Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0 OR MIT
   2/*
   3 * Copyright 2020-2021 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
  24#include <linux/types.h>
  25#include <linux/sched/task.h>
  26#include <linux/dynamic_debug.h>
  27#include <drm/ttm/ttm_tt.h>
  28#include <drm/drm_exec.h>
  29
  30#include "amdgpu_sync.h"
  31#include "amdgpu_object.h"
  32#include "amdgpu_vm.h"
  33#include "amdgpu_hmm.h"
  34#include "amdgpu.h"
  35#include "amdgpu_xgmi.h"
  36#include "kfd_priv.h"
  37#include "kfd_svm.h"
  38#include "kfd_migrate.h"
  39#include "kfd_smi_events.h"
  40
  41#ifdef dev_fmt
  42#undef dev_fmt
  43#endif
  44#define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
  45
  46#define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
  47
  48/* Long enough to ensure no retry fault comes after svm range is restored and
  49 * page table is updated.
  50 */
  51#define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING	(2UL * NSEC_PER_MSEC)
  52#if IS_ENABLED(CONFIG_DYNAMIC_DEBUG)
  53#define dynamic_svm_range_dump(svms) \
  54	_dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms)
  55#else
  56#define dynamic_svm_range_dump(svms) \
  57	do { if (0) svm_range_debug_dump(svms); } while (0)
  58#endif
  59
  60/* Giant svm range split into smaller ranges based on this, it is decided using
  61 * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to
  62 * power of 2MB.
  63 */
  64static uint64_t max_svm_range_pages;
  65
  66struct criu_svm_metadata {
  67	struct list_head list;
  68	struct kfd_criu_svm_range_priv_data data;
  69};
  70
  71static void svm_range_evict_svm_bo_worker(struct work_struct *work);
  72static bool
  73svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
  74				    const struct mmu_notifier_range *range,
  75				    unsigned long cur_seq);
  76static int
  77svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
  78		   uint64_t *bo_s, uint64_t *bo_l);
  79static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
  80	.invalidate = svm_range_cpu_invalidate_pagetables,
  81};
  82
  83/**
  84 * svm_range_unlink - unlink svm_range from lists and interval tree
  85 * @prange: svm range structure to be removed
  86 *
  87 * Remove the svm_range from the svms and svm_bo lists and the svms
  88 * interval tree.
  89 *
  90 * Context: The caller must hold svms->lock
  91 */
  92static void svm_range_unlink(struct svm_range *prange)
  93{
  94	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
  95		 prange, prange->start, prange->last);
  96
  97	if (prange->svm_bo) {
  98		spin_lock(&prange->svm_bo->list_lock);
  99		list_del(&prange->svm_bo_list);
 100		spin_unlock(&prange->svm_bo->list_lock);
 101	}
 102
 103	list_del(&prange->list);
 104	if (prange->it_node.start != 0 && prange->it_node.last != 0)
 105		interval_tree_remove(&prange->it_node, &prange->svms->objects);
 106}
 107
 108static void
 109svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
 110{
 111	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
 112		 prange, prange->start, prange->last);
 113
 114	mmu_interval_notifier_insert_locked(&prange->notifier, mm,
 115				     prange->start << PAGE_SHIFT,
 116				     prange->npages << PAGE_SHIFT,
 117				     &svm_range_mn_ops);
 118}
 119
 120/**
 121 * svm_range_add_to_svms - add svm range to svms
 122 * @prange: svm range structure to be added
 123 *
 124 * Add the svm range to svms interval tree and link list
 125 *
 126 * Context: The caller must hold svms->lock
 127 */
 128static void svm_range_add_to_svms(struct svm_range *prange)
 129{
 130	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
 131		 prange, prange->start, prange->last);
 132
 133	list_move_tail(&prange->list, &prange->svms->list);
 134	prange->it_node.start = prange->start;
 135	prange->it_node.last = prange->last;
 136	interval_tree_insert(&prange->it_node, &prange->svms->objects);
 137}
 138
 139static void svm_range_remove_notifier(struct svm_range *prange)
 140{
 141	pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
 142		 prange->svms, prange,
 143		 prange->notifier.interval_tree.start >> PAGE_SHIFT,
 144		 prange->notifier.interval_tree.last >> PAGE_SHIFT);
 145
 146	if (prange->notifier.interval_tree.start != 0 &&
 147	    prange->notifier.interval_tree.last != 0)
 148		mmu_interval_notifier_remove(&prange->notifier);
 149}
 150
 151static bool
 152svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
 153{
 154	return dma_addr && !dma_mapping_error(dev, dma_addr) &&
 155	       !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
 156}
 157
 158static int
 159svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
 160		      unsigned long offset, unsigned long npages,
 161		      unsigned long *hmm_pfns, uint32_t gpuidx)
 162{
 163	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
 164	dma_addr_t *addr = prange->dma_addr[gpuidx];
 165	struct device *dev = adev->dev;
 166	struct page *page;
 167	int i, r;
 168
 169	if (!addr) {
 170		addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL);
 171		if (!addr)
 172			return -ENOMEM;
 173		prange->dma_addr[gpuidx] = addr;
 174	}
 175
 176	addr += offset;
 177	for (i = 0; i < npages; i++) {
 178		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
 179			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
 180
 181		page = hmm_pfn_to_page(hmm_pfns[i]);
 182		if (is_zone_device_page(page)) {
 183			struct amdgpu_device *bo_adev = prange->svm_bo->node->adev;
 184
 185			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
 186				   bo_adev->vm_manager.vram_base_offset -
 187				   bo_adev->kfd.pgmap.range.start;
 188			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
 189			pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
 190			continue;
 191		}
 192		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
 193		r = dma_mapping_error(dev, addr[i]);
 194		if (r) {
 195			dev_err(dev, "failed %d dma_map_page\n", r);
 196			return r;
 197		}
 198		pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
 199				     addr[i] >> PAGE_SHIFT, page_to_pfn(page));
 200	}
 201
 202	return 0;
 203}
 204
 205static int
 206svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
 207		  unsigned long offset, unsigned long npages,
 208		  unsigned long *hmm_pfns)
 209{
 210	struct kfd_process *p;
 211	uint32_t gpuidx;
 212	int r;
 213
 214	p = container_of(prange->svms, struct kfd_process, svms);
 215
 216	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
 217		struct kfd_process_device *pdd;
 218
 219		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
 220		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
 221		if (!pdd) {
 222			pr_debug("failed to find device idx %d\n", gpuidx);
 223			return -EINVAL;
 224		}
 225
 226		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
 227					  hmm_pfns, gpuidx);
 228		if (r)
 229			break;
 230	}
 231
 232	return r;
 233}
 234
 235void svm_range_dma_unmap_dev(struct device *dev, dma_addr_t *dma_addr,
 236			 unsigned long offset, unsigned long npages)
 237{
 238	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
 239	int i;
 240
 241	if (!dma_addr)
 242		return;
 243
 244	for (i = offset; i < offset + npages; i++) {
 245		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
 246			continue;
 247		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
 248		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
 249		dma_addr[i] = 0;
 250	}
 251}
 252
 253void svm_range_dma_unmap(struct svm_range *prange)
 254{
 255	struct kfd_process_device *pdd;
 256	dma_addr_t *dma_addr;
 257	struct device *dev;
 258	struct kfd_process *p;
 259	uint32_t gpuidx;
 260
 261	p = container_of(prange->svms, struct kfd_process, svms);
 262
 263	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
 264		dma_addr = prange->dma_addr[gpuidx];
 265		if (!dma_addr)
 266			continue;
 267
 268		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
 269		if (!pdd) {
 270			pr_debug("failed to find device idx %d\n", gpuidx);
 271			continue;
 272		}
 273		dev = &pdd->dev->adev->pdev->dev;
 274
 275		svm_range_dma_unmap_dev(dev, dma_addr, 0, prange->npages);
 276	}
 277}
 278
 279static void svm_range_free(struct svm_range *prange, bool do_unmap)
 280{
 281	uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
 282	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 283	uint32_t gpuidx;
 284
 285	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
 286		 prange->start, prange->last);
 287
 288	svm_range_vram_node_free(prange);
 289	if (do_unmap)
 290		svm_range_dma_unmap(prange);
 291
 292	if (do_unmap && !p->xnack_enabled) {
 293		pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size);
 294		amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
 295					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
 296	}
 297
 298	/* free dma_addr array for each gpu */
 299	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
 300		if (prange->dma_addr[gpuidx]) {
 301			kvfree(prange->dma_addr[gpuidx]);
 302			prange->dma_addr[gpuidx] = NULL;
 303		}
 304	}
 305
 306	mutex_destroy(&prange->lock);
 307	mutex_destroy(&prange->migrate_mutex);
 308	kfree(prange);
 309}
 310
 311static void
 312svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
 313				 uint8_t *granularity, uint32_t *flags)
 314{
 315	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
 316	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
 317	*granularity = 9;
 318	*flags =
 319		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
 320}
 321
 322static struct
 323svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
 324			 uint64_t last, bool update_mem_usage)
 325{
 326	uint64_t size = last - start + 1;
 327	struct svm_range *prange;
 328	struct kfd_process *p;
 329
 330	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
 331	if (!prange)
 332		return NULL;
 333
 334	p = container_of(svms, struct kfd_process, svms);
 335	if (!p->xnack_enabled && update_mem_usage &&
 336	    amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
 337				    KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) {
 338		pr_info("SVM mapping failed, exceeds resident system memory limit\n");
 339		kfree(prange);
 340		return NULL;
 341	}
 342	prange->npages = size;
 343	prange->svms = svms;
 344	prange->start = start;
 345	prange->last = last;
 346	INIT_LIST_HEAD(&prange->list);
 347	INIT_LIST_HEAD(&prange->update_list);
 348	INIT_LIST_HEAD(&prange->svm_bo_list);
 349	INIT_LIST_HEAD(&prange->deferred_list);
 350	INIT_LIST_HEAD(&prange->child_list);
 351	atomic_set(&prange->invalid, 0);
 352	prange->validate_timestamp = 0;
 353	prange->vram_pages = 0;
 354	mutex_init(&prange->migrate_mutex);
 355	mutex_init(&prange->lock);
 356
 357	if (p->xnack_enabled)
 358		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
 359			    MAX_GPU_INSTANCE);
 360
 361	svm_range_set_default_attributes(&prange->preferred_loc,
 362					 &prange->prefetch_loc,
 363					 &prange->granularity, &prange->flags);
 364
 365	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
 366
 367	return prange;
 368}
 369
 370static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
 371{
 372	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
 373		return false;
 374
 375	return true;
 376}
 377
 378static void svm_range_bo_release(struct kref *kref)
 379{
 380	struct svm_range_bo *svm_bo;
 381
 382	svm_bo = container_of(kref, struct svm_range_bo, kref);
 383	pr_debug("svm_bo 0x%p\n", svm_bo);
 384
 385	spin_lock(&svm_bo->list_lock);
 386	while (!list_empty(&svm_bo->range_list)) {
 387		struct svm_range *prange =
 388				list_first_entry(&svm_bo->range_list,
 389						struct svm_range, svm_bo_list);
 390		/* list_del_init tells a concurrent svm_range_vram_node_new when
 391		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
 392		 */
 393		list_del_init(&prange->svm_bo_list);
 394		spin_unlock(&svm_bo->list_lock);
 395
 396		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
 397			 prange->start, prange->last);
 398		mutex_lock(&prange->lock);
 399		prange->svm_bo = NULL;
 400		/* prange should not hold vram page now */
 401		WARN_ONCE(prange->actual_loc, "prange should not hold vram page");
 402		mutex_unlock(&prange->lock);
 403
 404		spin_lock(&svm_bo->list_lock);
 405	}
 406	spin_unlock(&svm_bo->list_lock);
 407	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base))
 408		/* We're not in the eviction worker. Signal the fence. */
 409		dma_fence_signal(&svm_bo->eviction_fence->base);
 410	dma_fence_put(&svm_bo->eviction_fence->base);
 411	amdgpu_bo_unref(&svm_bo->bo);
 412	kfree(svm_bo);
 413}
 414
 415static void svm_range_bo_wq_release(struct work_struct *work)
 416{
 417	struct svm_range_bo *svm_bo;
 418
 419	svm_bo = container_of(work, struct svm_range_bo, release_work);
 420	svm_range_bo_release(&svm_bo->kref);
 421}
 422
 423static void svm_range_bo_release_async(struct kref *kref)
 424{
 425	struct svm_range_bo *svm_bo;
 426
 427	svm_bo = container_of(kref, struct svm_range_bo, kref);
 428	pr_debug("svm_bo 0x%p\n", svm_bo);
 429	INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
 430	schedule_work(&svm_bo->release_work);
 431}
 432
 433void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
 434{
 435	kref_put(&svm_bo->kref, svm_range_bo_release_async);
 436}
 437
 438static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
 439{
 440	if (svm_bo)
 441		kref_put(&svm_bo->kref, svm_range_bo_release);
 442}
 443
 444static bool
 445svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange)
 446{
 447	mutex_lock(&prange->lock);
 448	if (!prange->svm_bo) {
 449		mutex_unlock(&prange->lock);
 450		return false;
 451	}
 452	if (prange->ttm_res) {
 453		/* We still have a reference, all is well */
 454		mutex_unlock(&prange->lock);
 455		return true;
 456	}
 457	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
 458		/*
 459		 * Migrate from GPU to GPU, remove range from source svm_bo->node
 460		 * range list, and return false to allocate svm_bo from destination
 461		 * node.
 462		 */
 463		if (prange->svm_bo->node != node) {
 464			mutex_unlock(&prange->lock);
 465
 466			spin_lock(&prange->svm_bo->list_lock);
 467			list_del_init(&prange->svm_bo_list);
 468			spin_unlock(&prange->svm_bo->list_lock);
 469
 470			svm_range_bo_unref(prange->svm_bo);
 471			return false;
 472		}
 473		if (READ_ONCE(prange->svm_bo->evicting)) {
 474			struct dma_fence *f;
 475			struct svm_range_bo *svm_bo;
 476			/* The BO is getting evicted,
 477			 * we need to get a new one
 478			 */
 479			mutex_unlock(&prange->lock);
 480			svm_bo = prange->svm_bo;
 481			f = dma_fence_get(&svm_bo->eviction_fence->base);
 482			svm_range_bo_unref(prange->svm_bo);
 483			/* wait for the fence to avoid long spin-loop
 484			 * at list_empty_careful
 485			 */
 486			dma_fence_wait(f, false);
 487			dma_fence_put(f);
 488		} else {
 489			/* The BO was still around and we got
 490			 * a new reference to it
 491			 */
 492			mutex_unlock(&prange->lock);
 493			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
 494				 prange->svms, prange->start, prange->last);
 495
 496			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
 497			return true;
 498		}
 499
 500	} else {
 501		mutex_unlock(&prange->lock);
 502	}
 503
 504	/* We need a new svm_bo. Spin-loop to wait for concurrent
 505	 * svm_range_bo_release to finish removing this range from
 506	 * its range list and set prange->svm_bo to null. After this,
 507	 * it is safe to reuse the svm_bo pointer and svm_bo_list head.
 508	 */
 509	while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo)
 510		cond_resched();
 511
 512	return false;
 513}
 514
 515static struct svm_range_bo *svm_range_bo_new(void)
 516{
 517	struct svm_range_bo *svm_bo;
 518
 519	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
 520	if (!svm_bo)
 521		return NULL;
 522
 523	kref_init(&svm_bo->kref);
 524	INIT_LIST_HEAD(&svm_bo->range_list);
 525	spin_lock_init(&svm_bo->list_lock);
 526
 527	return svm_bo;
 528}
 529
 530int
 531svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
 532			bool clear)
 533{
 534	struct amdgpu_bo_param bp;
 535	struct svm_range_bo *svm_bo;
 536	struct amdgpu_bo_user *ubo;
 537	struct amdgpu_bo *bo;
 538	struct kfd_process *p;
 539	struct mm_struct *mm;
 540	int r;
 541
 542	p = container_of(prange->svms, struct kfd_process, svms);
 543	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
 544		 prange->start, prange->last);
 545
 546	if (svm_range_validate_svm_bo(node, prange))
 547		return 0;
 548
 549	svm_bo = svm_range_bo_new();
 550	if (!svm_bo) {
 551		pr_debug("failed to alloc svm bo\n");
 552		return -ENOMEM;
 553	}
 554	mm = get_task_mm(p->lead_thread);
 555	if (!mm) {
 556		pr_debug("failed to get mm\n");
 557		kfree(svm_bo);
 558		return -ESRCH;
 559	}
 560	svm_bo->node = node;
 561	svm_bo->eviction_fence =
 562		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
 563					   mm,
 564					   svm_bo);
 565	mmput(mm);
 566	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
 567	svm_bo->evicting = 0;
 568	memset(&bp, 0, sizeof(bp));
 569	bp.size = prange->npages * PAGE_SIZE;
 570	bp.byte_align = PAGE_SIZE;
 571	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
 572	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
 573	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
 574	bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
 575	bp.type = ttm_bo_type_device;
 576	bp.resv = NULL;
 577	if (node->xcp)
 578		bp.xcp_id_plus1 = node->xcp->id + 1;
 579
 580	r = amdgpu_bo_create_user(node->adev, &bp, &ubo);
 581	if (r) {
 582		pr_debug("failed %d to create bo\n", r);
 583		goto create_bo_failed;
 584	}
 585	bo = &ubo->bo;
 586
 587	pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n",
 588		 bo->tbo.resource->start << PAGE_SHIFT, bp.size,
 589		 bp.xcp_id_plus1 - 1);
 590
 591	r = amdgpu_bo_reserve(bo, true);
 592	if (r) {
 593		pr_debug("failed %d to reserve bo\n", r);
 594		goto reserve_bo_failed;
 595	}
 596
 597	if (clear) {
 598		r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
 599		if (r) {
 600			pr_debug("failed %d to sync bo\n", r);
 601			amdgpu_bo_unreserve(bo);
 602			goto reserve_bo_failed;
 603		}
 604	}
 605
 606	r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
 607	if (r) {
 608		pr_debug("failed %d to reserve bo\n", r);
 609		amdgpu_bo_unreserve(bo);
 610		goto reserve_bo_failed;
 611	}
 612	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
 613
 614	amdgpu_bo_unreserve(bo);
 615
 616	svm_bo->bo = bo;
 617	prange->svm_bo = svm_bo;
 618	prange->ttm_res = bo->tbo.resource;
 619	prange->offset = 0;
 620
 621	spin_lock(&svm_bo->list_lock);
 622	list_add(&prange->svm_bo_list, &svm_bo->range_list);
 623	spin_unlock(&svm_bo->list_lock);
 624
 625	return 0;
 626
 627reserve_bo_failed:
 628	amdgpu_bo_unref(&bo);
 629create_bo_failed:
 630	dma_fence_put(&svm_bo->eviction_fence->base);
 631	kfree(svm_bo);
 632	prange->ttm_res = NULL;
 633
 634	return r;
 635}
 636
 637void svm_range_vram_node_free(struct svm_range *prange)
 638{
 639	/* serialize prange->svm_bo unref */
 640	mutex_lock(&prange->lock);
 641	/* prange->svm_bo has not been unref */
 642	if (prange->ttm_res) {
 643		prange->ttm_res = NULL;
 644		mutex_unlock(&prange->lock);
 645		svm_range_bo_unref(prange->svm_bo);
 646	} else
 647		mutex_unlock(&prange->lock);
 648}
 649
 650struct kfd_node *
 651svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id)
 652{
 653	struct kfd_process *p;
 654	struct kfd_process_device *pdd;
 655
 656	p = container_of(prange->svms, struct kfd_process, svms);
 657	pdd = kfd_process_device_data_by_id(p, gpu_id);
 658	if (!pdd) {
 659		pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id);
 660		return NULL;
 661	}
 662
 663	return pdd->dev;
 664}
 665
 666struct kfd_process_device *
 667svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node)
 668{
 669	struct kfd_process *p;
 670
 671	p = container_of(prange->svms, struct kfd_process, svms);
 672
 673	return kfd_get_process_device_data(node, p);
 674}
 675
 676static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
 677{
 678	struct ttm_operation_ctx ctx = { false, false };
 679
 680	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
 681
 682	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 683}
 684
 685static int
 686svm_range_check_attr(struct kfd_process *p,
 687		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
 688{
 689	uint32_t i;
 690
 691	for (i = 0; i < nattr; i++) {
 692		uint32_t val = attrs[i].value;
 693		int gpuidx = MAX_GPU_INSTANCE;
 694
 695		switch (attrs[i].type) {
 696		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
 697			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
 698			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
 699				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
 700			break;
 701		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
 702			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
 703				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
 704			break;
 705		case KFD_IOCTL_SVM_ATTR_ACCESS:
 706		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
 707		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
 708			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
 709			break;
 710		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
 711			break;
 712		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
 713			break;
 714		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
 715			break;
 716		default:
 717			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
 718			return -EINVAL;
 719		}
 720
 721		if (gpuidx < 0) {
 722			pr_debug("no GPU 0x%x found\n", val);
 723			return -EINVAL;
 724		} else if (gpuidx < MAX_GPU_INSTANCE &&
 725			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
 726			pr_debug("GPU 0x%x not supported\n", val);
 727			return -EINVAL;
 728		}
 729	}
 730
 731	return 0;
 732}
 733
 734static void
 735svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
 736		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
 737		      bool *update_mapping)
 738{
 739	uint32_t i;
 740	int gpuidx;
 741
 742	for (i = 0; i < nattr; i++) {
 743		switch (attrs[i].type) {
 744		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
 745			prange->preferred_loc = attrs[i].value;
 746			break;
 747		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
 748			prange->prefetch_loc = attrs[i].value;
 749			break;
 750		case KFD_IOCTL_SVM_ATTR_ACCESS:
 751		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
 752		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
 753			if (!p->xnack_enabled)
 754				*update_mapping = true;
 755
 756			gpuidx = kfd_process_gpuidx_from_gpuid(p,
 757							       attrs[i].value);
 758			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
 759				bitmap_clear(prange->bitmap_access, gpuidx, 1);
 760				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
 761			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
 762				bitmap_set(prange->bitmap_access, gpuidx, 1);
 763				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
 764			} else {
 765				bitmap_clear(prange->bitmap_access, gpuidx, 1);
 766				bitmap_set(prange->bitmap_aip, gpuidx, 1);
 767			}
 768			break;
 769		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
 770			*update_mapping = true;
 771			prange->flags |= attrs[i].value;
 772			break;
 773		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
 774			*update_mapping = true;
 775			prange->flags &= ~attrs[i].value;
 776			break;
 777		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
 778			prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F);
 779			break;
 780		default:
 781			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
 782		}
 783	}
 784}
 785
 786static bool
 787svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
 788			uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
 789{
 790	uint32_t i;
 791	int gpuidx;
 792
 793	for (i = 0; i < nattr; i++) {
 794		switch (attrs[i].type) {
 795		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
 796			if (prange->preferred_loc != attrs[i].value)
 797				return false;
 798			break;
 799		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
 800			/* Prefetch should always trigger a migration even
 801			 * if the value of the attribute didn't change.
 802			 */
 803			return false;
 804		case KFD_IOCTL_SVM_ATTR_ACCESS:
 805		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
 806		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
 807			gpuidx = kfd_process_gpuidx_from_gpuid(p,
 808							       attrs[i].value);
 809			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
 810				if (test_bit(gpuidx, prange->bitmap_access) ||
 811				    test_bit(gpuidx, prange->bitmap_aip))
 812					return false;
 813			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
 814				if (!test_bit(gpuidx, prange->bitmap_access))
 815					return false;
 816			} else {
 817				if (!test_bit(gpuidx, prange->bitmap_aip))
 818					return false;
 819			}
 820			break;
 821		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
 822			if ((prange->flags & attrs[i].value) != attrs[i].value)
 823				return false;
 824			break;
 825		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
 826			if ((prange->flags & attrs[i].value) != 0)
 827				return false;
 828			break;
 829		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
 830			if (prange->granularity != attrs[i].value)
 831				return false;
 832			break;
 833		default:
 834			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
 835		}
 836	}
 837
 838	return true;
 839}
 840
 841/**
 842 * svm_range_debug_dump - print all range information from svms
 843 * @svms: svm range list header
 844 *
 845 * debug output svm range start, end, prefetch location from svms
 846 * interval tree and link list
 847 *
 848 * Context: The caller must hold svms->lock
 849 */
 850static void svm_range_debug_dump(struct svm_range_list *svms)
 851{
 852	struct interval_tree_node *node;
 853	struct svm_range *prange;
 854
 855	pr_debug("dump svms 0x%p list\n", svms);
 856	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
 857
 858	list_for_each_entry(prange, &svms->list, list) {
 859		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
 860			 prange, prange->start, prange->npages,
 861			 prange->start + prange->npages - 1,
 862			 prange->actual_loc);
 863	}
 864
 865	pr_debug("dump svms 0x%p interval tree\n", svms);
 866	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
 867	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
 868	while (node) {
 869		prange = container_of(node, struct svm_range, it_node);
 870		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
 871			 prange, prange->start, prange->npages,
 872			 prange->start + prange->npages - 1,
 873			 prange->actual_loc);
 874		node = interval_tree_iter_next(node, 0, ~0ULL);
 875	}
 876}
 877
 878static void *
 879svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements,
 880		     uint64_t offset, uint64_t *vram_pages)
 881{
 882	unsigned char *src = (unsigned char *)psrc + offset;
 883	unsigned char *dst;
 884	uint64_t i;
 885
 886	dst = kvmalloc_array(num_elements, size, GFP_KERNEL);
 887	if (!dst)
 888		return NULL;
 889
 890	if (!vram_pages) {
 891		memcpy(dst, src, num_elements * size);
 892		return (void *)dst;
 893	}
 894
 895	*vram_pages = 0;
 896	for (i = 0; i < num_elements; i++) {
 897		dma_addr_t *temp;
 898		temp = (dma_addr_t *)dst + i;
 899		*temp = *((dma_addr_t *)src + i);
 900		if (*temp&SVM_RANGE_VRAM_DOMAIN)
 901			(*vram_pages)++;
 902	}
 903
 904	return (void *)dst;
 905}
 906
 907static int
 908svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src)
 909{
 910	int i;
 911
 912	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
 913		if (!src->dma_addr[i])
 914			continue;
 915		dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i],
 916					sizeof(*src->dma_addr[i]), src->npages, 0, NULL);
 917		if (!dst->dma_addr[i])
 918			return -ENOMEM;
 919	}
 920
 921	return 0;
 922}
 923
 924static int
 925svm_range_split_array(void *ppnew, void *ppold, size_t size,
 926		      uint64_t old_start, uint64_t old_n,
 927		      uint64_t new_start, uint64_t new_n, uint64_t *new_vram_pages)
 928{
 929	unsigned char *new, *old, *pold;
 930	uint64_t d;
 931
 932	if (!ppold)
 933		return 0;
 934	pold = *(unsigned char **)ppold;
 935	if (!pold)
 936		return 0;
 937
 938	d = (new_start - old_start) * size;
 939	/* get dma addr array for new range and calculte its vram page number */
 940	new = svm_range_copy_array(pold, size, new_n, d, new_vram_pages);
 941	if (!new)
 942		return -ENOMEM;
 943	d = (new_start == old_start) ? new_n * size : 0;
 944	old = svm_range_copy_array(pold, size, old_n, d, NULL);
 945	if (!old) {
 946		kvfree(new);
 947		return -ENOMEM;
 948	}
 949	kvfree(pold);
 950	*(void **)ppold = old;
 951	*(void **)ppnew = new;
 952
 953	return 0;
 954}
 955
 956static int
 957svm_range_split_pages(struct svm_range *new, struct svm_range *old,
 958		      uint64_t start, uint64_t last)
 959{
 960	uint64_t npages = last - start + 1;
 961	int i, r;
 962
 963	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
 964		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
 965					  sizeof(*old->dma_addr[i]), old->start,
 966					  npages, new->start, new->npages,
 967					  old->actual_loc ? &new->vram_pages : NULL);
 968		if (r)
 969			return r;
 970	}
 971	if (old->actual_loc)
 972		old->vram_pages -= new->vram_pages;
 973
 974	return 0;
 975}
 976
 977static int
 978svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
 979		      uint64_t start, uint64_t last)
 980{
 981	uint64_t npages = last - start + 1;
 982
 983	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
 984		 new->svms, new, new->start, start, last);
 985
 986	if (new->start == old->start) {
 987		new->offset = old->offset;
 988		old->offset += new->npages;
 989	} else {
 990		new->offset = old->offset + npages;
 991	}
 992
 993	new->svm_bo = svm_range_bo_ref(old->svm_bo);
 994	new->ttm_res = old->ttm_res;
 995
 996	spin_lock(&new->svm_bo->list_lock);
 997	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
 998	spin_unlock(&new->svm_bo->list_lock);
 999
1000	return 0;
1001}
1002
1003/**
1004 * svm_range_split_adjust - split range and adjust
1005 *
1006 * @new: new range
1007 * @old: the old range
1008 * @start: the old range adjust to start address in pages
1009 * @last: the old range adjust to last address in pages
1010 *
1011 * Copy system memory dma_addr or vram ttm_res in old range to new
1012 * range from new_start up to size new->npages, the remaining old range is from
1013 * start to last
1014 *
1015 * Return:
1016 * 0 - OK, -ENOMEM - out of memory
1017 */
1018static int
1019svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
1020		      uint64_t start, uint64_t last)
1021{
1022	int r;
1023
1024	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
1025		 new->svms, new->start, old->start, old->last, start, last);
1026
1027	if (new->start < old->start ||
1028	    new->last > old->last) {
1029		WARN_ONCE(1, "invalid new range start or last\n");
1030		return -EINVAL;
1031	}
1032
1033	r = svm_range_split_pages(new, old, start, last);
1034	if (r)
1035		return r;
1036
1037	if (old->actual_loc && old->ttm_res) {
1038		r = svm_range_split_nodes(new, old, start, last);
1039		if (r)
1040			return r;
1041	}
1042
1043	old->npages = last - start + 1;
1044	old->start = start;
1045	old->last = last;
1046	new->flags = old->flags;
1047	new->preferred_loc = old->preferred_loc;
1048	new->prefetch_loc = old->prefetch_loc;
1049	new->actual_loc = old->actual_loc;
1050	new->granularity = old->granularity;
1051	new->mapped_to_gpu = old->mapped_to_gpu;
1052	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1053	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1054
1055	return 0;
1056}
1057
1058/**
1059 * svm_range_split - split a range in 2 ranges
1060 *
1061 * @prange: the svm range to split
1062 * @start: the remaining range start address in pages
1063 * @last: the remaining range last address in pages
1064 * @new: the result new range generated
1065 *
1066 * Two cases only:
1067 * case 1: if start == prange->start
1068 *         prange ==> prange[start, last]
1069 *         new range [last + 1, prange->last]
1070 *
1071 * case 2: if last == prange->last
1072 *         prange ==> prange[start, last]
1073 *         new range [prange->start, start - 1]
1074 *
1075 * Return:
1076 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1077 */
1078static int
1079svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1080		struct svm_range **new)
1081{
1082	uint64_t old_start = prange->start;
1083	uint64_t old_last = prange->last;
1084	struct svm_range_list *svms;
1085	int r = 0;
1086
1087	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1088		 old_start, old_last, start, last);
1089
1090	if (old_start != start && old_last != last)
1091		return -EINVAL;
1092	if (start < old_start || last > old_last)
1093		return -EINVAL;
1094
1095	svms = prange->svms;
1096	if (old_start == start)
1097		*new = svm_range_new(svms, last + 1, old_last, false);
1098	else
1099		*new = svm_range_new(svms, old_start, start - 1, false);
1100	if (!*new)
1101		return -ENOMEM;
1102
1103	r = svm_range_split_adjust(*new, prange, start, last);
1104	if (r) {
1105		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1106			 r, old_start, old_last, start, last);
1107		svm_range_free(*new, false);
1108		*new = NULL;
1109	}
1110
1111	return r;
1112}
1113
1114static int
1115svm_range_split_tail(struct svm_range *prange, uint64_t new_last,
1116		     struct list_head *insert_list, struct list_head *remap_list)
1117{
1118	struct svm_range *tail = NULL;
1119	int r = svm_range_split(prange, prange->start, new_last, &tail);
1120
1121	if (!r) {
1122		list_add(&tail->list, insert_list);
1123		if (!IS_ALIGNED(new_last + 1, 1UL << prange->granularity))
1124			list_add(&tail->update_list, remap_list);
1125	}
1126	return r;
1127}
1128
1129static int
1130svm_range_split_head(struct svm_range *prange, uint64_t new_start,
1131		     struct list_head *insert_list, struct list_head *remap_list)
1132{
1133	struct svm_range *head = NULL;
1134	int r = svm_range_split(prange, new_start, prange->last, &head);
1135
1136	if (!r) {
1137		list_add(&head->list, insert_list);
1138		if (!IS_ALIGNED(new_start, 1UL << prange->granularity))
1139			list_add(&head->update_list, remap_list);
1140	}
1141	return r;
1142}
1143
1144static void
1145svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1146		    struct svm_range *pchild, enum svm_work_list_ops op)
1147{
1148	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1149		 pchild, pchild->start, pchild->last, prange, op);
1150
1151	pchild->work_item.mm = mm;
1152	pchild->work_item.op = op;
1153	list_add_tail(&pchild->child_list, &prange->child_list);
1154}
1155
1156static bool
1157svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
1158{
1159	return (node_a->adev == node_b->adev ||
1160		amdgpu_xgmi_same_hive(node_a->adev, node_b->adev));
1161}
1162
1163static uint64_t
1164svm_range_get_pte_flags(struct kfd_node *node,
1165			struct svm_range *prange, int domain)
1166{
1167	struct kfd_node *bo_node;
1168	uint32_t flags = prange->flags;
1169	uint32_t mapping_flags = 0;
1170	uint64_t pte_flags;
1171	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1172	bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENT);
1173	bool ext_coherent = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENT;
1174	bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/
1175	unsigned int mtype_local;
1176
1177	if (domain == SVM_RANGE_VRAM_DOMAIN)
1178		bo_node = prange->svm_bo->node;
1179
1180	switch (amdgpu_ip_version(node->adev, GC_HWIP, 0)) {
1181	case IP_VERSION(9, 4, 1):
1182		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1183			if (bo_node == node) {
1184				mapping_flags |= coherent ?
1185					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1186			} else {
1187				mapping_flags |= coherent ?
1188					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1189				if (svm_nodes_in_same_hive(node, bo_node))
1190					snoop = true;
1191			}
1192		} else {
1193			mapping_flags |= coherent ?
1194				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1195		}
1196		break;
1197	case IP_VERSION(9, 4, 2):
1198		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1199			if (bo_node == node) {
1200				mapping_flags |= coherent ?
1201					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1202				if (node->adev->gmc.xgmi.connected_to_cpu)
1203					snoop = true;
1204			} else {
1205				mapping_flags |= coherent ?
1206					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1207				if (svm_nodes_in_same_hive(node, bo_node))
1208					snoop = true;
1209			}
1210		} else {
1211			mapping_flags |= coherent ?
1212				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1213		}
1214		break;
1215	case IP_VERSION(9, 4, 3):
1216		if (ext_coherent)
1217			mtype_local = node->adev->rev_id ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_UC;
1218		else
1219			mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC :
1220				amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1221		snoop = true;
1222		if (uncached) {
1223			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1224		} else if (domain == SVM_RANGE_VRAM_DOMAIN) {
1225			/* local HBM region close to partition */
1226			if (bo_node->adev == node->adev &&
1227			    (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
1228				mapping_flags |= mtype_local;
1229			/* local HBM region far from partition or remote XGMI GPU
1230			 * with regular system scope coherence
1231			 */
1232			else if (svm_nodes_in_same_hive(bo_node, node) && !ext_coherent)
1233				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1234			/* PCIe P2P or extended system scope coherence */
1235			else
1236				mapping_flags |= AMDGPU_VM_MTYPE_UC;
1237		/* system memory accessed by the APU */
1238		} else if (node->adev->flags & AMD_IS_APU) {
1239			/* On NUMA systems, locality is determined per-page
1240			 * in amdgpu_gmc_override_vm_pte_flags
1241			 */
1242			if (num_possible_nodes() <= 1)
1243				mapping_flags |= mtype_local;
1244			else
1245				mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1246		/* system memory accessed by the dGPU */
1247		} else {
1248			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1249		}
1250		break;
1251	default:
1252		mapping_flags |= coherent ?
1253			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1254	}
1255
1256	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1257
1258	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1259		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1260	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1261		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1262
1263	pte_flags = AMDGPU_PTE_VALID;
1264	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1265	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1266
1267	pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags);
1268	return pte_flags;
1269}
1270
1271static int
1272svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1273			 uint64_t start, uint64_t last,
1274			 struct dma_fence **fence)
1275{
1276	uint64_t init_pte_value = 0;
1277
1278	pr_debug("[0x%llx 0x%llx]\n", start, last);
1279
1280	return amdgpu_vm_update_range(adev, vm, false, true, true, false, NULL, start,
1281				      last, init_pte_value, 0, 0, NULL, NULL,
1282				      fence);
1283}
1284
1285static int
1286svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1287			  unsigned long last, uint32_t trigger)
1288{
1289	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1290	struct kfd_process_device *pdd;
1291	struct dma_fence *fence = NULL;
1292	struct kfd_process *p;
1293	uint32_t gpuidx;
1294	int r = 0;
1295
1296	if (!prange->mapped_to_gpu) {
1297		pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1298			 prange, prange->start, prange->last);
1299		return 0;
1300	}
1301
1302	if (prange->start == start && prange->last == last) {
1303		pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1304		prange->mapped_to_gpu = false;
1305	}
1306
1307	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1308		  MAX_GPU_INSTANCE);
1309	p = container_of(prange->svms, struct kfd_process, svms);
1310
1311	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1312		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1313		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1314		if (!pdd) {
1315			pr_debug("failed to find device idx %d\n", gpuidx);
1316			return -EINVAL;
1317		}
1318
1319		kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1320					     start, last, trigger);
1321
1322		r = svm_range_unmap_from_gpu(pdd->dev->adev,
1323					     drm_priv_to_vm(pdd->drm_priv),
1324					     start, last, &fence);
1325		if (r)
1326			break;
1327
1328		if (fence) {
1329			r = dma_fence_wait(fence, false);
1330			dma_fence_put(fence);
1331			fence = NULL;
1332			if (r)
1333				break;
1334		}
1335		kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1336	}
1337
1338	return r;
1339}
1340
1341static int
1342svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1343		     unsigned long offset, unsigned long npages, bool readonly,
1344		     dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1345		     struct dma_fence **fence, bool flush_tlb)
1346{
1347	struct amdgpu_device *adev = pdd->dev->adev;
1348	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1349	uint64_t pte_flags;
1350	unsigned long last_start;
1351	int last_domain;
1352	int r = 0;
1353	int64_t i, j;
1354
1355	last_start = prange->start + offset;
1356
1357	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1358		 last_start, last_start + npages - 1, readonly);
1359
1360	for (i = offset; i < offset + npages; i++) {
1361		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1362		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1363
1364		/* Collect all pages in the same address range and memory domain
1365		 * that can be mapped with a single call to update mapping.
1366		 */
1367		if (i < offset + npages - 1 &&
1368		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1369			continue;
1370
1371		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1372			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1373
1374		pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain);
1375		if (readonly)
1376			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1377
1378		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1379			 prange->svms, last_start, prange->start + i,
1380			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1381			 pte_flags);
1382
1383		/* For dGPU mode, we use same vm_manager to allocate VRAM for
1384		 * different memory partition based on fpfn/lpfn, we should use
1385		 * same vm_manager.vram_base_offset regardless memory partition.
1386		 */
1387		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, true,
1388					   NULL, last_start, prange->start + i,
1389					   pte_flags,
1390					   (last_start - prange->start) << PAGE_SHIFT,
1391					   bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1392					   NULL, dma_addr, &vm->last_update);
1393
1394		for (j = last_start - prange->start; j <= i; j++)
1395			dma_addr[j] |= last_domain;
1396
1397		if (r) {
1398			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1399			goto out;
1400		}
1401		last_start = prange->start + i + 1;
1402	}
1403
1404	r = amdgpu_vm_update_pdes(adev, vm, false);
1405	if (r) {
1406		pr_debug("failed %d to update directories 0x%lx\n", r,
1407			 prange->start);
1408		goto out;
1409	}
1410
1411	if (fence)
1412		*fence = dma_fence_get(vm->last_update);
1413
1414out:
1415	return r;
1416}
1417
1418static int
1419svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1420		      unsigned long npages, bool readonly,
1421		      unsigned long *bitmap, bool wait, bool flush_tlb)
1422{
1423	struct kfd_process_device *pdd;
1424	struct amdgpu_device *bo_adev = NULL;
1425	struct kfd_process *p;
1426	struct dma_fence *fence = NULL;
1427	uint32_t gpuidx;
1428	int r = 0;
1429
1430	if (prange->svm_bo && prange->ttm_res)
1431		bo_adev = prange->svm_bo->node->adev;
1432
1433	p = container_of(prange->svms, struct kfd_process, svms);
1434	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1435		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1436		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1437		if (!pdd) {
1438			pr_debug("failed to find device idx %d\n", gpuidx);
1439			return -EINVAL;
1440		}
1441
1442		pdd = kfd_bind_process_to_device(pdd->dev, p);
1443		if (IS_ERR(pdd))
1444			return -EINVAL;
1445
1446		if (bo_adev && pdd->dev->adev != bo_adev &&
1447		    !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1448			pr_debug("cannot map to device idx %d\n", gpuidx);
1449			continue;
1450		}
1451
1452		r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1453					 prange->dma_addr[gpuidx],
1454					 bo_adev, wait ? &fence : NULL,
1455					 flush_tlb);
1456		if (r)
1457			break;
1458
1459		if (fence) {
1460			r = dma_fence_wait(fence, false);
1461			dma_fence_put(fence);
1462			fence = NULL;
1463			if (r) {
1464				pr_debug("failed %d to dma fence wait\n", r);
1465				break;
1466			}
1467		}
1468
1469		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1470	}
1471
1472	return r;
1473}
1474
1475struct svm_validate_context {
1476	struct kfd_process *process;
1477	struct svm_range *prange;
1478	bool intr;
1479	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1480	struct drm_exec exec;
1481};
1482
1483static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr)
1484{
1485	struct kfd_process_device *pdd;
1486	struct amdgpu_vm *vm;
1487	uint32_t gpuidx;
1488	int r;
1489
1490	drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0, 0);
1491	drm_exec_until_all_locked(&ctx->exec) {
1492		for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1493			pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1494			if (!pdd) {
1495				pr_debug("failed to find device idx %d\n", gpuidx);
1496				r = -EINVAL;
1497				goto unreserve_out;
1498			}
1499			vm = drm_priv_to_vm(pdd->drm_priv);
1500
1501			r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1502			drm_exec_retry_on_contention(&ctx->exec);
1503			if (unlikely(r)) {
1504				pr_debug("failed %d to reserve bo\n", r);
1505				goto unreserve_out;
1506			}
1507		}
1508	}
1509
1510	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1511		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1512		if (!pdd) {
1513			pr_debug("failed to find device idx %d\n", gpuidx);
1514			r = -EINVAL;
1515			goto unreserve_out;
1516		}
1517
1518		r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1519					      drm_priv_to_vm(pdd->drm_priv),
1520					      svm_range_bo_validate, NULL);
1521		if (r) {
1522			pr_debug("failed %d validate pt bos\n", r);
1523			goto unreserve_out;
1524		}
1525	}
1526
1527	return 0;
1528
1529unreserve_out:
1530	drm_exec_fini(&ctx->exec);
1531	return r;
1532}
1533
1534static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1535{
1536	drm_exec_fini(&ctx->exec);
1537}
1538
1539static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1540{
1541	struct kfd_process_device *pdd;
1542
1543	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1544	if (!pdd)
1545		return NULL;
1546
1547	return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1548}
1549
1550/*
1551 * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1552 *
1553 * To prevent concurrent destruction or change of range attributes, the
1554 * svm_read_lock must be held. The caller must not hold the svm_write_lock
1555 * because that would block concurrent evictions and lead to deadlocks. To
1556 * serialize concurrent migrations or validations of the same range, the
1557 * prange->migrate_mutex must be held.
1558 *
1559 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1560 * eviction fence.
1561 *
1562 * The following sequence ensures race-free validation and GPU mapping:
1563 *
1564 * 1. Reserve page table (and SVM BO if range is in VRAM)
1565 * 2. hmm_range_fault to get page addresses (if system memory)
1566 * 3. DMA-map pages (if system memory)
1567 * 4-a. Take notifier lock
1568 * 4-b. Check that pages still valid (mmu_interval_read_retry)
1569 * 4-c. Check that the range was not split or otherwise invalidated
1570 * 4-d. Update GPU page table
1571 * 4.e. Release notifier lock
1572 * 5. Release page table (and SVM BO) reservation
1573 */
1574static int svm_range_validate_and_map(struct mm_struct *mm,
1575				      unsigned long map_start, unsigned long map_last,
1576				      struct svm_range *prange, int32_t gpuidx,
1577				      bool intr, bool wait, bool flush_tlb)
1578{
1579	struct svm_validate_context *ctx;
1580	unsigned long start, end, addr;
1581	struct kfd_process *p;
1582	void *owner;
1583	int32_t idx;
1584	int r = 0;
1585
1586	ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
1587	if (!ctx)
1588		return -ENOMEM;
1589	ctx->process = container_of(prange->svms, struct kfd_process, svms);
1590	ctx->prange = prange;
1591	ctx->intr = intr;
1592
1593	if (gpuidx < MAX_GPU_INSTANCE) {
1594		bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE);
1595		bitmap_set(ctx->bitmap, gpuidx, 1);
1596	} else if (ctx->process->xnack_enabled) {
1597		bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1598
1599		/* If prefetch range to GPU, or GPU retry fault migrate range to
1600		 * GPU, which has ACCESS attribute to the range, create mapping
1601		 * on that GPU.
1602		 */
1603		if (prange->actual_loc) {
1604			gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process,
1605							prange->actual_loc);
1606			if (gpuidx < 0) {
1607				WARN_ONCE(1, "failed get device by id 0x%x\n",
1608					 prange->actual_loc);
1609				r = -EINVAL;
1610				goto free_ctx;
1611			}
1612			if (test_bit(gpuidx, prange->bitmap_access))
1613				bitmap_set(ctx->bitmap, gpuidx, 1);
1614		}
1615
1616		/*
1617		 * If prange is already mapped or with always mapped flag,
1618		 * update mapping on GPUs with ACCESS attribute
1619		 */
1620		if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1621			if (prange->mapped_to_gpu ||
1622			    prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)
1623				bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1624		}
1625	} else {
1626		bitmap_or(ctx->bitmap, prange->bitmap_access,
1627			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1628	}
1629
1630	if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1631		r = 0;
1632		goto free_ctx;
1633	}
1634
1635	if (prange->actual_loc && !prange->ttm_res) {
1636		/* This should never happen. actual_loc gets set by
1637		 * svm_migrate_ram_to_vram after allocating a BO.
1638		 */
1639		WARN_ONCE(1, "VRAM BO missing during validation\n");
1640		r = -EINVAL;
1641		goto free_ctx;
1642	}
1643
1644	svm_range_reserve_bos(ctx, intr);
1645
1646	p = container_of(prange->svms, struct kfd_process, svms);
1647	owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
1648						MAX_GPU_INSTANCE));
1649	for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) {
1650		if (kfd_svm_page_owner(p, idx) != owner) {
1651			owner = NULL;
1652			break;
1653		}
1654	}
1655
1656	start = map_start << PAGE_SHIFT;
1657	end = (map_last + 1) << PAGE_SHIFT;
1658	for (addr = start; !r && addr < end; ) {
1659		struct hmm_range *hmm_range;
1660		unsigned long map_start_vma;
1661		unsigned long map_last_vma;
1662		struct vm_area_struct *vma;
1663		unsigned long next = 0;
1664		unsigned long offset;
1665		unsigned long npages;
1666		bool readonly;
1667
1668		vma = vma_lookup(mm, addr);
1669		if (vma) {
1670			readonly = !(vma->vm_flags & VM_WRITE);
1671
1672			next = min(vma->vm_end, end);
1673			npages = (next - addr) >> PAGE_SHIFT;
1674			WRITE_ONCE(p->svms.faulting_task, current);
1675			r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages,
1676						       readonly, owner, NULL,
1677						       &hmm_range);
1678			WRITE_ONCE(p->svms.faulting_task, NULL);
1679			if (r) {
1680				pr_debug("failed %d to get svm range pages\n", r);
1681				if (r == -EBUSY)
1682					r = -EAGAIN;
1683			}
1684		} else {
1685			r = -EFAULT;
1686		}
1687
1688		if (!r) {
1689			offset = (addr >> PAGE_SHIFT) - prange->start;
1690			r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
1691					      hmm_range->hmm_pfns);
1692			if (r)
1693				pr_debug("failed %d to dma map range\n", r);
1694		}
1695
1696		svm_range_lock(prange);
1697		if (!r && amdgpu_hmm_range_get_pages_done(hmm_range)) {
1698			pr_debug("hmm update the range, need validate again\n");
1699			r = -EAGAIN;
1700		}
1701
1702		if (!r && !list_empty(&prange->child_list)) {
1703			pr_debug("range split by unmap in parallel, validate again\n");
1704			r = -EAGAIN;
1705		}
1706
1707		if (!r) {
1708			map_start_vma = max(map_start, prange->start + offset);
1709			map_last_vma = min(map_last, prange->start + offset + npages - 1);
1710			if (map_start_vma <= map_last_vma) {
1711				offset = map_start_vma - prange->start;
1712				npages = map_last_vma - map_start_vma + 1;
1713				r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1714							  ctx->bitmap, wait, flush_tlb);
1715			}
1716		}
1717
1718		if (!r && next == end)
1719			prange->mapped_to_gpu = true;
1720
1721		svm_range_unlock(prange);
1722
1723		addr = next;
1724	}
1725
1726	svm_range_unreserve_bos(ctx);
1727	if (!r)
1728		prange->validate_timestamp = ktime_get_boottime();
1729
1730free_ctx:
1731	kfree(ctx);
1732
1733	return r;
1734}
1735
1736/**
1737 * svm_range_list_lock_and_flush_work - flush pending deferred work
1738 *
1739 * @svms: the svm range list
1740 * @mm: the mm structure
1741 *
1742 * Context: Returns with mmap write lock held, pending deferred work flushed
1743 *
1744 */
1745void
1746svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1747				   struct mm_struct *mm)
1748{
1749retry_flush_work:
1750	flush_work(&svms->deferred_list_work);
1751	mmap_write_lock(mm);
1752
1753	if (list_empty(&svms->deferred_range_list))
1754		return;
1755	mmap_write_unlock(mm);
1756	pr_debug("retry flush\n");
1757	goto retry_flush_work;
1758}
1759
1760static void svm_range_restore_work(struct work_struct *work)
1761{
1762	struct delayed_work *dwork = to_delayed_work(work);
1763	struct amdkfd_process_info *process_info;
1764	struct svm_range_list *svms;
1765	struct svm_range *prange;
1766	struct kfd_process *p;
1767	struct mm_struct *mm;
1768	int evicted_ranges;
1769	int invalid;
1770	int r;
1771
1772	svms = container_of(dwork, struct svm_range_list, restore_work);
1773	evicted_ranges = atomic_read(&svms->evicted_ranges);
1774	if (!evicted_ranges)
1775		return;
1776
1777	pr_debug("restore svm ranges\n");
1778
1779	p = container_of(svms, struct kfd_process, svms);
1780	process_info = p->kgd_process_info;
1781
1782	/* Keep mm reference when svm_range_validate_and_map ranges */
1783	mm = get_task_mm(p->lead_thread);
1784	if (!mm) {
1785		pr_debug("svms 0x%p process mm gone\n", svms);
1786		return;
1787	}
1788
1789	mutex_lock(&process_info->lock);
1790	svm_range_list_lock_and_flush_work(svms, mm);
1791	mutex_lock(&svms->lock);
1792
1793	evicted_ranges = atomic_read(&svms->evicted_ranges);
1794
1795	list_for_each_entry(prange, &svms->list, list) {
1796		invalid = atomic_read(&prange->invalid);
1797		if (!invalid)
1798			continue;
1799
1800		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1801			 prange->svms, prange, prange->start, prange->last,
1802			 invalid);
1803
1804		/*
1805		 * If range is migrating, wait for migration is done.
1806		 */
1807		mutex_lock(&prange->migrate_mutex);
1808
1809		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
1810					       MAX_GPU_INSTANCE, false, true, false);
1811		if (r)
1812			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1813				 prange->start);
1814
1815		mutex_unlock(&prange->migrate_mutex);
1816		if (r)
1817			goto out_reschedule;
1818
1819		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1820			goto out_reschedule;
1821	}
1822
1823	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1824	    evicted_ranges)
1825		goto out_reschedule;
1826
1827	evicted_ranges = 0;
1828
1829	r = kgd2kfd_resume_mm(mm);
1830	if (r) {
1831		/* No recovery from this failure. Probably the CP is
1832		 * hanging. No point trying again.
1833		 */
1834		pr_debug("failed %d to resume KFD\n", r);
1835	}
1836
1837	pr_debug("restore svm ranges successfully\n");
1838
1839out_reschedule:
1840	mutex_unlock(&svms->lock);
1841	mmap_write_unlock(mm);
1842	mutex_unlock(&process_info->lock);
1843
1844	/* If validation failed, reschedule another attempt */
1845	if (evicted_ranges) {
1846		pr_debug("reschedule to restore svm range\n");
1847		queue_delayed_work(system_freezable_wq, &svms->restore_work,
1848			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1849
1850		kfd_smi_event_queue_restore_rescheduled(mm);
1851	}
1852	mmput(mm);
1853}
1854
1855/**
1856 * svm_range_evict - evict svm range
1857 * @prange: svm range structure
1858 * @mm: current process mm_struct
1859 * @start: starting process queue number
1860 * @last: last process queue number
1861 * @event: mmu notifier event when range is evicted or migrated
1862 *
1863 * Stop all queues of the process to ensure GPU doesn't access the memory, then
1864 * return to let CPU evict the buffer and proceed CPU pagetable update.
1865 *
1866 * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1867 * If invalidation happens while restore work is running, restore work will
1868 * restart to ensure to get the latest CPU pages mapping to GPU, then start
1869 * the queues.
1870 */
1871static int
1872svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1873		unsigned long start, unsigned long last,
1874		enum mmu_notifier_event event)
1875{
1876	struct svm_range_list *svms = prange->svms;
1877	struct svm_range *pchild;
1878	struct kfd_process *p;
1879	int r = 0;
1880
1881	p = container_of(svms, struct kfd_process, svms);
1882
1883	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1884		 svms, prange->start, prange->last, start, last);
1885
1886	if (!p->xnack_enabled ||
1887	    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1888		int evicted_ranges;
1889		bool mapped = prange->mapped_to_gpu;
1890
1891		list_for_each_entry(pchild, &prange->child_list, child_list) {
1892			if (!pchild->mapped_to_gpu)
1893				continue;
1894			mapped = true;
1895			mutex_lock_nested(&pchild->lock, 1);
1896			if (pchild->start <= last && pchild->last >= start) {
1897				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1898					 pchild->start, pchild->last);
1899				atomic_inc(&pchild->invalid);
1900			}
1901			mutex_unlock(&pchild->lock);
1902		}
1903
1904		if (!mapped)
1905			return r;
1906
1907		if (prange->start <= last && prange->last >= start)
1908			atomic_inc(&prange->invalid);
1909
1910		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1911		if (evicted_ranges != 1)
1912			return r;
1913
1914		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1915			 prange->svms, prange->start, prange->last);
1916
1917		/* First eviction, stop the queues */
1918		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1919		if (r)
1920			pr_debug("failed to quiesce KFD\n");
1921
1922		pr_debug("schedule to restore svm %p ranges\n", svms);
1923		queue_delayed_work(system_freezable_wq, &svms->restore_work,
1924			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1925	} else {
1926		unsigned long s, l;
1927		uint32_t trigger;
1928
1929		if (event == MMU_NOTIFY_MIGRATE)
1930			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1931		else
1932			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1933
1934		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1935			 prange->svms, start, last);
1936		list_for_each_entry(pchild, &prange->child_list, child_list) {
1937			mutex_lock_nested(&pchild->lock, 1);
1938			s = max(start, pchild->start);
1939			l = min(last, pchild->last);
1940			if (l >= s)
1941				svm_range_unmap_from_gpus(pchild, s, l, trigger);
1942			mutex_unlock(&pchild->lock);
1943		}
1944		s = max(start, prange->start);
1945		l = min(last, prange->last);
1946		if (l >= s)
1947			svm_range_unmap_from_gpus(prange, s, l, trigger);
1948	}
1949
1950	return r;
1951}
1952
1953static struct svm_range *svm_range_clone(struct svm_range *old)
1954{
1955	struct svm_range *new;
1956
1957	new = svm_range_new(old->svms, old->start, old->last, false);
1958	if (!new)
1959		return NULL;
1960	if (svm_range_copy_dma_addrs(new, old)) {
1961		svm_range_free(new, false);
1962		return NULL;
1963	}
1964	if (old->svm_bo) {
1965		new->ttm_res = old->ttm_res;
1966		new->offset = old->offset;
1967		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1968		spin_lock(&new->svm_bo->list_lock);
1969		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1970		spin_unlock(&new->svm_bo->list_lock);
1971	}
1972	new->flags = old->flags;
1973	new->preferred_loc = old->preferred_loc;
1974	new->prefetch_loc = old->prefetch_loc;
1975	new->actual_loc = old->actual_loc;
1976	new->granularity = old->granularity;
1977	new->mapped_to_gpu = old->mapped_to_gpu;
1978	new->vram_pages = old->vram_pages;
1979	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1980	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1981
1982	return new;
1983}
1984
1985void svm_range_set_max_pages(struct amdgpu_device *adev)
1986{
1987	uint64_t max_pages;
1988	uint64_t pages, _pages;
1989	uint64_t min_pages = 0;
1990	int i, id;
1991
1992	for (i = 0; i < adev->kfd.dev->num_nodes; i++) {
1993		if (adev->kfd.dev->nodes[i]->xcp)
1994			id = adev->kfd.dev->nodes[i]->xcp->id;
1995		else
1996			id = -1;
1997		pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17;
1998		pages = clamp(pages, 1ULL << 9, 1ULL << 18);
1999		pages = rounddown_pow_of_two(pages);
2000		min_pages = min_not_zero(min_pages, pages);
2001	}
2002
2003	do {
2004		max_pages = READ_ONCE(max_svm_range_pages);
2005		_pages = min_not_zero(max_pages, min_pages);
2006	} while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
2007}
2008
2009static int
2010svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
2011		    uint64_t max_pages, struct list_head *insert_list,
2012		    struct list_head *update_list)
2013{
2014	struct svm_range *prange;
2015	uint64_t l;
2016
2017	pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
2018		 max_pages, start, last);
2019
2020	while (last >= start) {
2021		l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
2022
2023		prange = svm_range_new(svms, start, l, true);
2024		if (!prange)
2025			return -ENOMEM;
2026		list_add(&prange->list, insert_list);
2027		list_add(&prange->update_list, update_list);
2028
2029		start = l + 1;
2030	}
2031	return 0;
2032}
2033
2034/**
2035 * svm_range_add - add svm range and handle overlap
2036 * @p: the range add to this process svms
2037 * @start: page size aligned
2038 * @size: page size aligned
2039 * @nattr: number of attributes
2040 * @attrs: array of attributes
2041 * @update_list: output, the ranges need validate and update GPU mapping
2042 * @insert_list: output, the ranges need insert to svms
2043 * @remove_list: output, the ranges are replaced and need remove from svms
2044 * @remap_list: output, remap unaligned svm ranges
2045 *
2046 * Check if the virtual address range has overlap with any existing ranges,
2047 * split partly overlapping ranges and add new ranges in the gaps. All changes
2048 * should be applied to the range_list and interval tree transactionally. If
2049 * any range split or allocation fails, the entire update fails. Therefore any
2050 * existing overlapping svm_ranges are cloned and the original svm_ranges left
2051 * unchanged.
2052 *
2053 * If the transaction succeeds, the caller can update and insert clones and
2054 * new ranges, then free the originals.
2055 *
2056 * Otherwise the caller can free the clones and new ranges, while the old
2057 * svm_ranges remain unchanged.
2058 *
2059 * Context: Process context, caller must hold svms->lock
2060 *
2061 * Return:
2062 * 0 - OK, otherwise error code
2063 */
2064static int
2065svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2066	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2067	      struct list_head *update_list, struct list_head *insert_list,
2068	      struct list_head *remove_list, struct list_head *remap_list)
2069{
2070	unsigned long last = start + size - 1UL;
2071	struct svm_range_list *svms = &p->svms;
2072	struct interval_tree_node *node;
2073	struct svm_range *prange;
2074	struct svm_range *tmp;
2075	struct list_head new_list;
2076	int r = 0;
2077
2078	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
2079
2080	INIT_LIST_HEAD(update_list);
2081	INIT_LIST_HEAD(insert_list);
2082	INIT_LIST_HEAD(remove_list);
2083	INIT_LIST_HEAD(&new_list);
2084	INIT_LIST_HEAD(remap_list);
2085
2086	node = interval_tree_iter_first(&svms->objects, start, last);
2087	while (node) {
2088		struct interval_tree_node *next;
2089		unsigned long next_start;
2090
2091		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2092			 node->last);
2093
2094		prange = container_of(node, struct svm_range, it_node);
2095		next = interval_tree_iter_next(node, start, last);
2096		next_start = min(node->last, last) + 1;
2097
2098		if (svm_range_is_same_attrs(p, prange, nattr, attrs) &&
2099		    prange->mapped_to_gpu) {
2100			/* nothing to do */
2101		} else if (node->start < start || node->last > last) {
2102			/* node intersects the update range and its attributes
2103			 * will change. Clone and split it, apply updates only
2104			 * to the overlapping part
2105			 */
2106			struct svm_range *old = prange;
2107
2108			prange = svm_range_clone(old);
2109			if (!prange) {
2110				r = -ENOMEM;
2111				goto out;
2112			}
2113
2114			list_add(&old->update_list, remove_list);
2115			list_add(&prange->list, insert_list);
2116			list_add(&prange->update_list, update_list);
2117
2118			if (node->start < start) {
2119				pr_debug("change old range start\n");
2120				r = svm_range_split_head(prange, start,
2121							 insert_list, remap_list);
2122				if (r)
2123					goto out;
2124			}
2125			if (node->last > last) {
2126				pr_debug("change old range last\n");
2127				r = svm_range_split_tail(prange, last,
2128							 insert_list, remap_list);
2129				if (r)
2130					goto out;
2131			}
2132		} else {
2133			/* The node is contained within start..last,
2134			 * just update it
2135			 */
2136			list_add(&prange->update_list, update_list);
2137		}
2138
2139		/* insert a new node if needed */
2140		if (node->start > start) {
2141			r = svm_range_split_new(svms, start, node->start - 1,
2142						READ_ONCE(max_svm_range_pages),
2143						&new_list, update_list);
2144			if (r)
2145				goto out;
2146		}
2147
2148		node = next;
2149		start = next_start;
2150	}
2151
2152	/* add a final range at the end if needed */
2153	if (start <= last)
2154		r = svm_range_split_new(svms, start, last,
2155					READ_ONCE(max_svm_range_pages),
2156					&new_list, update_list);
2157
2158out:
2159	if (r) {
2160		list_for_each_entry_safe(prange, tmp, insert_list, list)
2161			svm_range_free(prange, false);
2162		list_for_each_entry_safe(prange, tmp, &new_list, list)
2163			svm_range_free(prange, true);
2164	} else {
2165		list_splice(&new_list, insert_list);
2166	}
2167
2168	return r;
2169}
2170
2171static void
2172svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2173					    struct svm_range *prange)
2174{
2175	unsigned long start;
2176	unsigned long last;
2177
2178	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2179	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2180
2181	if (prange->start == start && prange->last == last)
2182		return;
2183
2184	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2185		  prange->svms, prange, start, last, prange->start,
2186		  prange->last);
2187
2188	if (start != 0 && last != 0) {
2189		interval_tree_remove(&prange->it_node, &prange->svms->objects);
2190		svm_range_remove_notifier(prange);
2191	}
2192	prange->it_node.start = prange->start;
2193	prange->it_node.last = prange->last;
2194
2195	interval_tree_insert(&prange->it_node, &prange->svms->objects);
2196	svm_range_add_notifier_locked(mm, prange);
2197}
2198
2199static void
2200svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2201			 struct mm_struct *mm)
2202{
2203	switch (prange->work_item.op) {
2204	case SVM_OP_NULL:
2205		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2206			 svms, prange, prange->start, prange->last);
2207		break;
2208	case SVM_OP_UNMAP_RANGE:
2209		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2210			 svms, prange, prange->start, prange->last);
2211		svm_range_unlink(prange);
2212		svm_range_remove_notifier(prange);
2213		svm_range_free(prange, true);
2214		break;
2215	case SVM_OP_UPDATE_RANGE_NOTIFIER:
2216		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2217			 svms, prange, prange->start, prange->last);
2218		svm_range_update_notifier_and_interval_tree(mm, prange);
2219		break;
2220	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2221		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2222			 svms, prange, prange->start, prange->last);
2223		svm_range_update_notifier_and_interval_tree(mm, prange);
2224		/* TODO: implement deferred validation and mapping */
2225		break;
2226	case SVM_OP_ADD_RANGE:
2227		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2228			 prange->start, prange->last);
2229		svm_range_add_to_svms(prange);
2230		svm_range_add_notifier_locked(mm, prange);
2231		break;
2232	case SVM_OP_ADD_RANGE_AND_MAP:
2233		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2234			 prange, prange->start, prange->last);
2235		svm_range_add_to_svms(prange);
2236		svm_range_add_notifier_locked(mm, prange);
2237		/* TODO: implement deferred validation and mapping */
2238		break;
2239	default:
2240		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2241			 prange->work_item.op);
2242	}
2243}
2244
2245static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2246{
2247	struct kfd_process_device *pdd;
2248	struct kfd_process *p;
2249	int drain;
2250	uint32_t i;
2251
2252	p = container_of(svms, struct kfd_process, svms);
2253
2254restart:
2255	drain = atomic_read(&svms->drain_pagefaults);
2256	if (!drain)
2257		return;
2258
2259	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2260		pdd = p->pdds[i];
2261		if (!pdd)
2262			continue;
2263
2264		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2265
2266		amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2267				pdd->dev->adev->irq.retry_cam_enabled ?
2268				&pdd->dev->adev->irq.ih :
2269				&pdd->dev->adev->irq.ih1);
2270
2271		if (pdd->dev->adev->irq.retry_cam_enabled)
2272			amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2273				&pdd->dev->adev->irq.ih_soft);
2274
2275
2276		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2277	}
2278	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2279		goto restart;
2280}
2281
2282static void svm_range_deferred_list_work(struct work_struct *work)
2283{
2284	struct svm_range_list *svms;
2285	struct svm_range *prange;
2286	struct mm_struct *mm;
2287
2288	svms = container_of(work, struct svm_range_list, deferred_list_work);
2289	pr_debug("enter svms 0x%p\n", svms);
2290
2291	spin_lock(&svms->deferred_list_lock);
2292	while (!list_empty(&svms->deferred_range_list)) {
2293		prange = list_first_entry(&svms->deferred_range_list,
2294					  struct svm_range, deferred_list);
2295		spin_unlock(&svms->deferred_list_lock);
2296
2297		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2298			 prange->start, prange->last, prange->work_item.op);
2299
2300		mm = prange->work_item.mm;
2301retry:
2302		mmap_write_lock(mm);
2303
2304		/* Checking for the need to drain retry faults must be inside
2305		 * mmap write lock to serialize with munmap notifiers.
2306		 */
2307		if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2308			mmap_write_unlock(mm);
2309			svm_range_drain_retry_fault(svms);
2310			goto retry;
2311		}
2312
2313		/* Remove from deferred_list must be inside mmap write lock, for
2314		 * two race cases:
2315		 * 1. unmap_from_cpu may change work_item.op and add the range
2316		 *    to deferred_list again, cause use after free bug.
2317		 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2318		 *    lock and continue because deferred_list is empty, but
2319		 *    deferred_list work is actually waiting for mmap lock.
2320		 */
2321		spin_lock(&svms->deferred_list_lock);
2322		list_del_init(&prange->deferred_list);
2323		spin_unlock(&svms->deferred_list_lock);
2324
2325		mutex_lock(&svms->lock);
2326		mutex_lock(&prange->migrate_mutex);
2327		while (!list_empty(&prange->child_list)) {
2328			struct svm_range *pchild;
2329
2330			pchild = list_first_entry(&prange->child_list,
2331						struct svm_range, child_list);
2332			pr_debug("child prange 0x%p op %d\n", pchild,
2333				 pchild->work_item.op);
2334			list_del_init(&pchild->child_list);
2335			svm_range_handle_list_op(svms, pchild, mm);
2336		}
2337		mutex_unlock(&prange->migrate_mutex);
2338
2339		svm_range_handle_list_op(svms, prange, mm);
2340		mutex_unlock(&svms->lock);
2341		mmap_write_unlock(mm);
2342
2343		/* Pairs with mmget in svm_range_add_list_work. If dropping the
2344		 * last mm refcount, schedule release work to avoid circular locking
2345		 */
2346		mmput_async(mm);
2347
2348		spin_lock(&svms->deferred_list_lock);
2349	}
2350	spin_unlock(&svms->deferred_list_lock);
2351	pr_debug("exit svms 0x%p\n", svms);
2352}
2353
2354void
2355svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2356			struct mm_struct *mm, enum svm_work_list_ops op)
2357{
2358	spin_lock(&svms->deferred_list_lock);
2359	/* if prange is on the deferred list */
2360	if (!list_empty(&prange->deferred_list)) {
2361		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2362		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2363		if (op != SVM_OP_NULL &&
2364		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2365			prange->work_item.op = op;
2366	} else {
2367		prange->work_item.op = op;
2368
2369		/* Pairs with mmput in deferred_list_work */
2370		mmget(mm);
2371		prange->work_item.mm = mm;
2372		list_add_tail(&prange->deferred_list,
2373			      &prange->svms->deferred_range_list);
2374		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2375			 prange, prange->start, prange->last, op);
2376	}
2377	spin_unlock(&svms->deferred_list_lock);
2378}
2379
2380void schedule_deferred_list_work(struct svm_range_list *svms)
2381{
2382	spin_lock(&svms->deferred_list_lock);
2383	if (!list_empty(&svms->deferred_range_list))
2384		schedule_work(&svms->deferred_list_work);
2385	spin_unlock(&svms->deferred_list_lock);
2386}
2387
2388static void
2389svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2390		      struct svm_range *prange, unsigned long start,
2391		      unsigned long last)
2392{
2393	struct svm_range *head;
2394	struct svm_range *tail;
2395
2396	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2397		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2398			 prange->start, prange->last);
2399		return;
2400	}
2401	if (start > prange->last || last < prange->start)
2402		return;
2403
2404	head = tail = prange;
2405	if (start > prange->start)
2406		svm_range_split(prange, prange->start, start - 1, &tail);
2407	if (last < tail->last)
2408		svm_range_split(tail, last + 1, tail->last, &head);
2409
2410	if (head != prange && tail != prange) {
2411		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2412		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2413	} else if (tail != prange) {
2414		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2415	} else if (head != prange) {
2416		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2417	} else if (parent != prange) {
2418		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2419	}
2420}
2421
2422static void
2423svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2424			 unsigned long start, unsigned long last)
2425{
2426	uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU;
2427	struct svm_range_list *svms;
2428	struct svm_range *pchild;
2429	struct kfd_process *p;
2430	unsigned long s, l;
2431	bool unmap_parent;
2432
2433	p = kfd_lookup_process_by_mm(mm);
2434	if (!p)
2435		return;
2436	svms = &p->svms;
2437
2438	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2439		 prange, prange->start, prange->last, start, last);
2440
2441	/* Make sure pending page faults are drained in the deferred worker
2442	 * before the range is freed to avoid straggler interrupts on
2443	 * unmapped memory causing "phantom faults".
2444	 */
2445	atomic_inc(&svms->drain_pagefaults);
2446
2447	unmap_parent = start <= prange->start && last >= prange->last;
2448
2449	list_for_each_entry(pchild, &prange->child_list, child_list) {
2450		mutex_lock_nested(&pchild->lock, 1);
2451		s = max(start, pchild->start);
2452		l = min(last, pchild->last);
2453		if (l >= s)
2454			svm_range_unmap_from_gpus(pchild, s, l, trigger);
2455		svm_range_unmap_split(mm, prange, pchild, start, last);
2456		mutex_unlock(&pchild->lock);
2457	}
2458	s = max(start, prange->start);
2459	l = min(last, prange->last);
2460	if (l >= s)
2461		svm_range_unmap_from_gpus(prange, s, l, trigger);
2462	svm_range_unmap_split(mm, prange, prange, start, last);
2463
2464	if (unmap_parent)
2465		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2466	else
2467		svm_range_add_list_work(svms, prange, mm,
2468					SVM_OP_UPDATE_RANGE_NOTIFIER);
2469	schedule_deferred_list_work(svms);
2470
2471	kfd_unref_process(p);
2472}
2473
2474/**
2475 * svm_range_cpu_invalidate_pagetables - interval notifier callback
2476 * @mni: mmu_interval_notifier struct
2477 * @range: mmu_notifier_range struct
2478 * @cur_seq: value to pass to mmu_interval_set_seq()
2479 *
2480 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2481 * is from migration, or CPU page invalidation callback.
2482 *
2483 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2484 * work thread, and split prange if only part of prange is unmapped.
2485 *
2486 * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2487 * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2488 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2489 * update GPU mapping to recover.
2490 *
2491 * Context: mmap lock, notifier_invalidate_start lock are held
2492 *          for invalidate event, prange lock is held if this is from migration
2493 */
2494static bool
2495svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2496				    const struct mmu_notifier_range *range,
2497				    unsigned long cur_seq)
2498{
2499	struct svm_range *prange;
2500	unsigned long start;
2501	unsigned long last;
2502
2503	if (range->event == MMU_NOTIFY_RELEASE)
2504		return true;
2505	if (!mmget_not_zero(mni->mm))
2506		return true;
2507
2508	start = mni->interval_tree.start;
2509	last = mni->interval_tree.last;
2510	start = max(start, range->start) >> PAGE_SHIFT;
2511	last = min(last, range->end - 1) >> PAGE_SHIFT;
2512	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2513		 start, last, range->start >> PAGE_SHIFT,
2514		 (range->end - 1) >> PAGE_SHIFT,
2515		 mni->interval_tree.start >> PAGE_SHIFT,
2516		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2517
2518	prange = container_of(mni, struct svm_range, notifier);
2519
2520	svm_range_lock(prange);
2521	mmu_interval_set_seq(mni, cur_seq);
2522
2523	switch (range->event) {
2524	case MMU_NOTIFY_UNMAP:
2525		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2526		break;
2527	default:
2528		svm_range_evict(prange, mni->mm, start, last, range->event);
2529		break;
2530	}
2531
2532	svm_range_unlock(prange);
2533	mmput(mni->mm);
2534
2535	return true;
2536}
2537
2538/**
2539 * svm_range_from_addr - find svm range from fault address
2540 * @svms: svm range list header
2541 * @addr: address to search range interval tree, in pages
2542 * @parent: parent range if range is on child list
2543 *
2544 * Context: The caller must hold svms->lock
2545 *
2546 * Return: the svm_range found or NULL
2547 */
2548struct svm_range *
2549svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2550		    struct svm_range **parent)
2551{
2552	struct interval_tree_node *node;
2553	struct svm_range *prange;
2554	struct svm_range *pchild;
2555
2556	node = interval_tree_iter_first(&svms->objects, addr, addr);
2557	if (!node)
2558		return NULL;
2559
2560	prange = container_of(node, struct svm_range, it_node);
2561	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2562		 addr, prange->start, prange->last, node->start, node->last);
2563
2564	if (addr >= prange->start && addr <= prange->last) {
2565		if (parent)
2566			*parent = prange;
2567		return prange;
2568	}
2569	list_for_each_entry(pchild, &prange->child_list, child_list)
2570		if (addr >= pchild->start && addr <= pchild->last) {
2571			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2572				 addr, pchild->start, pchild->last);
2573			if (parent)
2574				*parent = prange;
2575			return pchild;
2576		}
2577
2578	return NULL;
2579}
2580
2581/* svm_range_best_restore_location - decide the best fault restore location
2582 * @prange: svm range structure
2583 * @adev: the GPU on which vm fault happened
2584 *
2585 * This is only called when xnack is on, to decide the best location to restore
2586 * the range mapping after GPU vm fault. Caller uses the best location to do
2587 * migration if actual loc is not best location, then update GPU page table
2588 * mapping to the best location.
2589 *
2590 * If the preferred loc is accessible by faulting GPU, use preferred loc.
2591 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2592 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2593 *    if range actual loc is cpu, best_loc is cpu
2594 *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2595 *    range actual loc.
2596 * Otherwise, GPU no access, best_loc is -1.
2597 *
2598 * Return:
2599 * -1 means vm fault GPU no access
2600 * 0 for CPU or GPU id
2601 */
2602static int32_t
2603svm_range_best_restore_location(struct svm_range *prange,
2604				struct kfd_node *node,
2605				int32_t *gpuidx)
2606{
2607	struct kfd_node *bo_node, *preferred_node;
2608	struct kfd_process *p;
2609	uint32_t gpuid;
2610	int r;
2611
2612	p = container_of(prange->svms, struct kfd_process, svms);
2613
2614	r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx);
2615	if (r < 0) {
2616		pr_debug("failed to get gpuid from kgd\n");
2617		return -1;
2618	}
2619
2620	if (node->adev->gmc.is_app_apu)
2621		return 0;
2622
2623	if (prange->preferred_loc == gpuid ||
2624	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2625		return prange->preferred_loc;
2626	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2627		preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc);
2628		if (preferred_node && svm_nodes_in_same_hive(node, preferred_node))
2629			return prange->preferred_loc;
2630		/* fall through */
2631	}
2632
2633	if (test_bit(*gpuidx, prange->bitmap_access))
2634		return gpuid;
2635
2636	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2637		if (!prange->actual_loc)
2638			return 0;
2639
2640		bo_node = svm_range_get_node_by_id(prange, prange->actual_loc);
2641		if (bo_node && svm_nodes_in_same_hive(node, bo_node))
2642			return prange->actual_loc;
2643		else
2644			return 0;
2645	}
2646
2647	return -1;
2648}
2649
2650static int
2651svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2652			       unsigned long *start, unsigned long *last,
2653			       bool *is_heap_stack)
2654{
2655	struct vm_area_struct *vma;
2656	struct interval_tree_node *node;
2657	struct rb_node *rb_node;
2658	unsigned long start_limit, end_limit;
2659
2660	vma = vma_lookup(p->mm, addr << PAGE_SHIFT);
2661	if (!vma) {
2662		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2663		return -EFAULT;
2664	}
2665
2666	*is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma);
2667
2668	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2669		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2670	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2671		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2672	/* First range that starts after the fault address */
2673	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2674	if (node) {
2675		end_limit = min(end_limit, node->start);
2676		/* Last range that ends before the fault address */
2677		rb_node = rb_prev(&node->rb);
2678	} else {
2679		/* Last range must end before addr because
2680		 * there was no range after addr
2681		 */
2682		rb_node = rb_last(&p->svms.objects.rb_root);
2683	}
2684	if (rb_node) {
2685		node = container_of(rb_node, struct interval_tree_node, rb);
2686		if (node->last >= addr) {
2687			WARN(1, "Overlap with prev node and page fault addr\n");
2688			return -EFAULT;
2689		}
2690		start_limit = max(start_limit, node->last + 1);
2691	}
2692
2693	*start = start_limit;
2694	*last = end_limit - 1;
2695
2696	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2697		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2698		 *start, *last, *is_heap_stack);
2699
2700	return 0;
2701}
2702
2703static int
2704svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2705			   uint64_t *bo_s, uint64_t *bo_l)
2706{
2707	struct amdgpu_bo_va_mapping *mapping;
2708	struct interval_tree_node *node;
2709	struct amdgpu_bo *bo = NULL;
2710	unsigned long userptr;
2711	uint32_t i;
2712	int r;
2713
2714	for (i = 0; i < p->n_pdds; i++) {
2715		struct amdgpu_vm *vm;
2716
2717		if (!p->pdds[i]->drm_priv)
2718			continue;
2719
2720		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2721		r = amdgpu_bo_reserve(vm->root.bo, false);
2722		if (r)
2723			return r;
2724
2725		/* Check userptr by searching entire vm->va interval tree */
2726		node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2727		while (node) {
2728			mapping = container_of((struct rb_node *)node,
2729					       struct amdgpu_bo_va_mapping, rb);
2730			bo = mapping->bo_va->base.bo;
2731
2732			if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2733							 start << PAGE_SHIFT,
2734							 last << PAGE_SHIFT,
2735							 &userptr)) {
2736				node = interval_tree_iter_next(node, 0, ~0ULL);
2737				continue;
2738			}
2739
2740			pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2741				 start, last);
2742			if (bo_s && bo_l) {
2743				*bo_s = userptr >> PAGE_SHIFT;
2744				*bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2745			}
2746			amdgpu_bo_unreserve(vm->root.bo);
2747			return -EADDRINUSE;
2748		}
2749		amdgpu_bo_unreserve(vm->root.bo);
2750	}
2751	return 0;
2752}
2753
2754static struct
2755svm_range *svm_range_create_unregistered_range(struct kfd_node *node,
2756						struct kfd_process *p,
2757						struct mm_struct *mm,
2758						int64_t addr)
2759{
2760	struct svm_range *prange = NULL;
2761	unsigned long start, last;
2762	uint32_t gpuid, gpuidx;
2763	bool is_heap_stack;
2764	uint64_t bo_s = 0;
2765	uint64_t bo_l = 0;
2766	int r;
2767
2768	if (svm_range_get_range_boundaries(p, addr, &start, &last,
2769					   &is_heap_stack))
2770		return NULL;
2771
2772	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2773	if (r != -EADDRINUSE)
2774		r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2775
2776	if (r == -EADDRINUSE) {
2777		if (addr >= bo_s && addr <= bo_l)
2778			return NULL;
2779
2780		/* Create one page svm range if 2MB range overlapping */
2781		start = addr;
2782		last = addr;
2783	}
2784
2785	prange = svm_range_new(&p->svms, start, last, true);
2786	if (!prange) {
2787		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2788		return NULL;
2789	}
2790	if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) {
2791		pr_debug("failed to get gpuid from kgd\n");
2792		svm_range_free(prange, true);
2793		return NULL;
2794	}
2795
2796	if (is_heap_stack)
2797		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2798
2799	svm_range_add_to_svms(prange);
2800	svm_range_add_notifier_locked(mm, prange);
2801
2802	return prange;
2803}
2804
2805/* svm_range_skip_recover - decide if prange can be recovered
2806 * @prange: svm range structure
2807 *
2808 * GPU vm retry fault handle skip recover the range for cases:
2809 * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2810 *    deferred list work will drain the stale fault before free the prange.
2811 * 2. prange is on deferred list to add interval notifier after split, or
2812 * 3. prange is child range, it is split from parent prange, recover later
2813 *    after interval notifier is added.
2814 *
2815 * Return: true to skip recover, false to recover
2816 */
2817static bool svm_range_skip_recover(struct svm_range *prange)
2818{
2819	struct svm_range_list *svms = prange->svms;
2820
2821	spin_lock(&svms->deferred_list_lock);
2822	if (list_empty(&prange->deferred_list) &&
2823	    list_empty(&prange->child_list)) {
2824		spin_unlock(&svms->deferred_list_lock);
2825		return false;
2826	}
2827	spin_unlock(&svms->deferred_list_lock);
2828
2829	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2830		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2831			 svms, prange, prange->start, prange->last);
2832		return true;
2833	}
2834	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2835	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2836		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2837			 svms, prange, prange->start, prange->last);
2838		return true;
2839	}
2840	return false;
2841}
2842
2843static void
2844svm_range_count_fault(struct kfd_node *node, struct kfd_process *p,
2845		      int32_t gpuidx)
2846{
2847	struct kfd_process_device *pdd;
2848
2849	/* fault is on different page of same range
2850	 * or fault is skipped to recover later
2851	 * or fault is on invalid virtual address
2852	 */
2853	if (gpuidx == MAX_GPU_INSTANCE) {
2854		uint32_t gpuid;
2855		int r;
2856
2857		r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx);
2858		if (r < 0)
2859			return;
2860	}
2861
2862	/* fault is recovered
2863	 * or fault cannot recover because GPU no access on the range
2864	 */
2865	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2866	if (pdd)
2867		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2868}
2869
2870static bool
2871svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2872{
2873	unsigned long requested = VM_READ;
2874
2875	if (write_fault)
2876		requested |= VM_WRITE;
2877
2878	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2879		vma->vm_flags);
2880	return (vma->vm_flags & requested) == requested;
2881}
2882
2883int
2884svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2885			uint32_t vmid, uint32_t node_id,
2886			uint64_t addr, bool write_fault)
2887{
2888	unsigned long start, last, size;
2889	struct mm_struct *mm = NULL;
2890	struct svm_range_list *svms;
2891	struct svm_range *prange;
2892	struct kfd_process *p;
2893	ktime_t timestamp = ktime_get_boottime();
2894	struct kfd_node *node;
2895	int32_t best_loc;
2896	int32_t gpuidx = MAX_GPU_INSTANCE;
2897	bool write_locked = false;
2898	struct vm_area_struct *vma;
2899	bool migration = false;
2900	int r = 0;
2901
2902	if (!KFD_IS_SVM_API_SUPPORTED(adev)) {
2903		pr_debug("device does not support SVM\n");
2904		return -EFAULT;
2905	}
2906
2907	p = kfd_lookup_process_by_pasid(pasid);
2908	if (!p) {
2909		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2910		return 0;
2911	}
2912	svms = &p->svms;
2913
2914	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2915
2916	if (atomic_read(&svms->drain_pagefaults)) {
2917		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2918		r = 0;
2919		goto out;
2920	}
2921
2922	if (!p->xnack_enabled) {
2923		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2924		r = -EFAULT;
2925		goto out;
2926	}
2927
2928	/* p->lead_thread is available as kfd_process_wq_release flush the work
2929	 * before releasing task ref.
2930	 */
2931	mm = get_task_mm(p->lead_thread);
2932	if (!mm) {
2933		pr_debug("svms 0x%p failed to get mm\n", svms);
2934		r = 0;
2935		goto out;
2936	}
2937
2938	node = kfd_node_by_irq_ids(adev, node_id, vmid);
2939	if (!node) {
2940		pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id,
2941			 vmid);
2942		r = -EFAULT;
2943		goto out;
2944	}
2945	mmap_read_lock(mm);
2946retry_write_locked:
2947	mutex_lock(&svms->lock);
2948	prange = svm_range_from_addr(svms, addr, NULL);
2949	if (!prange) {
2950		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2951			 svms, addr);
2952		if (!write_locked) {
2953			/* Need the write lock to create new range with MMU notifier.
2954			 * Also flush pending deferred work to make sure the interval
2955			 * tree is up to date before we add a new range
2956			 */
2957			mutex_unlock(&svms->lock);
2958			mmap_read_unlock(mm);
2959			mmap_write_lock(mm);
2960			write_locked = true;
2961			goto retry_write_locked;
2962		}
2963		prange = svm_range_create_unregistered_range(node, p, mm, addr);
2964		if (!prange) {
2965			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2966				 svms, addr);
2967			mmap_write_downgrade(mm);
2968			r = -EFAULT;
2969			goto out_unlock_svms;
2970		}
2971	}
2972	if (write_locked)
2973		mmap_write_downgrade(mm);
2974
2975	mutex_lock(&prange->migrate_mutex);
2976
2977	if (svm_range_skip_recover(prange)) {
2978		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
2979		r = 0;
2980		goto out_unlock_range;
2981	}
2982
2983	/* skip duplicate vm fault on different pages of same range */
2984	if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
2985				AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
2986		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2987			 svms, prange->start, prange->last);
2988		r = 0;
2989		goto out_unlock_range;
2990	}
2991
2992	/* __do_munmap removed VMA, return success as we are handling stale
2993	 * retry fault.
2994	 */
2995	vma = vma_lookup(mm, addr << PAGE_SHIFT);
2996	if (!vma) {
2997		pr_debug("address 0x%llx VMA is removed\n", addr);
2998		r = 0;
2999		goto out_unlock_range;
3000	}
3001
3002	if (!svm_fault_allowed(vma, write_fault)) {
3003		pr_debug("fault addr 0x%llx no %s permission\n", addr,
3004			write_fault ? "write" : "read");
3005		r = -EPERM;
3006		goto out_unlock_range;
3007	}
3008
3009	best_loc = svm_range_best_restore_location(prange, node, &gpuidx);
3010	if (best_loc == -1) {
3011		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
3012			 svms, prange->start, prange->last);
3013		r = -EACCES;
3014		goto out_unlock_range;
3015	}
3016
3017	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
3018		 svms, prange->start, prange->last, best_loc,
3019		 prange->actual_loc);
3020
3021	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
3022				       write_fault, timestamp);
3023
3024	/* Align migration range start and size to granularity size */
3025	size = 1UL << prange->granularity;
3026	start = max_t(unsigned long, ALIGN_DOWN(addr, size), prange->start);
3027	last = min_t(unsigned long, ALIGN(addr + 1, size) - 1, prange->last);
3028	if (prange->actual_loc != 0 || best_loc != 0) {
3029		migration = true;
3030
3031		if (best_loc) {
3032			r = svm_migrate_to_vram(prange, best_loc, start, last,
3033					mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
3034			if (r) {
3035				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
3036					 r, addr);
3037				/* Fallback to system memory if migration to
3038				 * VRAM failed
3039				 */
3040				if (prange->actual_loc && prange->actual_loc != best_loc)
3041					r = svm_migrate_vram_to_ram(prange, mm, start, last,
3042						KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
3043				else
3044					r = 0;
3045			}
3046		} else {
3047			r = svm_migrate_vram_to_ram(prange, mm, start, last,
3048					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
3049		}
3050		if (r) {
3051			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
3052				 r, svms, start, last);
3053			goto out_unlock_range;
3054		}
3055	}
3056
3057	r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
3058				       false, false);
3059	if (r)
3060		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
3061			 r, svms, start, last);
3062
3063	kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
3064				     migration);
3065
3066out_unlock_range:
3067	mutex_unlock(&prange->migrate_mutex);
3068out_unlock_svms:
3069	mutex_unlock(&svms->lock);
3070	mmap_read_unlock(mm);
3071
3072	svm_range_count_fault(node, p, gpuidx);
3073
3074	mmput(mm);
3075out:
3076	kfd_unref_process(p);
3077
3078	if (r == -EAGAIN) {
3079		pr_debug("recover vm fault later\n");
3080		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3081		r = 0;
3082	}
3083	return r;
3084}
3085
3086int
3087svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
3088{
3089	struct svm_range *prange, *pchild;
3090	uint64_t reserved_size = 0;
3091	uint64_t size;
3092	int r = 0;
3093
3094	pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
3095
3096	mutex_lock(&p->svms.lock);
3097
3098	list_for_each_entry(prange, &p->svms.list, list) {
3099		svm_range_lock(prange);
3100		list_for_each_entry(pchild, &prange->child_list, child_list) {
3101			size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
3102			if (xnack_enabled) {
3103				amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3104					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3105			} else {
3106				r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3107					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3108				if (r)
3109					goto out_unlock;
3110				reserved_size += size;
3111			}
3112		}
3113
3114		size = (prange->last - prange->start + 1) << PAGE_SHIFT;
3115		if (xnack_enabled) {
3116			amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3117					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3118		} else {
3119			r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3120					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3121			if (r)
3122				goto out_unlock;
3123			reserved_size += size;
3124		}
3125out_unlock:
3126		svm_range_unlock(prange);
3127		if (r)
3128			break;
3129	}
3130
3131	if (r)
3132		amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3133					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3134	else
3135		/* Change xnack mode must be inside svms lock, to avoid race with
3136		 * svm_range_deferred_list_work unreserve memory in parallel.
3137		 */
3138		p->xnack_enabled = xnack_enabled;
3139
3140	mutex_unlock(&p->svms.lock);
3141	return r;
3142}
3143
3144void svm_range_list_fini(struct kfd_process *p)
3145{
3146	struct svm_range *prange;
3147	struct svm_range *next;
3148
3149	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3150
3151	cancel_delayed_work_sync(&p->svms.restore_work);
3152
3153	/* Ensure list work is finished before process is destroyed */
3154	flush_work(&p->svms.deferred_list_work);
3155
3156	/*
3157	 * Ensure no retry fault comes in afterwards, as page fault handler will
3158	 * not find kfd process and take mm lock to recover fault.
3159	 */
3160	atomic_inc(&p->svms.drain_pagefaults);
3161	svm_range_drain_retry_fault(&p->svms);
3162
3163	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3164		svm_range_unlink(prange);
3165		svm_range_remove_notifier(prange);
3166		svm_range_free(prange, true);
3167	}
3168
3169	mutex_destroy(&p->svms.lock);
3170
3171	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3172}
3173
3174int svm_range_list_init(struct kfd_process *p)
3175{
3176	struct svm_range_list *svms = &p->svms;
3177	int i;
3178
3179	svms->objects = RB_ROOT_CACHED;
3180	mutex_init(&svms->lock);
3181	INIT_LIST_HEAD(&svms->list);
3182	atomic_set(&svms->evicted_ranges, 0);
3183	atomic_set(&svms->drain_pagefaults, 0);
3184	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3185	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3186	INIT_LIST_HEAD(&svms->deferred_range_list);
3187	INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3188	spin_lock_init(&svms->deferred_list_lock);
3189
3190	for (i = 0; i < p->n_pdds; i++)
3191		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev))
3192			bitmap_set(svms->bitmap_supported, i, 1);
3193
3194	return 0;
3195}
3196
3197/**
3198 * svm_range_check_vm - check if virtual address range mapped already
3199 * @p: current kfd_process
3200 * @start: range start address, in pages
3201 * @last: range last address, in pages
3202 * @bo_s: mapping start address in pages if address range already mapped
3203 * @bo_l: mapping last address in pages if address range already mapped
3204 *
3205 * The purpose is to avoid virtual address ranges already allocated by
3206 * kfd_ioctl_alloc_memory_of_gpu ioctl.
3207 * It looks for each pdd in the kfd_process.
3208 *
3209 * Context: Process context
3210 *
3211 * Return 0 - OK, if the range is not mapped.
3212 * Otherwise error code:
3213 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3214 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3215 * a signal. Release all buffer reservations and return to user-space.
3216 */
3217static int
3218svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3219		   uint64_t *bo_s, uint64_t *bo_l)
3220{
3221	struct amdgpu_bo_va_mapping *mapping;
3222	struct interval_tree_node *node;
3223	uint32_t i;
3224	int r;
3225
3226	for (i = 0; i < p->n_pdds; i++) {
3227		struct amdgpu_vm *vm;
3228
3229		if (!p->pdds[i]->drm_priv)
3230			continue;
3231
3232		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3233		r = amdgpu_bo_reserve(vm->root.bo, false);
3234		if (r)
3235			return r;
3236
3237		node = interval_tree_iter_first(&vm->va, start, last);
3238		if (node) {
3239			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3240				 start, last);
3241			mapping = container_of((struct rb_node *)node,
3242					       struct amdgpu_bo_va_mapping, rb);
3243			if (bo_s && bo_l) {
3244				*bo_s = mapping->start;
3245				*bo_l = mapping->last;
3246			}
3247			amdgpu_bo_unreserve(vm->root.bo);
3248			return -EADDRINUSE;
3249		}
3250		amdgpu_bo_unreserve(vm->root.bo);
3251	}
3252
3253	return 0;
3254}
3255
3256/**
3257 * svm_range_is_valid - check if virtual address range is valid
3258 * @p: current kfd_process
3259 * @start: range start address, in pages
3260 * @size: range size, in pages
3261 *
3262 * Valid virtual address range means it belongs to one or more VMAs
3263 *
3264 * Context: Process context
3265 *
3266 * Return:
3267 *  0 - OK, otherwise error code
3268 */
3269static int
3270svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3271{
3272	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3273	struct vm_area_struct *vma;
3274	unsigned long end;
3275	unsigned long start_unchg = start;
3276
3277	start <<= PAGE_SHIFT;
3278	end = start + (size << PAGE_SHIFT);
3279	do {
3280		vma = vma_lookup(p->mm, start);
3281		if (!vma || (vma->vm_flags & device_vma))
3282			return -EFAULT;
3283		start = min(end, vma->vm_end);
3284	} while (start < end);
3285
3286	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3287				  NULL);
3288}
3289
3290/**
3291 * svm_range_best_prefetch_location - decide the best prefetch location
3292 * @prange: svm range structure
3293 *
3294 * For xnack off:
3295 * If range map to single GPU, the best prefetch location is prefetch_loc, which
3296 * can be CPU or GPU.
3297 *
3298 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3299 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3300 * the best prefetch location is always CPU, because GPU can not have coherent
3301 * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3302 *
3303 * For xnack on:
3304 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3305 * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3306 *
3307 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3308 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3309 * prefetch location is always CPU.
3310 *
3311 * Context: Process context
3312 *
3313 * Return:
3314 * 0 for CPU or GPU id
3315 */
3316static uint32_t
3317svm_range_best_prefetch_location(struct svm_range *prange)
3318{
3319	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3320	uint32_t best_loc = prange->prefetch_loc;
3321	struct kfd_process_device *pdd;
3322	struct kfd_node *bo_node;
3323	struct kfd_process *p;
3324	uint32_t gpuidx;
3325
3326	p = container_of(prange->svms, struct kfd_process, svms);
3327
3328	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3329		goto out;
3330
3331	bo_node = svm_range_get_node_by_id(prange, best_loc);
3332	if (!bo_node) {
3333		WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc);
3334		best_loc = 0;
3335		goto out;
3336	}
3337
3338	if (bo_node->adev->gmc.is_app_apu) {
3339		best_loc = 0;
3340		goto out;
3341	}
3342
3343	if (p->xnack_enabled)
3344		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3345	else
3346		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3347			  MAX_GPU_INSTANCE);
3348
3349	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3350		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3351		if (!pdd) {
3352			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3353			continue;
3354		}
3355
3356		if (pdd->dev->adev == bo_node->adev)
3357			continue;
3358
3359		if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) {
3360			best_loc = 0;
3361			break;
3362		}
3363	}
3364
3365out:
3366	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3367		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3368		 best_loc);
3369
3370	return best_loc;
3371}
3372
3373/* svm_range_trigger_migration - start page migration if prefetch loc changed
3374 * @mm: current process mm_struct
3375 * @prange: svm range structure
3376 * @migrated: output, true if migration is triggered
3377 *
3378 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3379 * from ram to vram.
3380 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3381 * from vram to ram.
3382 *
3383 * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3384 * and restore work:
3385 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3386 *    stops all queues, schedule restore work
3387 * 2. svm_range_restore_work wait for migration is done by
3388 *    a. svm_range_validate_vram takes prange->migrate_mutex
3389 *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3390 * 3. restore work update mappings of GPU, resume all queues.
3391 *
3392 * Context: Process context
3393 *
3394 * Return:
3395 * 0 - OK, otherwise - error code of migration
3396 */
3397static int
3398svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3399			    bool *migrated)
3400{
3401	uint32_t best_loc;
3402	int r = 0;
3403
3404	*migrated = false;
3405	best_loc = svm_range_best_prefetch_location(prange);
3406
3407	/* when best_loc is a gpu node and same as prange->actual_loc
3408	 * we still need do migration as prange->actual_loc !=0 does
3409	 * not mean all pages in prange are vram. hmm migrate will pick
3410	 * up right pages during migration.
3411	 */
3412	if ((best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) ||
3413	    (best_loc == 0 && prange->actual_loc == 0))
3414		return 0;
3415
3416	if (!best_loc) {
3417		r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last,
3418					KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3419		*migrated = !r;
3420		return r;
3421	}
3422
3423	r = svm_migrate_to_vram(prange, best_loc, prange->start, prange->last,
3424				mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3425	*migrated = !r;
3426
3427	return r;
3428}
3429
3430int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3431{
3432	/* Dereferencing fence->svm_bo is safe here because the fence hasn't
3433	 * signaled yet and we're under the protection of the fence->lock.
3434	 * After the fence is signaled in svm_range_bo_release, we cannot get
3435	 * here any more.
3436	 *
3437	 * Reference is dropped in svm_range_evict_svm_bo_worker.
3438	 */
3439	if (svm_bo_ref_unless_zero(fence->svm_bo)) {
3440		WRITE_ONCE(fence->svm_bo->evicting, 1);
3441		schedule_work(&fence->svm_bo->eviction_work);
3442	}
3443
3444	return 0;
3445}
3446
3447static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3448{
3449	struct svm_range_bo *svm_bo;
3450	struct mm_struct *mm;
3451	int r = 0;
3452
3453	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3454
3455	if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3456		mm = svm_bo->eviction_fence->mm;
3457	} else {
3458		svm_range_bo_unref(svm_bo);
3459		return;
3460	}
3461
3462	mmap_read_lock(mm);
3463	spin_lock(&svm_bo->list_lock);
3464	while (!list_empty(&svm_bo->range_list) && !r) {
3465		struct svm_range *prange =
3466				list_first_entry(&svm_bo->range_list,
3467						struct svm_range, svm_bo_list);
3468		int retries = 3;
3469
3470		list_del_init(&prange->svm_bo_list);
3471		spin_unlock(&svm_bo->list_lock);
3472
3473		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3474			 prange->start, prange->last);
3475
3476		mutex_lock(&prange->migrate_mutex);
3477		do {
3478			/* migrate all vram pages in this prange to sys ram
3479			 * after that prange->actual_loc should be zero
3480			 */
3481			r = svm_migrate_vram_to_ram(prange, mm,
3482					prange->start, prange->last,
3483					KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3484		} while (!r && prange->actual_loc && --retries);
3485
3486		if (!r && prange->actual_loc)
3487			pr_info_once("Migration failed during eviction");
3488
3489		if (!prange->actual_loc) {
3490			mutex_lock(&prange->lock);
3491			prange->svm_bo = NULL;
3492			mutex_unlock(&prange->lock);
3493		}
3494		mutex_unlock(&prange->migrate_mutex);
3495
3496		spin_lock(&svm_bo->list_lock);
3497	}
3498	spin_unlock(&svm_bo->list_lock);
3499	mmap_read_unlock(mm);
3500	mmput(mm);
3501
3502	dma_fence_signal(&svm_bo->eviction_fence->base);
3503
3504	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3505	 * has been called in svm_migrate_vram_to_ram
3506	 */
3507	WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3508	svm_range_bo_unref(svm_bo);
3509}
3510
3511static int
3512svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3513		   uint64_t start, uint64_t size, uint32_t nattr,
3514		   struct kfd_ioctl_svm_attribute *attrs)
3515{
3516	struct amdkfd_process_info *process_info = p->kgd_process_info;
3517	struct list_head update_list;
3518	struct list_head insert_list;
3519	struct list_head remove_list;
3520	struct list_head remap_list;
3521	struct svm_range_list *svms;
3522	struct svm_range *prange;
3523	struct svm_range *next;
3524	bool update_mapping = false;
3525	bool flush_tlb;
3526	int r, ret = 0;
3527
3528	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3529		 p->pasid, &p->svms, start, start + size - 1, size);
3530
3531	r = svm_range_check_attr(p, nattr, attrs);
3532	if (r)
3533		return r;
3534
3535	svms = &p->svms;
3536
3537	mutex_lock(&process_info->lock);
3538
3539	svm_range_list_lock_and_flush_work(svms, mm);
3540
3541	r = svm_range_is_valid(p, start, size);
3542	if (r) {
3543		pr_debug("invalid range r=%d\n", r);
3544		mmap_write_unlock(mm);
3545		goto out;
3546	}
3547
3548	mutex_lock(&svms->lock);
3549
3550	/* Add new range and split existing ranges as needed */
3551	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3552			  &insert_list, &remove_list, &remap_list);
3553	if (r) {
3554		mutex_unlock(&svms->lock);
3555		mmap_write_unlock(mm);
3556		goto out;
3557	}
3558	/* Apply changes as a transaction */
3559	list_for_each_entry_safe(prange, next, &insert_list, list) {
3560		svm_range_add_to_svms(prange);
3561		svm_range_add_notifier_locked(mm, prange);
3562	}
3563	list_for_each_entry(prange, &update_list, update_list) {
3564		svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3565		/* TODO: unmap ranges from GPU that lost access */
3566	}
3567	list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3568		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3569			 prange->svms, prange, prange->start,
3570			 prange->last);
3571		svm_range_unlink(prange);
3572		svm_range_remove_notifier(prange);
3573		svm_range_free(prange, false);
3574	}
3575
3576	mmap_write_downgrade(mm);
3577	/* Trigger migrations and revalidate and map to GPUs as needed. If
3578	 * this fails we may be left with partially completed actions. There
3579	 * is no clean way of rolling back to the previous state in such a
3580	 * case because the rollback wouldn't be guaranteed to work either.
3581	 */
3582	list_for_each_entry(prange, &update_list, update_list) {
3583		bool migrated;
3584
3585		mutex_lock(&prange->migrate_mutex);
3586
3587		r = svm_range_trigger_migration(mm, prange, &migrated);
3588		if (r)
3589			goto out_unlock_range;
3590
3591		if (migrated && (!p->xnack_enabled ||
3592		    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3593		    prange->mapped_to_gpu) {
3594			pr_debug("restore_work will update mappings of GPUs\n");
3595			mutex_unlock(&prange->migrate_mutex);
3596			continue;
3597		}
3598
3599		if (!migrated && !update_mapping) {
3600			mutex_unlock(&prange->migrate_mutex);
3601			continue;
3602		}
3603
3604		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3605
3606		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
3607					       MAX_GPU_INSTANCE, true, true, flush_tlb);
3608		if (r)
3609			pr_debug("failed %d to map svm range\n", r);
3610
3611out_unlock_range:
3612		mutex_unlock(&prange->migrate_mutex);
3613		if (r)
3614			ret = r;
3615	}
3616
3617	list_for_each_entry(prange, &remap_list, update_list) {
3618		pr_debug("Remapping prange 0x%p [0x%lx 0x%lx]\n",
3619			 prange, prange->start, prange->last);
3620		mutex_lock(&prange->migrate_mutex);
3621		r = svm_range_validate_and_map(mm,  prange->start, prange->last, prange,
3622					       MAX_GPU_INSTANCE, true, true, prange->mapped_to_gpu);
3623		if (r)
3624			pr_debug("failed %d on remap svm range\n", r);
3625		mutex_unlock(&prange->migrate_mutex);
3626		if (r)
3627			ret = r;
3628	}
3629
3630	dynamic_svm_range_dump(svms);
3631
3632	mutex_unlock(&svms->lock);
3633	mmap_read_unlock(mm);
3634out:
3635	mutex_unlock(&process_info->lock);
3636
3637	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3638		 &p->svms, start, start + size - 1, r);
3639
3640	return ret ? ret : r;
3641}
3642
3643static int
3644svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3645		   uint64_t start, uint64_t size, uint32_t nattr,
3646		   struct kfd_ioctl_svm_attribute *attrs)
3647{
3648	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3649	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3650	bool get_preferred_loc = false;
3651	bool get_prefetch_loc = false;
3652	bool get_granularity = false;
3653	bool get_accessible = false;
3654	bool get_flags = false;
3655	uint64_t last = start + size - 1UL;
3656	uint8_t granularity = 0xff;
3657	struct interval_tree_node *node;
3658	struct svm_range_list *svms;
3659	struct svm_range *prange;
3660	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3661	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3662	uint32_t flags_and = 0xffffffff;
3663	uint32_t flags_or = 0;
3664	int gpuidx;
3665	uint32_t i;
3666	int r = 0;
3667
3668	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3669		 start + size - 1, nattr);
3670
3671	/* Flush pending deferred work to avoid racing with deferred actions from
3672	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3673	 * can still race with get_attr because we don't hold the mmap lock. But that
3674	 * would be a race condition in the application anyway, and undefined
3675	 * behaviour is acceptable in that case.
3676	 */
3677	flush_work(&p->svms.deferred_list_work);
3678
3679	mmap_read_lock(mm);
3680	r = svm_range_is_valid(p, start, size);
3681	mmap_read_unlock(mm);
3682	if (r) {
3683		pr_debug("invalid range r=%d\n", r);
3684		return r;
3685	}
3686
3687	for (i = 0; i < nattr; i++) {
3688		switch (attrs[i].type) {
3689		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3690			get_preferred_loc = true;
3691			break;
3692		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3693			get_prefetch_loc = true;
3694			break;
3695		case KFD_IOCTL_SVM_ATTR_ACCESS:
3696			get_accessible = true;
3697			break;
3698		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3699		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3700			get_flags = true;
3701			break;
3702		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3703			get_granularity = true;
3704			break;
3705		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3706		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3707			fallthrough;
3708		default:
3709			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3710			return -EINVAL;
3711		}
3712	}
3713
3714	svms = &p->svms;
3715
3716	mutex_lock(&svms->lock);
3717
3718	node = interval_tree_iter_first(&svms->objects, start, last);
3719	if (!node) {
3720		pr_debug("range attrs not found return default values\n");
3721		svm_range_set_default_attributes(&location, &prefetch_loc,
3722						 &granularity, &flags_and);
3723		flags_or = flags_and;
3724		if (p->xnack_enabled)
3725			bitmap_copy(bitmap_access, svms->bitmap_supported,
3726				    MAX_GPU_INSTANCE);
3727		else
3728			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3729		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3730		goto fill_values;
3731	}
3732	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3733	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3734
3735	while (node) {
3736		struct interval_tree_node *next;
3737
3738		prange = container_of(node, struct svm_range, it_node);
3739		next = interval_tree_iter_next(node, start, last);
3740
3741		if (get_preferred_loc) {
3742			if (prange->preferred_loc ==
3743					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3744			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3745			     location != prange->preferred_loc)) {
3746				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3747				get_preferred_loc = false;
3748			} else {
3749				location = prange->preferred_loc;
3750			}
3751		}
3752		if (get_prefetch_loc) {
3753			if (prange->prefetch_loc ==
3754					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3755			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3756			     prefetch_loc != prange->prefetch_loc)) {
3757				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3758				get_prefetch_loc = false;
3759			} else {
3760				prefetch_loc = prange->prefetch_loc;
3761			}
3762		}
3763		if (get_accessible) {
3764			bitmap_and(bitmap_access, bitmap_access,
3765				   prange->bitmap_access, MAX_GPU_INSTANCE);
3766			bitmap_and(bitmap_aip, bitmap_aip,
3767				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3768		}
3769		if (get_flags) {
3770			flags_and &= prange->flags;
3771			flags_or |= prange->flags;
3772		}
3773
3774		if (get_granularity && prange->granularity < granularity)
3775			granularity = prange->granularity;
3776
3777		node = next;
3778	}
3779fill_values:
3780	mutex_unlock(&svms->lock);
3781
3782	for (i = 0; i < nattr; i++) {
3783		switch (attrs[i].type) {
3784		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3785			attrs[i].value = location;
3786			break;
3787		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3788			attrs[i].value = prefetch_loc;
3789			break;
3790		case KFD_IOCTL_SVM_ATTR_ACCESS:
3791			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3792							       attrs[i].value);
3793			if (gpuidx < 0) {
3794				pr_debug("invalid gpuid %x\n", attrs[i].value);
3795				return -EINVAL;
3796			}
3797			if (test_bit(gpuidx, bitmap_access))
3798				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3799			else if (test_bit(gpuidx, bitmap_aip))
3800				attrs[i].type =
3801					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3802			else
3803				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3804			break;
3805		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3806			attrs[i].value = flags_and;
3807			break;
3808		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3809			attrs[i].value = ~flags_or;
3810			break;
3811		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3812			attrs[i].value = (uint32_t)granularity;
3813			break;
3814		}
3815	}
3816
3817	return 0;
3818}
3819
3820int kfd_criu_resume_svm(struct kfd_process *p)
3821{
3822	struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3823	int nattr_common = 4, nattr_accessibility = 1;
3824	struct criu_svm_metadata *criu_svm_md = NULL;
3825	struct svm_range_list *svms = &p->svms;
3826	struct criu_svm_metadata *next = NULL;
3827	uint32_t set_flags = 0xffffffff;
3828	int i, j, num_attrs, ret = 0;
3829	uint64_t set_attr_size;
3830	struct mm_struct *mm;
3831
3832	if (list_empty(&svms->criu_svm_metadata_list)) {
3833		pr_debug("No SVM data from CRIU restore stage 2\n");
3834		return ret;
3835	}
3836
3837	mm = get_task_mm(p->lead_thread);
3838	if (!mm) {
3839		pr_err("failed to get mm for the target process\n");
3840		return -ESRCH;
3841	}
3842
3843	num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3844
3845	i = j = 0;
3846	list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3847		pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3848			 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3849
3850		for (j = 0; j < num_attrs; j++) {
3851			pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3852				 i, j, criu_svm_md->data.attrs[j].type,
3853				 i, j, criu_svm_md->data.attrs[j].value);
3854			switch (criu_svm_md->data.attrs[j].type) {
3855			/* During Checkpoint operation, the query for
3856			 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3857			 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3858			 * not used by the range which was checkpointed. Care
3859			 * must be taken to not restore with an invalid value
3860			 * otherwise the gpuidx value will be invalid and
3861			 * set_attr would eventually fail so just replace those
3862			 * with another dummy attribute such as
3863			 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3864			 */
3865			case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3866				if (criu_svm_md->data.attrs[j].value ==
3867				    KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3868					criu_svm_md->data.attrs[j].type =
3869						KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3870					criu_svm_md->data.attrs[j].value = 0;
3871				}
3872				break;
3873			case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3874				set_flags = criu_svm_md->data.attrs[j].value;
3875				break;
3876			default:
3877				break;
3878			}
3879		}
3880
3881		/* CLR_FLAGS is not available via get_attr during checkpoint but
3882		 * it needs to be inserted before restoring the ranges so
3883		 * allocate extra space for it before calling set_attr
3884		 */
3885		set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3886						(num_attrs + 1);
3887		set_attr_new = krealloc(set_attr, set_attr_size,
3888					    GFP_KERNEL);
3889		if (!set_attr_new) {
3890			ret = -ENOMEM;
3891			goto exit;
3892		}
3893		set_attr = set_attr_new;
3894
3895		memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3896					sizeof(struct kfd_ioctl_svm_attribute));
3897		set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3898		set_attr[num_attrs].value = ~set_flags;
3899
3900		ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3901					 criu_svm_md->data.size, num_attrs + 1,
3902					 set_attr);
3903		if (ret) {
3904			pr_err("CRIU: failed to set range attributes\n");
3905			goto exit;
3906		}
3907
3908		i++;
3909	}
3910exit:
3911	kfree(set_attr);
3912	list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3913		pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3914						criu_svm_md->data.start_addr);
3915		kfree(criu_svm_md);
3916	}
3917
3918	mmput(mm);
3919	return ret;
3920
3921}
3922
3923int kfd_criu_restore_svm(struct kfd_process *p,
3924			 uint8_t __user *user_priv_ptr,
3925			 uint64_t *priv_data_offset,
3926			 uint64_t max_priv_data_size)
3927{
3928	uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3929	int nattr_common = 4, nattr_accessibility = 1;
3930	struct criu_svm_metadata *criu_svm_md = NULL;
3931	struct svm_range_list *svms = &p->svms;
3932	uint32_t num_devices;
3933	int ret = 0;
3934
3935	num_devices = p->n_pdds;
3936	/* Handle one SVM range object at a time, also the number of gpus are
3937	 * assumed to be same on the restore node, checking must be done while
3938	 * evaluating the topology earlier
3939	 */
3940
3941	svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3942		(nattr_common + nattr_accessibility * num_devices);
3943	svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3944
3945	svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3946								svm_attrs_size;
3947
3948	criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3949	if (!criu_svm_md) {
3950		pr_err("failed to allocate memory to store svm metadata\n");
3951		return -ENOMEM;
3952	}
3953	if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3954		ret = -EINVAL;
3955		goto exit;
3956	}
3957
3958	ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3959			     svm_priv_data_size);
3960	if (ret) {
3961		ret = -EFAULT;
3962		goto exit;
3963	}
3964	*priv_data_offset += svm_priv_data_size;
3965
3966	list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3967
3968	return 0;
3969
3970
3971exit:
3972	kfree(criu_svm_md);
3973	return ret;
3974}
3975
3976int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3977		       uint64_t *svm_priv_data_size)
3978{
3979	uint64_t total_size, accessibility_size, common_attr_size;
3980	int nattr_common = 4, nattr_accessibility = 1;
3981	int num_devices = p->n_pdds;
3982	struct svm_range_list *svms;
3983	struct svm_range *prange;
3984	uint32_t count = 0;
3985
3986	*svm_priv_data_size = 0;
3987
3988	svms = &p->svms;
3989	if (!svms)
3990		return -EINVAL;
3991
3992	mutex_lock(&svms->lock);
3993	list_for_each_entry(prange, &svms->list, list) {
3994		pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3995			 prange, prange->start, prange->npages,
3996			 prange->start + prange->npages - 1);
3997		count++;
3998	}
3999	mutex_unlock(&svms->lock);
4000
4001	*num_svm_ranges = count;
4002	/* Only the accessbility attributes need to be queried for all the gpus
4003	 * individually, remaining ones are spanned across the entire process
4004	 * regardless of the various gpu nodes. Of the remaining attributes,
4005	 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
4006	 *
4007	 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
4008	 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
4009	 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
4010	 * KFD_IOCTL_SVM_ATTR_GRANULARITY
4011	 *
4012	 * ** ACCESSBILITY ATTRIBUTES **
4013	 * (Considered as one, type is altered during query, value is gpuid)
4014	 * KFD_IOCTL_SVM_ATTR_ACCESS
4015	 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
4016	 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
4017	 */
4018	if (*num_svm_ranges > 0) {
4019		common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4020			nattr_common;
4021		accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
4022			nattr_accessibility * num_devices;
4023
4024		total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
4025			common_attr_size + accessibility_size;
4026
4027		*svm_priv_data_size = *num_svm_ranges * total_size;
4028	}
4029
4030	pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
4031		 *svm_priv_data_size);
4032	return 0;
4033}
4034
4035int kfd_criu_checkpoint_svm(struct kfd_process *p,
4036			    uint8_t __user *user_priv_data,
4037			    uint64_t *priv_data_offset)
4038{
4039	struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
4040	struct kfd_ioctl_svm_attribute *query_attr = NULL;
4041	uint64_t svm_priv_data_size, query_attr_size = 0;
4042	int index, nattr_common = 4, ret = 0;
4043	struct svm_range_list *svms;
4044	int num_devices = p->n_pdds;
4045	struct svm_range *prange;
4046	struct mm_struct *mm;
4047
4048	svms = &p->svms;
4049	if (!svms)
4050		return -EINVAL;
4051
4052	mm = get_task_mm(p->lead_thread);
4053	if (!mm) {
4054		pr_err("failed to get mm for the target process\n");
4055		return -ESRCH;
4056	}
4057
4058	query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4059				(nattr_common + num_devices);
4060
4061	query_attr = kzalloc(query_attr_size, GFP_KERNEL);
4062	if (!query_attr) {
4063		ret = -ENOMEM;
4064		goto exit;
4065	}
4066
4067	query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
4068	query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
4069	query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
4070	query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
4071
4072	for (index = 0; index < num_devices; index++) {
4073		struct kfd_process_device *pdd = p->pdds[index];
4074
4075		query_attr[index + nattr_common].type =
4076			KFD_IOCTL_SVM_ATTR_ACCESS;
4077		query_attr[index + nattr_common].value = pdd->user_gpu_id;
4078	}
4079
4080	svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
4081
4082	svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
4083	if (!svm_priv) {
4084		ret = -ENOMEM;
4085		goto exit_query;
4086	}
4087
4088	index = 0;
4089	list_for_each_entry(prange, &svms->list, list) {
4090
4091		svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
4092		svm_priv->start_addr = prange->start;
4093		svm_priv->size = prange->npages;
4094		memcpy(&svm_priv->attrs, query_attr, query_attr_size);
4095		pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
4096			 prange, prange->start, prange->npages,
4097			 prange->start + prange->npages - 1,
4098			 prange->npages * PAGE_SIZE);
4099
4100		ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
4101					 svm_priv->size,
4102					 (nattr_common + num_devices),
4103					 svm_priv->attrs);
4104		if (ret) {
4105			pr_err("CRIU: failed to obtain range attributes\n");
4106			goto exit_priv;
4107		}
4108
4109		if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
4110				 svm_priv_data_size)) {
4111			pr_err("Failed to copy svm priv to user\n");
4112			ret = -EFAULT;
4113			goto exit_priv;
4114		}
4115
4116		*priv_data_offset += svm_priv_data_size;
4117
4118	}
4119
4120
4121exit_priv:
4122	kfree(svm_priv);
4123exit_query:
4124	kfree(query_attr);
4125exit:
4126	mmput(mm);
4127	return ret;
4128}
4129
4130int
4131svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
4132	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
4133{
4134	struct mm_struct *mm = current->mm;
4135	int r;
4136
4137	start >>= PAGE_SHIFT;
4138	size >>= PAGE_SHIFT;
4139
4140	switch (op) {
4141	case KFD_IOCTL_SVM_OP_SET_ATTR:
4142		r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4143		break;
4144	case KFD_IOCTL_SVM_OP_GET_ATTR:
4145		r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);
4146		break;
4147	default:
4148		r = EINVAL;
4149		break;
4150	}
4151
4152	return r;
4153}