xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision fca7216b)
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 #ifdef dev_fmt
37 #undef dev_fmt
38 #endif
39 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
40 
41 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
42 
43 /* Long enough to ensure no retry fault comes after svm range is restored and
44  * page table is updated.
45  */
46 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING	2000
47 
48 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
49 static bool
50 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
51 				    const struct mmu_notifier_range *range,
52 				    unsigned long cur_seq);
53 static int
54 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
55 		   uint64_t *bo_s, uint64_t *bo_l);
56 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
57 	.invalidate = svm_range_cpu_invalidate_pagetables,
58 };
59 
60 /**
61  * svm_range_unlink - unlink svm_range from lists and interval tree
62  * @prange: svm range structure to be removed
63  *
64  * Remove the svm_range from the svms and svm_bo lists and the svms
65  * interval tree.
66  *
67  * Context: The caller must hold svms->lock
68  */
69 static void svm_range_unlink(struct svm_range *prange)
70 {
71 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
72 		 prange, prange->start, prange->last);
73 
74 	if (prange->svm_bo) {
75 		spin_lock(&prange->svm_bo->list_lock);
76 		list_del(&prange->svm_bo_list);
77 		spin_unlock(&prange->svm_bo->list_lock);
78 	}
79 
80 	list_del(&prange->list);
81 	if (prange->it_node.start != 0 && prange->it_node.last != 0)
82 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
83 }
84 
85 static void
86 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
87 {
88 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
89 		 prange, prange->start, prange->last);
90 
91 	mmu_interval_notifier_insert_locked(&prange->notifier, mm,
92 				     prange->start << PAGE_SHIFT,
93 				     prange->npages << PAGE_SHIFT,
94 				     &svm_range_mn_ops);
95 }
96 
97 /**
98  * svm_range_add_to_svms - add svm range to svms
99  * @prange: svm range structure to be added
100  *
101  * Add the svm range to svms interval tree and link list
102  *
103  * Context: The caller must hold svms->lock
104  */
105 static void svm_range_add_to_svms(struct svm_range *prange)
106 {
107 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
108 		 prange, prange->start, prange->last);
109 
110 	list_add_tail(&prange->list, &prange->svms->list);
111 	prange->it_node.start = prange->start;
112 	prange->it_node.last = prange->last;
113 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
114 }
115 
116 static void svm_range_remove_notifier(struct svm_range *prange)
117 {
118 	pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
119 		 prange->svms, prange,
120 		 prange->notifier.interval_tree.start >> PAGE_SHIFT,
121 		 prange->notifier.interval_tree.last >> PAGE_SHIFT);
122 
123 	if (prange->notifier.interval_tree.start != 0 &&
124 	    prange->notifier.interval_tree.last != 0)
125 		mmu_interval_notifier_remove(&prange->notifier);
126 }
127 
128 static bool
129 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
130 {
131 	return dma_addr && !dma_mapping_error(dev, dma_addr) &&
132 	       !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
133 }
134 
135 static int
136 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
137 		      unsigned long offset, unsigned long npages,
138 		      unsigned long *hmm_pfns, uint32_t gpuidx)
139 {
140 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
141 	dma_addr_t *addr = prange->dma_addr[gpuidx];
142 	struct device *dev = adev->dev;
143 	struct page *page;
144 	int i, r;
145 
146 	if (!addr) {
147 		addr = kvmalloc_array(prange->npages, sizeof(*addr),
148 				      GFP_KERNEL | __GFP_ZERO);
149 		if (!addr)
150 			return -ENOMEM;
151 		prange->dma_addr[gpuidx] = addr;
152 	}
153 
154 	addr += offset;
155 	for (i = 0; i < npages; i++) {
156 		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
157 			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
158 
159 		page = hmm_pfn_to_page(hmm_pfns[i]);
160 		if (is_zone_device_page(page)) {
161 			struct amdgpu_device *bo_adev =
162 					amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
163 
164 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
165 				   bo_adev->vm_manager.vram_base_offset -
166 				   bo_adev->kfd.dev->pgmap.range.start;
167 			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
168 			pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
169 			continue;
170 		}
171 		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
172 		r = dma_mapping_error(dev, addr[i]);
173 		if (r) {
174 			dev_err(dev, "failed %d dma_map_page\n", r);
175 			return r;
176 		}
177 		pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
178 				     addr[i] >> PAGE_SHIFT, page_to_pfn(page));
179 	}
180 	return 0;
181 }
182 
183 static int
184 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
185 		  unsigned long offset, unsigned long npages,
186 		  unsigned long *hmm_pfns)
187 {
188 	struct kfd_process *p;
189 	uint32_t gpuidx;
190 	int r;
191 
192 	p = container_of(prange->svms, struct kfd_process, svms);
193 
194 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
195 		struct kfd_process_device *pdd;
196 		struct amdgpu_device *adev;
197 
198 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
199 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
200 		if (!pdd) {
201 			pr_debug("failed to find device idx %d\n", gpuidx);
202 			return -EINVAL;
203 		}
204 		adev = (struct amdgpu_device *)pdd->dev->kgd;
205 
206 		r = svm_range_dma_map_dev(adev, prange, offset, npages,
207 					  hmm_pfns, gpuidx);
208 		if (r)
209 			break;
210 	}
211 
212 	return r;
213 }
214 
215 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
216 			 unsigned long offset, unsigned long npages)
217 {
218 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
219 	int i;
220 
221 	if (!dma_addr)
222 		return;
223 
224 	for (i = offset; i < offset + npages; i++) {
225 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
226 			continue;
227 		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
228 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
229 		dma_addr[i] = 0;
230 	}
231 }
232 
233 void svm_range_free_dma_mappings(struct svm_range *prange)
234 {
235 	struct kfd_process_device *pdd;
236 	dma_addr_t *dma_addr;
237 	struct device *dev;
238 	struct kfd_process *p;
239 	uint32_t gpuidx;
240 
241 	p = container_of(prange->svms, struct kfd_process, svms);
242 
243 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
244 		dma_addr = prange->dma_addr[gpuidx];
245 		if (!dma_addr)
246 			continue;
247 
248 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
249 		if (!pdd) {
250 			pr_debug("failed to find device idx %d\n", gpuidx);
251 			continue;
252 		}
253 		dev = &pdd->dev->pdev->dev;
254 		svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
255 		kvfree(dma_addr);
256 		prange->dma_addr[gpuidx] = NULL;
257 	}
258 }
259 
260 static void svm_range_free(struct svm_range *prange)
261 {
262 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
263 		 prange->start, prange->last);
264 
265 	svm_range_vram_node_free(prange);
266 	svm_range_free_dma_mappings(prange);
267 	mutex_destroy(&prange->lock);
268 	mutex_destroy(&prange->migrate_mutex);
269 	kfree(prange);
270 }
271 
272 static void
273 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
274 				 uint8_t *granularity, uint32_t *flags)
275 {
276 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
277 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
278 	*granularity = 9;
279 	*flags =
280 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
281 }
282 
283 static struct
284 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
285 			 uint64_t last)
286 {
287 	uint64_t size = last - start + 1;
288 	struct svm_range *prange;
289 	struct kfd_process *p;
290 
291 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
292 	if (!prange)
293 		return NULL;
294 	prange->npages = size;
295 	prange->svms = svms;
296 	prange->start = start;
297 	prange->last = last;
298 	INIT_LIST_HEAD(&prange->list);
299 	INIT_LIST_HEAD(&prange->update_list);
300 	INIT_LIST_HEAD(&prange->remove_list);
301 	INIT_LIST_HEAD(&prange->insert_list);
302 	INIT_LIST_HEAD(&prange->svm_bo_list);
303 	INIT_LIST_HEAD(&prange->deferred_list);
304 	INIT_LIST_HEAD(&prange->child_list);
305 	atomic_set(&prange->invalid, 0);
306 	prange->validate_timestamp = 0;
307 	mutex_init(&prange->migrate_mutex);
308 	mutex_init(&prange->lock);
309 
310 	p = container_of(svms, struct kfd_process, svms);
311 	if (p->xnack_enabled)
312 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
313 			    MAX_GPU_INSTANCE);
314 
315 	svm_range_set_default_attributes(&prange->preferred_loc,
316 					 &prange->prefetch_loc,
317 					 &prange->granularity, &prange->flags);
318 
319 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
320 
321 	return prange;
322 }
323 
324 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
325 {
326 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
327 		return false;
328 
329 	return true;
330 }
331 
332 static void svm_range_bo_release(struct kref *kref)
333 {
334 	struct svm_range_bo *svm_bo;
335 
336 	svm_bo = container_of(kref, struct svm_range_bo, kref);
337 	spin_lock(&svm_bo->list_lock);
338 	while (!list_empty(&svm_bo->range_list)) {
339 		struct svm_range *prange =
340 				list_first_entry(&svm_bo->range_list,
341 						struct svm_range, svm_bo_list);
342 		/* list_del_init tells a concurrent svm_range_vram_node_new when
343 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
344 		 */
345 		list_del_init(&prange->svm_bo_list);
346 		spin_unlock(&svm_bo->list_lock);
347 
348 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
349 			 prange->start, prange->last);
350 		mutex_lock(&prange->lock);
351 		prange->svm_bo = NULL;
352 		mutex_unlock(&prange->lock);
353 
354 		spin_lock(&svm_bo->list_lock);
355 	}
356 	spin_unlock(&svm_bo->list_lock);
357 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
358 		/* We're not in the eviction worker.
359 		 * Signal the fence and synchronize with any
360 		 * pending eviction work.
361 		 */
362 		dma_fence_signal(&svm_bo->eviction_fence->base);
363 		cancel_work_sync(&svm_bo->eviction_work);
364 	}
365 	dma_fence_put(&svm_bo->eviction_fence->base);
366 	amdgpu_bo_unref(&svm_bo->bo);
367 	kfree(svm_bo);
368 }
369 
370 void svm_range_bo_unref(struct svm_range_bo *svm_bo)
371 {
372 	if (!svm_bo)
373 		return;
374 
375 	kref_put(&svm_bo->kref, svm_range_bo_release);
376 }
377 
378 static bool
379 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
380 {
381 	struct amdgpu_device *bo_adev;
382 
383 	mutex_lock(&prange->lock);
384 	if (!prange->svm_bo) {
385 		mutex_unlock(&prange->lock);
386 		return false;
387 	}
388 	if (prange->ttm_res) {
389 		/* We still have a reference, all is well */
390 		mutex_unlock(&prange->lock);
391 		return true;
392 	}
393 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
394 		/*
395 		 * Migrate from GPU to GPU, remove range from source bo_adev
396 		 * svm_bo range list, and return false to allocate svm_bo from
397 		 * destination adev.
398 		 */
399 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
400 		if (bo_adev != adev) {
401 			mutex_unlock(&prange->lock);
402 
403 			spin_lock(&prange->svm_bo->list_lock);
404 			list_del_init(&prange->svm_bo_list);
405 			spin_unlock(&prange->svm_bo->list_lock);
406 
407 			svm_range_bo_unref(prange->svm_bo);
408 			return false;
409 		}
410 		if (READ_ONCE(prange->svm_bo->evicting)) {
411 			struct dma_fence *f;
412 			struct svm_range_bo *svm_bo;
413 			/* The BO is getting evicted,
414 			 * we need to get a new one
415 			 */
416 			mutex_unlock(&prange->lock);
417 			svm_bo = prange->svm_bo;
418 			f = dma_fence_get(&svm_bo->eviction_fence->base);
419 			svm_range_bo_unref(prange->svm_bo);
420 			/* wait for the fence to avoid long spin-loop
421 			 * at list_empty_careful
422 			 */
423 			dma_fence_wait(f, false);
424 			dma_fence_put(f);
425 		} else {
426 			/* The BO was still around and we got
427 			 * a new reference to it
428 			 */
429 			mutex_unlock(&prange->lock);
430 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
431 				 prange->svms, prange->start, prange->last);
432 
433 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
434 			return true;
435 		}
436 
437 	} else {
438 		mutex_unlock(&prange->lock);
439 	}
440 
441 	/* We need a new svm_bo. Spin-loop to wait for concurrent
442 	 * svm_range_bo_release to finish removing this range from
443 	 * its range list. After this, it is safe to reuse the
444 	 * svm_bo pointer and svm_bo_list head.
445 	 */
446 	while (!list_empty_careful(&prange->svm_bo_list))
447 		;
448 
449 	return false;
450 }
451 
452 static struct svm_range_bo *svm_range_bo_new(void)
453 {
454 	struct svm_range_bo *svm_bo;
455 
456 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
457 	if (!svm_bo)
458 		return NULL;
459 
460 	kref_init(&svm_bo->kref);
461 	INIT_LIST_HEAD(&svm_bo->range_list);
462 	spin_lock_init(&svm_bo->list_lock);
463 
464 	return svm_bo;
465 }
466 
467 int
468 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
469 			bool clear)
470 {
471 	struct amdgpu_bo_param bp;
472 	struct svm_range_bo *svm_bo;
473 	struct amdgpu_bo_user *ubo;
474 	struct amdgpu_bo *bo;
475 	struct kfd_process *p;
476 	struct mm_struct *mm;
477 	int r;
478 
479 	p = container_of(prange->svms, struct kfd_process, svms);
480 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
481 		 prange->start, prange->last);
482 
483 	if (svm_range_validate_svm_bo(adev, prange))
484 		return 0;
485 
486 	svm_bo = svm_range_bo_new();
487 	if (!svm_bo) {
488 		pr_debug("failed to alloc svm bo\n");
489 		return -ENOMEM;
490 	}
491 	mm = get_task_mm(p->lead_thread);
492 	if (!mm) {
493 		pr_debug("failed to get mm\n");
494 		kfree(svm_bo);
495 		return -ESRCH;
496 	}
497 	svm_bo->svms = prange->svms;
498 	svm_bo->eviction_fence =
499 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
500 					   mm,
501 					   svm_bo);
502 	mmput(mm);
503 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
504 	svm_bo->evicting = 0;
505 	memset(&bp, 0, sizeof(bp));
506 	bp.size = prange->npages * PAGE_SIZE;
507 	bp.byte_align = PAGE_SIZE;
508 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
509 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
510 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
511 	bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO;
512 	bp.type = ttm_bo_type_device;
513 	bp.resv = NULL;
514 
515 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
516 	if (r) {
517 		pr_debug("failed %d to create bo\n", r);
518 		goto create_bo_failed;
519 	}
520 	bo = &ubo->bo;
521 	r = amdgpu_bo_reserve(bo, true);
522 	if (r) {
523 		pr_debug("failed %d to reserve bo\n", r);
524 		goto reserve_bo_failed;
525 	}
526 
527 	r = dma_resv_reserve_shared(bo->tbo.base.resv, 1);
528 	if (r) {
529 		pr_debug("failed %d to reserve bo\n", r);
530 		amdgpu_bo_unreserve(bo);
531 		goto reserve_bo_failed;
532 	}
533 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
534 
535 	amdgpu_bo_unreserve(bo);
536 
537 	svm_bo->bo = bo;
538 	prange->svm_bo = svm_bo;
539 	prange->ttm_res = bo->tbo.resource;
540 	prange->offset = 0;
541 
542 	spin_lock(&svm_bo->list_lock);
543 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
544 	spin_unlock(&svm_bo->list_lock);
545 
546 	return 0;
547 
548 reserve_bo_failed:
549 	amdgpu_bo_unref(&bo);
550 create_bo_failed:
551 	dma_fence_put(&svm_bo->eviction_fence->base);
552 	kfree(svm_bo);
553 	prange->ttm_res = NULL;
554 
555 	return r;
556 }
557 
558 void svm_range_vram_node_free(struct svm_range *prange)
559 {
560 	svm_range_bo_unref(prange->svm_bo);
561 	prange->ttm_res = NULL;
562 }
563 
564 struct amdgpu_device *
565 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
566 {
567 	struct kfd_process_device *pdd;
568 	struct kfd_process *p;
569 	int32_t gpu_idx;
570 
571 	p = container_of(prange->svms, struct kfd_process, svms);
572 
573 	gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
574 	if (gpu_idx < 0) {
575 		pr_debug("failed to get device by id 0x%x\n", gpu_id);
576 		return NULL;
577 	}
578 	pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
579 	if (!pdd) {
580 		pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
581 		return NULL;
582 	}
583 
584 	return (struct amdgpu_device *)pdd->dev->kgd;
585 }
586 
587 struct kfd_process_device *
588 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
589 {
590 	struct kfd_process *p;
591 	int32_t gpu_idx, gpuid;
592 	int r;
593 
594 	p = container_of(prange->svms, struct kfd_process, svms);
595 
596 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
597 	if (r) {
598 		pr_debug("failed to get device id by adev %p\n", adev);
599 		return NULL;
600 	}
601 
602 	return kfd_process_device_from_gpuidx(p, gpu_idx);
603 }
604 
605 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
606 {
607 	struct ttm_operation_ctx ctx = { false, false };
608 
609 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
610 
611 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
612 }
613 
614 static int
615 svm_range_check_attr(struct kfd_process *p,
616 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
617 {
618 	uint32_t i;
619 
620 	for (i = 0; i < nattr; i++) {
621 		uint32_t val = attrs[i].value;
622 		int gpuidx = MAX_GPU_INSTANCE;
623 
624 		switch (attrs[i].type) {
625 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
626 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
627 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
628 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
629 			break;
630 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
631 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
632 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
633 			break;
634 		case KFD_IOCTL_SVM_ATTR_ACCESS:
635 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
636 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
637 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
638 			break;
639 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
640 			break;
641 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
642 			break;
643 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
644 			break;
645 		default:
646 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
647 			return -EINVAL;
648 		}
649 
650 		if (gpuidx < 0) {
651 			pr_debug("no GPU 0x%x found\n", val);
652 			return -EINVAL;
653 		} else if (gpuidx < MAX_GPU_INSTANCE &&
654 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
655 			pr_debug("GPU 0x%x not supported\n", val);
656 			return -EINVAL;
657 		}
658 	}
659 
660 	return 0;
661 }
662 
663 static void
664 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
665 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
666 {
667 	uint32_t i;
668 	int gpuidx;
669 
670 	for (i = 0; i < nattr; i++) {
671 		switch (attrs[i].type) {
672 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
673 			prange->preferred_loc = attrs[i].value;
674 			break;
675 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
676 			prange->prefetch_loc = attrs[i].value;
677 			break;
678 		case KFD_IOCTL_SVM_ATTR_ACCESS:
679 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
680 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
681 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
682 							       attrs[i].value);
683 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
684 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
685 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
686 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
687 				bitmap_set(prange->bitmap_access, gpuidx, 1);
688 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
689 			} else {
690 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
691 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
692 			}
693 			break;
694 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
695 			prange->flags |= attrs[i].value;
696 			break;
697 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
698 			prange->flags &= ~attrs[i].value;
699 			break;
700 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
701 			prange->granularity = attrs[i].value;
702 			break;
703 		default:
704 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
705 		}
706 	}
707 }
708 
709 /**
710  * svm_range_debug_dump - print all range information from svms
711  * @svms: svm range list header
712  *
713  * debug output svm range start, end, prefetch location from svms
714  * interval tree and link list
715  *
716  * Context: The caller must hold svms->lock
717  */
718 static void svm_range_debug_dump(struct svm_range_list *svms)
719 {
720 	struct interval_tree_node *node;
721 	struct svm_range *prange;
722 
723 	pr_debug("dump svms 0x%p list\n", svms);
724 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
725 
726 	list_for_each_entry(prange, &svms->list, list) {
727 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
728 			 prange, prange->start, prange->npages,
729 			 prange->start + prange->npages - 1,
730 			 prange->actual_loc);
731 	}
732 
733 	pr_debug("dump svms 0x%p interval tree\n", svms);
734 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
735 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
736 	while (node) {
737 		prange = container_of(node, struct svm_range, it_node);
738 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
739 			 prange, prange->start, prange->npages,
740 			 prange->start + prange->npages - 1,
741 			 prange->actual_loc);
742 		node = interval_tree_iter_next(node, 0, ~0ULL);
743 	}
744 }
745 
746 static bool
747 svm_range_is_same_attrs(struct svm_range *old, struct svm_range *new)
748 {
749 	return (old->prefetch_loc == new->prefetch_loc &&
750 		old->flags == new->flags &&
751 		old->granularity == new->granularity);
752 }
753 
754 static int
755 svm_range_split_array(void *ppnew, void *ppold, size_t size,
756 		      uint64_t old_start, uint64_t old_n,
757 		      uint64_t new_start, uint64_t new_n)
758 {
759 	unsigned char *new, *old, *pold;
760 	uint64_t d;
761 
762 	if (!ppold)
763 		return 0;
764 	pold = *(unsigned char **)ppold;
765 	if (!pold)
766 		return 0;
767 
768 	new = kvmalloc_array(new_n, size, GFP_KERNEL);
769 	if (!new)
770 		return -ENOMEM;
771 
772 	d = (new_start - old_start) * size;
773 	memcpy(new, pold + d, new_n * size);
774 
775 	old = kvmalloc_array(old_n, size, GFP_KERNEL);
776 	if (!old) {
777 		kvfree(new);
778 		return -ENOMEM;
779 	}
780 
781 	d = (new_start == old_start) ? new_n * size : 0;
782 	memcpy(old, pold + d, old_n * size);
783 
784 	kvfree(pold);
785 	*(void **)ppold = old;
786 	*(void **)ppnew = new;
787 
788 	return 0;
789 }
790 
791 static int
792 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
793 		      uint64_t start, uint64_t last)
794 {
795 	uint64_t npages = last - start + 1;
796 	int i, r;
797 
798 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
799 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
800 					  sizeof(*old->dma_addr[i]), old->start,
801 					  npages, new->start, new->npages);
802 		if (r)
803 			return r;
804 	}
805 
806 	return 0;
807 }
808 
809 static int
810 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
811 		      uint64_t start, uint64_t last)
812 {
813 	uint64_t npages = last - start + 1;
814 
815 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
816 		 new->svms, new, new->start, start, last);
817 
818 	if (new->start == old->start) {
819 		new->offset = old->offset;
820 		old->offset += new->npages;
821 	} else {
822 		new->offset = old->offset + npages;
823 	}
824 
825 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
826 	new->ttm_res = old->ttm_res;
827 
828 	spin_lock(&new->svm_bo->list_lock);
829 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
830 	spin_unlock(&new->svm_bo->list_lock);
831 
832 	return 0;
833 }
834 
835 /**
836  * svm_range_split_adjust - split range and adjust
837  *
838  * @new: new range
839  * @old: the old range
840  * @start: the old range adjust to start address in pages
841  * @last: the old range adjust to last address in pages
842  *
843  * Copy system memory dma_addr or vram ttm_res in old range to new
844  * range from new_start up to size new->npages, the remaining old range is from
845  * start to last
846  *
847  * Return:
848  * 0 - OK, -ENOMEM - out of memory
849  */
850 static int
851 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
852 		      uint64_t start, uint64_t last)
853 {
854 	int r;
855 
856 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
857 		 new->svms, new->start, old->start, old->last, start, last);
858 
859 	if (new->start < old->start ||
860 	    new->last > old->last) {
861 		WARN_ONCE(1, "invalid new range start or last\n");
862 		return -EINVAL;
863 	}
864 
865 	r = svm_range_split_pages(new, old, start, last);
866 	if (r)
867 		return r;
868 
869 	if (old->actual_loc && old->ttm_res) {
870 		r = svm_range_split_nodes(new, old, start, last);
871 		if (r)
872 			return r;
873 	}
874 
875 	old->npages = last - start + 1;
876 	old->start = start;
877 	old->last = last;
878 	new->flags = old->flags;
879 	new->preferred_loc = old->preferred_loc;
880 	new->prefetch_loc = old->prefetch_loc;
881 	new->actual_loc = old->actual_loc;
882 	new->granularity = old->granularity;
883 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
884 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
885 
886 	return 0;
887 }
888 
889 /**
890  * svm_range_split - split a range in 2 ranges
891  *
892  * @prange: the svm range to split
893  * @start: the remaining range start address in pages
894  * @last: the remaining range last address in pages
895  * @new: the result new range generated
896  *
897  * Two cases only:
898  * case 1: if start == prange->start
899  *         prange ==> prange[start, last]
900  *         new range [last + 1, prange->last]
901  *
902  * case 2: if last == prange->last
903  *         prange ==> prange[start, last]
904  *         new range [prange->start, start - 1]
905  *
906  * Return:
907  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
908  */
909 static int
910 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
911 		struct svm_range **new)
912 {
913 	uint64_t old_start = prange->start;
914 	uint64_t old_last = prange->last;
915 	struct svm_range_list *svms;
916 	int r = 0;
917 
918 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
919 		 old_start, old_last, start, last);
920 
921 	if (old_start != start && old_last != last)
922 		return -EINVAL;
923 	if (start < old_start || last > old_last)
924 		return -EINVAL;
925 
926 	svms = prange->svms;
927 	if (old_start == start)
928 		*new = svm_range_new(svms, last + 1, old_last);
929 	else
930 		*new = svm_range_new(svms, old_start, start - 1);
931 	if (!*new)
932 		return -ENOMEM;
933 
934 	r = svm_range_split_adjust(*new, prange, start, last);
935 	if (r) {
936 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
937 			 r, old_start, old_last, start, last);
938 		svm_range_free(*new);
939 		*new = NULL;
940 	}
941 
942 	return r;
943 }
944 
945 static int
946 svm_range_split_tail(struct svm_range *prange, struct svm_range *new,
947 		     uint64_t new_last, struct list_head *insert_list)
948 {
949 	struct svm_range *tail;
950 	int r = svm_range_split(prange, prange->start, new_last, &tail);
951 
952 	if (!r)
953 		list_add(&tail->insert_list, insert_list);
954 	return r;
955 }
956 
957 static int
958 svm_range_split_head(struct svm_range *prange, struct svm_range *new,
959 		     uint64_t new_start, struct list_head *insert_list)
960 {
961 	struct svm_range *head;
962 	int r = svm_range_split(prange, new_start, prange->last, &head);
963 
964 	if (!r)
965 		list_add(&head->insert_list, insert_list);
966 	return r;
967 }
968 
969 static void
970 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
971 		    struct svm_range *pchild, enum svm_work_list_ops op)
972 {
973 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
974 		 pchild, pchild->start, pchild->last, prange, op);
975 
976 	pchild->work_item.mm = mm;
977 	pchild->work_item.op = op;
978 	list_add_tail(&pchild->child_list, &prange->child_list);
979 }
980 
981 /**
982  * svm_range_split_by_granularity - collect ranges within granularity boundary
983  *
984  * @p: the process with svms list
985  * @mm: mm structure
986  * @addr: the vm fault address in pages, to split the prange
987  * @parent: parent range if prange is from child list
988  * @prange: prange to split
989  *
990  * Trims @prange to be a single aligned block of prange->granularity if
991  * possible. The head and tail are added to the child_list in @parent.
992  *
993  * Context: caller must hold mmap_read_lock and prange->lock
994  *
995  * Return:
996  * 0 - OK, otherwise error code
997  */
998 int
999 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1000 			       unsigned long addr, struct svm_range *parent,
1001 			       struct svm_range *prange)
1002 {
1003 	struct svm_range *head, *tail;
1004 	unsigned long start, last, size;
1005 	int r;
1006 
1007 	/* Align splited range start and size to granularity size, then a single
1008 	 * PTE will be used for whole range, this reduces the number of PTE
1009 	 * updated and the L1 TLB space used for translation.
1010 	 */
1011 	size = 1UL << prange->granularity;
1012 	start = ALIGN_DOWN(addr, size);
1013 	last = ALIGN(addr + 1, size) - 1;
1014 
1015 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1016 		 prange->svms, prange->start, prange->last, start, last, size);
1017 
1018 	if (start > prange->start) {
1019 		r = svm_range_split(prange, start, prange->last, &head);
1020 		if (r)
1021 			return r;
1022 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1023 	}
1024 
1025 	if (last < prange->last) {
1026 		r = svm_range_split(prange, prange->start, last, &tail);
1027 		if (r)
1028 			return r;
1029 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1030 	}
1031 
1032 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1033 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1034 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1035 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1036 			 prange, prange->start, prange->last,
1037 			 SVM_OP_ADD_RANGE_AND_MAP);
1038 	}
1039 	return 0;
1040 }
1041 
1042 static uint64_t
1043 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1044 			int domain)
1045 {
1046 	struct amdgpu_device *bo_adev;
1047 	uint32_t flags = prange->flags;
1048 	uint32_t mapping_flags = 0;
1049 	uint64_t pte_flags;
1050 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1051 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1052 
1053 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1054 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1055 
1056 	switch (adev->asic_type) {
1057 	case CHIP_ARCTURUS:
1058 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1059 			if (bo_adev == adev) {
1060 				mapping_flags |= coherent ?
1061 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1062 			} else {
1063 				mapping_flags |= coherent ?
1064 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1065 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1066 					snoop = true;
1067 			}
1068 		} else {
1069 			mapping_flags |= coherent ?
1070 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1071 		}
1072 		break;
1073 	case CHIP_ALDEBARAN:
1074 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1075 			if (bo_adev == adev) {
1076 				mapping_flags |= coherent ?
1077 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1078 				if (adev->gmc.xgmi.connected_to_cpu)
1079 					snoop = true;
1080 			} else {
1081 				mapping_flags |= coherent ?
1082 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1083 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1084 					snoop = true;
1085 			}
1086 		} else {
1087 			mapping_flags |= coherent ?
1088 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1089 		}
1090 		break;
1091 	default:
1092 		mapping_flags |= coherent ?
1093 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1094 	}
1095 
1096 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1097 
1098 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1099 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1100 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1101 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1102 
1103 	pte_flags = AMDGPU_PTE_VALID;
1104 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1105 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1106 
1107 	pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1108 	return pte_flags;
1109 }
1110 
1111 static int
1112 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1113 			 uint64_t start, uint64_t last,
1114 			 struct dma_fence **fence)
1115 {
1116 	uint64_t init_pte_value = 0;
1117 
1118 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1119 
1120 	return amdgpu_vm_bo_update_mapping(adev, adev, vm, false, true, NULL,
1121 					   start, last, init_pte_value, 0,
1122 					   NULL, NULL, fence, NULL);
1123 }
1124 
1125 static int
1126 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1127 			  unsigned long last)
1128 {
1129 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1130 	struct kfd_process_device *pdd;
1131 	struct dma_fence *fence = NULL;
1132 	struct amdgpu_device *adev;
1133 	struct kfd_process *p;
1134 	uint32_t gpuidx;
1135 	int r = 0;
1136 
1137 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1138 		  MAX_GPU_INSTANCE);
1139 	p = container_of(prange->svms, struct kfd_process, svms);
1140 
1141 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1142 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1143 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1144 		if (!pdd) {
1145 			pr_debug("failed to find device idx %d\n", gpuidx);
1146 			return -EINVAL;
1147 		}
1148 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1149 
1150 		r = svm_range_unmap_from_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1151 					     start, last, &fence);
1152 		if (r)
1153 			break;
1154 
1155 		if (fence) {
1156 			r = dma_fence_wait(fence, false);
1157 			dma_fence_put(fence);
1158 			fence = NULL;
1159 			if (r)
1160 				break;
1161 		}
1162 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1163 					p->pasid, TLB_FLUSH_HEAVYWEIGHT);
1164 	}
1165 
1166 	return r;
1167 }
1168 
1169 static int
1170 svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1171 		     struct svm_range *prange, unsigned long offset,
1172 		     unsigned long npages, bool readonly, dma_addr_t *dma_addr,
1173 		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
1174 {
1175 	struct amdgpu_bo_va bo_va;
1176 	bool table_freed = false;
1177 	uint64_t pte_flags;
1178 	unsigned long last_start;
1179 	int last_domain;
1180 	int r = 0;
1181 	int64_t i, j;
1182 
1183 	last_start = prange->start + offset;
1184 
1185 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1186 		 last_start, last_start + npages - 1, readonly);
1187 
1188 	if (prange->svm_bo && prange->ttm_res)
1189 		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
1190 
1191 	for (i = offset; i < offset + npages; i++) {
1192 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1193 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1194 
1195 		/* Collect all pages in the same address range and memory domain
1196 		 * that can be mapped with a single call to update mapping.
1197 		 */
1198 		if (i < offset + npages - 1 &&
1199 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1200 			continue;
1201 
1202 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1203 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1204 
1205 		pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1206 		if (readonly)
1207 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1208 
1209 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1210 			 prange->svms, last_start, prange->start + i,
1211 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1212 			 pte_flags);
1213 
1214 		r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false,
1215 						NULL, last_start,
1216 						prange->start + i, pte_flags,
1217 						last_start - prange->start,
1218 						NULL, dma_addr,
1219 						&vm->last_update,
1220 						&table_freed);
1221 
1222 		for (j = last_start - prange->start; j <= i; j++)
1223 			dma_addr[j] |= last_domain;
1224 
1225 		if (r) {
1226 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1227 			goto out;
1228 		}
1229 		last_start = prange->start + i + 1;
1230 	}
1231 
1232 	r = amdgpu_vm_update_pdes(adev, vm, false);
1233 	if (r) {
1234 		pr_debug("failed %d to update directories 0x%lx\n", r,
1235 			 prange->start);
1236 		goto out;
1237 	}
1238 
1239 	if (fence)
1240 		*fence = dma_fence_get(vm->last_update);
1241 
1242 	if (table_freed) {
1243 		struct kfd_process *p;
1244 
1245 		p = container_of(prange->svms, struct kfd_process, svms);
1246 		amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
1247 						p->pasid, TLB_FLUSH_LEGACY);
1248 	}
1249 out:
1250 	return r;
1251 }
1252 
1253 static int
1254 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1255 		      unsigned long npages, bool readonly,
1256 		      unsigned long *bitmap, bool wait)
1257 {
1258 	struct kfd_process_device *pdd;
1259 	struct amdgpu_device *bo_adev;
1260 	struct amdgpu_device *adev;
1261 	struct kfd_process *p;
1262 	struct dma_fence *fence = NULL;
1263 	uint32_t gpuidx;
1264 	int r = 0;
1265 
1266 	if (prange->svm_bo && prange->ttm_res)
1267 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1268 	else
1269 		bo_adev = NULL;
1270 
1271 	p = container_of(prange->svms, struct kfd_process, svms);
1272 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1273 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1274 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1275 		if (!pdd) {
1276 			pr_debug("failed to find device idx %d\n", gpuidx);
1277 			return -EINVAL;
1278 		}
1279 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1280 
1281 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1282 		if (IS_ERR(pdd))
1283 			return -EINVAL;
1284 
1285 		if (bo_adev && adev != bo_adev &&
1286 		    !amdgpu_xgmi_same_hive(adev, bo_adev)) {
1287 			pr_debug("cannot map to device idx %d\n", gpuidx);
1288 			continue;
1289 		}
1290 
1291 		r = svm_range_map_to_gpu(adev, drm_priv_to_vm(pdd->drm_priv),
1292 					 prange, offset, npages, readonly,
1293 					 prange->dma_addr[gpuidx],
1294 					 bo_adev, wait ? &fence : NULL);
1295 		if (r)
1296 			break;
1297 
1298 		if (fence) {
1299 			r = dma_fence_wait(fence, false);
1300 			dma_fence_put(fence);
1301 			fence = NULL;
1302 			if (r) {
1303 				pr_debug("failed %d to dma fence wait\n", r);
1304 				break;
1305 			}
1306 		}
1307 	}
1308 
1309 	return r;
1310 }
1311 
1312 struct svm_validate_context {
1313 	struct kfd_process *process;
1314 	struct svm_range *prange;
1315 	bool intr;
1316 	unsigned long bitmap[MAX_GPU_INSTANCE];
1317 	struct ttm_validate_buffer tv[MAX_GPU_INSTANCE];
1318 	struct list_head validate_list;
1319 	struct ww_acquire_ctx ticket;
1320 };
1321 
1322 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1323 {
1324 	struct kfd_process_device *pdd;
1325 	struct amdgpu_device *adev;
1326 	struct amdgpu_vm *vm;
1327 	uint32_t gpuidx;
1328 	int r;
1329 
1330 	INIT_LIST_HEAD(&ctx->validate_list);
1331 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1332 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1333 		if (!pdd) {
1334 			pr_debug("failed to find device idx %d\n", gpuidx);
1335 			return -EINVAL;
1336 		}
1337 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1338 		vm = drm_priv_to_vm(pdd->drm_priv);
1339 
1340 		ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1341 		ctx->tv[gpuidx].num_shared = 4;
1342 		list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1343 	}
1344 
1345 	r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1346 				   ctx->intr, NULL);
1347 	if (r) {
1348 		pr_debug("failed %d to reserve bo\n", r);
1349 		return r;
1350 	}
1351 
1352 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1353 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1354 		if (!pdd) {
1355 			pr_debug("failed to find device idx %d\n", gpuidx);
1356 			r = -EINVAL;
1357 			goto unreserve_out;
1358 		}
1359 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1360 
1361 		r = amdgpu_vm_validate_pt_bos(adev, drm_priv_to_vm(pdd->drm_priv),
1362 					      svm_range_bo_validate, NULL);
1363 		if (r) {
1364 			pr_debug("failed %d validate pt bos\n", r);
1365 			goto unreserve_out;
1366 		}
1367 	}
1368 
1369 	return 0;
1370 
1371 unreserve_out:
1372 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1373 	return r;
1374 }
1375 
1376 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1377 {
1378 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1379 }
1380 
1381 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1382 {
1383 	struct kfd_process_device *pdd;
1384 	struct amdgpu_device *adev;
1385 
1386 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1387 	adev = (struct amdgpu_device *)pdd->dev->kgd;
1388 
1389 	return SVM_ADEV_PGMAP_OWNER(adev);
1390 }
1391 
1392 /*
1393  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1394  *
1395  * To prevent concurrent destruction or change of range attributes, the
1396  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1397  * because that would block concurrent evictions and lead to deadlocks. To
1398  * serialize concurrent migrations or validations of the same range, the
1399  * prange->migrate_mutex must be held.
1400  *
1401  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1402  * eviction fence.
1403  *
1404  * The following sequence ensures race-free validation and GPU mapping:
1405  *
1406  * 1. Reserve page table (and SVM BO if range is in VRAM)
1407  * 2. hmm_range_fault to get page addresses (if system memory)
1408  * 3. DMA-map pages (if system memory)
1409  * 4-a. Take notifier lock
1410  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1411  * 4-c. Check that the range was not split or otherwise invalidated
1412  * 4-d. Update GPU page table
1413  * 4.e. Release notifier lock
1414  * 5. Release page table (and SVM BO) reservation
1415  */
1416 static int svm_range_validate_and_map(struct mm_struct *mm,
1417 				      struct svm_range *prange,
1418 				      int32_t gpuidx, bool intr, bool wait)
1419 {
1420 	struct svm_validate_context ctx;
1421 	unsigned long start, end, addr;
1422 	struct kfd_process *p;
1423 	void *owner;
1424 	int32_t idx;
1425 	int r = 0;
1426 
1427 	ctx.process = container_of(prange->svms, struct kfd_process, svms);
1428 	ctx.prange = prange;
1429 	ctx.intr = intr;
1430 
1431 	if (gpuidx < MAX_GPU_INSTANCE) {
1432 		bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1433 		bitmap_set(ctx.bitmap, gpuidx, 1);
1434 	} else if (ctx.process->xnack_enabled) {
1435 		bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1436 
1437 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1438 		 * GPU, which has ACCESS attribute to the range, create mapping
1439 		 * on that GPU.
1440 		 */
1441 		if (prange->actual_loc) {
1442 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1443 							prange->actual_loc);
1444 			if (gpuidx < 0) {
1445 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1446 					 prange->actual_loc);
1447 				return -EINVAL;
1448 			}
1449 			if (test_bit(gpuidx, prange->bitmap_access))
1450 				bitmap_set(ctx.bitmap, gpuidx, 1);
1451 		}
1452 	} else {
1453 		bitmap_or(ctx.bitmap, prange->bitmap_access,
1454 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1455 	}
1456 
1457 	if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE))
1458 		return 0;
1459 
1460 	if (prange->actual_loc && !prange->ttm_res) {
1461 		/* This should never happen. actual_loc gets set by
1462 		 * svm_migrate_ram_to_vram after allocating a BO.
1463 		 */
1464 		WARN_ONCE(1, "VRAM BO missing during validation\n");
1465 		return -EINVAL;
1466 	}
1467 
1468 	svm_range_reserve_bos(&ctx);
1469 
1470 	p = container_of(prange->svms, struct kfd_process, svms);
1471 	owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1472 						MAX_GPU_INSTANCE));
1473 	for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1474 		if (kfd_svm_page_owner(p, idx) != owner) {
1475 			owner = NULL;
1476 			break;
1477 		}
1478 	}
1479 
1480 	start = prange->start << PAGE_SHIFT;
1481 	end = (prange->last + 1) << PAGE_SHIFT;
1482 	for (addr = start; addr < end && !r; ) {
1483 		struct hmm_range *hmm_range;
1484 		struct vm_area_struct *vma;
1485 		unsigned long next;
1486 		unsigned long offset;
1487 		unsigned long npages;
1488 		bool readonly;
1489 
1490 		vma = find_vma(mm, addr);
1491 		if (!vma || addr < vma->vm_start) {
1492 			r = -EFAULT;
1493 			goto unreserve_out;
1494 		}
1495 		readonly = !(vma->vm_flags & VM_WRITE);
1496 
1497 		next = min(vma->vm_end, end);
1498 		npages = (next - addr) >> PAGE_SHIFT;
1499 		WRITE_ONCE(p->svms.faulting_task, current);
1500 		r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1501 					       addr, npages, &hmm_range,
1502 					       readonly, true, owner);
1503 		WRITE_ONCE(p->svms.faulting_task, NULL);
1504 		if (r) {
1505 			pr_debug("failed %d to get svm range pages\n", r);
1506 			goto unreserve_out;
1507 		}
1508 
1509 		offset = (addr - start) >> PAGE_SHIFT;
1510 		r = svm_range_dma_map(prange, ctx.bitmap, offset, npages,
1511 				      hmm_range->hmm_pfns);
1512 		if (r) {
1513 			pr_debug("failed %d to dma map range\n", r);
1514 			goto unreserve_out;
1515 		}
1516 
1517 		svm_range_lock(prange);
1518 		if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1519 			pr_debug("hmm update the range, need validate again\n");
1520 			r = -EAGAIN;
1521 			goto unlock_out;
1522 		}
1523 		if (!list_empty(&prange->child_list)) {
1524 			pr_debug("range split by unmap in parallel, validate again\n");
1525 			r = -EAGAIN;
1526 			goto unlock_out;
1527 		}
1528 
1529 		r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1530 					  ctx.bitmap, wait);
1531 
1532 unlock_out:
1533 		svm_range_unlock(prange);
1534 
1535 		addr = next;
1536 	}
1537 
1538 	if (addr == end)
1539 		prange->validated_once = true;
1540 
1541 unreserve_out:
1542 	svm_range_unreserve_bos(&ctx);
1543 
1544 	if (!r)
1545 		prange->validate_timestamp = ktime_to_us(ktime_get());
1546 
1547 	return r;
1548 }
1549 
1550 /**
1551  * svm_range_list_lock_and_flush_work - flush pending deferred work
1552  *
1553  * @svms: the svm range list
1554  * @mm: the mm structure
1555  *
1556  * Context: Returns with mmap write lock held, pending deferred work flushed
1557  *
1558  */
1559 void
1560 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1561 				   struct mm_struct *mm)
1562 {
1563 retry_flush_work:
1564 	flush_work(&svms->deferred_list_work);
1565 	mmap_write_lock(mm);
1566 
1567 	if (list_empty(&svms->deferred_range_list))
1568 		return;
1569 	mmap_write_unlock(mm);
1570 	pr_debug("retry flush\n");
1571 	goto retry_flush_work;
1572 }
1573 
1574 static void svm_range_restore_work(struct work_struct *work)
1575 {
1576 	struct delayed_work *dwork = to_delayed_work(work);
1577 	struct svm_range_list *svms;
1578 	struct svm_range *prange;
1579 	struct kfd_process *p;
1580 	struct mm_struct *mm;
1581 	int evicted_ranges;
1582 	int invalid;
1583 	int r;
1584 
1585 	svms = container_of(dwork, struct svm_range_list, restore_work);
1586 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1587 	if (!evicted_ranges)
1588 		return;
1589 
1590 	pr_debug("restore svm ranges\n");
1591 
1592 	/* kfd_process_notifier_release destroys this worker thread. So during
1593 	 * the lifetime of this thread, kfd_process and mm will be valid.
1594 	 */
1595 	p = container_of(svms, struct kfd_process, svms);
1596 	mm = p->mm;
1597 	if (!mm)
1598 		return;
1599 
1600 	svm_range_list_lock_and_flush_work(svms, mm);
1601 	mutex_lock(&svms->lock);
1602 
1603 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1604 
1605 	list_for_each_entry(prange, &svms->list, list) {
1606 		invalid = atomic_read(&prange->invalid);
1607 		if (!invalid)
1608 			continue;
1609 
1610 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1611 			 prange->svms, prange, prange->start, prange->last,
1612 			 invalid);
1613 
1614 		/*
1615 		 * If range is migrating, wait for migration is done.
1616 		 */
1617 		mutex_lock(&prange->migrate_mutex);
1618 
1619 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1620 					       false, true);
1621 		if (r)
1622 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1623 				 prange->start);
1624 
1625 		mutex_unlock(&prange->migrate_mutex);
1626 		if (r)
1627 			goto out_reschedule;
1628 
1629 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1630 			goto out_reschedule;
1631 	}
1632 
1633 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1634 	    evicted_ranges)
1635 		goto out_reschedule;
1636 
1637 	evicted_ranges = 0;
1638 
1639 	r = kgd2kfd_resume_mm(mm);
1640 	if (r) {
1641 		/* No recovery from this failure. Probably the CP is
1642 		 * hanging. No point trying again.
1643 		 */
1644 		pr_debug("failed %d to resume KFD\n", r);
1645 	}
1646 
1647 	pr_debug("restore svm ranges successfully\n");
1648 
1649 out_reschedule:
1650 	mutex_unlock(&svms->lock);
1651 	mmap_write_unlock(mm);
1652 
1653 	/* If validation failed, reschedule another attempt */
1654 	if (evicted_ranges) {
1655 		pr_debug("reschedule to restore svm range\n");
1656 		schedule_delayed_work(&svms->restore_work,
1657 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1658 	}
1659 }
1660 
1661 /**
1662  * svm_range_evict - evict svm range
1663  *
1664  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1665  * return to let CPU evict the buffer and proceed CPU pagetable update.
1666  *
1667  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1668  * If invalidation happens while restore work is running, restore work will
1669  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1670  * the queues.
1671  */
1672 static int
1673 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1674 		unsigned long start, unsigned long last)
1675 {
1676 	struct svm_range_list *svms = prange->svms;
1677 	struct svm_range *pchild;
1678 	struct kfd_process *p;
1679 	int r = 0;
1680 
1681 	p = container_of(svms, struct kfd_process, svms);
1682 
1683 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1684 		 svms, prange->start, prange->last, start, last);
1685 
1686 	if (!p->xnack_enabled) {
1687 		int evicted_ranges;
1688 
1689 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1690 			mutex_lock_nested(&pchild->lock, 1);
1691 			if (pchild->start <= last && pchild->last >= start) {
1692 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1693 					 pchild->start, pchild->last);
1694 				atomic_inc(&pchild->invalid);
1695 			}
1696 			mutex_unlock(&pchild->lock);
1697 		}
1698 
1699 		if (prange->start <= last && prange->last >= start)
1700 			atomic_inc(&prange->invalid);
1701 
1702 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1703 		if (evicted_ranges != 1)
1704 			return r;
1705 
1706 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1707 			 prange->svms, prange->start, prange->last);
1708 
1709 		/* First eviction, stop the queues */
1710 		r = kgd2kfd_quiesce_mm(mm);
1711 		if (r)
1712 			pr_debug("failed to quiesce KFD\n");
1713 
1714 		pr_debug("schedule to restore svm %p ranges\n", svms);
1715 		schedule_delayed_work(&svms->restore_work,
1716 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1717 	} else {
1718 		unsigned long s, l;
1719 
1720 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1721 			 prange->svms, start, last);
1722 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1723 			mutex_lock_nested(&pchild->lock, 1);
1724 			s = max(start, pchild->start);
1725 			l = min(last, pchild->last);
1726 			if (l >= s)
1727 				svm_range_unmap_from_gpus(pchild, s, l);
1728 			mutex_unlock(&pchild->lock);
1729 		}
1730 		s = max(start, prange->start);
1731 		l = min(last, prange->last);
1732 		if (l >= s)
1733 			svm_range_unmap_from_gpus(prange, s, l);
1734 	}
1735 
1736 	return r;
1737 }
1738 
1739 static struct svm_range *svm_range_clone(struct svm_range *old)
1740 {
1741 	struct svm_range *new;
1742 
1743 	new = svm_range_new(old->svms, old->start, old->last);
1744 	if (!new)
1745 		return NULL;
1746 
1747 	if (old->svm_bo) {
1748 		new->ttm_res = old->ttm_res;
1749 		new->offset = old->offset;
1750 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1751 		spin_lock(&new->svm_bo->list_lock);
1752 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1753 		spin_unlock(&new->svm_bo->list_lock);
1754 	}
1755 	new->flags = old->flags;
1756 	new->preferred_loc = old->preferred_loc;
1757 	new->prefetch_loc = old->prefetch_loc;
1758 	new->actual_loc = old->actual_loc;
1759 	new->granularity = old->granularity;
1760 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1761 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1762 
1763 	return new;
1764 }
1765 
1766 /**
1767  * svm_range_handle_overlap - split overlap ranges
1768  * @svms: svm range list header
1769  * @new: range added with this attributes
1770  * @start: range added start address, in pages
1771  * @last: range last address, in pages
1772  * @update_list: output, the ranges attributes are updated. For set_attr, this
1773  *               will do validation and map to GPUs. For unmap, this will be
1774  *               removed and unmap from GPUs
1775  * @insert_list: output, the ranges will be inserted into svms, attributes are
1776  *               not changes. For set_attr, this will add into svms.
1777  * @remove_list:output, the ranges will be removed from svms
1778  * @left: the remaining range after overlap, For set_attr, this will be added
1779  *        as new range.
1780  *
1781  * Total have 5 overlap cases.
1782  *
1783  * This function handles overlap of an address interval with existing
1784  * struct svm_ranges for applying new attributes. This may require
1785  * splitting existing struct svm_ranges. All changes should be applied to
1786  * the range_list and interval tree transactionally. If any split operation
1787  * fails, the entire update fails. Therefore the existing overlapping
1788  * svm_ranges are cloned and the original svm_ranges left unchanged. If the
1789  * transaction succeeds, the modified clones are added and the originals
1790  * freed. Otherwise the clones are removed and the old svm_ranges remain.
1791  *
1792  * Context: The caller must hold svms->lock
1793  */
1794 static int
1795 svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
1796 			 unsigned long start, unsigned long last,
1797 			 struct list_head *update_list,
1798 			 struct list_head *insert_list,
1799 			 struct list_head *remove_list,
1800 			 unsigned long *left)
1801 {
1802 	struct interval_tree_node *node;
1803 	struct svm_range *prange;
1804 	struct svm_range *tmp;
1805 	int r = 0;
1806 
1807 	INIT_LIST_HEAD(update_list);
1808 	INIT_LIST_HEAD(insert_list);
1809 	INIT_LIST_HEAD(remove_list);
1810 
1811 	node = interval_tree_iter_first(&svms->objects, start, last);
1812 	while (node) {
1813 		struct interval_tree_node *next;
1814 		struct svm_range *old;
1815 		unsigned long next_start;
1816 
1817 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
1818 			 node->last);
1819 
1820 		old = container_of(node, struct svm_range, it_node);
1821 		next = interval_tree_iter_next(node, start, last);
1822 		next_start = min(node->last, last) + 1;
1823 
1824 		if (node->start < start || node->last > last) {
1825 			/* node intersects the updated range, clone+split it */
1826 			prange = svm_range_clone(old);
1827 			if (!prange) {
1828 				r = -ENOMEM;
1829 				goto out;
1830 			}
1831 
1832 			list_add(&old->remove_list, remove_list);
1833 			list_add(&prange->insert_list, insert_list);
1834 
1835 			if (node->start < start) {
1836 				pr_debug("change old range start\n");
1837 				r = svm_range_split_head(prange, new, start,
1838 							 insert_list);
1839 				if (r)
1840 					goto out;
1841 			}
1842 			if (node->last > last) {
1843 				pr_debug("change old range last\n");
1844 				r = svm_range_split_tail(prange, new, last,
1845 							 insert_list);
1846 				if (r)
1847 					goto out;
1848 			}
1849 		} else {
1850 			/* The node is contained within start..last,
1851 			 * just update it
1852 			 */
1853 			prange = old;
1854 		}
1855 
1856 		if (!svm_range_is_same_attrs(prange, new))
1857 			list_add(&prange->update_list, update_list);
1858 
1859 		/* insert a new node if needed */
1860 		if (node->start > start) {
1861 			prange = svm_range_new(prange->svms, start,
1862 					       node->start - 1);
1863 			if (!prange) {
1864 				r = -ENOMEM;
1865 				goto out;
1866 			}
1867 
1868 			list_add(&prange->insert_list, insert_list);
1869 			list_add(&prange->update_list, update_list);
1870 		}
1871 
1872 		node = next;
1873 		start = next_start;
1874 	}
1875 
1876 	if (left && start <= last)
1877 		*left = last - start + 1;
1878 
1879 out:
1880 	if (r)
1881 		list_for_each_entry_safe(prange, tmp, insert_list, insert_list)
1882 			svm_range_free(prange);
1883 
1884 	return r;
1885 }
1886 
1887 static void
1888 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
1889 					    struct svm_range *prange)
1890 {
1891 	unsigned long start;
1892 	unsigned long last;
1893 
1894 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
1895 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
1896 
1897 	if (prange->start == start && prange->last == last)
1898 		return;
1899 
1900 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1901 		  prange->svms, prange, start, last, prange->start,
1902 		  prange->last);
1903 
1904 	if (start != 0 && last != 0) {
1905 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
1906 		svm_range_remove_notifier(prange);
1907 	}
1908 	prange->it_node.start = prange->start;
1909 	prange->it_node.last = prange->last;
1910 
1911 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
1912 	svm_range_add_notifier_locked(mm, prange);
1913 }
1914 
1915 static void
1916 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange)
1917 {
1918 	struct mm_struct *mm = prange->work_item.mm;
1919 
1920 	switch (prange->work_item.op) {
1921 	case SVM_OP_NULL:
1922 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1923 			 svms, prange, prange->start, prange->last);
1924 		break;
1925 	case SVM_OP_UNMAP_RANGE:
1926 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1927 			 svms, prange, prange->start, prange->last);
1928 		svm_range_unlink(prange);
1929 		svm_range_remove_notifier(prange);
1930 		svm_range_free(prange);
1931 		break;
1932 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
1933 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1934 			 svms, prange, prange->start, prange->last);
1935 		svm_range_update_notifier_and_interval_tree(mm, prange);
1936 		break;
1937 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
1938 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1939 			 svms, prange, prange->start, prange->last);
1940 		svm_range_update_notifier_and_interval_tree(mm, prange);
1941 		/* TODO: implement deferred validation and mapping */
1942 		break;
1943 	case SVM_OP_ADD_RANGE:
1944 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
1945 			 prange->start, prange->last);
1946 		svm_range_add_to_svms(prange);
1947 		svm_range_add_notifier_locked(mm, prange);
1948 		break;
1949 	case SVM_OP_ADD_RANGE_AND_MAP:
1950 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
1951 			 prange, prange->start, prange->last);
1952 		svm_range_add_to_svms(prange);
1953 		svm_range_add_notifier_locked(mm, prange);
1954 		/* TODO: implement deferred validation and mapping */
1955 		break;
1956 	default:
1957 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
1958 			 prange->work_item.op);
1959 	}
1960 }
1961 
1962 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
1963 {
1964 	struct kfd_process_device *pdd;
1965 	struct amdgpu_device *adev;
1966 	struct kfd_process *p;
1967 	int drain;
1968 	uint32_t i;
1969 
1970 	p = container_of(svms, struct kfd_process, svms);
1971 
1972 restart:
1973 	drain = atomic_read(&svms->drain_pagefaults);
1974 	if (!drain)
1975 		return;
1976 
1977 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
1978 		pdd = p->pdds[i];
1979 		if (!pdd)
1980 			continue;
1981 
1982 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
1983 		adev = (struct amdgpu_device *)pdd->dev->kgd;
1984 
1985 		amdgpu_ih_wait_on_checkpoint_process(adev, &adev->irq.ih1);
1986 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
1987 	}
1988 	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
1989 		goto restart;
1990 }
1991 
1992 static void svm_range_deferred_list_work(struct work_struct *work)
1993 {
1994 	struct svm_range_list *svms;
1995 	struct svm_range *prange;
1996 	struct mm_struct *mm;
1997 	struct kfd_process *p;
1998 
1999 	svms = container_of(work, struct svm_range_list, deferred_list_work);
2000 	pr_debug("enter svms 0x%p\n", svms);
2001 
2002 	p = container_of(svms, struct kfd_process, svms);
2003 	/* Avoid mm is gone when inserting mmu notifier */
2004 	mm = get_task_mm(p->lead_thread);
2005 	if (!mm) {
2006 		pr_debug("svms 0x%p process mm gone\n", svms);
2007 		return;
2008 	}
2009 retry:
2010 	mmap_write_lock(mm);
2011 
2012 	/* Checking for the need to drain retry faults must be inside
2013 	 * mmap write lock to serialize with munmap notifiers.
2014 	 */
2015 	if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2016 		mmap_write_unlock(mm);
2017 		svm_range_drain_retry_fault(svms);
2018 		goto retry;
2019 	}
2020 
2021 	spin_lock(&svms->deferred_list_lock);
2022 	while (!list_empty(&svms->deferred_range_list)) {
2023 		prange = list_first_entry(&svms->deferred_range_list,
2024 					  struct svm_range, deferred_list);
2025 		list_del_init(&prange->deferred_list);
2026 		spin_unlock(&svms->deferred_list_lock);
2027 
2028 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2029 			 prange->start, prange->last, prange->work_item.op);
2030 
2031 		mutex_lock(&svms->lock);
2032 		mutex_lock(&prange->migrate_mutex);
2033 		while (!list_empty(&prange->child_list)) {
2034 			struct svm_range *pchild;
2035 
2036 			pchild = list_first_entry(&prange->child_list,
2037 						struct svm_range, child_list);
2038 			pr_debug("child prange 0x%p op %d\n", pchild,
2039 				 pchild->work_item.op);
2040 			list_del_init(&pchild->child_list);
2041 			svm_range_handle_list_op(svms, pchild);
2042 		}
2043 		mutex_unlock(&prange->migrate_mutex);
2044 
2045 		svm_range_handle_list_op(svms, prange);
2046 		mutex_unlock(&svms->lock);
2047 
2048 		spin_lock(&svms->deferred_list_lock);
2049 	}
2050 	spin_unlock(&svms->deferred_list_lock);
2051 
2052 	mmap_write_unlock(mm);
2053 	mmput(mm);
2054 	pr_debug("exit svms 0x%p\n", svms);
2055 }
2056 
2057 void
2058 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2059 			struct mm_struct *mm, enum svm_work_list_ops op)
2060 {
2061 	spin_lock(&svms->deferred_list_lock);
2062 	/* if prange is on the deferred list */
2063 	if (!list_empty(&prange->deferred_list)) {
2064 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2065 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2066 		if (op != SVM_OP_NULL &&
2067 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2068 			prange->work_item.op = op;
2069 	} else {
2070 		prange->work_item.op = op;
2071 		prange->work_item.mm = mm;
2072 		list_add_tail(&prange->deferred_list,
2073 			      &prange->svms->deferred_range_list);
2074 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2075 			 prange, prange->start, prange->last, op);
2076 	}
2077 	spin_unlock(&svms->deferred_list_lock);
2078 }
2079 
2080 void schedule_deferred_list_work(struct svm_range_list *svms)
2081 {
2082 	spin_lock(&svms->deferred_list_lock);
2083 	if (!list_empty(&svms->deferred_range_list))
2084 		schedule_work(&svms->deferred_list_work);
2085 	spin_unlock(&svms->deferred_list_lock);
2086 }
2087 
2088 static void
2089 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2090 		      struct svm_range *prange, unsigned long start,
2091 		      unsigned long last)
2092 {
2093 	struct svm_range *head;
2094 	struct svm_range *tail;
2095 
2096 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2097 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2098 			 prange->start, prange->last);
2099 		return;
2100 	}
2101 	if (start > prange->last || last < prange->start)
2102 		return;
2103 
2104 	head = tail = prange;
2105 	if (start > prange->start)
2106 		svm_range_split(prange, prange->start, start - 1, &tail);
2107 	if (last < tail->last)
2108 		svm_range_split(tail, last + 1, tail->last, &head);
2109 
2110 	if (head != prange && tail != prange) {
2111 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2112 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2113 	} else if (tail != prange) {
2114 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2115 	} else if (head != prange) {
2116 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2117 	} else if (parent != prange) {
2118 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2119 	}
2120 }
2121 
2122 static void
2123 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2124 			 unsigned long start, unsigned long last)
2125 {
2126 	struct svm_range_list *svms;
2127 	struct svm_range *pchild;
2128 	struct kfd_process *p;
2129 	unsigned long s, l;
2130 	bool unmap_parent;
2131 
2132 	p = kfd_lookup_process_by_mm(mm);
2133 	if (!p)
2134 		return;
2135 	svms = &p->svms;
2136 
2137 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2138 		 prange, prange->start, prange->last, start, last);
2139 
2140 	/* Make sure pending page faults are drained in the deferred worker
2141 	 * before the range is freed to avoid straggler interrupts on
2142 	 * unmapped memory causing "phantom faults".
2143 	 */
2144 	atomic_inc(&svms->drain_pagefaults);
2145 
2146 	unmap_parent = start <= prange->start && last >= prange->last;
2147 
2148 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2149 		mutex_lock_nested(&pchild->lock, 1);
2150 		s = max(start, pchild->start);
2151 		l = min(last, pchild->last);
2152 		if (l >= s)
2153 			svm_range_unmap_from_gpus(pchild, s, l);
2154 		svm_range_unmap_split(mm, prange, pchild, start, last);
2155 		mutex_unlock(&pchild->lock);
2156 	}
2157 	s = max(start, prange->start);
2158 	l = min(last, prange->last);
2159 	if (l >= s)
2160 		svm_range_unmap_from_gpus(prange, s, l);
2161 	svm_range_unmap_split(mm, prange, prange, start, last);
2162 
2163 	if (unmap_parent)
2164 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2165 	else
2166 		svm_range_add_list_work(svms, prange, mm,
2167 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2168 	schedule_deferred_list_work(svms);
2169 
2170 	kfd_unref_process(p);
2171 }
2172 
2173 /**
2174  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2175  *
2176  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2177  * is from migration, or CPU page invalidation callback.
2178  *
2179  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2180  * work thread, and split prange if only part of prange is unmapped.
2181  *
2182  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2183  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2184  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2185  * update GPU mapping to recover.
2186  *
2187  * Context: mmap lock, notifier_invalidate_start lock are held
2188  *          for invalidate event, prange lock is held if this is from migration
2189  */
2190 static bool
2191 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2192 				    const struct mmu_notifier_range *range,
2193 				    unsigned long cur_seq)
2194 {
2195 	struct svm_range *prange;
2196 	unsigned long start;
2197 	unsigned long last;
2198 
2199 	if (range->event == MMU_NOTIFY_RELEASE)
2200 		return true;
2201 
2202 	start = mni->interval_tree.start;
2203 	last = mni->interval_tree.last;
2204 	start = (start > range->start ? start : range->start) >> PAGE_SHIFT;
2205 	last = (last < (range->end - 1) ? last : range->end - 1) >> PAGE_SHIFT;
2206 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2207 		 start, last, range->start >> PAGE_SHIFT,
2208 		 (range->end - 1) >> PAGE_SHIFT,
2209 		 mni->interval_tree.start >> PAGE_SHIFT,
2210 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2211 
2212 	prange = container_of(mni, struct svm_range, notifier);
2213 
2214 	svm_range_lock(prange);
2215 	mmu_interval_set_seq(mni, cur_seq);
2216 
2217 	switch (range->event) {
2218 	case MMU_NOTIFY_UNMAP:
2219 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2220 		break;
2221 	default:
2222 		svm_range_evict(prange, mni->mm, start, last);
2223 		break;
2224 	}
2225 
2226 	svm_range_unlock(prange);
2227 
2228 	return true;
2229 }
2230 
2231 /**
2232  * svm_range_from_addr - find svm range from fault address
2233  * @svms: svm range list header
2234  * @addr: address to search range interval tree, in pages
2235  * @parent: parent range if range is on child list
2236  *
2237  * Context: The caller must hold svms->lock
2238  *
2239  * Return: the svm_range found or NULL
2240  */
2241 struct svm_range *
2242 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2243 		    struct svm_range **parent)
2244 {
2245 	struct interval_tree_node *node;
2246 	struct svm_range *prange;
2247 	struct svm_range *pchild;
2248 
2249 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2250 	if (!node)
2251 		return NULL;
2252 
2253 	prange = container_of(node, struct svm_range, it_node);
2254 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2255 		 addr, prange->start, prange->last, node->start, node->last);
2256 
2257 	if (addr >= prange->start && addr <= prange->last) {
2258 		if (parent)
2259 			*parent = prange;
2260 		return prange;
2261 	}
2262 	list_for_each_entry(pchild, &prange->child_list, child_list)
2263 		if (addr >= pchild->start && addr <= pchild->last) {
2264 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2265 				 addr, pchild->start, pchild->last);
2266 			if (parent)
2267 				*parent = prange;
2268 			return pchild;
2269 		}
2270 
2271 	return NULL;
2272 }
2273 
2274 /* svm_range_best_restore_location - decide the best fault restore location
2275  * @prange: svm range structure
2276  * @adev: the GPU on which vm fault happened
2277  *
2278  * This is only called when xnack is on, to decide the best location to restore
2279  * the range mapping after GPU vm fault. Caller uses the best location to do
2280  * migration if actual loc is not best location, then update GPU page table
2281  * mapping to the best location.
2282  *
2283  * If the preferred loc is accessible by faulting GPU, use preferred loc.
2284  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2285  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2286  *    if range actual loc is cpu, best_loc is cpu
2287  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2288  *    range actual loc.
2289  * Otherwise, GPU no access, best_loc is -1.
2290  *
2291  * Return:
2292  * -1 means vm fault GPU no access
2293  * 0 for CPU or GPU id
2294  */
2295 static int32_t
2296 svm_range_best_restore_location(struct svm_range *prange,
2297 				struct amdgpu_device *adev,
2298 				int32_t *gpuidx)
2299 {
2300 	struct amdgpu_device *bo_adev, *preferred_adev;
2301 	struct kfd_process *p;
2302 	uint32_t gpuid;
2303 	int r;
2304 
2305 	p = container_of(prange->svms, struct kfd_process, svms);
2306 
2307 	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, gpuidx);
2308 	if (r < 0) {
2309 		pr_debug("failed to get gpuid from kgd\n");
2310 		return -1;
2311 	}
2312 
2313 	if (prange->preferred_loc == gpuid ||
2314 	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2315 		return prange->preferred_loc;
2316 	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2317 		preferred_adev = svm_range_get_adev_by_id(prange,
2318 							prange->preferred_loc);
2319 		if (amdgpu_xgmi_same_hive(adev, preferred_adev))
2320 			return prange->preferred_loc;
2321 		/* fall through */
2322 	}
2323 
2324 	if (test_bit(*gpuidx, prange->bitmap_access))
2325 		return gpuid;
2326 
2327 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2328 		if (!prange->actual_loc)
2329 			return 0;
2330 
2331 		bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2332 		if (amdgpu_xgmi_same_hive(adev, bo_adev))
2333 			return prange->actual_loc;
2334 		else
2335 			return 0;
2336 	}
2337 
2338 	return -1;
2339 }
2340 
2341 static int
2342 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2343 			       unsigned long *start, unsigned long *last,
2344 			       bool *is_heap_stack)
2345 {
2346 	struct vm_area_struct *vma;
2347 	struct interval_tree_node *node;
2348 	unsigned long start_limit, end_limit;
2349 
2350 	vma = find_vma(p->mm, addr << PAGE_SHIFT);
2351 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2352 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2353 		return -EFAULT;
2354 	}
2355 
2356 	*is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
2357 			  vma->vm_end >= vma->vm_mm->start_brk) ||
2358 			 (vma->vm_start <= vma->vm_mm->start_stack &&
2359 			  vma->vm_end >= vma->vm_mm->start_stack);
2360 
2361 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2362 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2363 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2364 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2365 	/* First range that starts after the fault address */
2366 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2367 	if (node) {
2368 		end_limit = min(end_limit, node->start);
2369 		/* Last range that ends before the fault address */
2370 		node = container_of(rb_prev(&node->rb),
2371 				    struct interval_tree_node, rb);
2372 	} else {
2373 		/* Last range must end before addr because
2374 		 * there was no range after addr
2375 		 */
2376 		node = container_of(rb_last(&p->svms.objects.rb_root),
2377 				    struct interval_tree_node, rb);
2378 	}
2379 	if (node) {
2380 		if (node->last >= addr) {
2381 			WARN(1, "Overlap with prev node and page fault addr\n");
2382 			return -EFAULT;
2383 		}
2384 		start_limit = max(start_limit, node->last + 1);
2385 	}
2386 
2387 	*start = start_limit;
2388 	*last = end_limit - 1;
2389 
2390 	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2391 		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2392 		 *start, *last, *is_heap_stack);
2393 
2394 	return 0;
2395 }
2396 
2397 static int
2398 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2399 			   uint64_t *bo_s, uint64_t *bo_l)
2400 {
2401 	struct amdgpu_bo_va_mapping *mapping;
2402 	struct interval_tree_node *node;
2403 	struct amdgpu_bo *bo = NULL;
2404 	unsigned long userptr;
2405 	uint32_t i;
2406 	int r;
2407 
2408 	for (i = 0; i < p->n_pdds; i++) {
2409 		struct amdgpu_vm *vm;
2410 
2411 		if (!p->pdds[i]->drm_priv)
2412 			continue;
2413 
2414 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2415 		r = amdgpu_bo_reserve(vm->root.bo, false);
2416 		if (r)
2417 			return r;
2418 
2419 		/* Check userptr by searching entire vm->va interval tree */
2420 		node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2421 		while (node) {
2422 			mapping = container_of((struct rb_node *)node,
2423 					       struct amdgpu_bo_va_mapping, rb);
2424 			bo = mapping->bo_va->base.bo;
2425 
2426 			if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2427 							 start << PAGE_SHIFT,
2428 							 last << PAGE_SHIFT,
2429 							 &userptr)) {
2430 				node = interval_tree_iter_next(node, 0, ~0ULL);
2431 				continue;
2432 			}
2433 
2434 			pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2435 				 start, last);
2436 			if (bo_s && bo_l) {
2437 				*bo_s = userptr >> PAGE_SHIFT;
2438 				*bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2439 			}
2440 			amdgpu_bo_unreserve(vm->root.bo);
2441 			return -EADDRINUSE;
2442 		}
2443 		amdgpu_bo_unreserve(vm->root.bo);
2444 	}
2445 	return 0;
2446 }
2447 
2448 static struct
2449 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2450 						struct kfd_process *p,
2451 						struct mm_struct *mm,
2452 						int64_t addr)
2453 {
2454 	struct svm_range *prange = NULL;
2455 	unsigned long start, last;
2456 	uint32_t gpuid, gpuidx;
2457 	bool is_heap_stack;
2458 	uint64_t bo_s = 0;
2459 	uint64_t bo_l = 0;
2460 	int r;
2461 
2462 	if (svm_range_get_range_boundaries(p, addr, &start, &last,
2463 					   &is_heap_stack))
2464 		return NULL;
2465 
2466 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2467 	if (r != -EADDRINUSE)
2468 		r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2469 
2470 	if (r == -EADDRINUSE) {
2471 		if (addr >= bo_s && addr <= bo_l)
2472 			return NULL;
2473 
2474 		/* Create one page svm range if 2MB range overlapping */
2475 		start = addr;
2476 		last = addr;
2477 	}
2478 
2479 	prange = svm_range_new(&p->svms, start, last);
2480 	if (!prange) {
2481 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2482 		return NULL;
2483 	}
2484 	if (kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx)) {
2485 		pr_debug("failed to get gpuid from kgd\n");
2486 		svm_range_free(prange);
2487 		return NULL;
2488 	}
2489 
2490 	if (is_heap_stack)
2491 		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2492 
2493 	svm_range_add_to_svms(prange);
2494 	svm_range_add_notifier_locked(mm, prange);
2495 
2496 	return prange;
2497 }
2498 
2499 /* svm_range_skip_recover - decide if prange can be recovered
2500  * @prange: svm range structure
2501  *
2502  * GPU vm retry fault handle skip recover the range for cases:
2503  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2504  *    deferred list work will drain the stale fault before free the prange.
2505  * 2. prange is on deferred list to add interval notifier after split, or
2506  * 3. prange is child range, it is split from parent prange, recover later
2507  *    after interval notifier is added.
2508  *
2509  * Return: true to skip recover, false to recover
2510  */
2511 static bool svm_range_skip_recover(struct svm_range *prange)
2512 {
2513 	struct svm_range_list *svms = prange->svms;
2514 
2515 	spin_lock(&svms->deferred_list_lock);
2516 	if (list_empty(&prange->deferred_list) &&
2517 	    list_empty(&prange->child_list)) {
2518 		spin_unlock(&svms->deferred_list_lock);
2519 		return false;
2520 	}
2521 	spin_unlock(&svms->deferred_list_lock);
2522 
2523 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2524 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2525 			 svms, prange, prange->start, prange->last);
2526 		return true;
2527 	}
2528 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2529 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2530 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2531 			 svms, prange, prange->start, prange->last);
2532 		return true;
2533 	}
2534 	return false;
2535 }
2536 
2537 static void
2538 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2539 		      int32_t gpuidx)
2540 {
2541 	struct kfd_process_device *pdd;
2542 
2543 	/* fault is on different page of same range
2544 	 * or fault is skipped to recover later
2545 	 * or fault is on invalid virtual address
2546 	 */
2547 	if (gpuidx == MAX_GPU_INSTANCE) {
2548 		uint32_t gpuid;
2549 		int r;
2550 
2551 		r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpuidx);
2552 		if (r < 0)
2553 			return;
2554 	}
2555 
2556 	/* fault is recovered
2557 	 * or fault cannot recover because GPU no access on the range
2558 	 */
2559 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2560 	if (pdd)
2561 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2562 }
2563 
2564 static bool
2565 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2566 {
2567 	unsigned long requested = VM_READ;
2568 
2569 	if (write_fault)
2570 		requested |= VM_WRITE;
2571 
2572 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2573 		vma->vm_flags);
2574 	return (vma->vm_flags & requested) == requested;
2575 }
2576 
2577 int
2578 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2579 			uint64_t addr, bool write_fault)
2580 {
2581 	struct mm_struct *mm = NULL;
2582 	struct svm_range_list *svms;
2583 	struct svm_range *prange;
2584 	struct kfd_process *p;
2585 	uint64_t timestamp;
2586 	int32_t best_loc;
2587 	int32_t gpuidx = MAX_GPU_INSTANCE;
2588 	bool write_locked = false;
2589 	struct vm_area_struct *vma;
2590 	int r = 0;
2591 
2592 	if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2593 		pr_debug("device does not support SVM\n");
2594 		return -EFAULT;
2595 	}
2596 
2597 	p = kfd_lookup_process_by_pasid(pasid);
2598 	if (!p) {
2599 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2600 		return 0;
2601 	}
2602 	if (!p->xnack_enabled) {
2603 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2604 		r = -EFAULT;
2605 		goto out;
2606 	}
2607 	svms = &p->svms;
2608 
2609 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2610 
2611 	if (atomic_read(&svms->drain_pagefaults)) {
2612 		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2613 		r = 0;
2614 		goto out;
2615 	}
2616 
2617 	/* p->lead_thread is available as kfd_process_wq_release flush the work
2618 	 * before releasing task ref.
2619 	 */
2620 	mm = get_task_mm(p->lead_thread);
2621 	if (!mm) {
2622 		pr_debug("svms 0x%p failed to get mm\n", svms);
2623 		r = 0;
2624 		goto out;
2625 	}
2626 
2627 	mmap_read_lock(mm);
2628 retry_write_locked:
2629 	mutex_lock(&svms->lock);
2630 	prange = svm_range_from_addr(svms, addr, NULL);
2631 	if (!prange) {
2632 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2633 			 svms, addr);
2634 		if (!write_locked) {
2635 			/* Need the write lock to create new range with MMU notifier.
2636 			 * Also flush pending deferred work to make sure the interval
2637 			 * tree is up to date before we add a new range
2638 			 */
2639 			mutex_unlock(&svms->lock);
2640 			mmap_read_unlock(mm);
2641 			mmap_write_lock(mm);
2642 			write_locked = true;
2643 			goto retry_write_locked;
2644 		}
2645 		prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2646 		if (!prange) {
2647 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2648 				 svms, addr);
2649 			mmap_write_downgrade(mm);
2650 			r = -EFAULT;
2651 			goto out_unlock_svms;
2652 		}
2653 	}
2654 	if (write_locked)
2655 		mmap_write_downgrade(mm);
2656 
2657 	mutex_lock(&prange->migrate_mutex);
2658 
2659 	if (svm_range_skip_recover(prange)) {
2660 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2661 		r = 0;
2662 		goto out_unlock_range;
2663 	}
2664 
2665 	timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp;
2666 	/* skip duplicate vm fault on different pages of same range */
2667 	if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
2668 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2669 			 svms, prange->start, prange->last);
2670 		r = 0;
2671 		goto out_unlock_range;
2672 	}
2673 
2674 	/* __do_munmap removed VMA, return success as we are handling stale
2675 	 * retry fault.
2676 	 */
2677 	vma = find_vma(mm, addr << PAGE_SHIFT);
2678 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2679 		pr_debug("address 0x%llx VMA is removed\n", addr);
2680 		r = 0;
2681 		goto out_unlock_range;
2682 	}
2683 
2684 	if (!svm_fault_allowed(vma, write_fault)) {
2685 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
2686 			write_fault ? "write" : "read");
2687 		r = -EPERM;
2688 		goto out_unlock_range;
2689 	}
2690 
2691 	best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2692 	if (best_loc == -1) {
2693 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2694 			 svms, prange->start, prange->last);
2695 		r = -EACCES;
2696 		goto out_unlock_range;
2697 	}
2698 
2699 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2700 		 svms, prange->start, prange->last, best_loc,
2701 		 prange->actual_loc);
2702 
2703 	if (prange->actual_loc != best_loc) {
2704 		if (best_loc) {
2705 			r = svm_migrate_to_vram(prange, best_loc, mm);
2706 			if (r) {
2707 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2708 					 r, addr);
2709 				/* Fallback to system memory if migration to
2710 				 * VRAM failed
2711 				 */
2712 				if (prange->actual_loc)
2713 					r = svm_migrate_vram_to_ram(prange, mm);
2714 				else
2715 					r = 0;
2716 			}
2717 		} else {
2718 			r = svm_migrate_vram_to_ram(prange, mm);
2719 		}
2720 		if (r) {
2721 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2722 				 r, svms, prange->start, prange->last);
2723 			goto out_unlock_range;
2724 		}
2725 	}
2726 
2727 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false);
2728 	if (r)
2729 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2730 			 r, svms, prange->start, prange->last);
2731 
2732 out_unlock_range:
2733 	mutex_unlock(&prange->migrate_mutex);
2734 out_unlock_svms:
2735 	mutex_unlock(&svms->lock);
2736 	mmap_read_unlock(mm);
2737 
2738 	svm_range_count_fault(adev, p, gpuidx);
2739 
2740 	mmput(mm);
2741 out:
2742 	kfd_unref_process(p);
2743 
2744 	if (r == -EAGAIN) {
2745 		pr_debug("recover vm fault later\n");
2746 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2747 		r = 0;
2748 	}
2749 	return r;
2750 }
2751 
2752 void svm_range_list_fini(struct kfd_process *p)
2753 {
2754 	struct svm_range *prange;
2755 	struct svm_range *next;
2756 
2757 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
2758 
2759 	/* Ensure list work is finished before process is destroyed */
2760 	flush_work(&p->svms.deferred_list_work);
2761 
2762 	/*
2763 	 * Ensure no retry fault comes in afterwards, as page fault handler will
2764 	 * not find kfd process and take mm lock to recover fault.
2765 	 */
2766 	atomic_inc(&p->svms.drain_pagefaults);
2767 	svm_range_drain_retry_fault(&p->svms);
2768 
2769 
2770 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
2771 		svm_range_unlink(prange);
2772 		svm_range_remove_notifier(prange);
2773 		svm_range_free(prange);
2774 	}
2775 
2776 	mutex_destroy(&p->svms.lock);
2777 
2778 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
2779 }
2780 
2781 int svm_range_list_init(struct kfd_process *p)
2782 {
2783 	struct svm_range_list *svms = &p->svms;
2784 	int i;
2785 
2786 	svms->objects = RB_ROOT_CACHED;
2787 	mutex_init(&svms->lock);
2788 	INIT_LIST_HEAD(&svms->list);
2789 	atomic_set(&svms->evicted_ranges, 0);
2790 	atomic_set(&svms->drain_pagefaults, 0);
2791 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
2792 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
2793 	INIT_LIST_HEAD(&svms->deferred_range_list);
2794 	spin_lock_init(&svms->deferred_list_lock);
2795 
2796 	for (i = 0; i < p->n_pdds; i++)
2797 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
2798 			bitmap_set(svms->bitmap_supported, i, 1);
2799 
2800 	return 0;
2801 }
2802 
2803 /**
2804  * svm_range_check_vm - check if virtual address range mapped already
2805  * @p: current kfd_process
2806  * @start: range start address, in pages
2807  * @last: range last address, in pages
2808  * @bo_s: mapping start address in pages if address range already mapped
2809  * @bo_l: mapping last address in pages if address range already mapped
2810  *
2811  * The purpose is to avoid virtual address ranges already allocated by
2812  * kfd_ioctl_alloc_memory_of_gpu ioctl.
2813  * It looks for each pdd in the kfd_process.
2814  *
2815  * Context: Process context
2816  *
2817  * Return 0 - OK, if the range is not mapped.
2818  * Otherwise error code:
2819  * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
2820  * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
2821  * a signal. Release all buffer reservations and return to user-space.
2822  */
2823 static int
2824 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
2825 		   uint64_t *bo_s, uint64_t *bo_l)
2826 {
2827 	struct amdgpu_bo_va_mapping *mapping;
2828 	struct interval_tree_node *node;
2829 	uint32_t i;
2830 	int r;
2831 
2832 	for (i = 0; i < p->n_pdds; i++) {
2833 		struct amdgpu_vm *vm;
2834 
2835 		if (!p->pdds[i]->drm_priv)
2836 			continue;
2837 
2838 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2839 		r = amdgpu_bo_reserve(vm->root.bo, false);
2840 		if (r)
2841 			return r;
2842 
2843 		node = interval_tree_iter_first(&vm->va, start, last);
2844 		if (node) {
2845 			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
2846 				 start, last);
2847 			mapping = container_of((struct rb_node *)node,
2848 					       struct amdgpu_bo_va_mapping, rb);
2849 			if (bo_s && bo_l) {
2850 				*bo_s = mapping->start;
2851 				*bo_l = mapping->last;
2852 			}
2853 			amdgpu_bo_unreserve(vm->root.bo);
2854 			return -EADDRINUSE;
2855 		}
2856 		amdgpu_bo_unreserve(vm->root.bo);
2857 	}
2858 
2859 	return 0;
2860 }
2861 
2862 /**
2863  * svm_range_is_valid - check if virtual address range is valid
2864  * @p: current kfd_process
2865  * @start: range start address, in pages
2866  * @size: range size, in pages
2867  *
2868  * Valid virtual address range means it belongs to one or more VMAs
2869  *
2870  * Context: Process context
2871  *
2872  * Return:
2873  *  0 - OK, otherwise error code
2874  */
2875 static int
2876 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
2877 {
2878 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
2879 	struct vm_area_struct *vma;
2880 	unsigned long end;
2881 	unsigned long start_unchg = start;
2882 
2883 	start <<= PAGE_SHIFT;
2884 	end = start + (size << PAGE_SHIFT);
2885 	do {
2886 		vma = find_vma(p->mm, start);
2887 		if (!vma || start < vma->vm_start ||
2888 		    (vma->vm_flags & device_vma))
2889 			return -EFAULT;
2890 		start = min(end, vma->vm_end);
2891 	} while (start < end);
2892 
2893 	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
2894 				  NULL);
2895 }
2896 
2897 /**
2898  * svm_range_add - add svm range and handle overlap
2899  * @p: the range add to this process svms
2900  * @start: page size aligned
2901  * @size: page size aligned
2902  * @nattr: number of attributes
2903  * @attrs: array of attributes
2904  * @update_list: output, the ranges need validate and update GPU mapping
2905  * @insert_list: output, the ranges need insert to svms
2906  * @remove_list: output, the ranges are replaced and need remove from svms
2907  *
2908  * Check if the virtual address range has overlap with the registered ranges,
2909  * split the overlapped range, copy and adjust pages address and vram nodes in
2910  * old and new ranges.
2911  *
2912  * Context: Process context, caller must hold svms->lock
2913  *
2914  * Return:
2915  * 0 - OK, otherwise error code
2916  */
2917 static int
2918 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2919 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2920 	      struct list_head *update_list, struct list_head *insert_list,
2921 	      struct list_head *remove_list)
2922 {
2923 	uint64_t last = start + size - 1UL;
2924 	struct svm_range_list *svms;
2925 	struct svm_range new = {0};
2926 	struct svm_range *prange;
2927 	unsigned long left = 0;
2928 	int r = 0;
2929 
2930 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", &p->svms, start, last);
2931 
2932 	svm_range_apply_attrs(p, &new, nattr, attrs);
2933 
2934 	svms = &p->svms;
2935 
2936 	r = svm_range_handle_overlap(svms, &new, start, last, update_list,
2937 				     insert_list, remove_list, &left);
2938 	if (r)
2939 		return r;
2940 
2941 	if (left) {
2942 		prange = svm_range_new(svms, last - left + 1, last);
2943 		list_add(&prange->insert_list, insert_list);
2944 		list_add(&prange->update_list, update_list);
2945 	}
2946 
2947 	return 0;
2948 }
2949 
2950 /**
2951  * svm_range_best_prefetch_location - decide the best prefetch location
2952  * @prange: svm range structure
2953  *
2954  * For xnack off:
2955  * If range map to single GPU, the best prefetch location is prefetch_loc, which
2956  * can be CPU or GPU.
2957  *
2958  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
2959  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
2960  * the best prefetch location is always CPU, because GPU can not have coherent
2961  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
2962  *
2963  * For xnack on:
2964  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
2965  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
2966  *
2967  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
2968  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
2969  * prefetch location is always CPU.
2970  *
2971  * Context: Process context
2972  *
2973  * Return:
2974  * 0 for CPU or GPU id
2975  */
2976 static uint32_t
2977 svm_range_best_prefetch_location(struct svm_range *prange)
2978 {
2979 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
2980 	uint32_t best_loc = prange->prefetch_loc;
2981 	struct kfd_process_device *pdd;
2982 	struct amdgpu_device *bo_adev;
2983 	struct amdgpu_device *adev;
2984 	struct kfd_process *p;
2985 	uint32_t gpuidx;
2986 
2987 	p = container_of(prange->svms, struct kfd_process, svms);
2988 
2989 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
2990 		goto out;
2991 
2992 	bo_adev = svm_range_get_adev_by_id(prange, best_loc);
2993 	if (!bo_adev) {
2994 		WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
2995 		best_loc = 0;
2996 		goto out;
2997 	}
2998 
2999 	if (p->xnack_enabled)
3000 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3001 	else
3002 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3003 			  MAX_GPU_INSTANCE);
3004 
3005 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3006 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3007 		if (!pdd) {
3008 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3009 			continue;
3010 		}
3011 		adev = (struct amdgpu_device *)pdd->dev->kgd;
3012 
3013 		if (adev == bo_adev)
3014 			continue;
3015 
3016 		if (!amdgpu_xgmi_same_hive(adev, bo_adev)) {
3017 			best_loc = 0;
3018 			break;
3019 		}
3020 	}
3021 
3022 out:
3023 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3024 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3025 		 best_loc);
3026 
3027 	return best_loc;
3028 }
3029 
3030 /* FIXME: This is a workaround for page locking bug when some pages are
3031  * invalid during migration to VRAM
3032  */
3033 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
3034 			void *owner)
3035 {
3036 	struct hmm_range *hmm_range;
3037 	int r;
3038 
3039 	if (prange->validated_once)
3040 		return;
3041 
3042 	r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
3043 				       prange->start << PAGE_SHIFT,
3044 				       prange->npages, &hmm_range,
3045 				       false, true, owner);
3046 	if (!r) {
3047 		amdgpu_hmm_range_get_pages_done(hmm_range);
3048 		prange->validated_once = true;
3049 	}
3050 }
3051 
3052 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3053  * @mm: current process mm_struct
3054  * @prange: svm range structure
3055  * @migrated: output, true if migration is triggered
3056  *
3057  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3058  * from ram to vram.
3059  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3060  * from vram to ram.
3061  *
3062  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3063  * and restore work:
3064  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3065  *    stops all queues, schedule restore work
3066  * 2. svm_range_restore_work wait for migration is done by
3067  *    a. svm_range_validate_vram takes prange->migrate_mutex
3068  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3069  * 3. restore work update mappings of GPU, resume all queues.
3070  *
3071  * Context: Process context
3072  *
3073  * Return:
3074  * 0 - OK, otherwise - error code of migration
3075  */
3076 static int
3077 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3078 			    bool *migrated)
3079 {
3080 	uint32_t best_loc;
3081 	int r = 0;
3082 
3083 	*migrated = false;
3084 	best_loc = svm_range_best_prefetch_location(prange);
3085 
3086 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3087 	    best_loc == prange->actual_loc)
3088 		return 0;
3089 
3090 	if (!best_loc) {
3091 		r = svm_migrate_vram_to_ram(prange, mm);
3092 		*migrated = !r;
3093 		return r;
3094 	}
3095 
3096 	r = svm_migrate_to_vram(prange, best_loc, mm);
3097 	*migrated = !r;
3098 
3099 	return r;
3100 }
3101 
3102 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3103 {
3104 	if (!fence)
3105 		return -EINVAL;
3106 
3107 	if (dma_fence_is_signaled(&fence->base))
3108 		return 0;
3109 
3110 	if (fence->svm_bo) {
3111 		WRITE_ONCE(fence->svm_bo->evicting, 1);
3112 		schedule_work(&fence->svm_bo->eviction_work);
3113 	}
3114 
3115 	return 0;
3116 }
3117 
3118 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3119 {
3120 	struct svm_range_bo *svm_bo;
3121 	struct kfd_process *p;
3122 	struct mm_struct *mm;
3123 
3124 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3125 	if (!svm_bo_ref_unless_zero(svm_bo))
3126 		return; /* svm_bo was freed while eviction was pending */
3127 
3128 	/* svm_range_bo_release destroys this worker thread. So during
3129 	 * the lifetime of this thread, kfd_process and mm will be valid.
3130 	 */
3131 	p = container_of(svm_bo->svms, struct kfd_process, svms);
3132 	mm = p->mm;
3133 	if (!mm)
3134 		return;
3135 
3136 	mmap_read_lock(mm);
3137 	spin_lock(&svm_bo->list_lock);
3138 	while (!list_empty(&svm_bo->range_list)) {
3139 		struct svm_range *prange =
3140 				list_first_entry(&svm_bo->range_list,
3141 						struct svm_range, svm_bo_list);
3142 		int retries = 3;
3143 
3144 		list_del_init(&prange->svm_bo_list);
3145 		spin_unlock(&svm_bo->list_lock);
3146 
3147 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3148 			 prange->start, prange->last);
3149 
3150 		mutex_lock(&prange->migrate_mutex);
3151 		do {
3152 			svm_migrate_vram_to_ram(prange,
3153 						svm_bo->eviction_fence->mm);
3154 		} while (prange->actual_loc && --retries);
3155 		WARN(prange->actual_loc, "Migration failed during eviction");
3156 
3157 		mutex_lock(&prange->lock);
3158 		prange->svm_bo = NULL;
3159 		mutex_unlock(&prange->lock);
3160 
3161 		mutex_unlock(&prange->migrate_mutex);
3162 
3163 		spin_lock(&svm_bo->list_lock);
3164 	}
3165 	spin_unlock(&svm_bo->list_lock);
3166 	mmap_read_unlock(mm);
3167 
3168 	dma_fence_signal(&svm_bo->eviction_fence->base);
3169 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3170 	 * has been called in svm_migrate_vram_to_ram
3171 	 */
3172 	WARN_ONCE(kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3173 	svm_range_bo_unref(svm_bo);
3174 }
3175 
3176 static int
3177 svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
3178 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
3179 {
3180 	struct mm_struct *mm = current->mm;
3181 	struct list_head update_list;
3182 	struct list_head insert_list;
3183 	struct list_head remove_list;
3184 	struct svm_range_list *svms;
3185 	struct svm_range *prange;
3186 	struct svm_range *next;
3187 	int r = 0;
3188 
3189 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3190 		 p->pasid, &p->svms, start, start + size - 1, size);
3191 
3192 	r = svm_range_check_attr(p, nattr, attrs);
3193 	if (r)
3194 		return r;
3195 
3196 	svms = &p->svms;
3197 
3198 	svm_range_list_lock_and_flush_work(svms, mm);
3199 
3200 	r = svm_range_is_valid(p, start, size);
3201 	if (r) {
3202 		pr_debug("invalid range r=%d\n", r);
3203 		mmap_write_unlock(mm);
3204 		goto out;
3205 	}
3206 
3207 	mutex_lock(&svms->lock);
3208 
3209 	/* Add new range and split existing ranges as needed */
3210 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3211 			  &insert_list, &remove_list);
3212 	if (r) {
3213 		mutex_unlock(&svms->lock);
3214 		mmap_write_unlock(mm);
3215 		goto out;
3216 	}
3217 	/* Apply changes as a transaction */
3218 	list_for_each_entry_safe(prange, next, &insert_list, insert_list) {
3219 		svm_range_add_to_svms(prange);
3220 		svm_range_add_notifier_locked(mm, prange);
3221 	}
3222 	list_for_each_entry(prange, &update_list, update_list) {
3223 		svm_range_apply_attrs(p, prange, nattr, attrs);
3224 		/* TODO: unmap ranges from GPU that lost access */
3225 	}
3226 	list_for_each_entry_safe(prange, next, &remove_list,
3227 				remove_list) {
3228 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3229 			 prange->svms, prange, prange->start,
3230 			 prange->last);
3231 		svm_range_unlink(prange);
3232 		svm_range_remove_notifier(prange);
3233 		svm_range_free(prange);
3234 	}
3235 
3236 	mmap_write_downgrade(mm);
3237 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3238 	 * this fails we may be left with partially completed actions. There
3239 	 * is no clean way of rolling back to the previous state in such a
3240 	 * case because the rollback wouldn't be guaranteed to work either.
3241 	 */
3242 	list_for_each_entry(prange, &update_list, update_list) {
3243 		bool migrated;
3244 
3245 		mutex_lock(&prange->migrate_mutex);
3246 
3247 		r = svm_range_trigger_migration(mm, prange, &migrated);
3248 		if (r)
3249 			goto out_unlock_range;
3250 
3251 		if (migrated && !p->xnack_enabled) {
3252 			pr_debug("restore_work will update mappings of GPUs\n");
3253 			mutex_unlock(&prange->migrate_mutex);
3254 			continue;
3255 		}
3256 
3257 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3258 					       true, true);
3259 		if (r)
3260 			pr_debug("failed %d to map svm range\n", r);
3261 
3262 out_unlock_range:
3263 		mutex_unlock(&prange->migrate_mutex);
3264 		if (r)
3265 			break;
3266 	}
3267 
3268 	svm_range_debug_dump(svms);
3269 
3270 	mutex_unlock(&svms->lock);
3271 	mmap_read_unlock(mm);
3272 out:
3273 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3274 		 &p->svms, start, start + size - 1, r);
3275 
3276 	return r;
3277 }
3278 
3279 static int
3280 svm_range_get_attr(struct kfd_process *p, uint64_t start, uint64_t size,
3281 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
3282 {
3283 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3284 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3285 	bool get_preferred_loc = false;
3286 	bool get_prefetch_loc = false;
3287 	bool get_granularity = false;
3288 	bool get_accessible = false;
3289 	bool get_flags = false;
3290 	uint64_t last = start + size - 1UL;
3291 	struct mm_struct *mm = current->mm;
3292 	uint8_t granularity = 0xff;
3293 	struct interval_tree_node *node;
3294 	struct svm_range_list *svms;
3295 	struct svm_range *prange;
3296 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3297 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3298 	uint32_t flags_and = 0xffffffff;
3299 	uint32_t flags_or = 0;
3300 	int gpuidx;
3301 	uint32_t i;
3302 	int r = 0;
3303 
3304 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3305 		 start + size - 1, nattr);
3306 
3307 	/* Flush pending deferred work to avoid racing with deferred actions from
3308 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3309 	 * can still race with get_attr because we don't hold the mmap lock. But that
3310 	 * would be a race condition in the application anyway, and undefined
3311 	 * behaviour is acceptable in that case.
3312 	 */
3313 	flush_work(&p->svms.deferred_list_work);
3314 
3315 	mmap_read_lock(mm);
3316 	r = svm_range_is_valid(p, start, size);
3317 	mmap_read_unlock(mm);
3318 	if (r) {
3319 		pr_debug("invalid range r=%d\n", r);
3320 		return r;
3321 	}
3322 
3323 	for (i = 0; i < nattr; i++) {
3324 		switch (attrs[i].type) {
3325 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3326 			get_preferred_loc = true;
3327 			break;
3328 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3329 			get_prefetch_loc = true;
3330 			break;
3331 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3332 			get_accessible = true;
3333 			break;
3334 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3335 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3336 			get_flags = true;
3337 			break;
3338 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3339 			get_granularity = true;
3340 			break;
3341 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3342 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3343 			fallthrough;
3344 		default:
3345 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3346 			return -EINVAL;
3347 		}
3348 	}
3349 
3350 	svms = &p->svms;
3351 
3352 	mutex_lock(&svms->lock);
3353 
3354 	node = interval_tree_iter_first(&svms->objects, start, last);
3355 	if (!node) {
3356 		pr_debug("range attrs not found return default values\n");
3357 		svm_range_set_default_attributes(&location, &prefetch_loc,
3358 						 &granularity, &flags_and);
3359 		flags_or = flags_and;
3360 		if (p->xnack_enabled)
3361 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3362 				    MAX_GPU_INSTANCE);
3363 		else
3364 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3365 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3366 		goto fill_values;
3367 	}
3368 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3369 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3370 
3371 	while (node) {
3372 		struct interval_tree_node *next;
3373 
3374 		prange = container_of(node, struct svm_range, it_node);
3375 		next = interval_tree_iter_next(node, start, last);
3376 
3377 		if (get_preferred_loc) {
3378 			if (prange->preferred_loc ==
3379 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3380 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3381 			     location != prange->preferred_loc)) {
3382 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3383 				get_preferred_loc = false;
3384 			} else {
3385 				location = prange->preferred_loc;
3386 			}
3387 		}
3388 		if (get_prefetch_loc) {
3389 			if (prange->prefetch_loc ==
3390 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3391 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3392 			     prefetch_loc != prange->prefetch_loc)) {
3393 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3394 				get_prefetch_loc = false;
3395 			} else {
3396 				prefetch_loc = prange->prefetch_loc;
3397 			}
3398 		}
3399 		if (get_accessible) {
3400 			bitmap_and(bitmap_access, bitmap_access,
3401 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3402 			bitmap_and(bitmap_aip, bitmap_aip,
3403 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3404 		}
3405 		if (get_flags) {
3406 			flags_and &= prange->flags;
3407 			flags_or |= prange->flags;
3408 		}
3409 
3410 		if (get_granularity && prange->granularity < granularity)
3411 			granularity = prange->granularity;
3412 
3413 		node = next;
3414 	}
3415 fill_values:
3416 	mutex_unlock(&svms->lock);
3417 
3418 	for (i = 0; i < nattr; i++) {
3419 		switch (attrs[i].type) {
3420 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3421 			attrs[i].value = location;
3422 			break;
3423 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3424 			attrs[i].value = prefetch_loc;
3425 			break;
3426 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3427 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3428 							       attrs[i].value);
3429 			if (gpuidx < 0) {
3430 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3431 				return -EINVAL;
3432 			}
3433 			if (test_bit(gpuidx, bitmap_access))
3434 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3435 			else if (test_bit(gpuidx, bitmap_aip))
3436 				attrs[i].type =
3437 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3438 			else
3439 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3440 			break;
3441 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3442 			attrs[i].value = flags_and;
3443 			break;
3444 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3445 			attrs[i].value = ~flags_or;
3446 			break;
3447 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3448 			attrs[i].value = (uint32_t)granularity;
3449 			break;
3450 		}
3451 	}
3452 
3453 	return 0;
3454 }
3455 
3456 int
3457 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3458 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3459 {
3460 	int r;
3461 
3462 	start >>= PAGE_SHIFT;
3463 	size >>= PAGE_SHIFT;
3464 
3465 	switch (op) {
3466 	case KFD_IOCTL_SVM_OP_SET_ATTR:
3467 		r = svm_range_set_attr(p, start, size, nattrs, attrs);
3468 		break;
3469 	case KFD_IOCTL_SVM_OP_GET_ATTR:
3470 		r = svm_range_get_attr(p, start, size, nattrs, attrs);
3471 		break;
3472 	default:
3473 		r = EINVAL;
3474 		break;
3475 	}
3476 
3477 	return r;
3478 }
3479