xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 61ae993c)
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  */
64 static uint64_t max_svm_range_pages;
65 
66 struct criu_svm_metadata {
67 	struct list_head list;
68 	struct kfd_criu_svm_range_priv_data data;
69 };
70 
71 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
72 static bool
73 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
74 				    const struct mmu_notifier_range *range,
75 				    unsigned long cur_seq);
76 static int
77 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
78 		   uint64_t *bo_s, uint64_t *bo_l);
79 static 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  */
92 static 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 
108 static void
109 svm_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  */
128 static 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 
139 static 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 
151 static bool
152 svm_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 
158 static int
159 svm_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 	return 0;
202 }
203 
204 static int
205 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
206 		  unsigned long offset, unsigned long npages,
207 		  unsigned long *hmm_pfns)
208 {
209 	struct kfd_process *p;
210 	uint32_t gpuidx;
211 	int r;
212 
213 	p = container_of(prange->svms, struct kfd_process, svms);
214 
215 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
216 		struct kfd_process_device *pdd;
217 
218 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
219 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
220 		if (!pdd) {
221 			pr_debug("failed to find device idx %d\n", gpuidx);
222 			return -EINVAL;
223 		}
224 
225 		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
226 					  hmm_pfns, gpuidx);
227 		if (r)
228 			break;
229 	}
230 
231 	return r;
232 }
233 
234 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
235 			 unsigned long offset, unsigned long npages)
236 {
237 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
238 	int i;
239 
240 	if (!dma_addr)
241 		return;
242 
243 	for (i = offset; i < offset + npages; i++) {
244 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
245 			continue;
246 		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
247 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
248 		dma_addr[i] = 0;
249 	}
250 }
251 
252 void svm_range_free_dma_mappings(struct svm_range *prange, bool unmap_dma)
253 {
254 	struct kfd_process_device *pdd;
255 	dma_addr_t *dma_addr;
256 	struct device *dev;
257 	struct kfd_process *p;
258 	uint32_t gpuidx;
259 
260 	p = container_of(prange->svms, struct kfd_process, svms);
261 
262 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
263 		dma_addr = prange->dma_addr[gpuidx];
264 		if (!dma_addr)
265 			continue;
266 
267 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
268 		if (!pdd) {
269 			pr_debug("failed to find device idx %d\n", gpuidx);
270 			continue;
271 		}
272 		dev = &pdd->dev->adev->pdev->dev;
273 		if (unmap_dma)
274 			svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
275 		kvfree(dma_addr);
276 		prange->dma_addr[gpuidx] = NULL;
277 	}
278 }
279 
280 static void svm_range_free(struct svm_range *prange, bool do_unmap)
281 {
282 	uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
283 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
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 	svm_range_free_dma_mappings(prange, do_unmap);
290 
291 	if (do_unmap && !p->xnack_enabled) {
292 		pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size);
293 		amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
294 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
295 	}
296 	mutex_destroy(&prange->lock);
297 	mutex_destroy(&prange->migrate_mutex);
298 	kfree(prange);
299 }
300 
301 static void
302 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
303 				 uint8_t *granularity, uint32_t *flags)
304 {
305 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
306 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
307 	*granularity = 9;
308 	*flags =
309 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
310 }
311 
312 static struct
313 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
314 			 uint64_t last, bool update_mem_usage)
315 {
316 	uint64_t size = last - start + 1;
317 	struct svm_range *prange;
318 	struct kfd_process *p;
319 
320 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
321 	if (!prange)
322 		return NULL;
323 
324 	p = container_of(svms, struct kfd_process, svms);
325 	if (!p->xnack_enabled && update_mem_usage &&
326 	    amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
327 				    KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) {
328 		pr_info("SVM mapping failed, exceeds resident system memory limit\n");
329 		kfree(prange);
330 		return NULL;
331 	}
332 	prange->npages = size;
333 	prange->svms = svms;
334 	prange->start = start;
335 	prange->last = last;
336 	INIT_LIST_HEAD(&prange->list);
337 	INIT_LIST_HEAD(&prange->update_list);
338 	INIT_LIST_HEAD(&prange->svm_bo_list);
339 	INIT_LIST_HEAD(&prange->deferred_list);
340 	INIT_LIST_HEAD(&prange->child_list);
341 	atomic_set(&prange->invalid, 0);
342 	prange->validate_timestamp = 0;
343 	mutex_init(&prange->migrate_mutex);
344 	mutex_init(&prange->lock);
345 
346 	if (p->xnack_enabled)
347 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
348 			    MAX_GPU_INSTANCE);
349 
350 	svm_range_set_default_attributes(&prange->preferred_loc,
351 					 &prange->prefetch_loc,
352 					 &prange->granularity, &prange->flags);
353 
354 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
355 
356 	return prange;
357 }
358 
359 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
360 {
361 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
362 		return false;
363 
364 	return true;
365 }
366 
367 static void svm_range_bo_release(struct kref *kref)
368 {
369 	struct svm_range_bo *svm_bo;
370 
371 	svm_bo = container_of(kref, struct svm_range_bo, kref);
372 	pr_debug("svm_bo 0x%p\n", svm_bo);
373 
374 	spin_lock(&svm_bo->list_lock);
375 	while (!list_empty(&svm_bo->range_list)) {
376 		struct svm_range *prange =
377 				list_first_entry(&svm_bo->range_list,
378 						struct svm_range, svm_bo_list);
379 		/* list_del_init tells a concurrent svm_range_vram_node_new when
380 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
381 		 */
382 		list_del_init(&prange->svm_bo_list);
383 		spin_unlock(&svm_bo->list_lock);
384 
385 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
386 			 prange->start, prange->last);
387 		mutex_lock(&prange->lock);
388 		prange->svm_bo = NULL;
389 		mutex_unlock(&prange->lock);
390 
391 		spin_lock(&svm_bo->list_lock);
392 	}
393 	spin_unlock(&svm_bo->list_lock);
394 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
395 		/* We're not in the eviction worker.
396 		 * Signal the fence and synchronize with any
397 		 * pending eviction work.
398 		 */
399 		dma_fence_signal(&svm_bo->eviction_fence->base);
400 		cancel_work_sync(&svm_bo->eviction_work);
401 	}
402 	dma_fence_put(&svm_bo->eviction_fence->base);
403 	amdgpu_bo_unref(&svm_bo->bo);
404 	kfree(svm_bo);
405 }
406 
407 static void svm_range_bo_wq_release(struct work_struct *work)
408 {
409 	struct svm_range_bo *svm_bo;
410 
411 	svm_bo = container_of(work, struct svm_range_bo, release_work);
412 	svm_range_bo_release(&svm_bo->kref);
413 }
414 
415 static void svm_range_bo_release_async(struct kref *kref)
416 {
417 	struct svm_range_bo *svm_bo;
418 
419 	svm_bo = container_of(kref, struct svm_range_bo, kref);
420 	pr_debug("svm_bo 0x%p\n", svm_bo);
421 	INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
422 	schedule_work(&svm_bo->release_work);
423 }
424 
425 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
426 {
427 	kref_put(&svm_bo->kref, svm_range_bo_release_async);
428 }
429 
430 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
431 {
432 	if (svm_bo)
433 		kref_put(&svm_bo->kref, svm_range_bo_release);
434 }
435 
436 static bool
437 svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange)
438 {
439 	mutex_lock(&prange->lock);
440 	if (!prange->svm_bo) {
441 		mutex_unlock(&prange->lock);
442 		return false;
443 	}
444 	if (prange->ttm_res) {
445 		/* We still have a reference, all is well */
446 		mutex_unlock(&prange->lock);
447 		return true;
448 	}
449 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
450 		/*
451 		 * Migrate from GPU to GPU, remove range from source svm_bo->node
452 		 * range list, and return false to allocate svm_bo from destination
453 		 * node.
454 		 */
455 		if (prange->svm_bo->node != node) {
456 			mutex_unlock(&prange->lock);
457 
458 			spin_lock(&prange->svm_bo->list_lock);
459 			list_del_init(&prange->svm_bo_list);
460 			spin_unlock(&prange->svm_bo->list_lock);
461 
462 			svm_range_bo_unref(prange->svm_bo);
463 			return false;
464 		}
465 		if (READ_ONCE(prange->svm_bo->evicting)) {
466 			struct dma_fence *f;
467 			struct svm_range_bo *svm_bo;
468 			/* The BO is getting evicted,
469 			 * we need to get a new one
470 			 */
471 			mutex_unlock(&prange->lock);
472 			svm_bo = prange->svm_bo;
473 			f = dma_fence_get(&svm_bo->eviction_fence->base);
474 			svm_range_bo_unref(prange->svm_bo);
475 			/* wait for the fence to avoid long spin-loop
476 			 * at list_empty_careful
477 			 */
478 			dma_fence_wait(f, false);
479 			dma_fence_put(f);
480 		} else {
481 			/* The BO was still around and we got
482 			 * a new reference to it
483 			 */
484 			mutex_unlock(&prange->lock);
485 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
486 				 prange->svms, prange->start, prange->last);
487 
488 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
489 			return true;
490 		}
491 
492 	} else {
493 		mutex_unlock(&prange->lock);
494 	}
495 
496 	/* We need a new svm_bo. Spin-loop to wait for concurrent
497 	 * svm_range_bo_release to finish removing this range from
498 	 * its range list and set prange->svm_bo to null. After this,
499 	 * it is safe to reuse the svm_bo pointer and svm_bo_list head.
500 	 */
501 	while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo)
502 		cond_resched();
503 
504 	return false;
505 }
506 
507 static struct svm_range_bo *svm_range_bo_new(void)
508 {
509 	struct svm_range_bo *svm_bo;
510 
511 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
512 	if (!svm_bo)
513 		return NULL;
514 
515 	kref_init(&svm_bo->kref);
516 	INIT_LIST_HEAD(&svm_bo->range_list);
517 	spin_lock_init(&svm_bo->list_lock);
518 
519 	return svm_bo;
520 }
521 
522 int
523 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
524 			bool clear)
525 {
526 	struct amdgpu_bo_param bp;
527 	struct svm_range_bo *svm_bo;
528 	struct amdgpu_bo_user *ubo;
529 	struct amdgpu_bo *bo;
530 	struct kfd_process *p;
531 	struct mm_struct *mm;
532 	int r;
533 
534 	p = container_of(prange->svms, struct kfd_process, svms);
535 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
536 		 prange->start, prange->last);
537 
538 	if (svm_range_validate_svm_bo(node, prange))
539 		return 0;
540 
541 	svm_bo = svm_range_bo_new();
542 	if (!svm_bo) {
543 		pr_debug("failed to alloc svm bo\n");
544 		return -ENOMEM;
545 	}
546 	mm = get_task_mm(p->lead_thread);
547 	if (!mm) {
548 		pr_debug("failed to get mm\n");
549 		kfree(svm_bo);
550 		return -ESRCH;
551 	}
552 	svm_bo->node = node;
553 	svm_bo->eviction_fence =
554 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
555 					   mm,
556 					   svm_bo);
557 	mmput(mm);
558 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
559 	svm_bo->evicting = 0;
560 	memset(&bp, 0, sizeof(bp));
561 	bp.size = prange->npages * PAGE_SIZE;
562 	bp.byte_align = PAGE_SIZE;
563 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
564 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
565 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
566 	bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
567 	bp.type = ttm_bo_type_device;
568 	bp.resv = NULL;
569 	if (node->xcp)
570 		bp.xcp_id_plus1 = node->xcp->id + 1;
571 
572 	r = amdgpu_bo_create_user(node->adev, &bp, &ubo);
573 	if (r) {
574 		pr_debug("failed %d to create bo\n", r);
575 		goto create_bo_failed;
576 	}
577 	bo = &ubo->bo;
578 
579 	pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n",
580 		 bo->tbo.resource->start << PAGE_SHIFT, bp.size,
581 		 bp.xcp_id_plus1 - 1);
582 
583 	r = amdgpu_bo_reserve(bo, true);
584 	if (r) {
585 		pr_debug("failed %d to reserve bo\n", r);
586 		goto reserve_bo_failed;
587 	}
588 
589 	if (clear) {
590 		r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
591 		if (r) {
592 			pr_debug("failed %d to sync bo\n", r);
593 			amdgpu_bo_unreserve(bo);
594 			goto reserve_bo_failed;
595 		}
596 	}
597 
598 	r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
599 	if (r) {
600 		pr_debug("failed %d to reserve bo\n", r);
601 		amdgpu_bo_unreserve(bo);
602 		goto reserve_bo_failed;
603 	}
604 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
605 
606 	amdgpu_bo_unreserve(bo);
607 
608 	svm_bo->bo = bo;
609 	prange->svm_bo = svm_bo;
610 	prange->ttm_res = bo->tbo.resource;
611 	prange->offset = 0;
612 
613 	spin_lock(&svm_bo->list_lock);
614 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
615 	spin_unlock(&svm_bo->list_lock);
616 
617 	return 0;
618 
619 reserve_bo_failed:
620 	amdgpu_bo_unref(&bo);
621 create_bo_failed:
622 	dma_fence_put(&svm_bo->eviction_fence->base);
623 	kfree(svm_bo);
624 	prange->ttm_res = NULL;
625 
626 	return r;
627 }
628 
629 void svm_range_vram_node_free(struct svm_range *prange)
630 {
631 	/* serialize prange->svm_bo unref */
632 	mutex_lock(&prange->lock);
633 	/* prange->svm_bo has not been unref */
634 	if (prange->ttm_res) {
635 		prange->ttm_res = NULL;
636 		mutex_unlock(&prange->lock);
637 		svm_range_bo_unref(prange->svm_bo);
638 	} else
639 		mutex_unlock(&prange->lock);
640 }
641 
642 struct kfd_node *
643 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id)
644 {
645 	struct kfd_process *p;
646 	struct kfd_process_device *pdd;
647 
648 	p = container_of(prange->svms, struct kfd_process, svms);
649 	pdd = kfd_process_device_data_by_id(p, gpu_id);
650 	if (!pdd) {
651 		pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id);
652 		return NULL;
653 	}
654 
655 	return pdd->dev;
656 }
657 
658 struct kfd_process_device *
659 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node)
660 {
661 	struct kfd_process *p;
662 
663 	p = container_of(prange->svms, struct kfd_process, svms);
664 
665 	return kfd_get_process_device_data(node, p);
666 }
667 
668 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
669 {
670 	struct ttm_operation_ctx ctx = { false, false };
671 
672 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
673 
674 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
675 }
676 
677 static int
678 svm_range_check_attr(struct kfd_process *p,
679 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
680 {
681 	uint32_t i;
682 
683 	for (i = 0; i < nattr; i++) {
684 		uint32_t val = attrs[i].value;
685 		int gpuidx = MAX_GPU_INSTANCE;
686 
687 		switch (attrs[i].type) {
688 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
689 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
690 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
691 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
692 			break;
693 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
694 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
695 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
696 			break;
697 		case KFD_IOCTL_SVM_ATTR_ACCESS:
698 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
699 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
700 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
701 			break;
702 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
703 			break;
704 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
705 			break;
706 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
707 			break;
708 		default:
709 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
710 			return -EINVAL;
711 		}
712 
713 		if (gpuidx < 0) {
714 			pr_debug("no GPU 0x%x found\n", val);
715 			return -EINVAL;
716 		} else if (gpuidx < MAX_GPU_INSTANCE &&
717 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
718 			pr_debug("GPU 0x%x not supported\n", val);
719 			return -EINVAL;
720 		}
721 	}
722 
723 	return 0;
724 }
725 
726 static void
727 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
728 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
729 		      bool *update_mapping)
730 {
731 	uint32_t i;
732 	int gpuidx;
733 
734 	for (i = 0; i < nattr; i++) {
735 		switch (attrs[i].type) {
736 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
737 			prange->preferred_loc = attrs[i].value;
738 			break;
739 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
740 			prange->prefetch_loc = attrs[i].value;
741 			break;
742 		case KFD_IOCTL_SVM_ATTR_ACCESS:
743 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
744 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
745 			if (!p->xnack_enabled)
746 				*update_mapping = true;
747 
748 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
749 							       attrs[i].value);
750 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
751 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
752 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
753 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
754 				bitmap_set(prange->bitmap_access, gpuidx, 1);
755 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
756 			} else {
757 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
758 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
759 			}
760 			break;
761 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
762 			*update_mapping = true;
763 			prange->flags |= attrs[i].value;
764 			break;
765 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
766 			*update_mapping = true;
767 			prange->flags &= ~attrs[i].value;
768 			break;
769 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
770 			prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F);
771 			break;
772 		default:
773 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
774 		}
775 	}
776 }
777 
778 static bool
779 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
780 			uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
781 {
782 	uint32_t i;
783 	int gpuidx;
784 
785 	for (i = 0; i < nattr; i++) {
786 		switch (attrs[i].type) {
787 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
788 			if (prange->preferred_loc != attrs[i].value)
789 				return false;
790 			break;
791 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
792 			/* Prefetch should always trigger a migration even
793 			 * if the value of the attribute didn't change.
794 			 */
795 			return false;
796 		case KFD_IOCTL_SVM_ATTR_ACCESS:
797 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
798 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
799 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
800 							       attrs[i].value);
801 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
802 				if (test_bit(gpuidx, prange->bitmap_access) ||
803 				    test_bit(gpuidx, prange->bitmap_aip))
804 					return false;
805 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
806 				if (!test_bit(gpuidx, prange->bitmap_access))
807 					return false;
808 			} else {
809 				if (!test_bit(gpuidx, prange->bitmap_aip))
810 					return false;
811 			}
812 			break;
813 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
814 			if ((prange->flags & attrs[i].value) != attrs[i].value)
815 				return false;
816 			break;
817 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
818 			if ((prange->flags & attrs[i].value) != 0)
819 				return false;
820 			break;
821 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
822 			if (prange->granularity != attrs[i].value)
823 				return false;
824 			break;
825 		default:
826 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
827 		}
828 	}
829 
830 	return true;
831 }
832 
833 /**
834  * svm_range_debug_dump - print all range information from svms
835  * @svms: svm range list header
836  *
837  * debug output svm range start, end, prefetch location from svms
838  * interval tree and link list
839  *
840  * Context: The caller must hold svms->lock
841  */
842 static void svm_range_debug_dump(struct svm_range_list *svms)
843 {
844 	struct interval_tree_node *node;
845 	struct svm_range *prange;
846 
847 	pr_debug("dump svms 0x%p list\n", svms);
848 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
849 
850 	list_for_each_entry(prange, &svms->list, list) {
851 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
852 			 prange, prange->start, prange->npages,
853 			 prange->start + prange->npages - 1,
854 			 prange->actual_loc);
855 	}
856 
857 	pr_debug("dump svms 0x%p interval tree\n", svms);
858 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
859 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
860 	while (node) {
861 		prange = container_of(node, struct svm_range, it_node);
862 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
863 			 prange, prange->start, prange->npages,
864 			 prange->start + prange->npages - 1,
865 			 prange->actual_loc);
866 		node = interval_tree_iter_next(node, 0, ~0ULL);
867 	}
868 }
869 
870 static void *
871 svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements,
872 		     uint64_t offset)
873 {
874 	unsigned char *dst;
875 
876 	dst = kvmalloc_array(num_elements, size, GFP_KERNEL);
877 	if (!dst)
878 		return NULL;
879 	memcpy(dst, (unsigned char *)psrc + offset, num_elements * size);
880 
881 	return (void *)dst;
882 }
883 
884 static int
885 svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src)
886 {
887 	int i;
888 
889 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
890 		if (!src->dma_addr[i])
891 			continue;
892 		dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i],
893 					sizeof(*src->dma_addr[i]), src->npages, 0);
894 		if (!dst->dma_addr[i])
895 			return -ENOMEM;
896 	}
897 
898 	return 0;
899 }
900 
901 static int
902 svm_range_split_array(void *ppnew, void *ppold, size_t size,
903 		      uint64_t old_start, uint64_t old_n,
904 		      uint64_t new_start, uint64_t new_n)
905 {
906 	unsigned char *new, *old, *pold;
907 	uint64_t d;
908 
909 	if (!ppold)
910 		return 0;
911 	pold = *(unsigned char **)ppold;
912 	if (!pold)
913 		return 0;
914 
915 	d = (new_start - old_start) * size;
916 	new = svm_range_copy_array(pold, size, new_n, d);
917 	if (!new)
918 		return -ENOMEM;
919 	d = (new_start == old_start) ? new_n * size : 0;
920 	old = svm_range_copy_array(pold, size, old_n, d);
921 	if (!old) {
922 		kvfree(new);
923 		return -ENOMEM;
924 	}
925 	kvfree(pold);
926 	*(void **)ppold = old;
927 	*(void **)ppnew = new;
928 
929 	return 0;
930 }
931 
932 static int
933 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
934 		      uint64_t start, uint64_t last)
935 {
936 	uint64_t npages = last - start + 1;
937 	int i, r;
938 
939 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
940 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
941 					  sizeof(*old->dma_addr[i]), old->start,
942 					  npages, new->start, new->npages);
943 		if (r)
944 			return r;
945 	}
946 
947 	return 0;
948 }
949 
950 static int
951 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
952 		      uint64_t start, uint64_t last)
953 {
954 	uint64_t npages = last - start + 1;
955 
956 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
957 		 new->svms, new, new->start, start, last);
958 
959 	if (new->start == old->start) {
960 		new->offset = old->offset;
961 		old->offset += new->npages;
962 	} else {
963 		new->offset = old->offset + npages;
964 	}
965 
966 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
967 	new->ttm_res = old->ttm_res;
968 
969 	spin_lock(&new->svm_bo->list_lock);
970 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
971 	spin_unlock(&new->svm_bo->list_lock);
972 
973 	return 0;
974 }
975 
976 /**
977  * svm_range_split_adjust - split range and adjust
978  *
979  * @new: new range
980  * @old: the old range
981  * @start: the old range adjust to start address in pages
982  * @last: the old range adjust to last address in pages
983  *
984  * Copy system memory dma_addr or vram ttm_res in old range to new
985  * range from new_start up to size new->npages, the remaining old range is from
986  * start to last
987  *
988  * Return:
989  * 0 - OK, -ENOMEM - out of memory
990  */
991 static int
992 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
993 		      uint64_t start, uint64_t last)
994 {
995 	int r;
996 
997 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
998 		 new->svms, new->start, old->start, old->last, start, last);
999 
1000 	if (new->start < old->start ||
1001 	    new->last > old->last) {
1002 		WARN_ONCE(1, "invalid new range start or last\n");
1003 		return -EINVAL;
1004 	}
1005 
1006 	r = svm_range_split_pages(new, old, start, last);
1007 	if (r)
1008 		return r;
1009 
1010 	if (old->actual_loc && old->ttm_res) {
1011 		r = svm_range_split_nodes(new, old, start, last);
1012 		if (r)
1013 			return r;
1014 	}
1015 
1016 	old->npages = last - start + 1;
1017 	old->start = start;
1018 	old->last = last;
1019 	new->flags = old->flags;
1020 	new->preferred_loc = old->preferred_loc;
1021 	new->prefetch_loc = old->prefetch_loc;
1022 	new->actual_loc = old->actual_loc;
1023 	new->granularity = old->granularity;
1024 	new->mapped_to_gpu = old->mapped_to_gpu;
1025 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1026 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1027 
1028 	return 0;
1029 }
1030 
1031 /**
1032  * svm_range_split - split a range in 2 ranges
1033  *
1034  * @prange: the svm range to split
1035  * @start: the remaining range start address in pages
1036  * @last: the remaining range last address in pages
1037  * @new: the result new range generated
1038  *
1039  * Two cases only:
1040  * case 1: if start == prange->start
1041  *         prange ==> prange[start, last]
1042  *         new range [last + 1, prange->last]
1043  *
1044  * case 2: if last == prange->last
1045  *         prange ==> prange[start, last]
1046  *         new range [prange->start, start - 1]
1047  *
1048  * Return:
1049  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1050  */
1051 static int
1052 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1053 		struct svm_range **new)
1054 {
1055 	uint64_t old_start = prange->start;
1056 	uint64_t old_last = prange->last;
1057 	struct svm_range_list *svms;
1058 	int r = 0;
1059 
1060 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1061 		 old_start, old_last, start, last);
1062 
1063 	if (old_start != start && old_last != last)
1064 		return -EINVAL;
1065 	if (start < old_start || last > old_last)
1066 		return -EINVAL;
1067 
1068 	svms = prange->svms;
1069 	if (old_start == start)
1070 		*new = svm_range_new(svms, last + 1, old_last, false);
1071 	else
1072 		*new = svm_range_new(svms, old_start, start - 1, false);
1073 	if (!*new)
1074 		return -ENOMEM;
1075 
1076 	r = svm_range_split_adjust(*new, prange, start, last);
1077 	if (r) {
1078 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1079 			 r, old_start, old_last, start, last);
1080 		svm_range_free(*new, false);
1081 		*new = NULL;
1082 	}
1083 
1084 	return r;
1085 }
1086 
1087 static int
1088 svm_range_split_tail(struct svm_range *prange,
1089 		     uint64_t new_last, struct list_head *insert_list)
1090 {
1091 	struct svm_range *tail;
1092 	int r = svm_range_split(prange, prange->start, new_last, &tail);
1093 
1094 	if (!r)
1095 		list_add(&tail->list, insert_list);
1096 	return r;
1097 }
1098 
1099 static int
1100 svm_range_split_head(struct svm_range *prange,
1101 		     uint64_t new_start, struct list_head *insert_list)
1102 {
1103 	struct svm_range *head;
1104 	int r = svm_range_split(prange, new_start, prange->last, &head);
1105 
1106 	if (!r)
1107 		list_add(&head->list, insert_list);
1108 	return r;
1109 }
1110 
1111 static void
1112 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1113 		    struct svm_range *pchild, enum svm_work_list_ops op)
1114 {
1115 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1116 		 pchild, pchild->start, pchild->last, prange, op);
1117 
1118 	pchild->work_item.mm = mm;
1119 	pchild->work_item.op = op;
1120 	list_add_tail(&pchild->child_list, &prange->child_list);
1121 }
1122 
1123 /**
1124  * svm_range_split_by_granularity - collect ranges within granularity boundary
1125  *
1126  * @p: the process with svms list
1127  * @mm: mm structure
1128  * @addr: the vm fault address in pages, to split the prange
1129  * @parent: parent range if prange is from child list
1130  * @prange: prange to split
1131  *
1132  * Trims @prange to be a single aligned block of prange->granularity if
1133  * possible. The head and tail are added to the child_list in @parent.
1134  *
1135  * Context: caller must hold mmap_read_lock and prange->lock
1136  *
1137  * Return:
1138  * 0 - OK, otherwise error code
1139  */
1140 int
1141 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1142 			       unsigned long addr, struct svm_range *parent,
1143 			       struct svm_range *prange)
1144 {
1145 	struct svm_range *head, *tail;
1146 	unsigned long start, last, size;
1147 	int r;
1148 
1149 	/* Align splited range start and size to granularity size, then a single
1150 	 * PTE will be used for whole range, this reduces the number of PTE
1151 	 * updated and the L1 TLB space used for translation.
1152 	 */
1153 	size = 1UL << prange->granularity;
1154 	start = ALIGN_DOWN(addr, size);
1155 	last = ALIGN(addr + 1, size) - 1;
1156 
1157 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1158 		 prange->svms, prange->start, prange->last, start, last, size);
1159 
1160 	if (start > prange->start) {
1161 		r = svm_range_split(prange, start, prange->last, &head);
1162 		if (r)
1163 			return r;
1164 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1165 	}
1166 
1167 	if (last < prange->last) {
1168 		r = svm_range_split(prange, prange->start, last, &tail);
1169 		if (r)
1170 			return r;
1171 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1172 	}
1173 
1174 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1175 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1176 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1177 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1178 			 prange, prange->start, prange->last,
1179 			 SVM_OP_ADD_RANGE_AND_MAP);
1180 	}
1181 	return 0;
1182 }
1183 static bool
1184 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
1185 {
1186 	return (node_a->adev == node_b->adev ||
1187 		amdgpu_xgmi_same_hive(node_a->adev, node_b->adev));
1188 }
1189 
1190 static uint64_t
1191 svm_range_get_pte_flags(struct kfd_node *node,
1192 			struct svm_range *prange, int domain)
1193 {
1194 	struct kfd_node *bo_node;
1195 	uint32_t flags = prange->flags;
1196 	uint32_t mapping_flags = 0;
1197 	uint64_t pte_flags;
1198 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1199 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1200 	bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/
1201 	unsigned int mtype_local;
1202 
1203 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1204 		bo_node = prange->svm_bo->node;
1205 
1206 	switch (node->adev->ip_versions[GC_HWIP][0]) {
1207 	case IP_VERSION(9, 4, 1):
1208 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1209 			if (bo_node == node) {
1210 				mapping_flags |= coherent ?
1211 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1212 			} else {
1213 				mapping_flags |= coherent ?
1214 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1215 				if (svm_nodes_in_same_hive(node, bo_node))
1216 					snoop = true;
1217 			}
1218 		} else {
1219 			mapping_flags |= coherent ?
1220 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1221 		}
1222 		break;
1223 	case IP_VERSION(9, 4, 2):
1224 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1225 			if (bo_node == node) {
1226 				mapping_flags |= coherent ?
1227 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1228 				if (node->adev->gmc.xgmi.connected_to_cpu)
1229 					snoop = true;
1230 			} else {
1231 				mapping_flags |= coherent ?
1232 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1233 				if (svm_nodes_in_same_hive(node, bo_node))
1234 					snoop = true;
1235 			}
1236 		} else {
1237 			mapping_flags |= coherent ?
1238 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1239 		}
1240 		break;
1241 	case IP_VERSION(9, 4, 3):
1242 		mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC :
1243 			     (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW);
1244 		snoop = true;
1245 		if (uncached) {
1246 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1247 		} else if (domain == SVM_RANGE_VRAM_DOMAIN) {
1248 			/* local HBM region close to partition */
1249 			if (bo_node->adev == node->adev &&
1250 			    (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
1251 				mapping_flags |= mtype_local;
1252 			/* local HBM region far from partition or remote XGMI GPU */
1253 			else if (svm_nodes_in_same_hive(bo_node, node))
1254 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1255 			/* PCIe P2P */
1256 			else
1257 				mapping_flags |= AMDGPU_VM_MTYPE_UC;
1258 		/* system memory accessed by the APU */
1259 		} else if (node->adev->flags & AMD_IS_APU) {
1260 			/* On NUMA systems, locality is determined per-page
1261 			 * in amdgpu_gmc_override_vm_pte_flags
1262 			 */
1263 			if (num_possible_nodes() <= 1)
1264 				mapping_flags |= mtype_local;
1265 			else
1266 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1267 		/* system memory accessed by the dGPU */
1268 		} else {
1269 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1270 		}
1271 		break;
1272 	default:
1273 		mapping_flags |= coherent ?
1274 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1275 	}
1276 
1277 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1278 
1279 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1280 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1281 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1282 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1283 
1284 	pte_flags = AMDGPU_PTE_VALID;
1285 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1286 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1287 
1288 	pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags);
1289 	return pte_flags;
1290 }
1291 
1292 static int
1293 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1294 			 uint64_t start, uint64_t last,
1295 			 struct dma_fence **fence)
1296 {
1297 	uint64_t init_pte_value = 0;
1298 
1299 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1300 
1301 	return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1302 				      last, init_pte_value, 0, 0, NULL, NULL,
1303 				      fence);
1304 }
1305 
1306 static int
1307 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1308 			  unsigned long last, uint32_t trigger)
1309 {
1310 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1311 	struct kfd_process_device *pdd;
1312 	struct dma_fence *fence = NULL;
1313 	struct kfd_process *p;
1314 	uint32_t gpuidx;
1315 	int r = 0;
1316 
1317 	if (!prange->mapped_to_gpu) {
1318 		pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1319 			 prange, prange->start, prange->last);
1320 		return 0;
1321 	}
1322 
1323 	if (prange->start == start && prange->last == last) {
1324 		pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1325 		prange->mapped_to_gpu = false;
1326 	}
1327 
1328 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1329 		  MAX_GPU_INSTANCE);
1330 	p = container_of(prange->svms, struct kfd_process, svms);
1331 
1332 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1333 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1334 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1335 		if (!pdd) {
1336 			pr_debug("failed to find device idx %d\n", gpuidx);
1337 			return -EINVAL;
1338 		}
1339 
1340 		kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1341 					     start, last, trigger);
1342 
1343 		r = svm_range_unmap_from_gpu(pdd->dev->adev,
1344 					     drm_priv_to_vm(pdd->drm_priv),
1345 					     start, last, &fence);
1346 		if (r)
1347 			break;
1348 
1349 		if (fence) {
1350 			r = dma_fence_wait(fence, false);
1351 			dma_fence_put(fence);
1352 			fence = NULL;
1353 			if (r)
1354 				break;
1355 		}
1356 		kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1357 	}
1358 
1359 	return r;
1360 }
1361 
1362 static int
1363 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1364 		     unsigned long offset, unsigned long npages, bool readonly,
1365 		     dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1366 		     struct dma_fence **fence, bool flush_tlb)
1367 {
1368 	struct amdgpu_device *adev = pdd->dev->adev;
1369 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1370 	uint64_t pte_flags;
1371 	unsigned long last_start;
1372 	int last_domain;
1373 	int r = 0;
1374 	int64_t i, j;
1375 
1376 	last_start = prange->start + offset;
1377 
1378 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1379 		 last_start, last_start + npages - 1, readonly);
1380 
1381 	for (i = offset; i < offset + npages; i++) {
1382 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1383 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1384 
1385 		/* Collect all pages in the same address range and memory domain
1386 		 * that can be mapped with a single call to update mapping.
1387 		 */
1388 		if (i < offset + npages - 1 &&
1389 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1390 			continue;
1391 
1392 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1393 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1394 
1395 		pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain);
1396 		if (readonly)
1397 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1398 
1399 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1400 			 prange->svms, last_start, prange->start + i,
1401 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1402 			 pte_flags);
1403 
1404 		/* For dGPU mode, we use same vm_manager to allocate VRAM for
1405 		 * different memory partition based on fpfn/lpfn, we should use
1406 		 * same vm_manager.vram_base_offset regardless memory partition.
1407 		 */
1408 		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL,
1409 					   last_start, prange->start + i,
1410 					   pte_flags,
1411 					   (last_start - prange->start) << PAGE_SHIFT,
1412 					   bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1413 					   NULL, dma_addr, &vm->last_update);
1414 
1415 		for (j = last_start - prange->start; j <= i; j++)
1416 			dma_addr[j] |= last_domain;
1417 
1418 		if (r) {
1419 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1420 			goto out;
1421 		}
1422 		last_start = prange->start + i + 1;
1423 	}
1424 
1425 	r = amdgpu_vm_update_pdes(adev, vm, false);
1426 	if (r) {
1427 		pr_debug("failed %d to update directories 0x%lx\n", r,
1428 			 prange->start);
1429 		goto out;
1430 	}
1431 
1432 	if (fence)
1433 		*fence = dma_fence_get(vm->last_update);
1434 
1435 out:
1436 	return r;
1437 }
1438 
1439 static int
1440 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1441 		      unsigned long npages, bool readonly,
1442 		      unsigned long *bitmap, bool wait, bool flush_tlb)
1443 {
1444 	struct kfd_process_device *pdd;
1445 	struct amdgpu_device *bo_adev = NULL;
1446 	struct kfd_process *p;
1447 	struct dma_fence *fence = NULL;
1448 	uint32_t gpuidx;
1449 	int r = 0;
1450 
1451 	if (prange->svm_bo && prange->ttm_res)
1452 		bo_adev = prange->svm_bo->node->adev;
1453 
1454 	p = container_of(prange->svms, struct kfd_process, svms);
1455 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1456 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1457 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1458 		if (!pdd) {
1459 			pr_debug("failed to find device idx %d\n", gpuidx);
1460 			return -EINVAL;
1461 		}
1462 
1463 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1464 		if (IS_ERR(pdd))
1465 			return -EINVAL;
1466 
1467 		if (bo_adev && pdd->dev->adev != bo_adev &&
1468 		    !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1469 			pr_debug("cannot map to device idx %d\n", gpuidx);
1470 			continue;
1471 		}
1472 
1473 		r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1474 					 prange->dma_addr[gpuidx],
1475 					 bo_adev, wait ? &fence : NULL,
1476 					 flush_tlb);
1477 		if (r)
1478 			break;
1479 
1480 		if (fence) {
1481 			r = dma_fence_wait(fence, false);
1482 			dma_fence_put(fence);
1483 			fence = NULL;
1484 			if (r) {
1485 				pr_debug("failed %d to dma fence wait\n", r);
1486 				break;
1487 			}
1488 		}
1489 
1490 		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1491 	}
1492 
1493 	return r;
1494 }
1495 
1496 struct svm_validate_context {
1497 	struct kfd_process *process;
1498 	struct svm_range *prange;
1499 	bool intr;
1500 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1501 	struct drm_exec exec;
1502 };
1503 
1504 static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr)
1505 {
1506 	struct kfd_process_device *pdd;
1507 	struct amdgpu_vm *vm;
1508 	uint32_t gpuidx;
1509 	int r;
1510 
1511 	drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0);
1512 	drm_exec_until_all_locked(&ctx->exec) {
1513 		for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1514 			pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1515 			if (!pdd) {
1516 				pr_debug("failed to find device idx %d\n", gpuidx);
1517 				r = -EINVAL;
1518 				goto unreserve_out;
1519 			}
1520 			vm = drm_priv_to_vm(pdd->drm_priv);
1521 
1522 			r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1523 			drm_exec_retry_on_contention(&ctx->exec);
1524 			if (unlikely(r)) {
1525 				pr_debug("failed %d to reserve bo\n", r);
1526 				goto unreserve_out;
1527 			}
1528 		}
1529 	}
1530 
1531 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1532 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1533 		if (!pdd) {
1534 			pr_debug("failed to find device idx %d\n", gpuidx);
1535 			r = -EINVAL;
1536 			goto unreserve_out;
1537 		}
1538 
1539 		r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1540 					      drm_priv_to_vm(pdd->drm_priv),
1541 					      svm_range_bo_validate, NULL);
1542 		if (r) {
1543 			pr_debug("failed %d validate pt bos\n", r);
1544 			goto unreserve_out;
1545 		}
1546 	}
1547 
1548 	return 0;
1549 
1550 unreserve_out:
1551 	drm_exec_fini(&ctx->exec);
1552 	return r;
1553 }
1554 
1555 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1556 {
1557 	drm_exec_fini(&ctx->exec);
1558 }
1559 
1560 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1561 {
1562 	struct kfd_process_device *pdd;
1563 
1564 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1565 	if (!pdd)
1566 		return NULL;
1567 
1568 	return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1569 }
1570 
1571 /*
1572  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1573  *
1574  * To prevent concurrent destruction or change of range attributes, the
1575  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1576  * because that would block concurrent evictions and lead to deadlocks. To
1577  * serialize concurrent migrations or validations of the same range, the
1578  * prange->migrate_mutex must be held.
1579  *
1580  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1581  * eviction fence.
1582  *
1583  * The following sequence ensures race-free validation and GPU mapping:
1584  *
1585  * 1. Reserve page table (and SVM BO if range is in VRAM)
1586  * 2. hmm_range_fault to get page addresses (if system memory)
1587  * 3. DMA-map pages (if system memory)
1588  * 4-a. Take notifier lock
1589  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1590  * 4-c. Check that the range was not split or otherwise invalidated
1591  * 4-d. Update GPU page table
1592  * 4.e. Release notifier lock
1593  * 5. Release page table (and SVM BO) reservation
1594  */
1595 static int svm_range_validate_and_map(struct mm_struct *mm,
1596 				      struct svm_range *prange, int32_t gpuidx,
1597 				      bool intr, bool wait, bool flush_tlb)
1598 {
1599 	struct svm_validate_context *ctx;
1600 	unsigned long start, end, addr;
1601 	struct kfd_process *p;
1602 	void *owner;
1603 	int32_t idx;
1604 	int r = 0;
1605 
1606 	ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
1607 	if (!ctx)
1608 		return -ENOMEM;
1609 	ctx->process = container_of(prange->svms, struct kfd_process, svms);
1610 	ctx->prange = prange;
1611 	ctx->intr = intr;
1612 
1613 	if (gpuidx < MAX_GPU_INSTANCE) {
1614 		bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE);
1615 		bitmap_set(ctx->bitmap, gpuidx, 1);
1616 	} else if (ctx->process->xnack_enabled) {
1617 		bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1618 
1619 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1620 		 * GPU, which has ACCESS attribute to the range, create mapping
1621 		 * on that GPU.
1622 		 */
1623 		if (prange->actual_loc) {
1624 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process,
1625 							prange->actual_loc);
1626 			if (gpuidx < 0) {
1627 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1628 					 prange->actual_loc);
1629 				r = -EINVAL;
1630 				goto free_ctx;
1631 			}
1632 			if (test_bit(gpuidx, prange->bitmap_access))
1633 				bitmap_set(ctx->bitmap, gpuidx, 1);
1634 		}
1635 	} else {
1636 		bitmap_or(ctx->bitmap, prange->bitmap_access,
1637 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1638 	}
1639 
1640 	if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1641 		bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1642 		if (!prange->mapped_to_gpu ||
1643 		    bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1644 			r = 0;
1645 			goto free_ctx;
1646 		}
1647 	}
1648 
1649 	if (prange->actual_loc && !prange->ttm_res) {
1650 		/* This should never happen. actual_loc gets set by
1651 		 * svm_migrate_ram_to_vram after allocating a BO.
1652 		 */
1653 		WARN_ONCE(1, "VRAM BO missing during validation\n");
1654 		r = -EINVAL;
1655 		goto free_ctx;
1656 	}
1657 
1658 	svm_range_reserve_bos(ctx, intr);
1659 
1660 	p = container_of(prange->svms, struct kfd_process, svms);
1661 	owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
1662 						MAX_GPU_INSTANCE));
1663 	for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) {
1664 		if (kfd_svm_page_owner(p, idx) != owner) {
1665 			owner = NULL;
1666 			break;
1667 		}
1668 	}
1669 
1670 	start = prange->start << PAGE_SHIFT;
1671 	end = (prange->last + 1) << PAGE_SHIFT;
1672 	for (addr = start; !r && addr < end; ) {
1673 		struct hmm_range *hmm_range;
1674 		struct vm_area_struct *vma;
1675 		unsigned long next = 0;
1676 		unsigned long offset;
1677 		unsigned long npages;
1678 		bool readonly;
1679 
1680 		vma = vma_lookup(mm, addr);
1681 		if (vma) {
1682 			readonly = !(vma->vm_flags & VM_WRITE);
1683 
1684 			next = min(vma->vm_end, end);
1685 			npages = (next - addr) >> PAGE_SHIFT;
1686 			WRITE_ONCE(p->svms.faulting_task, current);
1687 			r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages,
1688 						       readonly, owner, NULL,
1689 						       &hmm_range);
1690 			WRITE_ONCE(p->svms.faulting_task, NULL);
1691 			if (r) {
1692 				pr_debug("failed %d to get svm range pages\n", r);
1693 				if (r == -EBUSY)
1694 					r = -EAGAIN;
1695 			}
1696 		} else {
1697 			r = -EFAULT;
1698 		}
1699 
1700 		if (!r) {
1701 			offset = (addr - start) >> PAGE_SHIFT;
1702 			r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
1703 					      hmm_range->hmm_pfns);
1704 			if (r)
1705 				pr_debug("failed %d to dma map range\n", r);
1706 		}
1707 
1708 		svm_range_lock(prange);
1709 		if (!r && amdgpu_hmm_range_get_pages_done(hmm_range)) {
1710 			pr_debug("hmm update the range, need validate again\n");
1711 			r = -EAGAIN;
1712 		}
1713 
1714 		if (!r && !list_empty(&prange->child_list)) {
1715 			pr_debug("range split by unmap in parallel, validate again\n");
1716 			r = -EAGAIN;
1717 		}
1718 
1719 		if (!r)
1720 			r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1721 						  ctx->bitmap, wait, flush_tlb);
1722 
1723 		if (!r && next == end)
1724 			prange->mapped_to_gpu = true;
1725 
1726 		svm_range_unlock(prange);
1727 
1728 		addr = next;
1729 	}
1730 
1731 	svm_range_unreserve_bos(ctx);
1732 	if (!r)
1733 		prange->validate_timestamp = ktime_get_boottime();
1734 
1735 free_ctx:
1736 	kfree(ctx);
1737 
1738 	return r;
1739 }
1740 
1741 /**
1742  * svm_range_list_lock_and_flush_work - flush pending deferred work
1743  *
1744  * @svms: the svm range list
1745  * @mm: the mm structure
1746  *
1747  * Context: Returns with mmap write lock held, pending deferred work flushed
1748  *
1749  */
1750 void
1751 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1752 				   struct mm_struct *mm)
1753 {
1754 retry_flush_work:
1755 	flush_work(&svms->deferred_list_work);
1756 	mmap_write_lock(mm);
1757 
1758 	if (list_empty(&svms->deferred_range_list))
1759 		return;
1760 	mmap_write_unlock(mm);
1761 	pr_debug("retry flush\n");
1762 	goto retry_flush_work;
1763 }
1764 
1765 static void svm_range_restore_work(struct work_struct *work)
1766 {
1767 	struct delayed_work *dwork = to_delayed_work(work);
1768 	struct amdkfd_process_info *process_info;
1769 	struct svm_range_list *svms;
1770 	struct svm_range *prange;
1771 	struct kfd_process *p;
1772 	struct mm_struct *mm;
1773 	int evicted_ranges;
1774 	int invalid;
1775 	int r;
1776 
1777 	svms = container_of(dwork, struct svm_range_list, restore_work);
1778 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1779 	if (!evicted_ranges)
1780 		return;
1781 
1782 	pr_debug("restore svm ranges\n");
1783 
1784 	p = container_of(svms, struct kfd_process, svms);
1785 	process_info = p->kgd_process_info;
1786 
1787 	/* Keep mm reference when svm_range_validate_and_map ranges */
1788 	mm = get_task_mm(p->lead_thread);
1789 	if (!mm) {
1790 		pr_debug("svms 0x%p process mm gone\n", svms);
1791 		return;
1792 	}
1793 
1794 	mutex_lock(&process_info->lock);
1795 	svm_range_list_lock_and_flush_work(svms, mm);
1796 	mutex_lock(&svms->lock);
1797 
1798 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1799 
1800 	list_for_each_entry(prange, &svms->list, list) {
1801 		invalid = atomic_read(&prange->invalid);
1802 		if (!invalid)
1803 			continue;
1804 
1805 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1806 			 prange->svms, prange, prange->start, prange->last,
1807 			 invalid);
1808 
1809 		/*
1810 		 * If range is migrating, wait for migration is done.
1811 		 */
1812 		mutex_lock(&prange->migrate_mutex);
1813 
1814 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1815 					       false, true, false);
1816 		if (r)
1817 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1818 				 prange->start);
1819 
1820 		mutex_unlock(&prange->migrate_mutex);
1821 		if (r)
1822 			goto out_reschedule;
1823 
1824 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1825 			goto out_reschedule;
1826 	}
1827 
1828 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1829 	    evicted_ranges)
1830 		goto out_reschedule;
1831 
1832 	evicted_ranges = 0;
1833 
1834 	r = kgd2kfd_resume_mm(mm);
1835 	if (r) {
1836 		/* No recovery from this failure. Probably the CP is
1837 		 * hanging. No point trying again.
1838 		 */
1839 		pr_debug("failed %d to resume KFD\n", r);
1840 	}
1841 
1842 	pr_debug("restore svm ranges successfully\n");
1843 
1844 out_reschedule:
1845 	mutex_unlock(&svms->lock);
1846 	mmap_write_unlock(mm);
1847 	mutex_unlock(&process_info->lock);
1848 
1849 	/* If validation failed, reschedule another attempt */
1850 	if (evicted_ranges) {
1851 		pr_debug("reschedule to restore svm range\n");
1852 		schedule_delayed_work(&svms->restore_work,
1853 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1854 
1855 		kfd_smi_event_queue_restore_rescheduled(mm);
1856 	}
1857 	mmput(mm);
1858 }
1859 
1860 /**
1861  * svm_range_evict - evict svm range
1862  * @prange: svm range structure
1863  * @mm: current process mm_struct
1864  * @start: starting process queue number
1865  * @last: last process queue number
1866  * @event: mmu notifier event when range is evicted or migrated
1867  *
1868  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1869  * return to let CPU evict the buffer and proceed CPU pagetable update.
1870  *
1871  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1872  * If invalidation happens while restore work is running, restore work will
1873  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1874  * the queues.
1875  */
1876 static int
1877 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1878 		unsigned long start, unsigned long last,
1879 		enum mmu_notifier_event event)
1880 {
1881 	struct svm_range_list *svms = prange->svms;
1882 	struct svm_range *pchild;
1883 	struct kfd_process *p;
1884 	int r = 0;
1885 
1886 	p = container_of(svms, struct kfd_process, svms);
1887 
1888 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1889 		 svms, prange->start, prange->last, start, last);
1890 
1891 	if (!p->xnack_enabled ||
1892 	    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1893 		int evicted_ranges;
1894 		bool mapped = prange->mapped_to_gpu;
1895 
1896 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1897 			if (!pchild->mapped_to_gpu)
1898 				continue;
1899 			mapped = true;
1900 			mutex_lock_nested(&pchild->lock, 1);
1901 			if (pchild->start <= last && pchild->last >= start) {
1902 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1903 					 pchild->start, pchild->last);
1904 				atomic_inc(&pchild->invalid);
1905 			}
1906 			mutex_unlock(&pchild->lock);
1907 		}
1908 
1909 		if (!mapped)
1910 			return r;
1911 
1912 		if (prange->start <= last && prange->last >= start)
1913 			atomic_inc(&prange->invalid);
1914 
1915 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1916 		if (evicted_ranges != 1)
1917 			return r;
1918 
1919 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1920 			 prange->svms, prange->start, prange->last);
1921 
1922 		/* First eviction, stop the queues */
1923 		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1924 		if (r)
1925 			pr_debug("failed to quiesce KFD\n");
1926 
1927 		pr_debug("schedule to restore svm %p ranges\n", svms);
1928 		schedule_delayed_work(&svms->restore_work,
1929 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1930 	} else {
1931 		unsigned long s, l;
1932 		uint32_t trigger;
1933 
1934 		if (event == MMU_NOTIFY_MIGRATE)
1935 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1936 		else
1937 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1938 
1939 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1940 			 prange->svms, start, last);
1941 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1942 			mutex_lock_nested(&pchild->lock, 1);
1943 			s = max(start, pchild->start);
1944 			l = min(last, pchild->last);
1945 			if (l >= s)
1946 				svm_range_unmap_from_gpus(pchild, s, l, trigger);
1947 			mutex_unlock(&pchild->lock);
1948 		}
1949 		s = max(start, prange->start);
1950 		l = min(last, prange->last);
1951 		if (l >= s)
1952 			svm_range_unmap_from_gpus(prange, s, l, trigger);
1953 	}
1954 
1955 	return r;
1956 }
1957 
1958 static struct svm_range *svm_range_clone(struct svm_range *old)
1959 {
1960 	struct svm_range *new;
1961 
1962 	new = svm_range_new(old->svms, old->start, old->last, false);
1963 	if (!new)
1964 		return NULL;
1965 	if (svm_range_copy_dma_addrs(new, old)) {
1966 		svm_range_free(new, false);
1967 		return NULL;
1968 	}
1969 	if (old->svm_bo) {
1970 		new->ttm_res = old->ttm_res;
1971 		new->offset = old->offset;
1972 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1973 		spin_lock(&new->svm_bo->list_lock);
1974 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1975 		spin_unlock(&new->svm_bo->list_lock);
1976 	}
1977 	new->flags = old->flags;
1978 	new->preferred_loc = old->preferred_loc;
1979 	new->prefetch_loc = old->prefetch_loc;
1980 	new->actual_loc = old->actual_loc;
1981 	new->granularity = old->granularity;
1982 	new->mapped_to_gpu = old->mapped_to_gpu;
1983 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1984 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1985 
1986 	return new;
1987 }
1988 
1989 void svm_range_set_max_pages(struct amdgpu_device *adev)
1990 {
1991 	uint64_t max_pages;
1992 	uint64_t pages, _pages;
1993 	uint64_t min_pages = 0;
1994 	int i, id;
1995 
1996 	for (i = 0; i < adev->kfd.dev->num_nodes; i++) {
1997 		if (adev->kfd.dev->nodes[i]->xcp)
1998 			id = adev->kfd.dev->nodes[i]->xcp->id;
1999 		else
2000 			id = -1;
2001 		pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17;
2002 		pages = clamp(pages, 1ULL << 9, 1ULL << 18);
2003 		pages = rounddown_pow_of_two(pages);
2004 		min_pages = min_not_zero(min_pages, pages);
2005 	}
2006 
2007 	do {
2008 		max_pages = READ_ONCE(max_svm_range_pages);
2009 		_pages = min_not_zero(max_pages, min_pages);
2010 	} while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
2011 }
2012 
2013 static int
2014 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
2015 		    uint64_t max_pages, struct list_head *insert_list,
2016 		    struct list_head *update_list)
2017 {
2018 	struct svm_range *prange;
2019 	uint64_t l;
2020 
2021 	pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
2022 		 max_pages, start, last);
2023 
2024 	while (last >= start) {
2025 		l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
2026 
2027 		prange = svm_range_new(svms, start, l, true);
2028 		if (!prange)
2029 			return -ENOMEM;
2030 		list_add(&prange->list, insert_list);
2031 		list_add(&prange->update_list, update_list);
2032 
2033 		start = l + 1;
2034 	}
2035 	return 0;
2036 }
2037 
2038 /**
2039  * svm_range_add - add svm range and handle overlap
2040  * @p: the range add to this process svms
2041  * @start: page size aligned
2042  * @size: page size aligned
2043  * @nattr: number of attributes
2044  * @attrs: array of attributes
2045  * @update_list: output, the ranges need validate and update GPU mapping
2046  * @insert_list: output, the ranges need insert to svms
2047  * @remove_list: output, the ranges are replaced and need remove from svms
2048  *
2049  * Check if the virtual address range has overlap with any existing ranges,
2050  * split partly overlapping ranges and add new ranges in the gaps. All changes
2051  * should be applied to the range_list and interval tree transactionally. If
2052  * any range split or allocation fails, the entire update fails. Therefore any
2053  * existing overlapping svm_ranges are cloned and the original svm_ranges left
2054  * unchanged.
2055  *
2056  * If the transaction succeeds, the caller can update and insert clones and
2057  * new ranges, then free the originals.
2058  *
2059  * Otherwise the caller can free the clones and new ranges, while the old
2060  * svm_ranges remain unchanged.
2061  *
2062  * Context: Process context, caller must hold svms->lock
2063  *
2064  * Return:
2065  * 0 - OK, otherwise error code
2066  */
2067 static int
2068 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2069 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2070 	      struct list_head *update_list, struct list_head *insert_list,
2071 	      struct list_head *remove_list)
2072 {
2073 	unsigned long last = start + size - 1UL;
2074 	struct svm_range_list *svms = &p->svms;
2075 	struct interval_tree_node *node;
2076 	struct svm_range *prange;
2077 	struct svm_range *tmp;
2078 	struct list_head new_list;
2079 	int r = 0;
2080 
2081 	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
2082 
2083 	INIT_LIST_HEAD(update_list);
2084 	INIT_LIST_HEAD(insert_list);
2085 	INIT_LIST_HEAD(remove_list);
2086 	INIT_LIST_HEAD(&new_list);
2087 
2088 	node = interval_tree_iter_first(&svms->objects, start, last);
2089 	while (node) {
2090 		struct interval_tree_node *next;
2091 		unsigned long next_start;
2092 
2093 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2094 			 node->last);
2095 
2096 		prange = container_of(node, struct svm_range, it_node);
2097 		next = interval_tree_iter_next(node, start, last);
2098 		next_start = min(node->last, last) + 1;
2099 
2100 		if (svm_range_is_same_attrs(p, prange, nattr, attrs) &&
2101 		    prange->mapped_to_gpu) {
2102 			/* nothing to do */
2103 		} else if (node->start < start || node->last > last) {
2104 			/* node intersects the update range and its attributes
2105 			 * will change. Clone and split it, apply updates only
2106 			 * to the overlapping part
2107 			 */
2108 			struct svm_range *old = prange;
2109 
2110 			prange = svm_range_clone(old);
2111 			if (!prange) {
2112 				r = -ENOMEM;
2113 				goto out;
2114 			}
2115 
2116 			list_add(&old->update_list, remove_list);
2117 			list_add(&prange->list, insert_list);
2118 			list_add(&prange->update_list, update_list);
2119 
2120 			if (node->start < start) {
2121 				pr_debug("change old range start\n");
2122 				r = svm_range_split_head(prange, start,
2123 							 insert_list);
2124 				if (r)
2125 					goto out;
2126 			}
2127 			if (node->last > last) {
2128 				pr_debug("change old range last\n");
2129 				r = svm_range_split_tail(prange, last,
2130 							 insert_list);
2131 				if (r)
2132 					goto out;
2133 			}
2134 		} else {
2135 			/* The node is contained within start..last,
2136 			 * just update it
2137 			 */
2138 			list_add(&prange->update_list, update_list);
2139 		}
2140 
2141 		/* insert a new node if needed */
2142 		if (node->start > start) {
2143 			r = svm_range_split_new(svms, start, node->start - 1,
2144 						READ_ONCE(max_svm_range_pages),
2145 						&new_list, update_list);
2146 			if (r)
2147 				goto out;
2148 		}
2149 
2150 		node = next;
2151 		start = next_start;
2152 	}
2153 
2154 	/* add a final range at the end if needed */
2155 	if (start <= last)
2156 		r = svm_range_split_new(svms, start, last,
2157 					READ_ONCE(max_svm_range_pages),
2158 					&new_list, update_list);
2159 
2160 out:
2161 	if (r) {
2162 		list_for_each_entry_safe(prange, tmp, insert_list, list)
2163 			svm_range_free(prange, false);
2164 		list_for_each_entry_safe(prange, tmp, &new_list, list)
2165 			svm_range_free(prange, true);
2166 	} else {
2167 		list_splice(&new_list, insert_list);
2168 	}
2169 
2170 	return r;
2171 }
2172 
2173 static void
2174 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2175 					    struct svm_range *prange)
2176 {
2177 	unsigned long start;
2178 	unsigned long last;
2179 
2180 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2181 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2182 
2183 	if (prange->start == start && prange->last == last)
2184 		return;
2185 
2186 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2187 		  prange->svms, prange, start, last, prange->start,
2188 		  prange->last);
2189 
2190 	if (start != 0 && last != 0) {
2191 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
2192 		svm_range_remove_notifier(prange);
2193 	}
2194 	prange->it_node.start = prange->start;
2195 	prange->it_node.last = prange->last;
2196 
2197 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
2198 	svm_range_add_notifier_locked(mm, prange);
2199 }
2200 
2201 static void
2202 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2203 			 struct mm_struct *mm)
2204 {
2205 	switch (prange->work_item.op) {
2206 	case SVM_OP_NULL:
2207 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2208 			 svms, prange, prange->start, prange->last);
2209 		break;
2210 	case SVM_OP_UNMAP_RANGE:
2211 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2212 			 svms, prange, prange->start, prange->last);
2213 		svm_range_unlink(prange);
2214 		svm_range_remove_notifier(prange);
2215 		svm_range_free(prange, true);
2216 		break;
2217 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
2218 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2219 			 svms, prange, prange->start, prange->last);
2220 		svm_range_update_notifier_and_interval_tree(mm, prange);
2221 		break;
2222 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2223 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2224 			 svms, prange, prange->start, prange->last);
2225 		svm_range_update_notifier_and_interval_tree(mm, prange);
2226 		/* TODO: implement deferred validation and mapping */
2227 		break;
2228 	case SVM_OP_ADD_RANGE:
2229 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2230 			 prange->start, prange->last);
2231 		svm_range_add_to_svms(prange);
2232 		svm_range_add_notifier_locked(mm, prange);
2233 		break;
2234 	case SVM_OP_ADD_RANGE_AND_MAP:
2235 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2236 			 prange, prange->start, prange->last);
2237 		svm_range_add_to_svms(prange);
2238 		svm_range_add_notifier_locked(mm, prange);
2239 		/* TODO: implement deferred validation and mapping */
2240 		break;
2241 	default:
2242 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2243 			 prange->work_item.op);
2244 	}
2245 }
2246 
2247 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2248 {
2249 	struct kfd_process_device *pdd;
2250 	struct kfd_process *p;
2251 	int drain;
2252 	uint32_t i;
2253 
2254 	p = container_of(svms, struct kfd_process, svms);
2255 
2256 restart:
2257 	drain = atomic_read(&svms->drain_pagefaults);
2258 	if (!drain)
2259 		return;
2260 
2261 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2262 		pdd = p->pdds[i];
2263 		if (!pdd)
2264 			continue;
2265 
2266 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2267 
2268 		amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2269 				pdd->dev->adev->irq.retry_cam_enabled ?
2270 				&pdd->dev->adev->irq.ih :
2271 				&pdd->dev->adev->irq.ih1);
2272 
2273 		if (pdd->dev->adev->irq.retry_cam_enabled)
2274 			amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2275 				&pdd->dev->adev->irq.ih_soft);
2276 
2277 
2278 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2279 	}
2280 	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2281 		goto restart;
2282 }
2283 
2284 static void svm_range_deferred_list_work(struct work_struct *work)
2285 {
2286 	struct svm_range_list *svms;
2287 	struct svm_range *prange;
2288 	struct mm_struct *mm;
2289 
2290 	svms = container_of(work, struct svm_range_list, deferred_list_work);
2291 	pr_debug("enter svms 0x%p\n", svms);
2292 
2293 	spin_lock(&svms->deferred_list_lock);
2294 	while (!list_empty(&svms->deferred_range_list)) {
2295 		prange = list_first_entry(&svms->deferred_range_list,
2296 					  struct svm_range, deferred_list);
2297 		spin_unlock(&svms->deferred_list_lock);
2298 
2299 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2300 			 prange->start, prange->last, prange->work_item.op);
2301 
2302 		mm = prange->work_item.mm;
2303 retry:
2304 		mmap_write_lock(mm);
2305 
2306 		/* Checking for the need to drain retry faults must be inside
2307 		 * mmap write lock to serialize with munmap notifiers.
2308 		 */
2309 		if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2310 			mmap_write_unlock(mm);
2311 			svm_range_drain_retry_fault(svms);
2312 			goto retry;
2313 		}
2314 
2315 		/* Remove from deferred_list must be inside mmap write lock, for
2316 		 * two race cases:
2317 		 * 1. unmap_from_cpu may change work_item.op and add the range
2318 		 *    to deferred_list again, cause use after free bug.
2319 		 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2320 		 *    lock and continue because deferred_list is empty, but
2321 		 *    deferred_list work is actually waiting for mmap lock.
2322 		 */
2323 		spin_lock(&svms->deferred_list_lock);
2324 		list_del_init(&prange->deferred_list);
2325 		spin_unlock(&svms->deferred_list_lock);
2326 
2327 		mutex_lock(&svms->lock);
2328 		mutex_lock(&prange->migrate_mutex);
2329 		while (!list_empty(&prange->child_list)) {
2330 			struct svm_range *pchild;
2331 
2332 			pchild = list_first_entry(&prange->child_list,
2333 						struct svm_range, child_list);
2334 			pr_debug("child prange 0x%p op %d\n", pchild,
2335 				 pchild->work_item.op);
2336 			list_del_init(&pchild->child_list);
2337 			svm_range_handle_list_op(svms, pchild, mm);
2338 		}
2339 		mutex_unlock(&prange->migrate_mutex);
2340 
2341 		svm_range_handle_list_op(svms, prange, mm);
2342 		mutex_unlock(&svms->lock);
2343 		mmap_write_unlock(mm);
2344 
2345 		/* Pairs with mmget in svm_range_add_list_work */
2346 		mmput(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 
2354 void
2355 svm_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 
2380 void 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 
2388 static void
2389 svm_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 
2422 static void
2423 svm_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  */
2494 static bool
2495 svm_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  */
2548 struct svm_range *
2549 svm_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  */
2602 static int32_t
2603 svm_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 
2650 static int
2651 svm_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 	unsigned long start_limit, end_limit;
2658 
2659 	vma = vma_lookup(p->mm, addr << PAGE_SHIFT);
2660 	if (!vma) {
2661 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2662 		return -EFAULT;
2663 	}
2664 
2665 	*is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma);
2666 
2667 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2668 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2669 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2670 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2671 	/* First range that starts after the fault address */
2672 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2673 	if (node) {
2674 		end_limit = min(end_limit, node->start);
2675 		/* Last range that ends before the fault address */
2676 		node = container_of(rb_prev(&node->rb),
2677 				    struct interval_tree_node, rb);
2678 	} else {
2679 		/* Last range must end before addr because
2680 		 * there was no range after addr
2681 		 */
2682 		node = container_of(rb_last(&p->svms.objects.rb_root),
2683 				    struct interval_tree_node, rb);
2684 	}
2685 	if (node) {
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 
2703 static int
2704 svm_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 
2754 static struct
2755 svm_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  */
2817 static 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 
2843 static void
2844 svm_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 
2870 static bool
2871 svm_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 
2883 int
2884 svm_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 	struct mm_struct *mm = NULL;
2889 	struct svm_range_list *svms;
2890 	struct svm_range *prange;
2891 	struct kfd_process *p;
2892 	ktime_t timestamp = ktime_get_boottime();
2893 	struct kfd_node *node;
2894 	int32_t best_loc;
2895 	int32_t gpuidx = MAX_GPU_INSTANCE;
2896 	bool write_locked = false;
2897 	struct vm_area_struct *vma;
2898 	bool migration = false;
2899 	int r = 0;
2900 
2901 	if (!KFD_IS_SVM_API_SUPPORTED(adev)) {
2902 		pr_debug("device does not support SVM\n");
2903 		return -EFAULT;
2904 	}
2905 
2906 	p = kfd_lookup_process_by_pasid(pasid);
2907 	if (!p) {
2908 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2909 		return 0;
2910 	}
2911 	svms = &p->svms;
2912 
2913 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2914 
2915 	if (atomic_read(&svms->drain_pagefaults)) {
2916 		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2917 		r = 0;
2918 		goto out;
2919 	}
2920 
2921 	if (!p->xnack_enabled) {
2922 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2923 		r = -EFAULT;
2924 		goto out;
2925 	}
2926 
2927 	/* p->lead_thread is available as kfd_process_wq_release flush the work
2928 	 * before releasing task ref.
2929 	 */
2930 	mm = get_task_mm(p->lead_thread);
2931 	if (!mm) {
2932 		pr_debug("svms 0x%p failed to get mm\n", svms);
2933 		r = 0;
2934 		goto out;
2935 	}
2936 
2937 	node = kfd_node_by_irq_ids(adev, node_id, vmid);
2938 	if (!node) {
2939 		pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id,
2940 			 vmid);
2941 		r = -EFAULT;
2942 		goto out;
2943 	}
2944 	mmap_read_lock(mm);
2945 retry_write_locked:
2946 	mutex_lock(&svms->lock);
2947 	prange = svm_range_from_addr(svms, addr, NULL);
2948 	if (!prange) {
2949 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2950 			 svms, addr);
2951 		if (!write_locked) {
2952 			/* Need the write lock to create new range with MMU notifier.
2953 			 * Also flush pending deferred work to make sure the interval
2954 			 * tree is up to date before we add a new range
2955 			 */
2956 			mutex_unlock(&svms->lock);
2957 			mmap_read_unlock(mm);
2958 			mmap_write_lock(mm);
2959 			write_locked = true;
2960 			goto retry_write_locked;
2961 		}
2962 		prange = svm_range_create_unregistered_range(node, p, mm, addr);
2963 		if (!prange) {
2964 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2965 				 svms, addr);
2966 			mmap_write_downgrade(mm);
2967 			r = -EFAULT;
2968 			goto out_unlock_svms;
2969 		}
2970 	}
2971 	if (write_locked)
2972 		mmap_write_downgrade(mm);
2973 
2974 	mutex_lock(&prange->migrate_mutex);
2975 
2976 	if (svm_range_skip_recover(prange)) {
2977 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
2978 		r = 0;
2979 		goto out_unlock_range;
2980 	}
2981 
2982 	/* skip duplicate vm fault on different pages of same range */
2983 	if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
2984 				AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
2985 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2986 			 svms, prange->start, prange->last);
2987 		r = 0;
2988 		goto out_unlock_range;
2989 	}
2990 
2991 	/* __do_munmap removed VMA, return success as we are handling stale
2992 	 * retry fault.
2993 	 */
2994 	vma = vma_lookup(mm, addr << PAGE_SHIFT);
2995 	if (!vma) {
2996 		pr_debug("address 0x%llx VMA is removed\n", addr);
2997 		r = 0;
2998 		goto out_unlock_range;
2999 	}
3000 
3001 	if (!svm_fault_allowed(vma, write_fault)) {
3002 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
3003 			write_fault ? "write" : "read");
3004 		r = -EPERM;
3005 		goto out_unlock_range;
3006 	}
3007 
3008 	best_loc = svm_range_best_restore_location(prange, node, &gpuidx);
3009 	if (best_loc == -1) {
3010 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
3011 			 svms, prange->start, prange->last);
3012 		r = -EACCES;
3013 		goto out_unlock_range;
3014 	}
3015 
3016 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
3017 		 svms, prange->start, prange->last, best_loc,
3018 		 prange->actual_loc);
3019 
3020 	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
3021 				       write_fault, timestamp);
3022 
3023 	if (prange->actual_loc != best_loc) {
3024 		migration = true;
3025 		if (best_loc) {
3026 			r = svm_migrate_to_vram(prange, best_loc, mm,
3027 					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
3028 			if (r) {
3029 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
3030 					 r, addr);
3031 				/* Fallback to system memory if migration to
3032 				 * VRAM failed
3033 				 */
3034 				if (prange->actual_loc)
3035 					r = svm_migrate_vram_to_ram(prange, mm,
3036 					   KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3037 					   NULL);
3038 				else
3039 					r = 0;
3040 			}
3041 		} else {
3042 			r = svm_migrate_vram_to_ram(prange, mm,
3043 					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3044 					NULL);
3045 		}
3046 		if (r) {
3047 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
3048 				 r, svms, prange->start, prange->last);
3049 			goto out_unlock_range;
3050 		}
3051 	}
3052 
3053 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
3054 	if (r)
3055 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
3056 			 r, svms, prange->start, prange->last);
3057 
3058 	kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
3059 				     migration);
3060 
3061 out_unlock_range:
3062 	mutex_unlock(&prange->migrate_mutex);
3063 out_unlock_svms:
3064 	mutex_unlock(&svms->lock);
3065 	mmap_read_unlock(mm);
3066 
3067 	svm_range_count_fault(node, p, gpuidx);
3068 
3069 	mmput(mm);
3070 out:
3071 	kfd_unref_process(p);
3072 
3073 	if (r == -EAGAIN) {
3074 		pr_debug("recover vm fault later\n");
3075 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3076 		r = 0;
3077 	}
3078 	return r;
3079 }
3080 
3081 int
3082 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
3083 {
3084 	struct svm_range *prange, *pchild;
3085 	uint64_t reserved_size = 0;
3086 	uint64_t size;
3087 	int r = 0;
3088 
3089 	pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
3090 
3091 	mutex_lock(&p->svms.lock);
3092 
3093 	list_for_each_entry(prange, &p->svms.list, list) {
3094 		svm_range_lock(prange);
3095 		list_for_each_entry(pchild, &prange->child_list, child_list) {
3096 			size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
3097 			if (xnack_enabled) {
3098 				amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3099 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3100 			} else {
3101 				r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3102 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3103 				if (r)
3104 					goto out_unlock;
3105 				reserved_size += size;
3106 			}
3107 		}
3108 
3109 		size = (prange->last - prange->start + 1) << PAGE_SHIFT;
3110 		if (xnack_enabled) {
3111 			amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3112 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3113 		} else {
3114 			r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3115 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3116 			if (r)
3117 				goto out_unlock;
3118 			reserved_size += size;
3119 		}
3120 out_unlock:
3121 		svm_range_unlock(prange);
3122 		if (r)
3123 			break;
3124 	}
3125 
3126 	if (r)
3127 		amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3128 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3129 	else
3130 		/* Change xnack mode must be inside svms lock, to avoid race with
3131 		 * svm_range_deferred_list_work unreserve memory in parallel.
3132 		 */
3133 		p->xnack_enabled = xnack_enabled;
3134 
3135 	mutex_unlock(&p->svms.lock);
3136 	return r;
3137 }
3138 
3139 void svm_range_list_fini(struct kfd_process *p)
3140 {
3141 	struct svm_range *prange;
3142 	struct svm_range *next;
3143 
3144 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3145 
3146 	cancel_delayed_work_sync(&p->svms.restore_work);
3147 
3148 	/* Ensure list work is finished before process is destroyed */
3149 	flush_work(&p->svms.deferred_list_work);
3150 
3151 	/*
3152 	 * Ensure no retry fault comes in afterwards, as page fault handler will
3153 	 * not find kfd process and take mm lock to recover fault.
3154 	 */
3155 	atomic_inc(&p->svms.drain_pagefaults);
3156 	svm_range_drain_retry_fault(&p->svms);
3157 
3158 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3159 		svm_range_unlink(prange);
3160 		svm_range_remove_notifier(prange);
3161 		svm_range_free(prange, true);
3162 	}
3163 
3164 	mutex_destroy(&p->svms.lock);
3165 
3166 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3167 }
3168 
3169 int svm_range_list_init(struct kfd_process *p)
3170 {
3171 	struct svm_range_list *svms = &p->svms;
3172 	int i;
3173 
3174 	svms->objects = RB_ROOT_CACHED;
3175 	mutex_init(&svms->lock);
3176 	INIT_LIST_HEAD(&svms->list);
3177 	atomic_set(&svms->evicted_ranges, 0);
3178 	atomic_set(&svms->drain_pagefaults, 0);
3179 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3180 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3181 	INIT_LIST_HEAD(&svms->deferred_range_list);
3182 	INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3183 	spin_lock_init(&svms->deferred_list_lock);
3184 
3185 	for (i = 0; i < p->n_pdds; i++)
3186 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev))
3187 			bitmap_set(svms->bitmap_supported, i, 1);
3188 
3189 	return 0;
3190 }
3191 
3192 /**
3193  * svm_range_check_vm - check if virtual address range mapped already
3194  * @p: current kfd_process
3195  * @start: range start address, in pages
3196  * @last: range last address, in pages
3197  * @bo_s: mapping start address in pages if address range already mapped
3198  * @bo_l: mapping last address in pages if address range already mapped
3199  *
3200  * The purpose is to avoid virtual address ranges already allocated by
3201  * kfd_ioctl_alloc_memory_of_gpu ioctl.
3202  * It looks for each pdd in the kfd_process.
3203  *
3204  * Context: Process context
3205  *
3206  * Return 0 - OK, if the range is not mapped.
3207  * Otherwise error code:
3208  * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3209  * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3210  * a signal. Release all buffer reservations and return to user-space.
3211  */
3212 static int
3213 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3214 		   uint64_t *bo_s, uint64_t *bo_l)
3215 {
3216 	struct amdgpu_bo_va_mapping *mapping;
3217 	struct interval_tree_node *node;
3218 	uint32_t i;
3219 	int r;
3220 
3221 	for (i = 0; i < p->n_pdds; i++) {
3222 		struct amdgpu_vm *vm;
3223 
3224 		if (!p->pdds[i]->drm_priv)
3225 			continue;
3226 
3227 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3228 		r = amdgpu_bo_reserve(vm->root.bo, false);
3229 		if (r)
3230 			return r;
3231 
3232 		node = interval_tree_iter_first(&vm->va, start, last);
3233 		if (node) {
3234 			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3235 				 start, last);
3236 			mapping = container_of((struct rb_node *)node,
3237 					       struct amdgpu_bo_va_mapping, rb);
3238 			if (bo_s && bo_l) {
3239 				*bo_s = mapping->start;
3240 				*bo_l = mapping->last;
3241 			}
3242 			amdgpu_bo_unreserve(vm->root.bo);
3243 			return -EADDRINUSE;
3244 		}
3245 		amdgpu_bo_unreserve(vm->root.bo);
3246 	}
3247 
3248 	return 0;
3249 }
3250 
3251 /**
3252  * svm_range_is_valid - check if virtual address range is valid
3253  * @p: current kfd_process
3254  * @start: range start address, in pages
3255  * @size: range size, in pages
3256  *
3257  * Valid virtual address range means it belongs to one or more VMAs
3258  *
3259  * Context: Process context
3260  *
3261  * Return:
3262  *  0 - OK, otherwise error code
3263  */
3264 static int
3265 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3266 {
3267 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3268 	struct vm_area_struct *vma;
3269 	unsigned long end;
3270 	unsigned long start_unchg = start;
3271 
3272 	start <<= PAGE_SHIFT;
3273 	end = start + (size << PAGE_SHIFT);
3274 	do {
3275 		vma = vma_lookup(p->mm, start);
3276 		if (!vma || (vma->vm_flags & device_vma))
3277 			return -EFAULT;
3278 		start = min(end, vma->vm_end);
3279 	} while (start < end);
3280 
3281 	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3282 				  NULL);
3283 }
3284 
3285 /**
3286  * svm_range_best_prefetch_location - decide the best prefetch location
3287  * @prange: svm range structure
3288  *
3289  * For xnack off:
3290  * If range map to single GPU, the best prefetch location is prefetch_loc, which
3291  * can be CPU or GPU.
3292  *
3293  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3294  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3295  * the best prefetch location is always CPU, because GPU can not have coherent
3296  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3297  *
3298  * For xnack on:
3299  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3300  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3301  *
3302  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3303  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3304  * prefetch location is always CPU.
3305  *
3306  * Context: Process context
3307  *
3308  * Return:
3309  * 0 for CPU or GPU id
3310  */
3311 static uint32_t
3312 svm_range_best_prefetch_location(struct svm_range *prange)
3313 {
3314 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3315 	uint32_t best_loc = prange->prefetch_loc;
3316 	struct kfd_process_device *pdd;
3317 	struct kfd_node *bo_node;
3318 	struct kfd_process *p;
3319 	uint32_t gpuidx;
3320 
3321 	p = container_of(prange->svms, struct kfd_process, svms);
3322 
3323 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3324 		goto out;
3325 
3326 	bo_node = svm_range_get_node_by_id(prange, best_loc);
3327 	if (!bo_node) {
3328 		WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc);
3329 		best_loc = 0;
3330 		goto out;
3331 	}
3332 
3333 	if (bo_node->adev->gmc.is_app_apu) {
3334 		best_loc = 0;
3335 		goto out;
3336 	}
3337 
3338 	if (p->xnack_enabled)
3339 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3340 	else
3341 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3342 			  MAX_GPU_INSTANCE);
3343 
3344 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3345 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3346 		if (!pdd) {
3347 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3348 			continue;
3349 		}
3350 
3351 		if (pdd->dev->adev == bo_node->adev)
3352 			continue;
3353 
3354 		if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) {
3355 			best_loc = 0;
3356 			break;
3357 		}
3358 	}
3359 
3360 out:
3361 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3362 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3363 		 best_loc);
3364 
3365 	return best_loc;
3366 }
3367 
3368 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3369  * @mm: current process mm_struct
3370  * @prange: svm range structure
3371  * @migrated: output, true if migration is triggered
3372  *
3373  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3374  * from ram to vram.
3375  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3376  * from vram to ram.
3377  *
3378  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3379  * and restore work:
3380  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3381  *    stops all queues, schedule restore work
3382  * 2. svm_range_restore_work wait for migration is done by
3383  *    a. svm_range_validate_vram takes prange->migrate_mutex
3384  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3385  * 3. restore work update mappings of GPU, resume all queues.
3386  *
3387  * Context: Process context
3388  *
3389  * Return:
3390  * 0 - OK, otherwise - error code of migration
3391  */
3392 static int
3393 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3394 			    bool *migrated)
3395 {
3396 	uint32_t best_loc;
3397 	int r = 0;
3398 
3399 	*migrated = false;
3400 	best_loc = svm_range_best_prefetch_location(prange);
3401 
3402 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3403 	    best_loc == prange->actual_loc)
3404 		return 0;
3405 
3406 	if (!best_loc) {
3407 		r = svm_migrate_vram_to_ram(prange, mm,
3408 					KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3409 		*migrated = !r;
3410 		return r;
3411 	}
3412 
3413 	r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3414 	*migrated = !r;
3415 
3416 	return r;
3417 }
3418 
3419 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3420 {
3421 	if (!fence)
3422 		return -EINVAL;
3423 
3424 	if (dma_fence_is_signaled(&fence->base))
3425 		return 0;
3426 
3427 	if (fence->svm_bo) {
3428 		WRITE_ONCE(fence->svm_bo->evicting, 1);
3429 		schedule_work(&fence->svm_bo->eviction_work);
3430 	}
3431 
3432 	return 0;
3433 }
3434 
3435 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3436 {
3437 	struct svm_range_bo *svm_bo;
3438 	struct mm_struct *mm;
3439 	int r = 0;
3440 
3441 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3442 	if (!svm_bo_ref_unless_zero(svm_bo))
3443 		return; /* svm_bo was freed while eviction was pending */
3444 
3445 	if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3446 		mm = svm_bo->eviction_fence->mm;
3447 	} else {
3448 		svm_range_bo_unref(svm_bo);
3449 		return;
3450 	}
3451 
3452 	mmap_read_lock(mm);
3453 	spin_lock(&svm_bo->list_lock);
3454 	while (!list_empty(&svm_bo->range_list) && !r) {
3455 		struct svm_range *prange =
3456 				list_first_entry(&svm_bo->range_list,
3457 						struct svm_range, svm_bo_list);
3458 		int retries = 3;
3459 
3460 		list_del_init(&prange->svm_bo_list);
3461 		spin_unlock(&svm_bo->list_lock);
3462 
3463 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3464 			 prange->start, prange->last);
3465 
3466 		mutex_lock(&prange->migrate_mutex);
3467 		do {
3468 			r = svm_migrate_vram_to_ram(prange, mm,
3469 					KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3470 		} while (!r && prange->actual_loc && --retries);
3471 
3472 		if (!r && prange->actual_loc)
3473 			pr_info_once("Migration failed during eviction");
3474 
3475 		if (!prange->actual_loc) {
3476 			mutex_lock(&prange->lock);
3477 			prange->svm_bo = NULL;
3478 			mutex_unlock(&prange->lock);
3479 		}
3480 		mutex_unlock(&prange->migrate_mutex);
3481 
3482 		spin_lock(&svm_bo->list_lock);
3483 	}
3484 	spin_unlock(&svm_bo->list_lock);
3485 	mmap_read_unlock(mm);
3486 	mmput(mm);
3487 
3488 	dma_fence_signal(&svm_bo->eviction_fence->base);
3489 
3490 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3491 	 * has been called in svm_migrate_vram_to_ram
3492 	 */
3493 	WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3494 	svm_range_bo_unref(svm_bo);
3495 }
3496 
3497 static int
3498 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3499 		   uint64_t start, uint64_t size, uint32_t nattr,
3500 		   struct kfd_ioctl_svm_attribute *attrs)
3501 {
3502 	struct amdkfd_process_info *process_info = p->kgd_process_info;
3503 	struct list_head update_list;
3504 	struct list_head insert_list;
3505 	struct list_head remove_list;
3506 	struct svm_range_list *svms;
3507 	struct svm_range *prange;
3508 	struct svm_range *next;
3509 	bool update_mapping = false;
3510 	bool flush_tlb;
3511 	int r, ret = 0;
3512 
3513 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3514 		 p->pasid, &p->svms, start, start + size - 1, size);
3515 
3516 	r = svm_range_check_attr(p, nattr, attrs);
3517 	if (r)
3518 		return r;
3519 
3520 	svms = &p->svms;
3521 
3522 	mutex_lock(&process_info->lock);
3523 
3524 	svm_range_list_lock_and_flush_work(svms, mm);
3525 
3526 	r = svm_range_is_valid(p, start, size);
3527 	if (r) {
3528 		pr_debug("invalid range r=%d\n", r);
3529 		mmap_write_unlock(mm);
3530 		goto out;
3531 	}
3532 
3533 	mutex_lock(&svms->lock);
3534 
3535 	/* Add new range and split existing ranges as needed */
3536 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3537 			  &insert_list, &remove_list);
3538 	if (r) {
3539 		mutex_unlock(&svms->lock);
3540 		mmap_write_unlock(mm);
3541 		goto out;
3542 	}
3543 	/* Apply changes as a transaction */
3544 	list_for_each_entry_safe(prange, next, &insert_list, list) {
3545 		svm_range_add_to_svms(prange);
3546 		svm_range_add_notifier_locked(mm, prange);
3547 	}
3548 	list_for_each_entry(prange, &update_list, update_list) {
3549 		svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3550 		/* TODO: unmap ranges from GPU that lost access */
3551 	}
3552 	list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3553 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3554 			 prange->svms, prange, prange->start,
3555 			 prange->last);
3556 		svm_range_unlink(prange);
3557 		svm_range_remove_notifier(prange);
3558 		svm_range_free(prange, false);
3559 	}
3560 
3561 	mmap_write_downgrade(mm);
3562 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3563 	 * this fails we may be left with partially completed actions. There
3564 	 * is no clean way of rolling back to the previous state in such a
3565 	 * case because the rollback wouldn't be guaranteed to work either.
3566 	 */
3567 	list_for_each_entry(prange, &update_list, update_list) {
3568 		bool migrated;
3569 
3570 		mutex_lock(&prange->migrate_mutex);
3571 
3572 		r = svm_range_trigger_migration(mm, prange, &migrated);
3573 		if (r)
3574 			goto out_unlock_range;
3575 
3576 		if (migrated && (!p->xnack_enabled ||
3577 		    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3578 		    prange->mapped_to_gpu) {
3579 			pr_debug("restore_work will update mappings of GPUs\n");
3580 			mutex_unlock(&prange->migrate_mutex);
3581 			continue;
3582 		}
3583 
3584 		if (!migrated && !update_mapping) {
3585 			mutex_unlock(&prange->migrate_mutex);
3586 			continue;
3587 		}
3588 
3589 		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3590 
3591 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3592 					       true, true, flush_tlb);
3593 		if (r)
3594 			pr_debug("failed %d to map svm range\n", r);
3595 
3596 out_unlock_range:
3597 		mutex_unlock(&prange->migrate_mutex);
3598 		if (r)
3599 			ret = r;
3600 	}
3601 
3602 	dynamic_svm_range_dump(svms);
3603 
3604 	mutex_unlock(&svms->lock);
3605 	mmap_read_unlock(mm);
3606 out:
3607 	mutex_unlock(&process_info->lock);
3608 
3609 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3610 		 &p->svms, start, start + size - 1, r);
3611 
3612 	return ret ? ret : r;
3613 }
3614 
3615 static int
3616 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3617 		   uint64_t start, uint64_t size, uint32_t nattr,
3618 		   struct kfd_ioctl_svm_attribute *attrs)
3619 {
3620 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3621 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3622 	bool get_preferred_loc = false;
3623 	bool get_prefetch_loc = false;
3624 	bool get_granularity = false;
3625 	bool get_accessible = false;
3626 	bool get_flags = false;
3627 	uint64_t last = start + size - 1UL;
3628 	uint8_t granularity = 0xff;
3629 	struct interval_tree_node *node;
3630 	struct svm_range_list *svms;
3631 	struct svm_range *prange;
3632 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3633 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3634 	uint32_t flags_and = 0xffffffff;
3635 	uint32_t flags_or = 0;
3636 	int gpuidx;
3637 	uint32_t i;
3638 	int r = 0;
3639 
3640 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3641 		 start + size - 1, nattr);
3642 
3643 	/* Flush pending deferred work to avoid racing with deferred actions from
3644 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3645 	 * can still race with get_attr because we don't hold the mmap lock. But that
3646 	 * would be a race condition in the application anyway, and undefined
3647 	 * behaviour is acceptable in that case.
3648 	 */
3649 	flush_work(&p->svms.deferred_list_work);
3650 
3651 	mmap_read_lock(mm);
3652 	r = svm_range_is_valid(p, start, size);
3653 	mmap_read_unlock(mm);
3654 	if (r) {
3655 		pr_debug("invalid range r=%d\n", r);
3656 		return r;
3657 	}
3658 
3659 	for (i = 0; i < nattr; i++) {
3660 		switch (attrs[i].type) {
3661 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3662 			get_preferred_loc = true;
3663 			break;
3664 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3665 			get_prefetch_loc = true;
3666 			break;
3667 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3668 			get_accessible = true;
3669 			break;
3670 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3671 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3672 			get_flags = true;
3673 			break;
3674 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3675 			get_granularity = true;
3676 			break;
3677 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3678 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3679 			fallthrough;
3680 		default:
3681 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3682 			return -EINVAL;
3683 		}
3684 	}
3685 
3686 	svms = &p->svms;
3687 
3688 	mutex_lock(&svms->lock);
3689 
3690 	node = interval_tree_iter_first(&svms->objects, start, last);
3691 	if (!node) {
3692 		pr_debug("range attrs not found return default values\n");
3693 		svm_range_set_default_attributes(&location, &prefetch_loc,
3694 						 &granularity, &flags_and);
3695 		flags_or = flags_and;
3696 		if (p->xnack_enabled)
3697 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3698 				    MAX_GPU_INSTANCE);
3699 		else
3700 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3701 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3702 		goto fill_values;
3703 	}
3704 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3705 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3706 
3707 	while (node) {
3708 		struct interval_tree_node *next;
3709 
3710 		prange = container_of(node, struct svm_range, it_node);
3711 		next = interval_tree_iter_next(node, start, last);
3712 
3713 		if (get_preferred_loc) {
3714 			if (prange->preferred_loc ==
3715 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3716 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3717 			     location != prange->preferred_loc)) {
3718 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3719 				get_preferred_loc = false;
3720 			} else {
3721 				location = prange->preferred_loc;
3722 			}
3723 		}
3724 		if (get_prefetch_loc) {
3725 			if (prange->prefetch_loc ==
3726 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3727 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3728 			     prefetch_loc != prange->prefetch_loc)) {
3729 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3730 				get_prefetch_loc = false;
3731 			} else {
3732 				prefetch_loc = prange->prefetch_loc;
3733 			}
3734 		}
3735 		if (get_accessible) {
3736 			bitmap_and(bitmap_access, bitmap_access,
3737 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3738 			bitmap_and(bitmap_aip, bitmap_aip,
3739 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3740 		}
3741 		if (get_flags) {
3742 			flags_and &= prange->flags;
3743 			flags_or |= prange->flags;
3744 		}
3745 
3746 		if (get_granularity && prange->granularity < granularity)
3747 			granularity = prange->granularity;
3748 
3749 		node = next;
3750 	}
3751 fill_values:
3752 	mutex_unlock(&svms->lock);
3753 
3754 	for (i = 0; i < nattr; i++) {
3755 		switch (attrs[i].type) {
3756 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3757 			attrs[i].value = location;
3758 			break;
3759 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3760 			attrs[i].value = prefetch_loc;
3761 			break;
3762 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3763 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3764 							       attrs[i].value);
3765 			if (gpuidx < 0) {
3766 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3767 				return -EINVAL;
3768 			}
3769 			if (test_bit(gpuidx, bitmap_access))
3770 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3771 			else if (test_bit(gpuidx, bitmap_aip))
3772 				attrs[i].type =
3773 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3774 			else
3775 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3776 			break;
3777 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3778 			attrs[i].value = flags_and;
3779 			break;
3780 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3781 			attrs[i].value = ~flags_or;
3782 			break;
3783 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3784 			attrs[i].value = (uint32_t)granularity;
3785 			break;
3786 		}
3787 	}
3788 
3789 	return 0;
3790 }
3791 
3792 int kfd_criu_resume_svm(struct kfd_process *p)
3793 {
3794 	struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3795 	int nattr_common = 4, nattr_accessibility = 1;
3796 	struct criu_svm_metadata *criu_svm_md = NULL;
3797 	struct svm_range_list *svms = &p->svms;
3798 	struct criu_svm_metadata *next = NULL;
3799 	uint32_t set_flags = 0xffffffff;
3800 	int i, j, num_attrs, ret = 0;
3801 	uint64_t set_attr_size;
3802 	struct mm_struct *mm;
3803 
3804 	if (list_empty(&svms->criu_svm_metadata_list)) {
3805 		pr_debug("No SVM data from CRIU restore stage 2\n");
3806 		return ret;
3807 	}
3808 
3809 	mm = get_task_mm(p->lead_thread);
3810 	if (!mm) {
3811 		pr_err("failed to get mm for the target process\n");
3812 		return -ESRCH;
3813 	}
3814 
3815 	num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3816 
3817 	i = j = 0;
3818 	list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3819 		pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3820 			 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3821 
3822 		for (j = 0; j < num_attrs; j++) {
3823 			pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3824 				 i, j, criu_svm_md->data.attrs[j].type,
3825 				 i, j, criu_svm_md->data.attrs[j].value);
3826 			switch (criu_svm_md->data.attrs[j].type) {
3827 			/* During Checkpoint operation, the query for
3828 			 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3829 			 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3830 			 * not used by the range which was checkpointed. Care
3831 			 * must be taken to not restore with an invalid value
3832 			 * otherwise the gpuidx value will be invalid and
3833 			 * set_attr would eventually fail so just replace those
3834 			 * with another dummy attribute such as
3835 			 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3836 			 */
3837 			case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3838 				if (criu_svm_md->data.attrs[j].value ==
3839 				    KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3840 					criu_svm_md->data.attrs[j].type =
3841 						KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3842 					criu_svm_md->data.attrs[j].value = 0;
3843 				}
3844 				break;
3845 			case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3846 				set_flags = criu_svm_md->data.attrs[j].value;
3847 				break;
3848 			default:
3849 				break;
3850 			}
3851 		}
3852 
3853 		/* CLR_FLAGS is not available via get_attr during checkpoint but
3854 		 * it needs to be inserted before restoring the ranges so
3855 		 * allocate extra space for it before calling set_attr
3856 		 */
3857 		set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3858 						(num_attrs + 1);
3859 		set_attr_new = krealloc(set_attr, set_attr_size,
3860 					    GFP_KERNEL);
3861 		if (!set_attr_new) {
3862 			ret = -ENOMEM;
3863 			goto exit;
3864 		}
3865 		set_attr = set_attr_new;
3866 
3867 		memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3868 					sizeof(struct kfd_ioctl_svm_attribute));
3869 		set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3870 		set_attr[num_attrs].value = ~set_flags;
3871 
3872 		ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3873 					 criu_svm_md->data.size, num_attrs + 1,
3874 					 set_attr);
3875 		if (ret) {
3876 			pr_err("CRIU: failed to set range attributes\n");
3877 			goto exit;
3878 		}
3879 
3880 		i++;
3881 	}
3882 exit:
3883 	kfree(set_attr);
3884 	list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3885 		pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3886 						criu_svm_md->data.start_addr);
3887 		kfree(criu_svm_md);
3888 	}
3889 
3890 	mmput(mm);
3891 	return ret;
3892 
3893 }
3894 
3895 int kfd_criu_restore_svm(struct kfd_process *p,
3896 			 uint8_t __user *user_priv_ptr,
3897 			 uint64_t *priv_data_offset,
3898 			 uint64_t max_priv_data_size)
3899 {
3900 	uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3901 	int nattr_common = 4, nattr_accessibility = 1;
3902 	struct criu_svm_metadata *criu_svm_md = NULL;
3903 	struct svm_range_list *svms = &p->svms;
3904 	uint32_t num_devices;
3905 	int ret = 0;
3906 
3907 	num_devices = p->n_pdds;
3908 	/* Handle one SVM range object at a time, also the number of gpus are
3909 	 * assumed to be same on the restore node, checking must be done while
3910 	 * evaluating the topology earlier
3911 	 */
3912 
3913 	svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3914 		(nattr_common + nattr_accessibility * num_devices);
3915 	svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3916 
3917 	svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3918 								svm_attrs_size;
3919 
3920 	criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3921 	if (!criu_svm_md) {
3922 		pr_err("failed to allocate memory to store svm metadata\n");
3923 		return -ENOMEM;
3924 	}
3925 	if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3926 		ret = -EINVAL;
3927 		goto exit;
3928 	}
3929 
3930 	ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3931 			     svm_priv_data_size);
3932 	if (ret) {
3933 		ret = -EFAULT;
3934 		goto exit;
3935 	}
3936 	*priv_data_offset += svm_priv_data_size;
3937 
3938 	list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3939 
3940 	return 0;
3941 
3942 
3943 exit:
3944 	kfree(criu_svm_md);
3945 	return ret;
3946 }
3947 
3948 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3949 		       uint64_t *svm_priv_data_size)
3950 {
3951 	uint64_t total_size, accessibility_size, common_attr_size;
3952 	int nattr_common = 4, nattr_accessibility = 1;
3953 	int num_devices = p->n_pdds;
3954 	struct svm_range_list *svms;
3955 	struct svm_range *prange;
3956 	uint32_t count = 0;
3957 
3958 	*svm_priv_data_size = 0;
3959 
3960 	svms = &p->svms;
3961 	if (!svms)
3962 		return -EINVAL;
3963 
3964 	mutex_lock(&svms->lock);
3965 	list_for_each_entry(prange, &svms->list, list) {
3966 		pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3967 			 prange, prange->start, prange->npages,
3968 			 prange->start + prange->npages - 1);
3969 		count++;
3970 	}
3971 	mutex_unlock(&svms->lock);
3972 
3973 	*num_svm_ranges = count;
3974 	/* Only the accessbility attributes need to be queried for all the gpus
3975 	 * individually, remaining ones are spanned across the entire process
3976 	 * regardless of the various gpu nodes. Of the remaining attributes,
3977 	 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3978 	 *
3979 	 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3980 	 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3981 	 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3982 	 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3983 	 *
3984 	 * ** ACCESSBILITY ATTRIBUTES **
3985 	 * (Considered as one, type is altered during query, value is gpuid)
3986 	 * KFD_IOCTL_SVM_ATTR_ACCESS
3987 	 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3988 	 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3989 	 */
3990 	if (*num_svm_ranges > 0) {
3991 		common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3992 			nattr_common;
3993 		accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3994 			nattr_accessibility * num_devices;
3995 
3996 		total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3997 			common_attr_size + accessibility_size;
3998 
3999 		*svm_priv_data_size = *num_svm_ranges * total_size;
4000 	}
4001 
4002 	pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
4003 		 *svm_priv_data_size);
4004 	return 0;
4005 }
4006 
4007 int kfd_criu_checkpoint_svm(struct kfd_process *p,
4008 			    uint8_t __user *user_priv_data,
4009 			    uint64_t *priv_data_offset)
4010 {
4011 	struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
4012 	struct kfd_ioctl_svm_attribute *query_attr = NULL;
4013 	uint64_t svm_priv_data_size, query_attr_size = 0;
4014 	int index, nattr_common = 4, ret = 0;
4015 	struct svm_range_list *svms;
4016 	int num_devices = p->n_pdds;
4017 	struct svm_range *prange;
4018 	struct mm_struct *mm;
4019 
4020 	svms = &p->svms;
4021 	if (!svms)
4022 		return -EINVAL;
4023 
4024 	mm = get_task_mm(p->lead_thread);
4025 	if (!mm) {
4026 		pr_err("failed to get mm for the target process\n");
4027 		return -ESRCH;
4028 	}
4029 
4030 	query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4031 				(nattr_common + num_devices);
4032 
4033 	query_attr = kzalloc(query_attr_size, GFP_KERNEL);
4034 	if (!query_attr) {
4035 		ret = -ENOMEM;
4036 		goto exit;
4037 	}
4038 
4039 	query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
4040 	query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
4041 	query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
4042 	query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
4043 
4044 	for (index = 0; index < num_devices; index++) {
4045 		struct kfd_process_device *pdd = p->pdds[index];
4046 
4047 		query_attr[index + nattr_common].type =
4048 			KFD_IOCTL_SVM_ATTR_ACCESS;
4049 		query_attr[index + nattr_common].value = pdd->user_gpu_id;
4050 	}
4051 
4052 	svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
4053 
4054 	svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
4055 	if (!svm_priv) {
4056 		ret = -ENOMEM;
4057 		goto exit_query;
4058 	}
4059 
4060 	index = 0;
4061 	list_for_each_entry(prange, &svms->list, list) {
4062 
4063 		svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
4064 		svm_priv->start_addr = prange->start;
4065 		svm_priv->size = prange->npages;
4066 		memcpy(&svm_priv->attrs, query_attr, query_attr_size);
4067 		pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
4068 			 prange, prange->start, prange->npages,
4069 			 prange->start + prange->npages - 1,
4070 			 prange->npages * PAGE_SIZE);
4071 
4072 		ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
4073 					 svm_priv->size,
4074 					 (nattr_common + num_devices),
4075 					 svm_priv->attrs);
4076 		if (ret) {
4077 			pr_err("CRIU: failed to obtain range attributes\n");
4078 			goto exit_priv;
4079 		}
4080 
4081 		if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
4082 				 svm_priv_data_size)) {
4083 			pr_err("Failed to copy svm priv to user\n");
4084 			ret = -EFAULT;
4085 			goto exit_priv;
4086 		}
4087 
4088 		*priv_data_offset += svm_priv_data_size;
4089 
4090 	}
4091 
4092 
4093 exit_priv:
4094 	kfree(svm_priv);
4095 exit_query:
4096 	kfree(query_attr);
4097 exit:
4098 	mmput(mm);
4099 	return ret;
4100 }
4101 
4102 int
4103 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
4104 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
4105 {
4106 	struct mm_struct *mm = current->mm;
4107 	int r;
4108 
4109 	start >>= PAGE_SHIFT;
4110 	size >>= PAGE_SHIFT;
4111 
4112 	switch (op) {
4113 	case KFD_IOCTL_SVM_OP_SET_ATTR:
4114 		r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4115 		break;
4116 	case KFD_IOCTL_SVM_OP_GET_ATTR:
4117 		r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);
4118 		break;
4119 	default:
4120 		r = EINVAL;
4121 		break;
4122 	}
4123 
4124 	return r;
4125 }
4126