Linux Audio

Check our new training course

Loading...
v6.9.4
   1// SPDX-License-Identifier: MIT
   2/*
   3 * Copyright 2014-2018 Advanced Micro Devices, Inc.
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice shall be included in
  13 * all copies or substantial portions of the Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21 * OTHER DEALINGS IN THE SOFTWARE.
  22 */
  23#include <linux/dma-buf.h>
  24#include <linux/list.h>
  25#include <linux/pagemap.h>
  26#include <linux/sched/mm.h>
  27#include <linux/sched/task.h>
  28#include <linux/fdtable.h>
  29#include <drm/ttm/ttm_tt.h>
  30
  31#include <drm/drm_exec.h>
  32
  33#include "amdgpu_object.h"
  34#include "amdgpu_gem.h"
  35#include "amdgpu_vm.h"
  36#include "amdgpu_hmm.h"
  37#include "amdgpu_amdkfd.h"
  38#include "amdgpu_dma_buf.h"
  39#include <uapi/linux/kfd_ioctl.h>
  40#include "amdgpu_xgmi.h"
  41#include "kfd_priv.h"
  42#include "kfd_smi_events.h"
  43
  44/* Userptr restore delay, just long enough to allow consecutive VM
  45 * changes to accumulate
  46 */
  47#define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
  48#define AMDGPU_RESERVE_MEM_LIMIT			(3UL << 29)
  49
  50/*
  51 * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
  52 * BO chunk
  53 */
  54#define VRAM_AVAILABLITY_ALIGN (1 << 21)
  55
  56/* Impose limit on how much memory KFD can use */
  57static struct {
  58	uint64_t max_system_mem_limit;
  59	uint64_t max_ttm_mem_limit;
  60	int64_t system_mem_used;
  61	int64_t ttm_mem_used;
  62	spinlock_t mem_limit_lock;
  63} kfd_mem_limit;
  64
  65static const char * const domain_bit_to_string[] = {
  66		"CPU",
  67		"GTT",
  68		"VRAM",
  69		"GDS",
  70		"GWS",
  71		"OA"
  72};
  73
  74#define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  75
  76static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
  77
  78static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
  79		struct kgd_mem *mem)
  80{
  81	struct kfd_mem_attachment *entry;
  82
  83	list_for_each_entry(entry, &mem->attachments, list)
  84		if (entry->bo_va->base.vm == avm)
  85			return true;
  86
  87	return false;
  88}
  89
  90/**
  91 * reuse_dmamap() - Check whether adev can share the original
  92 * userptr BO
  93 *
  94 * If both adev and bo_adev are in direct mapping or
  95 * in the same iommu group, they can share the original BO.
  96 *
  97 * @adev: Device to which can or cannot share the original BO
  98 * @bo_adev: Device to which allocated BO belongs to
  99 *
 100 * Return: returns true if adev can share original userptr BO,
 101 * false otherwise.
 102 */
 103static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
 104{
 105	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
 106			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
 107}
 108
 109/* Set memory usage limits. Current, limits are
 110 *  System (TTM + userptr) memory - 15/16th System RAM
 111 *  TTM memory - 3/8th System RAM
 112 */
 113void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 114{
 115	struct sysinfo si;
 116	uint64_t mem;
 117
 118	if (kfd_mem_limit.max_system_mem_limit)
 119		return;
 120
 121	si_meminfo(&si);
 122	mem = si.totalram - si.totalhigh;
 123	mem *= si.mem_unit;
 124
 125	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
 126	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6);
 127	if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT)
 128		kfd_mem_limit.max_system_mem_limit >>= 1;
 129	else
 130		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
 131
 132	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
 133	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
 134		(kfd_mem_limit.max_system_mem_limit >> 20),
 135		(kfd_mem_limit.max_ttm_mem_limit >> 20));
 136}
 137
 138void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
 139{
 140	kfd_mem_limit.system_mem_used += size;
 141}
 142
 143/* Estimate page table size needed to represent a given memory size
 144 *
 145 * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
 146 * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
 147 * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
 148 * for 2MB pages for TLB efficiency. However, small allocations and
 149 * fragmented system memory still need some 4KB pages. We choose a
 150 * compromise that should work in most cases without reserving too
 151 * much memory for page tables unnecessarily (factor 16K, >> 14).
 152 */
 153
 154#define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
 155
 156/**
 157 * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
 158 * of buffer.
 159 *
 160 * @adev: Device to which allocated BO belongs to
 161 * @size: Size of buffer, in bytes, encapsulated by B0. This should be
 162 * equivalent to amdgpu_bo_size(BO)
 163 * @alloc_flag: Flag used in allocating a BO as noted above
 164 * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
 165 * managed as one compute node in driver for app
 166 *
 167 * Return:
 168 *	returns -ENOMEM in case of error, ZERO otherwise
 169 */
 170int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 171		uint64_t size, u32 alloc_flag, int8_t xcp_id)
 172{
 173	uint64_t reserved_for_pt =
 174		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
 175	size_t system_mem_needed, ttm_mem_needed, vram_needed;
 176	int ret = 0;
 177	uint64_t vram_size = 0;
 178
 179	system_mem_needed = 0;
 180	ttm_mem_needed = 0;
 181	vram_needed = 0;
 182	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 183		system_mem_needed = size;
 184		ttm_mem_needed = size;
 185	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 186		/*
 187		 * Conservatively round up the allocation requirement to 2 MB
 188		 * to avoid fragmentation caused by 4K allocations in the tail
 189		 * 2M BO chunk.
 190		 */
 191		vram_needed = size;
 192		/*
 193		 * For GFX 9.4.3, get the VRAM size from XCP structs
 194		 */
 195		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 196			return -EINVAL;
 197
 198		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
 199		if (adev->gmc.is_app_apu) {
 200			system_mem_needed = size;
 201			ttm_mem_needed = size;
 202		}
 203	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 204		system_mem_needed = size;
 205	} else if (!(alloc_flag &
 206				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 207				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 208		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 209		return -ENOMEM;
 210	}
 211
 212	spin_lock(&kfd_mem_limit.mem_limit_lock);
 213
 214	if (kfd_mem_limit.system_mem_used + system_mem_needed >
 215	    kfd_mem_limit.max_system_mem_limit)
 216		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
 217
 218	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
 219	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
 220	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 221	     kfd_mem_limit.max_ttm_mem_limit) ||
 222	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
 223	     vram_size - reserved_for_pt - atomic64_read(&adev->vram_pin_size))) {
 224		ret = -ENOMEM;
 225		goto release;
 226	}
 227
 228	/* Update memory accounting by decreasing available system
 229	 * memory, TTM memory and GPU memory as computed above
 230	 */
 231	WARN_ONCE(vram_needed && !adev,
 232		  "adev reference can't be null when vram is used");
 233	if (adev && xcp_id >= 0) {
 234		adev->kfd.vram_used[xcp_id] += vram_needed;
 235		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
 236				vram_needed :
 237				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 238	}
 239	kfd_mem_limit.system_mem_used += system_mem_needed;
 240	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 241
 242release:
 243	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 244	return ret;
 245}
 246
 247void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 248		uint64_t size, u32 alloc_flag, int8_t xcp_id)
 249{
 250	spin_lock(&kfd_mem_limit.mem_limit_lock);
 251
 252	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 253		kfd_mem_limit.system_mem_used -= size;
 254		kfd_mem_limit.ttm_mem_used -= size;
 255	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 256		WARN_ONCE(!adev,
 257			  "adev reference can't be null when alloc mem flags vram is set");
 258		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 259			goto release;
 260
 261		if (adev) {
 262			adev->kfd.vram_used[xcp_id] -= size;
 263			if (adev->gmc.is_app_apu) {
 264				adev->kfd.vram_used_aligned[xcp_id] -= size;
 265				kfd_mem_limit.system_mem_used -= size;
 266				kfd_mem_limit.ttm_mem_used -= size;
 267			} else {
 268				adev->kfd.vram_used_aligned[xcp_id] -=
 269					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
 270			}
 271		}
 272	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 273		kfd_mem_limit.system_mem_used -= size;
 274	} else if (!(alloc_flag &
 275				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 276				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 277		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 278		goto release;
 279	}
 280	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
 281		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
 282	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 283		  "KFD TTM memory accounting unbalanced");
 284	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 285		  "KFD system memory accounting unbalanced");
 286
 287release:
 288	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 289}
 290
 291void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
 292{
 293	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 294	u32 alloc_flags = bo->kfd_bo->alloc_flags;
 295	u64 size = amdgpu_bo_size(bo);
 296
 297	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
 298					  bo->xcp_id);
 299
 300	kfree(bo->kfd_bo);
 301}
 302
 303/**
 304 * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
 305 * about USERPTR or DOOREBELL or MMIO BO.
 306 *
 307 * @adev: Device for which dmamap BO is being created
 308 * @mem: BO of peer device that is being DMA mapped. Provides parameters
 309 *	 in building the dmamap BO
 310 * @bo_out: Output parameter updated with handle of dmamap BO
 311 */
 312static int
 313create_dmamap_sg_bo(struct amdgpu_device *adev,
 314		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
 315{
 316	struct drm_gem_object *gem_obj;
 317	int ret;
 318	uint64_t flags = 0;
 319
 320	ret = amdgpu_bo_reserve(mem->bo, false);
 321	if (ret)
 322		return ret;
 323
 324	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
 325		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
 326					AMDGPU_GEM_CREATE_UNCACHED);
 327
 328	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
 329			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
 330			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
 331
 332	amdgpu_bo_unreserve(mem->bo);
 333
 334	if (ret) {
 335		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
 336		return -EINVAL;
 337	}
 338
 339	*bo_out = gem_to_amdgpu_bo(gem_obj);
 340	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
 341	return ret;
 342}
 343
 344/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
 345 *  reservation object.
 346 *
 347 * @bo: [IN] Remove eviction fence(s) from this BO
 348 * @ef: [IN] This eviction fence is removed if it
 349 *  is present in the shared list.
 350 *
 351 * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 352 */
 353static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 354					struct amdgpu_amdkfd_fence *ef)
 355{
 356	struct dma_fence *replacement;
 357
 358	if (!ef)
 359		return -EINVAL;
 360
 361	/* TODO: Instead of block before we should use the fence of the page
 362	 * table update and TLB flush here directly.
 363	 */
 364	replacement = dma_fence_get_stub();
 365	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
 366				replacement, DMA_RESV_USAGE_BOOKKEEP);
 367	dma_fence_put(replacement);
 368	return 0;
 369}
 370
 371int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
 372{
 373	struct amdgpu_bo *root = bo;
 374	struct amdgpu_vm_bo_base *vm_bo;
 375	struct amdgpu_vm *vm;
 376	struct amdkfd_process_info *info;
 377	struct amdgpu_amdkfd_fence *ef;
 378	int ret;
 379
 380	/* we can always get vm_bo from root PD bo.*/
 381	while (root->parent)
 382		root = root->parent;
 383
 384	vm_bo = root->vm_bo;
 385	if (!vm_bo)
 386		return 0;
 387
 388	vm = vm_bo->vm;
 389	if (!vm)
 390		return 0;
 391
 392	info = vm->process_info;
 393	if (!info || !info->eviction_fence)
 394		return 0;
 395
 396	ef = container_of(dma_fence_get(&info->eviction_fence->base),
 397			struct amdgpu_amdkfd_fence, base);
 398
 399	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
 400	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
 401	dma_resv_unlock(bo->tbo.base.resv);
 402
 403	dma_fence_put(&ef->base);
 404	return ret;
 405}
 406
 407static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 408				     bool wait)
 409{
 410	struct ttm_operation_ctx ctx = { false, false };
 411	int ret;
 412
 413	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 414		 "Called with userptr BO"))
 415		return -EINVAL;
 416
 417	amdgpu_bo_placement_from_domain(bo, domain);
 418
 419	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 420	if (ret)
 421		goto validate_fail;
 422	if (wait)
 423		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 424
 425validate_fail:
 426	return ret;
 427}
 428
 429int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
 430					uint32_t domain,
 431					struct dma_fence *fence)
 432{
 433	int ret = amdgpu_bo_reserve(bo, false);
 434
 435	if (ret)
 436		return ret;
 437
 438	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
 439	if (ret)
 440		goto unreserve_out;
 441
 442	ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
 443	if (ret)
 444		goto unreserve_out;
 445
 446	dma_resv_add_fence(bo->tbo.base.resv, fence,
 447			   DMA_RESV_USAGE_BOOKKEEP);
 448
 449unreserve_out:
 450	amdgpu_bo_unreserve(bo);
 451
 452	return ret;
 453}
 454
 455static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
 456{
 457	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
 458}
 459
 460/* vm_validate_pt_pd_bos - Validate page table and directory BOs
 461 *
 462 * Page directories are not updated here because huge page handling
 463 * during page table updates can invalidate page directory entries
 464 * again. Page directories are only updated after updating page
 465 * tables.
 466 */
 467static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm,
 468				 struct ww_acquire_ctx *ticket)
 469{
 470	struct amdgpu_bo *pd = vm->root.bo;
 471	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 472	int ret;
 473
 474	ret = amdgpu_vm_validate(adev, vm, ticket,
 475				 amdgpu_amdkfd_validate_vm_bo, NULL);
 476	if (ret) {
 477		pr_err("failed to validate PT BOs\n");
 478		return ret;
 479	}
 480
 481	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 482
 483	return 0;
 484}
 485
 486static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 487{
 488	struct amdgpu_bo *pd = vm->root.bo;
 489	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 490	int ret;
 491
 492	ret = amdgpu_vm_update_pdes(adev, vm, false);
 493	if (ret)
 494		return ret;
 495
 496	return amdgpu_sync_fence(sync, vm->last_update);
 497}
 498
 499static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
 500{
 501	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
 502				 AMDGPU_VM_MTYPE_DEFAULT;
 503
 504	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
 505		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
 506	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
 507		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
 508
 509	return amdgpu_gem_va_map_flags(adev, mapping_flags);
 510}
 511
 512/**
 513 * create_sg_table() - Create an sg_table for a contiguous DMA addr range
 514 * @addr: The starting address to point to
 515 * @size: Size of memory area in bytes being pointed to
 516 *
 517 * Allocates an instance of sg_table and initializes it to point to memory
 518 * area specified by input parameters. The address used to build is assumed
 519 * to be DMA mapped, if needed.
 520 *
 521 * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
 522 * because they are physically contiguous.
 523 *
 524 * Return: Initialized instance of SG Table or NULL
 525 */
 526static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
 527{
 528	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
 529
 530	if (!sg)
 531		return NULL;
 532	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
 533		kfree(sg);
 534		return NULL;
 535	}
 536	sg_dma_address(sg->sgl) = addr;
 537	sg->sgl->length = size;
 538#ifdef CONFIG_NEED_SG_DMA_LENGTH
 539	sg->sgl->dma_length = size;
 540#endif
 541	return sg;
 542}
 543
 544static int
 545kfd_mem_dmamap_userptr(struct kgd_mem *mem,
 546		       struct kfd_mem_attachment *attachment)
 547{
 548	enum dma_data_direction direction =
 549		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 550		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 551	struct ttm_operation_ctx ctx = {.interruptible = true};
 552	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 553	struct amdgpu_device *adev = attachment->adev;
 554	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
 555	struct ttm_tt *ttm = bo->tbo.ttm;
 556	int ret;
 557
 558	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
 559		return -EINVAL;
 560
 561	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
 562	if (unlikely(!ttm->sg))
 563		return -ENOMEM;
 564
 565	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
 566	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
 567					ttm->num_pages, 0,
 568					(u64)ttm->num_pages << PAGE_SHIFT,
 569					GFP_KERNEL);
 570	if (unlikely(ret))
 571		goto free_sg;
 572
 573	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 574	if (unlikely(ret))
 575		goto release_sg;
 576
 577	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 578	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 579	if (ret)
 580		goto unmap_sg;
 581
 582	return 0;
 583
 584unmap_sg:
 585	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 586release_sg:
 587	pr_err("DMA map userptr failed: %d\n", ret);
 588	sg_free_table(ttm->sg);
 589free_sg:
 590	kfree(ttm->sg);
 591	ttm->sg = NULL;
 592	return ret;
 593}
 594
 595static int
 596kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
 597{
 598	struct ttm_operation_ctx ctx = {.interruptible = true};
 599	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 600	int ret;
 601
 602	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 603	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 604	if (ret)
 605		return ret;
 606
 607	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 608	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 609}
 610
 611/**
 612 * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
 613 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 614 * @attachment: Virtual address attachment of the BO on accessing device
 615 *
 616 * An access request from the device that owns DOORBELL does not require DMA mapping.
 617 * This is because the request doesn't go through PCIe root complex i.e. it instead
 618 * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
 619 *
 620 * In contrast, all access requests for MMIO need to be DMA mapped without regard to
 621 * device ownership. This is because access requests for MMIO go through PCIe root
 622 * complex.
 623 *
 624 * This is accomplished in two steps:
 625 *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
 626 *         in updating requesting device's page table
 627 *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
 628 *         accessible. This allows an update of requesting device's page table
 629 *         with entries associated with DOOREBELL or MMIO memory
 630 *
 631 * This method is invoked in the following contexts:
 632 *   - Mapping of DOORBELL or MMIO BO of same or peer device
 633 *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
 634 *
 635 * Return: ZERO if successful, NON-ZERO otherwise
 636 */
 637static int
 638kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
 639		     struct kfd_mem_attachment *attachment)
 640{
 641	struct ttm_operation_ctx ctx = {.interruptible = true};
 642	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 643	struct amdgpu_device *adev = attachment->adev;
 644	struct ttm_tt *ttm = bo->tbo.ttm;
 645	enum dma_data_direction dir;
 646	dma_addr_t dma_addr;
 647	bool mmio;
 648	int ret;
 649
 650	/* Expect SG Table of dmapmap BO to be NULL */
 651	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
 652	if (unlikely(ttm->sg)) {
 653		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
 654		return -EINVAL;
 655	}
 656
 657	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 658			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 659	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
 660	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
 661	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
 662	dma_addr = dma_map_resource(adev->dev, dma_addr,
 663			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 664	ret = dma_mapping_error(adev->dev, dma_addr);
 665	if (unlikely(ret))
 666		return ret;
 667	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
 668
 669	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
 670	if (unlikely(!ttm->sg)) {
 671		ret = -ENOMEM;
 672		goto unmap_sg;
 673	}
 674
 675	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 676	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 677	if (unlikely(ret))
 678		goto free_sg;
 679
 680	return ret;
 681
 682free_sg:
 683	sg_free_table(ttm->sg);
 684	kfree(ttm->sg);
 685	ttm->sg = NULL;
 686unmap_sg:
 687	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
 688			   dir, DMA_ATTR_SKIP_CPU_SYNC);
 689	return ret;
 690}
 691
 692static int
 693kfd_mem_dmamap_attachment(struct kgd_mem *mem,
 694			  struct kfd_mem_attachment *attachment)
 695{
 696	switch (attachment->type) {
 697	case KFD_MEM_ATT_SHARED:
 698		return 0;
 699	case KFD_MEM_ATT_USERPTR:
 700		return kfd_mem_dmamap_userptr(mem, attachment);
 701	case KFD_MEM_ATT_DMABUF:
 702		return kfd_mem_dmamap_dmabuf(attachment);
 703	case KFD_MEM_ATT_SG:
 704		return kfd_mem_dmamap_sg_bo(mem, attachment);
 705	default:
 706		WARN_ON_ONCE(1);
 707	}
 708	return -EINVAL;
 709}
 710
 711static void
 712kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
 713			 struct kfd_mem_attachment *attachment)
 714{
 715	enum dma_data_direction direction =
 716		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 717		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 718	struct ttm_operation_ctx ctx = {.interruptible = false};
 719	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 720	struct amdgpu_device *adev = attachment->adev;
 721	struct ttm_tt *ttm = bo->tbo.ttm;
 722
 723	if (unlikely(!ttm->sg))
 724		return;
 725
 726	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 727	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 728
 729	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 730	sg_free_table(ttm->sg);
 731	kfree(ttm->sg);
 732	ttm->sg = NULL;
 733}
 734
 735static void
 736kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
 737{
 738	/* This is a no-op. We don't want to trigger eviction fences when
 739	 * unmapping DMABufs. Therefore the invalidation (moving to system
 740	 * domain) is done in kfd_mem_dmamap_dmabuf.
 741	 */
 742}
 743
 744/**
 745 * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
 746 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 747 * @attachment: Virtual address attachment of the BO on accessing device
 748 *
 749 * The method performs following steps:
 750 *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
 751 *   - Free SG Table that is used to encapsulate DMA mapped memory of
 752 *          peer device's DOORBELL or MMIO memory
 753 *
 754 * This method is invoked in the following contexts:
 755 *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
 756 *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
 757 *
 758 * Return: void
 759 */
 760static void
 761kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
 762		       struct kfd_mem_attachment *attachment)
 763{
 764	struct ttm_operation_ctx ctx = {.interruptible = true};
 765	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 766	struct amdgpu_device *adev = attachment->adev;
 767	struct ttm_tt *ttm = bo->tbo.ttm;
 768	enum dma_data_direction dir;
 769
 770	if (unlikely(!ttm->sg)) {
 771		pr_debug("SG Table of BO is NULL");
 772		return;
 773	}
 774
 775	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 776	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 777
 778	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 779				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 780	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
 781			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 782	sg_free_table(ttm->sg);
 783	kfree(ttm->sg);
 784	ttm->sg = NULL;
 785	bo->tbo.sg = NULL;
 786}
 787
 788static void
 789kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
 790			    struct kfd_mem_attachment *attachment)
 791{
 792	switch (attachment->type) {
 793	case KFD_MEM_ATT_SHARED:
 794		break;
 795	case KFD_MEM_ATT_USERPTR:
 796		kfd_mem_dmaunmap_userptr(mem, attachment);
 797		break;
 798	case KFD_MEM_ATT_DMABUF:
 799		kfd_mem_dmaunmap_dmabuf(attachment);
 800		break;
 801	case KFD_MEM_ATT_SG:
 802		kfd_mem_dmaunmap_sg_bo(mem, attachment);
 803		break;
 804	default:
 805		WARN_ON_ONCE(1);
 806	}
 807}
 808
 809static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
 810{
 811	if (!mem->dmabuf) {
 812		struct amdgpu_device *bo_adev;
 813		struct dma_buf *dmabuf;
 814		int r, fd;
 815
 816		bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 817		r = drm_gem_prime_handle_to_fd(&bo_adev->ddev, bo_adev->kfd.client.file,
 818					       mem->gem_handle,
 819			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 820					       DRM_RDWR : 0, &fd);
 821		if (r)
 822			return r;
 823		dmabuf = dma_buf_get(fd);
 824		close_fd(fd);
 825		if (WARN_ON_ONCE(IS_ERR(dmabuf)))
 826			return PTR_ERR(dmabuf);
 827		mem->dmabuf = dmabuf;
 828	}
 829
 830	return 0;
 831}
 832
 833static int
 834kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
 835		      struct amdgpu_bo **bo)
 836{
 837	struct drm_gem_object *gobj;
 838	int ret;
 839
 840	ret = kfd_mem_export_dmabuf(mem);
 841	if (ret)
 842		return ret;
 843
 844	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
 845	if (IS_ERR(gobj))
 846		return PTR_ERR(gobj);
 847
 848	*bo = gem_to_amdgpu_bo(gobj);
 849	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
 850
 851	return 0;
 852}
 853
 854/* kfd_mem_attach - Add a BO to a VM
 855 *
 856 * Everything that needs to bo done only once when a BO is first added
 857 * to a VM. It can later be mapped and unmapped many times without
 858 * repeating these steps.
 859 *
 860 * 0. Create BO for DMA mapping, if needed
 861 * 1. Allocate and initialize BO VA entry data structure
 862 * 2. Add BO to the VM
 863 * 3. Determine ASIC-specific PTE flags
 864 * 4. Alloc page tables and directories if needed
 865 * 4a.  Validate new page tables and directories
 866 */
 867static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 868		struct amdgpu_vm *vm, bool is_aql)
 869{
 870	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 871	unsigned long bo_size = mem->bo->tbo.base.size;
 872	uint64_t va = mem->va;
 873	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
 874	struct amdgpu_bo *bo[2] = {NULL, NULL};
 875	struct amdgpu_bo_va *bo_va;
 876	bool same_hive = false;
 877	int i, ret;
 878
 879	if (!va) {
 880		pr_err("Invalid VA when adding BO to VM\n");
 881		return -EINVAL;
 882	}
 883
 884	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
 885	 *
 886	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
 887	 * In contrast the access path of VRAM BOs depens upon the type of
 888	 * link that connects the peer device. Access over PCIe is allowed
 889	 * if peer device has large BAR. In contrast, access over xGMI is
 890	 * allowed for both small and large BAR configurations of peer device
 891	 */
 892	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
 893	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
 894	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
 895	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 896		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
 897			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
 898		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
 899			return -EINVAL;
 900	}
 901
 902	for (i = 0; i <= is_aql; i++) {
 903		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
 904		if (unlikely(!attachment[i])) {
 905			ret = -ENOMEM;
 906			goto unwind;
 907		}
 908
 909		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 910			 va + bo_size, vm);
 911
 912		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
 913		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
 914		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
 915		    same_hive) {
 916			/* Mappings on the local GPU, or VRAM mappings in the
 917			 * local hive, or userptr, or GTT mapping can reuse dma map
 918			 * address space share the original BO
 919			 */
 920			attachment[i]->type = KFD_MEM_ATT_SHARED;
 921			bo[i] = mem->bo;
 922			drm_gem_object_get(&bo[i]->tbo.base);
 923		} else if (i > 0) {
 924			/* Multiple mappings on the same GPU share the BO */
 925			attachment[i]->type = KFD_MEM_ATT_SHARED;
 926			bo[i] = bo[0];
 927			drm_gem_object_get(&bo[i]->tbo.base);
 928		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
 929			/* Create an SG BO to DMA-map userptrs on other GPUs */
 930			attachment[i]->type = KFD_MEM_ATT_USERPTR;
 931			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 932			if (ret)
 933				goto unwind;
 934		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
 935		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
 936			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
 937				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
 938				  "Handing invalid SG BO in ATTACH request");
 939			attachment[i]->type = KFD_MEM_ATT_SG;
 940			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 941			if (ret)
 942				goto unwind;
 943		/* Enable acces to GTT and VRAM BOs of peer devices */
 944		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
 945			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
 946			attachment[i]->type = KFD_MEM_ATT_DMABUF;
 947			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
 948			if (ret)
 949				goto unwind;
 950			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
 951		} else {
 952			WARN_ONCE(true, "Handling invalid ATTACH request");
 953			ret = -EINVAL;
 954			goto unwind;
 955		}
 956
 957		/* Add BO to VM internal data structures */
 958		ret = amdgpu_bo_reserve(bo[i], false);
 959		if (ret) {
 960			pr_debug("Unable to reserve BO during memory attach");
 961			goto unwind;
 962		}
 963		bo_va = amdgpu_vm_bo_find(vm, bo[i]);
 964		if (!bo_va)
 965			bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
 966		else
 967			++bo_va->ref_count;
 968		attachment[i]->bo_va = bo_va;
 969		amdgpu_bo_unreserve(bo[i]);
 970		if (unlikely(!attachment[i]->bo_va)) {
 971			ret = -ENOMEM;
 972			pr_err("Failed to add BO object to VM. ret == %d\n",
 973			       ret);
 974			goto unwind;
 975		}
 976		attachment[i]->va = va;
 977		attachment[i]->pte_flags = get_pte_flags(adev, mem);
 978		attachment[i]->adev = adev;
 979		list_add(&attachment[i]->list, &mem->attachments);
 980
 981		va += bo_size;
 982	}
 983
 984	return 0;
 985
 986unwind:
 987	for (; i >= 0; i--) {
 988		if (!attachment[i])
 989			continue;
 990		if (attachment[i]->bo_va) {
 991			amdgpu_bo_reserve(bo[i], true);
 992			if (--attachment[i]->bo_va->ref_count == 0)
 993				amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
 994			amdgpu_bo_unreserve(bo[i]);
 995			list_del(&attachment[i]->list);
 996		}
 997		if (bo[i])
 998			drm_gem_object_put(&bo[i]->tbo.base);
 999		kfree(attachment[i]);
1000	}
1001	return ret;
1002}
1003
1004static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
1005{
1006	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
1007
1008	pr_debug("\t remove VA 0x%llx in entry %p\n",
1009			attachment->va, attachment);
1010	if (--attachment->bo_va->ref_count == 0)
1011		amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
1012	drm_gem_object_put(&bo->tbo.base);
1013	list_del(&attachment->list);
1014	kfree(attachment);
1015}
1016
1017static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
1018				struct amdkfd_process_info *process_info,
1019				bool userptr)
1020{
1021	mutex_lock(&process_info->lock);
1022	if (userptr)
1023		list_add_tail(&mem->validate_list,
1024			      &process_info->userptr_valid_list);
1025	else
1026		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
1027	mutex_unlock(&process_info->lock);
1028}
1029
1030static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
1031		struct amdkfd_process_info *process_info)
1032{
1033	mutex_lock(&process_info->lock);
1034	list_del(&mem->validate_list);
1035	mutex_unlock(&process_info->lock);
1036}
1037
1038/* Initializes user pages. It registers the MMU notifier and validates
1039 * the userptr BO in the GTT domain.
1040 *
1041 * The BO must already be on the userptr_valid_list. Otherwise an
1042 * eviction and restore may happen that leaves the new BO unmapped
1043 * with the user mode queues running.
1044 *
1045 * Takes the process_info->lock to protect against concurrent restore
1046 * workers.
1047 *
1048 * Returns 0 for success, negative errno for errors.
1049 */
1050static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
1051			   bool criu_resume)
1052{
1053	struct amdkfd_process_info *process_info = mem->process_info;
1054	struct amdgpu_bo *bo = mem->bo;
1055	struct ttm_operation_ctx ctx = { true, false };
1056	struct hmm_range *range;
1057	int ret = 0;
1058
1059	mutex_lock(&process_info->lock);
1060
1061	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1062	if (ret) {
1063		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1064		goto out;
1065	}
1066
1067	ret = amdgpu_hmm_register(bo, user_addr);
1068	if (ret) {
1069		pr_err("%s: Failed to register MMU notifier: %d\n",
1070		       __func__, ret);
1071		goto out;
1072	}
1073
1074	if (criu_resume) {
1075		/*
1076		 * During a CRIU restore operation, the userptr buffer objects
1077		 * will be validated in the restore_userptr_work worker at a
1078		 * later stage when it is scheduled by another ioctl called by
1079		 * CRIU master process for the target pid for restore.
1080		 */
1081		mutex_lock(&process_info->notifier_lock);
1082		mem->invalid++;
1083		mutex_unlock(&process_info->notifier_lock);
1084		mutex_unlock(&process_info->lock);
1085		return 0;
1086	}
1087
1088	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1089	if (ret) {
1090		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1091		goto unregister_out;
1092	}
1093
1094	ret = amdgpu_bo_reserve(bo, true);
1095	if (ret) {
1096		pr_err("%s: Failed to reserve BO\n", __func__);
1097		goto release_out;
1098	}
1099	amdgpu_bo_placement_from_domain(bo, mem->domain);
1100	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1101	if (ret)
1102		pr_err("%s: failed to validate BO\n", __func__);
1103	amdgpu_bo_unreserve(bo);
1104
1105release_out:
1106	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1107unregister_out:
1108	if (ret)
1109		amdgpu_hmm_unregister(bo);
1110out:
1111	mutex_unlock(&process_info->lock);
1112	return ret;
1113}
1114
1115/* Reserving a BO and its page table BOs must happen atomically to
1116 * avoid deadlocks. Some operations update multiple VMs at once. Track
1117 * all the reservation info in a context structure. Optionally a sync
1118 * object can track VM updates.
1119 */
1120struct bo_vm_reservation_context {
1121	/* DRM execution context for the reservation */
1122	struct drm_exec exec;
1123	/* Number of VMs reserved */
1124	unsigned int n_vms;
1125	/* Pointer to sync object */
1126	struct amdgpu_sync *sync;
1127};
1128
1129enum bo_vm_match {
1130	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1131	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1132	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1133};
1134
1135/**
1136 * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1137 * @mem: KFD BO structure.
1138 * @vm: the VM to reserve.
1139 * @ctx: the struct that will be used in unreserve_bo_and_vms().
1140 */
1141static int reserve_bo_and_vm(struct kgd_mem *mem,
1142			      struct amdgpu_vm *vm,
1143			      struct bo_vm_reservation_context *ctx)
1144{
1145	struct amdgpu_bo *bo = mem->bo;
1146	int ret;
1147
1148	WARN_ON(!vm);
1149
1150	ctx->n_vms = 1;
1151	ctx->sync = &mem->sync;
1152	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1153	drm_exec_until_all_locked(&ctx->exec) {
1154		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1155		drm_exec_retry_on_contention(&ctx->exec);
1156		if (unlikely(ret))
1157			goto error;
1158
1159		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1160		drm_exec_retry_on_contention(&ctx->exec);
1161		if (unlikely(ret))
1162			goto error;
1163	}
1164	return 0;
1165
1166error:
1167	pr_err("Failed to reserve buffers in ttm.\n");
1168	drm_exec_fini(&ctx->exec);
1169	return ret;
1170}
1171
1172/**
1173 * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1174 * @mem: KFD BO structure.
1175 * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1176 * is used. Otherwise, a single VM associated with the BO.
1177 * @map_type: the mapping status that will be used to filter the VMs.
1178 * @ctx: the struct that will be used in unreserve_bo_and_vms().
1179 *
1180 * Returns 0 for success, negative for failure.
1181 */
1182static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1183				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1184				struct bo_vm_reservation_context *ctx)
1185{
1186	struct kfd_mem_attachment *entry;
1187	struct amdgpu_bo *bo = mem->bo;
1188	int ret;
1189
1190	ctx->sync = &mem->sync;
1191	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1192	drm_exec_until_all_locked(&ctx->exec) {
1193		ctx->n_vms = 0;
1194		list_for_each_entry(entry, &mem->attachments, list) {
1195			if ((vm && vm != entry->bo_va->base.vm) ||
1196				(entry->is_mapped != map_type
1197				&& map_type != BO_VM_ALL))
1198				continue;
1199
1200			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1201						&ctx->exec, 2);
1202			drm_exec_retry_on_contention(&ctx->exec);
1203			if (unlikely(ret))
1204				goto error;
1205			++ctx->n_vms;
1206		}
1207
1208		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1209		drm_exec_retry_on_contention(&ctx->exec);
1210		if (unlikely(ret))
1211			goto error;
1212	}
1213	return 0;
1214
1215error:
1216	pr_err("Failed to reserve buffers in ttm.\n");
1217	drm_exec_fini(&ctx->exec);
1218	return ret;
1219}
1220
1221/**
1222 * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1223 * @ctx: Reservation context to unreserve
1224 * @wait: Optionally wait for a sync object representing pending VM updates
1225 * @intr: Whether the wait is interruptible
1226 *
1227 * Also frees any resources allocated in
1228 * reserve_bo_and_(cond_)vm(s). Returns the status from
1229 * amdgpu_sync_wait.
1230 */
1231static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1232				 bool wait, bool intr)
1233{
1234	int ret = 0;
1235
1236	if (wait)
1237		ret = amdgpu_sync_wait(ctx->sync, intr);
1238
1239	drm_exec_fini(&ctx->exec);
1240	ctx->sync = NULL;
1241	return ret;
1242}
1243
1244static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1245				struct kfd_mem_attachment *entry,
1246				struct amdgpu_sync *sync)
1247{
1248	struct amdgpu_bo_va *bo_va = entry->bo_va;
1249	struct amdgpu_device *adev = entry->adev;
1250	struct amdgpu_vm *vm = bo_va->base.vm;
1251
1252	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1253
1254	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1255
1256	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1257}
1258
1259static int update_gpuvm_pte(struct kgd_mem *mem,
1260			    struct kfd_mem_attachment *entry,
1261			    struct amdgpu_sync *sync)
1262{
1263	struct amdgpu_bo_va *bo_va = entry->bo_va;
1264	struct amdgpu_device *adev = entry->adev;
1265	int ret;
1266
1267	ret = kfd_mem_dmamap_attachment(mem, entry);
1268	if (ret)
1269		return ret;
1270
1271	/* Update the page tables  */
1272	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1273	if (ret) {
1274		pr_err("amdgpu_vm_bo_update failed\n");
1275		return ret;
1276	}
1277
1278	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1279}
1280
1281static int map_bo_to_gpuvm(struct kgd_mem *mem,
1282			   struct kfd_mem_attachment *entry,
1283			   struct amdgpu_sync *sync,
1284			   bool no_update_pte)
1285{
1286	int ret;
1287
1288	/* Set virtual address for the allocation */
1289	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1290			       amdgpu_bo_size(entry->bo_va->base.bo),
1291			       entry->pte_flags);
1292	if (ret) {
1293		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1294				entry->va, ret);
1295		return ret;
1296	}
1297
1298	if (no_update_pte)
1299		return 0;
1300
1301	ret = update_gpuvm_pte(mem, entry, sync);
1302	if (ret) {
1303		pr_err("update_gpuvm_pte() failed\n");
1304		goto update_gpuvm_pte_failed;
1305	}
1306
1307	return 0;
1308
1309update_gpuvm_pte_failed:
1310	unmap_bo_from_gpuvm(mem, entry, sync);
1311	kfd_mem_dmaunmap_attachment(mem, entry);
1312	return ret;
1313}
1314
1315static int process_validate_vms(struct amdkfd_process_info *process_info,
1316				struct ww_acquire_ctx *ticket)
1317{
1318	struct amdgpu_vm *peer_vm;
1319	int ret;
1320
1321	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1322			    vm_list_node) {
1323		ret = vm_validate_pt_pd_bos(peer_vm, ticket);
1324		if (ret)
1325			return ret;
1326	}
1327
1328	return 0;
1329}
1330
1331static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1332				 struct amdgpu_sync *sync)
1333{
1334	struct amdgpu_vm *peer_vm;
1335	int ret;
1336
1337	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1338			    vm_list_node) {
1339		struct amdgpu_bo *pd = peer_vm->root.bo;
1340
1341		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1342				       AMDGPU_SYNC_NE_OWNER,
1343				       AMDGPU_FENCE_OWNER_KFD);
1344		if (ret)
1345			return ret;
1346	}
1347
1348	return 0;
1349}
1350
1351static int process_update_pds(struct amdkfd_process_info *process_info,
1352			      struct amdgpu_sync *sync)
1353{
1354	struct amdgpu_vm *peer_vm;
1355	int ret;
1356
1357	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1358			    vm_list_node) {
1359		ret = vm_update_pds(peer_vm, sync);
1360		if (ret)
1361			return ret;
1362	}
1363
1364	return 0;
1365}
1366
1367static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1368		       struct dma_fence **ef)
1369{
1370	struct amdkfd_process_info *info = NULL;
1371	int ret;
1372
1373	if (!*process_info) {
1374		info = kzalloc(sizeof(*info), GFP_KERNEL);
1375		if (!info)
1376			return -ENOMEM;
1377
1378		mutex_init(&info->lock);
1379		mutex_init(&info->notifier_lock);
1380		INIT_LIST_HEAD(&info->vm_list_head);
1381		INIT_LIST_HEAD(&info->kfd_bo_list);
1382		INIT_LIST_HEAD(&info->userptr_valid_list);
1383		INIT_LIST_HEAD(&info->userptr_inval_list);
1384
1385		info->eviction_fence =
1386			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1387						   current->mm,
1388						   NULL);
1389		if (!info->eviction_fence) {
1390			pr_err("Failed to create eviction fence\n");
1391			ret = -ENOMEM;
1392			goto create_evict_fence_fail;
1393		}
1394
1395		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1396		INIT_DELAYED_WORK(&info->restore_userptr_work,
1397				  amdgpu_amdkfd_restore_userptr_worker);
1398
1399		*process_info = info;
1400	}
1401
1402	vm->process_info = *process_info;
1403
1404	/* Validate page directory and attach eviction fence */
1405	ret = amdgpu_bo_reserve(vm->root.bo, true);
1406	if (ret)
1407		goto reserve_pd_fail;
1408	ret = vm_validate_pt_pd_bos(vm, NULL);
1409	if (ret) {
1410		pr_err("validate_pt_pd_bos() failed\n");
1411		goto validate_pd_fail;
1412	}
1413	ret = amdgpu_bo_sync_wait(vm->root.bo,
1414				  AMDGPU_FENCE_OWNER_KFD, false);
1415	if (ret)
1416		goto wait_pd_fail;
1417	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1418	if (ret)
1419		goto reserve_shared_fail;
1420	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1421			   &vm->process_info->eviction_fence->base,
1422			   DMA_RESV_USAGE_BOOKKEEP);
1423	amdgpu_bo_unreserve(vm->root.bo);
1424
1425	/* Update process info */
1426	mutex_lock(&vm->process_info->lock);
1427	list_add_tail(&vm->vm_list_node,
1428			&(vm->process_info->vm_list_head));
1429	vm->process_info->n_vms++;
1430
1431	*ef = dma_fence_get(&vm->process_info->eviction_fence->base);
1432	mutex_unlock(&vm->process_info->lock);
1433
1434	return 0;
1435
1436reserve_shared_fail:
1437wait_pd_fail:
1438validate_pd_fail:
1439	amdgpu_bo_unreserve(vm->root.bo);
1440reserve_pd_fail:
1441	vm->process_info = NULL;
1442	if (info) {
1443		dma_fence_put(&info->eviction_fence->base);
1444		*process_info = NULL;
1445		put_pid(info->pid);
1446create_evict_fence_fail:
1447		mutex_destroy(&info->lock);
1448		mutex_destroy(&info->notifier_lock);
1449		kfree(info);
1450	}
1451	return ret;
1452}
1453
1454/**
1455 * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1456 * @bo: Handle of buffer object being pinned
1457 * @domain: Domain into which BO should be pinned
1458 *
1459 *   - USERPTR BOs are UNPINNABLE and will return error
1460 *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1461 *     PIN count incremented. It is valid to PIN a BO multiple times
1462 *
1463 * Return: ZERO if successful in pinning, Non-Zero in case of error.
1464 */
1465static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1466{
1467	int ret = 0;
1468
1469	ret = amdgpu_bo_reserve(bo, false);
1470	if (unlikely(ret))
1471		return ret;
1472
1473	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1474	if (ret)
1475		pr_err("Error in Pinning BO to domain: %d\n", domain);
1476
1477	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1478	amdgpu_bo_unreserve(bo);
1479
1480	return ret;
1481}
1482
1483/**
1484 * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1485 * @bo: Handle of buffer object being unpinned
1486 *
1487 *   - Is a illegal request for USERPTR BOs and is ignored
1488 *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1489 *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1490 */
1491static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1492{
1493	int ret = 0;
1494
1495	ret = amdgpu_bo_reserve(bo, false);
1496	if (unlikely(ret))
1497		return;
1498
1499	amdgpu_bo_unpin(bo);
1500	amdgpu_bo_unreserve(bo);
1501}
1502
1503int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1504				     struct amdgpu_vm *avm, u32 pasid)
1505
1506{
1507	int ret;
1508
1509	/* Free the original amdgpu allocated pasid,
1510	 * will be replaced with kfd allocated pasid.
1511	 */
1512	if (avm->pasid) {
1513		amdgpu_pasid_free(avm->pasid);
1514		amdgpu_vm_set_pasid(adev, avm, 0);
1515	}
1516
1517	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1518	if (ret)
1519		return ret;
1520
1521	return 0;
1522}
1523
1524int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1525					   struct amdgpu_vm *avm,
1526					   void **process_info,
1527					   struct dma_fence **ef)
1528{
1529	int ret;
1530
1531	/* Already a compute VM? */
1532	if (avm->process_info)
1533		return -EINVAL;
1534
1535	/* Convert VM into a compute VM */
1536	ret = amdgpu_vm_make_compute(adev, avm);
1537	if (ret)
1538		return ret;
1539
1540	/* Initialize KFD part of the VM and process info */
1541	ret = init_kfd_vm(avm, process_info, ef);
1542	if (ret)
1543		return ret;
1544
1545	amdgpu_vm_set_task_info(avm);
1546
1547	return 0;
1548}
1549
1550void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1551				    struct amdgpu_vm *vm)
1552{
1553	struct amdkfd_process_info *process_info = vm->process_info;
1554
1555	if (!process_info)
1556		return;
1557
1558	/* Update process info */
1559	mutex_lock(&process_info->lock);
1560	process_info->n_vms--;
1561	list_del(&vm->vm_list_node);
1562	mutex_unlock(&process_info->lock);
1563
1564	vm->process_info = NULL;
1565
1566	/* Release per-process resources when last compute VM is destroyed */
1567	if (!process_info->n_vms) {
1568		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1569		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1570		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1571
1572		dma_fence_put(&process_info->eviction_fence->base);
1573		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1574		put_pid(process_info->pid);
1575		mutex_destroy(&process_info->lock);
1576		mutex_destroy(&process_info->notifier_lock);
1577		kfree(process_info);
1578	}
1579}
1580
1581void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1582					    void *drm_priv)
1583{
1584	struct amdgpu_vm *avm;
1585
1586	if (WARN_ON(!adev || !drm_priv))
1587		return;
1588
1589	avm = drm_priv_to_vm(drm_priv);
1590
1591	pr_debug("Releasing process vm %p\n", avm);
1592
1593	/* The original pasid of amdgpu vm has already been
1594	 * released during making a amdgpu vm to a compute vm
1595	 * The current pasid is managed by kfd and will be
1596	 * released on kfd process destroy. Set amdgpu pasid
1597	 * to 0 to avoid duplicate release.
1598	 */
1599	amdgpu_vm_release_compute(adev, avm);
1600}
1601
1602uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1603{
1604	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1605	struct amdgpu_bo *pd = avm->root.bo;
1606	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1607
1608	if (adev->asic_type < CHIP_VEGA10)
1609		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1610	return avm->pd_phys_addr;
1611}
1612
1613void amdgpu_amdkfd_block_mmu_notifications(void *p)
1614{
1615	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1616
1617	mutex_lock(&pinfo->lock);
1618	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1619	mutex_unlock(&pinfo->lock);
1620}
1621
1622int amdgpu_amdkfd_criu_resume(void *p)
1623{
1624	int ret = 0;
1625	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1626
1627	mutex_lock(&pinfo->lock);
1628	pr_debug("scheduling work\n");
1629	mutex_lock(&pinfo->notifier_lock);
1630	pinfo->evicted_bos++;
1631	mutex_unlock(&pinfo->notifier_lock);
1632	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1633		ret = -EINVAL;
1634		goto out_unlock;
1635	}
1636	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1637	queue_delayed_work(system_freezable_wq,
1638			   &pinfo->restore_userptr_work, 0);
1639
1640out_unlock:
1641	mutex_unlock(&pinfo->lock);
1642	return ret;
1643}
1644
1645size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1646					  uint8_t xcp_id)
1647{
1648	uint64_t reserved_for_pt =
1649		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1650	ssize_t available;
1651	uint64_t vram_available, system_mem_available, ttm_mem_available;
1652
1653	spin_lock(&kfd_mem_limit.mem_limit_lock);
1654	vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1655		- adev->kfd.vram_used_aligned[xcp_id]
1656		- atomic64_read(&adev->vram_pin_size)
1657		- reserved_for_pt;
1658
1659	if (adev->gmc.is_app_apu) {
1660		system_mem_available = no_system_mem_limit ?
1661					kfd_mem_limit.max_system_mem_limit :
1662					kfd_mem_limit.max_system_mem_limit -
1663					kfd_mem_limit.system_mem_used;
1664
1665		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1666				kfd_mem_limit.ttm_mem_used;
1667
1668		available = min3(system_mem_available, ttm_mem_available,
1669				 vram_available);
1670		available = ALIGN_DOWN(available, PAGE_SIZE);
1671	} else {
1672		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1673	}
1674
1675	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1676
1677	if (available < 0)
1678		available = 0;
1679
1680	return available;
1681}
1682
1683int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1684		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1685		void *drm_priv, struct kgd_mem **mem,
1686		uint64_t *offset, uint32_t flags, bool criu_resume)
1687{
1688	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1689	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1690	enum ttm_bo_type bo_type = ttm_bo_type_device;
1691	struct sg_table *sg = NULL;
1692	uint64_t user_addr = 0;
1693	struct amdgpu_bo *bo;
1694	struct drm_gem_object *gobj = NULL;
1695	u32 domain, alloc_domain;
1696	uint64_t aligned_size;
1697	int8_t xcp_id = -1;
1698	u64 alloc_flags;
1699	int ret;
1700
1701	/*
1702	 * Check on which domain to allocate BO
1703	 */
1704	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1705		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1706
1707		if (adev->gmc.is_app_apu) {
1708			domain = AMDGPU_GEM_DOMAIN_GTT;
1709			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1710			alloc_flags = 0;
1711		} else {
1712			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1713			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1714			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1715		}
1716		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1717					0 : fpriv->xcp_id;
1718	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1719		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1720		alloc_flags = 0;
1721	} else {
1722		domain = AMDGPU_GEM_DOMAIN_GTT;
1723		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1724		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1725
1726		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1727			if (!offset || !*offset)
1728				return -EINVAL;
1729			user_addr = untagged_addr(*offset);
1730		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1731				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1732			bo_type = ttm_bo_type_sg;
1733			if (size > UINT_MAX)
1734				return -EINVAL;
1735			sg = create_sg_table(*offset, size);
1736			if (!sg)
1737				return -ENOMEM;
1738		} else {
1739			return -EINVAL;
1740		}
1741	}
1742
1743	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1744		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1745	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1746		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1747	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1748		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1749
1750	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1751	if (!*mem) {
1752		ret = -ENOMEM;
1753		goto err;
1754	}
1755	INIT_LIST_HEAD(&(*mem)->attachments);
1756	mutex_init(&(*mem)->lock);
1757	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1758
1759	/* Workaround for AQL queue wraparound bug. Map the same
1760	 * memory twice. That means we only actually allocate half
1761	 * the memory.
1762	 */
1763	if ((*mem)->aql_queue)
1764		size >>= 1;
1765	aligned_size = PAGE_ALIGN(size);
1766
1767	(*mem)->alloc_flags = flags;
1768
1769	amdgpu_sync_create(&(*mem)->sync);
1770
1771	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1772					      xcp_id);
1773	if (ret) {
1774		pr_debug("Insufficient memory\n");
1775		goto err_reserve_limit;
1776	}
1777
1778	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1779		 va, (*mem)->aql_queue ? size << 1 : size,
1780		 domain_string(alloc_domain), xcp_id);
1781
1782	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1783				       bo_type, NULL, &gobj, xcp_id + 1);
1784	if (ret) {
1785		pr_debug("Failed to create BO on domain %s. ret %d\n",
1786			 domain_string(alloc_domain), ret);
1787		goto err_bo_create;
1788	}
1789	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1790	if (ret) {
1791		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1792		goto err_node_allow;
1793	}
1794	ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
1795	if (ret)
1796		goto err_gem_handle_create;
1797	bo = gem_to_amdgpu_bo(gobj);
1798	if (bo_type == ttm_bo_type_sg) {
1799		bo->tbo.sg = sg;
1800		bo->tbo.ttm->sg = sg;
1801	}
1802	bo->kfd_bo = *mem;
1803	(*mem)->bo = bo;
1804	if (user_addr)
1805		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1806
1807	(*mem)->va = va;
1808	(*mem)->domain = domain;
1809	(*mem)->mapped_to_gpu_memory = 0;
1810	(*mem)->process_info = avm->process_info;
1811
1812	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1813
1814	if (user_addr) {
1815		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1816		ret = init_user_pages(*mem, user_addr, criu_resume);
1817		if (ret)
1818			goto allocate_init_user_pages_failed;
1819	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1820				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1821		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1822		if (ret) {
1823			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1824			goto err_pin_bo;
1825		}
1826		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1827		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1828	} else {
1829		mutex_lock(&avm->process_info->lock);
1830		if (avm->process_info->eviction_fence &&
1831		    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
1832			ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
1833				&avm->process_info->eviction_fence->base);
1834		mutex_unlock(&avm->process_info->lock);
1835		if (ret)
1836			goto err_validate_bo;
1837	}
1838
1839	if (offset)
1840		*offset = amdgpu_bo_mmap_offset(bo);
1841
1842	return 0;
1843
1844allocate_init_user_pages_failed:
1845err_pin_bo:
1846err_validate_bo:
1847	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1848	drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
1849err_gem_handle_create:
1850	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1851err_node_allow:
1852	/* Don't unreserve system mem limit twice */
1853	goto err_reserve_limit;
1854err_bo_create:
1855	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1856err_reserve_limit:
1857	amdgpu_sync_free(&(*mem)->sync);
1858	mutex_destroy(&(*mem)->lock);
1859	if (gobj)
1860		drm_gem_object_put(gobj);
1861	else
1862		kfree(*mem);
1863err:
1864	if (sg) {
1865		sg_free_table(sg);
1866		kfree(sg);
1867	}
1868	return ret;
1869}
1870
1871int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1872		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1873		uint64_t *size)
1874{
1875	struct amdkfd_process_info *process_info = mem->process_info;
1876	unsigned long bo_size = mem->bo->tbo.base.size;
1877	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1878	struct kfd_mem_attachment *entry, *tmp;
1879	struct bo_vm_reservation_context ctx;
1880	unsigned int mapped_to_gpu_memory;
1881	int ret;
1882	bool is_imported = false;
1883
1884	mutex_lock(&mem->lock);
1885
1886	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1887	if (mem->alloc_flags &
1888	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1889	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1890		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1891	}
1892
1893	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1894	is_imported = mem->is_imported;
1895	mutex_unlock(&mem->lock);
1896	/* lock is not needed after this, since mem is unused and will
1897	 * be freed anyway
1898	 */
1899
1900	if (mapped_to_gpu_memory > 0) {
1901		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1902				mem->va, bo_size);
1903		return -EBUSY;
1904	}
1905
1906	/* Make sure restore workers don't access the BO any more */
1907	mutex_lock(&process_info->lock);
1908	list_del(&mem->validate_list);
1909	mutex_unlock(&process_info->lock);
1910
1911	/* Cleanup user pages and MMU notifiers */
1912	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1913		amdgpu_hmm_unregister(mem->bo);
1914		mutex_lock(&process_info->notifier_lock);
1915		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1916		mutex_unlock(&process_info->notifier_lock);
1917	}
1918
1919	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1920	if (unlikely(ret))
1921		return ret;
1922
1923	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1924					process_info->eviction_fence);
1925	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1926		mem->va + bo_size * (1 + mem->aql_queue));
1927
1928	/* Remove from VM internal data structures */
1929	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
1930		kfd_mem_dmaunmap_attachment(mem, entry);
1931		kfd_mem_detach(entry);
1932	}
1933
1934	ret = unreserve_bo_and_vms(&ctx, false, false);
1935
1936	/* Free the sync object */
1937	amdgpu_sync_free(&mem->sync);
1938
1939	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1940	 * remap BO. We need to free it.
1941	 */
1942	if (mem->bo->tbo.sg) {
1943		sg_free_table(mem->bo->tbo.sg);
1944		kfree(mem->bo->tbo.sg);
1945	}
1946
1947	/* Update the size of the BO being freed if it was allocated from
1948	 * VRAM and is not imported. For APP APU VRAM allocations are done
1949	 * in GTT domain
1950	 */
1951	if (size) {
1952		if (!is_imported &&
1953		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1954		   (adev->gmc.is_app_apu &&
1955		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1956			*size = bo_size;
1957		else
1958			*size = 0;
1959	}
1960
1961	/* Free the BO*/
1962	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1963	drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
1964	if (mem->dmabuf) {
1965		dma_buf_put(mem->dmabuf);
1966		mem->dmabuf = NULL;
1967	}
1968	mutex_destroy(&mem->lock);
1969
1970	/* If this releases the last reference, it will end up calling
1971	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1972	 * this needs to be the last call here.
1973	 */
1974	drm_gem_object_put(&mem->bo->tbo.base);
1975
1976	/*
1977	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1978	 * explicitly free it here.
1979	 */
1980	if (!use_release_notifier)
1981		kfree(mem);
1982
1983	return ret;
1984}
1985
1986int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1987		struct amdgpu_device *adev, struct kgd_mem *mem,
1988		void *drm_priv)
1989{
1990	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1991	int ret;
1992	struct amdgpu_bo *bo;
1993	uint32_t domain;
1994	struct kfd_mem_attachment *entry;
1995	struct bo_vm_reservation_context ctx;
1996	unsigned long bo_size;
1997	bool is_invalid_userptr = false;
1998
1999	bo = mem->bo;
2000	if (!bo) {
2001		pr_err("Invalid BO when mapping memory to GPU\n");
2002		return -EINVAL;
2003	}
2004
2005	/* Make sure restore is not running concurrently. Since we
2006	 * don't map invalid userptr BOs, we rely on the next restore
2007	 * worker to do the mapping
2008	 */
2009	mutex_lock(&mem->process_info->lock);
2010
2011	/* Lock notifier lock. If we find an invalid userptr BO, we can be
2012	 * sure that the MMU notifier is no longer running
2013	 * concurrently and the queues are actually stopped
2014	 */
2015	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2016		mutex_lock(&mem->process_info->notifier_lock);
2017		is_invalid_userptr = !!mem->invalid;
2018		mutex_unlock(&mem->process_info->notifier_lock);
2019	}
2020
2021	mutex_lock(&mem->lock);
2022
2023	domain = mem->domain;
2024	bo_size = bo->tbo.base.size;
2025
2026	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
2027			mem->va,
2028			mem->va + bo_size * (1 + mem->aql_queue),
2029			avm, domain_string(domain));
2030
2031	if (!kfd_mem_is_attached(avm, mem)) {
2032		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
2033		if (ret)
2034			goto out;
2035	}
2036
2037	ret = reserve_bo_and_vm(mem, avm, &ctx);
2038	if (unlikely(ret))
2039		goto out;
2040
2041	/* Userptr can be marked as "not invalid", but not actually be
2042	 * validated yet (still in the system domain). In that case
2043	 * the queues are still stopped and we can leave mapping for
2044	 * the next restore worker
2045	 */
2046	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
2047	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
2048		is_invalid_userptr = true;
2049
2050	ret = vm_validate_pt_pd_bos(avm, NULL);
2051	if (unlikely(ret))
2052		goto out_unreserve;
2053
2054	list_for_each_entry(entry, &mem->attachments, list) {
2055		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2056			continue;
2057
2058		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2059			 entry->va, entry->va + bo_size, entry);
2060
2061		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2062				      is_invalid_userptr);
2063		if (ret) {
2064			pr_err("Failed to map bo to gpuvm\n");
2065			goto out_unreserve;
2066		}
2067
2068		ret = vm_update_pds(avm, ctx.sync);
2069		if (ret) {
2070			pr_err("Failed to update page directories\n");
2071			goto out_unreserve;
2072		}
2073
2074		entry->is_mapped = true;
2075		mem->mapped_to_gpu_memory++;
2076		pr_debug("\t INC mapping count %d\n",
2077			 mem->mapped_to_gpu_memory);
2078	}
2079
2080	ret = unreserve_bo_and_vms(&ctx, false, false);
2081
2082	goto out;
2083
2084out_unreserve:
2085	unreserve_bo_and_vms(&ctx, false, false);
2086out:
2087	mutex_unlock(&mem->process_info->lock);
2088	mutex_unlock(&mem->lock);
2089	return ret;
2090}
2091
2092int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv)
2093{
2094	struct kfd_mem_attachment *entry;
2095	struct amdgpu_vm *vm;
2096	int ret;
2097
2098	vm = drm_priv_to_vm(drm_priv);
2099
2100	mutex_lock(&mem->lock);
2101
2102	ret = amdgpu_bo_reserve(mem->bo, true);
2103	if (ret)
2104		goto out;
2105
2106	list_for_each_entry(entry, &mem->attachments, list) {
2107		if (entry->bo_va->base.vm != vm)
2108			continue;
2109		if (entry->bo_va->base.bo->tbo.ttm &&
2110		    !entry->bo_va->base.bo->tbo.ttm->sg)
2111			continue;
2112
2113		kfd_mem_dmaunmap_attachment(mem, entry);
2114	}
2115
2116	amdgpu_bo_unreserve(mem->bo);
2117out:
2118	mutex_unlock(&mem->lock);
2119
2120	return ret;
2121}
2122
2123int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2124		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2125{
2126	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2127	unsigned long bo_size = mem->bo->tbo.base.size;
2128	struct kfd_mem_attachment *entry;
2129	struct bo_vm_reservation_context ctx;
2130	int ret;
2131
2132	mutex_lock(&mem->lock);
2133
2134	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2135	if (unlikely(ret))
2136		goto out;
2137	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2138	if (ctx.n_vms == 0) {
2139		ret = -EINVAL;
2140		goto unreserve_out;
2141	}
2142
2143	ret = vm_validate_pt_pd_bos(avm, NULL);
2144	if (unlikely(ret))
2145		goto unreserve_out;
2146
2147	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2148		mem->va,
2149		mem->va + bo_size * (1 + mem->aql_queue),
2150		avm);
2151
2152	list_for_each_entry(entry, &mem->attachments, list) {
2153		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2154			continue;
2155
2156		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2157			 entry->va, entry->va + bo_size, entry);
2158
2159		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2160		entry->is_mapped = false;
2161
2162		mem->mapped_to_gpu_memory--;
2163		pr_debug("\t DEC mapping count %d\n",
2164			 mem->mapped_to_gpu_memory);
2165	}
2166
2167unreserve_out:
2168	unreserve_bo_and_vms(&ctx, false, false);
2169out:
2170	mutex_unlock(&mem->lock);
2171	return ret;
2172}
2173
2174int amdgpu_amdkfd_gpuvm_sync_memory(
2175		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2176{
2177	struct amdgpu_sync sync;
2178	int ret;
2179
2180	amdgpu_sync_create(&sync);
2181
2182	mutex_lock(&mem->lock);
2183	amdgpu_sync_clone(&mem->sync, &sync);
2184	mutex_unlock(&mem->lock);
2185
2186	ret = amdgpu_sync_wait(&sync, intr);
2187	amdgpu_sync_free(&sync);
2188	return ret;
2189}
2190
2191/**
2192 * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
 
2193 * @bo: Buffer object to be mapped
2194 *
2195 * Before return, bo reference count is incremented. To release the reference and unpin/
2196 * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2197 */
2198int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo)
2199{
2200	int ret;
2201
2202	ret = amdgpu_bo_reserve(bo, true);
2203	if (ret) {
2204		pr_err("Failed to reserve bo. ret %d\n", ret);
2205		goto err_reserve_bo_failed;
2206	}
2207
2208	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2209	if (ret) {
2210		pr_err("Failed to pin bo. ret %d\n", ret);
2211		goto err_pin_bo_failed;
2212	}
2213
2214	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2215	if (ret) {
2216		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2217		goto err_map_bo_gart_failed;
2218	}
2219
2220	amdgpu_amdkfd_remove_eviction_fence(
2221		bo, bo->vm_bo->vm->process_info->eviction_fence);
2222
2223	amdgpu_bo_unreserve(bo);
2224
2225	bo = amdgpu_bo_ref(bo);
2226
2227	return 0;
2228
2229err_map_bo_gart_failed:
2230	amdgpu_bo_unpin(bo);
2231err_pin_bo_failed:
2232	amdgpu_bo_unreserve(bo);
2233err_reserve_bo_failed:
2234
2235	return ret;
2236}
2237
2238/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2239 *
2240 * @mem: Buffer object to be mapped for CPU access
2241 * @kptr[out]: pointer in kernel CPU address space
2242 * @size[out]: size of the buffer
2243 *
2244 * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2245 * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2246 * validate_list, so the GPU mapping can be restored after a page table was
2247 * evicted.
2248 *
2249 * Return: 0 on success, error code on failure
2250 */
2251int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2252					     void **kptr, uint64_t *size)
2253{
2254	int ret;
2255	struct amdgpu_bo *bo = mem->bo;
2256
2257	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2258		pr_err("userptr can't be mapped to kernel\n");
2259		return -EINVAL;
2260	}
2261
2262	mutex_lock(&mem->process_info->lock);
2263
2264	ret = amdgpu_bo_reserve(bo, true);
2265	if (ret) {
2266		pr_err("Failed to reserve bo. ret %d\n", ret);
2267		goto bo_reserve_failed;
2268	}
2269
2270	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2271	if (ret) {
2272		pr_err("Failed to pin bo. ret %d\n", ret);
2273		goto pin_failed;
2274	}
2275
2276	ret = amdgpu_bo_kmap(bo, kptr);
2277	if (ret) {
2278		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2279		goto kmap_failed;
2280	}
2281
2282	amdgpu_amdkfd_remove_eviction_fence(
2283		bo, mem->process_info->eviction_fence);
2284
2285	if (size)
2286		*size = amdgpu_bo_size(bo);
2287
2288	amdgpu_bo_unreserve(bo);
2289
2290	mutex_unlock(&mem->process_info->lock);
2291	return 0;
2292
2293kmap_failed:
2294	amdgpu_bo_unpin(bo);
2295pin_failed:
2296	amdgpu_bo_unreserve(bo);
2297bo_reserve_failed:
2298	mutex_unlock(&mem->process_info->lock);
2299
2300	return ret;
2301}
2302
2303/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2304 *
2305 * @mem: Buffer object to be unmapped for CPU access
2306 *
2307 * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2308 * eviction fence, so this function should only be used for cleanup before the
2309 * BO is destroyed.
2310 */
2311void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2312{
2313	struct amdgpu_bo *bo = mem->bo;
2314
2315	amdgpu_bo_reserve(bo, true);
2316	amdgpu_bo_kunmap(bo);
2317	amdgpu_bo_unpin(bo);
2318	amdgpu_bo_unreserve(bo);
2319}
2320
2321int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2322					  struct kfd_vm_fault_info *mem)
2323{
2324	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2325		*mem = *adev->gmc.vm_fault_info;
2326		mb(); /* make sure read happened */
2327		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2328	}
2329	return 0;
2330}
2331
2332static int import_obj_create(struct amdgpu_device *adev,
2333			     struct dma_buf *dma_buf,
2334			     struct drm_gem_object *obj,
2335			     uint64_t va, void *drm_priv,
2336			     struct kgd_mem **mem, uint64_t *size,
2337			     uint64_t *mmap_offset)
2338{
2339	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2340	struct amdgpu_bo *bo;
2341	int ret;
2342
2343	bo = gem_to_amdgpu_bo(obj);
2344	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2345				    AMDGPU_GEM_DOMAIN_GTT)))
2346		/* Only VRAM and GTT BOs are supported */
2347		return -EINVAL;
2348
2349	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2350	if (!*mem)
2351		return -ENOMEM;
2352
2353	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2354	if (ret)
2355		goto err_free_mem;
2356
2357	if (size)
2358		*size = amdgpu_bo_size(bo);
2359
2360	if (mmap_offset)
2361		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2362
2363	INIT_LIST_HEAD(&(*mem)->attachments);
2364	mutex_init(&(*mem)->lock);
2365
2366	(*mem)->alloc_flags =
2367		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2368		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2369		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2370		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2371
2372	get_dma_buf(dma_buf);
2373	(*mem)->dmabuf = dma_buf;
2374	(*mem)->bo = bo;
2375	(*mem)->va = va;
2376	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
2377		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2378
2379	(*mem)->mapped_to_gpu_memory = 0;
2380	(*mem)->process_info = avm->process_info;
2381	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2382	amdgpu_sync_create(&(*mem)->sync);
2383	(*mem)->is_imported = true;
2384
2385	mutex_lock(&avm->process_info->lock);
2386	if (avm->process_info->eviction_fence &&
2387	    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
2388		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
2389				&avm->process_info->eviction_fence->base);
2390	mutex_unlock(&avm->process_info->lock);
2391	if (ret)
2392		goto err_remove_mem;
2393
2394	return 0;
2395
2396err_remove_mem:
2397	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
2398	drm_vma_node_revoke(&obj->vma_node, drm_priv);
2399err_free_mem:
2400	kfree(*mem);
2401	return ret;
2402}
2403
2404int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
2405					 uint64_t va, void *drm_priv,
2406					 struct kgd_mem **mem, uint64_t *size,
2407					 uint64_t *mmap_offset)
2408{
2409	struct drm_gem_object *obj;
2410	uint32_t handle;
2411	int ret;
2412
2413	ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
2414					 &handle);
2415	if (ret)
2416		return ret;
2417	obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
2418	if (!obj) {
2419		ret = -EINVAL;
2420		goto err_release_handle;
2421	}
2422
2423	ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
2424				mmap_offset);
2425	if (ret)
2426		goto err_put_obj;
2427
2428	(*mem)->gem_handle = handle;
2429
2430	return 0;
2431
2432err_put_obj:
2433	drm_gem_object_put(obj);
2434err_release_handle:
2435	drm_gem_handle_delete(adev->kfd.client.file, handle);
2436	return ret;
2437}
2438
2439int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2440				      struct dma_buf **dma_buf)
2441{
2442	int ret;
2443
2444	mutex_lock(&mem->lock);
2445	ret = kfd_mem_export_dmabuf(mem);
2446	if (ret)
2447		goto out;
2448
2449	get_dma_buf(mem->dmabuf);
2450	*dma_buf = mem->dmabuf;
2451out:
2452	mutex_unlock(&mem->lock);
2453	return ret;
2454}
2455
2456/* Evict a userptr BO by stopping the queues if necessary
2457 *
2458 * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2459 * cannot do any memory allocations, and cannot take any locks that
2460 * are held elsewhere while allocating memory.
2461 *
2462 * It doesn't do anything to the BO itself. The real work happens in
2463 * restore, where we get updated page addresses. This function only
2464 * ensures that GPU access to the BO is stopped.
2465 */
2466int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2467				unsigned long cur_seq, struct kgd_mem *mem)
2468{
2469	struct amdkfd_process_info *process_info = mem->process_info;
2470	int r = 0;
2471
2472	/* Do not process MMU notifications during CRIU restore until
2473	 * KFD_CRIU_OP_RESUME IOCTL is received
2474	 */
2475	if (READ_ONCE(process_info->block_mmu_notifications))
2476		return 0;
2477
2478	mutex_lock(&process_info->notifier_lock);
2479	mmu_interval_set_seq(mni, cur_seq);
2480
2481	mem->invalid++;
2482	if (++process_info->evicted_bos == 1) {
2483		/* First eviction, stop the queues */
2484		r = kgd2kfd_quiesce_mm(mni->mm,
2485				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2486		if (r)
2487			pr_err("Failed to quiesce KFD\n");
2488		queue_delayed_work(system_freezable_wq,
2489			&process_info->restore_userptr_work,
2490			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2491	}
2492	mutex_unlock(&process_info->notifier_lock);
2493
2494	return r;
2495}
2496
2497/* Update invalid userptr BOs
2498 *
2499 * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2500 * userptr_inval_list and updates user pages for all BOs that have
2501 * been invalidated since their last update.
2502 */
2503static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2504				     struct mm_struct *mm)
2505{
2506	struct kgd_mem *mem, *tmp_mem;
2507	struct amdgpu_bo *bo;
2508	struct ttm_operation_ctx ctx = { false, false };
2509	uint32_t invalid;
2510	int ret = 0;
2511
2512	mutex_lock(&process_info->notifier_lock);
2513
2514	/* Move all invalidated BOs to the userptr_inval_list */
2515	list_for_each_entry_safe(mem, tmp_mem,
2516				 &process_info->userptr_valid_list,
2517				 validate_list)
2518		if (mem->invalid)
2519			list_move_tail(&mem->validate_list,
2520				       &process_info->userptr_inval_list);
2521
2522	/* Go through userptr_inval_list and update any invalid user_pages */
2523	list_for_each_entry(mem, &process_info->userptr_inval_list,
2524			    validate_list) {
2525		invalid = mem->invalid;
2526		if (!invalid)
2527			/* BO hasn't been invalidated since the last
2528			 * revalidation attempt. Keep its page list.
2529			 */
2530			continue;
2531
2532		bo = mem->bo;
2533
2534		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2535		mem->range = NULL;
2536
2537		/* BO reservations and getting user pages (hmm_range_fault)
2538		 * must happen outside the notifier lock
2539		 */
2540		mutex_unlock(&process_info->notifier_lock);
2541
2542		/* Move the BO to system (CPU) domain if necessary to unmap
2543		 * and free the SG table
2544		 */
2545		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2546			if (amdgpu_bo_reserve(bo, true))
2547				return -EAGAIN;
2548			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2549			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2550			amdgpu_bo_unreserve(bo);
2551			if (ret) {
2552				pr_err("%s: Failed to invalidate userptr BO\n",
2553				       __func__);
2554				return -EAGAIN;
2555			}
2556		}
2557
2558		/* Get updated user pages */
2559		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2560						   &mem->range);
2561		if (ret) {
2562			pr_debug("Failed %d to get user pages\n", ret);
2563
2564			/* Return -EFAULT bad address error as success. It will
2565			 * fail later with a VM fault if the GPU tries to access
2566			 * it. Better than hanging indefinitely with stalled
2567			 * user mode queues.
2568			 *
2569			 * Return other error -EBUSY or -ENOMEM to retry restore
2570			 */
2571			if (ret != -EFAULT)
2572				return ret;
2573
2574			ret = 0;
2575		}
2576
2577		mutex_lock(&process_info->notifier_lock);
2578
2579		/* Mark the BO as valid unless it was invalidated
2580		 * again concurrently.
2581		 */
2582		if (mem->invalid != invalid) {
2583			ret = -EAGAIN;
2584			goto unlock_out;
2585		}
2586		 /* set mem valid if mem has hmm range associated */
2587		if (mem->range)
2588			mem->invalid = 0;
2589	}
2590
2591unlock_out:
2592	mutex_unlock(&process_info->notifier_lock);
2593
2594	return ret;
2595}
2596
2597/* Validate invalid userptr BOs
2598 *
2599 * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2600 * with new page addresses and waits for the page table updates to complete.
2601 */
2602static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2603{
2604	struct ttm_operation_ctx ctx = { false, false };
2605	struct amdgpu_sync sync;
2606	struct drm_exec exec;
2607
2608	struct amdgpu_vm *peer_vm;
2609	struct kgd_mem *mem, *tmp_mem;
2610	struct amdgpu_bo *bo;
2611	int ret;
2612
2613	amdgpu_sync_create(&sync);
2614
2615	drm_exec_init(&exec, 0, 0);
2616	/* Reserve all BOs and page tables for validation */
2617	drm_exec_until_all_locked(&exec) {
2618		/* Reserve all the page directories */
2619		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2620				    vm_list_node) {
2621			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2622			drm_exec_retry_on_contention(&exec);
2623			if (unlikely(ret))
2624				goto unreserve_out;
2625		}
2626
2627		/* Reserve the userptr_inval_list entries to resv_list */
2628		list_for_each_entry(mem, &process_info->userptr_inval_list,
2629				    validate_list) {
2630			struct drm_gem_object *gobj;
2631
2632			gobj = &mem->bo->tbo.base;
2633			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2634			drm_exec_retry_on_contention(&exec);
2635			if (unlikely(ret))
2636				goto unreserve_out;
2637		}
2638	}
2639
2640	ret = process_validate_vms(process_info, NULL);
2641	if (ret)
2642		goto unreserve_out;
2643
2644	/* Validate BOs and update GPUVM page tables */
2645	list_for_each_entry_safe(mem, tmp_mem,
2646				 &process_info->userptr_inval_list,
2647				 validate_list) {
2648		struct kfd_mem_attachment *attachment;
2649
2650		bo = mem->bo;
2651
2652		/* Validate the BO if we got user pages */
2653		if (bo->tbo.ttm->pages[0]) {
2654			amdgpu_bo_placement_from_domain(bo, mem->domain);
2655			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2656			if (ret) {
2657				pr_err("%s: failed to validate BO\n", __func__);
2658				goto unreserve_out;
2659			}
2660		}
2661
2662		/* Update mapping. If the BO was not validated
2663		 * (because we couldn't get user pages), this will
2664		 * clear the page table entries, which will result in
2665		 * VM faults if the GPU tries to access the invalid
2666		 * memory.
2667		 */
2668		list_for_each_entry(attachment, &mem->attachments, list) {
2669			if (!attachment->is_mapped)
2670				continue;
2671
2672			kfd_mem_dmaunmap_attachment(mem, attachment);
2673			ret = update_gpuvm_pte(mem, attachment, &sync);
2674			if (ret) {
2675				pr_err("%s: update PTE failed\n", __func__);
2676				/* make sure this gets validated again */
2677				mutex_lock(&process_info->notifier_lock);
2678				mem->invalid++;
2679				mutex_unlock(&process_info->notifier_lock);
2680				goto unreserve_out;
2681			}
2682		}
2683	}
2684
2685	/* Update page directories */
2686	ret = process_update_pds(process_info, &sync);
2687
2688unreserve_out:
2689	drm_exec_fini(&exec);
2690	amdgpu_sync_wait(&sync, false);
2691	amdgpu_sync_free(&sync);
2692
2693	return ret;
2694}
2695
2696/* Confirm that all user pages are valid while holding the notifier lock
2697 *
2698 * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2699 */
2700static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2701{
2702	struct kgd_mem *mem, *tmp_mem;
2703	int ret = 0;
2704
2705	list_for_each_entry_safe(mem, tmp_mem,
2706				 &process_info->userptr_inval_list,
2707				 validate_list) {
2708		bool valid;
2709
2710		/* keep mem without hmm range at userptr_inval_list */
2711		if (!mem->range)
2712			 continue;
2713
2714		/* Only check mem with hmm range associated */
2715		valid = amdgpu_ttm_tt_get_user_pages_done(
2716					mem->bo->tbo.ttm, mem->range);
2717
2718		mem->range = NULL;
2719		if (!valid) {
2720			WARN(!mem->invalid, "Invalid BO not marked invalid");
2721			ret = -EAGAIN;
2722			continue;
2723		}
2724
2725		if (mem->invalid) {
2726			WARN(1, "Valid BO is marked invalid");
2727			ret = -EAGAIN;
2728			continue;
2729		}
2730
2731		list_move_tail(&mem->validate_list,
2732			       &process_info->userptr_valid_list);
2733	}
2734
2735	return ret;
2736}
2737
2738/* Worker callback to restore evicted userptr BOs
2739 *
2740 * Tries to update and validate all userptr BOs. If successful and no
2741 * concurrent evictions happened, the queues are restarted. Otherwise,
2742 * reschedule for another attempt later.
2743 */
2744static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2745{
2746	struct delayed_work *dwork = to_delayed_work(work);
2747	struct amdkfd_process_info *process_info =
2748		container_of(dwork, struct amdkfd_process_info,
2749			     restore_userptr_work);
2750	struct task_struct *usertask;
2751	struct mm_struct *mm;
2752	uint32_t evicted_bos;
2753
2754	mutex_lock(&process_info->notifier_lock);
2755	evicted_bos = process_info->evicted_bos;
2756	mutex_unlock(&process_info->notifier_lock);
2757	if (!evicted_bos)
2758		return;
2759
2760	/* Reference task and mm in case of concurrent process termination */
2761	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2762	if (!usertask)
2763		return;
2764	mm = get_task_mm(usertask);
2765	if (!mm) {
2766		put_task_struct(usertask);
2767		return;
2768	}
2769
2770	mutex_lock(&process_info->lock);
2771
2772	if (update_invalid_user_pages(process_info, mm))
2773		goto unlock_out;
2774	/* userptr_inval_list can be empty if all evicted userptr BOs
2775	 * have been freed. In that case there is nothing to validate
2776	 * and we can just restart the queues.
2777	 */
2778	if (!list_empty(&process_info->userptr_inval_list)) {
2779		if (validate_invalid_user_pages(process_info))
2780			goto unlock_out;
2781	}
2782	/* Final check for concurrent evicton and atomic update. If
2783	 * another eviction happens after successful update, it will
2784	 * be a first eviction that calls quiesce_mm. The eviction
2785	 * reference counting inside KFD will handle this case.
2786	 */
2787	mutex_lock(&process_info->notifier_lock);
2788	if (process_info->evicted_bos != evicted_bos)
2789		goto unlock_notifier_out;
2790
2791	if (confirm_valid_user_pages_locked(process_info)) {
2792		WARN(1, "User pages unexpectedly invalid");
2793		goto unlock_notifier_out;
2794	}
2795
2796	process_info->evicted_bos = evicted_bos = 0;
2797
2798	if (kgd2kfd_resume_mm(mm)) {
2799		pr_err("%s: Failed to resume KFD\n", __func__);
2800		/* No recovery from this failure. Probably the CP is
2801		 * hanging. No point trying again.
2802		 */
2803	}
2804
2805unlock_notifier_out:
2806	mutex_unlock(&process_info->notifier_lock);
2807unlock_out:
2808	mutex_unlock(&process_info->lock);
2809
2810	/* If validation failed, reschedule another attempt */
2811	if (evicted_bos) {
2812		queue_delayed_work(system_freezable_wq,
2813			&process_info->restore_userptr_work,
2814			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2815
2816		kfd_smi_event_queue_restore_rescheduled(mm);
2817	}
2818	mmput(mm);
2819	put_task_struct(usertask);
2820}
2821
2822static void replace_eviction_fence(struct dma_fence __rcu **ef,
2823				   struct dma_fence *new_ef)
2824{
2825	struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
2826		/* protected by process_info->lock */);
2827
2828	/* If we're replacing an unsignaled eviction fence, that fence will
2829	 * never be signaled, and if anyone is still waiting on that fence,
2830	 * they will hang forever. This should never happen. We should only
2831	 * replace the fence in restore_work that only gets scheduled after
2832	 * eviction work signaled the fence.
2833	 */
2834	WARN_ONCE(!dma_fence_is_signaled(old_ef),
2835		  "Replacing unsignaled eviction fence");
2836	dma_fence_put(old_ef);
2837}
2838
2839/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2840 *   KFD process identified by process_info
2841 *
2842 * @process_info: amdkfd_process_info of the KFD process
2843 *
2844 * After memory eviction, restore thread calls this function. The function
2845 * should be called when the Process is still valid. BO restore involves -
2846 *
2847 * 1.  Release old eviction fence and create new one
2848 * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2849 * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2850 *     BOs that need to be reserved.
2851 * 4.  Reserve all the BOs
2852 * 5.  Validate of PD and PT BOs.
2853 * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2854 * 7.  Add fence to all PD and PT BOs.
2855 * 8.  Unreserve all BOs
2856 */
2857int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef)
2858{
2859	struct amdkfd_process_info *process_info = info;
2860	struct amdgpu_vm *peer_vm;
2861	struct kgd_mem *mem;
2862	struct list_head duplicate_save;
2863	struct amdgpu_sync sync_obj;
2864	unsigned long failed_size = 0;
2865	unsigned long total_size = 0;
2866	struct drm_exec exec;
2867	int ret;
2868
2869	INIT_LIST_HEAD(&duplicate_save);
2870
2871	mutex_lock(&process_info->lock);
2872
2873	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
2874	drm_exec_until_all_locked(&exec) {
2875		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2876				    vm_list_node) {
2877			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2878			drm_exec_retry_on_contention(&exec);
2879			if (unlikely(ret)) {
2880				pr_err("Locking VM PD failed, ret: %d\n", ret);
2881				goto ttm_reserve_fail;
2882			}
2883		}
2884
2885		/* Reserve all BOs and page tables/directory. Add all BOs from
2886		 * kfd_bo_list to ctx.list
2887		 */
2888		list_for_each_entry(mem, &process_info->kfd_bo_list,
2889				    validate_list) {
2890			struct drm_gem_object *gobj;
2891
2892			gobj = &mem->bo->tbo.base;
2893			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2894			drm_exec_retry_on_contention(&exec);
2895			if (unlikely(ret)) {
2896				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);
2897				goto ttm_reserve_fail;
2898			}
2899		}
2900	}
2901
2902	amdgpu_sync_create(&sync_obj);
2903
2904	/* Validate BOs managed by KFD */
 
 
 
 
 
2905	list_for_each_entry(mem, &process_info->kfd_bo_list,
2906			    validate_list) {
2907
2908		struct amdgpu_bo *bo = mem->bo;
2909		uint32_t domain = mem->domain;
 
2910		struct dma_resv_iter cursor;
2911		struct dma_fence *fence;
2912
2913		total_size += amdgpu_bo_size(bo);
2914
2915		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2916		if (ret) {
2917			pr_debug("Memory eviction: Validate BOs failed\n");
2918			failed_size += amdgpu_bo_size(bo);
2919			ret = amdgpu_amdkfd_bo_validate(bo,
2920						AMDGPU_GEM_DOMAIN_GTT, false);
2921			if (ret) {
2922				pr_debug("Memory eviction: Try again\n");
2923				goto validate_map_fail;
2924			}
2925		}
2926		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2927					DMA_RESV_USAGE_KERNEL, fence) {
2928			ret = amdgpu_sync_fence(&sync_obj, fence);
2929			if (ret) {
2930				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2931				goto validate_map_fail;
2932			}
2933		}
2934	}
2935
2936	if (failed_size)
2937		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2938
2939	/* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO
2940	 * validations above would invalidate DMABuf imports again.
2941	 */
2942	ret = process_validate_vms(process_info, &exec.ticket);
2943	if (ret) {
2944		pr_debug("Validating VMs failed, ret: %d\n", ret);
2945		goto validate_map_fail;
2946	}
2947
2948	/* Update mappings managed by KFD. */
2949	list_for_each_entry(mem, &process_info->kfd_bo_list,
2950			    validate_list) {
2951		struct kfd_mem_attachment *attachment;
2952
2953		list_for_each_entry(attachment, &mem->attachments, list) {
2954			if (!attachment->is_mapped)
2955				continue;
2956
2957			if (attachment->bo_va->base.bo->tbo.pin_count)
2958				continue;
2959
2960			kfd_mem_dmaunmap_attachment(mem, attachment);
2961			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2962			if (ret) {
2963				pr_debug("Memory eviction: update PTE failed. Try again\n");
2964				goto validate_map_fail;
2965			}
2966		}
2967	}
2968
 
 
 
2969	/* Update mappings not managed by KFD */
2970	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2971			vm_list_node) {
2972		struct amdgpu_device *adev = amdgpu_ttm_adev(
2973			peer_vm->root.bo->tbo.bdev);
2974
2975		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
2976		if (ret) {
2977			pr_debug("Memory eviction: handle moved failed. Try again\n");
2978			goto validate_map_fail;
2979		}
2980	}
2981
2982	/* Update page directories */
2983	ret = process_update_pds(process_info, &sync_obj);
2984	if (ret) {
2985		pr_debug("Memory eviction: update PDs failed. Try again\n");
2986		goto validate_map_fail;
2987	}
2988
2989	/* Sync with fences on all the page tables. They implicitly depend on any
2990	 * move fences from amdgpu_vm_handle_moved above.
2991	 */
2992	ret = process_sync_pds_resv(process_info, &sync_obj);
2993	if (ret) {
2994		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2995		goto validate_map_fail;
2996	}
2997
2998	/* Wait for validate and PT updates to finish */
2999	amdgpu_sync_wait(&sync_obj, false);
3000
3001	/* The old eviction fence may be unsignaled if restore happens
3002	 * after a GPU reset or suspend/resume. Keep the old fence in that
3003	 * case. Otherwise release the old eviction fence and create new
3004	 * one, because fence only goes from unsignaled to signaled once
3005	 * and cannot be reused. Use context and mm from the old fence.
3006	 *
3007	 * If an old eviction fence signals after this check, that's OK.
3008	 * Anyone signaling an eviction fence must stop the queues first
3009	 * and schedule another restore worker.
3010	 */
3011	if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
3012		struct amdgpu_amdkfd_fence *new_fence =
3013			amdgpu_amdkfd_fence_create(
3014				process_info->eviction_fence->base.context,
3015				process_info->eviction_fence->mm,
3016				NULL);
3017
3018		if (!new_fence) {
3019			pr_err("Failed to create eviction fence\n");
3020			ret = -ENOMEM;
3021			goto validate_map_fail;
3022		}
3023		dma_fence_put(&process_info->eviction_fence->base);
3024		process_info->eviction_fence = new_fence;
3025		replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
3026	} else {
3027		WARN_ONCE(*ef != &process_info->eviction_fence->base,
3028			  "KFD eviction fence doesn't match KGD process_info");
3029	}
3030
3031	/* Attach new eviction fence to all BOs except pinned ones */
3032	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
3033		if (mem->bo->tbo.pin_count)
3034			continue;
3035
3036		dma_resv_add_fence(mem->bo->tbo.base.resv,
3037				   &process_info->eviction_fence->base,
3038				   DMA_RESV_USAGE_BOOKKEEP);
3039	}
3040	/* Attach eviction fence to PD / PT BOs and DMABuf imports */
3041	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3042			    vm_list_node) {
3043		struct amdgpu_bo *bo = peer_vm->root.bo;
3044
3045		dma_resv_add_fence(bo->tbo.base.resv,
3046				   &process_info->eviction_fence->base,
3047				   DMA_RESV_USAGE_BOOKKEEP);
3048	}
3049
3050validate_map_fail:
3051	amdgpu_sync_free(&sync_obj);
3052ttm_reserve_fail:
3053	drm_exec_fini(&exec);
3054	mutex_unlock(&process_info->lock);
3055	return ret;
3056}
3057
3058int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
3059{
3060	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3061	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
3062	int ret;
3063
3064	if (!info || !gws)
3065		return -EINVAL;
3066
3067	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
3068	if (!*mem)
3069		return -ENOMEM;
3070
3071	mutex_init(&(*mem)->lock);
3072	INIT_LIST_HEAD(&(*mem)->attachments);
3073	(*mem)->bo = amdgpu_bo_ref(gws_bo);
3074	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
3075	(*mem)->process_info = process_info;
3076	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
3077	amdgpu_sync_create(&(*mem)->sync);
3078
3079
3080	/* Validate gws bo the first time it is added to process */
3081	mutex_lock(&(*mem)->process_info->lock);
3082	ret = amdgpu_bo_reserve(gws_bo, false);
3083	if (unlikely(ret)) {
3084		pr_err("Reserve gws bo failed %d\n", ret);
3085		goto bo_reservation_failure;
3086	}
3087
3088	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
3089	if (ret) {
3090		pr_err("GWS BO validate failed %d\n", ret);
3091		goto bo_validation_failure;
3092	}
3093	/* GWS resource is shared b/t amdgpu and amdkfd
3094	 * Add process eviction fence to bo so they can
3095	 * evict each other.
3096	 */
3097	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
3098	if (ret)
3099		goto reserve_shared_fail;
3100	dma_resv_add_fence(gws_bo->tbo.base.resv,
3101			   &process_info->eviction_fence->base,
3102			   DMA_RESV_USAGE_BOOKKEEP);
3103	amdgpu_bo_unreserve(gws_bo);
3104	mutex_unlock(&(*mem)->process_info->lock);
3105
3106	return ret;
3107
3108reserve_shared_fail:
3109bo_validation_failure:
3110	amdgpu_bo_unreserve(gws_bo);
3111bo_reservation_failure:
3112	mutex_unlock(&(*mem)->process_info->lock);
3113	amdgpu_sync_free(&(*mem)->sync);
3114	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
3115	amdgpu_bo_unref(&gws_bo);
3116	mutex_destroy(&(*mem)->lock);
3117	kfree(*mem);
3118	*mem = NULL;
3119	return ret;
3120}
3121
3122int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
3123{
3124	int ret;
3125	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3126	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
3127	struct amdgpu_bo *gws_bo = kgd_mem->bo;
3128
3129	/* Remove BO from process's validate list so restore worker won't touch
3130	 * it anymore
3131	 */
3132	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
3133
3134	ret = amdgpu_bo_reserve(gws_bo, false);
3135	if (unlikely(ret)) {
3136		pr_err("Reserve gws bo failed %d\n", ret);
3137		//TODO add BO back to validate_list?
3138		return ret;
3139	}
3140	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
3141			process_info->eviction_fence);
3142	amdgpu_bo_unreserve(gws_bo);
3143	amdgpu_sync_free(&kgd_mem->sync);
3144	amdgpu_bo_unref(&gws_bo);
3145	mutex_destroy(&kgd_mem->lock);
3146	kfree(mem);
3147	return 0;
3148}
3149
3150/* Returns GPU-specific tiling mode information */
3151int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
3152				struct tile_config *config)
3153{
3154	config->gb_addr_config = adev->gfx.config.gb_addr_config;
3155	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
3156	config->num_tile_configs =
3157			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
3158	config->macro_tile_config_ptr =
3159			adev->gfx.config.macrotile_mode_array;
3160	config->num_macro_tile_configs =
3161			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
3162
3163	/* Those values are not set from GFX9 onwards */
3164	config->num_banks = adev->gfx.config.num_banks;
3165	config->num_ranks = adev->gfx.config.num_ranks;
3166
3167	return 0;
3168}
3169
3170bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
3171{
3172	struct kfd_mem_attachment *entry;
3173
3174	list_for_each_entry(entry, &mem->attachments, list) {
3175		if (entry->is_mapped && entry->adev == adev)
3176			return true;
3177	}
3178	return false;
3179}
3180
3181#if defined(CONFIG_DEBUG_FS)
3182
3183int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3184{
3185
3186	spin_lock(&kfd_mem_limit.mem_limit_lock);
3187	seq_printf(m, "System mem used %lldM out of %lluM\n",
3188		  (kfd_mem_limit.system_mem_used >> 20),
3189		  (kfd_mem_limit.max_system_mem_limit >> 20));
3190	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3191		  (kfd_mem_limit.ttm_mem_used >> 20),
3192		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3193	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3194
3195	return 0;
3196}
3197
3198#endif
v6.8
   1// SPDX-License-Identifier: MIT
   2/*
   3 * Copyright 2014-2018 Advanced Micro Devices, Inc.
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice shall be included in
  13 * all copies or substantial portions of the Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21 * OTHER DEALINGS IN THE SOFTWARE.
  22 */
  23#include <linux/dma-buf.h>
  24#include <linux/list.h>
  25#include <linux/pagemap.h>
  26#include <linux/sched/mm.h>
  27#include <linux/sched/task.h>
  28#include <linux/fdtable.h>
  29#include <drm/ttm/ttm_tt.h>
  30
  31#include <drm/drm_exec.h>
  32
  33#include "amdgpu_object.h"
  34#include "amdgpu_gem.h"
  35#include "amdgpu_vm.h"
  36#include "amdgpu_hmm.h"
  37#include "amdgpu_amdkfd.h"
  38#include "amdgpu_dma_buf.h"
  39#include <uapi/linux/kfd_ioctl.h>
  40#include "amdgpu_xgmi.h"
  41#include "kfd_priv.h"
  42#include "kfd_smi_events.h"
  43
  44/* Userptr restore delay, just long enough to allow consecutive VM
  45 * changes to accumulate
  46 */
  47#define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
  48#define AMDGPU_RESERVE_MEM_LIMIT			(3UL << 29)
  49
  50/*
  51 * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
  52 * BO chunk
  53 */
  54#define VRAM_AVAILABLITY_ALIGN (1 << 21)
  55
  56/* Impose limit on how much memory KFD can use */
  57static struct {
  58	uint64_t max_system_mem_limit;
  59	uint64_t max_ttm_mem_limit;
  60	int64_t system_mem_used;
  61	int64_t ttm_mem_used;
  62	spinlock_t mem_limit_lock;
  63} kfd_mem_limit;
  64
  65static const char * const domain_bit_to_string[] = {
  66		"CPU",
  67		"GTT",
  68		"VRAM",
  69		"GDS",
  70		"GWS",
  71		"OA"
  72};
  73
  74#define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
  75
  76static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
  77
  78static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
  79		struct kgd_mem *mem)
  80{
  81	struct kfd_mem_attachment *entry;
  82
  83	list_for_each_entry(entry, &mem->attachments, list)
  84		if (entry->bo_va->base.vm == avm)
  85			return true;
  86
  87	return false;
  88}
  89
  90/**
  91 * reuse_dmamap() - Check whether adev can share the original
  92 * userptr BO
  93 *
  94 * If both adev and bo_adev are in direct mapping or
  95 * in the same iommu group, they can share the original BO.
  96 *
  97 * @adev: Device to which can or cannot share the original BO
  98 * @bo_adev: Device to which allocated BO belongs to
  99 *
 100 * Return: returns true if adev can share original userptr BO,
 101 * false otherwise.
 102 */
 103static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
 104{
 105	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
 106			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
 107}
 108
 109/* Set memory usage limits. Current, limits are
 110 *  System (TTM + userptr) memory - 15/16th System RAM
 111 *  TTM memory - 3/8th System RAM
 112 */
 113void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 114{
 115	struct sysinfo si;
 116	uint64_t mem;
 117
 118	if (kfd_mem_limit.max_system_mem_limit)
 119		return;
 120
 121	si_meminfo(&si);
 122	mem = si.totalram - si.totalhigh;
 123	mem *= si.mem_unit;
 124
 125	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
 126	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6);
 127	if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT)
 128		kfd_mem_limit.max_system_mem_limit >>= 1;
 129	else
 130		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
 131
 132	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
 133	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
 134		(kfd_mem_limit.max_system_mem_limit >> 20),
 135		(kfd_mem_limit.max_ttm_mem_limit >> 20));
 136}
 137
 138void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
 139{
 140	kfd_mem_limit.system_mem_used += size;
 141}
 142
 143/* Estimate page table size needed to represent a given memory size
 144 *
 145 * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
 146 * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
 147 * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
 148 * for 2MB pages for TLB efficiency. However, small allocations and
 149 * fragmented system memory still need some 4KB pages. We choose a
 150 * compromise that should work in most cases without reserving too
 151 * much memory for page tables unnecessarily (factor 16K, >> 14).
 152 */
 153
 154#define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
 155
 156/**
 157 * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
 158 * of buffer.
 159 *
 160 * @adev: Device to which allocated BO belongs to
 161 * @size: Size of buffer, in bytes, encapsulated by B0. This should be
 162 * equivalent to amdgpu_bo_size(BO)
 163 * @alloc_flag: Flag used in allocating a BO as noted above
 164 * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
 165 * managed as one compute node in driver for app
 166 *
 167 * Return:
 168 *	returns -ENOMEM in case of error, ZERO otherwise
 169 */
 170int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 171		uint64_t size, u32 alloc_flag, int8_t xcp_id)
 172{
 173	uint64_t reserved_for_pt =
 174		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
 175	size_t system_mem_needed, ttm_mem_needed, vram_needed;
 176	int ret = 0;
 177	uint64_t vram_size = 0;
 178
 179	system_mem_needed = 0;
 180	ttm_mem_needed = 0;
 181	vram_needed = 0;
 182	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 183		system_mem_needed = size;
 184		ttm_mem_needed = size;
 185	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 186		/*
 187		 * Conservatively round up the allocation requirement to 2 MB
 188		 * to avoid fragmentation caused by 4K allocations in the tail
 189		 * 2M BO chunk.
 190		 */
 191		vram_needed = size;
 192		/*
 193		 * For GFX 9.4.3, get the VRAM size from XCP structs
 194		 */
 195		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 196			return -EINVAL;
 197
 198		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
 199		if (adev->gmc.is_app_apu) {
 200			system_mem_needed = size;
 201			ttm_mem_needed = size;
 202		}
 203	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 204		system_mem_needed = size;
 205	} else if (!(alloc_flag &
 206				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 207				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 208		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 209		return -ENOMEM;
 210	}
 211
 212	spin_lock(&kfd_mem_limit.mem_limit_lock);
 213
 214	if (kfd_mem_limit.system_mem_used + system_mem_needed >
 215	    kfd_mem_limit.max_system_mem_limit)
 216		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
 217
 218	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
 219	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
 220	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
 221	     kfd_mem_limit.max_ttm_mem_limit) ||
 222	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
 223	     vram_size - reserved_for_pt)) {
 224		ret = -ENOMEM;
 225		goto release;
 226	}
 227
 228	/* Update memory accounting by decreasing available system
 229	 * memory, TTM memory and GPU memory as computed above
 230	 */
 231	WARN_ONCE(vram_needed && !adev,
 232		  "adev reference can't be null when vram is used");
 233	if (adev && xcp_id >= 0) {
 234		adev->kfd.vram_used[xcp_id] += vram_needed;
 235		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
 236				vram_needed :
 237				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 238	}
 239	kfd_mem_limit.system_mem_used += system_mem_needed;
 240	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 241
 242release:
 243	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 244	return ret;
 245}
 246
 247void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 248		uint64_t size, u32 alloc_flag, int8_t xcp_id)
 249{
 250	spin_lock(&kfd_mem_limit.mem_limit_lock);
 251
 252	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
 253		kfd_mem_limit.system_mem_used -= size;
 254		kfd_mem_limit.ttm_mem_used -= size;
 255	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 256		WARN_ONCE(!adev,
 257			  "adev reference can't be null when alloc mem flags vram is set");
 258		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
 259			goto release;
 260
 261		if (adev) {
 262			adev->kfd.vram_used[xcp_id] -= size;
 263			if (adev->gmc.is_app_apu) {
 264				adev->kfd.vram_used_aligned[xcp_id] -= size;
 265				kfd_mem_limit.system_mem_used -= size;
 266				kfd_mem_limit.ttm_mem_used -= size;
 267			} else {
 268				adev->kfd.vram_used_aligned[xcp_id] -=
 269					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
 270			}
 271		}
 272	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 273		kfd_mem_limit.system_mem_used -= size;
 274	} else if (!(alloc_flag &
 275				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
 276				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 277		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
 278		goto release;
 279	}
 280	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
 281		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
 282	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
 283		  "KFD TTM memory accounting unbalanced");
 284	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 285		  "KFD system memory accounting unbalanced");
 286
 287release:
 288	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 289}
 290
 291void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
 292{
 293	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 294	u32 alloc_flags = bo->kfd_bo->alloc_flags;
 295	u64 size = amdgpu_bo_size(bo);
 296
 297	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
 298					  bo->xcp_id);
 299
 300	kfree(bo->kfd_bo);
 301}
 302
 303/**
 304 * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
 305 * about USERPTR or DOOREBELL or MMIO BO.
 306 *
 307 * @adev: Device for which dmamap BO is being created
 308 * @mem: BO of peer device that is being DMA mapped. Provides parameters
 309 *	 in building the dmamap BO
 310 * @bo_out: Output parameter updated with handle of dmamap BO
 311 */
 312static int
 313create_dmamap_sg_bo(struct amdgpu_device *adev,
 314		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
 315{
 316	struct drm_gem_object *gem_obj;
 317	int ret;
 318	uint64_t flags = 0;
 319
 320	ret = amdgpu_bo_reserve(mem->bo, false);
 321	if (ret)
 322		return ret;
 323
 324	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
 325		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
 326					AMDGPU_GEM_CREATE_UNCACHED);
 327
 328	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
 329			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
 330			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
 331
 332	amdgpu_bo_unreserve(mem->bo);
 333
 334	if (ret) {
 335		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
 336		return -EINVAL;
 337	}
 338
 339	*bo_out = gem_to_amdgpu_bo(gem_obj);
 340	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
 341	return ret;
 342}
 343
 344/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
 345 *  reservation object.
 346 *
 347 * @bo: [IN] Remove eviction fence(s) from this BO
 348 * @ef: [IN] This eviction fence is removed if it
 349 *  is present in the shared list.
 350 *
 351 * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
 352 */
 353static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
 354					struct amdgpu_amdkfd_fence *ef)
 355{
 356	struct dma_fence *replacement;
 357
 358	if (!ef)
 359		return -EINVAL;
 360
 361	/* TODO: Instead of block before we should use the fence of the page
 362	 * table update and TLB flush here directly.
 363	 */
 364	replacement = dma_fence_get_stub();
 365	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
 366				replacement, DMA_RESV_USAGE_BOOKKEEP);
 367	dma_fence_put(replacement);
 368	return 0;
 369}
 370
 371int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
 372{
 373	struct amdgpu_bo *root = bo;
 374	struct amdgpu_vm_bo_base *vm_bo;
 375	struct amdgpu_vm *vm;
 376	struct amdkfd_process_info *info;
 377	struct amdgpu_amdkfd_fence *ef;
 378	int ret;
 379
 380	/* we can always get vm_bo from root PD bo.*/
 381	while (root->parent)
 382		root = root->parent;
 383
 384	vm_bo = root->vm_bo;
 385	if (!vm_bo)
 386		return 0;
 387
 388	vm = vm_bo->vm;
 389	if (!vm)
 390		return 0;
 391
 392	info = vm->process_info;
 393	if (!info || !info->eviction_fence)
 394		return 0;
 395
 396	ef = container_of(dma_fence_get(&info->eviction_fence->base),
 397			struct amdgpu_amdkfd_fence, base);
 398
 399	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
 400	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
 401	dma_resv_unlock(bo->tbo.base.resv);
 402
 403	dma_fence_put(&ef->base);
 404	return ret;
 405}
 406
 407static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
 408				     bool wait)
 409{
 410	struct ttm_operation_ctx ctx = { false, false };
 411	int ret;
 412
 413	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
 414		 "Called with userptr BO"))
 415		return -EINVAL;
 416
 417	amdgpu_bo_placement_from_domain(bo, domain);
 418
 419	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 420	if (ret)
 421		goto validate_fail;
 422	if (wait)
 423		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 424
 425validate_fail:
 426	return ret;
 427}
 428
 429static int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
 430					       uint32_t domain,
 431					       struct dma_fence *fence)
 432{
 433	int ret = amdgpu_bo_reserve(bo, false);
 434
 435	if (ret)
 436		return ret;
 437
 438	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
 439	if (ret)
 440		goto unreserve_out;
 441
 442	ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
 443	if (ret)
 444		goto unreserve_out;
 445
 446	dma_resv_add_fence(bo->tbo.base.resv, fence,
 447			   DMA_RESV_USAGE_BOOKKEEP);
 448
 449unreserve_out:
 450	amdgpu_bo_unreserve(bo);
 451
 452	return ret;
 453}
 454
 455static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
 456{
 457	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
 458}
 459
 460/* vm_validate_pt_pd_bos - Validate page table and directory BOs
 461 *
 462 * Page directories are not updated here because huge page handling
 463 * during page table updates can invalidate page directory entries
 464 * again. Page directories are only updated after updating page
 465 * tables.
 466 */
 467static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
 
 468{
 469	struct amdgpu_bo *pd = vm->root.bo;
 470	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 471	int ret;
 472
 473	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
 
 474	if (ret) {
 475		pr_err("failed to validate PT BOs\n");
 476		return ret;
 477	}
 478
 479	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 480
 481	return 0;
 482}
 483
 484static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
 485{
 486	struct amdgpu_bo *pd = vm->root.bo;
 487	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
 488	int ret;
 489
 490	ret = amdgpu_vm_update_pdes(adev, vm, false);
 491	if (ret)
 492		return ret;
 493
 494	return amdgpu_sync_fence(sync, vm->last_update);
 495}
 496
 497static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
 498{
 499	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
 500				 AMDGPU_VM_MTYPE_DEFAULT;
 501
 502	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
 503		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
 504	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
 505		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
 506
 507	return amdgpu_gem_va_map_flags(adev, mapping_flags);
 508}
 509
 510/**
 511 * create_sg_table() - Create an sg_table for a contiguous DMA addr range
 512 * @addr: The starting address to point to
 513 * @size: Size of memory area in bytes being pointed to
 514 *
 515 * Allocates an instance of sg_table and initializes it to point to memory
 516 * area specified by input parameters. The address used to build is assumed
 517 * to be DMA mapped, if needed.
 518 *
 519 * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
 520 * because they are physically contiguous.
 521 *
 522 * Return: Initialized instance of SG Table or NULL
 523 */
 524static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
 525{
 526	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
 527
 528	if (!sg)
 529		return NULL;
 530	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
 531		kfree(sg);
 532		return NULL;
 533	}
 534	sg_dma_address(sg->sgl) = addr;
 535	sg->sgl->length = size;
 536#ifdef CONFIG_NEED_SG_DMA_LENGTH
 537	sg->sgl->dma_length = size;
 538#endif
 539	return sg;
 540}
 541
 542static int
 543kfd_mem_dmamap_userptr(struct kgd_mem *mem,
 544		       struct kfd_mem_attachment *attachment)
 545{
 546	enum dma_data_direction direction =
 547		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 548		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 549	struct ttm_operation_ctx ctx = {.interruptible = true};
 550	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 551	struct amdgpu_device *adev = attachment->adev;
 552	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
 553	struct ttm_tt *ttm = bo->tbo.ttm;
 554	int ret;
 555
 556	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
 557		return -EINVAL;
 558
 559	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
 560	if (unlikely(!ttm->sg))
 561		return -ENOMEM;
 562
 563	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
 564	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
 565					ttm->num_pages, 0,
 566					(u64)ttm->num_pages << PAGE_SHIFT,
 567					GFP_KERNEL);
 568	if (unlikely(ret))
 569		goto free_sg;
 570
 571	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 572	if (unlikely(ret))
 573		goto release_sg;
 574
 575	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 576	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 577	if (ret)
 578		goto unmap_sg;
 579
 580	return 0;
 581
 582unmap_sg:
 583	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 584release_sg:
 585	pr_err("DMA map userptr failed: %d\n", ret);
 586	sg_free_table(ttm->sg);
 587free_sg:
 588	kfree(ttm->sg);
 589	ttm->sg = NULL;
 590	return ret;
 591}
 592
 593static int
 594kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
 595{
 596	struct ttm_operation_ctx ctx = {.interruptible = true};
 597	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 598	int ret;
 599
 600	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 601	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 602	if (ret)
 603		return ret;
 604
 605	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 606	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 607}
 608
 609/**
 610 * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
 611 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 612 * @attachment: Virtual address attachment of the BO on accessing device
 613 *
 614 * An access request from the device that owns DOORBELL does not require DMA mapping.
 615 * This is because the request doesn't go through PCIe root complex i.e. it instead
 616 * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
 617 *
 618 * In contrast, all access requests for MMIO need to be DMA mapped without regard to
 619 * device ownership. This is because access requests for MMIO go through PCIe root
 620 * complex.
 621 *
 622 * This is accomplished in two steps:
 623 *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
 624 *         in updating requesting device's page table
 625 *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
 626 *         accessible. This allows an update of requesting device's page table
 627 *         with entries associated with DOOREBELL or MMIO memory
 628 *
 629 * This method is invoked in the following contexts:
 630 *   - Mapping of DOORBELL or MMIO BO of same or peer device
 631 *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
 632 *
 633 * Return: ZERO if successful, NON-ZERO otherwise
 634 */
 635static int
 636kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
 637		     struct kfd_mem_attachment *attachment)
 638{
 639	struct ttm_operation_ctx ctx = {.interruptible = true};
 640	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 641	struct amdgpu_device *adev = attachment->adev;
 642	struct ttm_tt *ttm = bo->tbo.ttm;
 643	enum dma_data_direction dir;
 644	dma_addr_t dma_addr;
 645	bool mmio;
 646	int ret;
 647
 648	/* Expect SG Table of dmapmap BO to be NULL */
 649	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
 650	if (unlikely(ttm->sg)) {
 651		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
 652		return -EINVAL;
 653	}
 654
 655	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 656			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 657	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
 658	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
 659	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
 660	dma_addr = dma_map_resource(adev->dev, dma_addr,
 661			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 662	ret = dma_mapping_error(adev->dev, dma_addr);
 663	if (unlikely(ret))
 664		return ret;
 665	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
 666
 667	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
 668	if (unlikely(!ttm->sg)) {
 669		ret = -ENOMEM;
 670		goto unmap_sg;
 671	}
 672
 673	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
 674	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 675	if (unlikely(ret))
 676		goto free_sg;
 677
 678	return ret;
 679
 680free_sg:
 681	sg_free_table(ttm->sg);
 682	kfree(ttm->sg);
 683	ttm->sg = NULL;
 684unmap_sg:
 685	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
 686			   dir, DMA_ATTR_SKIP_CPU_SYNC);
 687	return ret;
 688}
 689
 690static int
 691kfd_mem_dmamap_attachment(struct kgd_mem *mem,
 692			  struct kfd_mem_attachment *attachment)
 693{
 694	switch (attachment->type) {
 695	case KFD_MEM_ATT_SHARED:
 696		return 0;
 697	case KFD_MEM_ATT_USERPTR:
 698		return kfd_mem_dmamap_userptr(mem, attachment);
 699	case KFD_MEM_ATT_DMABUF:
 700		return kfd_mem_dmamap_dmabuf(attachment);
 701	case KFD_MEM_ATT_SG:
 702		return kfd_mem_dmamap_sg_bo(mem, attachment);
 703	default:
 704		WARN_ON_ONCE(1);
 705	}
 706	return -EINVAL;
 707}
 708
 709static void
 710kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
 711			 struct kfd_mem_attachment *attachment)
 712{
 713	enum dma_data_direction direction =
 714		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 715		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 716	struct ttm_operation_ctx ctx = {.interruptible = false};
 717	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 718	struct amdgpu_device *adev = attachment->adev;
 719	struct ttm_tt *ttm = bo->tbo.ttm;
 720
 721	if (unlikely(!ttm->sg))
 722		return;
 723
 724	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 725	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 726
 727	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 728	sg_free_table(ttm->sg);
 729	kfree(ttm->sg);
 730	ttm->sg = NULL;
 731}
 732
 733static void
 734kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
 735{
 736	/* This is a no-op. We don't want to trigger eviction fences when
 737	 * unmapping DMABufs. Therefore the invalidation (moving to system
 738	 * domain) is done in kfd_mem_dmamap_dmabuf.
 739	 */
 740}
 741
 742/**
 743 * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
 744 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
 745 * @attachment: Virtual address attachment of the BO on accessing device
 746 *
 747 * The method performs following steps:
 748 *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
 749 *   - Free SG Table that is used to encapsulate DMA mapped memory of
 750 *          peer device's DOORBELL or MMIO memory
 751 *
 752 * This method is invoked in the following contexts:
 753 *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
 754 *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
 755 *
 756 * Return: void
 757 */
 758static void
 759kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
 760		       struct kfd_mem_attachment *attachment)
 761{
 762	struct ttm_operation_ctx ctx = {.interruptible = true};
 763	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
 764	struct amdgpu_device *adev = attachment->adev;
 765	struct ttm_tt *ttm = bo->tbo.ttm;
 766	enum dma_data_direction dir;
 767
 768	if (unlikely(!ttm->sg)) {
 769		pr_debug("SG Table of BO is NULL");
 770		return;
 771	}
 772
 773	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
 774	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 775
 776	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 777				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 778	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
 779			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
 780	sg_free_table(ttm->sg);
 781	kfree(ttm->sg);
 782	ttm->sg = NULL;
 783	bo->tbo.sg = NULL;
 784}
 785
 786static void
 787kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
 788			    struct kfd_mem_attachment *attachment)
 789{
 790	switch (attachment->type) {
 791	case KFD_MEM_ATT_SHARED:
 792		break;
 793	case KFD_MEM_ATT_USERPTR:
 794		kfd_mem_dmaunmap_userptr(mem, attachment);
 795		break;
 796	case KFD_MEM_ATT_DMABUF:
 797		kfd_mem_dmaunmap_dmabuf(attachment);
 798		break;
 799	case KFD_MEM_ATT_SG:
 800		kfd_mem_dmaunmap_sg_bo(mem, attachment);
 801		break;
 802	default:
 803		WARN_ON_ONCE(1);
 804	}
 805}
 806
 807static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
 808{
 809	if (!mem->dmabuf) {
 810		struct amdgpu_device *bo_adev;
 811		struct dma_buf *dmabuf;
 812		int r, fd;
 813
 814		bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 815		r = drm_gem_prime_handle_to_fd(&bo_adev->ddev, bo_adev->kfd.client.file,
 816					       mem->gem_handle,
 817			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
 818					       DRM_RDWR : 0, &fd);
 819		if (r)
 820			return r;
 821		dmabuf = dma_buf_get(fd);
 822		close_fd(fd);
 823		if (WARN_ON_ONCE(IS_ERR(dmabuf)))
 824			return PTR_ERR(dmabuf);
 825		mem->dmabuf = dmabuf;
 826	}
 827
 828	return 0;
 829}
 830
 831static int
 832kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
 833		      struct amdgpu_bo **bo)
 834{
 835	struct drm_gem_object *gobj;
 836	int ret;
 837
 838	ret = kfd_mem_export_dmabuf(mem);
 839	if (ret)
 840		return ret;
 841
 842	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
 843	if (IS_ERR(gobj))
 844		return PTR_ERR(gobj);
 845
 846	*bo = gem_to_amdgpu_bo(gobj);
 847	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
 848
 849	return 0;
 850}
 851
 852/* kfd_mem_attach - Add a BO to a VM
 853 *
 854 * Everything that needs to bo done only once when a BO is first added
 855 * to a VM. It can later be mapped and unmapped many times without
 856 * repeating these steps.
 857 *
 858 * 0. Create BO for DMA mapping, if needed
 859 * 1. Allocate and initialize BO VA entry data structure
 860 * 2. Add BO to the VM
 861 * 3. Determine ASIC-specific PTE flags
 862 * 4. Alloc page tables and directories if needed
 863 * 4a.  Validate new page tables and directories
 864 */
 865static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 866		struct amdgpu_vm *vm, bool is_aql)
 867{
 868	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
 869	unsigned long bo_size = mem->bo->tbo.base.size;
 870	uint64_t va = mem->va;
 871	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
 872	struct amdgpu_bo *bo[2] = {NULL, NULL};
 873	struct amdgpu_bo_va *bo_va;
 874	bool same_hive = false;
 875	int i, ret;
 876
 877	if (!va) {
 878		pr_err("Invalid VA when adding BO to VM\n");
 879		return -EINVAL;
 880	}
 881
 882	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
 883	 *
 884	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
 885	 * In contrast the access path of VRAM BOs depens upon the type of
 886	 * link that connects the peer device. Access over PCIe is allowed
 887	 * if peer device has large BAR. In contrast, access over xGMI is
 888	 * allowed for both small and large BAR configurations of peer device
 889	 */
 890	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
 891	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
 892	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
 893	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
 894		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
 895			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
 896		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
 897			return -EINVAL;
 898	}
 899
 900	for (i = 0; i <= is_aql; i++) {
 901		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
 902		if (unlikely(!attachment[i])) {
 903			ret = -ENOMEM;
 904			goto unwind;
 905		}
 906
 907		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
 908			 va + bo_size, vm);
 909
 910		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
 911		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
 912		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
 913		    same_hive) {
 914			/* Mappings on the local GPU, or VRAM mappings in the
 915			 * local hive, or userptr, or GTT mapping can reuse dma map
 916			 * address space share the original BO
 917			 */
 918			attachment[i]->type = KFD_MEM_ATT_SHARED;
 919			bo[i] = mem->bo;
 920			drm_gem_object_get(&bo[i]->tbo.base);
 921		} else if (i > 0) {
 922			/* Multiple mappings on the same GPU share the BO */
 923			attachment[i]->type = KFD_MEM_ATT_SHARED;
 924			bo[i] = bo[0];
 925			drm_gem_object_get(&bo[i]->tbo.base);
 926		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
 927			/* Create an SG BO to DMA-map userptrs on other GPUs */
 928			attachment[i]->type = KFD_MEM_ATT_USERPTR;
 929			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 930			if (ret)
 931				goto unwind;
 932		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
 933		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
 934			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
 935				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
 936				  "Handing invalid SG BO in ATTACH request");
 937			attachment[i]->type = KFD_MEM_ATT_SG;
 938			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
 939			if (ret)
 940				goto unwind;
 941		/* Enable acces to GTT and VRAM BOs of peer devices */
 942		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
 943			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
 944			attachment[i]->type = KFD_MEM_ATT_DMABUF;
 945			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
 946			if (ret)
 947				goto unwind;
 948			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
 949		} else {
 950			WARN_ONCE(true, "Handling invalid ATTACH request");
 951			ret = -EINVAL;
 952			goto unwind;
 953		}
 954
 955		/* Add BO to VM internal data structures */
 956		ret = amdgpu_bo_reserve(bo[i], false);
 957		if (ret) {
 958			pr_debug("Unable to reserve BO during memory attach");
 959			goto unwind;
 960		}
 961		bo_va = amdgpu_vm_bo_find(vm, bo[i]);
 962		if (!bo_va)
 963			bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
 964		else
 965			++bo_va->ref_count;
 966		attachment[i]->bo_va = bo_va;
 967		amdgpu_bo_unreserve(bo[i]);
 968		if (unlikely(!attachment[i]->bo_va)) {
 969			ret = -ENOMEM;
 970			pr_err("Failed to add BO object to VM. ret == %d\n",
 971			       ret);
 972			goto unwind;
 973		}
 974		attachment[i]->va = va;
 975		attachment[i]->pte_flags = get_pte_flags(adev, mem);
 976		attachment[i]->adev = adev;
 977		list_add(&attachment[i]->list, &mem->attachments);
 978
 979		va += bo_size;
 980	}
 981
 982	return 0;
 983
 984unwind:
 985	for (; i >= 0; i--) {
 986		if (!attachment[i])
 987			continue;
 988		if (attachment[i]->bo_va) {
 989			amdgpu_bo_reserve(bo[i], true);
 990			if (--attachment[i]->bo_va->ref_count == 0)
 991				amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
 992			amdgpu_bo_unreserve(bo[i]);
 993			list_del(&attachment[i]->list);
 994		}
 995		if (bo[i])
 996			drm_gem_object_put(&bo[i]->tbo.base);
 997		kfree(attachment[i]);
 998	}
 999	return ret;
1000}
1001
1002static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
1003{
1004	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
1005
1006	pr_debug("\t remove VA 0x%llx in entry %p\n",
1007			attachment->va, attachment);
1008	if (--attachment->bo_va->ref_count == 0)
1009		amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
1010	drm_gem_object_put(&bo->tbo.base);
1011	list_del(&attachment->list);
1012	kfree(attachment);
1013}
1014
1015static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
1016				struct amdkfd_process_info *process_info,
1017				bool userptr)
1018{
1019	mutex_lock(&process_info->lock);
1020	if (userptr)
1021		list_add_tail(&mem->validate_list,
1022			      &process_info->userptr_valid_list);
1023	else
1024		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
1025	mutex_unlock(&process_info->lock);
1026}
1027
1028static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
1029		struct amdkfd_process_info *process_info)
1030{
1031	mutex_lock(&process_info->lock);
1032	list_del(&mem->validate_list);
1033	mutex_unlock(&process_info->lock);
1034}
1035
1036/* Initializes user pages. It registers the MMU notifier and validates
1037 * the userptr BO in the GTT domain.
1038 *
1039 * The BO must already be on the userptr_valid_list. Otherwise an
1040 * eviction and restore may happen that leaves the new BO unmapped
1041 * with the user mode queues running.
1042 *
1043 * Takes the process_info->lock to protect against concurrent restore
1044 * workers.
1045 *
1046 * Returns 0 for success, negative errno for errors.
1047 */
1048static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
1049			   bool criu_resume)
1050{
1051	struct amdkfd_process_info *process_info = mem->process_info;
1052	struct amdgpu_bo *bo = mem->bo;
1053	struct ttm_operation_ctx ctx = { true, false };
1054	struct hmm_range *range;
1055	int ret = 0;
1056
1057	mutex_lock(&process_info->lock);
1058
1059	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1060	if (ret) {
1061		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1062		goto out;
1063	}
1064
1065	ret = amdgpu_hmm_register(bo, user_addr);
1066	if (ret) {
1067		pr_err("%s: Failed to register MMU notifier: %d\n",
1068		       __func__, ret);
1069		goto out;
1070	}
1071
1072	if (criu_resume) {
1073		/*
1074		 * During a CRIU restore operation, the userptr buffer objects
1075		 * will be validated in the restore_userptr_work worker at a
1076		 * later stage when it is scheduled by another ioctl called by
1077		 * CRIU master process for the target pid for restore.
1078		 */
1079		mutex_lock(&process_info->notifier_lock);
1080		mem->invalid++;
1081		mutex_unlock(&process_info->notifier_lock);
1082		mutex_unlock(&process_info->lock);
1083		return 0;
1084	}
1085
1086	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1087	if (ret) {
1088		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1089		goto unregister_out;
1090	}
1091
1092	ret = amdgpu_bo_reserve(bo, true);
1093	if (ret) {
1094		pr_err("%s: Failed to reserve BO\n", __func__);
1095		goto release_out;
1096	}
1097	amdgpu_bo_placement_from_domain(bo, mem->domain);
1098	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1099	if (ret)
1100		pr_err("%s: failed to validate BO\n", __func__);
1101	amdgpu_bo_unreserve(bo);
1102
1103release_out:
1104	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1105unregister_out:
1106	if (ret)
1107		amdgpu_hmm_unregister(bo);
1108out:
1109	mutex_unlock(&process_info->lock);
1110	return ret;
1111}
1112
1113/* Reserving a BO and its page table BOs must happen atomically to
1114 * avoid deadlocks. Some operations update multiple VMs at once. Track
1115 * all the reservation info in a context structure. Optionally a sync
1116 * object can track VM updates.
1117 */
1118struct bo_vm_reservation_context {
1119	/* DRM execution context for the reservation */
1120	struct drm_exec exec;
1121	/* Number of VMs reserved */
1122	unsigned int n_vms;
1123	/* Pointer to sync object */
1124	struct amdgpu_sync *sync;
1125};
1126
1127enum bo_vm_match {
1128	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1129	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1130	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1131};
1132
1133/**
1134 * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1135 * @mem: KFD BO structure.
1136 * @vm: the VM to reserve.
1137 * @ctx: the struct that will be used in unreserve_bo_and_vms().
1138 */
1139static int reserve_bo_and_vm(struct kgd_mem *mem,
1140			      struct amdgpu_vm *vm,
1141			      struct bo_vm_reservation_context *ctx)
1142{
1143	struct amdgpu_bo *bo = mem->bo;
1144	int ret;
1145
1146	WARN_ON(!vm);
1147
1148	ctx->n_vms = 1;
1149	ctx->sync = &mem->sync;
1150	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1151	drm_exec_until_all_locked(&ctx->exec) {
1152		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1153		drm_exec_retry_on_contention(&ctx->exec);
1154		if (unlikely(ret))
1155			goto error;
1156
1157		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1158		drm_exec_retry_on_contention(&ctx->exec);
1159		if (unlikely(ret))
1160			goto error;
1161	}
1162	return 0;
1163
1164error:
1165	pr_err("Failed to reserve buffers in ttm.\n");
1166	drm_exec_fini(&ctx->exec);
1167	return ret;
1168}
1169
1170/**
1171 * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1172 * @mem: KFD BO structure.
1173 * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1174 * is used. Otherwise, a single VM associated with the BO.
1175 * @map_type: the mapping status that will be used to filter the VMs.
1176 * @ctx: the struct that will be used in unreserve_bo_and_vms().
1177 *
1178 * Returns 0 for success, negative for failure.
1179 */
1180static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1181				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1182				struct bo_vm_reservation_context *ctx)
1183{
1184	struct kfd_mem_attachment *entry;
1185	struct amdgpu_bo *bo = mem->bo;
1186	int ret;
1187
1188	ctx->sync = &mem->sync;
1189	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1190	drm_exec_until_all_locked(&ctx->exec) {
1191		ctx->n_vms = 0;
1192		list_for_each_entry(entry, &mem->attachments, list) {
1193			if ((vm && vm != entry->bo_va->base.vm) ||
1194				(entry->is_mapped != map_type
1195				&& map_type != BO_VM_ALL))
1196				continue;
1197
1198			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1199						&ctx->exec, 2);
1200			drm_exec_retry_on_contention(&ctx->exec);
1201			if (unlikely(ret))
1202				goto error;
1203			++ctx->n_vms;
1204		}
1205
1206		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1207		drm_exec_retry_on_contention(&ctx->exec);
1208		if (unlikely(ret))
1209			goto error;
1210	}
1211	return 0;
1212
1213error:
1214	pr_err("Failed to reserve buffers in ttm.\n");
1215	drm_exec_fini(&ctx->exec);
1216	return ret;
1217}
1218
1219/**
1220 * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1221 * @ctx: Reservation context to unreserve
1222 * @wait: Optionally wait for a sync object representing pending VM updates
1223 * @intr: Whether the wait is interruptible
1224 *
1225 * Also frees any resources allocated in
1226 * reserve_bo_and_(cond_)vm(s). Returns the status from
1227 * amdgpu_sync_wait.
1228 */
1229static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1230				 bool wait, bool intr)
1231{
1232	int ret = 0;
1233
1234	if (wait)
1235		ret = amdgpu_sync_wait(ctx->sync, intr);
1236
1237	drm_exec_fini(&ctx->exec);
1238	ctx->sync = NULL;
1239	return ret;
1240}
1241
1242static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1243				struct kfd_mem_attachment *entry,
1244				struct amdgpu_sync *sync)
1245{
1246	struct amdgpu_bo_va *bo_va = entry->bo_va;
1247	struct amdgpu_device *adev = entry->adev;
1248	struct amdgpu_vm *vm = bo_va->base.vm;
1249
1250	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1251
1252	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1253
1254	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1255}
1256
1257static int update_gpuvm_pte(struct kgd_mem *mem,
1258			    struct kfd_mem_attachment *entry,
1259			    struct amdgpu_sync *sync)
1260{
1261	struct amdgpu_bo_va *bo_va = entry->bo_va;
1262	struct amdgpu_device *adev = entry->adev;
1263	int ret;
1264
1265	ret = kfd_mem_dmamap_attachment(mem, entry);
1266	if (ret)
1267		return ret;
1268
1269	/* Update the page tables  */
1270	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1271	if (ret) {
1272		pr_err("amdgpu_vm_bo_update failed\n");
1273		return ret;
1274	}
1275
1276	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1277}
1278
1279static int map_bo_to_gpuvm(struct kgd_mem *mem,
1280			   struct kfd_mem_attachment *entry,
1281			   struct amdgpu_sync *sync,
1282			   bool no_update_pte)
1283{
1284	int ret;
1285
1286	/* Set virtual address for the allocation */
1287	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1288			       amdgpu_bo_size(entry->bo_va->base.bo),
1289			       entry->pte_flags);
1290	if (ret) {
1291		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1292				entry->va, ret);
1293		return ret;
1294	}
1295
1296	if (no_update_pte)
1297		return 0;
1298
1299	ret = update_gpuvm_pte(mem, entry, sync);
1300	if (ret) {
1301		pr_err("update_gpuvm_pte() failed\n");
1302		goto update_gpuvm_pte_failed;
1303	}
1304
1305	return 0;
1306
1307update_gpuvm_pte_failed:
1308	unmap_bo_from_gpuvm(mem, entry, sync);
1309	kfd_mem_dmaunmap_attachment(mem, entry);
1310	return ret;
1311}
1312
1313static int process_validate_vms(struct amdkfd_process_info *process_info)
 
1314{
1315	struct amdgpu_vm *peer_vm;
1316	int ret;
1317
1318	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1319			    vm_list_node) {
1320		ret = vm_validate_pt_pd_bos(peer_vm);
1321		if (ret)
1322			return ret;
1323	}
1324
1325	return 0;
1326}
1327
1328static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1329				 struct amdgpu_sync *sync)
1330{
1331	struct amdgpu_vm *peer_vm;
1332	int ret;
1333
1334	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1335			    vm_list_node) {
1336		struct amdgpu_bo *pd = peer_vm->root.bo;
1337
1338		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1339				       AMDGPU_SYNC_NE_OWNER,
1340				       AMDGPU_FENCE_OWNER_KFD);
1341		if (ret)
1342			return ret;
1343	}
1344
1345	return 0;
1346}
1347
1348static int process_update_pds(struct amdkfd_process_info *process_info,
1349			      struct amdgpu_sync *sync)
1350{
1351	struct amdgpu_vm *peer_vm;
1352	int ret;
1353
1354	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1355			    vm_list_node) {
1356		ret = vm_update_pds(peer_vm, sync);
1357		if (ret)
1358			return ret;
1359	}
1360
1361	return 0;
1362}
1363
1364static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1365		       struct dma_fence **ef)
1366{
1367	struct amdkfd_process_info *info = NULL;
1368	int ret;
1369
1370	if (!*process_info) {
1371		info = kzalloc(sizeof(*info), GFP_KERNEL);
1372		if (!info)
1373			return -ENOMEM;
1374
1375		mutex_init(&info->lock);
1376		mutex_init(&info->notifier_lock);
1377		INIT_LIST_HEAD(&info->vm_list_head);
1378		INIT_LIST_HEAD(&info->kfd_bo_list);
1379		INIT_LIST_HEAD(&info->userptr_valid_list);
1380		INIT_LIST_HEAD(&info->userptr_inval_list);
1381
1382		info->eviction_fence =
1383			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1384						   current->mm,
1385						   NULL);
1386		if (!info->eviction_fence) {
1387			pr_err("Failed to create eviction fence\n");
1388			ret = -ENOMEM;
1389			goto create_evict_fence_fail;
1390		}
1391
1392		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1393		INIT_DELAYED_WORK(&info->restore_userptr_work,
1394				  amdgpu_amdkfd_restore_userptr_worker);
1395
1396		*process_info = info;
1397	}
1398
1399	vm->process_info = *process_info;
1400
1401	/* Validate page directory and attach eviction fence */
1402	ret = amdgpu_bo_reserve(vm->root.bo, true);
1403	if (ret)
1404		goto reserve_pd_fail;
1405	ret = vm_validate_pt_pd_bos(vm);
1406	if (ret) {
1407		pr_err("validate_pt_pd_bos() failed\n");
1408		goto validate_pd_fail;
1409	}
1410	ret = amdgpu_bo_sync_wait(vm->root.bo,
1411				  AMDGPU_FENCE_OWNER_KFD, false);
1412	if (ret)
1413		goto wait_pd_fail;
1414	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1415	if (ret)
1416		goto reserve_shared_fail;
1417	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1418			   &vm->process_info->eviction_fence->base,
1419			   DMA_RESV_USAGE_BOOKKEEP);
1420	amdgpu_bo_unreserve(vm->root.bo);
1421
1422	/* Update process info */
1423	mutex_lock(&vm->process_info->lock);
1424	list_add_tail(&vm->vm_list_node,
1425			&(vm->process_info->vm_list_head));
1426	vm->process_info->n_vms++;
1427
1428	*ef = dma_fence_get(&vm->process_info->eviction_fence->base);
1429	mutex_unlock(&vm->process_info->lock);
1430
1431	return 0;
1432
1433reserve_shared_fail:
1434wait_pd_fail:
1435validate_pd_fail:
1436	amdgpu_bo_unreserve(vm->root.bo);
1437reserve_pd_fail:
1438	vm->process_info = NULL;
1439	if (info) {
1440		dma_fence_put(&info->eviction_fence->base);
1441		*process_info = NULL;
1442		put_pid(info->pid);
1443create_evict_fence_fail:
1444		mutex_destroy(&info->lock);
1445		mutex_destroy(&info->notifier_lock);
1446		kfree(info);
1447	}
1448	return ret;
1449}
1450
1451/**
1452 * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1453 * @bo: Handle of buffer object being pinned
1454 * @domain: Domain into which BO should be pinned
1455 *
1456 *   - USERPTR BOs are UNPINNABLE and will return error
1457 *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1458 *     PIN count incremented. It is valid to PIN a BO multiple times
1459 *
1460 * Return: ZERO if successful in pinning, Non-Zero in case of error.
1461 */
1462static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1463{
1464	int ret = 0;
1465
1466	ret = amdgpu_bo_reserve(bo, false);
1467	if (unlikely(ret))
1468		return ret;
1469
1470	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1471	if (ret)
1472		pr_err("Error in Pinning BO to domain: %d\n", domain);
1473
1474	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1475	amdgpu_bo_unreserve(bo);
1476
1477	return ret;
1478}
1479
1480/**
1481 * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1482 * @bo: Handle of buffer object being unpinned
1483 *
1484 *   - Is a illegal request for USERPTR BOs and is ignored
1485 *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1486 *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1487 */
1488static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1489{
1490	int ret = 0;
1491
1492	ret = amdgpu_bo_reserve(bo, false);
1493	if (unlikely(ret))
1494		return;
1495
1496	amdgpu_bo_unpin(bo);
1497	amdgpu_bo_unreserve(bo);
1498}
1499
1500int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1501				     struct amdgpu_vm *avm, u32 pasid)
1502
1503{
1504	int ret;
1505
1506	/* Free the original amdgpu allocated pasid,
1507	 * will be replaced with kfd allocated pasid.
1508	 */
1509	if (avm->pasid) {
1510		amdgpu_pasid_free(avm->pasid);
1511		amdgpu_vm_set_pasid(adev, avm, 0);
1512	}
1513
1514	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1515	if (ret)
1516		return ret;
1517
1518	return 0;
1519}
1520
1521int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1522					   struct amdgpu_vm *avm,
1523					   void **process_info,
1524					   struct dma_fence **ef)
1525{
1526	int ret;
1527
1528	/* Already a compute VM? */
1529	if (avm->process_info)
1530		return -EINVAL;
1531
1532	/* Convert VM into a compute VM */
1533	ret = amdgpu_vm_make_compute(adev, avm);
1534	if (ret)
1535		return ret;
1536
1537	/* Initialize KFD part of the VM and process info */
1538	ret = init_kfd_vm(avm, process_info, ef);
1539	if (ret)
1540		return ret;
1541
1542	amdgpu_vm_set_task_info(avm);
1543
1544	return 0;
1545}
1546
1547void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1548				    struct amdgpu_vm *vm)
1549{
1550	struct amdkfd_process_info *process_info = vm->process_info;
1551
1552	if (!process_info)
1553		return;
1554
1555	/* Update process info */
1556	mutex_lock(&process_info->lock);
1557	process_info->n_vms--;
1558	list_del(&vm->vm_list_node);
1559	mutex_unlock(&process_info->lock);
1560
1561	vm->process_info = NULL;
1562
1563	/* Release per-process resources when last compute VM is destroyed */
1564	if (!process_info->n_vms) {
1565		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1566		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1567		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1568
1569		dma_fence_put(&process_info->eviction_fence->base);
1570		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1571		put_pid(process_info->pid);
1572		mutex_destroy(&process_info->lock);
1573		mutex_destroy(&process_info->notifier_lock);
1574		kfree(process_info);
1575	}
1576}
1577
1578void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1579					    void *drm_priv)
1580{
1581	struct amdgpu_vm *avm;
1582
1583	if (WARN_ON(!adev || !drm_priv))
1584		return;
1585
1586	avm = drm_priv_to_vm(drm_priv);
1587
1588	pr_debug("Releasing process vm %p\n", avm);
1589
1590	/* The original pasid of amdgpu vm has already been
1591	 * released during making a amdgpu vm to a compute vm
1592	 * The current pasid is managed by kfd and will be
1593	 * released on kfd process destroy. Set amdgpu pasid
1594	 * to 0 to avoid duplicate release.
1595	 */
1596	amdgpu_vm_release_compute(adev, avm);
1597}
1598
1599uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1600{
1601	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1602	struct amdgpu_bo *pd = avm->root.bo;
1603	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1604
1605	if (adev->asic_type < CHIP_VEGA10)
1606		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1607	return avm->pd_phys_addr;
1608}
1609
1610void amdgpu_amdkfd_block_mmu_notifications(void *p)
1611{
1612	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1613
1614	mutex_lock(&pinfo->lock);
1615	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1616	mutex_unlock(&pinfo->lock);
1617}
1618
1619int amdgpu_amdkfd_criu_resume(void *p)
1620{
1621	int ret = 0;
1622	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1623
1624	mutex_lock(&pinfo->lock);
1625	pr_debug("scheduling work\n");
1626	mutex_lock(&pinfo->notifier_lock);
1627	pinfo->evicted_bos++;
1628	mutex_unlock(&pinfo->notifier_lock);
1629	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1630		ret = -EINVAL;
1631		goto out_unlock;
1632	}
1633	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1634	queue_delayed_work(system_freezable_wq,
1635			   &pinfo->restore_userptr_work, 0);
1636
1637out_unlock:
1638	mutex_unlock(&pinfo->lock);
1639	return ret;
1640}
1641
1642size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1643					  uint8_t xcp_id)
1644{
1645	uint64_t reserved_for_pt =
1646		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1647	ssize_t available;
1648	uint64_t vram_available, system_mem_available, ttm_mem_available;
1649
1650	spin_lock(&kfd_mem_limit.mem_limit_lock);
1651	vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1652		- adev->kfd.vram_used_aligned[xcp_id]
1653		- atomic64_read(&adev->vram_pin_size)
1654		- reserved_for_pt;
1655
1656	if (adev->gmc.is_app_apu) {
1657		system_mem_available = no_system_mem_limit ?
1658					kfd_mem_limit.max_system_mem_limit :
1659					kfd_mem_limit.max_system_mem_limit -
1660					kfd_mem_limit.system_mem_used;
1661
1662		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1663				kfd_mem_limit.ttm_mem_used;
1664
1665		available = min3(system_mem_available, ttm_mem_available,
1666				 vram_available);
1667		available = ALIGN_DOWN(available, PAGE_SIZE);
1668	} else {
1669		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1670	}
1671
1672	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1673
1674	if (available < 0)
1675		available = 0;
1676
1677	return available;
1678}
1679
1680int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1681		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1682		void *drm_priv, struct kgd_mem **mem,
1683		uint64_t *offset, uint32_t flags, bool criu_resume)
1684{
1685	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1686	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1687	enum ttm_bo_type bo_type = ttm_bo_type_device;
1688	struct sg_table *sg = NULL;
1689	uint64_t user_addr = 0;
1690	struct amdgpu_bo *bo;
1691	struct drm_gem_object *gobj = NULL;
1692	u32 domain, alloc_domain;
1693	uint64_t aligned_size;
1694	int8_t xcp_id = -1;
1695	u64 alloc_flags;
1696	int ret;
1697
1698	/*
1699	 * Check on which domain to allocate BO
1700	 */
1701	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1702		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1703
1704		if (adev->gmc.is_app_apu) {
1705			domain = AMDGPU_GEM_DOMAIN_GTT;
1706			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1707			alloc_flags = 0;
1708		} else {
1709			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1710			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1711			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1712		}
1713		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1714					0 : fpriv->xcp_id;
1715	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1716		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1717		alloc_flags = 0;
1718	} else {
1719		domain = AMDGPU_GEM_DOMAIN_GTT;
1720		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1721		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1722
1723		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1724			if (!offset || !*offset)
1725				return -EINVAL;
1726			user_addr = untagged_addr(*offset);
1727		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1728				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1729			bo_type = ttm_bo_type_sg;
1730			if (size > UINT_MAX)
1731				return -EINVAL;
1732			sg = create_sg_table(*offset, size);
1733			if (!sg)
1734				return -ENOMEM;
1735		} else {
1736			return -EINVAL;
1737		}
1738	}
1739
1740	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1741		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1742	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1743		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1744	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1745		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1746
1747	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1748	if (!*mem) {
1749		ret = -ENOMEM;
1750		goto err;
1751	}
1752	INIT_LIST_HEAD(&(*mem)->attachments);
1753	mutex_init(&(*mem)->lock);
1754	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1755
1756	/* Workaround for AQL queue wraparound bug. Map the same
1757	 * memory twice. That means we only actually allocate half
1758	 * the memory.
1759	 */
1760	if ((*mem)->aql_queue)
1761		size >>= 1;
1762	aligned_size = PAGE_ALIGN(size);
1763
1764	(*mem)->alloc_flags = flags;
1765
1766	amdgpu_sync_create(&(*mem)->sync);
1767
1768	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1769					      xcp_id);
1770	if (ret) {
1771		pr_debug("Insufficient memory\n");
1772		goto err_reserve_limit;
1773	}
1774
1775	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1776		 va, (*mem)->aql_queue ? size << 1 : size,
1777		 domain_string(alloc_domain), xcp_id);
1778
1779	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1780				       bo_type, NULL, &gobj, xcp_id + 1);
1781	if (ret) {
1782		pr_debug("Failed to create BO on domain %s. ret %d\n",
1783			 domain_string(alloc_domain), ret);
1784		goto err_bo_create;
1785	}
1786	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1787	if (ret) {
1788		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1789		goto err_node_allow;
1790	}
1791	ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
1792	if (ret)
1793		goto err_gem_handle_create;
1794	bo = gem_to_amdgpu_bo(gobj);
1795	if (bo_type == ttm_bo_type_sg) {
1796		bo->tbo.sg = sg;
1797		bo->tbo.ttm->sg = sg;
1798	}
1799	bo->kfd_bo = *mem;
1800	(*mem)->bo = bo;
1801	if (user_addr)
1802		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1803
1804	(*mem)->va = va;
1805	(*mem)->domain = domain;
1806	(*mem)->mapped_to_gpu_memory = 0;
1807	(*mem)->process_info = avm->process_info;
1808
1809	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1810
1811	if (user_addr) {
1812		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1813		ret = init_user_pages(*mem, user_addr, criu_resume);
1814		if (ret)
1815			goto allocate_init_user_pages_failed;
1816	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1817				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1818		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1819		if (ret) {
1820			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1821			goto err_pin_bo;
1822		}
1823		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1824		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1825	} else {
1826		mutex_lock(&avm->process_info->lock);
1827		if (avm->process_info->eviction_fence &&
1828		    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
1829			ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
1830				&avm->process_info->eviction_fence->base);
1831		mutex_unlock(&avm->process_info->lock);
1832		if (ret)
1833			goto err_validate_bo;
1834	}
1835
1836	if (offset)
1837		*offset = amdgpu_bo_mmap_offset(bo);
1838
1839	return 0;
1840
1841allocate_init_user_pages_failed:
1842err_pin_bo:
1843err_validate_bo:
1844	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1845	drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
1846err_gem_handle_create:
1847	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1848err_node_allow:
1849	/* Don't unreserve system mem limit twice */
1850	goto err_reserve_limit;
1851err_bo_create:
1852	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1853err_reserve_limit:
 
1854	mutex_destroy(&(*mem)->lock);
1855	if (gobj)
1856		drm_gem_object_put(gobj);
1857	else
1858		kfree(*mem);
1859err:
1860	if (sg) {
1861		sg_free_table(sg);
1862		kfree(sg);
1863	}
1864	return ret;
1865}
1866
1867int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1868		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1869		uint64_t *size)
1870{
1871	struct amdkfd_process_info *process_info = mem->process_info;
1872	unsigned long bo_size = mem->bo->tbo.base.size;
1873	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1874	struct kfd_mem_attachment *entry, *tmp;
1875	struct bo_vm_reservation_context ctx;
1876	unsigned int mapped_to_gpu_memory;
1877	int ret;
1878	bool is_imported = false;
1879
1880	mutex_lock(&mem->lock);
1881
1882	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1883	if (mem->alloc_flags &
1884	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1885	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1886		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1887	}
1888
1889	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1890	is_imported = mem->is_imported;
1891	mutex_unlock(&mem->lock);
1892	/* lock is not needed after this, since mem is unused and will
1893	 * be freed anyway
1894	 */
1895
1896	if (mapped_to_gpu_memory > 0) {
1897		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1898				mem->va, bo_size);
1899		return -EBUSY;
1900	}
1901
1902	/* Make sure restore workers don't access the BO any more */
1903	mutex_lock(&process_info->lock);
1904	list_del(&mem->validate_list);
1905	mutex_unlock(&process_info->lock);
1906
1907	/* Cleanup user pages and MMU notifiers */
1908	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1909		amdgpu_hmm_unregister(mem->bo);
1910		mutex_lock(&process_info->notifier_lock);
1911		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1912		mutex_unlock(&process_info->notifier_lock);
1913	}
1914
1915	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1916	if (unlikely(ret))
1917		return ret;
1918
1919	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1920					process_info->eviction_fence);
1921	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1922		mem->va + bo_size * (1 + mem->aql_queue));
1923
1924	/* Remove from VM internal data structures */
1925	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
1926		kfd_mem_dmaunmap_attachment(mem, entry);
1927		kfd_mem_detach(entry);
1928	}
1929
1930	ret = unreserve_bo_and_vms(&ctx, false, false);
1931
1932	/* Free the sync object */
1933	amdgpu_sync_free(&mem->sync);
1934
1935	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1936	 * remap BO. We need to free it.
1937	 */
1938	if (mem->bo->tbo.sg) {
1939		sg_free_table(mem->bo->tbo.sg);
1940		kfree(mem->bo->tbo.sg);
1941	}
1942
1943	/* Update the size of the BO being freed if it was allocated from
1944	 * VRAM and is not imported. For APP APU VRAM allocations are done
1945	 * in GTT domain
1946	 */
1947	if (size) {
1948		if (!is_imported &&
1949		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1950		   (adev->gmc.is_app_apu &&
1951		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1952			*size = bo_size;
1953		else
1954			*size = 0;
1955	}
1956
1957	/* Free the BO*/
1958	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1959	drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
1960	if (mem->dmabuf) {
1961		dma_buf_put(mem->dmabuf);
1962		mem->dmabuf = NULL;
1963	}
1964	mutex_destroy(&mem->lock);
1965
1966	/* If this releases the last reference, it will end up calling
1967	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1968	 * this needs to be the last call here.
1969	 */
1970	drm_gem_object_put(&mem->bo->tbo.base);
1971
1972	/*
1973	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1974	 * explicitly free it here.
1975	 */
1976	if (!use_release_notifier)
1977		kfree(mem);
1978
1979	return ret;
1980}
1981
1982int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1983		struct amdgpu_device *adev, struct kgd_mem *mem,
1984		void *drm_priv)
1985{
1986	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1987	int ret;
1988	struct amdgpu_bo *bo;
1989	uint32_t domain;
1990	struct kfd_mem_attachment *entry;
1991	struct bo_vm_reservation_context ctx;
1992	unsigned long bo_size;
1993	bool is_invalid_userptr = false;
1994
1995	bo = mem->bo;
1996	if (!bo) {
1997		pr_err("Invalid BO when mapping memory to GPU\n");
1998		return -EINVAL;
1999	}
2000
2001	/* Make sure restore is not running concurrently. Since we
2002	 * don't map invalid userptr BOs, we rely on the next restore
2003	 * worker to do the mapping
2004	 */
2005	mutex_lock(&mem->process_info->lock);
2006
2007	/* Lock notifier lock. If we find an invalid userptr BO, we can be
2008	 * sure that the MMU notifier is no longer running
2009	 * concurrently and the queues are actually stopped
2010	 */
2011	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2012		mutex_lock(&mem->process_info->notifier_lock);
2013		is_invalid_userptr = !!mem->invalid;
2014		mutex_unlock(&mem->process_info->notifier_lock);
2015	}
2016
2017	mutex_lock(&mem->lock);
2018
2019	domain = mem->domain;
2020	bo_size = bo->tbo.base.size;
2021
2022	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
2023			mem->va,
2024			mem->va + bo_size * (1 + mem->aql_queue),
2025			avm, domain_string(domain));
2026
2027	if (!kfd_mem_is_attached(avm, mem)) {
2028		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
2029		if (ret)
2030			goto out;
2031	}
2032
2033	ret = reserve_bo_and_vm(mem, avm, &ctx);
2034	if (unlikely(ret))
2035		goto out;
2036
2037	/* Userptr can be marked as "not invalid", but not actually be
2038	 * validated yet (still in the system domain). In that case
2039	 * the queues are still stopped and we can leave mapping for
2040	 * the next restore worker
2041	 */
2042	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
2043	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
2044		is_invalid_userptr = true;
2045
2046	ret = vm_validate_pt_pd_bos(avm);
2047	if (unlikely(ret))
2048		goto out_unreserve;
2049
2050	list_for_each_entry(entry, &mem->attachments, list) {
2051		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2052			continue;
2053
2054		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2055			 entry->va, entry->va + bo_size, entry);
2056
2057		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2058				      is_invalid_userptr);
2059		if (ret) {
2060			pr_err("Failed to map bo to gpuvm\n");
2061			goto out_unreserve;
2062		}
2063
2064		ret = vm_update_pds(avm, ctx.sync);
2065		if (ret) {
2066			pr_err("Failed to update page directories\n");
2067			goto out_unreserve;
2068		}
2069
2070		entry->is_mapped = true;
2071		mem->mapped_to_gpu_memory++;
2072		pr_debug("\t INC mapping count %d\n",
2073			 mem->mapped_to_gpu_memory);
2074	}
2075
2076	ret = unreserve_bo_and_vms(&ctx, false, false);
2077
2078	goto out;
2079
2080out_unreserve:
2081	unreserve_bo_and_vms(&ctx, false, false);
2082out:
2083	mutex_unlock(&mem->process_info->lock);
2084	mutex_unlock(&mem->lock);
2085	return ret;
2086}
2087
2088int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv)
2089{
2090	struct kfd_mem_attachment *entry;
2091	struct amdgpu_vm *vm;
2092	int ret;
2093
2094	vm = drm_priv_to_vm(drm_priv);
2095
2096	mutex_lock(&mem->lock);
2097
2098	ret = amdgpu_bo_reserve(mem->bo, true);
2099	if (ret)
2100		goto out;
2101
2102	list_for_each_entry(entry, &mem->attachments, list) {
2103		if (entry->bo_va->base.vm != vm)
2104			continue;
2105		if (entry->bo_va->base.bo->tbo.ttm &&
2106		    !entry->bo_va->base.bo->tbo.ttm->sg)
2107			continue;
2108
2109		kfd_mem_dmaunmap_attachment(mem, entry);
2110	}
2111
2112	amdgpu_bo_unreserve(mem->bo);
2113out:
2114	mutex_unlock(&mem->lock);
2115
2116	return ret;
2117}
2118
2119int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2120		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2121{
2122	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2123	unsigned long bo_size = mem->bo->tbo.base.size;
2124	struct kfd_mem_attachment *entry;
2125	struct bo_vm_reservation_context ctx;
2126	int ret;
2127
2128	mutex_lock(&mem->lock);
2129
2130	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2131	if (unlikely(ret))
2132		goto out;
2133	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2134	if (ctx.n_vms == 0) {
2135		ret = -EINVAL;
2136		goto unreserve_out;
2137	}
2138
2139	ret = vm_validate_pt_pd_bos(avm);
2140	if (unlikely(ret))
2141		goto unreserve_out;
2142
2143	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2144		mem->va,
2145		mem->va + bo_size * (1 + mem->aql_queue),
2146		avm);
2147
2148	list_for_each_entry(entry, &mem->attachments, list) {
2149		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2150			continue;
2151
2152		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2153			 entry->va, entry->va + bo_size, entry);
2154
2155		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2156		entry->is_mapped = false;
2157
2158		mem->mapped_to_gpu_memory--;
2159		pr_debug("\t DEC mapping count %d\n",
2160			 mem->mapped_to_gpu_memory);
2161	}
2162
2163unreserve_out:
2164	unreserve_bo_and_vms(&ctx, false, false);
2165out:
2166	mutex_unlock(&mem->lock);
2167	return ret;
2168}
2169
2170int amdgpu_amdkfd_gpuvm_sync_memory(
2171		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2172{
2173	struct amdgpu_sync sync;
2174	int ret;
2175
2176	amdgpu_sync_create(&sync);
2177
2178	mutex_lock(&mem->lock);
2179	amdgpu_sync_clone(&mem->sync, &sync);
2180	mutex_unlock(&mem->lock);
2181
2182	ret = amdgpu_sync_wait(&sync, intr);
2183	amdgpu_sync_free(&sync);
2184	return ret;
2185}
2186
2187/**
2188 * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2189 * @adev: Device to which allocated BO belongs
2190 * @bo: Buffer object to be mapped
2191 *
2192 * Before return, bo reference count is incremented. To release the reference and unpin/
2193 * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2194 */
2195int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
2196{
2197	int ret;
2198
2199	ret = amdgpu_bo_reserve(bo, true);
2200	if (ret) {
2201		pr_err("Failed to reserve bo. ret %d\n", ret);
2202		goto err_reserve_bo_failed;
2203	}
2204
2205	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2206	if (ret) {
2207		pr_err("Failed to pin bo. ret %d\n", ret);
2208		goto err_pin_bo_failed;
2209	}
2210
2211	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2212	if (ret) {
2213		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2214		goto err_map_bo_gart_failed;
2215	}
2216
2217	amdgpu_amdkfd_remove_eviction_fence(
2218		bo, bo->vm_bo->vm->process_info->eviction_fence);
2219
2220	amdgpu_bo_unreserve(bo);
2221
2222	bo = amdgpu_bo_ref(bo);
2223
2224	return 0;
2225
2226err_map_bo_gart_failed:
2227	amdgpu_bo_unpin(bo);
2228err_pin_bo_failed:
2229	amdgpu_bo_unreserve(bo);
2230err_reserve_bo_failed:
2231
2232	return ret;
2233}
2234
2235/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2236 *
2237 * @mem: Buffer object to be mapped for CPU access
2238 * @kptr[out]: pointer in kernel CPU address space
2239 * @size[out]: size of the buffer
2240 *
2241 * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2242 * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2243 * validate_list, so the GPU mapping can be restored after a page table was
2244 * evicted.
2245 *
2246 * Return: 0 on success, error code on failure
2247 */
2248int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2249					     void **kptr, uint64_t *size)
2250{
2251	int ret;
2252	struct amdgpu_bo *bo = mem->bo;
2253
2254	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2255		pr_err("userptr can't be mapped to kernel\n");
2256		return -EINVAL;
2257	}
2258
2259	mutex_lock(&mem->process_info->lock);
2260
2261	ret = amdgpu_bo_reserve(bo, true);
2262	if (ret) {
2263		pr_err("Failed to reserve bo. ret %d\n", ret);
2264		goto bo_reserve_failed;
2265	}
2266
2267	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2268	if (ret) {
2269		pr_err("Failed to pin bo. ret %d\n", ret);
2270		goto pin_failed;
2271	}
2272
2273	ret = amdgpu_bo_kmap(bo, kptr);
2274	if (ret) {
2275		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2276		goto kmap_failed;
2277	}
2278
2279	amdgpu_amdkfd_remove_eviction_fence(
2280		bo, mem->process_info->eviction_fence);
2281
2282	if (size)
2283		*size = amdgpu_bo_size(bo);
2284
2285	amdgpu_bo_unreserve(bo);
2286
2287	mutex_unlock(&mem->process_info->lock);
2288	return 0;
2289
2290kmap_failed:
2291	amdgpu_bo_unpin(bo);
2292pin_failed:
2293	amdgpu_bo_unreserve(bo);
2294bo_reserve_failed:
2295	mutex_unlock(&mem->process_info->lock);
2296
2297	return ret;
2298}
2299
2300/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2301 *
2302 * @mem: Buffer object to be unmapped for CPU access
2303 *
2304 * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2305 * eviction fence, so this function should only be used for cleanup before the
2306 * BO is destroyed.
2307 */
2308void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2309{
2310	struct amdgpu_bo *bo = mem->bo;
2311
2312	amdgpu_bo_reserve(bo, true);
2313	amdgpu_bo_kunmap(bo);
2314	amdgpu_bo_unpin(bo);
2315	amdgpu_bo_unreserve(bo);
2316}
2317
2318int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2319					  struct kfd_vm_fault_info *mem)
2320{
2321	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2322		*mem = *adev->gmc.vm_fault_info;
2323		mb(); /* make sure read happened */
2324		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2325	}
2326	return 0;
2327}
2328
2329static int import_obj_create(struct amdgpu_device *adev,
2330			     struct dma_buf *dma_buf,
2331			     struct drm_gem_object *obj,
2332			     uint64_t va, void *drm_priv,
2333			     struct kgd_mem **mem, uint64_t *size,
2334			     uint64_t *mmap_offset)
2335{
2336	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2337	struct amdgpu_bo *bo;
2338	int ret;
2339
2340	bo = gem_to_amdgpu_bo(obj);
2341	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2342				    AMDGPU_GEM_DOMAIN_GTT)))
2343		/* Only VRAM and GTT BOs are supported */
2344		return -EINVAL;
2345
2346	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2347	if (!*mem)
2348		return -ENOMEM;
2349
2350	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2351	if (ret)
2352		goto err_free_mem;
2353
2354	if (size)
2355		*size = amdgpu_bo_size(bo);
2356
2357	if (mmap_offset)
2358		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2359
2360	INIT_LIST_HEAD(&(*mem)->attachments);
2361	mutex_init(&(*mem)->lock);
2362
2363	(*mem)->alloc_flags =
2364		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2365		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2366		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2367		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2368
2369	get_dma_buf(dma_buf);
2370	(*mem)->dmabuf = dma_buf;
2371	(*mem)->bo = bo;
2372	(*mem)->va = va;
2373	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
2374		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2375
2376	(*mem)->mapped_to_gpu_memory = 0;
2377	(*mem)->process_info = avm->process_info;
2378	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2379	amdgpu_sync_create(&(*mem)->sync);
2380	(*mem)->is_imported = true;
2381
2382	mutex_lock(&avm->process_info->lock);
2383	if (avm->process_info->eviction_fence &&
2384	    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
2385		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
2386				&avm->process_info->eviction_fence->base);
2387	mutex_unlock(&avm->process_info->lock);
2388	if (ret)
2389		goto err_remove_mem;
2390
2391	return 0;
2392
2393err_remove_mem:
2394	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
2395	drm_vma_node_revoke(&obj->vma_node, drm_priv);
2396err_free_mem:
2397	kfree(*mem);
2398	return ret;
2399}
2400
2401int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
2402					 uint64_t va, void *drm_priv,
2403					 struct kgd_mem **mem, uint64_t *size,
2404					 uint64_t *mmap_offset)
2405{
2406	struct drm_gem_object *obj;
2407	uint32_t handle;
2408	int ret;
2409
2410	ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
2411					 &handle);
2412	if (ret)
2413		return ret;
2414	obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
2415	if (!obj) {
2416		ret = -EINVAL;
2417		goto err_release_handle;
2418	}
2419
2420	ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
2421				mmap_offset);
2422	if (ret)
2423		goto err_put_obj;
2424
2425	(*mem)->gem_handle = handle;
2426
2427	return 0;
2428
2429err_put_obj:
2430	drm_gem_object_put(obj);
2431err_release_handle:
2432	drm_gem_handle_delete(adev->kfd.client.file, handle);
2433	return ret;
2434}
2435
2436int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2437				      struct dma_buf **dma_buf)
2438{
2439	int ret;
2440
2441	mutex_lock(&mem->lock);
2442	ret = kfd_mem_export_dmabuf(mem);
2443	if (ret)
2444		goto out;
2445
2446	get_dma_buf(mem->dmabuf);
2447	*dma_buf = mem->dmabuf;
2448out:
2449	mutex_unlock(&mem->lock);
2450	return ret;
2451}
2452
2453/* Evict a userptr BO by stopping the queues if necessary
2454 *
2455 * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2456 * cannot do any memory allocations, and cannot take any locks that
2457 * are held elsewhere while allocating memory.
2458 *
2459 * It doesn't do anything to the BO itself. The real work happens in
2460 * restore, where we get updated page addresses. This function only
2461 * ensures that GPU access to the BO is stopped.
2462 */
2463int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2464				unsigned long cur_seq, struct kgd_mem *mem)
2465{
2466	struct amdkfd_process_info *process_info = mem->process_info;
2467	int r = 0;
2468
2469	/* Do not process MMU notifications during CRIU restore until
2470	 * KFD_CRIU_OP_RESUME IOCTL is received
2471	 */
2472	if (READ_ONCE(process_info->block_mmu_notifications))
2473		return 0;
2474
2475	mutex_lock(&process_info->notifier_lock);
2476	mmu_interval_set_seq(mni, cur_seq);
2477
2478	mem->invalid++;
2479	if (++process_info->evicted_bos == 1) {
2480		/* First eviction, stop the queues */
2481		r = kgd2kfd_quiesce_mm(mni->mm,
2482				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2483		if (r)
2484			pr_err("Failed to quiesce KFD\n");
2485		queue_delayed_work(system_freezable_wq,
2486			&process_info->restore_userptr_work,
2487			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2488	}
2489	mutex_unlock(&process_info->notifier_lock);
2490
2491	return r;
2492}
2493
2494/* Update invalid userptr BOs
2495 *
2496 * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2497 * userptr_inval_list and updates user pages for all BOs that have
2498 * been invalidated since their last update.
2499 */
2500static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2501				     struct mm_struct *mm)
2502{
2503	struct kgd_mem *mem, *tmp_mem;
2504	struct amdgpu_bo *bo;
2505	struct ttm_operation_ctx ctx = { false, false };
2506	uint32_t invalid;
2507	int ret = 0;
2508
2509	mutex_lock(&process_info->notifier_lock);
2510
2511	/* Move all invalidated BOs to the userptr_inval_list */
2512	list_for_each_entry_safe(mem, tmp_mem,
2513				 &process_info->userptr_valid_list,
2514				 validate_list)
2515		if (mem->invalid)
2516			list_move_tail(&mem->validate_list,
2517				       &process_info->userptr_inval_list);
2518
2519	/* Go through userptr_inval_list and update any invalid user_pages */
2520	list_for_each_entry(mem, &process_info->userptr_inval_list,
2521			    validate_list) {
2522		invalid = mem->invalid;
2523		if (!invalid)
2524			/* BO hasn't been invalidated since the last
2525			 * revalidation attempt. Keep its page list.
2526			 */
2527			continue;
2528
2529		bo = mem->bo;
2530
2531		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2532		mem->range = NULL;
2533
2534		/* BO reservations and getting user pages (hmm_range_fault)
2535		 * must happen outside the notifier lock
2536		 */
2537		mutex_unlock(&process_info->notifier_lock);
2538
2539		/* Move the BO to system (CPU) domain if necessary to unmap
2540		 * and free the SG table
2541		 */
2542		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2543			if (amdgpu_bo_reserve(bo, true))
2544				return -EAGAIN;
2545			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2546			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2547			amdgpu_bo_unreserve(bo);
2548			if (ret) {
2549				pr_err("%s: Failed to invalidate userptr BO\n",
2550				       __func__);
2551				return -EAGAIN;
2552			}
2553		}
2554
2555		/* Get updated user pages */
2556		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2557						   &mem->range);
2558		if (ret) {
2559			pr_debug("Failed %d to get user pages\n", ret);
2560
2561			/* Return -EFAULT bad address error as success. It will
2562			 * fail later with a VM fault if the GPU tries to access
2563			 * it. Better than hanging indefinitely with stalled
2564			 * user mode queues.
2565			 *
2566			 * Return other error -EBUSY or -ENOMEM to retry restore
2567			 */
2568			if (ret != -EFAULT)
2569				return ret;
2570
2571			ret = 0;
2572		}
2573
2574		mutex_lock(&process_info->notifier_lock);
2575
2576		/* Mark the BO as valid unless it was invalidated
2577		 * again concurrently.
2578		 */
2579		if (mem->invalid != invalid) {
2580			ret = -EAGAIN;
2581			goto unlock_out;
2582		}
2583		 /* set mem valid if mem has hmm range associated */
2584		if (mem->range)
2585			mem->invalid = 0;
2586	}
2587
2588unlock_out:
2589	mutex_unlock(&process_info->notifier_lock);
2590
2591	return ret;
2592}
2593
2594/* Validate invalid userptr BOs
2595 *
2596 * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2597 * with new page addresses and waits for the page table updates to complete.
2598 */
2599static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2600{
2601	struct ttm_operation_ctx ctx = { false, false };
2602	struct amdgpu_sync sync;
2603	struct drm_exec exec;
2604
2605	struct amdgpu_vm *peer_vm;
2606	struct kgd_mem *mem, *tmp_mem;
2607	struct amdgpu_bo *bo;
2608	int ret;
2609
2610	amdgpu_sync_create(&sync);
2611
2612	drm_exec_init(&exec, 0, 0);
2613	/* Reserve all BOs and page tables for validation */
2614	drm_exec_until_all_locked(&exec) {
2615		/* Reserve all the page directories */
2616		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2617				    vm_list_node) {
2618			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2619			drm_exec_retry_on_contention(&exec);
2620			if (unlikely(ret))
2621				goto unreserve_out;
2622		}
2623
2624		/* Reserve the userptr_inval_list entries to resv_list */
2625		list_for_each_entry(mem, &process_info->userptr_inval_list,
2626				    validate_list) {
2627			struct drm_gem_object *gobj;
2628
2629			gobj = &mem->bo->tbo.base;
2630			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2631			drm_exec_retry_on_contention(&exec);
2632			if (unlikely(ret))
2633				goto unreserve_out;
2634		}
2635	}
2636
2637	ret = process_validate_vms(process_info);
2638	if (ret)
2639		goto unreserve_out;
2640
2641	/* Validate BOs and update GPUVM page tables */
2642	list_for_each_entry_safe(mem, tmp_mem,
2643				 &process_info->userptr_inval_list,
2644				 validate_list) {
2645		struct kfd_mem_attachment *attachment;
2646
2647		bo = mem->bo;
2648
2649		/* Validate the BO if we got user pages */
2650		if (bo->tbo.ttm->pages[0]) {
2651			amdgpu_bo_placement_from_domain(bo, mem->domain);
2652			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2653			if (ret) {
2654				pr_err("%s: failed to validate BO\n", __func__);
2655				goto unreserve_out;
2656			}
2657		}
2658
2659		/* Update mapping. If the BO was not validated
2660		 * (because we couldn't get user pages), this will
2661		 * clear the page table entries, which will result in
2662		 * VM faults if the GPU tries to access the invalid
2663		 * memory.
2664		 */
2665		list_for_each_entry(attachment, &mem->attachments, list) {
2666			if (!attachment->is_mapped)
2667				continue;
2668
2669			kfd_mem_dmaunmap_attachment(mem, attachment);
2670			ret = update_gpuvm_pte(mem, attachment, &sync);
2671			if (ret) {
2672				pr_err("%s: update PTE failed\n", __func__);
2673				/* make sure this gets validated again */
2674				mutex_lock(&process_info->notifier_lock);
2675				mem->invalid++;
2676				mutex_unlock(&process_info->notifier_lock);
2677				goto unreserve_out;
2678			}
2679		}
2680	}
2681
2682	/* Update page directories */
2683	ret = process_update_pds(process_info, &sync);
2684
2685unreserve_out:
2686	drm_exec_fini(&exec);
2687	amdgpu_sync_wait(&sync, false);
2688	amdgpu_sync_free(&sync);
2689
2690	return ret;
2691}
2692
2693/* Confirm that all user pages are valid while holding the notifier lock
2694 *
2695 * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2696 */
2697static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2698{
2699	struct kgd_mem *mem, *tmp_mem;
2700	int ret = 0;
2701
2702	list_for_each_entry_safe(mem, tmp_mem,
2703				 &process_info->userptr_inval_list,
2704				 validate_list) {
2705		bool valid;
2706
2707		/* keep mem without hmm range at userptr_inval_list */
2708		if (!mem->range)
2709			 continue;
2710
2711		/* Only check mem with hmm range associated */
2712		valid = amdgpu_ttm_tt_get_user_pages_done(
2713					mem->bo->tbo.ttm, mem->range);
2714
2715		mem->range = NULL;
2716		if (!valid) {
2717			WARN(!mem->invalid, "Invalid BO not marked invalid");
2718			ret = -EAGAIN;
2719			continue;
2720		}
2721
2722		if (mem->invalid) {
2723			WARN(1, "Valid BO is marked invalid");
2724			ret = -EAGAIN;
2725			continue;
2726		}
2727
2728		list_move_tail(&mem->validate_list,
2729			       &process_info->userptr_valid_list);
2730	}
2731
2732	return ret;
2733}
2734
2735/* Worker callback to restore evicted userptr BOs
2736 *
2737 * Tries to update and validate all userptr BOs. If successful and no
2738 * concurrent evictions happened, the queues are restarted. Otherwise,
2739 * reschedule for another attempt later.
2740 */
2741static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2742{
2743	struct delayed_work *dwork = to_delayed_work(work);
2744	struct amdkfd_process_info *process_info =
2745		container_of(dwork, struct amdkfd_process_info,
2746			     restore_userptr_work);
2747	struct task_struct *usertask;
2748	struct mm_struct *mm;
2749	uint32_t evicted_bos;
2750
2751	mutex_lock(&process_info->notifier_lock);
2752	evicted_bos = process_info->evicted_bos;
2753	mutex_unlock(&process_info->notifier_lock);
2754	if (!evicted_bos)
2755		return;
2756
2757	/* Reference task and mm in case of concurrent process termination */
2758	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2759	if (!usertask)
2760		return;
2761	mm = get_task_mm(usertask);
2762	if (!mm) {
2763		put_task_struct(usertask);
2764		return;
2765	}
2766
2767	mutex_lock(&process_info->lock);
2768
2769	if (update_invalid_user_pages(process_info, mm))
2770		goto unlock_out;
2771	/* userptr_inval_list can be empty if all evicted userptr BOs
2772	 * have been freed. In that case there is nothing to validate
2773	 * and we can just restart the queues.
2774	 */
2775	if (!list_empty(&process_info->userptr_inval_list)) {
2776		if (validate_invalid_user_pages(process_info))
2777			goto unlock_out;
2778	}
2779	/* Final check for concurrent evicton and atomic update. If
2780	 * another eviction happens after successful update, it will
2781	 * be a first eviction that calls quiesce_mm. The eviction
2782	 * reference counting inside KFD will handle this case.
2783	 */
2784	mutex_lock(&process_info->notifier_lock);
2785	if (process_info->evicted_bos != evicted_bos)
2786		goto unlock_notifier_out;
2787
2788	if (confirm_valid_user_pages_locked(process_info)) {
2789		WARN(1, "User pages unexpectedly invalid");
2790		goto unlock_notifier_out;
2791	}
2792
2793	process_info->evicted_bos = evicted_bos = 0;
2794
2795	if (kgd2kfd_resume_mm(mm)) {
2796		pr_err("%s: Failed to resume KFD\n", __func__);
2797		/* No recovery from this failure. Probably the CP is
2798		 * hanging. No point trying again.
2799		 */
2800	}
2801
2802unlock_notifier_out:
2803	mutex_unlock(&process_info->notifier_lock);
2804unlock_out:
2805	mutex_unlock(&process_info->lock);
2806
2807	/* If validation failed, reschedule another attempt */
2808	if (evicted_bos) {
2809		queue_delayed_work(system_freezable_wq,
2810			&process_info->restore_userptr_work,
2811			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2812
2813		kfd_smi_event_queue_restore_rescheduled(mm);
2814	}
2815	mmput(mm);
2816	put_task_struct(usertask);
2817}
2818
2819static void replace_eviction_fence(struct dma_fence __rcu **ef,
2820				   struct dma_fence *new_ef)
2821{
2822	struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
2823		/* protected by process_info->lock */);
2824
2825	/* If we're replacing an unsignaled eviction fence, that fence will
2826	 * never be signaled, and if anyone is still waiting on that fence,
2827	 * they will hang forever. This should never happen. We should only
2828	 * replace the fence in restore_work that only gets scheduled after
2829	 * eviction work signaled the fence.
2830	 */
2831	WARN_ONCE(!dma_fence_is_signaled(old_ef),
2832		  "Replacing unsignaled eviction fence");
2833	dma_fence_put(old_ef);
2834}
2835
2836/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2837 *   KFD process identified by process_info
2838 *
2839 * @process_info: amdkfd_process_info of the KFD process
2840 *
2841 * After memory eviction, restore thread calls this function. The function
2842 * should be called when the Process is still valid. BO restore involves -
2843 *
2844 * 1.  Release old eviction fence and create new one
2845 * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2846 * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2847 *     BOs that need to be reserved.
2848 * 4.  Reserve all the BOs
2849 * 5.  Validate of PD and PT BOs.
2850 * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2851 * 7.  Add fence to all PD and PT BOs.
2852 * 8.  Unreserve all BOs
2853 */
2854int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef)
2855{
2856	struct amdkfd_process_info *process_info = info;
2857	struct amdgpu_vm *peer_vm;
2858	struct kgd_mem *mem;
2859	struct list_head duplicate_save;
2860	struct amdgpu_sync sync_obj;
2861	unsigned long failed_size = 0;
2862	unsigned long total_size = 0;
2863	struct drm_exec exec;
2864	int ret;
2865
2866	INIT_LIST_HEAD(&duplicate_save);
2867
2868	mutex_lock(&process_info->lock);
2869
2870	drm_exec_init(&exec, 0, 0);
2871	drm_exec_until_all_locked(&exec) {
2872		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2873				    vm_list_node) {
2874			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2875			drm_exec_retry_on_contention(&exec);
2876			if (unlikely(ret))
 
2877				goto ttm_reserve_fail;
 
2878		}
2879
2880		/* Reserve all BOs and page tables/directory. Add all BOs from
2881		 * kfd_bo_list to ctx.list
2882		 */
2883		list_for_each_entry(mem, &process_info->kfd_bo_list,
2884				    validate_list) {
2885			struct drm_gem_object *gobj;
2886
2887			gobj = &mem->bo->tbo.base;
2888			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2889			drm_exec_retry_on_contention(&exec);
2890			if (unlikely(ret))
 
2891				goto ttm_reserve_fail;
 
2892		}
2893	}
2894
2895	amdgpu_sync_create(&sync_obj);
2896
2897	/* Validate PDs and PTs */
2898	ret = process_validate_vms(process_info);
2899	if (ret)
2900		goto validate_map_fail;
2901
2902	/* Validate BOs and map them to GPUVM (update VM page tables). */
2903	list_for_each_entry(mem, &process_info->kfd_bo_list,
2904			    validate_list) {
2905
2906		struct amdgpu_bo *bo = mem->bo;
2907		uint32_t domain = mem->domain;
2908		struct kfd_mem_attachment *attachment;
2909		struct dma_resv_iter cursor;
2910		struct dma_fence *fence;
2911
2912		total_size += amdgpu_bo_size(bo);
2913
2914		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2915		if (ret) {
2916			pr_debug("Memory eviction: Validate BOs failed\n");
2917			failed_size += amdgpu_bo_size(bo);
2918			ret = amdgpu_amdkfd_bo_validate(bo,
2919						AMDGPU_GEM_DOMAIN_GTT, false);
2920			if (ret) {
2921				pr_debug("Memory eviction: Try again\n");
2922				goto validate_map_fail;
2923			}
2924		}
2925		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2926					DMA_RESV_USAGE_KERNEL, fence) {
2927			ret = amdgpu_sync_fence(&sync_obj, fence);
2928			if (ret) {
2929				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2930				goto validate_map_fail;
2931			}
2932		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2933		list_for_each_entry(attachment, &mem->attachments, list) {
2934			if (!attachment->is_mapped)
2935				continue;
2936
2937			if (attachment->bo_va->base.bo->tbo.pin_count)
2938				continue;
2939
2940			kfd_mem_dmaunmap_attachment(mem, attachment);
2941			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2942			if (ret) {
2943				pr_debug("Memory eviction: update PTE failed. Try again\n");
2944				goto validate_map_fail;
2945			}
2946		}
2947	}
2948
2949	if (failed_size)
2950		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2951
2952	/* Update mappings not managed by KFD */
2953	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2954			vm_list_node) {
2955		struct amdgpu_device *adev = amdgpu_ttm_adev(
2956			peer_vm->root.bo->tbo.bdev);
2957
2958		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
2959		if (ret) {
2960			pr_debug("Memory eviction: handle moved failed. Try again\n");
2961			goto validate_map_fail;
2962		}
2963	}
2964
2965	/* Update page directories */
2966	ret = process_update_pds(process_info, &sync_obj);
2967	if (ret) {
2968		pr_debug("Memory eviction: update PDs failed. Try again\n");
2969		goto validate_map_fail;
2970	}
2971
2972	/* Sync with fences on all the page tables. They implicitly depend on any
2973	 * move fences from amdgpu_vm_handle_moved above.
2974	 */
2975	ret = process_sync_pds_resv(process_info, &sync_obj);
2976	if (ret) {
2977		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2978		goto validate_map_fail;
2979	}
2980
2981	/* Wait for validate and PT updates to finish */
2982	amdgpu_sync_wait(&sync_obj, false);
2983
2984	/* The old eviction fence may be unsignaled if restore happens
2985	 * after a GPU reset or suspend/resume. Keep the old fence in that
2986	 * case. Otherwise release the old eviction fence and create new
2987	 * one, because fence only goes from unsignaled to signaled once
2988	 * and cannot be reused. Use context and mm from the old fence.
2989	 *
2990	 * If an old eviction fence signals after this check, that's OK.
2991	 * Anyone signaling an eviction fence must stop the queues first
2992	 * and schedule another restore worker.
2993	 */
2994	if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
2995		struct amdgpu_amdkfd_fence *new_fence =
2996			amdgpu_amdkfd_fence_create(
2997				process_info->eviction_fence->base.context,
2998				process_info->eviction_fence->mm,
2999				NULL);
3000
3001		if (!new_fence) {
3002			pr_err("Failed to create eviction fence\n");
3003			ret = -ENOMEM;
3004			goto validate_map_fail;
3005		}
3006		dma_fence_put(&process_info->eviction_fence->base);
3007		process_info->eviction_fence = new_fence;
3008		replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
3009	} else {
3010		WARN_ONCE(*ef != &process_info->eviction_fence->base,
3011			  "KFD eviction fence doesn't match KGD process_info");
3012	}
3013
3014	/* Attach new eviction fence to all BOs except pinned ones */
3015	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
3016		if (mem->bo->tbo.pin_count)
3017			continue;
3018
3019		dma_resv_add_fence(mem->bo->tbo.base.resv,
3020				   &process_info->eviction_fence->base,
3021				   DMA_RESV_USAGE_BOOKKEEP);
3022	}
3023	/* Attach eviction fence to PD / PT BOs */
3024	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3025			    vm_list_node) {
3026		struct amdgpu_bo *bo = peer_vm->root.bo;
3027
3028		dma_resv_add_fence(bo->tbo.base.resv,
3029				   &process_info->eviction_fence->base,
3030				   DMA_RESV_USAGE_BOOKKEEP);
3031	}
3032
3033validate_map_fail:
3034	amdgpu_sync_free(&sync_obj);
3035ttm_reserve_fail:
3036	drm_exec_fini(&exec);
3037	mutex_unlock(&process_info->lock);
3038	return ret;
3039}
3040
3041int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
3042{
3043	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3044	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
3045	int ret;
3046
3047	if (!info || !gws)
3048		return -EINVAL;
3049
3050	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
3051	if (!*mem)
3052		return -ENOMEM;
3053
3054	mutex_init(&(*mem)->lock);
3055	INIT_LIST_HEAD(&(*mem)->attachments);
3056	(*mem)->bo = amdgpu_bo_ref(gws_bo);
3057	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
3058	(*mem)->process_info = process_info;
3059	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
3060	amdgpu_sync_create(&(*mem)->sync);
3061
3062
3063	/* Validate gws bo the first time it is added to process */
3064	mutex_lock(&(*mem)->process_info->lock);
3065	ret = amdgpu_bo_reserve(gws_bo, false);
3066	if (unlikely(ret)) {
3067		pr_err("Reserve gws bo failed %d\n", ret);
3068		goto bo_reservation_failure;
3069	}
3070
3071	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
3072	if (ret) {
3073		pr_err("GWS BO validate failed %d\n", ret);
3074		goto bo_validation_failure;
3075	}
3076	/* GWS resource is shared b/t amdgpu and amdkfd
3077	 * Add process eviction fence to bo so they can
3078	 * evict each other.
3079	 */
3080	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
3081	if (ret)
3082		goto reserve_shared_fail;
3083	dma_resv_add_fence(gws_bo->tbo.base.resv,
3084			   &process_info->eviction_fence->base,
3085			   DMA_RESV_USAGE_BOOKKEEP);
3086	amdgpu_bo_unreserve(gws_bo);
3087	mutex_unlock(&(*mem)->process_info->lock);
3088
3089	return ret;
3090
3091reserve_shared_fail:
3092bo_validation_failure:
3093	amdgpu_bo_unreserve(gws_bo);
3094bo_reservation_failure:
3095	mutex_unlock(&(*mem)->process_info->lock);
3096	amdgpu_sync_free(&(*mem)->sync);
3097	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
3098	amdgpu_bo_unref(&gws_bo);
3099	mutex_destroy(&(*mem)->lock);
3100	kfree(*mem);
3101	*mem = NULL;
3102	return ret;
3103}
3104
3105int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
3106{
3107	int ret;
3108	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3109	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
3110	struct amdgpu_bo *gws_bo = kgd_mem->bo;
3111
3112	/* Remove BO from process's validate list so restore worker won't touch
3113	 * it anymore
3114	 */
3115	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
3116
3117	ret = amdgpu_bo_reserve(gws_bo, false);
3118	if (unlikely(ret)) {
3119		pr_err("Reserve gws bo failed %d\n", ret);
3120		//TODO add BO back to validate_list?
3121		return ret;
3122	}
3123	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
3124			process_info->eviction_fence);
3125	amdgpu_bo_unreserve(gws_bo);
3126	amdgpu_sync_free(&kgd_mem->sync);
3127	amdgpu_bo_unref(&gws_bo);
3128	mutex_destroy(&kgd_mem->lock);
3129	kfree(mem);
3130	return 0;
3131}
3132
3133/* Returns GPU-specific tiling mode information */
3134int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
3135				struct tile_config *config)
3136{
3137	config->gb_addr_config = adev->gfx.config.gb_addr_config;
3138	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
3139	config->num_tile_configs =
3140			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
3141	config->macro_tile_config_ptr =
3142			adev->gfx.config.macrotile_mode_array;
3143	config->num_macro_tile_configs =
3144			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
3145
3146	/* Those values are not set from GFX9 onwards */
3147	config->num_banks = adev->gfx.config.num_banks;
3148	config->num_ranks = adev->gfx.config.num_ranks;
3149
3150	return 0;
3151}
3152
3153bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
3154{
3155	struct kfd_mem_attachment *entry;
3156
3157	list_for_each_entry(entry, &mem->attachments, list) {
3158		if (entry->is_mapped && entry->adev == adev)
3159			return true;
3160	}
3161	return false;
3162}
3163
3164#if defined(CONFIG_DEBUG_FS)
3165
3166int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3167{
3168
3169	spin_lock(&kfd_mem_limit.mem_limit_lock);
3170	seq_printf(m, "System mem used %lldM out of %lluM\n",
3171		  (kfd_mem_limit.system_mem_used >> 20),
3172		  (kfd_mem_limit.max_system_mem_limit >> 20));
3173	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3174		  (kfd_mem_limit.ttm_mem_used >> 20),
3175		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3176	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3177
3178	return 0;
3179}
3180
3181#endif