108	if (!ifence->fence->error) {
1109		queue_work(system_wq, &ifence->work);
1110	} else {
1111		ifence->base.base.error = ifence->fence->error;
1112		dma_fence_signal(&ifence->base.base);
1113		dma_fence_put(&ifence->base.base);
1114	}
1115	dma_fence_put(ifence->fence);
1116}
1117
1118static void invalidation_fence_work_func(struct work_struct *w)
1119{
1120	struct invalidation_fence *ifence =
1121		container_of(w, struct invalidation_fence, work);
1122
1123	trace_xe_gt_tlb_invalidation_fence_work_func(&ifence->base);
1124	xe_gt_tlb_invalidation_vma(ifence->gt, &ifence->base, ifence->vma);
1125}
1126
1127static int invalidation_fence_init(struct xe_gt *gt,
1128				   struct invalidation_fence *ifence,
1129				   struct dma_fence *fence,
1130				   struct xe_vma *vma)
1131{
1132	int ret;
1133
1134	trace_xe_gt_tlb_invalidation_fence_create(&ifence->base);
1135
1136	spin_lock_irq(&gt->tlb_invalidation.lock);
1137	dma_fence_init(&ifence->base.base, &invalidation_fence_ops,
1138		       &gt->tlb_invalidation.lock,
1139		       gt->tlb_invalidation.fence_context,
1140		       ++gt->tlb_invalidation.fence_seqno);
1141	spin_unlock_irq(&gt->tlb_invalidation.lock);
1142
1143	INIT_LIST_HEAD(&ifence->base.link);
1144
1145	dma_fence_get(&ifence->base.base);	/* Ref for caller */
1146	ifence->fence = fence;
1147	ifence->gt = gt;
1148	ifence->vma = vma;
1149
1150	INIT_WORK(&ifence->work, invalidation_fence_work_func);
1151	ret = dma_fence_add_callback(fence, &ifence->cb, invalidation_fence_cb);
1152	if (ret == -ENOENT) {
1153		dma_fence_put(ifence->fence);	/* Usually dropped in CB */
1154		invalidation_fence_work_func(&ifence->work);
1155	} else if (ret) {
1156		dma_fence_put(&ifence->base.base);	/* Caller ref */
1157		dma_fence_put(&ifence->base.base);	/* Creation ref */
1158	}
1159
1160	xe_gt_assert(gt, !ret || ret == -ENOENT);
1161
1162	return ret && ret != -ENOENT ? ret : 0;
1163}
1164
1165static void xe_pt_calc_rfence_interval(struct xe_vma *vma,
1166				       struct xe_pt_migrate_pt_update *update,
1167				       struct xe_vm_pgtable_update *entries,
1168				       u32 num_entries)
1169{
1170	int i, level = 0;
1171
1172	for (i = 0; i < num_entries; i++) {
1173		const struct xe_vm_pgtable_update *entry = &entries[i];
1174
1175		if (entry->pt->level > level)
1176			level = entry->pt->level;
1177	}
1178
1179	/* Greedy (non-optimal) calculation but simple */
1180	update->base.start = ALIGN_DOWN(xe_vma_start(vma),
1181					0x1ull << xe_pt_shift(level));
1182	update->base.last = ALIGN(xe_vma_end(vma),
1183				  0x1ull << xe_pt_shift(level)) - 1;
1184}
1185
1186/**
1187 * __xe_pt_bind_vma() - Build and connect a page-table tree for the vma
1188 * address range.
1189 * @tile: The tile to bind for.
1190 * @vma: The vma to bind.
1191 * @q: The exec_queue with which to do pipelined page-table updates.
1192 * @syncs: Entries to sync on before binding the built tree to the live vm tree.
1193 * @num_syncs: Number of @sync entries.
1194 * @rebind: Whether we're rebinding this vma to the same address range without
1195 * an unbind in-between.
1196 *
1197 * This function builds a page-table tree (see xe_pt_stage_bind() for more
1198 * information on page-table building), and the xe_vm_pgtable_update entries
1199 * abstracting the operations needed to attach it to the main vm tree. It
1200 * then takes the relevant locks and updates the metadata side of the main
1201 * vm tree and submits the operations for pipelined attachment of the
1202 * gpu page-table to the vm main tree, (which can be done either by the
1203 * cpu and the GPU).
1204 *
1205 * Return: A valid dma-fence representing the pipelined attachment operation
1206 * on success, an error pointer on error.
1207 */
1208struct dma_fence *
1209__xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue *q,
1210		 struct xe_sync_entry *syncs, u32 num_syncs,
1211		 bool rebind)
1212{
1213	struct xe_vm_pgtable_update entries[XE_VM_MAX_LEVEL * 2 + 1];
1214	struct xe_pt_migrate_pt_update bind_pt_update = {
1215		.base = {
1216			.ops = xe_vma_is_userptr(vma) ? &userptr_bind_ops : &bind_ops,
1217			.vma = vma,
1218			.tile_id = tile->id,
1219		},
1220		.bind = true,
1221	};
1222	struct xe_vm *vm = xe_vma_vm(vma);
1223	u32 num_entries;
1224	struct dma_fence *fence;
1225	struct invalidation_fence *ifence = NULL;
1226	struct xe_range_fence *rfence;
1227	int err;
1228
1229	bind_pt_update.locked = false;
1230	xe_bo_assert_held(xe_vma_bo(vma));
1231	xe_vm_assert_held(vm);
1232
1233	vm_dbg(&xe_vma_vm(vma)->xe->drm,
1234	       "Preparing bind, with range [%llx...%llx) engine %p.\n",
1235	       xe_vma_start(vma), xe_vma_end(vma), q);
1236
1237	err = xe_pt_prepare_bind(tile, vma, entries, &num_entries, rebind);
1238	if (err)
1239		goto err;
1240	xe_tile_assert(tile, num_entries <= ARRAY_SIZE(entries));
1241
1242	xe_vm_dbg_print_entries(tile_to_xe(tile), entries, num_entries);
1243	xe_pt_calc_rfence_interval(vma, &bind_pt_update, entries,
1244				   num_entries);
1245
1246	/*
1247	 * If rebind, we have to invalidate TLB on !LR vms to invalidate
1248	 * cached PTEs point to freed memory. on LR vms this is done
1249	 * automatically when the context is re-enabled by the rebind worker,
1250	 * or in fault mode it was invalidated on PTE zapping.
1251	 *
1252	 * If !rebind, and scratch enabled VMs, there is a chance the scratch
1253	 * PTE is already cached in the TLB so it needs to be invalidated.
1254	 * on !LR VMs this is done in the ring ops preceding a batch, but on
1255	 * non-faulting LR, in particular on user-space batch buffer chaining,
1256	 * it needs to be done here.
1257	 */
1258	if ((rebind && !xe_vm_in_lr_mode(vm) && !vm->batch_invalidate_tlb) ||
1259	    (!rebind && xe_vm_has_scratch(vm) && xe_vm_in_preempt_fence_mode(vm))) {
1260		ifence = kzalloc(sizeof(*ifence), GFP_KERNEL);
1261		if (!ifence)
1262			return ERR_PTR(-ENOMEM);
1263	}
1264
1265	rfence = kzalloc(sizeof(*rfence), GFP_KERNEL);
1266	if (!rfence) {
1267		kfree(ifence);
1268		return ERR_PTR(-ENOMEM);
1269	}
1270
1271	fence = xe_migrate_update_pgtables(tile->migrate,
1272					   vm, xe_vma_bo(vma), q,
1273					   entries, num_entries,
1274					   syncs, num_syncs,
1275					   &bind_pt_update.base);
1276	if (!IS_ERR(fence)) {
1277		bool last_munmap_rebind = vma->gpuva.flags & XE_VMA_LAST_REBIND;
1278		LLIST_HEAD(deferred);
1279		int err;
1280
1281		err = xe_range_fence_insert(&vm->rftree[tile->id], rfence,
1282					    &xe_range_fence_kfree_ops,
1283					    bind_pt_update.base.start,
1284					    bind_pt_update.base.last, fence);
1285		if (err)
1286			dma_fence_wait(fence, false);
1287
1288		/* TLB invalidation must be done before signaling rebind */
1289		if (ifence) {
1290			int err = invalidation_fence_init(tile->primary_gt, ifence, fence,
1291							  vma);
1292			if (err) {
1293				dma_fence_put(fence);
1294				kfree(ifence);
1295				return ERR_PTR(err);
1296			}
1297			fence = &ifence->base.base;
1298		}
1299
1300		/* add shared fence now for pagetable delayed destroy */
1301		dma_resv_add_fence(xe_vm_resv(vm), fence, !rebind &&
1302				   last_munmap_rebind ?
1303				   DMA_RESV_USAGE_KERNEL :
1304				   DMA_RESV_USAGE_BOOKKEEP);
1305
1306		if (!xe_vma_has_no_bo(vma) && !xe_vma_bo(vma)->vm)
1307			dma_resv_add_fence(xe_vma_bo(vma)->ttm.base.resv, fence,
1308					   DMA_RESV_USAGE_BOOKKEEP);
1309		xe_pt_commit_bind(vma, entries, num_entries, rebind,
1310				  bind_pt_update.locked ? &deferred : NULL);
1311
1312		/* This vma is live (again?) now */
1313		vma->tile_present |= BIT(tile->id);
1314
1315		if (bind_pt_update.locked) {
1316			to_userptr_vma(vma)->userptr.initial_bind = true;
1317			up_read(&vm->userptr.notifier_lock);
1318			xe_bo_put_commit(&deferred);
1319		}
1320		if (!rebind && last_munmap_rebind &&
1321		    xe_vm_in_preempt_fence_mode(vm))
1322			xe_vm_queue_rebind_worker(vm);
1323	} else {
1324		kfree(rfence);
1325		kfree(ifence);
1326		if (bind_pt_update.locked)
1327			up_read(&vm->userptr.notifier_lock);
1328		xe_pt_abort_bind(vma, entries, num_entries);
1329	}
1330
1331	return fence;
1332
1333err:
1334	return ERR_PTR(err);
1335}
1336
1337struct xe_pt_stage_unbind_walk {
1338	/** @base: The pagewalk base-class. */
1339	struct xe_pt_walk base;
1340
1341	/* Input parameters for the walk */
1342	/** @tile: The tile we're unbinding from. */
1343	struct xe_tile *tile;
1344
1345	/**
1346	 * @modified_start: Walk range start, modified to include any
1347	 * shared pagetables that we're the only user of and can thus
1348	 * treat as private.
1349	 */
1350	u64 modified_start;
1351	/** @modified_end: Walk range start, modified like @modified_start. */
1352	u64 modified_end;
1353
1354	/* Output */
1355	/* @wupd: Structure to track the page-table updates we're building */
1356	struct xe_walk_update wupd;
1357};
1358
1359/*
1360 * Check whether this range is the only one populating this pagetable,
1361 * and in that case, update the walk range checks so that higher levels don't
1362 * view us as a shared pagetable.
1363 */
1364static bool xe_pt_check_kill(u64 addr, u64 next, unsigned int level,
1365			     const struct xe_pt *child,
1366			     enum page_walk_action *action,
1367			     struct xe_pt_walk *walk)
1368{
1369	struct xe_pt_stage_unbind_walk *xe_walk =
1370		container_of(walk, typeof(*xe_walk), base);
1371	unsigned int shift = walk->shifts[level];
1372	u64 size = 1ull << shift;
1373
1374	if (IS_ALIGNED(addr, size) && IS_ALIGNED(next, size) &&
1375	    ((next - addr) >> shift) == child->num_live) {
1376		u64 size = 1ull << walk->shifts[level + 1];
1377
1378		*action = ACTION_CONTINUE;
1379
1380		if (xe_walk->modified_start >= addr)
1381			xe_walk->modified_start = round_down(addr, size);
1382		if (xe_walk->modified_end <= next)
1383			xe_walk->modified_end = round_up(next, size);
1384
1385		return true;
1386	}
1387
1388	return false;
1389}
1390
1391static int xe_pt_stage_unbind_entry(struct xe_ptw *parent, pgoff_t offset,
1392				    unsigned int level, u64 addr, u64 next,
1393				    struct xe_ptw **child,
1394				    enum page_walk_action *action,
1395				    struct xe_pt_walk *walk)
1396{
1397	struct xe_pt *xe_child = container_of(*child, typeof(*xe_child), base);
1398
1399	XE_WARN_ON(!*child);
1400	XE_WARN_ON(!level && xe_child->is_compact);
1401
1402	xe_pt_check_kill(addr, next, level - 1, xe_child, action, walk);
1403
1404	return 0;
1405}
1406
1407static int
1408xe_pt_stage_unbind_post_descend(struct xe_ptw *parent, pgoff_t offset,
1409				unsigned int level, u64 addr, u64 next,
1410				struct xe_ptw **child,
1411				enum page_walk_action *action,
1412				struct xe_pt_walk *walk)
1413{
1414	struct xe_pt_stage_unbind_walk *xe_walk =
1415		container_of(walk, typeof(*xe_walk), base);
1416	struct xe_pt *xe_child = container_of(*child, typeof(*xe_child), base);
1417	pgoff_t end_offset;
1418	u64 size = 1ull << walk->shifts[--level];
1419
1420	if (!IS_ALIGNED(addr, size))
1421		addr = xe_walk->modified_start;
1422	if (!IS_ALIGNED(next, size))
1423		next = xe_walk->modified_end;
1424
1425	/* Parent == *child is the root pt. Don't kill it. */
1426	if (parent != *child &&
1427	    xe_pt_check_kill(addr, next, level, xe_child, action, walk))
1428		return 0;
1429
1430	if (!xe_pt_nonshared_offsets(addr, next, level, walk, action, &offset,
1431				     &end_offset))
1432		return 0;
1433
1434	(void)xe_pt_new_shared(&xe_walk->wupd, xe_child, offset, false);
1435	xe_walk->wupd.updates[level].update->qwords = end_offset - offset;
1436
1437	return 0;
1438}
1439
1440static const struct xe_pt_walk_ops xe_pt_stage_unbind_ops = {
1441	.pt_entry = xe_pt_stage_unbind_entry,
1442	.pt_post_descend = xe_pt_stage_unbind_post_descend,
1443};
1444
1445/**
1446 * xe_pt_stage_unbind() - Build page-table update structures for an unbind
1447 * operation
1448 * @tile: The tile we're unbinding for.
1449 * @vma: The vma we're unbinding.
1450 * @entries: Caller-provided storage for the update structures.
1451 *
1452 * Builds page-table update structures for an unbind operation. The function
1453 * will attempt to remove all page-tables that we're the only user
1454 * of, and for that to work, the unbind operation must be committed in the
1455 * same critical section that blocks racing binds to the same page-table tree.
1456 *
1457 * Return: The number of entries used.
1458 */
1459static unsigned int xe_pt_stage_unbind(struct xe_tile *tile, struct xe_vma *vma,
1460				       struct xe_vm_pgtable_update *entries)
1461{
1462	struct xe_pt_stage_unbind_walk xe_walk = {
1463		.base = {
1464			.ops = &xe_pt_stage_unbind_ops,
1465			.shifts = xe_normal_pt_shifts,
1466			.max_level = XE_PT_HIGHEST_LEVEL,
1467		},
1468		.tile = tile,
1469		.modified_start = xe_vma_start(vma),
1470		.modified_end = xe_vma_end(vma),
1471		.wupd.entries = entries,
1472	};
1473	struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
1474
1475	(void)xe_pt_walk_shared(&pt->base, pt->level, xe_vma_start(vma),
1476				xe_vma_end(vma), &xe_walk.base);
1477
1478	return xe_walk.wupd.num_used_entries;
1479}
1480
1481static void
1482xe_migrate_clear_pgtable_callback(struct xe_migrate_pt_update *pt_update,
1483				  struct xe_tile *tile, struct iosys_map *map,
1484				  void *ptr, u32 qword_ofs, u32 num_qwords,
1485				  const struct xe_vm_pgtable_update *update)
1486{
1487	struct xe_vma *vma = pt_update->vma;
1488	u64 empty = __xe_pt_empty_pte(tile, xe_vma_vm(vma), update->pt->level);
1489	int i;
1490
1491	if (map && map->is_iomem)
1492		for (i = 0; i < num_qwords; ++i)
1493			xe_map_wr(tile_to_xe(tile), map, (qword_ofs + i) *
1494				  sizeof(u64), u64, empty);
1495	else if (map)
1496		memset64(map->vaddr + qword_ofs * sizeof(u64), empty,
1497			 num_qwords);
1498	else
1499		memset64(ptr, empty, num_qwords);
1500}
1501
1502static void
1503xe_pt_commit_unbind(struct xe_vma *vma,
1504		    struct xe_vm_pgtable_update *entries, u32 num_entries,
1505		    struct llist_head *deferred)
1506{
1507	u32 j;
1508
1509	xe_pt_commit_locks_assert(vma);
1510
1511	for (j = 0; j < num_entries; ++j) {
1512		struct xe_vm_pgtable_update *entry = &entries[j];
1513		struct xe_pt *pt = entry->pt;
1514
1515		pt->num_live -= entry->qwords;
1516		if (pt->level) {
1517			struct xe_pt_dir *pt_dir = as_xe_pt_dir(pt);
1518			u32 i;
1519
1520			for (i = entry->ofs; i < entry->ofs + entry->qwords;
1521			     i++) {
1522				if (xe_pt_entry(pt_dir, i))
1523					xe_pt_destroy(xe_pt_entry(pt_dir, i),
1524						      xe_vma_vm(vma)->flags, deferred);
1525
1526				pt_dir->children[i] = NULL;
1527			}
1528		}
1529	}
1530}
1531
1532static const struct xe_migrate_pt_update_ops unbind_ops = {
1533	.populate = xe_migrate_clear_pgtable_callback,
1534	.pre_commit = xe_pt_pre_commit,
1535};
1536
1537static const struct xe_migrate_pt_update_ops userptr_unbind_ops = {
1538	.populate = xe_migrate_clear_pgtable_callback,
1539	.pre_commit = xe_pt_userptr_pre_commit,
1540};
1541
1542/**
1543 * __xe_pt_unbind_vma() - Disconnect and free a page-table tree for the vma
1544 * address range.
1545 * @tile: The tile to unbind for.
1546 * @vma: The vma to unbind.
1547 * @q: The exec_queue with which to do pipelined page-table updates.
1548 * @syncs: Entries to sync on before disconnecting the tree to be destroyed.
1549 * @num_syncs: Number of @sync entries.
1550 *
1551 * This function builds a the xe_vm_pgtable_update entries abstracting the
1552 * operations needed to detach the page-table tree to be destroyed from the
1553 * man vm tree.
1554 * It then takes the relevant locks and submits the operations for
1555 * pipelined detachment of the gpu page-table from  the vm main tree,
1556 * (which can be done either by the cpu and the GPU), Finally it frees the
1557 * detached page-table tree.
1558 *
1559 * Return: A valid dma-fence representing the pipelined detachment operation
1560 * on success, an error pointer on error.
1561 */
1562struct dma_fence *
1563__xe_pt_unbind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue *q,
1564		   struct xe_sync_entry *syncs, u32 num_syncs)
1565{
1566	struct xe_vm_pgtable_update entries[XE_VM_MAX_LEVEL * 2 + 1];
1567	struct xe_pt_migrate_pt_update unbind_pt_update = {
1568		.base = {
1569			.ops = xe_vma_is_userptr(vma) ? &userptr_unbind_ops :
1570			&unbind_ops,
1571			.vma = vma,
1572			.tile_id = tile->id,
1573		},
1574	};
1575	struct xe_vm *vm = xe_vma_vm(vma);
1576	u32 num_entries;
1577	struct dma_fence *fence = NULL;
1578	struct invalidation_fence *ifence;
1579	struct xe_range_fence *rfence;
1580
1581	LLIST_HEAD(deferred);
1582
1583	xe_bo_assert_held(xe_vma_bo(vma));
1584	xe_vm_assert_held(vm);
1585
1586	vm_dbg(&xe_vma_vm(vma)->xe->drm,
1587	       "Preparing unbind, with range [%llx...%llx) engine %p.\n",
1588	       xe_vma_start(vma), xe_vma_end(vma), q);
1589
1590	num_entries = xe_pt_stage_unbind(tile, vma, entries);
1591	xe_tile_assert(tile, num_entries <= ARRAY_SIZE(entries));
1592
1593	xe_vm_dbg_print_entries(tile_to_xe(tile), entries, num_entries);
1594	xe_pt_calc_rfence_interval(vma, &unbind_pt_update, entries,
1595				   num_entries);
1596
1597	ifence = kzalloc(sizeof(*ifence), GFP_KERNEL);
1598	if (!ifence)
1599		return ERR_PTR(-ENOMEM);
1600
1601	rfence = kzalloc(sizeof(*rfence), GFP_KERNEL);
1602	if (!rfence) {
1603		kfree(ifence);
1604		return ERR_PTR(-ENOMEM);
1605	}
1606
1607	/*
1608	 * Even if we were already evicted and unbind to destroy, we need to
1609	 * clear again here. The eviction may have updated pagetables at a
1610	 * lower level, because it needs to be more conservative.
1611	 */
1612	fence = xe_migrate_update_pgtables(tile->migrate,
1613					   vm, NULL, q ? q :
1614					   vm->q[tile->id],
1615					   entries, num_entries,
1616					   syncs, num_syncs,
1617					   &unbind_pt_update.base);
1618	if (!IS_ERR(fence)) {
1619		int err;
1620
1621		err = xe_range_fence_insert(&vm->rftree[tile->id], rfence,
1622					    &xe_range_fence_kfree_ops,
1623					    unbind_pt_update.base.start,
1624					    unbind_pt_update.base.last, fence);
1625		if (err)
1626			dma_fence_wait(fence, false);
1627
1628		/* TLB invalidation must be done before signaling unbind */
1629		err = invalidation_fence_init(tile->primary_gt, ifence, fence, vma);
1630		if (err) {
1631			dma_fence_put(fence);
1632			kfree(ifence);
1633			return ERR_PTR(err);
1634		}
1635		fence = &ifence->base.base;
1636
1637		/* add shared fence now for pagetable delayed destroy */
1638		dma_resv_add_fence(xe_vm_resv(vm), fence,
1639				   DMA_RESV_USAGE_BOOKKEEP);
1640
1641		/* This fence will be installed by caller when doing eviction */
1642		if (!xe_vma_has_no_bo(vma) && !xe_vma_bo(vma)->vm)
1643			dma_resv_add_fence(xe_vma_bo(vma)->ttm.base.resv, fence,
1644					   DMA_RESV_USAGE_BOOKKEEP);
1645		xe_pt_commit_unbind(vma, entries, num_entries,
1646				    unbind_pt_update.locked ? &deferred : NULL);
1647		vma->tile_present &= ~BIT(tile->id);
1648	} else {
1649		kfree(rfence);
1650		kfree(ifence);
1651	}
1652
1653	if (!vma->tile_present)
1654		list_del_init(&vma->combined_links.rebind);
1655
1656	if (unbind_pt_update.locked) {
1657		xe_tile_assert(tile, xe_vma_is_userptr(vma));
1658
1659		if (!vma->tile_present) {
1660			spin_lock(&vm->userptr.invalidated_lock);
1661			list_del_init(&to_userptr_vma(vma)->userptr.invalidate_link);
1662			spin_unlock(&vm->userptr.invalidated_lock);
1663		}
1664		up_read(&vm->userptr.notifier_lock);
1665		xe_bo_put_commit(&deferred);
1666	}
1667
1668	return fence;
1669}