xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 464a5728)
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 int
122 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
123 		      unsigned long *hmm_pfns, uint32_t gpuidx)
124 {
125 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
126 	dma_addr_t *addr = prange->dma_addr[gpuidx];
127 	struct device *dev = adev->dev;
128 	struct page *page;
129 	int i, r;
130 
131 	if (!addr) {
132 		addr = kvmalloc_array(prange->npages, sizeof(*addr),
133 				      GFP_KERNEL | __GFP_ZERO);
134 		if (!addr)
135 			return -ENOMEM;
136 		prange->dma_addr[gpuidx] = addr;
137 	}
138 
139 	for (i = 0; i < prange->npages; i++) {
140 		if (WARN_ONCE(addr[i] && !dma_mapping_error(dev, addr[i]),
141 			      "leaking dma mapping\n"))
142 			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
143 
144 		page = hmm_pfn_to_page(hmm_pfns[i]);
145 		if (is_zone_device_page(page)) {
146 			struct amdgpu_device *bo_adev =
147 					amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
148 
149 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
150 				   bo_adev->vm_manager.vram_base_offset -
151 				   bo_adev->kfd.dev->pgmap.range.start;
152 			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
153 			pr_debug("vram address detected: 0x%llx\n", addr[i]);
154 			continue;
155 		}
156 		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
157 		r = dma_mapping_error(dev, addr[i]);
158 		if (r) {
159 			pr_debug("failed %d dma_map_page\n", r);
160 			return r;
161 		}
162 		pr_debug("dma mapping 0x%llx for page addr 0x%lx\n",
163 			 addr[i] >> PAGE_SHIFT, page_to_pfn(page));
164 	}
165 	return 0;
166 }
167 
168 static int
169 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
170 		  unsigned long *hmm_pfns)
171 {
172 	struct kfd_process *p;
173 	uint32_t gpuidx;
174 	int r;
175 
176 	p = container_of(prange->svms, struct kfd_process, svms);
177 
178 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
179 		struct kfd_process_device *pdd;
180 		struct amdgpu_device *adev;
181 
182 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
183 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
184 		if (!pdd) {
185 			pr_debug("failed to find device idx %d\n", gpuidx);
186 			return -EINVAL;
187 		}
188 		adev = (struct amdgpu_device *)pdd->dev->kgd;
189 
190 		r = svm_range_dma_map_dev(adev, prange, hmm_pfns, gpuidx);
191 		if (r)
192 			break;
193 	}
194 
195 	return r;
196 }
197 
198 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
199 			 unsigned long offset, unsigned long npages)
200 {
201 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
202 	int i;
203 
204 	if (!dma_addr)
205 		return;
206 
207 	for (i = offset; i < offset + npages; i++) {
208 		if (!dma_addr[i] || dma_mapping_error(dev, dma_addr[i]))
209 			continue;
210 		pr_debug("dma unmapping 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
211 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
212 		dma_addr[i] = 0;
213 	}
214 }
215 
216 void svm_range_free_dma_mappings(struct svm_range *prange)
217 {
218 	struct kfd_process_device *pdd;
219 	dma_addr_t *dma_addr;
220 	struct device *dev;
221 	struct kfd_process *p;
222 	uint32_t gpuidx;
223 
224 	p = container_of(prange->svms, struct kfd_process, svms);
225 
226 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
227 		dma_addr = prange->dma_addr[gpuidx];
228 		if (!dma_addr)
229 			continue;
230 
231 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
232 		if (!pdd) {
233 			pr_debug("failed to find device idx %d\n", gpuidx);
234 			continue;
235 		}
236 		dev = &pdd->dev->pdev->dev;
237 		svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
238 		kvfree(dma_addr);
239 		prange->dma_addr[gpuidx] = NULL;
240 	}
241 }
242 
243 static void svm_range_free(struct svm_range *prange)
244 {
245 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
246 		 prange->start, prange->last);
247 
248 	svm_range_vram_node_free(prange);
249 	svm_range_free_dma_mappings(prange);
250 	mutex_destroy(&prange->lock);
251 	mutex_destroy(&prange->migrate_mutex);
252 	kfree(prange);
253 }
254 
255 static void
256 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
257 				 uint8_t *granularity, uint32_t *flags)
258 {
259 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
260 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
261 	*granularity = 9;
262 	*flags =
263 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
264 }
265 
266 static struct
267 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
268 			 uint64_t last)
269 {
270 	uint64_t size = last - start + 1;
271 	struct svm_range *prange;
272 	struct kfd_process *p;
273 
274 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
275 	if (!prange)
276 		return NULL;
277 	prange->npages = size;
278 	prange->svms = svms;
279 	prange->start = start;
280 	prange->last = last;
281 	INIT_LIST_HEAD(&prange->list);
282 	INIT_LIST_HEAD(&prange->update_list);
283 	INIT_LIST_HEAD(&prange->remove_list);
284 	INIT_LIST_HEAD(&prange->insert_list);
285 	INIT_LIST_HEAD(&prange->svm_bo_list);
286 	INIT_LIST_HEAD(&prange->deferred_list);
287 	INIT_LIST_HEAD(&prange->child_list);
288 	atomic_set(&prange->invalid, 0);
289 	prange->validate_timestamp = 0;
290 	mutex_init(&prange->migrate_mutex);
291 	mutex_init(&prange->lock);
292 
293 	p = container_of(svms, struct kfd_process, svms);
294 	if (p->xnack_enabled)
295 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
296 			    MAX_GPU_INSTANCE);
297 
298 	svm_range_set_default_attributes(&prange->preferred_loc,
299 					 &prange->prefetch_loc,
300 					 &prange->granularity, &prange->flags);
301 
302 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
303 
304 	return prange;
305 }
306 
307 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
308 {
309 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
310 		return false;
311 
312 	return true;
313 }
314 
315 static void svm_range_bo_release(struct kref *kref)
316 {
317 	struct svm_range_bo *svm_bo;
318 
319 	svm_bo = container_of(kref, struct svm_range_bo, kref);
320 	spin_lock(&svm_bo->list_lock);
321 	while (!list_empty(&svm_bo->range_list)) {
322 		struct svm_range *prange =
323 				list_first_entry(&svm_bo->range_list,
324 						struct svm_range, svm_bo_list);
325 		/* list_del_init tells a concurrent svm_range_vram_node_new when
326 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
327 		 */
328 		list_del_init(&prange->svm_bo_list);
329 		spin_unlock(&svm_bo->list_lock);
330 
331 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
332 			 prange->start, prange->last);
333 		mutex_lock(&prange->lock);
334 		prange->svm_bo = NULL;
335 		mutex_unlock(&prange->lock);
336 
337 		spin_lock(&svm_bo->list_lock);
338 	}
339 	spin_unlock(&svm_bo->list_lock);
340 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
341 		/* We're not in the eviction worker.
342 		 * Signal the fence and synchronize with any
343 		 * pending eviction work.
344 		 */
345 		dma_fence_signal(&svm_bo->eviction_fence->base);
346 		cancel_work_sync(&svm_bo->eviction_work);
347 	}
348 	dma_fence_put(&svm_bo->eviction_fence->base);
349 	amdgpu_bo_unref(&svm_bo->bo);
350 	kfree(svm_bo);
351 }
352 
353 void svm_range_bo_unref(struct svm_range_bo *svm_bo)
354 {
355 	if (!svm_bo)
356 		return;
357 
358 	kref_put(&svm_bo->kref, svm_range_bo_release);
359 }
360 
361 static bool
362 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
363 {
364 	struct amdgpu_device *bo_adev;
365 
366 	mutex_lock(&prange->lock);
367 	if (!prange->svm_bo) {
368 		mutex_unlock(&prange->lock);
369 		return false;
370 	}
371 	if (prange->ttm_res) {
372 		/* We still have a reference, all is well */
373 		mutex_unlock(&prange->lock);
374 		return true;
375 	}
376 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
377 		/*
378 		 * Migrate from GPU to GPU, remove range from source bo_adev
379 		 * svm_bo range list, and return false to allocate svm_bo from
380 		 * destination adev.
381 		 */
382 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
383 		if (bo_adev != adev) {
384 			mutex_unlock(&prange->lock);
385 
386 			spin_lock(&prange->svm_bo->list_lock);
387 			list_del_init(&prange->svm_bo_list);
388 			spin_unlock(&prange->svm_bo->list_lock);
389 
390 			svm_range_bo_unref(prange->svm_bo);
391 			return false;
392 		}
393 		if (READ_ONCE(prange->svm_bo->evicting)) {
394 			struct dma_fence *f;
395 			struct svm_range_bo *svm_bo;
396 			/* The BO is getting evicted,
397 			 * we need to get a new one
398 			 */
399 			mutex_unlock(&prange->lock);
400 			svm_bo = prange->svm_bo;
401 			f = dma_fence_get(&svm_bo->eviction_fence->base);
402 			svm_range_bo_unref(prange->svm_bo);
403 			/* wait for the fence to avoid long spin-loop
404 			 * at list_empty_careful
405 			 */
406 			dma_fence_wait(f, false);
407 			dma_fence_put(f);
408 		} else {
409 			/* The BO was still around and we got
410 			 * a new reference to it
411 			 */
412 			mutex_unlock(&prange->lock);
413 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
414 				 prange->svms, prange->start, prange->last);
415 
416 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
417 			return true;
418 		}
419 
420 	} else {
421 		mutex_unlock(&prange->lock);
422 	}
423 
424 	/* We need a new svm_bo. Spin-loop to wait for concurrent
425 	 * svm_range_bo_release to finish removing this range from
426 	 * its range list. After this, it is safe to reuse the
427 	 * svm_bo pointer and svm_bo_list head.
428 	 */
429 	while (!list_empty_careful(&prange->svm_bo_list))
430 		;
431 
432 	return false;
433 }
434 
435 static struct svm_range_bo *svm_range_bo_new(void)
436 {
437 	struct svm_range_bo *svm_bo;
438 
439 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
440 	if (!svm_bo)
441 		return NULL;
442 
443 	kref_init(&svm_bo->kref);
444 	INIT_LIST_HEAD(&svm_bo->range_list);
445 	spin_lock_init(&svm_bo->list_lock);
446 
447 	return svm_bo;
448 }
449 
450 int
451 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
452 			bool clear)
453 {
454 	struct amdgpu_bo_param bp;
455 	struct svm_range_bo *svm_bo;
456 	struct amdgpu_bo_user *ubo;
457 	struct amdgpu_bo *bo;
458 	struct kfd_process *p;
459 	struct mm_struct *mm;
460 	int r;
461 
462 	p = container_of(prange->svms, struct kfd_process, svms);
463 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
464 		 prange->start, prange->last);
465 
466 	if (svm_range_validate_svm_bo(adev, prange))
467 		return 0;
468 
469 	svm_bo = svm_range_bo_new();
470 	if (!svm_bo) {
471 		pr_debug("failed to alloc svm bo\n");
472 		return -ENOMEM;
473 	}
474 	mm = get_task_mm(p->lead_thread);
475 	if (!mm) {
476 		pr_debug("failed to get mm\n");
477 		kfree(svm_bo);
478 		return -ESRCH;
479 	}
480 	svm_bo->svms = prange->svms;
481 	svm_bo->eviction_fence =
482 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
483 					   mm,
484 					   svm_bo);
485 	mmput(mm);
486 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
487 	svm_bo->evicting = 0;
488 	memset(&bp, 0, sizeof(bp));
489 	bp.size = prange->npages * PAGE_SIZE;
490 	bp.byte_align = PAGE_SIZE;
491 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
492 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
493 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
494 	bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO;
495 	bp.type = ttm_bo_type_device;
496 	bp.resv = NULL;
497 
498 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
499 	if (r) {
500 		pr_debug("failed %d to create bo\n", r);
501 		goto create_bo_failed;
502 	}
503 	bo = &ubo->bo;
504 	r = amdgpu_bo_reserve(bo, true);
505 	if (r) {
506 		pr_debug("failed %d to reserve bo\n", r);
507 		goto reserve_bo_failed;
508 	}
509 
510 	r = dma_resv_reserve_shared(bo->tbo.base.resv, 1);
511 	if (r) {
512 		pr_debug("failed %d to reserve bo\n", r);
513 		amdgpu_bo_unreserve(bo);
514 		goto reserve_bo_failed;
515 	}
516 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
517 
518 	amdgpu_bo_unreserve(bo);
519 
520 	svm_bo->bo = bo;
521 	prange->svm_bo = svm_bo;
522 	prange->ttm_res = bo->tbo.resource;
523 	prange->offset = 0;
524 
525 	spin_lock(&svm_bo->list_lock);
526 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
527 	spin_unlock(&svm_bo->list_lock);
528 
529 	return 0;
530 
531 reserve_bo_failed:
532 	amdgpu_bo_unref(&bo);
533 create_bo_failed:
534 	dma_fence_put(&svm_bo->eviction_fence->base);
535 	kfree(svm_bo);
536 	prange->ttm_res = NULL;
537 
538 	return r;
539 }
540 
541 void svm_range_vram_node_free(struct svm_range *prange)
542 {
543 	svm_range_bo_unref(prange->svm_bo);
544 	prange->ttm_res = NULL;
545 }
546 
547 struct amdgpu_device *
548 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
549 {
550 	struct kfd_process_device *pdd;
551 	struct kfd_process *p;
552 	int32_t gpu_idx;
553 
554 	p = container_of(prange->svms, struct kfd_process, svms);
555 
556 	gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
557 	if (gpu_idx < 0) {
558 		pr_debug("failed to get device by id 0x%x\n", gpu_id);
559 		return NULL;
560 	}
561 	pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
562 	if (!pdd) {
563 		pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
564 		return NULL;
565 	}
566 
567 	return (struct amdgpu_device *)pdd->dev->kgd;
568 }
569 
570 struct kfd_process_device *
571 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
572 {
573 	struct kfd_process *p;
574 	int32_t gpu_idx, gpuid;
575 	int r;
576 
577 	p = container_of(prange->svms, struct kfd_process, svms);
578 
579 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
580 	if (r) {
581 		pr_debug("failed to get device id by adev %p\n", adev);
582 		return NULL;
583 	}
584 
585 	return kfd_process_device_from_gpuidx(p, gpu_idx);
586 }
587 
588 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
589 {
590 	struct ttm_operation_ctx ctx = { false, false };
591 
592 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
593 
594 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
595 }
596 
597 static int
598 svm_range_check_attr(struct kfd_process *p,
599 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
600 {
601 	uint32_t i;
602 
603 	for (i = 0; i < nattr; i++) {
604 		uint32_t val = attrs[i].value;
605 		int gpuidx = MAX_GPU_INSTANCE;
606 
607 		switch (attrs[i].type) {
608 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
609 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
610 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
611 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
612 			break;
613 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
614 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
615 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
616 			break;
617 		case KFD_IOCTL_SVM_ATTR_ACCESS:
618 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
619 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
620 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
621 			break;
622 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
623 			break;
624 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
625 			break;
626 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
627 			break;
628 		default:
629 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
630 			return -EINVAL;
631 		}
632 
633 		if (gpuidx < 0) {
634 			pr_debug("no GPU 0x%x found\n", val);
635 			return -EINVAL;
636 		} else if (gpuidx < MAX_GPU_INSTANCE &&
637 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
638 			pr_debug("GPU 0x%x not supported\n", val);
639 			return -EINVAL;
640 		}
641 	}
642 
643 	return 0;
644 }
645 
646 static void
647 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
648 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
649 {
650 	uint32_t i;
651 	int gpuidx;
652 
653 	for (i = 0; i < nattr; i++) {
654 		switch (attrs[i].type) {
655 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
656 			prange->preferred_loc = attrs[i].value;
657 			break;
658 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
659 			prange->prefetch_loc = attrs[i].value;
660 			break;
661 		case KFD_IOCTL_SVM_ATTR_ACCESS:
662 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
663 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
664 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
665 							       attrs[i].value);
666 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
667 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
668 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
669 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
670 				bitmap_set(prange->bitmap_access, gpuidx, 1);
671 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
672 			} else {
673 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
674 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
675 			}
676 			break;
677 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
678 			prange->flags |= attrs[i].value;
679 			break;
680 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
681 			prange->flags &= ~attrs[i].value;
682 			break;
683 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
684 			prange->granularity = attrs[i].value;
685 			break;
686 		default:
687 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
688 		}
689 	}
690 }
691 
692 /**
693  * svm_range_debug_dump - print all range information from svms
694  * @svms: svm range list header
695  *
696  * debug output svm range start, end, prefetch location from svms
697  * interval tree and link list
698  *
699  * Context: The caller must hold svms->lock
700  */
701 static void svm_range_debug_dump(struct svm_range_list *svms)
702 {
703 	struct interval_tree_node *node;
704 	struct svm_range *prange;
705 
706 	pr_debug("dump svms 0x%p list\n", svms);
707 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
708 
709 	list_for_each_entry(prange, &svms->list, list) {
710 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
711 			 prange, prange->start, prange->npages,
712 			 prange->start + prange->npages - 1,
713 			 prange->actual_loc);
714 	}
715 
716 	pr_debug("dump svms 0x%p interval tree\n", svms);
717 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
718 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
719 	while (node) {
720 		prange = container_of(node, struct svm_range, it_node);
721 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
722 			 prange, prange->start, prange->npages,
723 			 prange->start + prange->npages - 1,
724 			 prange->actual_loc);
725 		node = interval_tree_iter_next(node, 0, ~0ULL);
726 	}
727 }
728 
729 static bool
730 svm_range_is_same_attrs(struct svm_range *old, struct svm_range *new)
731 {
732 	return (old->prefetch_loc == new->prefetch_loc &&
733 		old->flags == new->flags &&
734 		old->granularity == new->granularity);
735 }
736 
737 static int
738 svm_range_split_array(void *ppnew, void *ppold, size_t size,
739 		      uint64_t old_start, uint64_t old_n,
740 		      uint64_t new_start, uint64_t new_n)
741 {
742 	unsigned char *new, *old, *pold;
743 	uint64_t d;
744 
745 	if (!ppold)
746 		return 0;
747 	pold = *(unsigned char **)ppold;
748 	if (!pold)
749 		return 0;
750 
751 	new = kvmalloc_array(new_n, size, GFP_KERNEL);
752 	if (!new)
753 		return -ENOMEM;
754 
755 	d = (new_start - old_start) * size;
756 	memcpy(new, pold + d, new_n * size);
757 
758 	old = kvmalloc_array(old_n, size, GFP_KERNEL);
759 	if (!old) {
760 		kvfree(new);
761 		return -ENOMEM;
762 	}
763 
764 	d = (new_start == old_start) ? new_n * size : 0;
765 	memcpy(old, pold + d, old_n * size);
766 
767 	kvfree(pold);
768 	*(void **)ppold = old;
769 	*(void **)ppnew = new;
770 
771 	return 0;
772 }
773 
774 static int
775 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
776 		      uint64_t start, uint64_t last)
777 {
778 	uint64_t npages = last - start + 1;
779 	int i, r;
780 
781 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
782 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
783 					  sizeof(*old->dma_addr[i]), old->start,
784 					  npages, new->start, new->npages);
785 		if (r)
786 			return r;
787 	}
788 
789 	return 0;
790 }
791 
792 static int
793 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
794 		      uint64_t start, uint64_t last)
795 {
796 	uint64_t npages = last - start + 1;
797 
798 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
799 		 new->svms, new, new->start, start, last);
800 
801 	if (new->start == old->start) {
802 		new->offset = old->offset;
803 		old->offset += new->npages;
804 	} else {
805 		new->offset = old->offset + npages;
806 	}
807 
808 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
809 	new->ttm_res = old->ttm_res;
810 
811 	spin_lock(&new->svm_bo->list_lock);
812 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
813 	spin_unlock(&new->svm_bo->list_lock);
814 
815 	return 0;
816 }
817 
818 /**
819  * svm_range_split_adjust - split range and adjust
820  *
821  * @new: new range
822  * @old: the old range
823  * @start: the old range adjust to start address in pages
824  * @last: the old range adjust to last address in pages
825  *
826  * Copy system memory dma_addr or vram ttm_res in old range to new
827  * range from new_start up to size new->npages, the remaining old range is from
828  * start to last
829  *
830  * Return:
831  * 0 - OK, -ENOMEM - out of memory
832  */
833 static int
834 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
835 		      uint64_t start, uint64_t last)
836 {
837 	int r;
838 
839 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
840 		 new->svms, new->start, old->start, old->last, start, last);
841 
842 	if (new->start < old->start ||
843 	    new->last > old->last) {
844 		WARN_ONCE(1, "invalid new range start or last\n");
845 		return -EINVAL;
846 	}
847 
848 	r = svm_range_split_pages(new, old, start, last);
849 	if (r)
850 		return r;
851 
852 	if (old->actual_loc && old->ttm_res) {
853 		r = svm_range_split_nodes(new, old, start, last);
854 		if (r)
855 			return r;
856 	}
857 
858 	old->npages = last - start + 1;
859 	old->start = start;
860 	old->last = last;
861 	new->flags = old->flags;
862 	new->preferred_loc = old->preferred_loc;
863 	new->prefetch_loc = old->prefetch_loc;
864 	new->actual_loc = old->actual_loc;
865 	new->granularity = old->granularity;
866 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
867 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
868 
869 	return 0;
870 }
871 
872 /**
873  * svm_range_split - split a range in 2 ranges
874  *
875  * @prange: the svm range to split
876  * @start: the remaining range start address in pages
877  * @last: the remaining range last address in pages
878  * @new: the result new range generated
879  *
880  * Two cases only:
881  * case 1: if start == prange->start
882  *         prange ==> prange[start, last]
883  *         new range [last + 1, prange->last]
884  *
885  * case 2: if last == prange->last
886  *         prange ==> prange[start, last]
887  *         new range [prange->start, start - 1]
888  *
889  * Return:
890  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
891  */
892 static int
893 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
894 		struct svm_range **new)
895 {
896 	uint64_t old_start = prange->start;
897 	uint64_t old_last = prange->last;
898 	struct svm_range_list *svms;
899 	int r = 0;
900 
901 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
902 		 old_start, old_last, start, last);
903 
904 	if (old_start != start && old_last != last)
905 		return -EINVAL;
906 	if (start < old_start || last > old_last)
907 		return -EINVAL;
908 
909 	svms = prange->svms;
910 	if (old_start == start)
911 		*new = svm_range_new(svms, last + 1, old_last);
912 	else
913 		*new = svm_range_new(svms, old_start, start - 1);
914 	if (!*new)
915 		return -ENOMEM;
916 
917 	r = svm_range_split_adjust(*new, prange, start, last);
918 	if (r) {
919 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
920 			 r, old_start, old_last, start, last);
921 		svm_range_free(*new);
922 		*new = NULL;
923 	}
924 
925 	return r;
926 }
927 
928 static int
929 svm_range_split_tail(struct svm_range *prange, struct svm_range *new,
930 		     uint64_t new_last, struct list_head *insert_list)
931 {
932 	struct svm_range *tail;
933 	int r = svm_range_split(prange, prange->start, new_last, &tail);
934 
935 	if (!r)
936 		list_add(&tail->insert_list, insert_list);
937 	return r;
938 }
939 
940 static int
941 svm_range_split_head(struct svm_range *prange, struct svm_range *new,
942 		     uint64_t new_start, struct list_head *insert_list)
943 {
944 	struct svm_range *head;
945 	int r = svm_range_split(prange, new_start, prange->last, &head);
946 
947 	if (!r)
948 		list_add(&head->insert_list, insert_list);
949 	return r;
950 }
951 
952 static void
953 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
954 		    struct svm_range *pchild, enum svm_work_list_ops op)
955 {
956 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
957 		 pchild, pchild->start, pchild->last, prange, op);
958 
959 	pchild->work_item.mm = mm;
960 	pchild->work_item.op = op;
961 	list_add_tail(&pchild->child_list, &prange->child_list);
962 }
963 
964 /**
965  * svm_range_split_by_granularity - collect ranges within granularity boundary
966  *
967  * @p: the process with svms list
968  * @mm: mm structure
969  * @addr: the vm fault address in pages, to split the prange
970  * @parent: parent range if prange is from child list
971  * @prange: prange to split
972  *
973  * Trims @prange to be a single aligned block of prange->granularity if
974  * possible. The head and tail are added to the child_list in @parent.
975  *
976  * Context: caller must hold mmap_read_lock and prange->lock
977  *
978  * Return:
979  * 0 - OK, otherwise error code
980  */
981 int
982 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
983 			       unsigned long addr, struct svm_range *parent,
984 			       struct svm_range *prange)
985 {
986 	struct svm_range *head, *tail;
987 	unsigned long start, last, size;
988 	int r;
989 
990 	/* Align splited range start and size to granularity size, then a single
991 	 * PTE will be used for whole range, this reduces the number of PTE
992 	 * updated and the L1 TLB space used for translation.
993 	 */
994 	size = 1UL << prange->granularity;
995 	start = ALIGN_DOWN(addr, size);
996 	last = ALIGN(addr + 1, size) - 1;
997 
998 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
999 		 prange->svms, prange->start, prange->last, start, last, size);
1000 
1001 	if (start > prange->start) {
1002 		r = svm_range_split(prange, start, prange->last, &head);
1003 		if (r)
1004 			return r;
1005 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1006 	}
1007 
1008 	if (last < prange->last) {
1009 		r = svm_range_split(prange, prange->start, last, &tail);
1010 		if (r)
1011 			return r;
1012 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1013 	}
1014 
1015 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1016 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1017 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1018 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1019 			 prange, prange->start, prange->last,
1020 			 SVM_OP_ADD_RANGE_AND_MAP);
1021 	}
1022 	return 0;
1023 }
1024 
1025 static uint64_t
1026 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1027 			int domain)
1028 {
1029 	struct amdgpu_device *bo_adev;
1030 	uint32_t flags = prange->flags;
1031 	uint32_t mapping_flags = 0;
1032 	uint64_t pte_flags;
1033 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1034 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1035 
1036 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1037 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1038 
1039 	switch (adev->asic_type) {
1040 	case CHIP_ARCTURUS:
1041 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1042 			if (bo_adev == adev) {
1043 				mapping_flags |= coherent ?
1044 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1045 			} else {
1046 				mapping_flags |= coherent ?
1047 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1048 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1049 					snoop = true;
1050 			}
1051 		} else {
1052 			mapping_flags |= coherent ?
1053 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1054 		}
1055 		break;
1056 	case CHIP_ALDEBARAN:
1057 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1058 			if (bo_adev == adev) {
1059 				mapping_flags |= coherent ?
1060 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1061 				if (adev->gmc.xgmi.connected_to_cpu)
1062 					snoop = true;
1063 			} else {
1064 				mapping_flags |= coherent ?
1065 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1066 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1067 					snoop = true;
1068 			}
1069 		} else {
1070 			mapping_flags |= coherent ?
1071 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1072 		}
1073 		break;
1074 	default:
1075 		mapping_flags |= coherent ?
1076 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1077 	}
1078 
1079 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1080 
1081 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1082 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1083 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1084 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1085 
1086 	pte_flags = AMDGPU_PTE_VALID;
1087 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1088 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1089 
1090 	pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1091 
1092 	pr_debug("svms 0x%p [0x%lx 0x%lx] vram %d PTE 0x%llx mapping 0x%x\n",
1093 		 prange->svms, prange->start, prange->last,
1094 		 (domain == SVM_RANGE_VRAM_DOMAIN) ? 1:0, pte_flags, mapping_flags);
1095 
1096 	return pte_flags;
1097 }
1098 
1099 static int
1100 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1101 			 uint64_t start, uint64_t last,
1102 			 struct dma_fence **fence)
1103 {
1104 	uint64_t init_pte_value = 0;
1105 
1106 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1107 
1108 	return amdgpu_vm_bo_update_mapping(adev, adev, vm, false, true, NULL,
1109 					   start, last, init_pte_value, 0,
1110 					   NULL, NULL, fence, NULL);
1111 }
1112 
1113 static int
1114 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1115 			  unsigned long last)
1116 {
1117 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1118 	struct kfd_process_device *pdd;
1119 	struct dma_fence *fence = NULL;
1120 	struct amdgpu_device *adev;
1121 	struct kfd_process *p;
1122 	uint32_t gpuidx;
1123 	int r = 0;
1124 
1125 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1126 		  MAX_GPU_INSTANCE);
1127 	p = container_of(prange->svms, struct kfd_process, svms);
1128 
1129 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1130 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1131 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1132 		if (!pdd) {
1133 			pr_debug("failed to find device idx %d\n", gpuidx);
1134 			return -EINVAL;
1135 		}
1136 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1137 
1138 		r = svm_range_unmap_from_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1139 					     start, last, &fence);
1140 		if (r)
1141 			break;
1142 
1143 		if (fence) {
1144 			r = dma_fence_wait(fence, false);
1145 			dma_fence_put(fence);
1146 			fence = NULL;
1147 			if (r)
1148 				break;
1149 		}
1150 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1151 					p->pasid, TLB_FLUSH_HEAVYWEIGHT);
1152 	}
1153 
1154 	return r;
1155 }
1156 
1157 static int
1158 svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1159 		     struct svm_range *prange, dma_addr_t *dma_addr,
1160 		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
1161 {
1162 	struct amdgpu_bo_va bo_va;
1163 	bool table_freed = false;
1164 	uint64_t pte_flags;
1165 	unsigned long last_start;
1166 	int last_domain;
1167 	int r = 0;
1168 	int64_t i;
1169 
1170 	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
1171 		 prange->last);
1172 
1173 	if (prange->svm_bo && prange->ttm_res)
1174 		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
1175 
1176 	last_start = prange->start;
1177 	for (i = 0; i < prange->npages; i++) {
1178 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1179 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1180 		if ((prange->start + i) < prange->last &&
1181 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1182 			continue;
1183 
1184 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1185 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1186 		pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1187 		r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false, NULL,
1188 						last_start,
1189 						prange->start + i, pte_flags,
1190 						last_start - prange->start,
1191 						NULL,
1192 						dma_addr,
1193 						&vm->last_update,
1194 						&table_freed);
1195 		if (r) {
1196 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1197 			goto out;
1198 		}
1199 		last_start = prange->start + i + 1;
1200 	}
1201 
1202 	r = amdgpu_vm_update_pdes(adev, vm, false);
1203 	if (r) {
1204 		pr_debug("failed %d to update directories 0x%lx\n", r,
1205 			 prange->start);
1206 		goto out;
1207 	}
1208 
1209 	if (fence)
1210 		*fence = dma_fence_get(vm->last_update);
1211 
1212 	if (table_freed) {
1213 		struct kfd_process *p;
1214 
1215 		p = container_of(prange->svms, struct kfd_process, svms);
1216 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1217 						p->pasid, TLB_FLUSH_LEGACY);
1218 	}
1219 out:
1220 	return r;
1221 }
1222 
1223 static int svm_range_map_to_gpus(struct svm_range *prange,
1224 				 unsigned long *bitmap, bool wait)
1225 {
1226 	struct kfd_process_device *pdd;
1227 	struct amdgpu_device *bo_adev;
1228 	struct amdgpu_device *adev;
1229 	struct kfd_process *p;
1230 	struct dma_fence *fence = NULL;
1231 	uint32_t gpuidx;
1232 	int r = 0;
1233 
1234 	if (prange->svm_bo && prange->ttm_res)
1235 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1236 	else
1237 		bo_adev = NULL;
1238 
1239 	p = container_of(prange->svms, struct kfd_process, svms);
1240 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1241 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1242 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1243 		if (!pdd) {
1244 			pr_debug("failed to find device idx %d\n", gpuidx);
1245 			return -EINVAL;
1246 		}
1247 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1248 
1249 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1250 		if (IS_ERR(pdd))
1251 			return -EINVAL;
1252 
1253 		if (bo_adev && adev != bo_adev &&
1254 		    !amdgpu_xgmi_same_hive(adev, bo_adev)) {
1255 			pr_debug("cannot map to device idx %d\n", gpuidx);
1256 			continue;
1257 		}
1258 
1259 		r = svm_range_map_to_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1260 					 prange, prange->dma_addr[gpuidx],
1261 					 bo_adev, wait ? &fence : NULL);
1262 		if (r)
1263 			break;
1264 
1265 		if (fence) {
1266 			r = dma_fence_wait(fence, false);
1267 			dma_fence_put(fence);
1268 			fence = NULL;
1269 			if (r) {
1270 				pr_debug("failed %d to dma fence wait\n", r);
1271 				break;
1272 			}
1273 		}
1274 	}
1275 
1276 	return r;
1277 }
1278 
1279 struct svm_validate_context {
1280 	struct kfd_process *process;
1281 	struct svm_range *prange;
1282 	bool intr;
1283 	unsigned long bitmap[MAX_GPU_INSTANCE];
1284 	struct ttm_validate_buffer tv[MAX_GPU_INSTANCE+1];
1285 	struct list_head validate_list;
1286 	struct ww_acquire_ctx ticket;
1287 };
1288 
1289 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1290 {
1291 	struct kfd_process_device *pdd;
1292 	struct amdgpu_device *adev;
1293 	struct amdgpu_vm *vm;
1294 	uint32_t gpuidx;
1295 	int r;
1296 
1297 	INIT_LIST_HEAD(&ctx->validate_list);
1298 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1299 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1300 		if (!pdd) {
1301 			pr_debug("failed to find device idx %d\n", gpuidx);
1302 			return -EINVAL;
1303 		}
1304 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1305 		vm = drm_priv_to_vm(pdd->drm_priv);
1306 
1307 		ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1308 		ctx->tv[gpuidx].num_shared = 4;
1309 		list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1310 	}
1311 	if (ctx->prange->svm_bo && ctx->prange->ttm_res) {
1312 		ctx->tv[MAX_GPU_INSTANCE].bo = &ctx->prange->svm_bo->bo->tbo;
1313 		ctx->tv[MAX_GPU_INSTANCE].num_shared = 1;
1314 		list_add(&ctx->tv[MAX_GPU_INSTANCE].head, &ctx->validate_list);
1315 	}
1316 
1317 	r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1318 				   ctx->intr, NULL);
1319 	if (r) {
1320 		pr_debug("failed %d to reserve bo\n", r);
1321 		return r;
1322 	}
1323 
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 			r = -EINVAL;
1329 			goto unreserve_out;
1330 		}
1331 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1332 
1333 		r = amdgpu_vm_validate_pt_bos(adev, drm_priv_to_vm(pdd->drm_priv),
1334 					      svm_range_bo_validate, NULL);
1335 		if (r) {
1336 			pr_debug("failed %d validate pt bos\n", r);
1337 			goto unreserve_out;
1338 		}
1339 	}
1340 
1341 	return 0;
1342 
1343 unreserve_out:
1344 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1345 	return r;
1346 }
1347 
1348 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1349 {
1350 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1351 }
1352 
1353 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1354 {
1355 	struct kfd_process_device *pdd;
1356 	struct amdgpu_device *adev;
1357 
1358 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1359 	adev = (struct amdgpu_device *)pdd->dev->kgd;
1360 
1361 	return SVM_ADEV_PGMAP_OWNER(adev);
1362 }
1363 
1364 /*
1365  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1366  *
1367  * To prevent concurrent destruction or change of range attributes, the
1368  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1369  * because that would block concurrent evictions and lead to deadlocks. To
1370  * serialize concurrent migrations or validations of the same range, the
1371  * prange->migrate_mutex must be held.
1372  *
1373  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1374  * eviction fence.
1375  *
1376  * The following sequence ensures race-free validation and GPU mapping:
1377  *
1378  * 1. Reserve page table (and SVM BO if range is in VRAM)
1379  * 2. hmm_range_fault to get page addresses (if system memory)
1380  * 3. DMA-map pages (if system memory)
1381  * 4-a. Take notifier lock
1382  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1383  * 4-c. Check that the range was not split or otherwise invalidated
1384  * 4-d. Update GPU page table
1385  * 4.e. Release notifier lock
1386  * 5. Release page table (and SVM BO) reservation
1387  */
1388 static int svm_range_validate_and_map(struct mm_struct *mm,
1389 				      struct svm_range *prange,
1390 				      int32_t gpuidx, bool intr, bool wait)
1391 {
1392 	struct svm_validate_context ctx;
1393 	struct hmm_range *hmm_range;
1394 	struct kfd_process *p;
1395 	void *owner;
1396 	int32_t idx;
1397 	int r = 0;
1398 
1399 	ctx.process = container_of(prange->svms, struct kfd_process, svms);
1400 	ctx.prange = prange;
1401 	ctx.intr = intr;
1402 
1403 	if (gpuidx < MAX_GPU_INSTANCE) {
1404 		bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1405 		bitmap_set(ctx.bitmap, gpuidx, 1);
1406 	} else if (ctx.process->xnack_enabled) {
1407 		bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1408 
1409 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1410 		 * GPU, which has ACCESS attribute to the range, create mapping
1411 		 * on that GPU.
1412 		 */
1413 		if (prange->actual_loc) {
1414 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1415 							prange->actual_loc);
1416 			if (gpuidx < 0) {
1417 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1418 					 prange->actual_loc);
1419 				return -EINVAL;
1420 			}
1421 			if (test_bit(gpuidx, prange->bitmap_access))
1422 				bitmap_set(ctx.bitmap, gpuidx, 1);
1423 		}
1424 	} else {
1425 		bitmap_or(ctx.bitmap, prange->bitmap_access,
1426 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1427 	}
1428 
1429 	if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE))
1430 		return 0;
1431 
1432 	if (prange->actual_loc && !prange->ttm_res) {
1433 		/* This should never happen. actual_loc gets set by
1434 		 * svm_migrate_ram_to_vram after allocating a BO.
1435 		 */
1436 		WARN(1, "VRAM BO missing during validation\n");
1437 		return -EINVAL;
1438 	}
1439 
1440 	svm_range_reserve_bos(&ctx);
1441 
1442 	p = container_of(prange->svms, struct kfd_process, svms);
1443 	owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1444 						MAX_GPU_INSTANCE));
1445 	for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1446 		if (kfd_svm_page_owner(p, idx) != owner) {
1447 			owner = NULL;
1448 			break;
1449 		}
1450 	}
1451 	r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1452 				       prange->start << PAGE_SHIFT,
1453 				       prange->npages, &hmm_range,
1454 				       false, true, owner);
1455 	if (r) {
1456 		pr_debug("failed %d to get svm range pages\n", r);
1457 		goto unreserve_out;
1458 	}
1459 
1460 	r = svm_range_dma_map(prange, ctx.bitmap,
1461 			      hmm_range->hmm_pfns);
1462 	if (r) {
1463 		pr_debug("failed %d to dma map range\n", r);
1464 		goto unreserve_out;
1465 	}
1466 
1467 	prange->validated_once = true;
1468 
1469 	svm_range_lock(prange);
1470 	if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1471 		pr_debug("hmm update the range, need validate again\n");
1472 		r = -EAGAIN;
1473 		goto unlock_out;
1474 	}
1475 	if (!list_empty(&prange->child_list)) {
1476 		pr_debug("range split by unmap in parallel, validate again\n");
1477 		r = -EAGAIN;
1478 		goto unlock_out;
1479 	}
1480 
1481 	r = svm_range_map_to_gpus(prange, ctx.bitmap, wait);
1482 
1483 unlock_out:
1484 	svm_range_unlock(prange);
1485 unreserve_out:
1486 	svm_range_unreserve_bos(&ctx);
1487 
1488 	if (!r)
1489 		prange->validate_timestamp = ktime_to_us(ktime_get());
1490 
1491 	return r;
1492 }
1493 
1494 /**
1495  * svm_range_list_lock_and_flush_work - flush pending deferred work
1496  *
1497  * @svms: the svm range list
1498  * @mm: the mm structure
1499  *
1500  * Context: Returns with mmap write lock held, pending deferred work flushed
1501  *
1502  */
1503 static void
1504 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1505 				   struct mm_struct *mm)
1506 {
1507 retry_flush_work:
1508 	flush_work(&svms->deferred_list_work);
1509 	mmap_write_lock(mm);
1510 
1511 	if (list_empty(&svms->deferred_range_list))
1512 		return;
1513 	mmap_write_unlock(mm);
1514 	pr_debug("retry flush\n");
1515 	goto retry_flush_work;
1516 }
1517 
1518 static void svm_range_restore_work(struct work_struct *work)
1519 {
1520 	struct delayed_work *dwork = to_delayed_work(work);
1521 	struct amdkfd_process_info *process_info;
1522 	struct svm_range_list *svms;
1523 	struct svm_range *prange;
1524 	struct kfd_process *p;
1525 	struct mm_struct *mm;
1526 	int evicted_ranges;
1527 	int invalid;
1528 	int r;
1529 
1530 	svms = container_of(dwork, struct svm_range_list, restore_work);
1531 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1532 	if (!evicted_ranges)
1533 		return;
1534 
1535 	pr_debug("restore svm ranges\n");
1536 
1537 	/* kfd_process_notifier_release destroys this worker thread. So during
1538 	 * the lifetime of this thread, kfd_process and mm will be valid.
1539 	 */
1540 	p = container_of(svms, struct kfd_process, svms);
1541 	process_info = p->kgd_process_info;
1542 	mm = p->mm;
1543 	if (!mm)
1544 		return;
1545 
1546 	mutex_lock(&process_info->lock);
1547 	svm_range_list_lock_and_flush_work(svms, mm);
1548 	mutex_lock(&svms->lock);
1549 
1550 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1551 
1552 	list_for_each_entry(prange, &svms->list, list) {
1553 		invalid = atomic_read(&prange->invalid);
1554 		if (!invalid)
1555 			continue;
1556 
1557 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1558 			 prange->svms, prange, prange->start, prange->last,
1559 			 invalid);
1560 
1561 		/*
1562 		 * If range is migrating, wait for migration is done.
1563 		 */
1564 		mutex_lock(&prange->migrate_mutex);
1565 
1566 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1567 					       false, true);
1568 		if (r)
1569 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1570 				 prange->start);
1571 
1572 		mutex_unlock(&prange->migrate_mutex);
1573 		if (r)
1574 			goto out_reschedule;
1575 
1576 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1577 			goto out_reschedule;
1578 	}
1579 
1580 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1581 	    evicted_ranges)
1582 		goto out_reschedule;
1583 
1584 	evicted_ranges = 0;
1585 
1586 	r = kgd2kfd_resume_mm(mm);
1587 	if (r) {
1588 		/* No recovery from this failure. Probably the CP is
1589 		 * hanging. No point trying again.
1590 		 */
1591 		pr_debug("failed %d to resume KFD\n", r);
1592 	}
1593 
1594 	pr_debug("restore svm ranges successfully\n");
1595 
1596 out_reschedule:
1597 	mutex_unlock(&svms->lock);
1598 	mmap_write_unlock(mm);
1599 	mutex_unlock(&process_info->lock);
1600 
1601 	/* If validation failed, reschedule another attempt */
1602 	if (evicted_ranges) {
1603 		pr_debug("reschedule to restore svm range\n");
1604 		schedule_delayed_work(&svms->restore_work,
1605 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1606 	}
1607 }
1608 
1609 /**
1610  * svm_range_evict - evict svm range
1611  *
1612  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1613  * return to let CPU evict the buffer and proceed CPU pagetable update.
1614  *
1615  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1616  * If invalidation happens while restore work is running, restore work will
1617  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1618  * the queues.
1619  */
1620 static int
1621 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1622 		unsigned long start, unsigned long last)
1623 {
1624 	struct svm_range_list *svms = prange->svms;
1625 	struct svm_range *pchild;
1626 	struct kfd_process *p;
1627 	int r = 0;
1628 
1629 	p = container_of(svms, struct kfd_process, svms);
1630 
1631 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1632 		 svms, prange->start, prange->last, start, last);
1633 
1634 	if (!p->xnack_enabled) {
1635 		int evicted_ranges;
1636 
1637 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1638 			mutex_lock_nested(&pchild->lock, 1);
1639 			if (pchild->start <= last && pchild->last >= start) {
1640 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1641 					 pchild->start, pchild->last);
1642 				atomic_inc(&pchild->invalid);
1643 			}
1644 			mutex_unlock(&pchild->lock);
1645 		}
1646 
1647 		if (prange->start <= last && prange->last >= start)
1648 			atomic_inc(&prange->invalid);
1649 
1650 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1651 		if (evicted_ranges != 1)
1652 			return r;
1653 
1654 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1655 			 prange->svms, prange->start, prange->last);
1656 
1657 		/* First eviction, stop the queues */
1658 		r = kgd2kfd_quiesce_mm(mm);
1659 		if (r)
1660 			pr_debug("failed to quiesce KFD\n");
1661 
1662 		pr_debug("schedule to restore svm %p ranges\n", svms);
1663 		schedule_delayed_work(&svms->restore_work,
1664 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1665 	} else {
1666 		unsigned long s, l;
1667 
1668 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1669 			 prange->svms, start, last);
1670 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1671 			mutex_lock_nested(&pchild->lock, 1);
1672 			s = max(start, pchild->start);
1673 			l = min(last, pchild->last);
1674 			if (l >= s)
1675 				svm_range_unmap_from_gpus(pchild, s, l);
1676 			mutex_unlock(&pchild->lock);
1677 		}
1678 		s = max(start, prange->start);
1679 		l = min(last, prange->last);
1680 		if (l >= s)
1681 			svm_range_unmap_from_gpus(prange, s, l);
1682 	}
1683 
1684 	return r;
1685 }
1686 
1687 static struct svm_range *svm_range_clone(struct svm_range *old)
1688 {
1689 	struct svm_range *new;
1690 
1691 	new = svm_range_new(old->svms, old->start, old->last);
1692 	if (!new)
1693 		return NULL;
1694 
1695 	if (old->svm_bo) {
1696 		new->ttm_res = old->ttm_res;
1697 		new->offset = old->offset;
1698 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1699 		spin_lock(&new->svm_bo->list_lock);
1700 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1701 		spin_unlock(&new->svm_bo->list_lock);
1702 	}
1703 	new->flags = old->flags;
1704 	new->preferred_loc = old->preferred_loc;
1705 	new->prefetch_loc = old->prefetch_loc;
1706 	new->actual_loc = old->actual_loc;
1707 	new->granularity = old->granularity;
1708 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1709 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1710 
1711 	return new;
1712 }
1713 
1714 /**
1715  * svm_range_handle_overlap - split overlap ranges
1716  * @svms: svm range list header
1717  * @new: range added with this attributes
1718  * @start: range added start address, in pages
1719  * @last: range last address, in pages
1720  * @update_list: output, the ranges attributes are updated. For set_attr, this
1721  *               will do validation and map to GPUs. For unmap, this will be
1722  *               removed and unmap from GPUs
1723  * @insert_list: output, the ranges will be inserted into svms, attributes are
1724  *               not changes. For set_attr, this will add into svms.
1725  * @remove_list:output, the ranges will be removed from svms
1726  * @left: the remaining range after overlap, For set_attr, this will be added
1727  *        as new range.
1728  *
1729  * Total have 5 overlap cases.
1730  *
1731  * This function handles overlap of an address interval with existing
1732  * struct svm_ranges for applying new attributes. This may require
1733  * splitting existing struct svm_ranges. All changes should be applied to
1734  * the range_list and interval tree transactionally. If any split operation
1735  * fails, the entire update fails. Therefore the existing overlapping
1736  * svm_ranges are cloned and the original svm_ranges left unchanged. If the
1737  * transaction succeeds, the modified clones are added and the originals
1738  * freed. Otherwise the clones are removed and the old svm_ranges remain.
1739  *
1740  * Context: The caller must hold svms->lock
1741  */
1742 static int
1743 svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
1744 			 unsigned long start, unsigned long last,
1745 			 struct list_head *update_list,
1746 			 struct list_head *insert_list,
1747 			 struct list_head *remove_list,
1748 			 unsigned long *left)
1749 {
1750 	struct interval_tree_node *node;
1751 	struct svm_range *prange;
1752 	struct svm_range *tmp;
1753 	int r = 0;
1754 
1755 	INIT_LIST_HEAD(update_list);
1756 	INIT_LIST_HEAD(insert_list);
1757 	INIT_LIST_HEAD(remove_list);
1758 
1759 	node = interval_tree_iter_first(&svms->objects, start, last);
1760 	while (node) {
1761 		struct interval_tree_node *next;
1762 		struct svm_range *old;
1763 		unsigned long next_start;
1764 
1765 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
1766 			 node->last);
1767 
1768 		old = container_of(node, struct svm_range, it_node);
1769 		next = interval_tree_iter_next(node, start, last);
1770 		next_start = min(node->last, last) + 1;
1771 
1772 		if (node->start < start || node->last > last) {
1773 			/* node intersects the updated range, clone+split it */
1774 			prange = svm_range_clone(old);
1775 			if (!prange) {
1776 				r = -ENOMEM;
1777 				goto out;
1778 			}
1779 
1780 			list_add(&old->remove_list, remove_list);
1781 			list_add(&prange->insert_list, insert_list);
1782 
1783 			if (node->start < start) {
1784 				pr_debug("change old range start\n");
1785 				r = svm_range_split_head(prange, new, start,
1786 							 insert_list);
1787 				if (r)
1788 					goto out;
1789 			}
1790 			if (node->last > last) {
1791 				pr_debug("change old range last\n");
1792 				r = svm_range_split_tail(prange, new, last,
1793 							 insert_list);
1794 				if (r)
1795 					goto out;
1796 			}
1797 		} else {
1798 			/* The node is contained within start..last,
1799 			 * just update it
1800 			 */
1801 			prange = old;
1802 		}
1803 
1804 		if (!svm_range_is_same_attrs(prange, new))
1805 			list_add(&prange->update_list, update_list);
1806 
1807 		/* insert a new node if needed */
1808 		if (node->start > start) {
1809 			prange = svm_range_new(prange->svms, start,
1810 					       node->start - 1);
1811 			if (!prange) {
1812 				r = -ENOMEM;
1813 				goto out;
1814 			}
1815 
1816 			list_add(&prange->insert_list, insert_list);
1817 			list_add(&prange->update_list, update_list);
1818 		}
1819 
1820 		node = next;
1821 		start = next_start;
1822 	}
1823 
1824 	if (left && start <= last)
1825 		*left = last - start + 1;
1826 
1827 out:
1828 	if (r)
1829 		list_for_each_entry_safe(prange, tmp, insert_list, insert_list)
1830 			svm_range_free(prange);
1831 
1832 	return r;
1833 }
1834 
1835 static void
1836 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
1837 					    struct svm_range *prange)
1838 {
1839 	unsigned long start;
1840 	unsigned long last;
1841 
1842 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
1843 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
1844 
1845 	if (prange->start == start && prange->last == last)
1846 		return;
1847 
1848 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1849 		  prange->svms, prange, start, last, prange->start,
1850 		  prange->last);
1851 
1852 	if (start != 0 && last != 0) {
1853 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
1854 		svm_range_remove_notifier(prange);
1855 	}
1856 	prange->it_node.start = prange->start;
1857 	prange->it_node.last = prange->last;
1858 
1859 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
1860 	svm_range_add_notifier_locked(mm, prange);
1861 }
1862 
1863 static void
1864 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange)
1865 {
1866 	struct mm_struct *mm = prange->work_item.mm;
1867 
1868 	switch (prange->work_item.op) {
1869 	case SVM_OP_NULL:
1870 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1871 			 svms, prange, prange->start, prange->last);
1872 		break;
1873 	case SVM_OP_UNMAP_RANGE:
1874 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1875 			 svms, prange, prange->start, prange->last);
1876 		svm_range_unlink(prange);
1877 		svm_range_remove_notifier(prange);
1878 		svm_range_free(prange);
1879 		break;
1880 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
1881 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1882 			 svms, prange, prange->start, prange->last);
1883 		svm_range_update_notifier_and_interval_tree(mm, prange);
1884 		break;
1885 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
1886 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1887 			 svms, prange, prange->start, prange->last);
1888 		svm_range_update_notifier_and_interval_tree(mm, prange);
1889 		/* TODO: implement deferred validation and mapping */
1890 		break;
1891 	case SVM_OP_ADD_RANGE:
1892 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
1893 			 prange->start, prange->last);
1894 		svm_range_add_to_svms(prange);
1895 		svm_range_add_notifier_locked(mm, prange);
1896 		break;
1897 	case SVM_OP_ADD_RANGE_AND_MAP:
1898 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
1899 			 prange, prange->start, prange->last);
1900 		svm_range_add_to_svms(prange);
1901 		svm_range_add_notifier_locked(mm, prange);
1902 		/* TODO: implement deferred validation and mapping */
1903 		break;
1904 	default:
1905 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
1906 			 prange->work_item.op);
1907 	}
1908 }
1909 
1910 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
1911 {
1912 	struct kfd_process_device *pdd;
1913 	struct amdgpu_device *adev;
1914 	struct kfd_process *p;
1915 	uint32_t i;
1916 
1917 	p = container_of(svms, struct kfd_process, svms);
1918 
1919 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
1920 		pdd = p->pdds[i];
1921 		if (!pdd)
1922 			continue;
1923 
1924 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
1925 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1926 
1927 		amdgpu_ih_wait_on_checkpoint_process(adev, &adev->irq.ih1);
1928 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
1929 	}
1930 }
1931 
1932 static void svm_range_deferred_list_work(struct work_struct *work)
1933 {
1934 	struct svm_range_list *svms;
1935 	struct svm_range *prange;
1936 	struct mm_struct *mm;
1937 
1938 	svms = container_of(work, struct svm_range_list, deferred_list_work);
1939 	pr_debug("enter svms 0x%p\n", svms);
1940 
1941 	spin_lock(&svms->deferred_list_lock);
1942 	while (!list_empty(&svms->deferred_range_list)) {
1943 		prange = list_first_entry(&svms->deferred_range_list,
1944 					  struct svm_range, deferred_list);
1945 		spin_unlock(&svms->deferred_list_lock);
1946 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
1947 			 prange->start, prange->last, prange->work_item.op);
1948 
1949 		/* Make sure no stale retry fault coming after range is freed */
1950 		if (prange->work_item.op == SVM_OP_UNMAP_RANGE)
1951 			svm_range_drain_retry_fault(prange->svms);
1952 
1953 		mm = prange->work_item.mm;
1954 		mmap_write_lock(mm);
1955 		mutex_lock(&svms->lock);
1956 
1957 		/* Remove from deferred_list must be inside mmap write lock,
1958 		 * otherwise, svm_range_list_lock_and_flush_work may hold mmap
1959 		 * write lock, and continue because deferred_list is empty, then
1960 		 * deferred_list handle is blocked by mmap write lock.
1961 		 */
1962 		spin_lock(&svms->deferred_list_lock);
1963 		list_del_init(&prange->deferred_list);
1964 		spin_unlock(&svms->deferred_list_lock);
1965 
1966 		mutex_lock(&prange->migrate_mutex);
1967 		while (!list_empty(&prange->child_list)) {
1968 			struct svm_range *pchild;
1969 
1970 			pchild = list_first_entry(&prange->child_list,
1971 						struct svm_range, child_list);
1972 			pr_debug("child prange 0x%p op %d\n", pchild,
1973 				 pchild->work_item.op);
1974 			list_del_init(&pchild->child_list);
1975 			svm_range_handle_list_op(svms, pchild);
1976 		}
1977 		mutex_unlock(&prange->migrate_mutex);
1978 
1979 		svm_range_handle_list_op(svms, prange);
1980 		mutex_unlock(&svms->lock);
1981 		mmap_write_unlock(mm);
1982 
1983 		spin_lock(&svms->deferred_list_lock);
1984 	}
1985 	spin_unlock(&svms->deferred_list_lock);
1986 
1987 	pr_debug("exit svms 0x%p\n", svms);
1988 }
1989 
1990 void
1991 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
1992 			struct mm_struct *mm, enum svm_work_list_ops op)
1993 {
1994 	spin_lock(&svms->deferred_list_lock);
1995 	/* if prange is on the deferred list */
1996 	if (!list_empty(&prange->deferred_list)) {
1997 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
1998 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
1999 		if (op != SVM_OP_NULL &&
2000 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2001 			prange->work_item.op = op;
2002 	} else {
2003 		prange->work_item.op = op;
2004 		prange->work_item.mm = mm;
2005 		list_add_tail(&prange->deferred_list,
2006 			      &prange->svms->deferred_range_list);
2007 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2008 			 prange, prange->start, prange->last, op);
2009 	}
2010 	spin_unlock(&svms->deferred_list_lock);
2011 }
2012 
2013 void schedule_deferred_list_work(struct svm_range_list *svms)
2014 {
2015 	spin_lock(&svms->deferred_list_lock);
2016 	if (!list_empty(&svms->deferred_range_list))
2017 		schedule_work(&svms->deferred_list_work);
2018 	spin_unlock(&svms->deferred_list_lock);
2019 }
2020 
2021 static void
2022 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2023 		      struct svm_range *prange, unsigned long start,
2024 		      unsigned long last)
2025 {
2026 	struct svm_range *head;
2027 	struct svm_range *tail;
2028 
2029 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2030 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2031 			 prange->start, prange->last);
2032 		return;
2033 	}
2034 	if (start > prange->last || last < prange->start)
2035 		return;
2036 
2037 	head = tail = prange;
2038 	if (start > prange->start)
2039 		svm_range_split(prange, prange->start, start - 1, &tail);
2040 	if (last < tail->last)
2041 		svm_range_split(tail, last + 1, tail->last, &head);
2042 
2043 	if (head != prange && tail != prange) {
2044 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2045 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2046 	} else if (tail != prange) {
2047 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2048 	} else if (head != prange) {
2049 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2050 	} else if (parent != prange) {
2051 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2052 	}
2053 }
2054 
2055 static void
2056 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2057 			 unsigned long start, unsigned long last)
2058 {
2059 	struct svm_range_list *svms;
2060 	struct svm_range *pchild;
2061 	struct kfd_process *p;
2062 	unsigned long s, l;
2063 	bool unmap_parent;
2064 
2065 	p = kfd_lookup_process_by_mm(mm);
2066 	if (!p)
2067 		return;
2068 	svms = &p->svms;
2069 
2070 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2071 		 prange, prange->start, prange->last, start, last);
2072 
2073 	unmap_parent = start <= prange->start && last >= prange->last;
2074 
2075 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2076 		mutex_lock_nested(&pchild->lock, 1);
2077 		s = max(start, pchild->start);
2078 		l = min(last, pchild->last);
2079 		if (l >= s)
2080 			svm_range_unmap_from_gpus(pchild, s, l);
2081 		svm_range_unmap_split(mm, prange, pchild, start, last);
2082 		mutex_unlock(&pchild->lock);
2083 	}
2084 	s = max(start, prange->start);
2085 	l = min(last, prange->last);
2086 	if (l >= s)
2087 		svm_range_unmap_from_gpus(prange, s, l);
2088 	svm_range_unmap_split(mm, prange, prange, start, last);
2089 
2090 	if (unmap_parent)
2091 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2092 	else
2093 		svm_range_add_list_work(svms, prange, mm,
2094 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2095 	schedule_deferred_list_work(svms);
2096 
2097 	kfd_unref_process(p);
2098 }
2099 
2100 /**
2101  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2102  *
2103  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2104  * is from migration, or CPU page invalidation callback.
2105  *
2106  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2107  * work thread, and split prange if only part of prange is unmapped.
2108  *
2109  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2110  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2111  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2112  * update GPU mapping to recover.
2113  *
2114  * Context: mmap lock, notifier_invalidate_start lock are held
2115  *          for invalidate event, prange lock is held if this is from migration
2116  */
2117 static bool
2118 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2119 				    const struct mmu_notifier_range *range,
2120 				    unsigned long cur_seq)
2121 {
2122 	struct svm_range *prange;
2123 	unsigned long start;
2124 	unsigned long last;
2125 
2126 	if (range->event == MMU_NOTIFY_RELEASE)
2127 		return true;
2128 
2129 	start = mni->interval_tree.start;
2130 	last = mni->interval_tree.last;
2131 	start = (start > range->start ? start : range->start) >> PAGE_SHIFT;
2132 	last = (last < (range->end - 1) ? last : range->end - 1) >> PAGE_SHIFT;
2133 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2134 		 start, last, range->start >> PAGE_SHIFT,
2135 		 (range->end - 1) >> PAGE_SHIFT,
2136 		 mni->interval_tree.start >> PAGE_SHIFT,
2137 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2138 
2139 	prange = container_of(mni, struct svm_range, notifier);
2140 
2141 	svm_range_lock(prange);
2142 	mmu_interval_set_seq(mni, cur_seq);
2143 
2144 	switch (range->event) {
2145 	case MMU_NOTIFY_UNMAP:
2146 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2147 		break;
2148 	default:
2149 		svm_range_evict(prange, mni->mm, start, last);
2150 		break;
2151 	}
2152 
2153 	svm_range_unlock(prange);
2154 
2155 	return true;
2156 }
2157 
2158 /**
2159  * svm_range_from_addr - find svm range from fault address
2160  * @svms: svm range list header
2161  * @addr: address to search range interval tree, in pages
2162  * @parent: parent range if range is on child list
2163  *
2164  * Context: The caller must hold svms->lock
2165  *
2166  * Return: the svm_range found or NULL
2167  */
2168 struct svm_range *
2169 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2170 		    struct svm_range **parent)
2171 {
2172 	struct interval_tree_node *node;
2173 	struct svm_range *prange;
2174 	struct svm_range *pchild;
2175 
2176 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2177 	if (!node)
2178 		return NULL;
2179 
2180 	prange = container_of(node, struct svm_range, it_node);
2181 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2182 		 addr, prange->start, prange->last, node->start, node->last);
2183 
2184 	if (addr >= prange->start && addr <= prange->last) {
2185 		if (parent)
2186 			*parent = prange;
2187 		return prange;
2188 	}
2189 	list_for_each_entry(pchild, &prange->child_list, child_list)
2190 		if (addr >= pchild->start && addr <= pchild->last) {
2191 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2192 				 addr, pchild->start, pchild->last);
2193 			if (parent)
2194 				*parent = prange;
2195 			return pchild;
2196 		}
2197 
2198 	return NULL;
2199 }
2200 
2201 /* svm_range_best_restore_location - decide the best fault restore location
2202  * @prange: svm range structure
2203  * @adev: the GPU on which vm fault happened
2204  *
2205  * This is only called when xnack is on, to decide the best location to restore
2206  * the range mapping after GPU vm fault. Caller uses the best location to do
2207  * migration if actual loc is not best location, then update GPU page table
2208  * mapping to the best location.
2209  *
2210  * If vm fault gpu is range preferred loc, the best_loc is preferred loc.
2211  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2212  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2213  *    if range actual loc is cpu, best_loc is cpu
2214  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2215  *    range actual loc.
2216  * Otherwise, GPU no access, best_loc is -1.
2217  *
2218  * Return:
2219  * -1 means vm fault GPU no access
2220  * 0 for CPU or GPU id
2221  */
2222 static int32_t
2223 svm_range_best_restore_location(struct svm_range *prange,
2224 				struct amdgpu_device *adev,
2225 				int32_t *gpuidx)
2226 {
2227 	struct amdgpu_device *bo_adev;
2228 	struct kfd_process *p;
2229 	uint32_t gpuid;
2230 	int r;
2231 
2232 	p = container_of(prange->svms, struct kfd_process, svms);
2233 
2234 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, gpuidx);
2235 	if (r < 0) {
2236 		pr_debug("failed to get gpuid from kgd\n");
2237 		return -1;
2238 	}
2239 
2240 	if (prange->preferred_loc == gpuid)
2241 		return prange->preferred_loc;
2242 
2243 	if (test_bit(*gpuidx, prange->bitmap_access))
2244 		return gpuid;
2245 
2246 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2247 		if (!prange->actual_loc)
2248 			return 0;
2249 
2250 		bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2251 		if (amdgpu_xgmi_same_hive(adev, bo_adev))
2252 			return prange->actual_loc;
2253 		else
2254 			return 0;
2255 	}
2256 
2257 	return -1;
2258 }
2259 static int
2260 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2261 				unsigned long *start, unsigned long *last)
2262 {
2263 	struct vm_area_struct *vma;
2264 	struct interval_tree_node *node;
2265 	unsigned long start_limit, end_limit;
2266 
2267 	vma = find_vma(p->mm, addr << PAGE_SHIFT);
2268 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2269 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2270 		return -EFAULT;
2271 	}
2272 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2273 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2274 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2275 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2276 	/* First range that starts after the fault address */
2277 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2278 	if (node) {
2279 		end_limit = min(end_limit, node->start);
2280 		/* Last range that ends before the fault address */
2281 		node = container_of(rb_prev(&node->rb),
2282 				    struct interval_tree_node, rb);
2283 	} else {
2284 		/* Last range must end before addr because
2285 		 * there was no range after addr
2286 		 */
2287 		node = container_of(rb_last(&p->svms.objects.rb_root),
2288 				    struct interval_tree_node, rb);
2289 	}
2290 	if (node) {
2291 		if (node->last >= addr) {
2292 			WARN(1, "Overlap with prev node and page fault addr\n");
2293 			return -EFAULT;
2294 		}
2295 		start_limit = max(start_limit, node->last + 1);
2296 	}
2297 
2298 	*start = start_limit;
2299 	*last = end_limit - 1;
2300 
2301 	pr_debug("vma start: 0x%lx start: 0x%lx vma end: 0x%lx last: 0x%lx\n",
2302 		  vma->vm_start >> PAGE_SHIFT, *start,
2303 		  vma->vm_end >> PAGE_SHIFT, *last);
2304 
2305 	return 0;
2306 
2307 }
2308 static struct
2309 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2310 						struct kfd_process *p,
2311 						struct mm_struct *mm,
2312 						int64_t addr)
2313 {
2314 	struct svm_range *prange = NULL;
2315 	unsigned long start, last;
2316 	uint32_t gpuid, gpuidx;
2317 
2318 	if (svm_range_get_range_boundaries(p, addr, &start, &last))
2319 		return NULL;
2320 
2321 	prange = svm_range_new(&p->svms, start, last);
2322 	if (!prange) {
2323 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2324 		return NULL;
2325 	}
2326 	if (kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx)) {
2327 		pr_debug("failed to get gpuid from kgd\n");
2328 		svm_range_free(prange);
2329 		return NULL;
2330 	}
2331 
2332 	svm_range_add_to_svms(prange);
2333 	svm_range_add_notifier_locked(mm, prange);
2334 
2335 	return prange;
2336 }
2337 
2338 /* svm_range_skip_recover - decide if prange can be recovered
2339  * @prange: svm range structure
2340  *
2341  * GPU vm retry fault handle skip recover the range for cases:
2342  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2343  *    deferred list work will drain the stale fault before free the prange.
2344  * 2. prange is on deferred list to add interval notifier after split, or
2345  * 3. prange is child range, it is split from parent prange, recover later
2346  *    after interval notifier is added.
2347  *
2348  * Return: true to skip recover, false to recover
2349  */
2350 static bool svm_range_skip_recover(struct svm_range *prange)
2351 {
2352 	struct svm_range_list *svms = prange->svms;
2353 
2354 	spin_lock(&svms->deferred_list_lock);
2355 	if (list_empty(&prange->deferred_list) &&
2356 	    list_empty(&prange->child_list)) {
2357 		spin_unlock(&svms->deferred_list_lock);
2358 		return false;
2359 	}
2360 	spin_unlock(&svms->deferred_list_lock);
2361 
2362 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2363 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2364 			 svms, prange, prange->start, prange->last);
2365 		return true;
2366 	}
2367 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2368 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2369 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2370 			 svms, prange, prange->start, prange->last);
2371 		return true;
2372 	}
2373 	return false;
2374 }
2375 
2376 static void
2377 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2378 		      int32_t gpuidx)
2379 {
2380 	struct kfd_process_device *pdd;
2381 
2382 	/* fault is on different page of same range
2383 	 * or fault is skipped to recover later
2384 	 * or fault is on invalid virtual address
2385 	 */
2386 	if (gpuidx == MAX_GPU_INSTANCE) {
2387 		uint32_t gpuid;
2388 		int r;
2389 
2390 		r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx);
2391 		if (r < 0)
2392 			return;
2393 	}
2394 
2395 	/* fault is recovered
2396 	 * or fault cannot recover because GPU no access on the range
2397 	 */
2398 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2399 	if (pdd)
2400 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2401 }
2402 
2403 int
2404 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2405 			uint64_t addr)
2406 {
2407 	struct mm_struct *mm = NULL;
2408 	struct svm_range_list *svms;
2409 	struct svm_range *prange;
2410 	struct kfd_process *p;
2411 	uint64_t timestamp;
2412 	int32_t best_loc;
2413 	int32_t gpuidx = MAX_GPU_INSTANCE;
2414 	bool write_locked = false;
2415 	int r = 0;
2416 
2417 	if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2418 		pr_debug("device does not support SVM\n");
2419 		return -EFAULT;
2420 	}
2421 
2422 	p = kfd_lookup_process_by_pasid(pasid);
2423 	if (!p) {
2424 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2425 		return -ESRCH;
2426 	}
2427 	if (!p->xnack_enabled) {
2428 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2429 		return -EFAULT;
2430 	}
2431 	svms = &p->svms;
2432 
2433 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2434 
2435 	mm = get_task_mm(p->lead_thread);
2436 	if (!mm) {
2437 		pr_debug("svms 0x%p failed to get mm\n", svms);
2438 		r = -ESRCH;
2439 		goto out;
2440 	}
2441 
2442 	mmap_read_lock(mm);
2443 retry_write_locked:
2444 	mutex_lock(&svms->lock);
2445 	prange = svm_range_from_addr(svms, addr, NULL);
2446 	if (!prange) {
2447 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2448 			 svms, addr);
2449 		if (!write_locked) {
2450 			/* Need the write lock to create new range with MMU notifier.
2451 			 * Also flush pending deferred work to make sure the interval
2452 			 * tree is up to date before we add a new range
2453 			 */
2454 			mutex_unlock(&svms->lock);
2455 			mmap_read_unlock(mm);
2456 			mmap_write_lock(mm);
2457 			write_locked = true;
2458 			goto retry_write_locked;
2459 		}
2460 		prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2461 		if (!prange) {
2462 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2463 				 svms, addr);
2464 			mmap_write_downgrade(mm);
2465 			r = -EFAULT;
2466 			goto out_unlock_svms;
2467 		}
2468 	}
2469 	if (write_locked)
2470 		mmap_write_downgrade(mm);
2471 
2472 	mutex_lock(&prange->migrate_mutex);
2473 
2474 	if (svm_range_skip_recover(prange)) {
2475 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2476 		goto out_unlock_range;
2477 	}
2478 
2479 	timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp;
2480 	/* skip duplicate vm fault on different pages of same range */
2481 	if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
2482 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2483 			 svms, prange->start, prange->last);
2484 		goto out_unlock_range;
2485 	}
2486 
2487 	best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2488 	if (best_loc == -1) {
2489 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2490 			 svms, prange->start, prange->last);
2491 		r = -EACCES;
2492 		goto out_unlock_range;
2493 	}
2494 
2495 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2496 		 svms, prange->start, prange->last, best_loc,
2497 		 prange->actual_loc);
2498 
2499 	if (prange->actual_loc != best_loc) {
2500 		if (best_loc) {
2501 			r = svm_migrate_to_vram(prange, best_loc, mm);
2502 			if (r) {
2503 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2504 					 r, addr);
2505 				/* Fallback to system memory if migration to
2506 				 * VRAM failed
2507 				 */
2508 				if (prange->actual_loc)
2509 					r = svm_migrate_vram_to_ram(prange, mm);
2510 				else
2511 					r = 0;
2512 			}
2513 		} else {
2514 			r = svm_migrate_vram_to_ram(prange, mm);
2515 		}
2516 		if (r) {
2517 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2518 				 r, svms, prange->start, prange->last);
2519 			goto out_unlock_range;
2520 		}
2521 	}
2522 
2523 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false);
2524 	if (r)
2525 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2526 			 r, svms, prange->start, prange->last);
2527 
2528 out_unlock_range:
2529 	mutex_unlock(&prange->migrate_mutex);
2530 out_unlock_svms:
2531 	mutex_unlock(&svms->lock);
2532 	mmap_read_unlock(mm);
2533 
2534 	svm_range_count_fault(adev, p, gpuidx);
2535 
2536 	mmput(mm);
2537 out:
2538 	kfd_unref_process(p);
2539 
2540 	if (r == -EAGAIN) {
2541 		pr_debug("recover vm fault later\n");
2542 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2543 		r = 0;
2544 	}
2545 	return r;
2546 }
2547 
2548 void svm_range_list_fini(struct kfd_process *p)
2549 {
2550 	struct svm_range *prange;
2551 	struct svm_range *next;
2552 
2553 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
2554 
2555 	/* Ensure list work is finished before process is destroyed */
2556 	flush_work(&p->svms.deferred_list_work);
2557 
2558 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
2559 		svm_range_unlink(prange);
2560 		svm_range_remove_notifier(prange);
2561 		svm_range_free(prange);
2562 	}
2563 
2564 	mutex_destroy(&p->svms.lock);
2565 
2566 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
2567 }
2568 
2569 int svm_range_list_init(struct kfd_process *p)
2570 {
2571 	struct svm_range_list *svms = &p->svms;
2572 	int i;
2573 
2574 	svms->objects = RB_ROOT_CACHED;
2575 	mutex_init(&svms->lock);
2576 	INIT_LIST_HEAD(&svms->list);
2577 	atomic_set(&svms->evicted_ranges, 0);
2578 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
2579 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
2580 	INIT_LIST_HEAD(&svms->deferred_range_list);
2581 	spin_lock_init(&svms->deferred_list_lock);
2582 
2583 	for (i = 0; i < p->n_pdds; i++)
2584 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
2585 			bitmap_set(svms->bitmap_supported, i, 1);
2586 
2587 	return 0;
2588 }
2589 
2590 /**
2591  * svm_range_is_valid - check if virtual address range is valid
2592  * @mm: current process mm_struct
2593  * @start: range start address, in pages
2594  * @size: range size, in pages
2595  *
2596  * Valid virtual address range means it belongs to one or more VMAs
2597  *
2598  * Context: Process context
2599  *
2600  * Return:
2601  *  true - valid svm range
2602  *  false - invalid svm range
2603  */
2604 static bool
2605 svm_range_is_valid(struct mm_struct *mm, uint64_t start, uint64_t size)
2606 {
2607 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
2608 	struct vm_area_struct *vma;
2609 	unsigned long end;
2610 
2611 	start <<= PAGE_SHIFT;
2612 	end = start + (size << PAGE_SHIFT);
2613 
2614 	do {
2615 		vma = find_vma(mm, start);
2616 		if (!vma || start < vma->vm_start ||
2617 		    (vma->vm_flags & device_vma))
2618 			return false;
2619 		start = min(end, vma->vm_end);
2620 	} while (start < end);
2621 
2622 	return true;
2623 }
2624 
2625 /**
2626  * svm_range_add - add svm range and handle overlap
2627  * @p: the range add to this process svms
2628  * @start: page size aligned
2629  * @size: page size aligned
2630  * @nattr: number of attributes
2631  * @attrs: array of attributes
2632  * @update_list: output, the ranges need validate and update GPU mapping
2633  * @insert_list: output, the ranges need insert to svms
2634  * @remove_list: output, the ranges are replaced and need remove from svms
2635  *
2636  * Check if the virtual address range has overlap with the registered ranges,
2637  * split the overlapped range, copy and adjust pages address and vram nodes in
2638  * old and new ranges.
2639  *
2640  * Context: Process context, caller must hold svms->lock
2641  *
2642  * Return:
2643  * 0 - OK, otherwise error code
2644  */
2645 static int
2646 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2647 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2648 	      struct list_head *update_list, struct list_head *insert_list,
2649 	      struct list_head *remove_list)
2650 {
2651 	uint64_t last = start + size - 1UL;
2652 	struct svm_range_list *svms;
2653 	struct svm_range new = {0};
2654 	struct svm_range *prange;
2655 	unsigned long left = 0;
2656 	int r = 0;
2657 
2658 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", &p->svms, start, last);
2659 
2660 	svm_range_apply_attrs(p, &new, nattr, attrs);
2661 
2662 	svms = &p->svms;
2663 
2664 	r = svm_range_handle_overlap(svms, &new, start, last, update_list,
2665 				     insert_list, remove_list, &left);
2666 	if (r)
2667 		return r;
2668 
2669 	if (left) {
2670 		prange = svm_range_new(svms, last - left + 1, last);
2671 		list_add(&prange->insert_list, insert_list);
2672 		list_add(&prange->update_list, update_list);
2673 	}
2674 
2675 	return 0;
2676 }
2677 
2678 /* svm_range_best_prefetch_location - decide the best prefetch location
2679  * @prange: svm range structure
2680  *
2681  * For xnack off:
2682  * If range map to single GPU, the best acutal location is prefetch loc, which
2683  * can be CPU or GPU.
2684  *
2685  * If range map to multiple GPUs, only if mGPU connection on xgmi same hive,
2686  * the best actual location could be prefetch_loc GPU. If mGPU connection on
2687  * PCIe, the best actual location is always CPU, because GPU cannot access vram
2688  * of other GPUs, assuming PCIe small bar (large bar support is not upstream).
2689  *
2690  * For xnack on:
2691  * The best actual location is prefetch location. If mGPU connection on xgmi
2692  * same hive, range map to multiple GPUs. Otherwise, the range only map to
2693  * actual location GPU. Other GPU access vm fault will trigger migration.
2694  *
2695  * Context: Process context
2696  *
2697  * Return:
2698  * 0 for CPU or GPU id
2699  */
2700 static uint32_t
2701 svm_range_best_prefetch_location(struct svm_range *prange)
2702 {
2703 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
2704 	uint32_t best_loc = prange->prefetch_loc;
2705 	struct kfd_process_device *pdd;
2706 	struct amdgpu_device *bo_adev;
2707 	struct amdgpu_device *adev;
2708 	struct kfd_process *p;
2709 	uint32_t gpuidx;
2710 
2711 	p = container_of(prange->svms, struct kfd_process, svms);
2712 
2713 	/* xnack on */
2714 	if (p->xnack_enabled)
2715 		goto out;
2716 
2717 	/* xnack off */
2718 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
2719 		goto out;
2720 
2721 	bo_adev = svm_range_get_adev_by_id(prange, best_loc);
2722 	if (!bo_adev) {
2723 		WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
2724 		best_loc = 0;
2725 		goto out;
2726 	}
2727 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
2728 		  MAX_GPU_INSTANCE);
2729 
2730 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
2731 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2732 		if (!pdd) {
2733 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
2734 			continue;
2735 		}
2736 		adev = (struct amdgpu_device *)pdd->dev->kgd;
2737 
2738 		if (adev == bo_adev)
2739 			continue;
2740 
2741 		if (!amdgpu_xgmi_same_hive(adev, bo_adev)) {
2742 			best_loc = 0;
2743 			break;
2744 		}
2745 	}
2746 
2747 out:
2748 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
2749 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
2750 		 best_loc);
2751 
2752 	return best_loc;
2753 }
2754 
2755 /* FIXME: This is a workaround for page locking bug when some pages are
2756  * invalid during migration to VRAM
2757  */
2758 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
2759 			void *owner)
2760 {
2761 	struct hmm_range *hmm_range;
2762 	int r;
2763 
2764 	if (prange->validated_once)
2765 		return;
2766 
2767 	r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
2768 				       prange->start << PAGE_SHIFT,
2769 				       prange->npages, &hmm_range,
2770 				       false, true, owner);
2771 	if (!r) {
2772 		amdgpu_hmm_range_get_pages_done(hmm_range);
2773 		prange->validated_once = true;
2774 	}
2775 }
2776 
2777 /* svm_range_trigger_migration - start page migration if prefetch loc changed
2778  * @mm: current process mm_struct
2779  * @prange: svm range structure
2780  * @migrated: output, true if migration is triggered
2781  *
2782  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
2783  * from ram to vram.
2784  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
2785  * from vram to ram.
2786  *
2787  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
2788  * and restore work:
2789  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
2790  *    stops all queues, schedule restore work
2791  * 2. svm_range_restore_work wait for migration is done by
2792  *    a. svm_range_validate_vram takes prange->migrate_mutex
2793  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
2794  * 3. restore work update mappings of GPU, resume all queues.
2795  *
2796  * Context: Process context
2797  *
2798  * Return:
2799  * 0 - OK, otherwise - error code of migration
2800  */
2801 static int
2802 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
2803 			    bool *migrated)
2804 {
2805 	uint32_t best_loc;
2806 	int r = 0;
2807 
2808 	*migrated = false;
2809 	best_loc = svm_range_best_prefetch_location(prange);
2810 
2811 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
2812 	    best_loc == prange->actual_loc)
2813 		return 0;
2814 
2815 	if (!best_loc) {
2816 		r = svm_migrate_vram_to_ram(prange, mm);
2817 		*migrated = !r;
2818 		return r;
2819 	}
2820 
2821 	r = svm_migrate_to_vram(prange, best_loc, mm);
2822 	*migrated = !r;
2823 
2824 	return r;
2825 }
2826 
2827 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
2828 {
2829 	if (!fence)
2830 		return -EINVAL;
2831 
2832 	if (dma_fence_is_signaled(&fence->base))
2833 		return 0;
2834 
2835 	if (fence->svm_bo) {
2836 		WRITE_ONCE(fence->svm_bo->evicting, 1);
2837 		schedule_work(&fence->svm_bo->eviction_work);
2838 	}
2839 
2840 	return 0;
2841 }
2842 
2843 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
2844 {
2845 	struct svm_range_bo *svm_bo;
2846 	struct kfd_process *p;
2847 	struct mm_struct *mm;
2848 
2849 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
2850 	if (!svm_bo_ref_unless_zero(svm_bo))
2851 		return; /* svm_bo was freed while eviction was pending */
2852 
2853 	/* svm_range_bo_release destroys this worker thread. So during
2854 	 * the lifetime of this thread, kfd_process and mm will be valid.
2855 	 */
2856 	p = container_of(svm_bo->svms, struct kfd_process, svms);
2857 	mm = p->mm;
2858 	if (!mm)
2859 		return;
2860 
2861 	mmap_read_lock(mm);
2862 	spin_lock(&svm_bo->list_lock);
2863 	while (!list_empty(&svm_bo->range_list)) {
2864 		struct svm_range *prange =
2865 				list_first_entry(&svm_bo->range_list,
2866 						struct svm_range, svm_bo_list);
2867 		list_del_init(&prange->svm_bo_list);
2868 		spin_unlock(&svm_bo->list_lock);
2869 
2870 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
2871 			 prange->start, prange->last);
2872 
2873 		mutex_lock(&prange->migrate_mutex);
2874 		svm_migrate_vram_to_ram(prange, svm_bo->eviction_fence->mm);
2875 
2876 		mutex_lock(&prange->lock);
2877 		prange->svm_bo = NULL;
2878 		mutex_unlock(&prange->lock);
2879 
2880 		mutex_unlock(&prange->migrate_mutex);
2881 
2882 		spin_lock(&svm_bo->list_lock);
2883 	}
2884 	spin_unlock(&svm_bo->list_lock);
2885 	mmap_read_unlock(mm);
2886 
2887 	dma_fence_signal(&svm_bo->eviction_fence->base);
2888 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
2889 	 * has been called in svm_migrate_vram_to_ram
2890 	 */
2891 	WARN_ONCE(kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
2892 	svm_range_bo_unref(svm_bo);
2893 }
2894 
2895 static int
2896 svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
2897 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
2898 {
2899 	struct amdkfd_process_info *process_info = p->kgd_process_info;
2900 	struct mm_struct *mm = current->mm;
2901 	struct list_head update_list;
2902 	struct list_head insert_list;
2903 	struct list_head remove_list;
2904 	struct svm_range_list *svms;
2905 	struct svm_range *prange;
2906 	struct svm_range *next;
2907 	int r = 0;
2908 
2909 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
2910 		 p->pasid, &p->svms, start, start + size - 1, size);
2911 
2912 	r = svm_range_check_attr(p, nattr, attrs);
2913 	if (r)
2914 		return r;
2915 
2916 	svms = &p->svms;
2917 
2918 	mutex_lock(&process_info->lock);
2919 
2920 	svm_range_list_lock_and_flush_work(svms, mm);
2921 
2922 	if (!svm_range_is_valid(mm, start, size)) {
2923 		pr_debug("invalid range\n");
2924 		r = -EFAULT;
2925 		mmap_write_unlock(mm);
2926 		goto out;
2927 	}
2928 
2929 	mutex_lock(&svms->lock);
2930 
2931 	/* Add new range and split existing ranges as needed */
2932 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
2933 			  &insert_list, &remove_list);
2934 	if (r) {
2935 		mutex_unlock(&svms->lock);
2936 		mmap_write_unlock(mm);
2937 		goto out;
2938 	}
2939 	/* Apply changes as a transaction */
2940 	list_for_each_entry_safe(prange, next, &insert_list, insert_list) {
2941 		svm_range_add_to_svms(prange);
2942 		svm_range_add_notifier_locked(mm, prange);
2943 	}
2944 	list_for_each_entry(prange, &update_list, update_list) {
2945 		svm_range_apply_attrs(p, prange, nattr, attrs);
2946 		/* TODO: unmap ranges from GPU that lost access */
2947 	}
2948 	list_for_each_entry_safe(prange, next, &remove_list,
2949 				remove_list) {
2950 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2951 			 prange->svms, prange, prange->start,
2952 			 prange->last);
2953 		svm_range_unlink(prange);
2954 		svm_range_remove_notifier(prange);
2955 		svm_range_free(prange);
2956 	}
2957 
2958 	mmap_write_downgrade(mm);
2959 	/* Trigger migrations and revalidate and map to GPUs as needed. If
2960 	 * this fails we may be left with partially completed actions. There
2961 	 * is no clean way of rolling back to the previous state in such a
2962 	 * case because the rollback wouldn't be guaranteed to work either.
2963 	 */
2964 	list_for_each_entry(prange, &update_list, update_list) {
2965 		bool migrated;
2966 
2967 		mutex_lock(&prange->migrate_mutex);
2968 
2969 		r = svm_range_trigger_migration(mm, prange, &migrated);
2970 		if (r)
2971 			goto out_unlock_range;
2972 
2973 		if (migrated && !p->xnack_enabled) {
2974 			pr_debug("restore_work will update mappings of GPUs\n");
2975 			mutex_unlock(&prange->migrate_mutex);
2976 			continue;
2977 		}
2978 
2979 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
2980 					       true, true);
2981 		if (r)
2982 			pr_debug("failed %d to map svm range\n", r);
2983 
2984 out_unlock_range:
2985 		mutex_unlock(&prange->migrate_mutex);
2986 		if (r)
2987 			break;
2988 	}
2989 
2990 	svm_range_debug_dump(svms);
2991 
2992 	mutex_unlock(&svms->lock);
2993 	mmap_read_unlock(mm);
2994 out:
2995 	mutex_unlock(&process_info->lock);
2996 
2997 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
2998 		 &p->svms, start, start + size - 1, r);
2999 
3000 	return r;
3001 }
3002 
3003 static int
3004 svm_range_get_attr(struct kfd_process *p, uint64_t start, uint64_t size,
3005 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
3006 {
3007 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3008 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3009 	bool get_preferred_loc = false;
3010 	bool get_prefetch_loc = false;
3011 	bool get_granularity = false;
3012 	bool get_accessible = false;
3013 	bool get_flags = false;
3014 	uint64_t last = start + size - 1UL;
3015 	struct mm_struct *mm = current->mm;
3016 	uint8_t granularity = 0xff;
3017 	struct interval_tree_node *node;
3018 	struct svm_range_list *svms;
3019 	struct svm_range *prange;
3020 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3021 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3022 	uint32_t flags = 0xffffffff;
3023 	int gpuidx;
3024 	uint32_t i;
3025 
3026 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3027 		 start + size - 1, nattr);
3028 
3029 	/* Flush pending deferred work to avoid racing with deferred actions from
3030 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3031 	 * can still race with get_attr because we don't hold the mmap lock. But that
3032 	 * would be a race condition in the application anyway, and undefined
3033 	 * behaviour is acceptable in that case.
3034 	 */
3035 	flush_work(&p->svms.deferred_list_work);
3036 
3037 	mmap_read_lock(mm);
3038 	if (!svm_range_is_valid(mm, start, size)) {
3039 		pr_debug("invalid range\n");
3040 		mmap_read_unlock(mm);
3041 		return -EINVAL;
3042 	}
3043 	mmap_read_unlock(mm);
3044 
3045 	for (i = 0; i < nattr; i++) {
3046 		switch (attrs[i].type) {
3047 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3048 			get_preferred_loc = true;
3049 			break;
3050 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3051 			get_prefetch_loc = true;
3052 			break;
3053 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3054 			get_accessible = true;
3055 			break;
3056 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3057 			get_flags = true;
3058 			break;
3059 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3060 			get_granularity = true;
3061 			break;
3062 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3063 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3064 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3065 			fallthrough;
3066 		default:
3067 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3068 			return -EINVAL;
3069 		}
3070 	}
3071 
3072 	svms = &p->svms;
3073 
3074 	mutex_lock(&svms->lock);
3075 
3076 	node = interval_tree_iter_first(&svms->objects, start, last);
3077 	if (!node) {
3078 		pr_debug("range attrs not found return default values\n");
3079 		svm_range_set_default_attributes(&location, &prefetch_loc,
3080 						 &granularity, &flags);
3081 		if (p->xnack_enabled)
3082 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3083 				    MAX_GPU_INSTANCE);
3084 		else
3085 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3086 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3087 		goto fill_values;
3088 	}
3089 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3090 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3091 
3092 	while (node) {
3093 		struct interval_tree_node *next;
3094 
3095 		prange = container_of(node, struct svm_range, it_node);
3096 		next = interval_tree_iter_next(node, start, last);
3097 
3098 		if (get_preferred_loc) {
3099 			if (prange->preferred_loc ==
3100 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3101 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3102 			     location != prange->preferred_loc)) {
3103 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3104 				get_preferred_loc = false;
3105 			} else {
3106 				location = prange->preferred_loc;
3107 			}
3108 		}
3109 		if (get_prefetch_loc) {
3110 			if (prange->prefetch_loc ==
3111 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3112 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3113 			     prefetch_loc != prange->prefetch_loc)) {
3114 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3115 				get_prefetch_loc = false;
3116 			} else {
3117 				prefetch_loc = prange->prefetch_loc;
3118 			}
3119 		}
3120 		if (get_accessible) {
3121 			bitmap_and(bitmap_access, bitmap_access,
3122 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3123 			bitmap_and(bitmap_aip, bitmap_aip,
3124 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3125 		}
3126 		if (get_flags)
3127 			flags &= prange->flags;
3128 
3129 		if (get_granularity && prange->granularity < granularity)
3130 			granularity = prange->granularity;
3131 
3132 		node = next;
3133 	}
3134 fill_values:
3135 	mutex_unlock(&svms->lock);
3136 
3137 	for (i = 0; i < nattr; i++) {
3138 		switch (attrs[i].type) {
3139 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3140 			attrs[i].value = location;
3141 			break;
3142 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3143 			attrs[i].value = prefetch_loc;
3144 			break;
3145 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3146 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3147 							       attrs[i].value);
3148 			if (gpuidx < 0) {
3149 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3150 				return -EINVAL;
3151 			}
3152 			if (test_bit(gpuidx, bitmap_access))
3153 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3154 			else if (test_bit(gpuidx, bitmap_aip))
3155 				attrs[i].type =
3156 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3157 			else
3158 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3159 			break;
3160 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3161 			attrs[i].value = flags;
3162 			break;
3163 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3164 			attrs[i].value = (uint32_t)granularity;
3165 			break;
3166 		}
3167 	}
3168 
3169 	return 0;
3170 }
3171 
3172 int
3173 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3174 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3175 {
3176 	int r;
3177 
3178 	start >>= PAGE_SHIFT;
3179 	size >>= PAGE_SHIFT;
3180 
3181 	switch (op) {
3182 	case KFD_IOCTL_SVM_OP_SET_ATTR:
3183 		r = svm_range_set_attr(p, start, size, nattrs, attrs);
3184 		break;
3185 	case KFD_IOCTL_SVM_OP_GET_ATTR:
3186 		r = svm_range_get_attr(p, start, size, nattrs, attrs);
3187 		break;
3188 	default:
3189 		r = EINVAL;
3190 		break;
3191 	}
3192 
3193 	return r;
3194 }
3195