Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Mar 24-27, 2025, special US time zones
Register
Loading...
Note: File does not exist in v3.1.
   1/*
   2 * Copyright 2014-2018 Advanced Micro Devices, Inc.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice shall be included in
  12 * all copies or substantial portions of the Software.
  13 *
  14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 * OTHER DEALINGS IN THE SOFTWARE.
  21 */
  22
  23#define pr_fmt(fmt) "kfd2kgd: " fmt
  24
  25#include <linux/list.h>
  26#include <drm/drmP.h>
  27#include "amdgpu_object.h"
  28#include "amdgpu_vm.h"
  29#include "amdgpu_amdkfd.h"
  30
  31/* Special VM and GART address alignment needed for VI pre-Fiji due to
  32 * a HW bug.
  33 */
  34#define VI_BO_SIZE_ALIGN (0x8000)
  35
  36/* Impose limit on how much memory KFD can use */
  37static struct {
  38	uint64_t max_system_mem_limit;
  39	int64_t system_mem_used;
  40	spinlock_t mem_limit_lock;
  41} kfd_mem_limit;
  42
  43/* Struct used for amdgpu_amdkfd_bo_validate */
  44struct amdgpu_vm_parser {
  45	uint32_t        domain;
  46	bool            wait;
  47};
  48
  49static const char * const domain_bit_to_string[] = {
  50		"CPU",
  51		"GTT",
  52		"VRAM",
  53		"GDS",
  54		"GWS",
  55		"OA"
  56};
  57
  58#define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  59
  60
  61
  62static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
  63{
  64	return (struct amdgpu_device *)kgd;
  65}
  66
  67static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm,
  68		struct kgd_mem *mem)
  69{
  70	struct kfd_bo_va_list *entry;
  71
  72	list_for_each_entry(entry, &mem->bo_va_list, bo_list)
  73		if (entry->bo_va->base.vm == avm)
  74			return false;
  75
  76	return true;
  77}
  78
  79/* Set memory usage limits. Current, limits are
  80 *  System (kernel) memory - 3/8th System RAM
  81 */
  82void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
  83{
  84	struct sysinfo si;
  85	uint64_t mem;
  86
  87	si_meminfo(&si);
  88	mem = si.totalram - si.totalhigh;
  89	mem *= si.mem_unit;
  90
  91	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
  92	kfd_mem_limit.max_system_mem_limit = (mem >> 1) - (mem >> 3);
  93	pr_debug("Kernel memory limit %lluM\n",
  94		(kfd_mem_limit.max_system_mem_limit >> 20));
  95}
  96
  97static int amdgpu_amdkfd_reserve_system_mem_limit(struct amdgpu_device *adev,
  98					      uint64_t size, u32 domain)
  99{
 100	size_t acc_size;
 101	int ret = 0;
 102
 103	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 104				       sizeof(struct amdgpu_bo));
 105
 106	spin_lock(&kfd_mem_limit.mem_limit_lock);
 107	if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 108		if (kfd_mem_limit.system_mem_used + (acc_size + size) >
 109			kfd_mem_limit.max_system_mem_limit) {
 110			ret = -ENOMEM;
 111			goto err_no_mem;
 112		}
 113		kfd_mem_limit.system_mem_used += (acc_size + size);
 114	}
 115err_no_mem:
 116	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 117	return ret;
 118}
 119
 120static void unreserve_system_mem_limit(struct amdgpu_device *adev,
 121				       uint64_t size, u32 domain)
 122{
 123	size_t acc_size;
 124
 125	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 126				       sizeof(struct amdgpu_bo));
 127
 128	spin_lock(&kfd_mem_limit.mem_limit_lock);
 129	if (domain == AMDGPU_GEM_DOMAIN_GTT)
 130		kfd_mem_limit.system_mem_used -= (acc_size + size);
 131	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 132		  "kfd system memory accounting unbalanced");
 133
 134	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 135}
 136
 137void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo)
 138{
 139	spin_lock(&kfd_mem_limit.mem_limit_lock);
 140
 141	if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT) {
 142		kfd_mem_limit.system_mem_used -=
 143			(bo->tbo.acc_size + amdgpu_bo_size(bo));
 144	}
 145	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 146		  "kfd system memory accounting unbalanced");
 147
 148	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 149}
 150
 151
 152/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence(s) from BO's
 153 *  reservation object.
 154 *
 155 * @bo: [IN] Remove eviction fence(s) from this BO
 156 * @ef: [IN] If ef is specified, then this eviction fence is removed if it
 157 *  is present in the shared list.
 158 * @ef_list: [OUT] Returns list of eviction fences. These fences are removed
 159 *  from BO's reservation object shared list.
 160 * @ef_count: [OUT] Number of fences in ef_list.
 161 *
 162 * NOTE: If called with ef_list, then amdgpu_amdkfd_add_eviction_fence must be
 163 *  called to restore the eviction fences and to avoid memory leak. This is
 164 *  useful for shared BOs.
 165 * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 166 */
 167static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 168					struct amdgpu_amdkfd_fence *ef,
 169					struct amdgpu_amdkfd_fence ***ef_list,
 170					unsigned int *ef_count)
 171{
 172	struct reservation_object_list *fobj;
 173	struct reservation_object *resv;
 174	unsigned int i = 0, j = 0, k = 0, shared_count;
 175	unsigned int count = 0;
 176	struct amdgpu_amdkfd_fence **fence_list;
 177
 178	if (!ef && !ef_list)
 179		return -EINVAL;
 180
 181	if (ef_list) {
 182		*ef_list = NULL;
 183		*ef_count = 0;
 184	}
 185
 186	resv = bo->tbo.resv;
 187	fobj = reservation_object_get_list(resv);
 188
 189	if (!fobj)
 190		return 0;
 191
 192	preempt_disable();
 193	write_seqcount_begin(&resv->seq);
 194
 195	/* Go through all the shared fences in the resevation object. If
 196	 * ef is specified and it exists in the list, remove it and reduce the
 197	 * count. If ef is not specified, then get the count of eviction fences
 198	 * present.
 199	 */
 200	shared_count = fobj->shared_count;
 201	for (i = 0; i < shared_count; ++i) {
 202		struct dma_fence *f;
 203
 204		f = rcu_dereference_protected(fobj->shared[i],
 205					      reservation_object_held(resv));
 206
 207		if (ef) {
 208			if (f->context == ef->base.context) {
 209				dma_fence_put(f);
 210				fobj->shared_count--;
 211			} else {
 212				RCU_INIT_POINTER(fobj->shared[j++], f);
 213			}
 214		} else if (to_amdgpu_amdkfd_fence(f))
 215			count++;
 216	}
 217	write_seqcount_end(&resv->seq);
 218	preempt_enable();
 219
 220	if (ef || !count)
 221		return 0;
 222
 223	/* Alloc memory for count number of eviction fence pointers. Fill the
 224	 * ef_list array and ef_count
 225	 */
 226	fence_list = kcalloc(count, sizeof(struct amdgpu_amdkfd_fence *),
 227			     GFP_KERNEL);
 228	if (!fence_list)
 229		return -ENOMEM;
 230
 231	preempt_disable();
 232	write_seqcount_begin(&resv->seq);
 233
 234	j = 0;
 235	for (i = 0; i < shared_count; ++i) {
 236		struct dma_fence *f;
 237		struct amdgpu_amdkfd_fence *efence;
 238
 239		f = rcu_dereference_protected(fobj->shared[i],
 240			reservation_object_held(resv));
 241
 242		efence = to_amdgpu_amdkfd_fence(f);
 243		if (efence) {
 244			fence_list[k++] = efence;
 245			fobj->shared_count--;
 246		} else {
 247			RCU_INIT_POINTER(fobj->shared[j++], f);
 248		}
 249	}
 250
 251	write_seqcount_end(&resv->seq);
 252	preempt_enable();
 253
 254	*ef_list = fence_list;
 255	*ef_count = k;
 256
 257	return 0;
 258}
 259
 260/* amdgpu_amdkfd_add_eviction_fence - Adds eviction fence(s) back into BO's
 261 *  reservation object.
 262 *
 263 * @bo: [IN] Add eviction fences to this BO
 264 * @ef_list: [IN] List of eviction fences to be added
 265 * @ef_count: [IN] Number of fences in ef_list.
 266 *
 267 * NOTE: Must call amdgpu_amdkfd_remove_eviction_fence before calling this
 268 *  function.
 269 */
 270static void amdgpu_amdkfd_add_eviction_fence(struct amdgpu_bo *bo,
 271				struct amdgpu_amdkfd_fence **ef_list,
 272				unsigned int ef_count)
 273{
 274	int i;
 275
 276	if (!ef_list || !ef_count)
 277		return;
 278
 279	for (i = 0; i < ef_count; i++) {
 280		amdgpu_bo_fence(bo, &ef_list[i]->base, true);
 281		/* Re-adding the fence takes an additional reference. Drop that
 282		 * reference.
 283		 */
 284		dma_fence_put(&ef_list[i]->base);
 285	}
 286
 287	kfree(ef_list);
 288}
 289
 290static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 291				     bool wait)
 292{
 293	struct ttm_operation_ctx ctx = { false, false };
 294	int ret;
 295
 296	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 297		 "Called with userptr BO"))
 298		return -EINVAL;
 299
 300	amdgpu_ttm_placement_from_domain(bo, domain);
 301
 302	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 303	if (ret)
 304		goto validate_fail;
 305	if (wait) {
 306		struct amdgpu_amdkfd_fence **ef_list;
 307		unsigned int ef_count;
 308
 309		ret = amdgpu_amdkfd_remove_eviction_fence(bo, NULL, &ef_list,
 310							  &ef_count);
 311		if (ret)
 312			goto validate_fail;
 313
 314		ttm_bo_wait(&bo->tbo, false, false);
 315		amdgpu_amdkfd_add_eviction_fence(bo, ef_list, ef_count);
 316	}
 317
 318validate_fail:
 319	return ret;
 320}
 321
 322static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo)
 323{
 324	struct amdgpu_vm_parser *p = param;
 325
 326	return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait);
 327}
 328
 329/* vm_validate_pt_pd_bos - Validate page table and directory BOs
 330 *
 331 * Page directories are not updated here because huge page handling
 332 * during page table updates can invalidate page directory entries
 333 * again. Page directories are only updated after updating page
 334 * tables.
 335 */
 336static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
 337{
 338	struct amdgpu_bo *pd = vm->root.base.bo;
 339	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 340	struct amdgpu_vm_parser param;
 341	uint64_t addr, flags = AMDGPU_PTE_VALID;
 342	int ret;
 343
 344	param.domain = AMDGPU_GEM_DOMAIN_VRAM;
 345	param.wait = false;
 346
 347	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate,
 348					&param);
 349	if (ret) {
 350		pr_err("amdgpu: failed to validate PT BOs\n");
 351		return ret;
 352	}
 353
 354	ret = amdgpu_amdkfd_validate(&param, pd);
 355	if (ret) {
 356		pr_err("amdgpu: failed to validate PD\n");
 357		return ret;
 358	}
 359
 360	addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
 361	amdgpu_gmc_get_vm_pde(adev, -1, &addr, &flags);
 362	vm->pd_phys_addr = addr;
 363
 364	if (vm->use_cpu_for_update) {
 365		ret = amdgpu_bo_kmap(pd, NULL);
 366		if (ret) {
 367			pr_err("amdgpu: failed to kmap PD, ret=%d\n", ret);
 368			return ret;
 369		}
 370	}
 371
 372	return 0;
 373}
 374
 375static int sync_vm_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 376			 struct dma_fence *f)
 377{
 378	int ret = amdgpu_sync_fence(adev, sync, f, false);
 379
 380	/* Sync objects can't handle multiple GPUs (contexts) updating
 381	 * sync->last_vm_update. Fortunately we don't need it for
 382	 * KFD's purposes, so we can just drop that fence.
 383	 */
 384	if (sync->last_vm_update) {
 385		dma_fence_put(sync->last_vm_update);
 386		sync->last_vm_update = NULL;
 387	}
 388
 389	return ret;
 390}
 391
 392static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 393{
 394	struct amdgpu_bo *pd = vm->root.base.bo;
 395	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 396	int ret;
 397
 398	ret = amdgpu_vm_update_directories(adev, vm);
 399	if (ret)
 400		return ret;
 401
 402	return sync_vm_fence(adev, sync, vm->last_update);
 403}
 404
 405/* add_bo_to_vm - Add a BO to a VM
 406 *
 407 * Everything that needs to bo done only once when a BO is first added
 408 * to a VM. It can later be mapped and unmapped many times without
 409 * repeating these steps.
 410 *
 411 * 1. Allocate and initialize BO VA entry data structure
 412 * 2. Add BO to the VM
 413 * 3. Determine ASIC-specific PTE flags
 414 * 4. Alloc page tables and directories if needed
 415 * 4a.  Validate new page tables and directories
 416 */
 417static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem,
 418		struct amdgpu_vm *vm, bool is_aql,
 419		struct kfd_bo_va_list **p_bo_va_entry)
 420{
 421	int ret;
 422	struct kfd_bo_va_list *bo_va_entry;
 423	struct amdgpu_bo *pd = vm->root.base.bo;
 424	struct amdgpu_bo *bo = mem->bo;
 425	uint64_t va = mem->va;
 426	struct list_head *list_bo_va = &mem->bo_va_list;
 427	unsigned long bo_size = bo->tbo.mem.size;
 428
 429	if (!va) {
 430		pr_err("Invalid VA when adding BO to VM\n");
 431		return -EINVAL;
 432	}
 433
 434	if (is_aql)
 435		va += bo_size;
 436
 437	bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL);
 438	if (!bo_va_entry)
 439		return -ENOMEM;
 440
 441	pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 442			va + bo_size, vm);
 443
 444	/* Add BO to VM internal data structures*/
 445	bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, vm, bo);
 446	if (!bo_va_entry->bo_va) {
 447		ret = -EINVAL;
 448		pr_err("Failed to add BO object to VM. ret == %d\n",
 449				ret);
 450		goto err_vmadd;
 451	}
 452
 453	bo_va_entry->va = va;
 454	bo_va_entry->pte_flags = amdgpu_gmc_get_pte_flags(adev,
 455							 mem->mapping_flags);
 456	bo_va_entry->kgd_dev = (void *)adev;
 457	list_add(&bo_va_entry->bo_list, list_bo_va);
 458
 459	if (p_bo_va_entry)
 460		*p_bo_va_entry = bo_va_entry;
 461
 462	/* Allocate new page tables if needed and validate
 463	 * them. Clearing of new page tables and validate need to wait
 464	 * on move fences. We don't want that to trigger the eviction
 465	 * fence, so remove it temporarily.
 466	 */
 467	amdgpu_amdkfd_remove_eviction_fence(pd,
 468					vm->process_info->eviction_fence,
 469					NULL, NULL);
 470
 471	ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo));
 472	if (ret) {
 473		pr_err("Failed to allocate pts, err=%d\n", ret);
 474		goto err_alloc_pts;
 475	}
 476
 477	ret = vm_validate_pt_pd_bos(vm);
 478	if (ret) {
 479		pr_err("validate_pt_pd_bos() failed\n");
 480		goto err_alloc_pts;
 481	}
 482
 483	/* Add the eviction fence back */
 484	amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true);
 485
 486	return 0;
 487
 488err_alloc_pts:
 489	amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true);
 490	amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va);
 491	list_del(&bo_va_entry->bo_list);
 492err_vmadd:
 493	kfree(bo_va_entry);
 494	return ret;
 495}
 496
 497static void remove_bo_from_vm(struct amdgpu_device *adev,
 498		struct kfd_bo_va_list *entry, unsigned long size)
 499{
 500	pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n",
 501			entry->va,
 502			entry->va + size, entry);
 503	amdgpu_vm_bo_rmv(adev, entry->bo_va);
 504	list_del(&entry->bo_list);
 505	kfree(entry);
 506}
 507
 508static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
 509				struct amdkfd_process_info *process_info)
 510{
 511	struct ttm_validate_buffer *entry = &mem->validate_list;
 512	struct amdgpu_bo *bo = mem->bo;
 513
 514	INIT_LIST_HEAD(&entry->head);
 515	entry->shared = true;
 516	entry->bo = &bo->tbo;
 517	mutex_lock(&process_info->lock);
 518	list_add_tail(&entry->head, &process_info->kfd_bo_list);
 519	mutex_unlock(&process_info->lock);
 520}
 521
 522/* Reserving a BO and its page table BOs must happen atomically to
 523 * avoid deadlocks. Some operations update multiple VMs at once. Track
 524 * all the reservation info in a context structure. Optionally a sync
 525 * object can track VM updates.
 526 */
 527struct bo_vm_reservation_context {
 528	struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
 529	unsigned int n_vms;		    /* Number of VMs reserved	    */
 530	struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
 531	struct ww_acquire_ctx ticket;	    /* Reservation ticket	    */
 532	struct list_head list, duplicates;  /* BO lists			    */
 533	struct amdgpu_sync *sync;	    /* Pointer to sync object	    */
 534	bool reserved;			    /* Whether BOs are reserved	    */
 535};
 536
 537enum bo_vm_match {
 538	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
 539	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
 540	BO_VM_ALL,		/* Match all VMs a BO was added to    */
 541};
 542
 543/**
 544 * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
 545 * @mem: KFD BO structure.
 546 * @vm: the VM to reserve.
 547 * @ctx: the struct that will be used in unreserve_bo_and_vms().
 548 */
 549static int reserve_bo_and_vm(struct kgd_mem *mem,
 550			      struct amdgpu_vm *vm,
 551			      struct bo_vm_reservation_context *ctx)
 552{
 553	struct amdgpu_bo *bo = mem->bo;
 554	int ret;
 555
 556	WARN_ON(!vm);
 557
 558	ctx->reserved = false;
 559	ctx->n_vms = 1;
 560	ctx->sync = &mem->sync;
 561
 562	INIT_LIST_HEAD(&ctx->list);
 563	INIT_LIST_HEAD(&ctx->duplicates);
 564
 565	ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
 566	if (!ctx->vm_pd)
 567		return -ENOMEM;
 568
 569	ctx->kfd_bo.robj = bo;
 570	ctx->kfd_bo.priority = 0;
 571	ctx->kfd_bo.tv.bo = &bo->tbo;
 572	ctx->kfd_bo.tv.shared = true;
 573	ctx->kfd_bo.user_pages = NULL;
 574	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 575
 576	amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
 577
 578	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 579				     false, &ctx->duplicates);
 580	if (!ret)
 581		ctx->reserved = true;
 582	else {
 583		pr_err("Failed to reserve buffers in ttm\n");
 584		kfree(ctx->vm_pd);
 585		ctx->vm_pd = NULL;
 586	}
 587
 588	return ret;
 589}
 590
 591/**
 592 * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
 593 * @mem: KFD BO structure.
 594 * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
 595 * is used. Otherwise, a single VM associated with the BO.
 596 * @map_type: the mapping status that will be used to filter the VMs.
 597 * @ctx: the struct that will be used in unreserve_bo_and_vms().
 598 *
 599 * Returns 0 for success, negative for failure.
 600 */
 601static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
 602				struct amdgpu_vm *vm, enum bo_vm_match map_type,
 603				struct bo_vm_reservation_context *ctx)
 604{
 605	struct amdgpu_bo *bo = mem->bo;
 606	struct kfd_bo_va_list *entry;
 607	unsigned int i;
 608	int ret;
 609
 610	ctx->reserved = false;
 611	ctx->n_vms = 0;
 612	ctx->vm_pd = NULL;
 613	ctx->sync = &mem->sync;
 614
 615	INIT_LIST_HEAD(&ctx->list);
 616	INIT_LIST_HEAD(&ctx->duplicates);
 617
 618	list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
 619		if ((vm && vm != entry->bo_va->base.vm) ||
 620			(entry->is_mapped != map_type
 621			&& map_type != BO_VM_ALL))
 622			continue;
 623
 624		ctx->n_vms++;
 625	}
 626
 627	if (ctx->n_vms != 0) {
 628		ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
 629				     GFP_KERNEL);
 630		if (!ctx->vm_pd)
 631			return -ENOMEM;
 632	}
 633
 634	ctx->kfd_bo.robj = bo;
 635	ctx->kfd_bo.priority = 0;
 636	ctx->kfd_bo.tv.bo = &bo->tbo;
 637	ctx->kfd_bo.tv.shared = true;
 638	ctx->kfd_bo.user_pages = NULL;
 639	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
 640
 641	i = 0;
 642	list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
 643		if ((vm && vm != entry->bo_va->base.vm) ||
 644			(entry->is_mapped != map_type
 645			&& map_type != BO_VM_ALL))
 646			continue;
 647
 648		amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
 649				&ctx->vm_pd[i]);
 650		i++;
 651	}
 652
 653	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
 654				     false, &ctx->duplicates);
 655	if (!ret)
 656		ctx->reserved = true;
 657	else
 658		pr_err("Failed to reserve buffers in ttm.\n");
 659
 660	if (ret) {
 661		kfree(ctx->vm_pd);
 662		ctx->vm_pd = NULL;
 663	}
 664
 665	return ret;
 666}
 667
 668/**
 669 * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
 670 * @ctx: Reservation context to unreserve
 671 * @wait: Optionally wait for a sync object representing pending VM updates
 672 * @intr: Whether the wait is interruptible
 673 *
 674 * Also frees any resources allocated in
 675 * reserve_bo_and_(cond_)vm(s). Returns the status from
 676 * amdgpu_sync_wait.
 677 */
 678static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
 679				 bool wait, bool intr)
 680{
 681	int ret = 0;
 682
 683	if (wait)
 684		ret = amdgpu_sync_wait(ctx->sync, intr);
 685
 686	if (ctx->reserved)
 687		ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
 688	kfree(ctx->vm_pd);
 689
 690	ctx->sync = NULL;
 691
 692	ctx->reserved = false;
 693	ctx->vm_pd = NULL;
 694
 695	return ret;
 696}
 697
 698static int unmap_bo_from_gpuvm(struct amdgpu_device *adev,
 699				struct kfd_bo_va_list *entry,
 700				struct amdgpu_sync *sync)
 701{
 702	struct amdgpu_bo_va *bo_va = entry->bo_va;
 703	struct amdgpu_vm *vm = bo_va->base.vm;
 704	struct amdgpu_bo *pd = vm->root.base.bo;
 705
 706	/* Remove eviction fence from PD (and thereby from PTs too as
 707	 * they share the resv. object). Otherwise during PT update
 708	 * job (see amdgpu_vm_bo_update_mapping), eviction fence would
 709	 * get added to job->sync object and job execution would
 710	 * trigger the eviction fence.
 711	 */
 712	amdgpu_amdkfd_remove_eviction_fence(pd,
 713					    vm->process_info->eviction_fence,
 714					    NULL, NULL);
 715	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
 716
 717	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
 718
 719	/* Add the eviction fence back */
 720	amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true);
 721
 722	sync_vm_fence(adev, sync, bo_va->last_pt_update);
 723
 724	return 0;
 725}
 726
 727static int update_gpuvm_pte(struct amdgpu_device *adev,
 728		struct kfd_bo_va_list *entry,
 729		struct amdgpu_sync *sync)
 730{
 731	int ret;
 732	struct amdgpu_vm *vm;
 733	struct amdgpu_bo_va *bo_va;
 734	struct amdgpu_bo *bo;
 735
 736	bo_va = entry->bo_va;
 737	vm = bo_va->base.vm;
 738	bo = bo_va->base.bo;
 739
 740	/* Update the page tables  */
 741	ret = amdgpu_vm_bo_update(adev, bo_va, false);
 742	if (ret) {
 743		pr_err("amdgpu_vm_bo_update failed\n");
 744		return ret;
 745	}
 746
 747	return sync_vm_fence(adev, sync, bo_va->last_pt_update);
 748}
 749
 750static int map_bo_to_gpuvm(struct amdgpu_device *adev,
 751		struct kfd_bo_va_list *entry, struct amdgpu_sync *sync)
 752{
 753	int ret;
 754
 755	/* Set virtual address for the allocation */
 756	ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0,
 757			       amdgpu_bo_size(entry->bo_va->base.bo),
 758			       entry->pte_flags);
 759	if (ret) {
 760		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
 761				entry->va, ret);
 762		return ret;
 763	}
 764
 765	ret = update_gpuvm_pte(adev, entry, sync);
 766	if (ret) {
 767		pr_err("update_gpuvm_pte() failed\n");
 768		goto update_gpuvm_pte_failed;
 769	}
 770
 771	return 0;
 772
 773update_gpuvm_pte_failed:
 774	unmap_bo_from_gpuvm(adev, entry, sync);
 775	return ret;
 776}
 777
 778static int process_validate_vms(struct amdkfd_process_info *process_info)
 779{
 780	struct amdgpu_vm *peer_vm;
 781	int ret;
 782
 783	list_for_each_entry(peer_vm, &process_info->vm_list_head,
 784			    vm_list_node) {
 785		ret = vm_validate_pt_pd_bos(peer_vm);
 786		if (ret)
 787			return ret;
 788	}
 789
 790	return 0;
 791}
 792
 793static int process_update_pds(struct amdkfd_process_info *process_info,
 794			      struct amdgpu_sync *sync)
 795{
 796	struct amdgpu_vm *peer_vm;
 797	int ret;
 798
 799	list_for_each_entry(peer_vm, &process_info->vm_list_head,
 800			    vm_list_node) {
 801		ret = vm_update_pds(peer_vm, sync);
 802		if (ret)
 803			return ret;
 804	}
 805
 806	return 0;
 807}
 808
 809static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
 810		       struct dma_fence **ef)
 811{
 812	struct amdkfd_process_info *info = NULL;
 813	int ret;
 814
 815	if (!*process_info) {
 816		info = kzalloc(sizeof(*info), GFP_KERNEL);
 817		if (!info)
 818			return -ENOMEM;
 819
 820		mutex_init(&info->lock);
 821		INIT_LIST_HEAD(&info->vm_list_head);
 822		INIT_LIST_HEAD(&info->kfd_bo_list);
 823
 824		info->eviction_fence =
 825			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
 826						   current->mm);
 827		if (!info->eviction_fence) {
 828			pr_err("Failed to create eviction fence\n");
 829			ret = -ENOMEM;
 830			goto create_evict_fence_fail;
 831		}
 832
 833		*process_info = info;
 834		*ef = dma_fence_get(&info->eviction_fence->base);
 835	}
 836
 837	vm->process_info = *process_info;
 838
 839	/* Validate page directory and attach eviction fence */
 840	ret = amdgpu_bo_reserve(vm->root.base.bo, true);
 841	if (ret)
 842		goto reserve_pd_fail;
 843	ret = vm_validate_pt_pd_bos(vm);
 844	if (ret) {
 845		pr_err("validate_pt_pd_bos() failed\n");
 846		goto validate_pd_fail;
 847	}
 848	ret = ttm_bo_wait(&vm->root.base.bo->tbo, false, false);
 849	if (ret)
 850		goto wait_pd_fail;
 851	amdgpu_bo_fence(vm->root.base.bo,
 852			&vm->process_info->eviction_fence->base, true);
 853	amdgpu_bo_unreserve(vm->root.base.bo);
 854
 855	/* Update process info */
 856	mutex_lock(&vm->process_info->lock);
 857	list_add_tail(&vm->vm_list_node,
 858			&(vm->process_info->vm_list_head));
 859	vm->process_info->n_vms++;
 860	mutex_unlock(&vm->process_info->lock);
 861
 862	return 0;
 863
 864wait_pd_fail:
 865validate_pd_fail:
 866	amdgpu_bo_unreserve(vm->root.base.bo);
 867reserve_pd_fail:
 868	vm->process_info = NULL;
 869	if (info) {
 870		/* Two fence references: one in info and one in *ef */
 871		dma_fence_put(&info->eviction_fence->base);
 872		dma_fence_put(*ef);
 873		*ef = NULL;
 874		*process_info = NULL;
 875create_evict_fence_fail:
 876		mutex_destroy(&info->lock);
 877		kfree(info);
 878	}
 879	return ret;
 880}
 881
 882int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, void **vm,
 883					  void **process_info,
 884					  struct dma_fence **ef)
 885{
 886	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 887	struct amdgpu_vm *new_vm;
 888	int ret;
 889
 890	new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL);
 891	if (!new_vm)
 892		return -ENOMEM;
 893
 894	/* Initialize AMDGPU part of the VM */
 895	ret = amdgpu_vm_init(adev, new_vm, AMDGPU_VM_CONTEXT_COMPUTE, 0);
 896	if (ret) {
 897		pr_err("Failed init vm ret %d\n", ret);
 898		goto amdgpu_vm_init_fail;
 899	}
 900
 901	/* Initialize KFD part of the VM and process info */
 902	ret = init_kfd_vm(new_vm, process_info, ef);
 903	if (ret)
 904		goto init_kfd_vm_fail;
 905
 906	*vm = (void *) new_vm;
 907
 908	return 0;
 909
 910init_kfd_vm_fail:
 911	amdgpu_vm_fini(adev, new_vm);
 912amdgpu_vm_init_fail:
 913	kfree(new_vm);
 914	return ret;
 915}
 916
 917int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
 918					   struct file *filp,
 919					   void **vm, void **process_info,
 920					   struct dma_fence **ef)
 921{
 922	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 923	struct drm_file *drm_priv = filp->private_data;
 924	struct amdgpu_fpriv *drv_priv = drm_priv->driver_priv;
 925	struct amdgpu_vm *avm = &drv_priv->vm;
 926	int ret;
 927
 928	/* Already a compute VM? */
 929	if (avm->process_info)
 930		return -EINVAL;
 931
 932	/* Convert VM into a compute VM */
 933	ret = amdgpu_vm_make_compute(adev, avm);
 934	if (ret)
 935		return ret;
 936
 937	/* Initialize KFD part of the VM and process info */
 938	ret = init_kfd_vm(avm, process_info, ef);
 939	if (ret)
 940		return ret;
 941
 942	*vm = (void *)avm;
 943
 944	return 0;
 945}
 946
 947void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
 948				    struct amdgpu_vm *vm)
 949{
 950	struct amdkfd_process_info *process_info = vm->process_info;
 951	struct amdgpu_bo *pd = vm->root.base.bo;
 952
 953	if (!process_info)
 954		return;
 955
 956	/* Release eviction fence from PD */
 957	amdgpu_bo_reserve(pd, false);
 958	amdgpu_bo_fence(pd, NULL, false);
 959	amdgpu_bo_unreserve(pd);
 960
 961	/* Update process info */
 962	mutex_lock(&process_info->lock);
 963	process_info->n_vms--;
 964	list_del(&vm->vm_list_node);
 965	mutex_unlock(&process_info->lock);
 966
 967	/* Release per-process resources when last compute VM is destroyed */
 968	if (!process_info->n_vms) {
 969		WARN_ON(!list_empty(&process_info->kfd_bo_list));
 970
 971		dma_fence_put(&process_info->eviction_fence->base);
 972		mutex_destroy(&process_info->lock);
 973		kfree(process_info);
 974	}
 975}
 976
 977void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm)
 978{
 979	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 980	struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
 981
 982	if (WARN_ON(!kgd || !vm))
 983		return;
 984
 985	pr_debug("Destroying process vm %p\n", vm);
 986
 987	/* Release the VM context */
 988	amdgpu_vm_fini(adev, avm);
 989	kfree(vm);
 990}
 991
 992uint32_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm)
 993{
 994	struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
 995
 996	return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
 997}
 998
 999int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1000		struct kgd_dev *kgd, uint64_t va, uint64_t size,
1001		void *vm, struct kgd_mem **mem,
1002		uint64_t *offset, uint32_t flags)
1003{
1004	struct amdgpu_device *adev = get_amdgpu_device(kgd);
1005	struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1006	struct amdgpu_bo *bo;
1007	int byte_align;
1008	u32 alloc_domain;
1009	u64 alloc_flags;
1010	uint32_t mapping_flags;
1011	int ret;
1012
1013	/*
1014	 * Check on which domain to allocate BO
1015	 */
1016	if (flags & ALLOC_MEM_FLAGS_VRAM) {
1017		alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1018		alloc_flags = AMDGPU_GEM_CREATE_VRAM_CLEARED;
1019		alloc_flags |= (flags & ALLOC_MEM_FLAGS_PUBLIC) ?
1020			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED :
1021			AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
1022	} else if (flags & ALLOC_MEM_FLAGS_GTT) {
1023		alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1024		alloc_flags = 0;
1025	} else {
1026		return -EINVAL;
1027	}
1028
1029	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1030	if (!*mem)
1031		return -ENOMEM;
1032	INIT_LIST_HEAD(&(*mem)->bo_va_list);
1033	mutex_init(&(*mem)->lock);
1034	(*mem)->aql_queue = !!(flags & ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1035
1036	/* Workaround for AQL queue wraparound bug. Map the same
1037	 * memory twice. That means we only actually allocate half
1038	 * the memory.
1039	 */
1040	if ((*mem)->aql_queue)
1041		size = size >> 1;
1042
1043	/* Workaround for TLB bug on older VI chips */
1044	byte_align = (adev->family == AMDGPU_FAMILY_VI &&
1045			adev->asic_type != CHIP_FIJI &&
1046			adev->asic_type != CHIP_POLARIS10 &&
1047			adev->asic_type != CHIP_POLARIS11) ?
1048			VI_BO_SIZE_ALIGN : 1;
1049
1050	mapping_flags = AMDGPU_VM_PAGE_READABLE;
1051	if (flags & ALLOC_MEM_FLAGS_WRITABLE)
1052		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
1053	if (flags & ALLOC_MEM_FLAGS_EXECUTABLE)
1054		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1055	if (flags & ALLOC_MEM_FLAGS_COHERENT)
1056		mapping_flags |= AMDGPU_VM_MTYPE_UC;
1057	else
1058		mapping_flags |= AMDGPU_VM_MTYPE_NC;
1059	(*mem)->mapping_flags = mapping_flags;
1060
1061	amdgpu_sync_create(&(*mem)->sync);
1062
1063	ret = amdgpu_amdkfd_reserve_system_mem_limit(adev, size, alloc_domain);
1064	if (ret) {
1065		pr_debug("Insufficient system memory\n");
1066		goto err_reserve_system_mem;
1067	}
1068
1069	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1070			va, size, domain_string(alloc_domain));
1071
1072	ret = amdgpu_bo_create(adev, size, byte_align,
1073				alloc_domain, alloc_flags, ttm_bo_type_device, NULL, &bo);
1074	if (ret) {
1075		pr_debug("Failed to create BO on domain %s. ret %d\n",
1076				domain_string(alloc_domain), ret);
1077		goto err_bo_create;
1078	}
1079	bo->kfd_bo = *mem;
1080	(*mem)->bo = bo;
1081
1082	(*mem)->va = va;
1083	(*mem)->domain = alloc_domain;
1084	(*mem)->mapped_to_gpu_memory = 0;
1085	(*mem)->process_info = avm->process_info;
1086	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info);
1087
1088	if (offset)
1089		*offset = amdgpu_bo_mmap_offset(bo);
1090
1091	return 0;
1092
1093err_bo_create:
1094	unreserve_system_mem_limit(adev, size, alloc_domain);
1095err_reserve_system_mem:
1096	mutex_destroy(&(*mem)->lock);
1097	kfree(*mem);
1098	return ret;
1099}
1100
1101int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1102		struct kgd_dev *kgd, struct kgd_mem *mem)
1103{
1104	struct amdkfd_process_info *process_info = mem->process_info;
1105	unsigned long bo_size = mem->bo->tbo.mem.size;
1106	struct kfd_bo_va_list *entry, *tmp;
1107	struct bo_vm_reservation_context ctx;
1108	struct ttm_validate_buffer *bo_list_entry;
1109	int ret;
1110
1111	mutex_lock(&mem->lock);
1112
1113	if (mem->mapped_to_gpu_memory > 0) {
1114		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1115				mem->va, bo_size);
1116		mutex_unlock(&mem->lock);
1117		return -EBUSY;
1118	}
1119
1120	mutex_unlock(&mem->lock);
1121	/* lock is not needed after this, since mem is unused and will
1122	 * be freed anyway
1123	 */
1124
1125	/* Make sure restore workers don't access the BO any more */
1126	bo_list_entry = &mem->validate_list;
1127	mutex_lock(&process_info->lock);
1128	list_del(&bo_list_entry->head);
1129	mutex_unlock(&process_info->lock);
1130
1131	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1132	if (unlikely(ret))
1133		return ret;
1134
1135	/* The eviction fence should be removed by the last unmap.
1136	 * TODO: Log an error condition if the bo still has the eviction fence
1137	 * attached
1138	 */
1139	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1140					process_info->eviction_fence,
1141					NULL, NULL);
1142	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1143		mem->va + bo_size * (1 + mem->aql_queue));
1144
1145	/* Remove from VM internal data structures */
1146	list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list)
1147		remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev,
1148				entry, bo_size);
1149
1150	ret = unreserve_bo_and_vms(&ctx, false, false);
1151
1152	/* Free the sync object */
1153	amdgpu_sync_free(&mem->sync);
1154
1155	/* Free the BO*/
1156	amdgpu_bo_unref(&mem->bo);
1157	mutex_destroy(&mem->lock);
1158	kfree(mem);
1159
1160	return ret;
1161}
1162
1163int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1164		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1165{
1166	struct amdgpu_device *adev = get_amdgpu_device(kgd);
1167	struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1168	int ret;
1169	struct amdgpu_bo *bo;
1170	uint32_t domain;
1171	struct kfd_bo_va_list *entry;
1172	struct bo_vm_reservation_context ctx;
1173	struct kfd_bo_va_list *bo_va_entry = NULL;
1174	struct kfd_bo_va_list *bo_va_entry_aql = NULL;
1175	unsigned long bo_size;
1176
1177	/* Make sure restore is not running concurrently.
1178	 */
1179	mutex_lock(&mem->process_info->lock);
1180
1181	mutex_lock(&mem->lock);
1182
1183	bo = mem->bo;
1184
1185	if (!bo) {
1186		pr_err("Invalid BO when mapping memory to GPU\n");
1187		ret = -EINVAL;
1188		goto out;
1189	}
1190
1191	domain = mem->domain;
1192	bo_size = bo->tbo.mem.size;
1193
1194	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1195			mem->va,
1196			mem->va + bo_size * (1 + mem->aql_queue),
1197			vm, domain_string(domain));
1198
1199	ret = reserve_bo_and_vm(mem, vm, &ctx);
1200	if (unlikely(ret))
1201		goto out;
1202
1203	if (check_if_add_bo_to_vm(avm, mem)) {
1204		ret = add_bo_to_vm(adev, mem, avm, false,
1205				&bo_va_entry);
1206		if (ret)
1207			goto add_bo_to_vm_failed;
1208		if (mem->aql_queue) {
1209			ret = add_bo_to_vm(adev, mem, avm,
1210					true, &bo_va_entry_aql);
1211			if (ret)
1212				goto add_bo_to_vm_failed_aql;
1213		}
1214	} else {
1215		ret = vm_validate_pt_pd_bos(avm);
1216		if (unlikely(ret))
1217			goto add_bo_to_vm_failed;
1218	}
1219
1220	if (mem->mapped_to_gpu_memory == 0) {
1221		/* Validate BO only once. The eviction fence gets added to BO
1222		 * the first time it is mapped. Validate will wait for all
1223		 * background evictions to complete.
1224		 */
1225		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1226		if (ret) {
1227			pr_debug("Validate failed\n");
1228			goto map_bo_to_gpuvm_failed;
1229		}
1230	}
1231
1232	list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1233		if (entry->bo_va->base.vm == vm && !entry->is_mapped) {
1234			pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1235					entry->va, entry->va + bo_size,
1236					entry);
1237
1238			ret = map_bo_to_gpuvm(adev, entry, ctx.sync);
1239			if (ret) {
1240				pr_err("Failed to map radeon bo to gpuvm\n");
1241				goto map_bo_to_gpuvm_failed;
1242			}
1243
1244			ret = vm_update_pds(vm, ctx.sync);
1245			if (ret) {
1246				pr_err("Failed to update page directories\n");
1247				goto map_bo_to_gpuvm_failed;
1248			}
1249
1250			entry->is_mapped = true;
1251			mem->mapped_to_gpu_memory++;
1252			pr_debug("\t INC mapping count %d\n",
1253					mem->mapped_to_gpu_memory);
1254		}
1255	}
1256
1257	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->pin_count)
1258		amdgpu_bo_fence(bo,
1259				&avm->process_info->eviction_fence->base,
1260				true);
1261	ret = unreserve_bo_and_vms(&ctx, false, false);
1262
1263	goto out;
1264
1265map_bo_to_gpuvm_failed:
1266	if (bo_va_entry_aql)
1267		remove_bo_from_vm(adev, bo_va_entry_aql, bo_size);
1268add_bo_to_vm_failed_aql:
1269	if (bo_va_entry)
1270		remove_bo_from_vm(adev, bo_va_entry, bo_size);
1271add_bo_to_vm_failed:
1272	unreserve_bo_and_vms(&ctx, false, false);
1273out:
1274	mutex_unlock(&mem->process_info->lock);
1275	mutex_unlock(&mem->lock);
1276	return ret;
1277}
1278
1279int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1280		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1281{
1282	struct amdgpu_device *adev = get_amdgpu_device(kgd);
1283	struct amdkfd_process_info *process_info =
1284		((struct amdgpu_vm *)vm)->process_info;
1285	unsigned long bo_size = mem->bo->tbo.mem.size;
1286	struct kfd_bo_va_list *entry;
1287	struct bo_vm_reservation_context ctx;
1288	int ret;
1289
1290	mutex_lock(&mem->lock);
1291
1292	ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx);
1293	if (unlikely(ret))
1294		goto out;
1295	/* If no VMs were reserved, it means the BO wasn't actually mapped */
1296	if (ctx.n_vms == 0) {
1297		ret = -EINVAL;
1298		goto unreserve_out;
1299	}
1300
1301	ret = vm_validate_pt_pd_bos((struct amdgpu_vm *)vm);
1302	if (unlikely(ret))
1303		goto unreserve_out;
1304
1305	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1306		mem->va,
1307		mem->va + bo_size * (1 + mem->aql_queue),
1308		vm);
1309
1310	list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1311		if (entry->bo_va->base.vm == vm && entry->is_mapped) {
1312			pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1313					entry->va,
1314					entry->va + bo_size,
1315					entry);
1316
1317			ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync);
1318			if (ret == 0) {
1319				entry->is_mapped = false;
1320			} else {
1321				pr_err("failed to unmap VA 0x%llx\n",
1322						mem->va);
1323				goto unreserve_out;
1324			}
1325
1326			mem->mapped_to_gpu_memory--;
1327			pr_debug("\t DEC mapping count %d\n",
1328					mem->mapped_to_gpu_memory);
1329		}
1330	}
1331
1332	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
1333	 * required.
1334	 */
1335	if (mem->mapped_to_gpu_memory == 0 &&
1336	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && !mem->bo->pin_count)
1337		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1338						process_info->eviction_fence,
1339						    NULL, NULL);
1340
1341unreserve_out:
1342	unreserve_bo_and_vms(&ctx, false, false);
1343out:
1344	mutex_unlock(&mem->lock);
1345	return ret;
1346}
1347
1348int amdgpu_amdkfd_gpuvm_sync_memory(
1349		struct kgd_dev *kgd, struct kgd_mem *mem, bool intr)
1350{
1351	struct amdgpu_sync sync;
1352	int ret;
1353
1354	amdgpu_sync_create(&sync);
1355
1356	mutex_lock(&mem->lock);
1357	amdgpu_sync_clone(&mem->sync, &sync);
1358	mutex_unlock(&mem->lock);
1359
1360	ret = amdgpu_sync_wait(&sync, intr);
1361	amdgpu_sync_free(&sync);
1362	return ret;
1363}
1364
1365int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
1366		struct kgd_mem *mem, void **kptr, uint64_t *size)
1367{
1368	int ret;
1369	struct amdgpu_bo *bo = mem->bo;
1370
1371	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1372		pr_err("userptr can't be mapped to kernel\n");
1373		return -EINVAL;
1374	}
1375
1376	/* delete kgd_mem from kfd_bo_list to avoid re-validating
1377	 * this BO in BO's restoring after eviction.
1378	 */
1379	mutex_lock(&mem->process_info->lock);
1380
1381	ret = amdgpu_bo_reserve(bo, true);
1382	if (ret) {
1383		pr_err("Failed to reserve bo. ret %d\n", ret);
1384		goto bo_reserve_failed;
1385	}
1386
1387	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT, NULL);
1388	if (ret) {
1389		pr_err("Failed to pin bo. ret %d\n", ret);
1390		goto pin_failed;
1391	}
1392
1393	ret = amdgpu_bo_kmap(bo, kptr);
1394	if (ret) {
1395		pr_err("Failed to map bo to kernel. ret %d\n", ret);
1396		goto kmap_failed;
1397	}
1398
1399	amdgpu_amdkfd_remove_eviction_fence(
1400		bo, mem->process_info->eviction_fence, NULL, NULL);
1401	list_del_init(&mem->validate_list.head);
1402
1403	if (size)
1404		*size = amdgpu_bo_size(bo);
1405
1406	amdgpu_bo_unreserve(bo);
1407
1408	mutex_unlock(&mem->process_info->lock);
1409	return 0;
1410
1411kmap_failed:
1412	amdgpu_bo_unpin(bo);
1413pin_failed:
1414	amdgpu_bo_unreserve(bo);
1415bo_reserve_failed:
1416	mutex_unlock(&mem->process_info->lock);
1417
1418	return ret;
1419}
1420
1421/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
1422 *   KFD process identified by process_info
1423 *
1424 * @process_info: amdkfd_process_info of the KFD process
1425 *
1426 * After memory eviction, restore thread calls this function. The function
1427 * should be called when the Process is still valid. BO restore involves -
1428 *
1429 * 1.  Release old eviction fence and create new one
1430 * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
1431 * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
1432 *     BOs that need to be reserved.
1433 * 4.  Reserve all the BOs
1434 * 5.  Validate of PD and PT BOs.
1435 * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
1436 * 7.  Add fence to all PD and PT BOs.
1437 * 8.  Unreserve all BOs
1438 */
1439int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
1440{
1441	struct amdgpu_bo_list_entry *pd_bo_list;
1442	struct amdkfd_process_info *process_info = info;
1443	struct amdgpu_vm *peer_vm;
1444	struct kgd_mem *mem;
1445	struct bo_vm_reservation_context ctx;
1446	struct amdgpu_amdkfd_fence *new_fence;
1447	int ret = 0, i;
1448	struct list_head duplicate_save;
1449	struct amdgpu_sync sync_obj;
1450
1451	INIT_LIST_HEAD(&duplicate_save);
1452	INIT_LIST_HEAD(&ctx.list);
1453	INIT_LIST_HEAD(&ctx.duplicates);
1454
1455	pd_bo_list = kcalloc(process_info->n_vms,
1456			     sizeof(struct amdgpu_bo_list_entry),
1457			     GFP_KERNEL);
1458	if (!pd_bo_list)
1459		return -ENOMEM;
1460
1461	i = 0;
1462	mutex_lock(&process_info->lock);
1463	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1464			vm_list_node)
1465		amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
1466
1467	/* Reserve all BOs and page tables/directory. Add all BOs from
1468	 * kfd_bo_list to ctx.list
1469	 */
1470	list_for_each_entry(mem, &process_info->kfd_bo_list,
1471			    validate_list.head) {
1472
1473		list_add_tail(&mem->resv_list.head, &ctx.list);
1474		mem->resv_list.bo = mem->validate_list.bo;
1475		mem->resv_list.shared = mem->validate_list.shared;
1476	}
1477
1478	ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
1479				     false, &duplicate_save);
1480	if (ret) {
1481		pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
1482		goto ttm_reserve_fail;
1483	}
1484
1485	amdgpu_sync_create(&sync_obj);
1486
1487	/* Validate PDs and PTs */
1488	ret = process_validate_vms(process_info);
1489	if (ret)
1490		goto validate_map_fail;
1491
1492	/* Wait for PD/PTs validate to finish */
1493	/* FIXME: I think this isn't needed */
1494	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1495			    vm_list_node) {
1496		struct amdgpu_bo *bo = peer_vm->root.base.bo;
1497
1498		ttm_bo_wait(&bo->tbo, false, false);
1499	}
1500
1501	/* Validate BOs and map them to GPUVM (update VM page tables). */
1502	list_for_each_entry(mem, &process_info->kfd_bo_list,
1503			    validate_list.head) {
1504
1505		struct amdgpu_bo *bo = mem->bo;
1506		uint32_t domain = mem->domain;
1507		struct kfd_bo_va_list *bo_va_entry;
1508
1509		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
1510		if (ret) {
1511			pr_debug("Memory eviction: Validate BOs failed. Try again\n");
1512			goto validate_map_fail;
1513		}
1514
1515		list_for_each_entry(bo_va_entry, &mem->bo_va_list,
1516				    bo_list) {
1517			ret = update_gpuvm_pte((struct amdgpu_device *)
1518					      bo_va_entry->kgd_dev,
1519					      bo_va_entry,
1520					      &sync_obj);
1521			if (ret) {
1522				pr_debug("Memory eviction: update PTE failed. Try again\n");
1523				goto validate_map_fail;
1524			}
1525		}
1526	}
1527
1528	/* Update page directories */
1529	ret = process_update_pds(process_info, &sync_obj);
1530	if (ret) {
1531		pr_debug("Memory eviction: update PDs failed. Try again\n");
1532		goto validate_map_fail;
1533	}
1534
1535	amdgpu_sync_wait(&sync_obj, false);
1536
1537	/* Release old eviction fence and create new one, because fence only
1538	 * goes from unsignaled to signaled, fence cannot be reused.
1539	 * Use context and mm from the old fence.
1540	 */
1541	new_fence = amdgpu_amdkfd_fence_create(
1542				process_info->eviction_fence->base.context,
1543				process_info->eviction_fence->mm);
1544	if (!new_fence) {
1545		pr_err("Failed to create eviction fence\n");
1546		ret = -ENOMEM;
1547		goto validate_map_fail;
1548	}
1549	dma_fence_put(&process_info->eviction_fence->base);
1550	process_info->eviction_fence = new_fence;
1551	*ef = dma_fence_get(&new_fence->base);
1552
1553	/* Wait for validate to finish and attach new eviction fence */
1554	list_for_each_entry(mem, &process_info->kfd_bo_list,
1555		validate_list.head)
1556		ttm_bo_wait(&mem->bo->tbo, false, false);
1557	list_for_each_entry(mem, &process_info->kfd_bo_list,
1558		validate_list.head)
1559		amdgpu_bo_fence(mem->bo,
1560			&process_info->eviction_fence->base, true);
1561
1562	/* Attach eviction fence to PD / PT BOs */
1563	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1564			    vm_list_node) {
1565		struct amdgpu_bo *bo = peer_vm->root.base.bo;
1566
1567		amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
1568	}
1569
1570validate_map_fail:
1571	ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
1572	amdgpu_sync_free(&sync_obj);
1573ttm_reserve_fail:
1574	mutex_unlock(&process_info->lock);
1575	kfree(pd_bo_list);
1576	return ret;
1577}