xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 16b0314a)
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 "amdgpu_sync.h"
27 #include "amdgpu_object.h"
28 #include "amdgpu_vm.h"
29 #include "amdgpu_mn.h"
30 #include "amdgpu.h"
31 #include "amdgpu_xgmi.h"
32 #include "kfd_priv.h"
33 #include "kfd_svm.h"
34 #include "kfd_migrate.h"
35 
36 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
37 
38 /* Long enough to ensure no retry fault comes after svm range is restored and
39  * page table is updated.
40  */
41 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING	2000
42 
43 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
44 static bool
45 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
46 				    const struct mmu_notifier_range *range,
47 				    unsigned long cur_seq);
48 
49 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
50 	.invalidate = svm_range_cpu_invalidate_pagetables,
51 };
52 
53 /**
54  * svm_range_unlink - unlink svm_range from lists and interval tree
55  * @prange: svm range structure to be removed
56  *
57  * Remove the svm_range from the svms and svm_bo lists and the svms
58  * interval tree.
59  *
60  * Context: The caller must hold svms->lock
61  */
62 static void svm_range_unlink(struct svm_range *prange)
63 {
64 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
65 		 prange, prange->start, prange->last);
66 
67 	if (prange->svm_bo) {
68 		spin_lock(&prange->svm_bo->list_lock);
69 		list_del(&prange->svm_bo_list);
70 		spin_unlock(&prange->svm_bo->list_lock);
71 	}
72 
73 	list_del(&prange->list);
74 	if (prange->it_node.start != 0 && prange->it_node.last != 0)
75 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
76 }
77 
78 static void
79 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
80 {
81 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
82 		 prange, prange->start, prange->last);
83 
84 	mmu_interval_notifier_insert_locked(&prange->notifier, mm,
85 				     prange->start << PAGE_SHIFT,
86 				     prange->npages << PAGE_SHIFT,
87 				     &svm_range_mn_ops);
88 }
89 
90 /**
91  * svm_range_add_to_svms - add svm range to svms
92  * @prange: svm range structure to be added
93  *
94  * Add the svm range to svms interval tree and link list
95  *
96  * Context: The caller must hold svms->lock
97  */
98 static void svm_range_add_to_svms(struct svm_range *prange)
99 {
100 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
101 		 prange, prange->start, prange->last);
102 
103 	list_add_tail(&prange->list, &prange->svms->list);
104 	prange->it_node.start = prange->start;
105 	prange->it_node.last = prange->last;
106 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
107 }
108 
109 static void svm_range_remove_notifier(struct svm_range *prange)
110 {
111 	pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
112 		 prange->svms, prange,
113 		 prange->notifier.interval_tree.start >> PAGE_SHIFT,
114 		 prange->notifier.interval_tree.last >> PAGE_SHIFT);
115 
116 	if (prange->notifier.interval_tree.start != 0 &&
117 	    prange->notifier.interval_tree.last != 0)
118 		mmu_interval_notifier_remove(&prange->notifier);
119 }
120 
121 static bool
122 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
123 {
124 	return dma_addr && !dma_mapping_error(dev, dma_addr) &&
125 	       !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
126 }
127 
128 static int
129 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
130 		      unsigned long offset, unsigned long npages,
131 		      unsigned long *hmm_pfns, uint32_t gpuidx)
132 {
133 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
134 	dma_addr_t *addr = prange->dma_addr[gpuidx];
135 	struct device *dev = adev->dev;
136 	struct page *page;
137 	int i, r;
138 
139 	if (!addr) {
140 		addr = kvmalloc_array(prange->npages, sizeof(*addr),
141 				      GFP_KERNEL | __GFP_ZERO);
142 		if (!addr)
143 			return -ENOMEM;
144 		prange->dma_addr[gpuidx] = addr;
145 	}
146 
147 	addr += offset;
148 	for (i = 0; i < npages; i++) {
149 		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
150 			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
151 
152 		page = hmm_pfn_to_page(hmm_pfns[i]);
153 		if (is_zone_device_page(page)) {
154 			struct amdgpu_device *bo_adev =
155 					amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
156 
157 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
158 				   bo_adev->vm_manager.vram_base_offset -
159 				   bo_adev->kfd.dev->pgmap.range.start;
160 			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
161 			pr_debug("vram address detected: 0x%llx\n", addr[i]);
162 			continue;
163 		}
164 		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
165 		r = dma_mapping_error(dev, addr[i]);
166 		if (r) {
167 			pr_debug("failed %d dma_map_page\n", r);
168 			return r;
169 		}
170 		pr_debug("dma mapping 0x%llx for page addr 0x%lx\n",
171 			 addr[i] >> PAGE_SHIFT, page_to_pfn(page));
172 	}
173 	return 0;
174 }
175 
176 static int
177 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
178 		  unsigned long offset, unsigned long npages,
179 		  unsigned long *hmm_pfns)
180 {
181 	struct kfd_process *p;
182 	uint32_t gpuidx;
183 	int r;
184 
185 	p = container_of(prange->svms, struct kfd_process, svms);
186 
187 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
188 		struct kfd_process_device *pdd;
189 		struct amdgpu_device *adev;
190 
191 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
192 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
193 		if (!pdd) {
194 			pr_debug("failed to find device idx %d\n", gpuidx);
195 			return -EINVAL;
196 		}
197 		adev = (struct amdgpu_device *)pdd->dev->kgd;
198 
199 		r = svm_range_dma_map_dev(adev, prange, offset, npages,
200 					  hmm_pfns, gpuidx);
201 		if (r)
202 			break;
203 	}
204 
205 	return r;
206 }
207 
208 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
209 			 unsigned long offset, unsigned long npages)
210 {
211 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
212 	int i;
213 
214 	if (!dma_addr)
215 		return;
216 
217 	for (i = offset; i < offset + npages; i++) {
218 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
219 			continue;
220 		pr_debug("dma unmapping 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
221 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
222 		dma_addr[i] = 0;
223 	}
224 }
225 
226 void svm_range_free_dma_mappings(struct svm_range *prange)
227 {
228 	struct kfd_process_device *pdd;
229 	dma_addr_t *dma_addr;
230 	struct device *dev;
231 	struct kfd_process *p;
232 	uint32_t gpuidx;
233 
234 	p = container_of(prange->svms, struct kfd_process, svms);
235 
236 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
237 		dma_addr = prange->dma_addr[gpuidx];
238 		if (!dma_addr)
239 			continue;
240 
241 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
242 		if (!pdd) {
243 			pr_debug("failed to find device idx %d\n", gpuidx);
244 			continue;
245 		}
246 		dev = &pdd->dev->pdev->dev;
247 		svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
248 		kvfree(dma_addr);
249 		prange->dma_addr[gpuidx] = NULL;
250 	}
251 }
252 
253 static void svm_range_free(struct svm_range *prange)
254 {
255 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
256 		 prange->start, prange->last);
257 
258 	svm_range_vram_node_free(prange);
259 	svm_range_free_dma_mappings(prange);
260 	mutex_destroy(&prange->lock);
261 	mutex_destroy(&prange->migrate_mutex);
262 	kfree(prange);
263 }
264 
265 static void
266 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
267 				 uint8_t *granularity, uint32_t *flags)
268 {
269 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
270 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
271 	*granularity = 9;
272 	*flags =
273 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
274 }
275 
276 static struct
277 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
278 			 uint64_t last)
279 {
280 	uint64_t size = last - start + 1;
281 	struct svm_range *prange;
282 	struct kfd_process *p;
283 
284 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
285 	if (!prange)
286 		return NULL;
287 	prange->npages = size;
288 	prange->svms = svms;
289 	prange->start = start;
290 	prange->last = last;
291 	INIT_LIST_HEAD(&prange->list);
292 	INIT_LIST_HEAD(&prange->update_list);
293 	INIT_LIST_HEAD(&prange->remove_list);
294 	INIT_LIST_HEAD(&prange->insert_list);
295 	INIT_LIST_HEAD(&prange->svm_bo_list);
296 	INIT_LIST_HEAD(&prange->deferred_list);
297 	INIT_LIST_HEAD(&prange->child_list);
298 	atomic_set(&prange->invalid, 0);
299 	prange->validate_timestamp = 0;
300 	mutex_init(&prange->migrate_mutex);
301 	mutex_init(&prange->lock);
302 
303 	p = container_of(svms, struct kfd_process, svms);
304 	if (p->xnack_enabled)
305 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
306 			    MAX_GPU_INSTANCE);
307 
308 	svm_range_set_default_attributes(&prange->preferred_loc,
309 					 &prange->prefetch_loc,
310 					 &prange->granularity, &prange->flags);
311 
312 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
313 
314 	return prange;
315 }
316 
317 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
318 {
319 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
320 		return false;
321 
322 	return true;
323 }
324 
325 static void svm_range_bo_release(struct kref *kref)
326 {
327 	struct svm_range_bo *svm_bo;
328 
329 	svm_bo = container_of(kref, struct svm_range_bo, kref);
330 	spin_lock(&svm_bo->list_lock);
331 	while (!list_empty(&svm_bo->range_list)) {
332 		struct svm_range *prange =
333 				list_first_entry(&svm_bo->range_list,
334 						struct svm_range, svm_bo_list);
335 		/* list_del_init tells a concurrent svm_range_vram_node_new when
336 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
337 		 */
338 		list_del_init(&prange->svm_bo_list);
339 		spin_unlock(&svm_bo->list_lock);
340 
341 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
342 			 prange->start, prange->last);
343 		mutex_lock(&prange->lock);
344 		prange->svm_bo = NULL;
345 		mutex_unlock(&prange->lock);
346 
347 		spin_lock(&svm_bo->list_lock);
348 	}
349 	spin_unlock(&svm_bo->list_lock);
350 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
351 		/* We're not in the eviction worker.
352 		 * Signal the fence and synchronize with any
353 		 * pending eviction work.
354 		 */
355 		dma_fence_signal(&svm_bo->eviction_fence->base);
356 		cancel_work_sync(&svm_bo->eviction_work);
357 	}
358 	dma_fence_put(&svm_bo->eviction_fence->base);
359 	amdgpu_bo_unref(&svm_bo->bo);
360 	kfree(svm_bo);
361 }
362 
363 void svm_range_bo_unref(struct svm_range_bo *svm_bo)
364 {
365 	if (!svm_bo)
366 		return;
367 
368 	kref_put(&svm_bo->kref, svm_range_bo_release);
369 }
370 
371 static bool
372 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
373 {
374 	struct amdgpu_device *bo_adev;
375 
376 	mutex_lock(&prange->lock);
377 	if (!prange->svm_bo) {
378 		mutex_unlock(&prange->lock);
379 		return false;
380 	}
381 	if (prange->ttm_res) {
382 		/* We still have a reference, all is well */
383 		mutex_unlock(&prange->lock);
384 		return true;
385 	}
386 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
387 		/*
388 		 * Migrate from GPU to GPU, remove range from source bo_adev
389 		 * svm_bo range list, and return false to allocate svm_bo from
390 		 * destination adev.
391 		 */
392 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
393 		if (bo_adev != adev) {
394 			mutex_unlock(&prange->lock);
395 
396 			spin_lock(&prange->svm_bo->list_lock);
397 			list_del_init(&prange->svm_bo_list);
398 			spin_unlock(&prange->svm_bo->list_lock);
399 
400 			svm_range_bo_unref(prange->svm_bo);
401 			return false;
402 		}
403 		if (READ_ONCE(prange->svm_bo->evicting)) {
404 			struct dma_fence *f;
405 			struct svm_range_bo *svm_bo;
406 			/* The BO is getting evicted,
407 			 * we need to get a new one
408 			 */
409 			mutex_unlock(&prange->lock);
410 			svm_bo = prange->svm_bo;
411 			f = dma_fence_get(&svm_bo->eviction_fence->base);
412 			svm_range_bo_unref(prange->svm_bo);
413 			/* wait for the fence to avoid long spin-loop
414 			 * at list_empty_careful
415 			 */
416 			dma_fence_wait(f, false);
417 			dma_fence_put(f);
418 		} else {
419 			/* The BO was still around and we got
420 			 * a new reference to it
421 			 */
422 			mutex_unlock(&prange->lock);
423 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
424 				 prange->svms, prange->start, prange->last);
425 
426 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
427 			return true;
428 		}
429 
430 	} else {
431 		mutex_unlock(&prange->lock);
432 	}
433 
434 	/* We need a new svm_bo. Spin-loop to wait for concurrent
435 	 * svm_range_bo_release to finish removing this range from
436 	 * its range list. After this, it is safe to reuse the
437 	 * svm_bo pointer and svm_bo_list head.
438 	 */
439 	while (!list_empty_careful(&prange->svm_bo_list))
440 		;
441 
442 	return false;
443 }
444 
445 static struct svm_range_bo *svm_range_bo_new(void)
446 {
447 	struct svm_range_bo *svm_bo;
448 
449 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
450 	if (!svm_bo)
451 		return NULL;
452 
453 	kref_init(&svm_bo->kref);
454 	INIT_LIST_HEAD(&svm_bo->range_list);
455 	spin_lock_init(&svm_bo->list_lock);
456 
457 	return svm_bo;
458 }
459 
460 int
461 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
462 			bool clear)
463 {
464 	struct amdgpu_bo_param bp;
465 	struct svm_range_bo *svm_bo;
466 	struct amdgpu_bo_user *ubo;
467 	struct amdgpu_bo *bo;
468 	struct kfd_process *p;
469 	struct mm_struct *mm;
470 	int r;
471 
472 	p = container_of(prange->svms, struct kfd_process, svms);
473 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
474 		 prange->start, prange->last);
475 
476 	if (svm_range_validate_svm_bo(adev, prange))
477 		return 0;
478 
479 	svm_bo = svm_range_bo_new();
480 	if (!svm_bo) {
481 		pr_debug("failed to alloc svm bo\n");
482 		return -ENOMEM;
483 	}
484 	mm = get_task_mm(p->lead_thread);
485 	if (!mm) {
486 		pr_debug("failed to get mm\n");
487 		kfree(svm_bo);
488 		return -ESRCH;
489 	}
490 	svm_bo->svms = prange->svms;
491 	svm_bo->eviction_fence =
492 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
493 					   mm,
494 					   svm_bo);
495 	mmput(mm);
496 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
497 	svm_bo->evicting = 0;
498 	memset(&bp, 0, sizeof(bp));
499 	bp.size = prange->npages * PAGE_SIZE;
500 	bp.byte_align = PAGE_SIZE;
501 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
502 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
503 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
504 	bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO;
505 	bp.type = ttm_bo_type_device;
506 	bp.resv = NULL;
507 
508 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
509 	if (r) {
510 		pr_debug("failed %d to create bo\n", r);
511 		goto create_bo_failed;
512 	}
513 	bo = &ubo->bo;
514 	r = amdgpu_bo_reserve(bo, true);
515 	if (r) {
516 		pr_debug("failed %d to reserve bo\n", r);
517 		goto reserve_bo_failed;
518 	}
519 
520 	r = dma_resv_reserve_shared(bo->tbo.base.resv, 1);
521 	if (r) {
522 		pr_debug("failed %d to reserve bo\n", r);
523 		amdgpu_bo_unreserve(bo);
524 		goto reserve_bo_failed;
525 	}
526 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
527 
528 	amdgpu_bo_unreserve(bo);
529 
530 	svm_bo->bo = bo;
531 	prange->svm_bo = svm_bo;
532 	prange->ttm_res = bo->tbo.resource;
533 	prange->offset = 0;
534 
535 	spin_lock(&svm_bo->list_lock);
536 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
537 	spin_unlock(&svm_bo->list_lock);
538 
539 	return 0;
540 
541 reserve_bo_failed:
542 	amdgpu_bo_unref(&bo);
543 create_bo_failed:
544 	dma_fence_put(&svm_bo->eviction_fence->base);
545 	kfree(svm_bo);
546 	prange->ttm_res = NULL;
547 
548 	return r;
549 }
550 
551 void svm_range_vram_node_free(struct svm_range *prange)
552 {
553 	svm_range_bo_unref(prange->svm_bo);
554 	prange->ttm_res = NULL;
555 }
556 
557 struct amdgpu_device *
558 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
559 {
560 	struct kfd_process_device *pdd;
561 	struct kfd_process *p;
562 	int32_t gpu_idx;
563 
564 	p = container_of(prange->svms, struct kfd_process, svms);
565 
566 	gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
567 	if (gpu_idx < 0) {
568 		pr_debug("failed to get device by id 0x%x\n", gpu_id);
569 		return NULL;
570 	}
571 	pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
572 	if (!pdd) {
573 		pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
574 		return NULL;
575 	}
576 
577 	return (struct amdgpu_device *)pdd->dev->kgd;
578 }
579 
580 struct kfd_process_device *
581 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
582 {
583 	struct kfd_process *p;
584 	int32_t gpu_idx, gpuid;
585 	int r;
586 
587 	p = container_of(prange->svms, struct kfd_process, svms);
588 
589 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
590 	if (r) {
591 		pr_debug("failed to get device id by adev %p\n", adev);
592 		return NULL;
593 	}
594 
595 	return kfd_process_device_from_gpuidx(p, gpu_idx);
596 }
597 
598 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
599 {
600 	struct ttm_operation_ctx ctx = { false, false };
601 
602 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
603 
604 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
605 }
606 
607 static int
608 svm_range_check_attr(struct kfd_process *p,
609 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
610 {
611 	uint32_t i;
612 
613 	for (i = 0; i < nattr; i++) {
614 		uint32_t val = attrs[i].value;
615 		int gpuidx = MAX_GPU_INSTANCE;
616 
617 		switch (attrs[i].type) {
618 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
619 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
620 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
621 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
622 			break;
623 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
624 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
625 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
626 			break;
627 		case KFD_IOCTL_SVM_ATTR_ACCESS:
628 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
629 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
630 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
631 			break;
632 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
633 			break;
634 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
635 			break;
636 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
637 			break;
638 		default:
639 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
640 			return -EINVAL;
641 		}
642 
643 		if (gpuidx < 0) {
644 			pr_debug("no GPU 0x%x found\n", val);
645 			return -EINVAL;
646 		} else if (gpuidx < MAX_GPU_INSTANCE &&
647 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
648 			pr_debug("GPU 0x%x not supported\n", val);
649 			return -EINVAL;
650 		}
651 	}
652 
653 	return 0;
654 }
655 
656 static void
657 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
658 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
659 {
660 	uint32_t i;
661 	int gpuidx;
662 
663 	for (i = 0; i < nattr; i++) {
664 		switch (attrs[i].type) {
665 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
666 			prange->preferred_loc = attrs[i].value;
667 			break;
668 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
669 			prange->prefetch_loc = attrs[i].value;
670 			break;
671 		case KFD_IOCTL_SVM_ATTR_ACCESS:
672 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
673 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
674 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
675 							       attrs[i].value);
676 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
677 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
678 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
679 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
680 				bitmap_set(prange->bitmap_access, gpuidx, 1);
681 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
682 			} else {
683 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
684 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
685 			}
686 			break;
687 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
688 			prange->flags |= attrs[i].value;
689 			break;
690 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
691 			prange->flags &= ~attrs[i].value;
692 			break;
693 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
694 			prange->granularity = attrs[i].value;
695 			break;
696 		default:
697 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
698 		}
699 	}
700 }
701 
702 /**
703  * svm_range_debug_dump - print all range information from svms
704  * @svms: svm range list header
705  *
706  * debug output svm range start, end, prefetch location from svms
707  * interval tree and link list
708  *
709  * Context: The caller must hold svms->lock
710  */
711 static void svm_range_debug_dump(struct svm_range_list *svms)
712 {
713 	struct interval_tree_node *node;
714 	struct svm_range *prange;
715 
716 	pr_debug("dump svms 0x%p list\n", svms);
717 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
718 
719 	list_for_each_entry(prange, &svms->list, list) {
720 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
721 			 prange, prange->start, prange->npages,
722 			 prange->start + prange->npages - 1,
723 			 prange->actual_loc);
724 	}
725 
726 	pr_debug("dump svms 0x%p interval tree\n", svms);
727 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
728 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
729 	while (node) {
730 		prange = container_of(node, struct svm_range, it_node);
731 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
732 			 prange, prange->start, prange->npages,
733 			 prange->start + prange->npages - 1,
734 			 prange->actual_loc);
735 		node = interval_tree_iter_next(node, 0, ~0ULL);
736 	}
737 }
738 
739 static bool
740 svm_range_is_same_attrs(struct svm_range *old, struct svm_range *new)
741 {
742 	return (old->prefetch_loc == new->prefetch_loc &&
743 		old->flags == new->flags &&
744 		old->granularity == new->granularity);
745 }
746 
747 static int
748 svm_range_split_array(void *ppnew, void *ppold, size_t size,
749 		      uint64_t old_start, uint64_t old_n,
750 		      uint64_t new_start, uint64_t new_n)
751 {
752 	unsigned char *new, *old, *pold;
753 	uint64_t d;
754 
755 	if (!ppold)
756 		return 0;
757 	pold = *(unsigned char **)ppold;
758 	if (!pold)
759 		return 0;
760 
761 	new = kvmalloc_array(new_n, size, GFP_KERNEL);
762 	if (!new)
763 		return -ENOMEM;
764 
765 	d = (new_start - old_start) * size;
766 	memcpy(new, pold + d, new_n * size);
767 
768 	old = kvmalloc_array(old_n, size, GFP_KERNEL);
769 	if (!old) {
770 		kvfree(new);
771 		return -ENOMEM;
772 	}
773 
774 	d = (new_start == old_start) ? new_n * size : 0;
775 	memcpy(old, pold + d, old_n * size);
776 
777 	kvfree(pold);
778 	*(void **)ppold = old;
779 	*(void **)ppnew = new;
780 
781 	return 0;
782 }
783 
784 static int
785 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
786 		      uint64_t start, uint64_t last)
787 {
788 	uint64_t npages = last - start + 1;
789 	int i, r;
790 
791 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
792 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
793 					  sizeof(*old->dma_addr[i]), old->start,
794 					  npages, new->start, new->npages);
795 		if (r)
796 			return r;
797 	}
798 
799 	return 0;
800 }
801 
802 static int
803 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
804 		      uint64_t start, uint64_t last)
805 {
806 	uint64_t npages = last - start + 1;
807 
808 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
809 		 new->svms, new, new->start, start, last);
810 
811 	if (new->start == old->start) {
812 		new->offset = old->offset;
813 		old->offset += new->npages;
814 	} else {
815 		new->offset = old->offset + npages;
816 	}
817 
818 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
819 	new->ttm_res = old->ttm_res;
820 
821 	spin_lock(&new->svm_bo->list_lock);
822 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
823 	spin_unlock(&new->svm_bo->list_lock);
824 
825 	return 0;
826 }
827 
828 /**
829  * svm_range_split_adjust - split range and adjust
830  *
831  * @new: new range
832  * @old: the old range
833  * @start: the old range adjust to start address in pages
834  * @last: the old range adjust to last address in pages
835  *
836  * Copy system memory dma_addr or vram ttm_res in old range to new
837  * range from new_start up to size new->npages, the remaining old range is from
838  * start to last
839  *
840  * Return:
841  * 0 - OK, -ENOMEM - out of memory
842  */
843 static int
844 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
845 		      uint64_t start, uint64_t last)
846 {
847 	int r;
848 
849 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
850 		 new->svms, new->start, old->start, old->last, start, last);
851 
852 	if (new->start < old->start ||
853 	    new->last > old->last) {
854 		WARN_ONCE(1, "invalid new range start or last\n");
855 		return -EINVAL;
856 	}
857 
858 	r = svm_range_split_pages(new, old, start, last);
859 	if (r)
860 		return r;
861 
862 	if (old->actual_loc && old->ttm_res) {
863 		r = svm_range_split_nodes(new, old, start, last);
864 		if (r)
865 			return r;
866 	}
867 
868 	old->npages = last - start + 1;
869 	old->start = start;
870 	old->last = last;
871 	new->flags = old->flags;
872 	new->preferred_loc = old->preferred_loc;
873 	new->prefetch_loc = old->prefetch_loc;
874 	new->actual_loc = old->actual_loc;
875 	new->granularity = old->granularity;
876 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
877 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
878 
879 	return 0;
880 }
881 
882 /**
883  * svm_range_split - split a range in 2 ranges
884  *
885  * @prange: the svm range to split
886  * @start: the remaining range start address in pages
887  * @last: the remaining range last address in pages
888  * @new: the result new range generated
889  *
890  * Two cases only:
891  * case 1: if start == prange->start
892  *         prange ==> prange[start, last]
893  *         new range [last + 1, prange->last]
894  *
895  * case 2: if last == prange->last
896  *         prange ==> prange[start, last]
897  *         new range [prange->start, start - 1]
898  *
899  * Return:
900  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
901  */
902 static int
903 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
904 		struct svm_range **new)
905 {
906 	uint64_t old_start = prange->start;
907 	uint64_t old_last = prange->last;
908 	struct svm_range_list *svms;
909 	int r = 0;
910 
911 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
912 		 old_start, old_last, start, last);
913 
914 	if (old_start != start && old_last != last)
915 		return -EINVAL;
916 	if (start < old_start || last > old_last)
917 		return -EINVAL;
918 
919 	svms = prange->svms;
920 	if (old_start == start)
921 		*new = svm_range_new(svms, last + 1, old_last);
922 	else
923 		*new = svm_range_new(svms, old_start, start - 1);
924 	if (!*new)
925 		return -ENOMEM;
926 
927 	r = svm_range_split_adjust(*new, prange, start, last);
928 	if (r) {
929 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
930 			 r, old_start, old_last, start, last);
931 		svm_range_free(*new);
932 		*new = NULL;
933 	}
934 
935 	return r;
936 }
937 
938 static int
939 svm_range_split_tail(struct svm_range *prange, struct svm_range *new,
940 		     uint64_t new_last, struct list_head *insert_list)
941 {
942 	struct svm_range *tail;
943 	int r = svm_range_split(prange, prange->start, new_last, &tail);
944 
945 	if (!r)
946 		list_add(&tail->insert_list, insert_list);
947 	return r;
948 }
949 
950 static int
951 svm_range_split_head(struct svm_range *prange, struct svm_range *new,
952 		     uint64_t new_start, struct list_head *insert_list)
953 {
954 	struct svm_range *head;
955 	int r = svm_range_split(prange, new_start, prange->last, &head);
956 
957 	if (!r)
958 		list_add(&head->insert_list, insert_list);
959 	return r;
960 }
961 
962 static void
963 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
964 		    struct svm_range *pchild, enum svm_work_list_ops op)
965 {
966 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
967 		 pchild, pchild->start, pchild->last, prange, op);
968 
969 	pchild->work_item.mm = mm;
970 	pchild->work_item.op = op;
971 	list_add_tail(&pchild->child_list, &prange->child_list);
972 }
973 
974 /**
975  * svm_range_split_by_granularity - collect ranges within granularity boundary
976  *
977  * @p: the process with svms list
978  * @mm: mm structure
979  * @addr: the vm fault address in pages, to split the prange
980  * @parent: parent range if prange is from child list
981  * @prange: prange to split
982  *
983  * Trims @prange to be a single aligned block of prange->granularity if
984  * possible. The head and tail are added to the child_list in @parent.
985  *
986  * Context: caller must hold mmap_read_lock and prange->lock
987  *
988  * Return:
989  * 0 - OK, otherwise error code
990  */
991 int
992 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
993 			       unsigned long addr, struct svm_range *parent,
994 			       struct svm_range *prange)
995 {
996 	struct svm_range *head, *tail;
997 	unsigned long start, last, size;
998 	int r;
999 
1000 	/* Align splited range start and size to granularity size, then a single
1001 	 * PTE will be used for whole range, this reduces the number of PTE
1002 	 * updated and the L1 TLB space used for translation.
1003 	 */
1004 	size = 1UL << prange->granularity;
1005 	start = ALIGN_DOWN(addr, size);
1006 	last = ALIGN(addr + 1, size) - 1;
1007 
1008 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1009 		 prange->svms, prange->start, prange->last, start, last, size);
1010 
1011 	if (start > prange->start) {
1012 		r = svm_range_split(prange, start, prange->last, &head);
1013 		if (r)
1014 			return r;
1015 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1016 	}
1017 
1018 	if (last < prange->last) {
1019 		r = svm_range_split(prange, prange->start, last, &tail);
1020 		if (r)
1021 			return r;
1022 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1023 	}
1024 
1025 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1026 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1027 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1028 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1029 			 prange, prange->start, prange->last,
1030 			 SVM_OP_ADD_RANGE_AND_MAP);
1031 	}
1032 	return 0;
1033 }
1034 
1035 static uint64_t
1036 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1037 			int domain)
1038 {
1039 	struct amdgpu_device *bo_adev;
1040 	uint32_t flags = prange->flags;
1041 	uint32_t mapping_flags = 0;
1042 	uint64_t pte_flags;
1043 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1044 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1045 
1046 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1047 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1048 
1049 	switch (adev->asic_type) {
1050 	case CHIP_ARCTURUS:
1051 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1052 			if (bo_adev == adev) {
1053 				mapping_flags |= coherent ?
1054 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1055 			} else {
1056 				mapping_flags |= coherent ?
1057 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1058 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1059 					snoop = true;
1060 			}
1061 		} else {
1062 			mapping_flags |= coherent ?
1063 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1064 		}
1065 		break;
1066 	case CHIP_ALDEBARAN:
1067 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1068 			if (bo_adev == adev) {
1069 				mapping_flags |= coherent ?
1070 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1071 				if (adev->gmc.xgmi.connected_to_cpu)
1072 					snoop = true;
1073 			} else {
1074 				mapping_flags |= coherent ?
1075 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1076 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1077 					snoop = true;
1078 			}
1079 		} else {
1080 			mapping_flags |= coherent ?
1081 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1082 		}
1083 		break;
1084 	default:
1085 		mapping_flags |= coherent ?
1086 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1087 	}
1088 
1089 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1090 
1091 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1092 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1093 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1094 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1095 
1096 	pte_flags = AMDGPU_PTE_VALID;
1097 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1098 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1099 
1100 	pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1101 	return pte_flags;
1102 }
1103 
1104 static int
1105 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1106 			 uint64_t start, uint64_t last,
1107 			 struct dma_fence **fence)
1108 {
1109 	uint64_t init_pte_value = 0;
1110 
1111 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1112 
1113 	return amdgpu_vm_bo_update_mapping(adev, adev, vm, false, true, NULL,
1114 					   start, last, init_pte_value, 0,
1115 					   NULL, NULL, fence, NULL);
1116 }
1117 
1118 static int
1119 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1120 			  unsigned long last)
1121 {
1122 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1123 	struct kfd_process_device *pdd;
1124 	struct dma_fence *fence = NULL;
1125 	struct amdgpu_device *adev;
1126 	struct kfd_process *p;
1127 	uint32_t gpuidx;
1128 	int r = 0;
1129 
1130 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1131 		  MAX_GPU_INSTANCE);
1132 	p = container_of(prange->svms, struct kfd_process, svms);
1133 
1134 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1135 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1136 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1137 		if (!pdd) {
1138 			pr_debug("failed to find device idx %d\n", gpuidx);
1139 			return -EINVAL;
1140 		}
1141 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1142 
1143 		r = svm_range_unmap_from_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1144 					     start, last, &fence);
1145 		if (r)
1146 			break;
1147 
1148 		if (fence) {
1149 			r = dma_fence_wait(fence, false);
1150 			dma_fence_put(fence);
1151 			fence = NULL;
1152 			if (r)
1153 				break;
1154 		}
1155 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1156 					p->pasid, TLB_FLUSH_HEAVYWEIGHT);
1157 	}
1158 
1159 	return r;
1160 }
1161 
1162 static int
1163 svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1164 		     struct svm_range *prange, unsigned long offset,
1165 		     unsigned long npages, bool readonly, dma_addr_t *dma_addr,
1166 		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
1167 {
1168 	struct amdgpu_bo_va bo_va;
1169 	bool table_freed = false;
1170 	uint64_t pte_flags;
1171 	unsigned long last_start;
1172 	int last_domain;
1173 	int r = 0;
1174 	int64_t i, j;
1175 
1176 	last_start = prange->start + offset;
1177 
1178 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1179 		 last_start, last_start + npages - 1, readonly);
1180 
1181 	if (prange->svm_bo && prange->ttm_res)
1182 		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
1183 
1184 	for (i = offset; i < offset + npages; i++) {
1185 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1186 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1187 
1188 		/* Collect all pages in the same address range and memory domain
1189 		 * that can be mapped with a single call to update mapping.
1190 		 */
1191 		if (i < offset + npages - 1 &&
1192 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1193 			continue;
1194 
1195 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1196 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1197 
1198 		pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1199 		if (readonly)
1200 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1201 
1202 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1203 			 prange->svms, last_start, prange->start + i,
1204 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1205 			 pte_flags);
1206 
1207 		r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false,
1208 						NULL, last_start,
1209 						prange->start + i, pte_flags,
1210 						last_start - prange->start,
1211 						NULL, dma_addr,
1212 						&vm->last_update,
1213 						&table_freed);
1214 
1215 		for (j = last_start - prange->start; j <= i; j++)
1216 			dma_addr[j] |= last_domain;
1217 
1218 		if (r) {
1219 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1220 			goto out;
1221 		}
1222 		last_start = prange->start + i + 1;
1223 	}
1224 
1225 	r = amdgpu_vm_update_pdes(adev, vm, false);
1226 	if (r) {
1227 		pr_debug("failed %d to update directories 0x%lx\n", r,
1228 			 prange->start);
1229 		goto out;
1230 	}
1231 
1232 	if (fence)
1233 		*fence = dma_fence_get(vm->last_update);
1234 
1235 	if (table_freed) {
1236 		struct kfd_process *p;
1237 
1238 		p = container_of(prange->svms, struct kfd_process, svms);
1239 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1240 						p->pasid, TLB_FLUSH_LEGACY);
1241 	}
1242 out:
1243 	return r;
1244 }
1245 
1246 static int
1247 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1248 		      unsigned long npages, bool readonly,
1249 		      unsigned long *bitmap, bool wait)
1250 {
1251 	struct kfd_process_device *pdd;
1252 	struct amdgpu_device *bo_adev;
1253 	struct amdgpu_device *adev;
1254 	struct kfd_process *p;
1255 	struct dma_fence *fence = NULL;
1256 	uint32_t gpuidx;
1257 	int r = 0;
1258 
1259 	if (prange->svm_bo && prange->ttm_res)
1260 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1261 	else
1262 		bo_adev = NULL;
1263 
1264 	p = container_of(prange->svms, struct kfd_process, svms);
1265 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1266 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1267 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1268 		if (!pdd) {
1269 			pr_debug("failed to find device idx %d\n", gpuidx);
1270 			return -EINVAL;
1271 		}
1272 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1273 
1274 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1275 		if (IS_ERR(pdd))
1276 			return -EINVAL;
1277 
1278 		if (bo_adev && adev != bo_adev &&
1279 		    !amdgpu_xgmi_same_hive(adev, bo_adev)) {
1280 			pr_debug("cannot map to device idx %d\n", gpuidx);
1281 			continue;
1282 		}
1283 
1284 		r = svm_range_map_to_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1285 					 prange, offset, npages, readonly,
1286 					 prange->dma_addr[gpuidx],
1287 					 bo_adev, wait ? &fence : NULL);
1288 		if (r)
1289 			break;
1290 
1291 		if (fence) {
1292 			r = dma_fence_wait(fence, false);
1293 			dma_fence_put(fence);
1294 			fence = NULL;
1295 			if (r) {
1296 				pr_debug("failed %d to dma fence wait\n", r);
1297 				break;
1298 			}
1299 		}
1300 	}
1301 
1302 	return r;
1303 }
1304 
1305 struct svm_validate_context {
1306 	struct kfd_process *process;
1307 	struct svm_range *prange;
1308 	bool intr;
1309 	unsigned long bitmap[MAX_GPU_INSTANCE];
1310 	struct ttm_validate_buffer tv[MAX_GPU_INSTANCE+1];
1311 	struct list_head validate_list;
1312 	struct ww_acquire_ctx ticket;
1313 };
1314 
1315 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1316 {
1317 	struct kfd_process_device *pdd;
1318 	struct amdgpu_device *adev;
1319 	struct amdgpu_vm *vm;
1320 	uint32_t gpuidx;
1321 	int r;
1322 
1323 	INIT_LIST_HEAD(&ctx->validate_list);
1324 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1325 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1326 		if (!pdd) {
1327 			pr_debug("failed to find device idx %d\n", gpuidx);
1328 			return -EINVAL;
1329 		}
1330 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1331 		vm = drm_priv_to_vm(pdd->drm_priv);
1332 
1333 		ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1334 		ctx->tv[gpuidx].num_shared = 4;
1335 		list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1336 	}
1337 	if (ctx->prange->svm_bo && ctx->prange->ttm_res) {
1338 		ctx->tv[MAX_GPU_INSTANCE].bo = &ctx->prange->svm_bo->bo->tbo;
1339 		ctx->tv[MAX_GPU_INSTANCE].num_shared = 1;
1340 		list_add(&ctx->tv[MAX_GPU_INSTANCE].head, &ctx->validate_list);
1341 	}
1342 
1343 	r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1344 				   ctx->intr, NULL);
1345 	if (r) {
1346 		pr_debug("failed %d to reserve bo\n", r);
1347 		return r;
1348 	}
1349 
1350 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1351 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1352 		if (!pdd) {
1353 			pr_debug("failed to find device idx %d\n", gpuidx);
1354 			r = -EINVAL;
1355 			goto unreserve_out;
1356 		}
1357 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1358 
1359 		r = amdgpu_vm_validate_pt_bos(adev, drm_priv_to_vm(pdd->drm_priv),
1360 					      svm_range_bo_validate, NULL);
1361 		if (r) {
1362 			pr_debug("failed %d validate pt bos\n", r);
1363 			goto unreserve_out;
1364 		}
1365 	}
1366 
1367 	return 0;
1368 
1369 unreserve_out:
1370 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1371 	return r;
1372 }
1373 
1374 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1375 {
1376 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1377 }
1378 
1379 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1380 {
1381 	struct kfd_process_device *pdd;
1382 	struct amdgpu_device *adev;
1383 
1384 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1385 	adev = (struct amdgpu_device *)pdd->dev->kgd;
1386 
1387 	return SVM_ADEV_PGMAP_OWNER(adev);
1388 }
1389 
1390 /*
1391  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1392  *
1393  * To prevent concurrent destruction or change of range attributes, the
1394  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1395  * because that would block concurrent evictions and lead to deadlocks. To
1396  * serialize concurrent migrations or validations of the same range, the
1397  * prange->migrate_mutex must be held.
1398  *
1399  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1400  * eviction fence.
1401  *
1402  * The following sequence ensures race-free validation and GPU mapping:
1403  *
1404  * 1. Reserve page table (and SVM BO if range is in VRAM)
1405  * 2. hmm_range_fault to get page addresses (if system memory)
1406  * 3. DMA-map pages (if system memory)
1407  * 4-a. Take notifier lock
1408  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1409  * 4-c. Check that the range was not split or otherwise invalidated
1410  * 4-d. Update GPU page table
1411  * 4.e. Release notifier lock
1412  * 5. Release page table (and SVM BO) reservation
1413  */
1414 static int svm_range_validate_and_map(struct mm_struct *mm,
1415 				      struct svm_range *prange,
1416 				      int32_t gpuidx, bool intr, bool wait)
1417 {
1418 	struct svm_validate_context ctx;
1419 	unsigned long start, end, addr;
1420 	struct kfd_process *p;
1421 	void *owner;
1422 	int32_t idx;
1423 	int r = 0;
1424 
1425 	ctx.process = container_of(prange->svms, struct kfd_process, svms);
1426 	ctx.prange = prange;
1427 	ctx.intr = intr;
1428 
1429 	if (gpuidx < MAX_GPU_INSTANCE) {
1430 		bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1431 		bitmap_set(ctx.bitmap, gpuidx, 1);
1432 	} else if (ctx.process->xnack_enabled) {
1433 		bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1434 
1435 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1436 		 * GPU, which has ACCESS attribute to the range, create mapping
1437 		 * on that GPU.
1438 		 */
1439 		if (prange->actual_loc) {
1440 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1441 							prange->actual_loc);
1442 			if (gpuidx < 0) {
1443 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1444 					 prange->actual_loc);
1445 				return -EINVAL;
1446 			}
1447 			if (test_bit(gpuidx, prange->bitmap_access))
1448 				bitmap_set(ctx.bitmap, gpuidx, 1);
1449 		}
1450 	} else {
1451 		bitmap_or(ctx.bitmap, prange->bitmap_access,
1452 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1453 	}
1454 
1455 	if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE))
1456 		return 0;
1457 
1458 	if (prange->actual_loc && !prange->ttm_res) {
1459 		/* This should never happen. actual_loc gets set by
1460 		 * svm_migrate_ram_to_vram after allocating a BO.
1461 		 */
1462 		WARN(1, "VRAM BO missing during validation\n");
1463 		return -EINVAL;
1464 	}
1465 
1466 	svm_range_reserve_bos(&ctx);
1467 
1468 	p = container_of(prange->svms, struct kfd_process, svms);
1469 	owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1470 						MAX_GPU_INSTANCE));
1471 	for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1472 		if (kfd_svm_page_owner(p, idx) != owner) {
1473 			owner = NULL;
1474 			break;
1475 		}
1476 	}
1477 
1478 	start = prange->start << PAGE_SHIFT;
1479 	end = (prange->last + 1) << PAGE_SHIFT;
1480 	for (addr = start; addr < end && !r; ) {
1481 		struct hmm_range *hmm_range;
1482 		struct vm_area_struct *vma;
1483 		unsigned long next;
1484 		unsigned long offset;
1485 		unsigned long npages;
1486 		bool readonly;
1487 
1488 		vma = find_vma(mm, addr);
1489 		if (!vma || addr < vma->vm_start) {
1490 			r = -EFAULT;
1491 			goto unreserve_out;
1492 		}
1493 		readonly = !(vma->vm_flags & VM_WRITE);
1494 
1495 		next = min(vma->vm_end, end);
1496 		npages = (next - addr) >> PAGE_SHIFT;
1497 		r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1498 					       addr, npages, &hmm_range,
1499 					       readonly, true, owner);
1500 		if (r) {
1501 			pr_debug("failed %d to get svm range pages\n", r);
1502 			goto unreserve_out;
1503 		}
1504 
1505 		offset = (addr - start) >> PAGE_SHIFT;
1506 		r = svm_range_dma_map(prange, ctx.bitmap, offset, npages,
1507 				      hmm_range->hmm_pfns);
1508 		if (r) {
1509 			pr_debug("failed %d to dma map range\n", r);
1510 			goto unreserve_out;
1511 		}
1512 
1513 		svm_range_lock(prange);
1514 		if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1515 			pr_debug("hmm update the range, need validate again\n");
1516 			r = -EAGAIN;
1517 			goto unlock_out;
1518 		}
1519 		if (!list_empty(&prange->child_list)) {
1520 			pr_debug("range split by unmap in parallel, validate again\n");
1521 			r = -EAGAIN;
1522 			goto unlock_out;
1523 		}
1524 
1525 		r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1526 					  ctx.bitmap, wait);
1527 
1528 unlock_out:
1529 		svm_range_unlock(prange);
1530 
1531 		addr = next;
1532 	}
1533 
1534 	if (addr == end)
1535 		prange->validated_once = true;
1536 
1537 unreserve_out:
1538 	svm_range_unreserve_bos(&ctx);
1539 
1540 	if (!r)
1541 		prange->validate_timestamp = ktime_to_us(ktime_get());
1542 
1543 	return r;
1544 }
1545 
1546 /**
1547  * svm_range_list_lock_and_flush_work - flush pending deferred work
1548  *
1549  * @svms: the svm range list
1550  * @mm: the mm structure
1551  *
1552  * Context: Returns with mmap write lock held, pending deferred work flushed
1553  *
1554  */
1555 static void
1556 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1557 				   struct mm_struct *mm)
1558 {
1559 retry_flush_work:
1560 	flush_work(&svms->deferred_list_work);
1561 	mmap_write_lock(mm);
1562 
1563 	if (list_empty(&svms->deferred_range_list))
1564 		return;
1565 	mmap_write_unlock(mm);
1566 	pr_debug("retry flush\n");
1567 	goto retry_flush_work;
1568 }
1569 
1570 static void svm_range_restore_work(struct work_struct *work)
1571 {
1572 	struct delayed_work *dwork = to_delayed_work(work);
1573 	struct amdkfd_process_info *process_info;
1574 	struct svm_range_list *svms;
1575 	struct svm_range *prange;
1576 	struct kfd_process *p;
1577 	struct mm_struct *mm;
1578 	int evicted_ranges;
1579 	int invalid;
1580 	int r;
1581 
1582 	svms = container_of(dwork, struct svm_range_list, restore_work);
1583 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1584 	if (!evicted_ranges)
1585 		return;
1586 
1587 	pr_debug("restore svm ranges\n");
1588 
1589 	/* kfd_process_notifier_release destroys this worker thread. So during
1590 	 * the lifetime of this thread, kfd_process and mm will be valid.
1591 	 */
1592 	p = container_of(svms, struct kfd_process, svms);
1593 	process_info = p->kgd_process_info;
1594 	mm = p->mm;
1595 	if (!mm)
1596 		return;
1597 
1598 	mutex_lock(&process_info->lock);
1599 	svm_range_list_lock_and_flush_work(svms, mm);
1600 	mutex_lock(&svms->lock);
1601 
1602 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1603 
1604 	list_for_each_entry(prange, &svms->list, list) {
1605 		invalid = atomic_read(&prange->invalid);
1606 		if (!invalid)
1607 			continue;
1608 
1609 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1610 			 prange->svms, prange, prange->start, prange->last,
1611 			 invalid);
1612 
1613 		/*
1614 		 * If range is migrating, wait for migration is done.
1615 		 */
1616 		mutex_lock(&prange->migrate_mutex);
1617 
1618 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1619 					       false, true);
1620 		if (r)
1621 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1622 				 prange->start);
1623 
1624 		mutex_unlock(&prange->migrate_mutex);
1625 		if (r)
1626 			goto out_reschedule;
1627 
1628 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1629 			goto out_reschedule;
1630 	}
1631 
1632 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1633 	    evicted_ranges)
1634 		goto out_reschedule;
1635 
1636 	evicted_ranges = 0;
1637 
1638 	r = kgd2kfd_resume_mm(mm);
1639 	if (r) {
1640 		/* No recovery from this failure. Probably the CP is
1641 		 * hanging. No point trying again.
1642 		 */
1643 		pr_debug("failed %d to resume KFD\n", r);
1644 	}
1645 
1646 	pr_debug("restore svm ranges successfully\n");
1647 
1648 out_reschedule:
1649 	mutex_unlock(&svms->lock);
1650 	mmap_write_unlock(mm);
1651 	mutex_unlock(&process_info->lock);
1652 
1653 	/* If validation failed, reschedule another attempt */
1654 	if (evicted_ranges) {
1655 		pr_debug("reschedule to restore svm range\n");
1656 		schedule_delayed_work(&svms->restore_work,
1657 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1658 	}
1659 }
1660 
1661 /**
1662  * svm_range_evict - evict svm range
1663  *
1664  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1665  * return to let CPU evict the buffer and proceed CPU pagetable update.
1666  *
1667  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1668  * If invalidation happens while restore work is running, restore work will
1669  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1670  * the queues.
1671  */
1672 static int
1673 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1674 		unsigned long start, unsigned long last)
1675 {
1676 	struct svm_range_list *svms = prange->svms;
1677 	struct svm_range *pchild;
1678 	struct kfd_process *p;
1679 	int r = 0;
1680 
1681 	p = container_of(svms, struct kfd_process, svms);
1682 
1683 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1684 		 svms, prange->start, prange->last, start, last);
1685 
1686 	if (!p->xnack_enabled) {
1687 		int evicted_ranges;
1688 
1689 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1690 			mutex_lock_nested(&pchild->lock, 1);
1691 			if (pchild->start <= last && pchild->last >= start) {
1692 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1693 					 pchild->start, pchild->last);
1694 				atomic_inc(&pchild->invalid);
1695 			}
1696 			mutex_unlock(&pchild->lock);
1697 		}
1698 
1699 		if (prange->start <= last && prange->last >= start)
1700 			atomic_inc(&prange->invalid);
1701 
1702 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1703 		if (evicted_ranges != 1)
1704 			return r;
1705 
1706 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1707 			 prange->svms, prange->start, prange->last);
1708 
1709 		/* First eviction, stop the queues */
1710 		r = kgd2kfd_quiesce_mm(mm);
1711 		if (r)
1712 			pr_debug("failed to quiesce KFD\n");
1713 
1714 		pr_debug("schedule to restore svm %p ranges\n", svms);
1715 		schedule_delayed_work(&svms->restore_work,
1716 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1717 	} else {
1718 		unsigned long s, l;
1719 
1720 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1721 			 prange->svms, start, last);
1722 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1723 			mutex_lock_nested(&pchild->lock, 1);
1724 			s = max(start, pchild->start);
1725 			l = min(last, pchild->last);
1726 			if (l >= s)
1727 				svm_range_unmap_from_gpus(pchild, s, l);
1728 			mutex_unlock(&pchild->lock);
1729 		}
1730 		s = max(start, prange->start);
1731 		l = min(last, prange->last);
1732 		if (l >= s)
1733 			svm_range_unmap_from_gpus(prange, s, l);
1734 	}
1735 
1736 	return r;
1737 }
1738 
1739 static struct svm_range *svm_range_clone(struct svm_range *old)
1740 {
1741 	struct svm_range *new;
1742 
1743 	new = svm_range_new(old->svms, old->start, old->last);
1744 	if (!new)
1745 		return NULL;
1746 
1747 	if (old->svm_bo) {
1748 		new->ttm_res = old->ttm_res;
1749 		new->offset = old->offset;
1750 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1751 		spin_lock(&new->svm_bo->list_lock);
1752 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1753 		spin_unlock(&new->svm_bo->list_lock);
1754 	}
1755 	new->flags = old->flags;
1756 	new->preferred_loc = old->preferred_loc;
1757 	new->prefetch_loc = old->prefetch_loc;
1758 	new->actual_loc = old->actual_loc;
1759 	new->granularity = old->granularity;
1760 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1761 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1762 
1763 	return new;
1764 }
1765 
1766 /**
1767  * svm_range_handle_overlap - split overlap ranges
1768  * @svms: svm range list header
1769  * @new: range added with this attributes
1770  * @start: range added start address, in pages
1771  * @last: range last address, in pages
1772  * @update_list: output, the ranges attributes are updated. For set_attr, this
1773  *               will do validation and map to GPUs. For unmap, this will be
1774  *               removed and unmap from GPUs
1775  * @insert_list: output, the ranges will be inserted into svms, attributes are
1776  *               not changes. For set_attr, this will add into svms.
1777  * @remove_list:output, the ranges will be removed from svms
1778  * @left: the remaining range after overlap, For set_attr, this will be added
1779  *        as new range.
1780  *
1781  * Total have 5 overlap cases.
1782  *
1783  * This function handles overlap of an address interval with existing
1784  * struct svm_ranges for applying new attributes. This may require
1785  * splitting existing struct svm_ranges. All changes should be applied to
1786  * the range_list and interval tree transactionally. If any split operation
1787  * fails, the entire update fails. Therefore the existing overlapping
1788  * svm_ranges are cloned and the original svm_ranges left unchanged. If the
1789  * transaction succeeds, the modified clones are added and the originals
1790  * freed. Otherwise the clones are removed and the old svm_ranges remain.
1791  *
1792  * Context: The caller must hold svms->lock
1793  */
1794 static int
1795 svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
1796 			 unsigned long start, unsigned long last,
1797 			 struct list_head *update_list,
1798 			 struct list_head *insert_list,
1799 			 struct list_head *remove_list,
1800 			 unsigned long *left)
1801 {
1802 	struct interval_tree_node *node;
1803 	struct svm_range *prange;
1804 	struct svm_range *tmp;
1805 	int r = 0;
1806 
1807 	INIT_LIST_HEAD(update_list);
1808 	INIT_LIST_HEAD(insert_list);
1809 	INIT_LIST_HEAD(remove_list);
1810 
1811 	node = interval_tree_iter_first(&svms->objects, start, last);
1812 	while (node) {
1813 		struct interval_tree_node *next;
1814 		struct svm_range *old;
1815 		unsigned long next_start;
1816 
1817 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
1818 			 node->last);
1819 
1820 		old = container_of(node, struct svm_range, it_node);
1821 		next = interval_tree_iter_next(node, start, last);
1822 		next_start = min(node->last, last) + 1;
1823 
1824 		if (node->start < start || node->last > last) {
1825 			/* node intersects the updated range, clone+split it */
1826 			prange = svm_range_clone(old);
1827 			if (!prange) {
1828 				r = -ENOMEM;
1829 				goto out;
1830 			}
1831 
1832 			list_add(&old->remove_list, remove_list);
1833 			list_add(&prange->insert_list, insert_list);
1834 
1835 			if (node->start < start) {
1836 				pr_debug("change old range start\n");
1837 				r = svm_range_split_head(prange, new, start,
1838 							 insert_list);
1839 				if (r)
1840 					goto out;
1841 			}
1842 			if (node->last > last) {
1843 				pr_debug("change old range last\n");
1844 				r = svm_range_split_tail(prange, new, last,
1845 							 insert_list);
1846 				if (r)
1847 					goto out;
1848 			}
1849 		} else {
1850 			/* The node is contained within start..last,
1851 			 * just update it
1852 			 */
1853 			prange = old;
1854 		}
1855 
1856 		if (!svm_range_is_same_attrs(prange, new))
1857 			list_add(&prange->update_list, update_list);
1858 
1859 		/* insert a new node if needed */
1860 		if (node->start > start) {
1861 			prange = svm_range_new(prange->svms, start,
1862 					       node->start - 1);
1863 			if (!prange) {
1864 				r = -ENOMEM;
1865 				goto out;
1866 			}
1867 
1868 			list_add(&prange->insert_list, insert_list);
1869 			list_add(&prange->update_list, update_list);
1870 		}
1871 
1872 		node = next;
1873 		start = next_start;
1874 	}
1875 
1876 	if (left && start <= last)
1877 		*left = last - start + 1;
1878 
1879 out:
1880 	if (r)
1881 		list_for_each_entry_safe(prange, tmp, insert_list, insert_list)
1882 			svm_range_free(prange);
1883 
1884 	return r;
1885 }
1886 
1887 static void
1888 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
1889 					    struct svm_range *prange)
1890 {
1891 	unsigned long start;
1892 	unsigned long last;
1893 
1894 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
1895 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
1896 
1897 	if (prange->start == start && prange->last == last)
1898 		return;
1899 
1900 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1901 		  prange->svms, prange, start, last, prange->start,
1902 		  prange->last);
1903 
1904 	if (start != 0 && last != 0) {
1905 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
1906 		svm_range_remove_notifier(prange);
1907 	}
1908 	prange->it_node.start = prange->start;
1909 	prange->it_node.last = prange->last;
1910 
1911 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
1912 	svm_range_add_notifier_locked(mm, prange);
1913 }
1914 
1915 static void
1916 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange)
1917 {
1918 	struct mm_struct *mm = prange->work_item.mm;
1919 
1920 	switch (prange->work_item.op) {
1921 	case SVM_OP_NULL:
1922 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1923 			 svms, prange, prange->start, prange->last);
1924 		break;
1925 	case SVM_OP_UNMAP_RANGE:
1926 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1927 			 svms, prange, prange->start, prange->last);
1928 		svm_range_unlink(prange);
1929 		svm_range_remove_notifier(prange);
1930 		svm_range_free(prange);
1931 		break;
1932 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
1933 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1934 			 svms, prange, prange->start, prange->last);
1935 		svm_range_update_notifier_and_interval_tree(mm, prange);
1936 		break;
1937 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
1938 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1939 			 svms, prange, prange->start, prange->last);
1940 		svm_range_update_notifier_and_interval_tree(mm, prange);
1941 		/* TODO: implement deferred validation and mapping */
1942 		break;
1943 	case SVM_OP_ADD_RANGE:
1944 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
1945 			 prange->start, prange->last);
1946 		svm_range_add_to_svms(prange);
1947 		svm_range_add_notifier_locked(mm, prange);
1948 		break;
1949 	case SVM_OP_ADD_RANGE_AND_MAP:
1950 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
1951 			 prange, prange->start, prange->last);
1952 		svm_range_add_to_svms(prange);
1953 		svm_range_add_notifier_locked(mm, prange);
1954 		/* TODO: implement deferred validation and mapping */
1955 		break;
1956 	default:
1957 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
1958 			 prange->work_item.op);
1959 	}
1960 }
1961 
1962 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
1963 {
1964 	struct kfd_process_device *pdd;
1965 	struct amdgpu_device *adev;
1966 	struct kfd_process *p;
1967 	uint32_t i;
1968 
1969 	p = container_of(svms, struct kfd_process, svms);
1970 
1971 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
1972 		pdd = p->pdds[i];
1973 		if (!pdd)
1974 			continue;
1975 
1976 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
1977 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1978 
1979 		amdgpu_ih_wait_on_checkpoint_process(adev, &adev->irq.ih1);
1980 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
1981 	}
1982 }
1983 
1984 static void svm_range_deferred_list_work(struct work_struct *work)
1985 {
1986 	struct svm_range_list *svms;
1987 	struct svm_range *prange;
1988 	struct mm_struct *mm;
1989 
1990 	svms = container_of(work, struct svm_range_list, deferred_list_work);
1991 	pr_debug("enter svms 0x%p\n", svms);
1992 
1993 	spin_lock(&svms->deferred_list_lock);
1994 	while (!list_empty(&svms->deferred_range_list)) {
1995 		prange = list_first_entry(&svms->deferred_range_list,
1996 					  struct svm_range, deferred_list);
1997 		spin_unlock(&svms->deferred_list_lock);
1998 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
1999 			 prange->start, prange->last, prange->work_item.op);
2000 
2001 		/* Make sure no stale retry fault coming after range is freed */
2002 		if (prange->work_item.op == SVM_OP_UNMAP_RANGE)
2003 			svm_range_drain_retry_fault(prange->svms);
2004 
2005 		mm = prange->work_item.mm;
2006 		mmap_write_lock(mm);
2007 		mutex_lock(&svms->lock);
2008 
2009 		/* Remove from deferred_list must be inside mmap write lock,
2010 		 * otherwise, svm_range_list_lock_and_flush_work may hold mmap
2011 		 * write lock, and continue because deferred_list is empty, then
2012 		 * deferred_list handle is blocked by mmap write lock.
2013 		 */
2014 		spin_lock(&svms->deferred_list_lock);
2015 		list_del_init(&prange->deferred_list);
2016 		spin_unlock(&svms->deferred_list_lock);
2017 
2018 		mutex_lock(&prange->migrate_mutex);
2019 		while (!list_empty(&prange->child_list)) {
2020 			struct svm_range *pchild;
2021 
2022 			pchild = list_first_entry(&prange->child_list,
2023 						struct svm_range, child_list);
2024 			pr_debug("child prange 0x%p op %d\n", pchild,
2025 				 pchild->work_item.op);
2026 			list_del_init(&pchild->child_list);
2027 			svm_range_handle_list_op(svms, pchild);
2028 		}
2029 		mutex_unlock(&prange->migrate_mutex);
2030 
2031 		svm_range_handle_list_op(svms, prange);
2032 		mutex_unlock(&svms->lock);
2033 		mmap_write_unlock(mm);
2034 
2035 		spin_lock(&svms->deferred_list_lock);
2036 	}
2037 	spin_unlock(&svms->deferred_list_lock);
2038 
2039 	pr_debug("exit svms 0x%p\n", svms);
2040 }
2041 
2042 void
2043 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2044 			struct mm_struct *mm, enum svm_work_list_ops op)
2045 {
2046 	spin_lock(&svms->deferred_list_lock);
2047 	/* if prange is on the deferred list */
2048 	if (!list_empty(&prange->deferred_list)) {
2049 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2050 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2051 		if (op != SVM_OP_NULL &&
2052 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2053 			prange->work_item.op = op;
2054 	} else {
2055 		prange->work_item.op = op;
2056 		prange->work_item.mm = mm;
2057 		list_add_tail(&prange->deferred_list,
2058 			      &prange->svms->deferred_range_list);
2059 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2060 			 prange, prange->start, prange->last, op);
2061 	}
2062 	spin_unlock(&svms->deferred_list_lock);
2063 }
2064 
2065 void schedule_deferred_list_work(struct svm_range_list *svms)
2066 {
2067 	spin_lock(&svms->deferred_list_lock);
2068 	if (!list_empty(&svms->deferred_range_list))
2069 		schedule_work(&svms->deferred_list_work);
2070 	spin_unlock(&svms->deferred_list_lock);
2071 }
2072 
2073 static void
2074 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2075 		      struct svm_range *prange, unsigned long start,
2076 		      unsigned long last)
2077 {
2078 	struct svm_range *head;
2079 	struct svm_range *tail;
2080 
2081 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2082 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2083 			 prange->start, prange->last);
2084 		return;
2085 	}
2086 	if (start > prange->last || last < prange->start)
2087 		return;
2088 
2089 	head = tail = prange;
2090 	if (start > prange->start)
2091 		svm_range_split(prange, prange->start, start - 1, &tail);
2092 	if (last < tail->last)
2093 		svm_range_split(tail, last + 1, tail->last, &head);
2094 
2095 	if (head != prange && tail != prange) {
2096 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2097 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2098 	} else if (tail != prange) {
2099 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2100 	} else if (head != prange) {
2101 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2102 	} else if (parent != prange) {
2103 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2104 	}
2105 }
2106 
2107 static void
2108 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2109 			 unsigned long start, unsigned long last)
2110 {
2111 	struct svm_range_list *svms;
2112 	struct svm_range *pchild;
2113 	struct kfd_process *p;
2114 	unsigned long s, l;
2115 	bool unmap_parent;
2116 
2117 	p = kfd_lookup_process_by_mm(mm);
2118 	if (!p)
2119 		return;
2120 	svms = &p->svms;
2121 
2122 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2123 		 prange, prange->start, prange->last, start, last);
2124 
2125 	unmap_parent = start <= prange->start && last >= prange->last;
2126 
2127 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2128 		mutex_lock_nested(&pchild->lock, 1);
2129 		s = max(start, pchild->start);
2130 		l = min(last, pchild->last);
2131 		if (l >= s)
2132 			svm_range_unmap_from_gpus(pchild, s, l);
2133 		svm_range_unmap_split(mm, prange, pchild, start, last);
2134 		mutex_unlock(&pchild->lock);
2135 	}
2136 	s = max(start, prange->start);
2137 	l = min(last, prange->last);
2138 	if (l >= s)
2139 		svm_range_unmap_from_gpus(prange, s, l);
2140 	svm_range_unmap_split(mm, prange, prange, start, last);
2141 
2142 	if (unmap_parent)
2143 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2144 	else
2145 		svm_range_add_list_work(svms, prange, mm,
2146 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2147 	schedule_deferred_list_work(svms);
2148 
2149 	kfd_unref_process(p);
2150 }
2151 
2152 /**
2153  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2154  *
2155  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2156  * is from migration, or CPU page invalidation callback.
2157  *
2158  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2159  * work thread, and split prange if only part of prange is unmapped.
2160  *
2161  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2162  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2163  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2164  * update GPU mapping to recover.
2165  *
2166  * Context: mmap lock, notifier_invalidate_start lock are held
2167  *          for invalidate event, prange lock is held if this is from migration
2168  */
2169 static bool
2170 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2171 				    const struct mmu_notifier_range *range,
2172 				    unsigned long cur_seq)
2173 {
2174 	struct svm_range *prange;
2175 	unsigned long start;
2176 	unsigned long last;
2177 
2178 	if (range->event == MMU_NOTIFY_RELEASE)
2179 		return true;
2180 
2181 	start = mni->interval_tree.start;
2182 	last = mni->interval_tree.last;
2183 	start = (start > range->start ? start : range->start) >> PAGE_SHIFT;
2184 	last = (last < (range->end - 1) ? last : range->end - 1) >> PAGE_SHIFT;
2185 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2186 		 start, last, range->start >> PAGE_SHIFT,
2187 		 (range->end - 1) >> PAGE_SHIFT,
2188 		 mni->interval_tree.start >> PAGE_SHIFT,
2189 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2190 
2191 	prange = container_of(mni, struct svm_range, notifier);
2192 
2193 	svm_range_lock(prange);
2194 	mmu_interval_set_seq(mni, cur_seq);
2195 
2196 	switch (range->event) {
2197 	case MMU_NOTIFY_UNMAP:
2198 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2199 		break;
2200 	default:
2201 		svm_range_evict(prange, mni->mm, start, last);
2202 		break;
2203 	}
2204 
2205 	svm_range_unlock(prange);
2206 
2207 	return true;
2208 }
2209 
2210 /**
2211  * svm_range_from_addr - find svm range from fault address
2212  * @svms: svm range list header
2213  * @addr: address to search range interval tree, in pages
2214  * @parent: parent range if range is on child list
2215  *
2216  * Context: The caller must hold svms->lock
2217  *
2218  * Return: the svm_range found or NULL
2219  */
2220 struct svm_range *
2221 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2222 		    struct svm_range **parent)
2223 {
2224 	struct interval_tree_node *node;
2225 	struct svm_range *prange;
2226 	struct svm_range *pchild;
2227 
2228 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2229 	if (!node)
2230 		return NULL;
2231 
2232 	prange = container_of(node, struct svm_range, it_node);
2233 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2234 		 addr, prange->start, prange->last, node->start, node->last);
2235 
2236 	if (addr >= prange->start && addr <= prange->last) {
2237 		if (parent)
2238 			*parent = prange;
2239 		return prange;
2240 	}
2241 	list_for_each_entry(pchild, &prange->child_list, child_list)
2242 		if (addr >= pchild->start && addr <= pchild->last) {
2243 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2244 				 addr, pchild->start, pchild->last);
2245 			if (parent)
2246 				*parent = prange;
2247 			return pchild;
2248 		}
2249 
2250 	return NULL;
2251 }
2252 
2253 /* svm_range_best_restore_location - decide the best fault restore location
2254  * @prange: svm range structure
2255  * @adev: the GPU on which vm fault happened
2256  *
2257  * This is only called when xnack is on, to decide the best location to restore
2258  * the range mapping after GPU vm fault. Caller uses the best location to do
2259  * migration if actual loc is not best location, then update GPU page table
2260  * mapping to the best location.
2261  *
2262  * If vm fault gpu is range preferred loc, the best_loc is preferred loc.
2263  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2264  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2265  *    if range actual loc is cpu, best_loc is cpu
2266  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2267  *    range actual loc.
2268  * Otherwise, GPU no access, best_loc is -1.
2269  *
2270  * Return:
2271  * -1 means vm fault GPU no access
2272  * 0 for CPU or GPU id
2273  */
2274 static int32_t
2275 svm_range_best_restore_location(struct svm_range *prange,
2276 				struct amdgpu_device *adev,
2277 				int32_t *gpuidx)
2278 {
2279 	struct amdgpu_device *bo_adev;
2280 	struct kfd_process *p;
2281 	uint32_t gpuid;
2282 	int r;
2283 
2284 	p = container_of(prange->svms, struct kfd_process, svms);
2285 
2286 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, gpuidx);
2287 	if (r < 0) {
2288 		pr_debug("failed to get gpuid from kgd\n");
2289 		return -1;
2290 	}
2291 
2292 	if (prange->preferred_loc == gpuid)
2293 		return prange->preferred_loc;
2294 
2295 	if (test_bit(*gpuidx, prange->bitmap_access))
2296 		return gpuid;
2297 
2298 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2299 		if (!prange->actual_loc)
2300 			return 0;
2301 
2302 		bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2303 		if (amdgpu_xgmi_same_hive(adev, bo_adev))
2304 			return prange->actual_loc;
2305 		else
2306 			return 0;
2307 	}
2308 
2309 	return -1;
2310 }
2311 static int
2312 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2313 				unsigned long *start, unsigned long *last)
2314 {
2315 	struct vm_area_struct *vma;
2316 	struct interval_tree_node *node;
2317 	unsigned long start_limit, end_limit;
2318 
2319 	vma = find_vma(p->mm, addr << PAGE_SHIFT);
2320 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2321 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2322 		return -EFAULT;
2323 	}
2324 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2325 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2326 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2327 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2328 	/* First range that starts after the fault address */
2329 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2330 	if (node) {
2331 		end_limit = min(end_limit, node->start);
2332 		/* Last range that ends before the fault address */
2333 		node = container_of(rb_prev(&node->rb),
2334 				    struct interval_tree_node, rb);
2335 	} else {
2336 		/* Last range must end before addr because
2337 		 * there was no range after addr
2338 		 */
2339 		node = container_of(rb_last(&p->svms.objects.rb_root),
2340 				    struct interval_tree_node, rb);
2341 	}
2342 	if (node) {
2343 		if (node->last >= addr) {
2344 			WARN(1, "Overlap with prev node and page fault addr\n");
2345 			return -EFAULT;
2346 		}
2347 		start_limit = max(start_limit, node->last + 1);
2348 	}
2349 
2350 	*start = start_limit;
2351 	*last = end_limit - 1;
2352 
2353 	pr_debug("vma start: 0x%lx start: 0x%lx vma end: 0x%lx last: 0x%lx\n",
2354 		  vma->vm_start >> PAGE_SHIFT, *start,
2355 		  vma->vm_end >> PAGE_SHIFT, *last);
2356 
2357 	return 0;
2358 
2359 }
2360 static struct
2361 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2362 						struct kfd_process *p,
2363 						struct mm_struct *mm,
2364 						int64_t addr)
2365 {
2366 	struct svm_range *prange = NULL;
2367 	unsigned long start, last;
2368 	uint32_t gpuid, gpuidx;
2369 
2370 	if (svm_range_get_range_boundaries(p, addr, &start, &last))
2371 		return NULL;
2372 
2373 	prange = svm_range_new(&p->svms, start, last);
2374 	if (!prange) {
2375 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2376 		return NULL;
2377 	}
2378 	if (kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx)) {
2379 		pr_debug("failed to get gpuid from kgd\n");
2380 		svm_range_free(prange);
2381 		return NULL;
2382 	}
2383 
2384 	svm_range_add_to_svms(prange);
2385 	svm_range_add_notifier_locked(mm, prange);
2386 
2387 	return prange;
2388 }
2389 
2390 /* svm_range_skip_recover - decide if prange can be recovered
2391  * @prange: svm range structure
2392  *
2393  * GPU vm retry fault handle skip recover the range for cases:
2394  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2395  *    deferred list work will drain the stale fault before free the prange.
2396  * 2. prange is on deferred list to add interval notifier after split, or
2397  * 3. prange is child range, it is split from parent prange, recover later
2398  *    after interval notifier is added.
2399  *
2400  * Return: true to skip recover, false to recover
2401  */
2402 static bool svm_range_skip_recover(struct svm_range *prange)
2403 {
2404 	struct svm_range_list *svms = prange->svms;
2405 
2406 	spin_lock(&svms->deferred_list_lock);
2407 	if (list_empty(&prange->deferred_list) &&
2408 	    list_empty(&prange->child_list)) {
2409 		spin_unlock(&svms->deferred_list_lock);
2410 		return false;
2411 	}
2412 	spin_unlock(&svms->deferred_list_lock);
2413 
2414 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2415 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2416 			 svms, prange, prange->start, prange->last);
2417 		return true;
2418 	}
2419 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2420 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2421 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2422 			 svms, prange, prange->start, prange->last);
2423 		return true;
2424 	}
2425 	return false;
2426 }
2427 
2428 static void
2429 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2430 		      int32_t gpuidx)
2431 {
2432 	struct kfd_process_device *pdd;
2433 
2434 	/* fault is on different page of same range
2435 	 * or fault is skipped to recover later
2436 	 * or fault is on invalid virtual address
2437 	 */
2438 	if (gpuidx == MAX_GPU_INSTANCE) {
2439 		uint32_t gpuid;
2440 		int r;
2441 
2442 		r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx);
2443 		if (r < 0)
2444 			return;
2445 	}
2446 
2447 	/* fault is recovered
2448 	 * or fault cannot recover because GPU no access on the range
2449 	 */
2450 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2451 	if (pdd)
2452 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2453 }
2454 
2455 static bool
2456 svm_fault_allowed(struct mm_struct *mm, uint64_t addr, bool write_fault)
2457 {
2458 	unsigned long requested = VM_READ;
2459 	struct vm_area_struct *vma;
2460 
2461 	if (write_fault)
2462 		requested |= VM_WRITE;
2463 
2464 	vma = find_vma(mm, addr << PAGE_SHIFT);
2465 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2466 		pr_debug("address 0x%llx VMA is removed\n", addr);
2467 		return true;
2468 	}
2469 
2470 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2471 		vma->vm_flags);
2472 	return (vma->vm_flags & requested) == requested;
2473 }
2474 
2475 int
2476 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2477 			uint64_t addr, bool write_fault)
2478 {
2479 	struct mm_struct *mm = NULL;
2480 	struct svm_range_list *svms;
2481 	struct svm_range *prange;
2482 	struct kfd_process *p;
2483 	uint64_t timestamp;
2484 	int32_t best_loc;
2485 	int32_t gpuidx = MAX_GPU_INSTANCE;
2486 	bool write_locked = false;
2487 	int r = 0;
2488 
2489 	if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2490 		pr_debug("device does not support SVM\n");
2491 		return -EFAULT;
2492 	}
2493 
2494 	p = kfd_lookup_process_by_pasid(pasid);
2495 	if (!p) {
2496 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2497 		return -ESRCH;
2498 	}
2499 	if (!p->xnack_enabled) {
2500 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2501 		r = -EFAULT;
2502 		goto out;
2503 	}
2504 	svms = &p->svms;
2505 
2506 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2507 
2508 	mm = get_task_mm(p->lead_thread);
2509 	if (!mm) {
2510 		pr_debug("svms 0x%p failed to get mm\n", svms);
2511 		r = -ESRCH;
2512 		goto out;
2513 	}
2514 
2515 	mmap_read_lock(mm);
2516 retry_write_locked:
2517 	mutex_lock(&svms->lock);
2518 	prange = svm_range_from_addr(svms, addr, NULL);
2519 	if (!prange) {
2520 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2521 			 svms, addr);
2522 		if (!write_locked) {
2523 			/* Need the write lock to create new range with MMU notifier.
2524 			 * Also flush pending deferred work to make sure the interval
2525 			 * tree is up to date before we add a new range
2526 			 */
2527 			mutex_unlock(&svms->lock);
2528 			mmap_read_unlock(mm);
2529 			mmap_write_lock(mm);
2530 			write_locked = true;
2531 			goto retry_write_locked;
2532 		}
2533 		prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2534 		if (!prange) {
2535 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2536 				 svms, addr);
2537 			mmap_write_downgrade(mm);
2538 			r = -EFAULT;
2539 			goto out_unlock_svms;
2540 		}
2541 	}
2542 	if (write_locked)
2543 		mmap_write_downgrade(mm);
2544 
2545 	mutex_lock(&prange->migrate_mutex);
2546 
2547 	if (svm_range_skip_recover(prange)) {
2548 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2549 		goto out_unlock_range;
2550 	}
2551 
2552 	timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp;
2553 	/* skip duplicate vm fault on different pages of same range */
2554 	if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
2555 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2556 			 svms, prange->start, prange->last);
2557 		goto out_unlock_range;
2558 	}
2559 
2560 	if (!svm_fault_allowed(mm, addr, write_fault)) {
2561 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
2562 			write_fault ? "write" : "read");
2563 		r = -EPERM;
2564 		goto out_unlock_range;
2565 	}
2566 
2567 	best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2568 	if (best_loc == -1) {
2569 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2570 			 svms, prange->start, prange->last);
2571 		r = -EACCES;
2572 		goto out_unlock_range;
2573 	}
2574 
2575 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2576 		 svms, prange->start, prange->last, best_loc,
2577 		 prange->actual_loc);
2578 
2579 	if (prange->actual_loc != best_loc) {
2580 		if (best_loc) {
2581 			r = svm_migrate_to_vram(prange, best_loc, mm);
2582 			if (r) {
2583 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2584 					 r, addr);
2585 				/* Fallback to system memory if migration to
2586 				 * VRAM failed
2587 				 */
2588 				if (prange->actual_loc)
2589 					r = svm_migrate_vram_to_ram(prange, mm);
2590 				else
2591 					r = 0;
2592 			}
2593 		} else {
2594 			r = svm_migrate_vram_to_ram(prange, mm);
2595 		}
2596 		if (r) {
2597 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2598 				 r, svms, prange->start, prange->last);
2599 			goto out_unlock_range;
2600 		}
2601 	}
2602 
2603 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false);
2604 	if (r)
2605 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2606 			 r, svms, prange->start, prange->last);
2607 
2608 out_unlock_range:
2609 	mutex_unlock(&prange->migrate_mutex);
2610 out_unlock_svms:
2611 	mutex_unlock(&svms->lock);
2612 	mmap_read_unlock(mm);
2613 
2614 	svm_range_count_fault(adev, p, gpuidx);
2615 
2616 	mmput(mm);
2617 out:
2618 	kfd_unref_process(p);
2619 
2620 	if (r == -EAGAIN) {
2621 		pr_debug("recover vm fault later\n");
2622 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2623 		r = 0;
2624 	}
2625 	return r;
2626 }
2627 
2628 void svm_range_list_fini(struct kfd_process *p)
2629 {
2630 	struct svm_range *prange;
2631 	struct svm_range *next;
2632 
2633 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
2634 
2635 	/* Ensure list work is finished before process is destroyed */
2636 	flush_work(&p->svms.deferred_list_work);
2637 
2638 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
2639 		svm_range_unlink(prange);
2640 		svm_range_remove_notifier(prange);
2641 		svm_range_free(prange);
2642 	}
2643 
2644 	mutex_destroy(&p->svms.lock);
2645 
2646 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
2647 }
2648 
2649 int svm_range_list_init(struct kfd_process *p)
2650 {
2651 	struct svm_range_list *svms = &p->svms;
2652 	int i;
2653 
2654 	svms->objects = RB_ROOT_CACHED;
2655 	mutex_init(&svms->lock);
2656 	INIT_LIST_HEAD(&svms->list);
2657 	atomic_set(&svms->evicted_ranges, 0);
2658 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
2659 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
2660 	INIT_LIST_HEAD(&svms->deferred_range_list);
2661 	spin_lock_init(&svms->deferred_list_lock);
2662 
2663 	for (i = 0; i < p->n_pdds; i++)
2664 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
2665 			bitmap_set(svms->bitmap_supported, i, 1);
2666 
2667 	return 0;
2668 }
2669 
2670 /**
2671  * svm_range_is_valid - check if virtual address range is valid
2672  * @mm: current process mm_struct
2673  * @start: range start address, in pages
2674  * @size: range size, in pages
2675  *
2676  * Valid virtual address range means it belongs to one or more VMAs
2677  *
2678  * Context: Process context
2679  *
2680  * Return:
2681  *  true - valid svm range
2682  *  false - invalid svm range
2683  */
2684 static bool
2685 svm_range_is_valid(struct mm_struct *mm, uint64_t start, uint64_t size)
2686 {
2687 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
2688 	struct vm_area_struct *vma;
2689 	unsigned long end;
2690 
2691 	start <<= PAGE_SHIFT;
2692 	end = start + (size << PAGE_SHIFT);
2693 
2694 	do {
2695 		vma = find_vma(mm, start);
2696 		if (!vma || start < vma->vm_start ||
2697 		    (vma->vm_flags & device_vma))
2698 			return false;
2699 		start = min(end, vma->vm_end);
2700 	} while (start < end);
2701 
2702 	return true;
2703 }
2704 
2705 /**
2706  * svm_range_add - add svm range and handle overlap
2707  * @p: the range add to this process svms
2708  * @start: page size aligned
2709  * @size: page size aligned
2710  * @nattr: number of attributes
2711  * @attrs: array of attributes
2712  * @update_list: output, the ranges need validate and update GPU mapping
2713  * @insert_list: output, the ranges need insert to svms
2714  * @remove_list: output, the ranges are replaced and need remove from svms
2715  *
2716  * Check if the virtual address range has overlap with the registered ranges,
2717  * split the overlapped range, copy and adjust pages address and vram nodes in
2718  * old and new ranges.
2719  *
2720  * Context: Process context, caller must hold svms->lock
2721  *
2722  * Return:
2723  * 0 - OK, otherwise error code
2724  */
2725 static int
2726 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2727 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2728 	      struct list_head *update_list, struct list_head *insert_list,
2729 	      struct list_head *remove_list)
2730 {
2731 	uint64_t last = start + size - 1UL;
2732 	struct svm_range_list *svms;
2733 	struct svm_range new = {0};
2734 	struct svm_range *prange;
2735 	unsigned long left = 0;
2736 	int r = 0;
2737 
2738 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", &p->svms, start, last);
2739 
2740 	svm_range_apply_attrs(p, &new, nattr, attrs);
2741 
2742 	svms = &p->svms;
2743 
2744 	r = svm_range_handle_overlap(svms, &new, start, last, update_list,
2745 				     insert_list, remove_list, &left);
2746 	if (r)
2747 		return r;
2748 
2749 	if (left) {
2750 		prange = svm_range_new(svms, last - left + 1, last);
2751 		list_add(&prange->insert_list, insert_list);
2752 		list_add(&prange->update_list, update_list);
2753 	}
2754 
2755 	return 0;
2756 }
2757 
2758 /**
2759  * svm_range_best_prefetch_location - decide the best prefetch location
2760  * @prange: svm range structure
2761  *
2762  * For xnack off:
2763  * If range map to single GPU, the best prefetch location is prefetch_loc, which
2764  * can be CPU or GPU.
2765  *
2766  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
2767  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
2768  * the best prefetch location is always CPU, because GPU can not have coherent
2769  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
2770  *
2771  * For xnack on:
2772  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
2773  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
2774  *
2775  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
2776  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
2777  * prefetch location is always CPU.
2778  *
2779  * Context: Process context
2780  *
2781  * Return:
2782  * 0 for CPU or GPU id
2783  */
2784 static uint32_t
2785 svm_range_best_prefetch_location(struct svm_range *prange)
2786 {
2787 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
2788 	uint32_t best_loc = prange->prefetch_loc;
2789 	struct kfd_process_device *pdd;
2790 	struct amdgpu_device *bo_adev;
2791 	struct amdgpu_device *adev;
2792 	struct kfd_process *p;
2793 	uint32_t gpuidx;
2794 
2795 	p = container_of(prange->svms, struct kfd_process, svms);
2796 
2797 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
2798 		goto out;
2799 
2800 	bo_adev = svm_range_get_adev_by_id(prange, best_loc);
2801 	if (!bo_adev) {
2802 		WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
2803 		best_loc = 0;
2804 		goto out;
2805 	}
2806 
2807 	if (p->xnack_enabled)
2808 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
2809 	else
2810 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
2811 			  MAX_GPU_INSTANCE);
2812 
2813 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
2814 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2815 		if (!pdd) {
2816 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
2817 			continue;
2818 		}
2819 		adev = (struct amdgpu_device *)pdd->dev->kgd;
2820 
2821 		if (adev == bo_adev)
2822 			continue;
2823 
2824 		if (!amdgpu_xgmi_same_hive(adev, bo_adev)) {
2825 			best_loc = 0;
2826 			break;
2827 		}
2828 	}
2829 
2830 out:
2831 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
2832 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
2833 		 best_loc);
2834 
2835 	return best_loc;
2836 }
2837 
2838 /* FIXME: This is a workaround for page locking bug when some pages are
2839  * invalid during migration to VRAM
2840  */
2841 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
2842 			void *owner)
2843 {
2844 	struct hmm_range *hmm_range;
2845 	int r;
2846 
2847 	if (prange->validated_once)
2848 		return;
2849 
2850 	r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
2851 				       prange->start << PAGE_SHIFT,
2852 				       prange->npages, &hmm_range,
2853 				       false, true, owner);
2854 	if (!r) {
2855 		amdgpu_hmm_range_get_pages_done(hmm_range);
2856 		prange->validated_once = true;
2857 	}
2858 }
2859 
2860 /* svm_range_trigger_migration - start page migration if prefetch loc changed
2861  * @mm: current process mm_struct
2862  * @prange: svm range structure
2863  * @migrated: output, true if migration is triggered
2864  *
2865  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
2866  * from ram to vram.
2867  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
2868  * from vram to ram.
2869  *
2870  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
2871  * and restore work:
2872  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
2873  *    stops all queues, schedule restore work
2874  * 2. svm_range_restore_work wait for migration is done by
2875  *    a. svm_range_validate_vram takes prange->migrate_mutex
2876  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
2877  * 3. restore work update mappings of GPU, resume all queues.
2878  *
2879  * Context: Process context
2880  *
2881  * Return:
2882  * 0 - OK, otherwise - error code of migration
2883  */
2884 static int
2885 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
2886 			    bool *migrated)
2887 {
2888 	uint32_t best_loc;
2889 	int r = 0;
2890 
2891 	*migrated = false;
2892 	best_loc = svm_range_best_prefetch_location(prange);
2893 
2894 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
2895 	    best_loc == prange->actual_loc)
2896 		return 0;
2897 
2898 	if (!best_loc) {
2899 		r = svm_migrate_vram_to_ram(prange, mm);
2900 		*migrated = !r;
2901 		return r;
2902 	}
2903 
2904 	r = svm_migrate_to_vram(prange, best_loc, mm);
2905 	*migrated = !r;
2906 
2907 	return r;
2908 }
2909 
2910 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
2911 {
2912 	if (!fence)
2913 		return -EINVAL;
2914 
2915 	if (dma_fence_is_signaled(&fence->base))
2916 		return 0;
2917 
2918 	if (fence->svm_bo) {
2919 		WRITE_ONCE(fence->svm_bo->evicting, 1);
2920 		schedule_work(&fence->svm_bo->eviction_work);
2921 	}
2922 
2923 	return 0;
2924 }
2925 
2926 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
2927 {
2928 	struct svm_range_bo *svm_bo;
2929 	struct kfd_process *p;
2930 	struct mm_struct *mm;
2931 
2932 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
2933 	if (!svm_bo_ref_unless_zero(svm_bo))
2934 		return; /* svm_bo was freed while eviction was pending */
2935 
2936 	/* svm_range_bo_release destroys this worker thread. So during
2937 	 * the lifetime of this thread, kfd_process and mm will be valid.
2938 	 */
2939 	p = container_of(svm_bo->svms, struct kfd_process, svms);
2940 	mm = p->mm;
2941 	if (!mm)
2942 		return;
2943 
2944 	mmap_read_lock(mm);
2945 	spin_lock(&svm_bo->list_lock);
2946 	while (!list_empty(&svm_bo->range_list)) {
2947 		struct svm_range *prange =
2948 				list_first_entry(&svm_bo->range_list,
2949 						struct svm_range, svm_bo_list);
2950 		list_del_init(&prange->svm_bo_list);
2951 		spin_unlock(&svm_bo->list_lock);
2952 
2953 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
2954 			 prange->start, prange->last);
2955 
2956 		mutex_lock(&prange->migrate_mutex);
2957 		svm_migrate_vram_to_ram(prange, svm_bo->eviction_fence->mm);
2958 
2959 		mutex_lock(&prange->lock);
2960 		prange->svm_bo = NULL;
2961 		mutex_unlock(&prange->lock);
2962 
2963 		mutex_unlock(&prange->migrate_mutex);
2964 
2965 		spin_lock(&svm_bo->list_lock);
2966 	}
2967 	spin_unlock(&svm_bo->list_lock);
2968 	mmap_read_unlock(mm);
2969 
2970 	dma_fence_signal(&svm_bo->eviction_fence->base);
2971 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
2972 	 * has been called in svm_migrate_vram_to_ram
2973 	 */
2974 	WARN_ONCE(kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
2975 	svm_range_bo_unref(svm_bo);
2976 }
2977 
2978 static int
2979 svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
2980 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
2981 {
2982 	struct amdkfd_process_info *process_info = p->kgd_process_info;
2983 	struct mm_struct *mm = current->mm;
2984 	struct list_head update_list;
2985 	struct list_head insert_list;
2986 	struct list_head remove_list;
2987 	struct svm_range_list *svms;
2988 	struct svm_range *prange;
2989 	struct svm_range *next;
2990 	int r = 0;
2991 
2992 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
2993 		 p->pasid, &p->svms, start, start + size - 1, size);
2994 
2995 	r = svm_range_check_attr(p, nattr, attrs);
2996 	if (r)
2997 		return r;
2998 
2999 	svms = &p->svms;
3000 
3001 	mutex_lock(&process_info->lock);
3002 
3003 	svm_range_list_lock_and_flush_work(svms, mm);
3004 
3005 	if (!svm_range_is_valid(mm, start, size)) {
3006 		pr_debug("invalid range\n");
3007 		r = -EFAULT;
3008 		mmap_write_unlock(mm);
3009 		goto out;
3010 	}
3011 
3012 	mutex_lock(&svms->lock);
3013 
3014 	/* Add new range and split existing ranges as needed */
3015 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3016 			  &insert_list, &remove_list);
3017 	if (r) {
3018 		mutex_unlock(&svms->lock);
3019 		mmap_write_unlock(mm);
3020 		goto out;
3021 	}
3022 	/* Apply changes as a transaction */
3023 	list_for_each_entry_safe(prange, next, &insert_list, insert_list) {
3024 		svm_range_add_to_svms(prange);
3025 		svm_range_add_notifier_locked(mm, prange);
3026 	}
3027 	list_for_each_entry(prange, &update_list, update_list) {
3028 		svm_range_apply_attrs(p, prange, nattr, attrs);
3029 		/* TODO: unmap ranges from GPU that lost access */
3030 	}
3031 	list_for_each_entry_safe(prange, next, &remove_list,
3032 				remove_list) {
3033 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3034 			 prange->svms, prange, prange->start,
3035 			 prange->last);
3036 		svm_range_unlink(prange);
3037 		svm_range_remove_notifier(prange);
3038 		svm_range_free(prange);
3039 	}
3040 
3041 	mmap_write_downgrade(mm);
3042 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3043 	 * this fails we may be left with partially completed actions. There
3044 	 * is no clean way of rolling back to the previous state in such a
3045 	 * case because the rollback wouldn't be guaranteed to work either.
3046 	 */
3047 	list_for_each_entry(prange, &update_list, update_list) {
3048 		bool migrated;
3049 
3050 		mutex_lock(&prange->migrate_mutex);
3051 
3052 		r = svm_range_trigger_migration(mm, prange, &migrated);
3053 		if (r)
3054 			goto out_unlock_range;
3055 
3056 		if (migrated && !p->xnack_enabled) {
3057 			pr_debug("restore_work will update mappings of GPUs\n");
3058 			mutex_unlock(&prange->migrate_mutex);
3059 			continue;
3060 		}
3061 
3062 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3063 					       true, true);
3064 		if (r)
3065 			pr_debug("failed %d to map svm range\n", r);
3066 
3067 out_unlock_range:
3068 		mutex_unlock(&prange->migrate_mutex);
3069 		if (r)
3070 			break;
3071 	}
3072 
3073 	svm_range_debug_dump(svms);
3074 
3075 	mutex_unlock(&svms->lock);
3076 	mmap_read_unlock(mm);
3077 out:
3078 	mutex_unlock(&process_info->lock);
3079 
3080 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3081 		 &p->svms, start, start + size - 1, r);
3082 
3083 	return r;
3084 }
3085 
3086 static int
3087 svm_range_get_attr(struct kfd_process *p, uint64_t start, uint64_t size,
3088 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
3089 {
3090 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3091 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3092 	bool get_preferred_loc = false;
3093 	bool get_prefetch_loc = false;
3094 	bool get_granularity = false;
3095 	bool get_accessible = false;
3096 	bool get_flags = false;
3097 	uint64_t last = start + size - 1UL;
3098 	struct mm_struct *mm = current->mm;
3099 	uint8_t granularity = 0xff;
3100 	struct interval_tree_node *node;
3101 	struct svm_range_list *svms;
3102 	struct svm_range *prange;
3103 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3104 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3105 	uint32_t flags_and = 0xffffffff;
3106 	uint32_t flags_or = 0;
3107 	int gpuidx;
3108 	uint32_t i;
3109 
3110 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3111 		 start + size - 1, nattr);
3112 
3113 	/* Flush pending deferred work to avoid racing with deferred actions from
3114 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3115 	 * can still race with get_attr because we don't hold the mmap lock. But that
3116 	 * would be a race condition in the application anyway, and undefined
3117 	 * behaviour is acceptable in that case.
3118 	 */
3119 	flush_work(&p->svms.deferred_list_work);
3120 
3121 	mmap_read_lock(mm);
3122 	if (!svm_range_is_valid(mm, start, size)) {
3123 		pr_debug("invalid range\n");
3124 		mmap_read_unlock(mm);
3125 		return -EINVAL;
3126 	}
3127 	mmap_read_unlock(mm);
3128 
3129 	for (i = 0; i < nattr; i++) {
3130 		switch (attrs[i].type) {
3131 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3132 			get_preferred_loc = true;
3133 			break;
3134 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3135 			get_prefetch_loc = true;
3136 			break;
3137 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3138 			get_accessible = true;
3139 			break;
3140 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3141 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3142 			get_flags = true;
3143 			break;
3144 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3145 			get_granularity = true;
3146 			break;
3147 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3148 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3149 			fallthrough;
3150 		default:
3151 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3152 			return -EINVAL;
3153 		}
3154 	}
3155 
3156 	svms = &p->svms;
3157 
3158 	mutex_lock(&svms->lock);
3159 
3160 	node = interval_tree_iter_first(&svms->objects, start, last);
3161 	if (!node) {
3162 		pr_debug("range attrs not found return default values\n");
3163 		svm_range_set_default_attributes(&location, &prefetch_loc,
3164 						 &granularity, &flags_and);
3165 		flags_or = flags_and;
3166 		if (p->xnack_enabled)
3167 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3168 				    MAX_GPU_INSTANCE);
3169 		else
3170 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3171 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3172 		goto fill_values;
3173 	}
3174 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3175 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3176 
3177 	while (node) {
3178 		struct interval_tree_node *next;
3179 
3180 		prange = container_of(node, struct svm_range, it_node);
3181 		next = interval_tree_iter_next(node, start, last);
3182 
3183 		if (get_preferred_loc) {
3184 			if (prange->preferred_loc ==
3185 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3186 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3187 			     location != prange->preferred_loc)) {
3188 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3189 				get_preferred_loc = false;
3190 			} else {
3191 				location = prange->preferred_loc;
3192 			}
3193 		}
3194 		if (get_prefetch_loc) {
3195 			if (prange->prefetch_loc ==
3196 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3197 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3198 			     prefetch_loc != prange->prefetch_loc)) {
3199 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3200 				get_prefetch_loc = false;
3201 			} else {
3202 				prefetch_loc = prange->prefetch_loc;
3203 			}
3204 		}
3205 		if (get_accessible) {
3206 			bitmap_and(bitmap_access, bitmap_access,
3207 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3208 			bitmap_and(bitmap_aip, bitmap_aip,
3209 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3210 		}
3211 		if (get_flags) {
3212 			flags_and &= prange->flags;
3213 			flags_or |= prange->flags;
3214 		}
3215 
3216 		if (get_granularity && prange->granularity < granularity)
3217 			granularity = prange->granularity;
3218 
3219 		node = next;
3220 	}
3221 fill_values:
3222 	mutex_unlock(&svms->lock);
3223 
3224 	for (i = 0; i < nattr; i++) {
3225 		switch (attrs[i].type) {
3226 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3227 			attrs[i].value = location;
3228 			break;
3229 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3230 			attrs[i].value = prefetch_loc;
3231 			break;
3232 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3233 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3234 							       attrs[i].value);
3235 			if (gpuidx < 0) {
3236 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3237 				return -EINVAL;
3238 			}
3239 			if (test_bit(gpuidx, bitmap_access))
3240 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3241 			else if (test_bit(gpuidx, bitmap_aip))
3242 				attrs[i].type =
3243 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3244 			else
3245 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3246 			break;
3247 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3248 			attrs[i].value = flags_and;
3249 			break;
3250 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3251 			attrs[i].value = ~flags_or;
3252 			break;
3253 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3254 			attrs[i].value = (uint32_t)granularity;
3255 			break;
3256 		}
3257 	}
3258 
3259 	return 0;
3260 }
3261 
3262 int
3263 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3264 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3265 {
3266 	int r;
3267 
3268 	start >>= PAGE_SHIFT;
3269 	size >>= PAGE_SHIFT;
3270 
3271 	switch (op) {
3272 	case KFD_IOCTL_SVM_OP_SET_ATTR:
3273 		r = svm_range_set_attr(p, start, size, nattrs, attrs);
3274 		break;
3275 	case KFD_IOCTL_SVM_OP_GET_ATTR:
3276 		r = svm_range_get_attr(p, start, size, nattrs, attrs);
3277 		break;
3278 	default:
3279 		r = EINVAL;
3280 		break;
3281 	}
3282 
3283 	return r;
3284 }
3285