Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: MIT
   2/*
   3 * Copyright © 2021 Intel Corporation
   4 */
   5
   6#include "xe_bo.h"
   7
   8#include <linux/dma-buf.h>
   9
  10#include <drm/drm_drv.h>
  11#include <drm/drm_gem_ttm_helper.h>
  12#include <drm/drm_managed.h>
  13#include <drm/ttm/ttm_device.h>
  14#include <drm/ttm/ttm_placement.h>
  15#include <drm/ttm/ttm_tt.h>
  16#include <drm/xe_drm.h>
  17
  18#include "xe_device.h"
  19#include "xe_dma_buf.h"
  20#include "xe_drm_client.h"
  21#include "xe_ggtt.h"
  22#include "xe_gt.h"
  23#include "xe_map.h"
  24#include "xe_migrate.h"
  25#include "xe_preempt_fence.h"
  26#include "xe_res_cursor.h"
  27#include "xe_trace.h"
  28#include "xe_ttm_stolen_mgr.h"
  29#include "xe_vm.h"
  30
  31const char *const xe_mem_type_to_name[TTM_NUM_MEM_TYPES]  = {
  32	[XE_PL_SYSTEM] = "system",
  33	[XE_PL_TT] = "gtt",
  34	[XE_PL_VRAM0] = "vram0",
  35	[XE_PL_VRAM1] = "vram1",
  36	[XE_PL_STOLEN] = "stolen"
  37};
  38
  39static const struct ttm_place sys_placement_flags = {
  40	.fpfn = 0,
  41	.lpfn = 0,
  42	.mem_type = XE_PL_SYSTEM,
  43	.flags = 0,
  44};
  45
  46static struct ttm_placement sys_placement = {
  47	.num_placement = 1,
  48	.placement = &sys_placement_flags,
  49};
  50
  51static const struct ttm_place tt_placement_flags[] = {
  52	{
  53		.fpfn = 0,
  54		.lpfn = 0,
  55		.mem_type = XE_PL_TT,
  56		.flags = TTM_PL_FLAG_DESIRED,
  57	},
  58	{
  59		.fpfn = 0,
  60		.lpfn = 0,
  61		.mem_type = XE_PL_SYSTEM,
  62		.flags = TTM_PL_FLAG_FALLBACK,
  63	}
  64};
  65
  66static struct ttm_placement tt_placement = {
  67	.num_placement = 2,
  68	.placement = tt_placement_flags,
  69};
  70
  71bool mem_type_is_vram(u32 mem_type)
  72{
  73	return mem_type >= XE_PL_VRAM0 && mem_type != XE_PL_STOLEN;
  74}
  75
  76static bool resource_is_stolen_vram(struct xe_device *xe, struct ttm_resource *res)
  77{
  78	return res->mem_type == XE_PL_STOLEN && IS_DGFX(xe);
  79}
  80
  81static bool resource_is_vram(struct ttm_resource *res)
  82{
  83	return mem_type_is_vram(res->mem_type);
  84}
  85
  86bool xe_bo_is_vram(struct xe_bo *bo)
  87{
  88	return resource_is_vram(bo->ttm.resource) ||
  89		resource_is_stolen_vram(xe_bo_device(bo), bo->ttm.resource);
  90}
  91
  92bool xe_bo_is_stolen(struct xe_bo *bo)
  93{
  94	return bo->ttm.resource->mem_type == XE_PL_STOLEN;
  95}
  96
  97/**
  98 * xe_bo_is_stolen_devmem - check if BO is of stolen type accessed via PCI BAR
  99 * @bo: The BO
 100 *
 101 * The stolen memory is accessed through the PCI BAR for both DGFX and some
 102 * integrated platforms that have a dedicated bit in the PTE for devmem (DM).
 103 *
 104 * Returns: true if it's stolen memory accessed via PCI BAR, false otherwise.
 105 */
 106bool xe_bo_is_stolen_devmem(struct xe_bo *bo)
 107{
 108	return xe_bo_is_stolen(bo) &&
 109		GRAPHICS_VERx100(xe_bo_device(bo)) >= 1270;
 110}
 111
 112static bool xe_bo_is_user(struct xe_bo *bo)
 113{
 114	return bo->flags & XE_BO_CREATE_USER_BIT;
 115}
 116
 117static struct xe_migrate *
 118mem_type_to_migrate(struct xe_device *xe, u32 mem_type)
 119{
 120	struct xe_tile *tile;
 121
 122	xe_assert(xe, mem_type == XE_PL_STOLEN || mem_type_is_vram(mem_type));
 123	tile = &xe->tiles[mem_type == XE_PL_STOLEN ? 0 : (mem_type - XE_PL_VRAM0)];
 124	return tile->migrate;
 125}
 126
 127static struct xe_mem_region *res_to_mem_region(struct ttm_resource *res)
 128{
 129	struct xe_device *xe = ttm_to_xe_device(res->bo->bdev);
 130	struct ttm_resource_manager *mgr;
 131
 132	xe_assert(xe, resource_is_vram(res));
 133	mgr = ttm_manager_type(&xe->ttm, res->mem_type);
 134	return to_xe_ttm_vram_mgr(mgr)->vram;
 135}
 136
 137static void try_add_system(struct xe_device *xe, struct xe_bo *bo,
 138			   u32 bo_flags, u32 *c)
 139{
 140	if (bo_flags & XE_BO_CREATE_SYSTEM_BIT) {
 141		xe_assert(xe, *c < ARRAY_SIZE(bo->placements));
 142
 143		bo->placements[*c] = (struct ttm_place) {
 144			.mem_type = XE_PL_TT,
 145		};
 146		*c += 1;
 147	}
 148}
 149
 150static void add_vram(struct xe_device *xe, struct xe_bo *bo,
 151		     struct ttm_place *places, u32 bo_flags, u32 mem_type, u32 *c)
 152{
 153	struct ttm_place place = { .mem_type = mem_type };
 154	struct xe_mem_region *vram;
 155	u64 io_size;
 156
 157	xe_assert(xe, *c < ARRAY_SIZE(bo->placements));
 158
 159	vram = to_xe_ttm_vram_mgr(ttm_manager_type(&xe->ttm, mem_type))->vram;
 160	xe_assert(xe, vram && vram->usable_size);
 161	io_size = vram->io_size;
 162
 163	/*
 164	 * For eviction / restore on suspend / resume objects
 165	 * pinned in VRAM must be contiguous
 166	 */
 167	if (bo_flags & (XE_BO_CREATE_PINNED_BIT |
 168			XE_BO_CREATE_GGTT_BIT))
 169		place.flags |= TTM_PL_FLAG_CONTIGUOUS;
 170
 171	if (io_size < vram->usable_size) {
 172		if (bo_flags & XE_BO_NEEDS_CPU_ACCESS) {
 173			place.fpfn = 0;
 174			place.lpfn = io_size >> PAGE_SHIFT;
 175		} else {
 176			place.flags |= TTM_PL_FLAG_TOPDOWN;
 177		}
 178	}
 179	places[*c] = place;
 180	*c += 1;
 181}
 182
 183static void try_add_vram(struct xe_device *xe, struct xe_bo *bo,
 184			 u32 bo_flags, u32 *c)
 185{
 186	if (bo_flags & XE_BO_CREATE_VRAM0_BIT)
 187		add_vram(xe, bo, bo->placements, bo_flags, XE_PL_VRAM0, c);
 188	if (bo_flags & XE_BO_CREATE_VRAM1_BIT)
 189		add_vram(xe, bo, bo->placements, bo_flags, XE_PL_VRAM1, c);
 190}
 191
 192static void try_add_stolen(struct xe_device *xe, struct xe_bo *bo,
 193			   u32 bo_flags, u32 *c)
 194{
 195	if (bo_flags & XE_BO_CREATE_STOLEN_BIT) {
 196		xe_assert(xe, *c < ARRAY_SIZE(bo->placements));
 197
 198		bo->placements[*c] = (struct ttm_place) {
 199			.mem_type = XE_PL_STOLEN,
 200			.flags = bo_flags & (XE_BO_CREATE_PINNED_BIT |
 201					     XE_BO_CREATE_GGTT_BIT) ?
 202				TTM_PL_FLAG_CONTIGUOUS : 0,
 203		};
 204		*c += 1;
 205	}
 206}
 207
 208static int __xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo,
 209				       u32 bo_flags)
 210{
 211	u32 c = 0;
 212
 213	try_add_vram(xe, bo, bo_flags, &c);
 214	try_add_system(xe, bo, bo_flags, &c);
 215	try_add_stolen(xe, bo, bo_flags, &c);
 216
 217	if (!c)
 218		return -EINVAL;
 219
 220	bo->placement = (struct ttm_placement) {
 221		.num_placement = c,
 222		.placement = bo->placements,
 223	};
 224
 225	return 0;
 226}
 227
 228int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo,
 229			      u32 bo_flags)
 230{
 231	xe_bo_assert_held(bo);
 232	return __xe_bo_placement_for_flags(xe, bo, bo_flags);
 233}
 234
 235static void xe_evict_flags(struct ttm_buffer_object *tbo,
 236			   struct ttm_placement *placement)
 237{
 238	if (!xe_bo_is_xe_bo(tbo)) {
 239		/* Don't handle scatter gather BOs */
 240		if (tbo->type == ttm_bo_type_sg) {
 241			placement->num_placement = 0;
 242			return;
 243		}
 244
 245		*placement = sys_placement;
 246		return;
 247	}
 248
 249	/*
 250	 * For xe, sg bos that are evicted to system just triggers a
 251	 * rebind of the sg list upon subsequent validation to XE_PL_TT.
 252	 */
 253	switch (tbo->resource->mem_type) {
 254	case XE_PL_VRAM0:
 255	case XE_PL_VRAM1:
 256	case XE_PL_STOLEN:
 257		*placement = tt_placement;
 258		break;
 259	case XE_PL_TT:
 260	default:
 261		*placement = sys_placement;
 262		break;
 263	}
 264}
 265
 266struct xe_ttm_tt {
 267	struct ttm_tt ttm;
 268	struct device *dev;
 269	struct sg_table sgt;
 270	struct sg_table *sg;
 271};
 272
 273static int xe_tt_map_sg(struct ttm_tt *tt)
 274{
 275	struct xe_ttm_tt *xe_tt = container_of(tt, struct xe_ttm_tt, ttm);
 276	unsigned long num_pages = tt->num_pages;
 277	int ret;
 278
 279	XE_WARN_ON(tt->page_flags & TTM_TT_FLAG_EXTERNAL);
 280
 281	if (xe_tt->sg)
 282		return 0;
 283
 284	ret = sg_alloc_table_from_pages_segment(&xe_tt->sgt, tt->pages,
 285						num_pages, 0,
 286						(u64)num_pages << PAGE_SHIFT,
 287						xe_sg_segment_size(xe_tt->dev),
 288						GFP_KERNEL);
 289	if (ret)
 290		return ret;
 291
 292	xe_tt->sg = &xe_tt->sgt;
 293	ret = dma_map_sgtable(xe_tt->dev, xe_tt->sg, DMA_BIDIRECTIONAL,
 294			      DMA_ATTR_SKIP_CPU_SYNC);
 295	if (ret) {
 296		sg_free_table(xe_tt->sg);
 297		xe_tt->sg = NULL;
 298		return ret;
 299	}
 300
 301	return 0;
 302}
 303
 304struct sg_table *xe_bo_sg(struct xe_bo *bo)
 305{
 306	struct ttm_tt *tt = bo->ttm.ttm;
 307	struct xe_ttm_tt *xe_tt = container_of(tt, struct xe_ttm_tt, ttm);
 308
 309	return xe_tt->sg;
 310}
 311
 312static struct ttm_tt *xe_ttm_tt_create(struct ttm_buffer_object *ttm_bo,
 313				       u32 page_flags)
 314{
 315	struct xe_bo *bo = ttm_to_xe_bo(ttm_bo);
 316	struct xe_device *xe = xe_bo_device(bo);
 317	struct xe_ttm_tt *tt;
 318	unsigned long extra_pages;
 319	enum ttm_caching caching;
 320	int err;
 321
 322	tt = kzalloc(sizeof(*tt), GFP_KERNEL);
 323	if (!tt)
 324		return NULL;
 325
 326	tt->dev = xe->drm.dev;
 327
 328	extra_pages = 0;
 329	if (xe_bo_needs_ccs_pages(bo))
 330		extra_pages = DIV_ROUND_UP(xe_device_ccs_bytes(xe, bo->size),
 331					   PAGE_SIZE);
 332
 333	switch (bo->cpu_caching) {
 334	case DRM_XE_GEM_CPU_CACHING_WC:
 335		caching = ttm_write_combined;
 336		break;
 337	default:
 338		caching = ttm_cached;
 339		break;
 340	}
 341
 342	WARN_ON((bo->flags & XE_BO_CREATE_USER_BIT) && !bo->cpu_caching);
 343
 344	/*
 345	 * Display scanout is always non-coherent with the CPU cache.
 346	 *
 347	 * For Xe_LPG and beyond, PPGTT PTE lookups are also non-coherent and
 348	 * require a CPU:WC mapping.
 349	 */
 350	if ((!bo->cpu_caching && bo->flags & XE_BO_SCANOUT_BIT) ||
 351	    (xe->info.graphics_verx100 >= 1270 && bo->flags & XE_BO_PAGETABLE))
 352		caching = ttm_write_combined;
 353
 354	err = ttm_tt_init(&tt->ttm, &bo->ttm, page_flags, caching, extra_pages);
 355	if (err) {
 356		kfree(tt);
 357		return NULL;
 358	}
 359
 360	return &tt->ttm;
 361}
 362
 363static int xe_ttm_tt_populate(struct ttm_device *ttm_dev, struct ttm_tt *tt,
 364			      struct ttm_operation_ctx *ctx)
 365{
 366	int err;
 367
 368	/*
 369	 * dma-bufs are not populated with pages, and the dma-
 370	 * addresses are set up when moved to XE_PL_TT.
 371	 */
 372	if (tt->page_flags & TTM_TT_FLAG_EXTERNAL)
 373		return 0;
 374
 375	err = ttm_pool_alloc(&ttm_dev->pool, tt, ctx);
 376	if (err)
 377		return err;
 378
 379	/* A follow up may move this xe_bo_move when BO is moved to XE_PL_TT */
 380	err = xe_tt_map_sg(tt);
 381	if (err)
 382		ttm_pool_free(&ttm_dev->pool, tt);
 383
 384	return err;
 385}
 386
 387static void xe_ttm_tt_unpopulate(struct ttm_device *ttm_dev, struct ttm_tt *tt)
 388{
 389	struct xe_ttm_tt *xe_tt = container_of(tt, struct xe_ttm_tt, ttm);
 390
 391	if (tt->page_flags & TTM_TT_FLAG_EXTERNAL)
 392		return;
 393
 394	if (xe_tt->sg) {
 395		dma_unmap_sgtable(xe_tt->dev, xe_tt->sg,
 396				  DMA_BIDIRECTIONAL, 0);
 397		sg_free_table(xe_tt->sg);
 398		xe_tt->sg = NULL;
 399	}
 400
 401	return ttm_pool_free(&ttm_dev->pool, tt);
 402}
 403
 404static void xe_ttm_tt_destroy(struct ttm_device *ttm_dev, struct ttm_tt *tt)
 405{
 406	ttm_tt_fini(tt);
 407	kfree(tt);
 408}
 409
 410static int xe_ttm_io_mem_reserve(struct ttm_device *bdev,
 411				 struct ttm_resource *mem)
 412{
 413	struct xe_device *xe = ttm_to_xe_device(bdev);
 414
 415	switch (mem->mem_type) {
 416	case XE_PL_SYSTEM:
 417	case XE_PL_TT:
 418		return 0;
 419	case XE_PL_VRAM0:
 420	case XE_PL_VRAM1: {
 421		struct xe_ttm_vram_mgr_resource *vres =
 422			to_xe_ttm_vram_mgr_resource(mem);
 423		struct xe_mem_region *vram = res_to_mem_region(mem);
 424
 425		if (vres->used_visible_size < mem->size)
 426			return -EINVAL;
 427
 428		mem->bus.offset = mem->start << PAGE_SHIFT;
 429
 430		if (vram->mapping &&
 431		    mem->placement & TTM_PL_FLAG_CONTIGUOUS)
 432			mem->bus.addr = (u8 __force *)vram->mapping +
 433				mem->bus.offset;
 434
 435		mem->bus.offset += vram->io_start;
 436		mem->bus.is_iomem = true;
 437
 438#if  !defined(CONFIG_X86)
 439		mem->bus.caching = ttm_write_combined;
 440#endif
 441		return 0;
 442	} case XE_PL_STOLEN:
 443		return xe_ttm_stolen_io_mem_reserve(xe, mem);
 444	default:
 445		return -EINVAL;
 446	}
 447}
 448
 449static int xe_bo_trigger_rebind(struct xe_device *xe, struct xe_bo *bo,
 450				const struct ttm_operation_ctx *ctx)
 451{
 452	struct dma_resv_iter cursor;
 453	struct dma_fence *fence;
 454	struct drm_gem_object *obj = &bo->ttm.base;
 455	struct drm_gpuvm_bo *vm_bo;
 456	bool idle = false;
 457	int ret = 0;
 458
 459	dma_resv_assert_held(bo->ttm.base.resv);
 460
 461	if (!list_empty(&bo->ttm.base.gpuva.list)) {
 462		dma_resv_iter_begin(&cursor, bo->ttm.base.resv,
 463				    DMA_RESV_USAGE_BOOKKEEP);
 464		dma_resv_for_each_fence_unlocked(&cursor, fence)
 465			dma_fence_enable_sw_signaling(fence);
 466		dma_resv_iter_end(&cursor);
 467	}
 468
 469	drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
 470		struct xe_vm *vm = gpuvm_to_vm(vm_bo->vm);
 471		struct drm_gpuva *gpuva;
 472
 473		if (!xe_vm_in_fault_mode(vm)) {
 474			drm_gpuvm_bo_evict(vm_bo, true);
 475			continue;
 476		}
 477
 478		if (!idle) {
 479			long timeout;
 480
 481			if (ctx->no_wait_gpu &&
 482			    !dma_resv_test_signaled(bo->ttm.base.resv,
 483						    DMA_RESV_USAGE_BOOKKEEP))
 484				return -EBUSY;
 485
 486			timeout = dma_resv_wait_timeout(bo->ttm.base.resv,
 487							DMA_RESV_USAGE_BOOKKEEP,
 488							ctx->interruptible,
 489							MAX_SCHEDULE_TIMEOUT);
 490			if (!timeout)
 491				return -ETIME;
 492			if (timeout < 0)
 493				return timeout;
 494
 495			idle = true;
 496		}
 497
 498		drm_gpuvm_bo_for_each_va(gpuva, vm_bo) {
 499			struct xe_vma *vma = gpuva_to_vma(gpuva);
 500
 501			trace_xe_vma_evict(vma);
 502			ret = xe_vm_invalidate_vma(vma);
 503			if (XE_WARN_ON(ret))
 504				return ret;
 505		}
 506	}
 507
 508	return ret;
 509}
 510
 511/*
 512 * The dma-buf map_attachment() / unmap_attachment() is hooked up here.
 513 * Note that unmapping the attachment is deferred to the next
 514 * map_attachment time, or to bo destroy (after idling) whichever comes first.
 515 * This is to avoid syncing before unmap_attachment(), assuming that the
 516 * caller relies on idling the reservation object before moving the
 517 * backing store out. Should that assumption not hold, then we will be able
 518 * to unconditionally call unmap_attachment() when moving out to system.
 519 */
 520static int xe_bo_move_dmabuf(struct ttm_buffer_object *ttm_bo,
 521			     struct ttm_resource *new_res)
 522{
 523	struct dma_buf_attachment *attach = ttm_bo->base.import_attach;
 524	struct xe_ttm_tt *xe_tt = container_of(ttm_bo->ttm, struct xe_ttm_tt,
 525					       ttm);
 526	struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
 527	struct sg_table *sg;
 528
 529	xe_assert(xe, attach);
 530	xe_assert(xe, ttm_bo->ttm);
 531
 532	if (new_res->mem_type == XE_PL_SYSTEM)
 533		goto out;
 534
 535	if (ttm_bo->sg) {
 536		dma_buf_unmap_attachment(attach, ttm_bo->sg, DMA_BIDIRECTIONAL);
 537		ttm_bo->sg = NULL;
 538	}
 539
 540	sg = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
 541	if (IS_ERR(sg))
 542		return PTR_ERR(sg);
 543
 544	ttm_bo->sg = sg;
 545	xe_tt->sg = sg;
 546
 547out:
 548	ttm_bo_move_null(ttm_bo, new_res);
 549
 550	return 0;
 551}
 552
 553/**
 554 * xe_bo_move_notify - Notify subsystems of a pending move
 555 * @bo: The buffer object
 556 * @ctx: The struct ttm_operation_ctx controlling locking and waits.
 557 *
 558 * This function notifies subsystems of an upcoming buffer move.
 559 * Upon receiving such a notification, subsystems should schedule
 560 * halting access to the underlying pages and optionally add a fence
 561 * to the buffer object's dma_resv object, that signals when access is
 562 * stopped. The caller will wait on all dma_resv fences before
 563 * starting the move.
 564 *
 565 * A subsystem may commence access to the object after obtaining
 566 * bindings to the new backing memory under the object lock.
 567 *
 568 * Return: 0 on success, -EINTR or -ERESTARTSYS if interrupted in fault mode,
 569 * negative error code on error.
 570 */
 571static int xe_bo_move_notify(struct xe_bo *bo,
 572			     const struct ttm_operation_ctx *ctx)
 573{
 574	struct ttm_buffer_object *ttm_bo = &bo->ttm;
 575	struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
 576	struct ttm_resource *old_mem = ttm_bo->resource;
 577	u32 old_mem_type = old_mem ? old_mem->mem_type : XE_PL_SYSTEM;
 578	int ret;
 579
 580	/*
 581	 * If this starts to call into many components, consider
 582	 * using a notification chain here.
 583	 */
 584
 585	if (xe_bo_is_pinned(bo))
 586		return -EINVAL;
 587
 588	xe_bo_vunmap(bo);
 589	ret = xe_bo_trigger_rebind(xe, bo, ctx);
 590	if (ret)
 591		return ret;
 592
 593	/* Don't call move_notify() for imported dma-bufs. */
 594	if (ttm_bo->base.dma_buf && !ttm_bo->base.import_attach)
 595		dma_buf_move_notify(ttm_bo->base.dma_buf);
 596
 597	/*
 598	 * TTM has already nuked the mmap for us (see ttm_bo_unmap_virtual),
 599	 * so if we moved from VRAM make sure to unlink this from the userfault
 600	 * tracking.
 601	 */
 602	if (mem_type_is_vram(old_mem_type)) {
 603		mutex_lock(&xe->mem_access.vram_userfault.lock);
 604		if (!list_empty(&bo->vram_userfault_link))
 605			list_del_init(&bo->vram_userfault_link);
 606		mutex_unlock(&xe->mem_access.vram_userfault.lock);
 607	}
 608
 609	return 0;
 610}
 611
 612static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict,
 613		      struct ttm_operation_ctx *ctx,
 614		      struct ttm_resource *new_mem,
 615		      struct ttm_place *hop)
 616{
 617	struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
 618	struct xe_bo *bo = ttm_to_xe_bo(ttm_bo);
 619	struct ttm_resource *old_mem = ttm_bo->resource;
 620	u32 old_mem_type = old_mem ? old_mem->mem_type : XE_PL_SYSTEM;
 621	struct ttm_tt *ttm = ttm_bo->ttm;
 622	struct xe_migrate *migrate = NULL;
 623	struct dma_fence *fence;
 624	bool move_lacks_source;
 625	bool tt_has_data;
 626	bool needs_clear;
 627	bool handle_system_ccs = (!IS_DGFX(xe) && xe_bo_needs_ccs_pages(bo) &&
 628				  ttm && ttm_tt_is_populated(ttm)) ? true : false;
 629	int ret = 0;
 630	/* Bo creation path, moving to system or TT. */
 631	if ((!old_mem && ttm) && !handle_system_ccs) {
 632		ttm_bo_move_null(ttm_bo, new_mem);
 633		return 0;
 634	}
 635
 636	if (ttm_bo->type == ttm_bo_type_sg) {
 637		ret = xe_bo_move_notify(bo, ctx);
 638		if (!ret)
 639			ret = xe_bo_move_dmabuf(ttm_bo, new_mem);
 640		goto out;
 641	}
 642
 643	tt_has_data = ttm && (ttm_tt_is_populated(ttm) ||
 644			      (ttm->page_flags & TTM_TT_FLAG_SWAPPED));
 645
 646	move_lacks_source = handle_system_ccs ? (!bo->ccs_cleared)  :
 647						(!mem_type_is_vram(old_mem_type) && !tt_has_data);
 648
 649	needs_clear = (ttm && ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC) ||
 650		(!ttm && ttm_bo->type == ttm_bo_type_device);
 651
 652	if ((move_lacks_source && !needs_clear)) {
 653		ttm_bo_move_null(ttm_bo, new_mem);
 654		goto out;
 655	}
 656
 657	if (old_mem_type == XE_PL_SYSTEM && new_mem->mem_type == XE_PL_TT && !handle_system_ccs) {
 658		ttm_bo_move_null(ttm_bo, new_mem);
 659		goto out;
 660	}
 661
 662	/*
 663	 * Failed multi-hop where the old_mem is still marked as
 664	 * TTM_PL_FLAG_TEMPORARY, should just be a dummy move.
 665	 */
 666	if (old_mem_type == XE_PL_TT &&
 667	    new_mem->mem_type == XE_PL_TT) {
 668		ttm_bo_move_null(ttm_bo, new_mem);
 669		goto out;
 670	}
 671
 672	if (!move_lacks_source && !xe_bo_is_pinned(bo)) {
 673		ret = xe_bo_move_notify(bo, ctx);
 674		if (ret)
 675			goto out;
 676	}
 677
 678	if (old_mem_type == XE_PL_TT &&
 679	    new_mem->mem_type == XE_PL_SYSTEM) {
 680		long timeout = dma_resv_wait_timeout(ttm_bo->base.resv,
 681						     DMA_RESV_USAGE_BOOKKEEP,
 682						     true,
 683						     MAX_SCHEDULE_TIMEOUT);
 684		if (timeout < 0) {
 685			ret = timeout;
 686			goto out;
 687		}
 688
 689		if (!handle_system_ccs) {
 690			ttm_bo_move_null(ttm_bo, new_mem);
 691			goto out;
 692		}
 693	}
 694
 695	if (!move_lacks_source &&
 696	    ((old_mem_type == XE_PL_SYSTEM && resource_is_vram(new_mem)) ||
 697	     (mem_type_is_vram(old_mem_type) &&
 698	      new_mem->mem_type == XE_PL_SYSTEM))) {
 699		hop->fpfn = 0;
 700		hop->lpfn = 0;
 701		hop->mem_type = XE_PL_TT;
 702		hop->flags = TTM_PL_FLAG_TEMPORARY;
 703		ret = -EMULTIHOP;
 704		goto out;
 705	}
 706
 707	if (bo->tile)
 708		migrate = bo->tile->migrate;
 709	else if (resource_is_vram(new_mem))
 710		migrate = mem_type_to_migrate(xe, new_mem->mem_type);
 711	else if (mem_type_is_vram(old_mem_type))
 712		migrate = mem_type_to_migrate(xe, old_mem_type);
 713	else
 714		migrate = xe->tiles[0].migrate;
 715
 716	xe_assert(xe, migrate);
 717	trace_xe_bo_move(bo, new_mem->mem_type, old_mem_type, move_lacks_source);
 718	xe_device_mem_access_get(xe);
 719
 720	if (xe_bo_is_pinned(bo) && !xe_bo_is_user(bo)) {
 721		/*
 722		 * Kernel memory that is pinned should only be moved on suspend
 723		 * / resume, some of the pinned memory is required for the
 724		 * device to resume / use the GPU to move other evicted memory
 725		 * (user memory) around. This likely could be optimized a bit
 726		 * futher where we find the minimum set of pinned memory
 727		 * required for resume but for simplity doing a memcpy for all
 728		 * pinned memory.
 729		 */
 730		ret = xe_bo_vmap(bo);
 731		if (!ret) {
 732			ret = ttm_bo_move_memcpy(ttm_bo, ctx, new_mem);
 733
 734			/* Create a new VMAP once kernel BO back in VRAM */
 735			if (!ret && resource_is_vram(new_mem)) {
 736				struct xe_mem_region *vram = res_to_mem_region(new_mem);
 737				void __iomem *new_addr = vram->mapping +
 738					(new_mem->start << PAGE_SHIFT);
 739
 740				if (XE_WARN_ON(new_mem->start == XE_BO_INVALID_OFFSET)) {
 741					ret = -EINVAL;
 742					xe_device_mem_access_put(xe);
 743					goto out;
 744				}
 745
 746				xe_assert(xe, new_mem->start ==
 747					  bo->placements->fpfn);
 748
 749				iosys_map_set_vaddr_iomem(&bo->vmap, new_addr);
 750			}
 751		}
 752	} else {
 753		if (move_lacks_source)
 754			fence = xe_migrate_clear(migrate, bo, new_mem);
 755		else
 756			fence = xe_migrate_copy(migrate, bo, bo, old_mem,
 757						new_mem, handle_system_ccs);
 758		if (IS_ERR(fence)) {
 759			ret = PTR_ERR(fence);
 760			xe_device_mem_access_put(xe);
 761			goto out;
 762		}
 763		if (!move_lacks_source) {
 764			ret = ttm_bo_move_accel_cleanup(ttm_bo, fence, evict,
 765							true, new_mem);
 766			if (ret) {
 767				dma_fence_wait(fence, false);
 768				ttm_bo_move_null(ttm_bo, new_mem);
 769				ret = 0;
 770			}
 771		} else {
 772			/*
 773			 * ttm_bo_move_accel_cleanup() may blow up if
 774			 * bo->resource == NULL, so just attach the
 775			 * fence and set the new resource.
 776			 */
 777			dma_resv_add_fence(ttm_bo->base.resv, fence,
 778					   DMA_RESV_USAGE_KERNEL);
 779			ttm_bo_move_null(ttm_bo, new_mem);
 780		}
 781
 782		dma_fence_put(fence);
 783	}
 784
 785	xe_device_mem_access_put(xe);
 786
 787out:
 788	return ret;
 789
 790}
 791
 792/**
 793 * xe_bo_evict_pinned() - Evict a pinned VRAM object to system memory
 794 * @bo: The buffer object to move.
 795 *
 796 * On successful completion, the object memory will be moved to sytem memory.
 797 * This function blocks until the object has been fully moved.
 798 *
 799 * This is needed to for special handling of pinned VRAM object during
 800 * suspend-resume.
 801 *
 802 * Return: 0 on success. Negative error code on failure.
 803 */
 804int xe_bo_evict_pinned(struct xe_bo *bo)
 805{
 806	struct ttm_place place = {
 807		.mem_type = XE_PL_TT,
 808	};
 809	struct ttm_placement placement = {
 810		.placement = &place,
 811		.num_placement = 1,
 812	};
 813	struct ttm_operation_ctx ctx = {
 814		.interruptible = false,
 815	};
 816	struct ttm_resource *new_mem;
 817	int ret;
 818
 819	xe_bo_assert_held(bo);
 820
 821	if (WARN_ON(!bo->ttm.resource))
 822		return -EINVAL;
 823
 824	if (WARN_ON(!xe_bo_is_pinned(bo)))
 825		return -EINVAL;
 826
 827	if (WARN_ON(!xe_bo_is_vram(bo)))
 828		return -EINVAL;
 829
 830	ret = ttm_bo_mem_space(&bo->ttm, &placement, &new_mem, &ctx);
 831	if (ret)
 832		return ret;
 833
 834	if (!bo->ttm.ttm) {
 835		bo->ttm.ttm = xe_ttm_tt_create(&bo->ttm, 0);
 836		if (!bo->ttm.ttm) {
 837			ret = -ENOMEM;
 838			goto err_res_free;
 839		}
 840	}
 841
 842	ret = ttm_tt_populate(bo->ttm.bdev, bo->ttm.ttm, &ctx);
 843	if (ret)
 844		goto err_res_free;
 845
 846	ret = dma_resv_reserve_fences(bo->ttm.base.resv, 1);
 847	if (ret)
 848		goto err_res_free;
 849
 850	ret = xe_bo_move(&bo->ttm, false, &ctx, new_mem, NULL);
 851	if (ret)
 852		goto err_res_free;
 853
 854	dma_resv_wait_timeout(bo->ttm.base.resv, DMA_RESV_USAGE_KERNEL,
 855			      false, MAX_SCHEDULE_TIMEOUT);
 856
 857	return 0;
 858
 859err_res_free:
 860	ttm_resource_free(&bo->ttm, &new_mem);
 861	return ret;
 862}
 863
 864/**
 865 * xe_bo_restore_pinned() - Restore a pinned VRAM object
 866 * @bo: The buffer object to move.
 867 *
 868 * On successful completion, the object memory will be moved back to VRAM.
 869 * This function blocks until the object has been fully moved.
 870 *
 871 * This is needed to for special handling of pinned VRAM object during
 872 * suspend-resume.
 873 *
 874 * Return: 0 on success. Negative error code on failure.
 875 */
 876int xe_bo_restore_pinned(struct xe_bo *bo)
 877{
 878	struct ttm_operation_ctx ctx = {
 879		.interruptible = false,
 880	};
 881	struct ttm_resource *new_mem;
 882	int ret;
 883
 884	xe_bo_assert_held(bo);
 885
 886	if (WARN_ON(!bo->ttm.resource))
 887		return -EINVAL;
 888
 889	if (WARN_ON(!xe_bo_is_pinned(bo)))
 890		return -EINVAL;
 891
 892	if (WARN_ON(xe_bo_is_vram(bo) || !bo->ttm.ttm))
 893		return -EINVAL;
 894
 895	ret = ttm_bo_mem_space(&bo->ttm, &bo->placement, &new_mem, &ctx);
 896	if (ret)
 897		return ret;
 898
 899	ret = ttm_tt_populate(bo->ttm.bdev, bo->ttm.ttm, &ctx);
 900	if (ret)
 901		goto err_res_free;
 902
 903	ret = dma_resv_reserve_fences(bo->ttm.base.resv, 1);
 904	if (ret)
 905		goto err_res_free;
 906
 907	ret = xe_bo_move(&bo->ttm, false, &ctx, new_mem, NULL);
 908	if (ret)
 909		goto err_res_free;
 910
 911	dma_resv_wait_timeout(bo->ttm.base.resv, DMA_RESV_USAGE_KERNEL,
 912			      false, MAX_SCHEDULE_TIMEOUT);
 913
 914	return 0;
 915
 916err_res_free:
 917	ttm_resource_free(&bo->ttm, &new_mem);
 918	return ret;
 919}
 920
 921static unsigned long xe_ttm_io_mem_pfn(struct ttm_buffer_object *ttm_bo,
 922				       unsigned long page_offset)
 923{
 924	struct xe_bo *bo = ttm_to_xe_bo(ttm_bo);
 925	struct xe_res_cursor cursor;
 926	struct xe_mem_region *vram;
 927
 928	if (ttm_bo->resource->mem_type == XE_PL_STOLEN)
 929		return xe_ttm_stolen_io_offset(bo, page_offset << PAGE_SHIFT) >> PAGE_SHIFT;
 930
 931	vram = res_to_mem_region(ttm_bo->resource);
 932	xe_res_first(ttm_bo->resource, (u64)page_offset << PAGE_SHIFT, 0, &cursor);
 933	return (vram->io_start + cursor.start) >> PAGE_SHIFT;
 934}
 935
 936static void __xe_bo_vunmap(struct xe_bo *bo);
 937
 938/*
 939 * TODO: Move this function to TTM so we don't rely on how TTM does its
 940 * locking, thereby abusing TTM internals.
 941 */
 942static bool xe_ttm_bo_lock_in_destructor(struct ttm_buffer_object *ttm_bo)
 943{
 944	struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
 945	bool locked;
 946
 947	xe_assert(xe, !kref_read(&ttm_bo->kref));
 948
 949	/*
 950	 * We can typically only race with TTM trylocking under the
 951	 * lru_lock, which will immediately be unlocked again since
 952	 * the ttm_bo refcount is zero at this point. So trylocking *should*
 953	 * always succeed here, as long as we hold the lru lock.
 954	 */
 955	spin_lock(&ttm_bo->bdev->lru_lock);
 956	locked = dma_resv_trylock(ttm_bo->base.resv);
 957	spin_unlock(&ttm_bo->bdev->lru_lock);
 958	xe_assert(xe, locked);
 959
 960	return locked;
 961}
 962
 963static void xe_ttm_bo_release_notify(struct ttm_buffer_object *ttm_bo)
 964{
 965	struct dma_resv_iter cursor;
 966	struct dma_fence *fence;
 967	struct dma_fence *replacement = NULL;
 968	struct xe_bo *bo;
 969
 970	if (!xe_bo_is_xe_bo(ttm_bo))
 971		return;
 972
 973	bo = ttm_to_xe_bo(ttm_bo);
 974	xe_assert(xe_bo_device(bo), !(bo->created && kref_read(&ttm_bo->base.refcount)));
 975
 976	/*
 977	 * Corner case where TTM fails to allocate memory and this BOs resv
 978	 * still points the VMs resv
 979	 */
 980	if (ttm_bo->base.resv != &ttm_bo->base._resv)
 981		return;
 982
 983	if (!xe_ttm_bo_lock_in_destructor(ttm_bo))
 984		return;
 985
 986	/*
 987	 * Scrub the preempt fences if any. The unbind fence is already
 988	 * attached to the resv.
 989	 * TODO: Don't do this for external bos once we scrub them after
 990	 * unbind.
 991	 */
 992	dma_resv_for_each_fence(&cursor, ttm_bo->base.resv,
 993				DMA_RESV_USAGE_BOOKKEEP, fence) {
 994		if (xe_fence_is_xe_preempt(fence) &&
 995		    !dma_fence_is_signaled(fence)) {
 996			if (!replacement)
 997				replacement = dma_fence_get_stub();
 998
 999			dma_resv_replace_fences(ttm_bo->base.resv,
1000						fence->context,
1001						replacement,
1002						DMA_RESV_USAGE_BOOKKEEP);
1003		}
1004	}
1005	dma_fence_put(replacement);
1006
1007	dma_resv_unlock(ttm_bo->base.resv);
1008}
1009
1010static void xe_ttm_bo_delete_mem_notify(struct ttm_buffer_object *ttm_bo)
1011{
1012	if (!xe_bo_is_xe_bo(ttm_bo))
1013		return;
1014
1015	/*
1016	 * Object is idle and about to be destroyed. Release the
1017	 * dma-buf attachment.
1018	 */
1019	if (ttm_bo->type == ttm_bo_type_sg && ttm_bo->sg) {
1020		struct xe_ttm_tt *xe_tt = container_of(ttm_bo->ttm,
1021						       struct xe_ttm_tt, ttm);
1022
1023		dma_buf_unmap_attachment(ttm_bo->base.import_attach, ttm_bo->sg,
1024					 DMA_BIDIRECTIONAL);
1025		ttm_bo->sg = NULL;
1026		xe_tt->sg = NULL;
1027	}
1028}
1029
1030const struct ttm_device_funcs xe_ttm_funcs = {
1031	.ttm_tt_create = xe_ttm_tt_create,
1032	.ttm_tt_populate = xe_ttm_tt_populate,
1033	.ttm_tt_unpopulate = xe_ttm_tt_unpopulate,
1034	.ttm_tt_destroy = xe_ttm_tt_destroy,
1035	.evict_flags = xe_evict_flags,
1036	.move = xe_bo_move,
1037	.io_mem_reserve = xe_ttm_io_mem_reserve,
1038	.io_mem_pfn = xe_ttm_io_mem_pfn,
1039	.release_notify = xe_ttm_bo_release_notify,
1040	.eviction_valuable = ttm_bo_eviction_valuable,
1041	.delete_mem_notify = xe_ttm_bo_delete_mem_notify,
1042};
1043
1044static void xe_ttm_bo_destroy(struct ttm_buffer_object *ttm_bo)
1045{
1046	struct xe_bo *bo = ttm_to_xe_bo(ttm_bo);
1047	struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
1048
1049	if (bo->ttm.base.import_attach)
1050		drm_prime_gem_destroy(&bo->ttm.base, NULL);
1051	drm_gem_object_release(&bo->ttm.base);
1052
1053	xe_assert(xe, list_empty(&ttm_bo->base.gpuva.list));
1054
1055	if (bo->ggtt_node.size)
1056		xe_ggtt_remove_bo(bo->tile->mem.ggtt, bo);
1057
1058#ifdef CONFIG_PROC_FS
1059	if (bo->client)
1060		xe_drm_client_remove_bo(bo);
1061#endif
1062
1063	if (bo->vm && xe_bo_is_user(bo))
1064		xe_vm_put(bo->vm);
1065
1066	mutex_lock(&xe->mem_access.vram_userfault.lock);
1067	if (!list_empty(&bo->vram_userfault_link))
1068		list_del(&bo->vram_userfault_link);
1069	mutex_unlock(&xe->mem_access.vram_userfault.lock);
1070
1071	kfree(bo);
1072}
1073
1074static void xe_gem_object_free(struct drm_gem_object *obj)
1075{
1076	/* Our BO reference counting scheme works as follows:
1077	 *
1078	 * The gem object kref is typically used throughout the driver,
1079	 * and the gem object holds a ttm_buffer_object refcount, so
1080	 * that when the last gem object reference is put, which is when
1081	 * we end up in this function, we put also that ttm_buffer_object
1082	 * refcount. Anything using gem interfaces is then no longer
1083	 * allowed to access the object in a way that requires a gem
1084	 * refcount, including locking the object.
1085	 *
1086	 * driver ttm callbacks is allowed to use the ttm_buffer_object
1087	 * refcount directly if needed.
1088	 */
1089	__xe_bo_vunmap(gem_to_xe_bo(obj));
1090	ttm_bo_put(container_of(obj, struct ttm_buffer_object, base));
1091}
1092
1093static void xe_gem_object_close(struct drm_gem_object *obj,
1094				struct drm_file *file_priv)
1095{
1096	struct xe_bo *bo = gem_to_xe_bo(obj);
1097
1098	if (bo->vm && !xe_vm_in_fault_mode(bo->vm)) {
1099		xe_assert(xe_bo_device(bo), xe_bo_is_user(bo));
1100
1101		xe_bo_lock(bo, false);
1102		ttm_bo_set_bulk_move(&bo->ttm, NULL);
1103		xe_bo_unlock(bo);
1104	}
1105}
1106
1107static vm_fault_t xe_gem_fault(struct vm_fault *vmf)
1108{
1109	struct ttm_buffer_object *tbo = vmf->vma->vm_private_data;
1110	struct drm_device *ddev = tbo->base.dev;
1111	struct xe_device *xe = to_xe_device(ddev);
1112	struct xe_bo *bo = ttm_to_xe_bo(tbo);
1113	bool needs_rpm = bo->flags & XE_BO_CREATE_VRAM_MASK;
1114	vm_fault_t ret;
1115	int idx;
1116
1117	if (needs_rpm)
1118		xe_device_mem_access_get(xe);
1119
1120	ret = ttm_bo_vm_reserve(tbo, vmf);
1121	if (ret)
1122		goto out;
1123
1124	if (drm_dev_enter(ddev, &idx)) {
1125		trace_xe_bo_cpu_fault(bo);
1126
1127		ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
1128					       TTM_BO_VM_NUM_PREFAULT);
1129		drm_dev_exit(idx);
1130	} else {
1131		ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
1132	}
1133
1134	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
1135		goto out;
1136	/*
1137	 * ttm_bo_vm_reserve() already has dma_resv_lock.
1138	 */
1139	if (ret == VM_FAULT_NOPAGE && mem_type_is_vram(tbo->resource->mem_type)) {
1140		mutex_lock(&xe->mem_access.vram_userfault.lock);
1141		if (list_empty(&bo->vram_userfault_link))
1142			list_add(&bo->vram_userfault_link, &xe->mem_access.vram_userfault.list);
1143		mutex_unlock(&xe->mem_access.vram_userfault.lock);
1144	}
1145
1146	dma_resv_unlock(tbo->base.resv);
1147out:
1148	if (needs_rpm)
1149		xe_device_mem_access_put(xe);
1150
1151	return ret;
1152}
1153
1154static const struct vm_operations_struct xe_gem_vm_ops = {
1155	.fault = xe_gem_fault,
1156	.open = ttm_bo_vm_open,
1157	.close = ttm_bo_vm_close,
1158	.access = ttm_bo_vm_access
1159};
1160
1161static const struct drm_gem_object_funcs xe_gem_object_funcs = {
1162	.free = xe_gem_object_free,
1163	.close = xe_gem_object_close,
1164	.mmap = drm_gem_ttm_mmap,
1165	.export = xe_gem_prime_export,
1166	.vm_ops = &xe_gem_vm_ops,
1167};
1168
1169/**
1170 * xe_bo_alloc - Allocate storage for a struct xe_bo
1171 *
1172 * This funcition is intended to allocate storage to be used for input
1173 * to __xe_bo_create_locked(), in the case a pointer to the bo to be
1174 * created is needed before the call to __xe_bo_create_locked().
1175 * If __xe_bo_create_locked ends up never to be called, then the
1176 * storage allocated with this function needs to be freed using
1177 * xe_bo_free().
1178 *
1179 * Return: A pointer to an uninitialized struct xe_bo on success,
1180 * ERR_PTR(-ENOMEM) on error.
1181 */
1182struct xe_bo *xe_bo_alloc(void)
1183{
1184	struct xe_bo *bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1185
1186	if (!bo)
1187		return ERR_PTR(-ENOMEM);
1188
1189	return bo;
1190}
1191
1192/**
1193 * xe_bo_free - Free storage allocated using xe_bo_alloc()
1194 * @bo: The buffer object storage.
1195 *
1196 * Refer to xe_bo_alloc() documentation for valid use-cases.
1197 */
1198void xe_bo_free(struct xe_bo *bo)
1199{
1200	kfree(bo);
1201}
1202
1203struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
1204				     struct xe_tile *tile, struct dma_resv *resv,
1205				     struct ttm_lru_bulk_move *bulk, size_t size,
1206				     u16 cpu_caching, enum ttm_bo_type type,
1207				     u32 flags)
1208{
1209	struct ttm_operation_ctx ctx = {
1210		.interruptible = true,
1211		.no_wait_gpu = false,
1212	};
1213	struct ttm_placement *placement;
1214	uint32_t alignment;
1215	size_t aligned_size;
1216	int err;
1217
1218	/* Only kernel objects should set GT */
1219	xe_assert(xe, !tile || type == ttm_bo_type_kernel);
1220
1221	if (XE_WARN_ON(!size)) {
1222		xe_bo_free(bo);
1223		return ERR_PTR(-EINVAL);
1224	}
1225
1226	if (flags & (XE_BO_CREATE_VRAM_MASK | XE_BO_CREATE_STOLEN_BIT) &&
1227	    !(flags & XE_BO_CREATE_IGNORE_MIN_PAGE_SIZE_BIT) &&
1228	    xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K) {
1229		aligned_size = ALIGN(size, SZ_64K);
1230		if (type != ttm_bo_type_device)
1231			size = ALIGN(size, SZ_64K);
1232		flags |= XE_BO_INTERNAL_64K;
1233		alignment = SZ_64K >> PAGE_SHIFT;
1234
1235	} else {
1236		aligned_size = ALIGN(size, SZ_4K);
1237		flags &= ~XE_BO_INTERNAL_64K;
1238		alignment = SZ_4K >> PAGE_SHIFT;
1239	}
1240
1241	if (type == ttm_bo_type_device && aligned_size != size)
1242		return ERR_PTR(-EINVAL);
1243
1244	if (!bo) {
1245		bo = xe_bo_alloc();
1246		if (IS_ERR(bo))
1247			return bo;
1248	}
1249
1250	bo->ccs_cleared = false;
1251	bo->tile = tile;
1252	bo->size = size;
1253	bo->flags = flags;
1254	bo->cpu_caching = cpu_caching;
1255	bo->ttm.base.funcs = &xe_gem_object_funcs;
1256	bo->ttm.priority = XE_BO_PRIORITY_NORMAL;
1257	INIT_LIST_HEAD(&bo->pinned_link);
1258#ifdef CONFIG_PROC_FS
1259	INIT_LIST_HEAD(&bo->client_link);
1260#endif
1261	INIT_LIST_HEAD(&bo->vram_userfault_link);
1262
1263	drm_gem_private_object_init(&xe->drm, &bo->ttm.base, size);
1264
1265	if (resv) {
1266		ctx.allow_res_evict = !(flags & XE_BO_CREATE_NO_RESV_EVICT);
1267		ctx.resv = resv;
1268	}
1269
1270	if (!(flags & XE_BO_FIXED_PLACEMENT_BIT)) {
1271		err = __xe_bo_placement_for_flags(xe, bo, bo->flags);
1272		if (WARN_ON(err)) {
1273			xe_ttm_bo_destroy(&bo->ttm);
1274			return ERR_PTR(err);
1275		}
1276	}
1277
1278	/* Defer populating type_sg bos */
1279	placement = (type == ttm_bo_type_sg ||
1280		     bo->flags & XE_BO_DEFER_BACKING) ? &sys_placement :
1281		&bo->placement;
1282	err = ttm_bo_init_reserved(&xe->ttm, &bo->ttm, type,
1283				   placement, alignment,
1284				   &ctx, NULL, resv, xe_ttm_bo_destroy);
1285	if (err)
1286		return ERR_PTR(err);
1287
1288	/*
1289	 * The VRAM pages underneath are potentially still being accessed by the
1290	 * GPU, as per async GPU clearing and async evictions. However TTM makes
1291	 * sure to add any corresponding move/clear fences into the objects
1292	 * dma-resv using the DMA_RESV_USAGE_KERNEL slot.
1293	 *
1294	 * For KMD internal buffers we don't care about GPU clearing, however we
1295	 * still need to handle async evictions, where the VRAM is still being
1296	 * accessed by the GPU. Most internal callers are not expecting this,
1297	 * since they are missing the required synchronisation before accessing
1298	 * the memory. To keep things simple just sync wait any kernel fences
1299	 * here, if the buffer is designated KMD internal.
1300	 *
1301	 * For normal userspace objects we should already have the required
1302	 * pipelining or sync waiting elsewhere, since we already have to deal
1303	 * with things like async GPU clearing.
1304	 */
1305	if (type == ttm_bo_type_kernel) {
1306		long timeout = dma_resv_wait_timeout(bo->ttm.base.resv,
1307						     DMA_RESV_USAGE_KERNEL,
1308						     ctx.interruptible,
1309						     MAX_SCHEDULE_TIMEOUT);
1310
1311		if (timeout < 0) {
1312			if (!resv)
1313				dma_resv_unlock(bo->ttm.base.resv);
1314			xe_bo_put(bo);
1315			return ERR_PTR(timeout);
1316		}
1317	}
1318
1319	bo->created = true;
1320	if (bulk)
1321		ttm_bo_set_bulk_move(&bo->ttm, bulk);
1322	else
1323		ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
1324
1325	return bo;
1326}
1327
1328static int __xe_bo_fixed_placement(struct xe_device *xe,
1329				   struct xe_bo *bo,
1330				   u32 flags,
1331				   u64 start, u64 end, u64 size)
1332{
1333	struct ttm_place *place = bo->placements;
1334
1335	if (flags & (XE_BO_CREATE_USER_BIT|XE_BO_CREATE_SYSTEM_BIT))
1336		return -EINVAL;
1337
1338	place->flags = TTM_PL_FLAG_CONTIGUOUS;
1339	place->fpfn = start >> PAGE_SHIFT;
1340	place->lpfn = end >> PAGE_SHIFT;
1341
1342	switch (flags & (XE_BO_CREATE_STOLEN_BIT | XE_BO_CREATE_VRAM_MASK)) {
1343	case XE_BO_CREATE_VRAM0_BIT:
1344		place->mem_type = XE_PL_VRAM0;
1345		break;
1346	case XE_BO_CREATE_VRAM1_BIT:
1347		place->mem_type = XE_PL_VRAM1;
1348		break;
1349	case XE_BO_CREATE_STOLEN_BIT:
1350		place->mem_type = XE_PL_STOLEN;
1351		break;
1352
1353	default:
1354		/* 0 or multiple of the above set */
1355		return -EINVAL;
1356	}
1357
1358	bo->placement = (struct ttm_placement) {
1359		.num_placement = 1,
1360		.placement = place,
1361	};
1362
1363	return 0;
1364}
1365
1366static struct xe_bo *
1367__xe_bo_create_locked(struct xe_device *xe,
1368		      struct xe_tile *tile, struct xe_vm *vm,
1369		      size_t size, u64 start, u64 end,
1370		      u16 cpu_caching, enum ttm_bo_type type, u32 flags)
1371{
1372	struct xe_bo *bo = NULL;
1373	int err;
1374
1375	if (vm)
1376		xe_vm_assert_held(vm);
1377
1378	if (start || end != ~0ULL) {
1379		bo = xe_bo_alloc();
1380		if (IS_ERR(bo))
1381			return bo;
1382
1383		flags |= XE_BO_FIXED_PLACEMENT_BIT;
1384		err = __xe_bo_fixed_placement(xe, bo, flags, start, end, size);
1385		if (err) {
1386			xe_bo_free(bo);
1387			return ERR_PTR(err);
1388		}
1389	}
1390
1391	bo = ___xe_bo_create_locked(xe, bo, tile, vm ? xe_vm_resv(vm) : NULL,
1392				    vm && !xe_vm_in_fault_mode(vm) &&
1393				    flags & XE_BO_CREATE_USER_BIT ?
1394				    &vm->lru_bulk_move : NULL, size,
1395				    cpu_caching, type, flags);
1396	if (IS_ERR(bo))
1397		return bo;
1398
1399	/*
1400	 * Note that instead of taking a reference no the drm_gpuvm_resv_bo(),
1401	 * to ensure the shared resv doesn't disappear under the bo, the bo
1402	 * will keep a reference to the vm, and avoid circular references
1403	 * by having all the vm's bo refereferences released at vm close
1404	 * time.
1405	 */
1406	if (vm && xe_bo_is_user(bo))
1407		xe_vm_get(vm);
1408	bo->vm = vm;
1409
1410	if (bo->flags & XE_BO_CREATE_GGTT_BIT) {
1411		if (!tile && flags & XE_BO_CREATE_STOLEN_BIT)
1412			tile = xe_device_get_root_tile(xe);
1413
1414		xe_assert(xe, tile);
1415
1416		if (flags & XE_BO_FIXED_PLACEMENT_BIT) {
1417			err = xe_ggtt_insert_bo_at(tile->mem.ggtt, bo,
1418						   start + bo->size, U64_MAX);
1419		} else {
1420			err = xe_ggtt_insert_bo(tile->mem.ggtt, bo);
1421		}
1422		if (err)
1423			goto err_unlock_put_bo;
1424	}
1425
1426	return bo;
1427
1428err_unlock_put_bo:
1429	__xe_bo_unset_bulk_move(bo);
1430	xe_bo_unlock_vm_held(bo);
1431	xe_bo_put(bo);
1432	return ERR_PTR(err);
1433}
1434
1435struct xe_bo *
1436xe_bo_create_locked_range(struct xe_device *xe,
1437			  struct xe_tile *tile, struct xe_vm *vm,
1438			  size_t size, u64 start, u64 end,
1439			  enum ttm_bo_type type, u32 flags)
1440{
1441	return __xe_bo_create_locked(xe, tile, vm, size, start, end, 0, type, flags);
1442}
1443
1444struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile,
1445				  struct xe_vm *vm, size_t size,
1446				  enum ttm_bo_type type, u32 flags)
1447{
1448	return __xe_bo_create_locked(xe, tile, vm, size, 0, ~0ULL, 0, type, flags);
1449}
1450
1451struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_tile *tile,
1452				struct xe_vm *vm, size_t size,
1453				u16 cpu_caching,
1454				enum ttm_bo_type type,
1455				u32 flags)
1456{
1457	struct xe_bo *bo = __xe_bo_create_locked(xe, tile, vm, size, 0, ~0ULL,
1458						 cpu_caching, type,
1459						 flags | XE_BO_CREATE_USER_BIT);
1460	if (!IS_ERR(bo))
1461		xe_bo_unlock_vm_held(bo);
1462
1463	return bo;
1464}
1465
1466struct xe_bo *xe_bo_create(struct xe_device *xe, struct xe_tile *tile,
1467			   struct xe_vm *vm, size_t size,
1468			   enum ttm_bo_type type, u32 flags)
1469{
1470	struct xe_bo *bo = xe_bo_create_locked(xe, tile, vm, size, type, flags);
1471
1472	if (!IS_ERR(bo))
1473		xe_bo_unlock_vm_held(bo);
1474
1475	return bo;
1476}
1477
1478struct xe_bo *xe_bo_create_pin_map_at(struct xe_device *xe, struct xe_tile *tile,
1479				      struct xe_vm *vm,
1480				      size_t size, u64 offset,
1481				      enum ttm_bo_type type, u32 flags)
1482{
1483	struct xe_bo *bo;
1484	int err;
1485	u64 start = offset == ~0ull ? 0 : offset;
1486	u64 end = offset == ~0ull ? offset : start + size;
1487
1488	if (flags & XE_BO_CREATE_STOLEN_BIT &&
1489	    xe_ttm_stolen_cpu_access_needs_ggtt(xe))
1490		flags |= XE_BO_CREATE_GGTT_BIT;
1491
1492	bo = xe_bo_create_locked_range(xe, tile, vm, size, start, end, type,
1493				       flags | XE_BO_NEEDS_CPU_ACCESS);
1494	if (IS_ERR(bo))
1495		return bo;
1496
1497	err = xe_bo_pin(bo);
1498	if (err)
1499		goto err_put;
1500
1501	err = xe_bo_vmap(bo);
1502	if (err)
1503		goto err_unpin;
1504
1505	xe_bo_unlock_vm_held(bo);
1506
1507	return bo;
1508
1509err_unpin:
1510	xe_bo_unpin(bo);
1511err_put:
1512	xe_bo_unlock_vm_held(bo);
1513	xe_bo_put(bo);
1514	return ERR_PTR(err);
1515}
1516
1517struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
1518				   struct xe_vm *vm, size_t size,
1519				   enum ttm_bo_type type, u32 flags)
1520{
1521	return xe_bo_create_pin_map_at(xe, tile, vm, size, ~0ull, type, flags);
1522}
1523
1524struct xe_bo *xe_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
1525				     const void *data, size_t size,
1526				     enum ttm_bo_type type, u32 flags)
1527{
1528	struct xe_bo *bo = xe_bo_create_pin_map(xe, tile, NULL,
1529						ALIGN(size, PAGE_SIZE),
1530						type, flags);
1531	if (IS_ERR(bo))
1532		return bo;
1533
1534	xe_map_memcpy_to(xe, &bo->vmap, 0, data, size);
1535
1536	return bo;
1537}
1538
1539static void __xe_bo_unpin_map_no_vm(struct drm_device *drm, void *arg)
1540{
1541	xe_bo_unpin_map_no_vm(arg);
1542}
1543
1544struct xe_bo *xe_managed_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile,
1545					   size_t size, u32 flags)
1546{
1547	struct xe_bo *bo;
1548	int ret;
1549
1550	bo = xe_bo_create_pin_map(xe, tile, NULL, size, ttm_bo_type_kernel, flags);
1551	if (IS_ERR(bo))
1552		return bo;
1553
1554	ret = drmm_add_action_or_reset(&xe->drm, __xe_bo_unpin_map_no_vm, bo);
1555	if (ret)
1556		return ERR_PTR(ret);
1557
1558	return bo;
1559}
1560
1561struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile,
1562					     const void *data, size_t size, u32 flags)
1563{
1564	struct xe_bo *bo = xe_managed_bo_create_pin_map(xe, tile, ALIGN(size, PAGE_SIZE), flags);
1565
1566	if (IS_ERR(bo))
1567		return bo;
1568
1569	xe_map_memcpy_to(xe, &bo->vmap, 0, data, size);
1570
1571	return bo;
1572}
1573
1574/**
1575 * xe_managed_bo_reinit_in_vram
1576 * @xe: xe device
1577 * @tile: Tile where the new buffer will be created
1578 * @src: Managed buffer object allocated in system memory
1579 *
1580 * Replace a managed src buffer object allocated in system memory with a new
1581 * one allocated in vram, copying the data between them.
1582 * Buffer object in VRAM is not going to have the same GGTT address, the caller
1583 * is responsible for making sure that any old references to it are updated.
1584 *
1585 * Returns 0 for success, negative error code otherwise.
1586 */
1587int xe_managed_bo_reinit_in_vram(struct xe_device *xe, struct xe_tile *tile, struct xe_bo **src)
1588{
1589	struct xe_bo *bo;
1590
1591	xe_assert(xe, IS_DGFX(xe));
1592	xe_assert(xe, !(*src)->vmap.is_iomem);
1593
1594	bo = xe_managed_bo_create_from_data(xe, tile, (*src)->vmap.vaddr, (*src)->size,
1595					    XE_BO_CREATE_VRAM_IF_DGFX(tile) |
1596					    XE_BO_CREATE_GGTT_BIT);
1597	if (IS_ERR(bo))
1598		return PTR_ERR(bo);
1599
1600	drmm_release_action(&xe->drm, __xe_bo_unpin_map_no_vm, *src);
1601	*src = bo;
1602
1603	return 0;
1604}
1605
1606/*
1607 * XXX: This is in the VM bind data path, likely should calculate this once and
1608 * store, with a recalculation if the BO is moved.
1609 */
1610uint64_t vram_region_gpu_offset(struct ttm_resource *res)
1611{
1612	struct xe_device *xe = ttm_to_xe_device(res->bo->bdev);
1613
1614	if (res->mem_type == XE_PL_STOLEN)
1615		return xe_ttm_stolen_gpu_offset(xe);
1616
1617	return res_to_mem_region(res)->dpa_base;
1618}
1619
1620/**
1621 * xe_bo_pin_external - pin an external BO
1622 * @bo: buffer object to be pinned
1623 *
1624 * Pin an external (not tied to a VM, can be exported via dma-buf / prime FD)
1625 * BO. Unique call compared to xe_bo_pin as this function has it own set of
1626 * asserts and code to ensure evict / restore on suspend / resume.
1627 *
1628 * Returns 0 for success, negative error code otherwise.
1629 */
1630int xe_bo_pin_external(struct xe_bo *bo)
1631{
1632	struct xe_device *xe = xe_bo_device(bo);
1633	int err;
1634
1635	xe_assert(xe, !bo->vm);
1636	xe_assert(xe, xe_bo_is_user(bo));
1637
1638	if (!xe_bo_is_pinned(bo)) {
1639		err = xe_bo_validate(bo, NULL, false);
1640		if (err)
1641			return err;
1642
1643		if (xe_bo_is_vram(bo)) {
1644			spin_lock(&xe->pinned.lock);
1645			list_add_tail(&bo->pinned_link,
1646				      &xe->pinned.external_vram);
1647			spin_unlock(&xe->pinned.lock);
1648		}
1649	}
1650
1651	ttm_bo_pin(&bo->ttm);
1652
1653	/*
1654	 * FIXME: If we always use the reserve / unreserve functions for locking
1655	 * we do not need this.
1656	 */
1657	ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
1658
1659	return 0;
1660}
1661
1662int xe_bo_pin(struct xe_bo *bo)
1663{
1664	struct xe_device *xe = xe_bo_device(bo);
1665	int err;
1666
1667	/* We currently don't expect user BO to be pinned */
1668	xe_assert(xe, !xe_bo_is_user(bo));
1669
1670	/* Pinned object must be in GGTT or have pinned flag */
1671	xe_assert(xe, bo->flags & (XE_BO_CREATE_PINNED_BIT |
1672				   XE_BO_CREATE_GGTT_BIT));
1673
1674	/*
1675	 * No reason we can't support pinning imported dma-bufs we just don't
1676	 * expect to pin an imported dma-buf.
1677	 */
1678	xe_assert(xe, !bo->ttm.base.import_attach);
1679
1680	/* We only expect at most 1 pin */
1681	xe_assert(xe, !xe_bo_is_pinned(bo));
1682
1683	err = xe_bo_validate(bo, NULL, false);
1684	if (err)
1685		return err;
1686
1687	/*
1688	 * For pinned objects in on DGFX, which are also in vram, we expect
1689	 * these to be in contiguous VRAM memory. Required eviction / restore
1690	 * during suspend / resume (force restore to same physical address).
1691	 */
1692	if (IS_DGFX(xe) && !(IS_ENABLED(CONFIG_DRM_XE_DEBUG) &&
1693	    bo->flags & XE_BO_INTERNAL_TEST)) {
1694		struct ttm_place *place = &(bo->placements[0]);
1695
1696		if (mem_type_is_vram(place->mem_type)) {
1697			xe_assert(xe, place->flags & TTM_PL_FLAG_CONTIGUOUS);
1698
1699			place->fpfn = (xe_bo_addr(bo, 0, PAGE_SIZE) -
1700				       vram_region_gpu_offset(bo->ttm.resource)) >> PAGE_SHIFT;
1701			place->lpfn = place->fpfn + (bo->size >> PAGE_SHIFT);
1702
1703			spin_lock(&xe->pinned.lock);
1704			list_add_tail(&bo->pinned_link, &xe->pinned.kernel_bo_present);
1705			spin_unlock(&xe->pinned.lock);
1706		}
1707	}
1708
1709	ttm_bo_pin(&bo->ttm);
1710
1711	/*
1712	 * FIXME: If we always use the reserve / unreserve functions for locking
1713	 * we do not need this.
1714	 */
1715	ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
1716
1717	return 0;
1718}
1719
1720/**
1721 * xe_bo_unpin_external - unpin an external BO
1722 * @bo: buffer object to be unpinned
1723 *
1724 * Unpin an external (not tied to a VM, can be exported via dma-buf / prime FD)
1725 * BO. Unique call compared to xe_bo_unpin as this function has it own set of
1726 * asserts and code to ensure evict / restore on suspend / resume.
1727 *
1728 * Returns 0 for success, negative error code otherwise.
1729 */
1730void xe_bo_unpin_external(struct xe_bo *bo)
1731{
1732	struct xe_device *xe = xe_bo_device(bo);
1733
1734	xe_assert(xe, !bo->vm);
1735	xe_assert(xe, xe_bo_is_pinned(bo));
1736	xe_assert(xe, xe_bo_is_user(bo));
1737
1738	if (bo->ttm.pin_count == 1 && !list_empty(&bo->pinned_link)) {
1739		spin_lock(&xe->pinned.lock);
1740		list_del_init(&bo->pinned_link);
1741		spin_unlock(&xe->pinned.lock);
1742	}
1743
1744	ttm_bo_unpin(&bo->ttm);
1745
1746	/*
1747	 * FIXME: If we always use the reserve / unreserve functions for locking
1748	 * we do not need this.
1749	 */
1750	ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
1751}
1752
1753void xe_bo_unpin(struct xe_bo *bo)
1754{
1755	struct xe_device *xe = xe_bo_device(bo);
1756
1757	xe_assert(xe, !bo->ttm.base.import_attach);
1758	xe_assert(xe, xe_bo_is_pinned(bo));
1759
1760	if (IS_DGFX(xe) && !(IS_ENABLED(CONFIG_DRM_XE_DEBUG) &&
1761	    bo->flags & XE_BO_INTERNAL_TEST)) {
1762		struct ttm_place *place = &(bo->placements[0]);
1763
1764		if (mem_type_is_vram(place->mem_type)) {
1765			xe_assert(xe, !list_empty(&bo->pinned_link));
1766
1767			spin_lock(&xe->pinned.lock);
1768			list_del_init(&bo->pinned_link);
1769			spin_unlock(&xe->pinned.lock);
1770		}
1771	}
1772
1773	ttm_bo_unpin(&bo->ttm);
1774}
1775
1776/**
1777 * xe_bo_validate() - Make sure the bo is in an allowed placement
1778 * @bo: The bo,
1779 * @vm: Pointer to a the vm the bo shares a locked dma_resv object with, or
1780 *      NULL. Used together with @allow_res_evict.
1781 * @allow_res_evict: Whether it's allowed to evict bos sharing @vm's
1782 *                   reservation object.
1783 *
1784 * Make sure the bo is in allowed placement, migrating it if necessary. If
1785 * needed, other bos will be evicted. If bos selected for eviction shares
1786 * the @vm's reservation object, they can be evicted iff @allow_res_evict is
1787 * set to true, otherwise they will be bypassed.
1788 *
1789 * Return: 0 on success, negative error code on failure. May return
1790 * -EINTR or -ERESTARTSYS if internal waits are interrupted by a signal.
1791 */
1792int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict)
1793{
1794	struct ttm_operation_ctx ctx = {
1795		.interruptible = true,
1796		.no_wait_gpu = false,
1797	};
1798
1799	if (vm) {
1800		lockdep_assert_held(&vm->lock);
1801		xe_vm_assert_held(vm);
1802
1803		ctx.allow_res_evict = allow_res_evict;
1804		ctx.resv = xe_vm_resv(vm);
1805	}
1806
1807	return ttm_bo_validate(&bo->ttm, &bo->placement, &ctx);
1808}
1809
1810bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo)
1811{
1812	if (bo->destroy == &xe_ttm_bo_destroy)
1813		return true;
1814
1815	return false;
1816}
1817
1818/*
1819 * Resolve a BO address. There is no assert to check if the proper lock is held
1820 * so it should only be used in cases where it is not fatal to get the wrong
1821 * address, such as printing debug information, but not in cases where memory is
1822 * written based on this result.
1823 */
1824dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size)
1825{
1826	struct xe_device *xe = xe_bo_device(bo);
1827	struct xe_res_cursor cur;
1828	u64 page;
1829
1830	xe_assert(xe, page_size <= PAGE_SIZE);
1831	page = offset >> PAGE_SHIFT;
1832	offset &= (PAGE_SIZE - 1);
1833
1834	if (!xe_bo_is_vram(bo) && !xe_bo_is_stolen(bo)) {
1835		xe_assert(xe, bo->ttm.ttm);
1836
1837		xe_res_first_sg(xe_bo_sg(bo), page << PAGE_SHIFT,
1838				page_size, &cur);
1839		return xe_res_dma(&cur) + offset;
1840	} else {
1841		struct xe_res_cursor cur;
1842
1843		xe_res_first(bo->ttm.resource, page << PAGE_SHIFT,
1844			     page_size, &cur);
1845		return cur.start + offset + vram_region_gpu_offset(bo->ttm.resource);
1846	}
1847}
1848
1849dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size)
1850{
1851	if (!READ_ONCE(bo->ttm.pin_count))
1852		xe_bo_assert_held(bo);
1853	return __xe_bo_addr(bo, offset, page_size);
1854}
1855
1856int xe_bo_vmap(struct xe_bo *bo)
1857{
1858	void *virtual;
1859	bool is_iomem;
1860	int ret;
1861
1862	xe_bo_assert_held(bo);
1863
1864	if (!(bo->flags & XE_BO_NEEDS_CPU_ACCESS))
1865		return -EINVAL;
1866
1867	if (!iosys_map_is_null(&bo->vmap))
1868		return 0;
1869
1870	/*
1871	 * We use this more or less deprecated interface for now since
1872	 * ttm_bo_vmap() doesn't offer the optimization of kmapping
1873	 * single page bos, which is done here.
1874	 * TODO: Fix up ttm_bo_vmap to do that, or fix up ttm_bo_kmap
1875	 * to use struct iosys_map.
1876	 */
1877	ret = ttm_bo_kmap(&bo->ttm, 0, bo->size >> PAGE_SHIFT, &bo->kmap);
1878	if (ret)
1879		return ret;
1880
1881	virtual = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
1882	if (is_iomem)
1883		iosys_map_set_vaddr_iomem(&bo->vmap, (void __iomem *)virtual);
1884	else
1885		iosys_map_set_vaddr(&bo->vmap, virtual);
1886
1887	return 0;
1888}
1889
1890static void __xe_bo_vunmap(struct xe_bo *bo)
1891{
1892	if (!iosys_map_is_null(&bo->vmap)) {
1893		iosys_map_clear(&bo->vmap);
1894		ttm_bo_kunmap(&bo->kmap);
1895	}
1896}
1897
1898void xe_bo_vunmap(struct xe_bo *bo)
1899{
1900	xe_bo_assert_held(bo);
1901	__xe_bo_vunmap(bo);
1902}
1903
1904int xe_gem_create_ioctl(struct drm_device *dev, void *data,
1905			struct drm_file *file)
1906{
1907	struct xe_device *xe = to_xe_device(dev);
1908	struct xe_file *xef = to_xe_file(file);
1909	struct drm_xe_gem_create *args = data;
1910	struct xe_vm *vm = NULL;
1911	struct xe_bo *bo;
1912	unsigned int bo_flags;
1913	u32 handle;
1914	int err;
1915
1916	if (XE_IOCTL_DBG(xe, args->extensions) ||
1917	    XE_IOCTL_DBG(xe, args->pad[0] || args->pad[1] || args->pad[2]) ||
1918	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
1919		return -EINVAL;
1920
1921	/* at least one valid memory placement must be specified */
1922	if (XE_IOCTL_DBG(xe, (args->placement & ~xe->info.mem_region_mask) ||
1923			 !args->placement))
1924		return -EINVAL;
1925
1926	if (XE_IOCTL_DBG(xe, args->flags &
1927			 ~(DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING |
1928			   DRM_XE_GEM_CREATE_FLAG_SCANOUT |
1929			   DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM)))
1930		return -EINVAL;
1931
1932	if (XE_IOCTL_DBG(xe, args->handle))
1933		return -EINVAL;
1934
1935	if (XE_IOCTL_DBG(xe, !args->size))
1936		return -EINVAL;
1937
1938	if (XE_IOCTL_DBG(xe, args->size > SIZE_MAX))
1939		return -EINVAL;
1940
1941	if (XE_IOCTL_DBG(xe, args->size & ~PAGE_MASK))
1942		return -EINVAL;
1943
1944	bo_flags = 0;
1945	if (args->flags & DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING)
1946		bo_flags |= XE_BO_DEFER_BACKING;
1947
1948	if (args->flags & DRM_XE_GEM_CREATE_FLAG_SCANOUT)
1949		bo_flags |= XE_BO_SCANOUT_BIT;
1950
1951	bo_flags |= args->placement << (ffs(XE_BO_CREATE_SYSTEM_BIT) - 1);
1952
1953	if (args->flags & DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM) {
1954		if (XE_IOCTL_DBG(xe, !(bo_flags & XE_BO_CREATE_VRAM_MASK)))
1955			return -EINVAL;
1956
1957		bo_flags |= XE_BO_NEEDS_CPU_ACCESS;
1958	}
1959
1960	if (XE_IOCTL_DBG(xe, !args->cpu_caching ||
1961			 args->cpu_caching > DRM_XE_GEM_CPU_CACHING_WC))
1962		return -EINVAL;
1963
1964	if (XE_IOCTL_DBG(xe, bo_flags & XE_BO_CREATE_VRAM_MASK &&
1965			 args->cpu_caching != DRM_XE_GEM_CPU_CACHING_WC))
1966		return -EINVAL;
1967
1968	if (XE_IOCTL_DBG(xe, bo_flags & XE_BO_SCANOUT_BIT &&
1969			 args->cpu_caching == DRM_XE_GEM_CPU_CACHING_WB))
1970		return -EINVAL;
1971
1972	if (args->vm_id) {
1973		vm = xe_vm_lookup(xef, args->vm_id);
1974		if (XE_IOCTL_DBG(xe, !vm))
1975			return -ENOENT;
1976		err = xe_vm_lock(vm, true);
1977		if (err)
1978			goto out_vm;
1979	}
1980
1981	bo = xe_bo_create_user(xe, NULL, vm, args->size, args->cpu_caching,
1982			       ttm_bo_type_device, bo_flags);
1983
1984	if (vm)
1985		xe_vm_unlock(vm);
1986
1987	if (IS_ERR(bo)) {
1988		err = PTR_ERR(bo);
1989		goto out_vm;
1990	}
1991
1992	err = drm_gem_handle_create(file, &bo->ttm.base, &handle);
1993	if (err)
1994		goto out_bulk;
1995
1996	args->handle = handle;
1997	goto out_put;
1998
1999out_bulk:
2000	if (vm && !xe_vm_in_fault_mode(vm)) {
2001		xe_vm_lock(vm, false);
2002		__xe_bo_unset_bulk_move(bo);
2003		xe_vm_unlock(vm);
2004	}
2005out_put:
2006	xe_bo_put(bo);
2007out_vm:
2008	if (vm)
2009		xe_vm_put(vm);
2010
2011	return err;
2012}
2013
2014int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data,
2015			     struct drm_file *file)
2016{
2017	struct xe_device *xe = to_xe_device(dev);
2018	struct drm_xe_gem_mmap_offset *args = data;
2019	struct drm_gem_object *gem_obj;
2020
2021	if (XE_IOCTL_DBG(xe, args->extensions) ||
2022	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
2023		return -EINVAL;
2024
2025	if (XE_IOCTL_DBG(xe, args->flags))
2026		return -EINVAL;
2027
2028	gem_obj = drm_gem_object_lookup(file, args->handle);
2029	if (XE_IOCTL_DBG(xe, !gem_obj))
2030		return -ENOENT;
2031
2032	/* The mmap offset was set up at BO allocation time. */
2033	args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
2034
2035	xe_bo_put(gem_to_xe_bo(gem_obj));
2036	return 0;
2037}
2038
2039/**
2040 * xe_bo_lock() - Lock the buffer object's dma_resv object
2041 * @bo: The struct xe_bo whose lock is to be taken
2042 * @intr: Whether to perform any wait interruptible
2043 *
2044 * Locks the buffer object's dma_resv object. If the buffer object is
2045 * pointing to a shared dma_resv object, that shared lock is locked.
2046 *
2047 * Return: 0 on success, -EINTR if @intr is true and the wait for a
2048 * contended lock was interrupted. If @intr is set to false, the
2049 * function always returns 0.
2050 */
2051int xe_bo_lock(struct xe_bo *bo, bool intr)
2052{
2053	if (intr)
2054		return dma_resv_lock_interruptible(bo->ttm.base.resv, NULL);
2055
2056	dma_resv_lock(bo->ttm.base.resv, NULL);
2057
2058	return 0;
2059}
2060
2061/**
2062 * xe_bo_unlock() - Unlock the buffer object's dma_resv object
2063 * @bo: The struct xe_bo whose lock is to be released.
2064 *
2065 * Unlock a buffer object lock that was locked by xe_bo_lock().
2066 */
2067void xe_bo_unlock(struct xe_bo *bo)
2068{
2069	dma_resv_unlock(bo->ttm.base.resv);
2070}
2071
2072/**
2073 * xe_bo_can_migrate - Whether a buffer object likely can be migrated
2074 * @bo: The buffer object to migrate
2075 * @mem_type: The TTM memory type intended to migrate to
2076 *
2077 * Check whether the buffer object supports migration to the
2078 * given memory type. Note that pinning may affect the ability to migrate as
2079 * returned by this function.
2080 *
2081 * This function is primarily intended as a helper for checking the
2082 * possibility to migrate buffer objects and can be called without
2083 * the object lock held.
2084 *
2085 * Return: true if migration is possible, false otherwise.
2086 */
2087bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type)
2088{
2089	unsigned int cur_place;
2090
2091	if (bo->ttm.type == ttm_bo_type_kernel)
2092		return true;
2093
2094	if (bo->ttm.type == ttm_bo_type_sg)
2095		return false;
2096
2097	for (cur_place = 0; cur_place < bo->placement.num_placement;
2098	     cur_place++) {
2099		if (bo->placements[cur_place].mem_type == mem_type)
2100			return true;
2101	}
2102
2103	return false;
2104}
2105
2106static void xe_place_from_ttm_type(u32 mem_type, struct ttm_place *place)
2107{
2108	memset(place, 0, sizeof(*place));
2109	place->mem_type = mem_type;
2110}
2111
2112/**
2113 * xe_bo_migrate - Migrate an object to the desired region id
2114 * @bo: The buffer object to migrate.
2115 * @mem_type: The TTM region type to migrate to.
2116 *
2117 * Attempt to migrate the buffer object to the desired memory region. The
2118 * buffer object may not be pinned, and must be locked.
2119 * On successful completion, the object memory type will be updated,
2120 * but an async migration task may not have completed yet, and to
2121 * accomplish that, the object's kernel fences must be signaled with
2122 * the object lock held.
2123 *
2124 * Return: 0 on success. Negative error code on failure. In particular may
2125 * return -EINTR or -ERESTARTSYS if signal pending.
2126 */
2127int xe_bo_migrate(struct xe_bo *bo, u32 mem_type)
2128{
2129	struct xe_device *xe = ttm_to_xe_device(bo->ttm.bdev);
2130	struct ttm_operation_ctx ctx = {
2131		.interruptible = true,
2132		.no_wait_gpu = false,
2133	};
2134	struct ttm_placement placement;
2135	struct ttm_place requested;
2136
2137	xe_bo_assert_held(bo);
2138
2139	if (bo->ttm.resource->mem_type == mem_type)
2140		return 0;
2141
2142	if (xe_bo_is_pinned(bo))
2143		return -EBUSY;
2144
2145	if (!xe_bo_can_migrate(bo, mem_type))
2146		return -EINVAL;
2147
2148	xe_place_from_ttm_type(mem_type, &requested);
2149	placement.num_placement = 1;
2150	placement.placement = &requested;
2151
2152	/*
2153	 * Stolen needs to be handled like below VRAM handling if we ever need
2154	 * to support it.
2155	 */
2156	drm_WARN_ON(&xe->drm, mem_type == XE_PL_STOLEN);
2157
2158	if (mem_type_is_vram(mem_type)) {
2159		u32 c = 0;
2160
2161		add_vram(xe, bo, &requested, bo->flags, mem_type, &c);
2162	}
2163
2164	return ttm_bo_validate(&bo->ttm, &placement, &ctx);
2165}
2166
2167/**
2168 * xe_bo_evict - Evict an object to evict placement
2169 * @bo: The buffer object to migrate.
2170 * @force_alloc: Set force_alloc in ttm_operation_ctx
2171 *
2172 * On successful completion, the object memory will be moved to evict
2173 * placement. Ths function blocks until the object has been fully moved.
2174 *
2175 * Return: 0 on success. Negative error code on failure.
2176 */
2177int xe_bo_evict(struct xe_bo *bo, bool force_alloc)
2178{
2179	struct ttm_operation_ctx ctx = {
2180		.interruptible = false,
2181		.no_wait_gpu = false,
2182		.force_alloc = force_alloc,
2183	};
2184	struct ttm_placement placement;
2185	int ret;
2186
2187	xe_evict_flags(&bo->ttm, &placement);
2188	ret = ttm_bo_validate(&bo->ttm, &placement, &ctx);
2189	if (ret)
2190		return ret;
2191
2192	dma_resv_wait_timeout(bo->ttm.base.resv, DMA_RESV_USAGE_KERNEL,
2193			      false, MAX_SCHEDULE_TIMEOUT);
2194
2195	return 0;
2196}
2197
2198/**
2199 * xe_bo_needs_ccs_pages - Whether a bo needs to back up CCS pages when
2200 * placed in system memory.
2201 * @bo: The xe_bo
2202 *
2203 * Return: true if extra pages need to be allocated, false otherwise.
2204 */
2205bool xe_bo_needs_ccs_pages(struct xe_bo *bo)
2206{
2207	struct xe_device *xe = xe_bo_device(bo);
2208
2209	if (!xe_device_has_flat_ccs(xe) || bo->ttm.type != ttm_bo_type_device)
2210		return false;
2211
2212	/* On discrete GPUs, if the GPU can access this buffer from
2213	 * system memory (i.e., it allows XE_PL_TT placement), FlatCCS
2214	 * can't be used since there's no CCS storage associated with
2215	 * non-VRAM addresses.
2216	 */
2217	if (IS_DGFX(xe) && (bo->flags & XE_BO_CREATE_SYSTEM_BIT))
2218		return false;
2219
2220	return true;
2221}
2222
2223/**
2224 * __xe_bo_release_dummy() - Dummy kref release function
2225 * @kref: The embedded struct kref.
2226 *
2227 * Dummy release function for xe_bo_put_deferred(). Keep off.
2228 */
2229void __xe_bo_release_dummy(struct kref *kref)
2230{
2231}
2232
2233/**
2234 * xe_bo_put_commit() - Put bos whose put was deferred by xe_bo_put_deferred().
2235 * @deferred: The lockless list used for the call to xe_bo_put_deferred().
2236 *
2237 * Puts all bos whose put was deferred by xe_bo_put_deferred().
2238 * The @deferred list can be either an onstack local list or a global
2239 * shared list used by a workqueue.
2240 */
2241void xe_bo_put_commit(struct llist_head *deferred)
2242{
2243	struct llist_node *freed;
2244	struct xe_bo *bo, *next;
2245
2246	if (!deferred)
2247		return;
2248
2249	freed = llist_del_all(deferred);
2250	if (!freed)
2251		return;
2252
2253	llist_for_each_entry_safe(bo, next, freed, freed)
2254		drm_gem_object_free(&bo->ttm.base.refcount);
2255}
2256
2257/**
2258 * xe_bo_dumb_create - Create a dumb bo as backing for a fb
2259 * @file_priv: ...
2260 * @dev: ...
2261 * @args: ...
2262 *
2263 * See dumb_create() hook in include/drm/drm_drv.h
2264 *
2265 * Return: ...
2266 */
2267int xe_bo_dumb_create(struct drm_file *file_priv,
2268		      struct drm_device *dev,
2269		      struct drm_mode_create_dumb *args)
2270{
2271	struct xe_device *xe = to_xe_device(dev);
2272	struct xe_bo *bo;
2273	uint32_t handle;
2274	int cpp = DIV_ROUND_UP(args->bpp, 8);
2275	int err;
2276	u32 page_size = max_t(u32, PAGE_SIZE,
2277		xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K ? SZ_64K : SZ_4K);
2278
2279	args->pitch = ALIGN(args->width * cpp, 64);
2280	args->size = ALIGN(mul_u32_u32(args->pitch, args->height),
2281			   page_size);
2282
2283	bo = xe_bo_create_user(xe, NULL, NULL, args->size,
2284			       DRM_XE_GEM_CPU_CACHING_WC,
2285			       ttm_bo_type_device,
2286			       XE_BO_CREATE_VRAM_IF_DGFX(xe_device_get_root_tile(xe)) |
2287			       XE_BO_CREATE_USER_BIT | XE_BO_SCANOUT_BIT |
2288			       XE_BO_NEEDS_CPU_ACCESS);
2289	if (IS_ERR(bo))
2290		return PTR_ERR(bo);
2291
2292	err = drm_gem_handle_create(file_priv, &bo->ttm.base, &handle);
2293	/* drop reference from allocate - handle holds it now */
2294	drm_gem_object_put(&bo->ttm.base);
2295	if (!err)
2296		args->handle = handle;
2297	return err;
2298}
2299
2300void xe_bo_runtime_pm_release_mmap_offset(struct xe_bo *bo)
2301{
2302	struct ttm_buffer_object *tbo = &bo->ttm;
2303	struct ttm_device *bdev = tbo->bdev;
2304
2305	drm_vma_node_unmap(&tbo->base.vma_node, bdev->dev_mapping);
2306
2307	list_del_init(&bo->vram_userfault_link);
2308}
2309
2310#if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
2311#include "tests/xe_bo.c"
2312#endif