xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 48ca54e3)
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 struct criu_svm_metadata {
49 	struct list_head list;
50 	struct kfd_criu_svm_range_priv_data data;
51 };
52 
53 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
54 static bool
55 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
56 				    const struct mmu_notifier_range *range,
57 				    unsigned long cur_seq);
58 static int
59 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
60 		   uint64_t *bo_s, uint64_t *bo_l);
61 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
62 	.invalidate = svm_range_cpu_invalidate_pagetables,
63 };
64 
65 /**
66  * svm_range_unlink - unlink svm_range from lists and interval tree
67  * @prange: svm range structure to be removed
68  *
69  * Remove the svm_range from the svms and svm_bo lists and the svms
70  * interval tree.
71  *
72  * Context: The caller must hold svms->lock
73  */
74 static void svm_range_unlink(struct svm_range *prange)
75 {
76 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
77 		 prange, prange->start, prange->last);
78 
79 	if (prange->svm_bo) {
80 		spin_lock(&prange->svm_bo->list_lock);
81 		list_del(&prange->svm_bo_list);
82 		spin_unlock(&prange->svm_bo->list_lock);
83 	}
84 
85 	list_del(&prange->list);
86 	if (prange->it_node.start != 0 && prange->it_node.last != 0)
87 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
88 }
89 
90 static void
91 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
92 {
93 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
94 		 prange, prange->start, prange->last);
95 
96 	mmu_interval_notifier_insert_locked(&prange->notifier, mm,
97 				     prange->start << PAGE_SHIFT,
98 				     prange->npages << PAGE_SHIFT,
99 				     &svm_range_mn_ops);
100 }
101 
102 /**
103  * svm_range_add_to_svms - add svm range to svms
104  * @prange: svm range structure to be added
105  *
106  * Add the svm range to svms interval tree and link list
107  *
108  * Context: The caller must hold svms->lock
109  */
110 static void svm_range_add_to_svms(struct svm_range *prange)
111 {
112 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
113 		 prange, prange->start, prange->last);
114 
115 	list_move_tail(&prange->list, &prange->svms->list);
116 	prange->it_node.start = prange->start;
117 	prange->it_node.last = prange->last;
118 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
119 }
120 
121 static void svm_range_remove_notifier(struct svm_range *prange)
122 {
123 	pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
124 		 prange->svms, prange,
125 		 prange->notifier.interval_tree.start >> PAGE_SHIFT,
126 		 prange->notifier.interval_tree.last >> PAGE_SHIFT);
127 
128 	if (prange->notifier.interval_tree.start != 0 &&
129 	    prange->notifier.interval_tree.last != 0)
130 		mmu_interval_notifier_remove(&prange->notifier);
131 }
132 
133 static bool
134 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
135 {
136 	return dma_addr && !dma_mapping_error(dev, dma_addr) &&
137 	       !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
138 }
139 
140 static int
141 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
142 		      unsigned long offset, unsigned long npages,
143 		      unsigned long *hmm_pfns, uint32_t gpuidx)
144 {
145 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
146 	dma_addr_t *addr = prange->dma_addr[gpuidx];
147 	struct device *dev = adev->dev;
148 	struct page *page;
149 	int i, r;
150 
151 	if (!addr) {
152 		addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL);
153 		if (!addr)
154 			return -ENOMEM;
155 		prange->dma_addr[gpuidx] = addr;
156 	}
157 
158 	addr += offset;
159 	for (i = 0; i < npages; i++) {
160 		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
161 			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
162 
163 		page = hmm_pfn_to_page(hmm_pfns[i]);
164 		if (is_zone_device_page(page)) {
165 			struct amdgpu_device *bo_adev =
166 					amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
167 
168 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
169 				   bo_adev->vm_manager.vram_base_offset -
170 				   bo_adev->kfd.dev->pgmap.range.start;
171 			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
172 			pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
173 			continue;
174 		}
175 		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
176 		r = dma_mapping_error(dev, addr[i]);
177 		if (r) {
178 			dev_err(dev, "failed %d dma_map_page\n", r);
179 			return r;
180 		}
181 		pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
182 				     addr[i] >> PAGE_SHIFT, page_to_pfn(page));
183 	}
184 	return 0;
185 }
186 
187 static int
188 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
189 		  unsigned long offset, unsigned long npages,
190 		  unsigned long *hmm_pfns)
191 {
192 	struct kfd_process *p;
193 	uint32_t gpuidx;
194 	int r;
195 
196 	p = container_of(prange->svms, struct kfd_process, svms);
197 
198 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
199 		struct kfd_process_device *pdd;
200 
201 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
202 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
203 		if (!pdd) {
204 			pr_debug("failed to find device idx %d\n", gpuidx);
205 			return -EINVAL;
206 		}
207 
208 		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
209 					  hmm_pfns, gpuidx);
210 		if (r)
211 			break;
212 	}
213 
214 	return r;
215 }
216 
217 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
218 			 unsigned long offset, unsigned long npages)
219 {
220 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
221 	int i;
222 
223 	if (!dma_addr)
224 		return;
225 
226 	for (i = offset; i < offset + npages; i++) {
227 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
228 			continue;
229 		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
230 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
231 		dma_addr[i] = 0;
232 	}
233 }
234 
235 void svm_range_free_dma_mappings(struct svm_range *prange)
236 {
237 	struct kfd_process_device *pdd;
238 	dma_addr_t *dma_addr;
239 	struct device *dev;
240 	struct kfd_process *p;
241 	uint32_t gpuidx;
242 
243 	p = container_of(prange->svms, struct kfd_process, svms);
244 
245 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
246 		dma_addr = prange->dma_addr[gpuidx];
247 		if (!dma_addr)
248 			continue;
249 
250 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
251 		if (!pdd) {
252 			pr_debug("failed to find device idx %d\n", gpuidx);
253 			continue;
254 		}
255 		dev = &pdd->dev->pdev->dev;
256 		svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
257 		kvfree(dma_addr);
258 		prange->dma_addr[gpuidx] = NULL;
259 	}
260 }
261 
262 static void svm_range_free(struct svm_range *prange)
263 {
264 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
265 		 prange->start, prange->last);
266 
267 	svm_range_vram_node_free(prange);
268 	svm_range_free_dma_mappings(prange);
269 	mutex_destroy(&prange->lock);
270 	mutex_destroy(&prange->migrate_mutex);
271 	kfree(prange);
272 }
273 
274 static void
275 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
276 				 uint8_t *granularity, uint32_t *flags)
277 {
278 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
279 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
280 	*granularity = 9;
281 	*flags =
282 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
283 }
284 
285 static struct
286 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
287 			 uint64_t last)
288 {
289 	uint64_t size = last - start + 1;
290 	struct svm_range *prange;
291 	struct kfd_process *p;
292 
293 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
294 	if (!prange)
295 		return NULL;
296 	prange->npages = size;
297 	prange->svms = svms;
298 	prange->start = start;
299 	prange->last = last;
300 	INIT_LIST_HEAD(&prange->list);
301 	INIT_LIST_HEAD(&prange->update_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 	pr_debug("svm_bo 0x%p\n", svm_bo);
338 
339 	spin_lock(&svm_bo->list_lock);
340 	while (!list_empty(&svm_bo->range_list)) {
341 		struct svm_range *prange =
342 				list_first_entry(&svm_bo->range_list,
343 						struct svm_range, svm_bo_list);
344 		/* list_del_init tells a concurrent svm_range_vram_node_new when
345 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
346 		 */
347 		list_del_init(&prange->svm_bo_list);
348 		spin_unlock(&svm_bo->list_lock);
349 
350 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
351 			 prange->start, prange->last);
352 		mutex_lock(&prange->lock);
353 		prange->svm_bo = NULL;
354 		mutex_unlock(&prange->lock);
355 
356 		spin_lock(&svm_bo->list_lock);
357 	}
358 	spin_unlock(&svm_bo->list_lock);
359 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
360 		/* We're not in the eviction worker.
361 		 * Signal the fence and synchronize with any
362 		 * pending eviction work.
363 		 */
364 		dma_fence_signal(&svm_bo->eviction_fence->base);
365 		cancel_work_sync(&svm_bo->eviction_work);
366 	}
367 	dma_fence_put(&svm_bo->eviction_fence->base);
368 	amdgpu_bo_unref(&svm_bo->bo);
369 	kfree(svm_bo);
370 }
371 
372 static void svm_range_bo_wq_release(struct work_struct *work)
373 {
374 	struct svm_range_bo *svm_bo;
375 
376 	svm_bo = container_of(work, struct svm_range_bo, release_work);
377 	svm_range_bo_release(&svm_bo->kref);
378 }
379 
380 static void svm_range_bo_release_async(struct kref *kref)
381 {
382 	struct svm_range_bo *svm_bo;
383 
384 	svm_bo = container_of(kref, struct svm_range_bo, kref);
385 	pr_debug("svm_bo 0x%p\n", svm_bo);
386 	INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
387 	schedule_work(&svm_bo->release_work);
388 }
389 
390 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
391 {
392 	kref_put(&svm_bo->kref, svm_range_bo_release_async);
393 }
394 
395 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
396 {
397 	if (svm_bo)
398 		kref_put(&svm_bo->kref, svm_range_bo_release);
399 }
400 
401 static bool
402 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
403 {
404 	struct amdgpu_device *bo_adev;
405 
406 	mutex_lock(&prange->lock);
407 	if (!prange->svm_bo) {
408 		mutex_unlock(&prange->lock);
409 		return false;
410 	}
411 	if (prange->ttm_res) {
412 		/* We still have a reference, all is well */
413 		mutex_unlock(&prange->lock);
414 		return true;
415 	}
416 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
417 		/*
418 		 * Migrate from GPU to GPU, remove range from source bo_adev
419 		 * svm_bo range list, and return false to allocate svm_bo from
420 		 * destination adev.
421 		 */
422 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
423 		if (bo_adev != adev) {
424 			mutex_unlock(&prange->lock);
425 
426 			spin_lock(&prange->svm_bo->list_lock);
427 			list_del_init(&prange->svm_bo_list);
428 			spin_unlock(&prange->svm_bo->list_lock);
429 
430 			svm_range_bo_unref(prange->svm_bo);
431 			return false;
432 		}
433 		if (READ_ONCE(prange->svm_bo->evicting)) {
434 			struct dma_fence *f;
435 			struct svm_range_bo *svm_bo;
436 			/* The BO is getting evicted,
437 			 * we need to get a new one
438 			 */
439 			mutex_unlock(&prange->lock);
440 			svm_bo = prange->svm_bo;
441 			f = dma_fence_get(&svm_bo->eviction_fence->base);
442 			svm_range_bo_unref(prange->svm_bo);
443 			/* wait for the fence to avoid long spin-loop
444 			 * at list_empty_careful
445 			 */
446 			dma_fence_wait(f, false);
447 			dma_fence_put(f);
448 		} else {
449 			/* The BO was still around and we got
450 			 * a new reference to it
451 			 */
452 			mutex_unlock(&prange->lock);
453 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
454 				 prange->svms, prange->start, prange->last);
455 
456 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
457 			return true;
458 		}
459 
460 	} else {
461 		mutex_unlock(&prange->lock);
462 	}
463 
464 	/* We need a new svm_bo. Spin-loop to wait for concurrent
465 	 * svm_range_bo_release to finish removing this range from
466 	 * its range list. After this, it is safe to reuse the
467 	 * svm_bo pointer and svm_bo_list head.
468 	 */
469 	while (!list_empty_careful(&prange->svm_bo_list))
470 		;
471 
472 	return false;
473 }
474 
475 static struct svm_range_bo *svm_range_bo_new(void)
476 {
477 	struct svm_range_bo *svm_bo;
478 
479 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
480 	if (!svm_bo)
481 		return NULL;
482 
483 	kref_init(&svm_bo->kref);
484 	INIT_LIST_HEAD(&svm_bo->range_list);
485 	spin_lock_init(&svm_bo->list_lock);
486 
487 	return svm_bo;
488 }
489 
490 int
491 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
492 			bool clear)
493 {
494 	struct amdgpu_bo_param bp;
495 	struct svm_range_bo *svm_bo;
496 	struct amdgpu_bo_user *ubo;
497 	struct amdgpu_bo *bo;
498 	struct kfd_process *p;
499 	struct mm_struct *mm;
500 	int r;
501 
502 	p = container_of(prange->svms, struct kfd_process, svms);
503 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
504 		 prange->start, prange->last);
505 
506 	if (svm_range_validate_svm_bo(adev, prange))
507 		return 0;
508 
509 	svm_bo = svm_range_bo_new();
510 	if (!svm_bo) {
511 		pr_debug("failed to alloc svm bo\n");
512 		return -ENOMEM;
513 	}
514 	mm = get_task_mm(p->lead_thread);
515 	if (!mm) {
516 		pr_debug("failed to get mm\n");
517 		kfree(svm_bo);
518 		return -ESRCH;
519 	}
520 	svm_bo->svms = prange->svms;
521 	svm_bo->eviction_fence =
522 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
523 					   mm,
524 					   svm_bo);
525 	mmput(mm);
526 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
527 	svm_bo->evicting = 0;
528 	memset(&bp, 0, sizeof(bp));
529 	bp.size = prange->npages * PAGE_SIZE;
530 	bp.byte_align = PAGE_SIZE;
531 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
532 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
533 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
534 	bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
535 	bp.type = ttm_bo_type_device;
536 	bp.resv = NULL;
537 
538 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
539 	if (r) {
540 		pr_debug("failed %d to create bo\n", r);
541 		goto create_bo_failed;
542 	}
543 	bo = &ubo->bo;
544 	r = amdgpu_bo_reserve(bo, true);
545 	if (r) {
546 		pr_debug("failed %d to reserve bo\n", r);
547 		goto reserve_bo_failed;
548 	}
549 
550 	r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
551 	if (r) {
552 		pr_debug("failed %d to reserve bo\n", r);
553 		amdgpu_bo_unreserve(bo);
554 		goto reserve_bo_failed;
555 	}
556 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
557 
558 	amdgpu_bo_unreserve(bo);
559 
560 	svm_bo->bo = bo;
561 	prange->svm_bo = svm_bo;
562 	prange->ttm_res = bo->tbo.resource;
563 	prange->offset = 0;
564 
565 	spin_lock(&svm_bo->list_lock);
566 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
567 	spin_unlock(&svm_bo->list_lock);
568 
569 	return 0;
570 
571 reserve_bo_failed:
572 	amdgpu_bo_unref(&bo);
573 create_bo_failed:
574 	dma_fence_put(&svm_bo->eviction_fence->base);
575 	kfree(svm_bo);
576 	prange->ttm_res = NULL;
577 
578 	return r;
579 }
580 
581 void svm_range_vram_node_free(struct svm_range *prange)
582 {
583 	svm_range_bo_unref(prange->svm_bo);
584 	prange->ttm_res = NULL;
585 }
586 
587 struct amdgpu_device *
588 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
589 {
590 	struct kfd_process_device *pdd;
591 	struct kfd_process *p;
592 	int32_t gpu_idx;
593 
594 	p = container_of(prange->svms, struct kfd_process, svms);
595 
596 	gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
597 	if (gpu_idx < 0) {
598 		pr_debug("failed to get device by id 0x%x\n", gpu_id);
599 		return NULL;
600 	}
601 	pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
602 	if (!pdd) {
603 		pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
604 		return NULL;
605 	}
606 
607 	return pdd->dev->adev;
608 }
609 
610 struct kfd_process_device *
611 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
612 {
613 	struct kfd_process *p;
614 	int32_t gpu_idx, gpuid;
615 	int r;
616 
617 	p = container_of(prange->svms, struct kfd_process, svms);
618 
619 	r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx);
620 	if (r) {
621 		pr_debug("failed to get device id by adev %p\n", adev);
622 		return NULL;
623 	}
624 
625 	return kfd_process_device_from_gpuidx(p, gpu_idx);
626 }
627 
628 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
629 {
630 	struct ttm_operation_ctx ctx = { false, false };
631 
632 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
633 
634 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
635 }
636 
637 static int
638 svm_range_check_attr(struct kfd_process *p,
639 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
640 {
641 	uint32_t i;
642 
643 	for (i = 0; i < nattr; i++) {
644 		uint32_t val = attrs[i].value;
645 		int gpuidx = MAX_GPU_INSTANCE;
646 
647 		switch (attrs[i].type) {
648 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
649 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
650 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
651 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
652 			break;
653 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
654 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
655 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
656 			break;
657 		case KFD_IOCTL_SVM_ATTR_ACCESS:
658 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
659 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
660 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
661 			break;
662 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
663 			break;
664 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
665 			break;
666 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
667 			break;
668 		default:
669 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
670 			return -EINVAL;
671 		}
672 
673 		if (gpuidx < 0) {
674 			pr_debug("no GPU 0x%x found\n", val);
675 			return -EINVAL;
676 		} else if (gpuidx < MAX_GPU_INSTANCE &&
677 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
678 			pr_debug("GPU 0x%x not supported\n", val);
679 			return -EINVAL;
680 		}
681 	}
682 
683 	return 0;
684 }
685 
686 static void
687 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
688 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
689 		      bool *update_mapping)
690 {
691 	uint32_t i;
692 	int gpuidx;
693 
694 	for (i = 0; i < nattr; i++) {
695 		switch (attrs[i].type) {
696 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
697 			prange->preferred_loc = attrs[i].value;
698 			break;
699 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
700 			prange->prefetch_loc = attrs[i].value;
701 			break;
702 		case KFD_IOCTL_SVM_ATTR_ACCESS:
703 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
704 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
705 			*update_mapping = true;
706 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
707 							       attrs[i].value);
708 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
709 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
710 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
711 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
712 				bitmap_set(prange->bitmap_access, gpuidx, 1);
713 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
714 			} else {
715 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
716 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
717 			}
718 			break;
719 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
720 			*update_mapping = true;
721 			prange->flags |= attrs[i].value;
722 			break;
723 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
724 			*update_mapping = true;
725 			prange->flags &= ~attrs[i].value;
726 			break;
727 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
728 			prange->granularity = attrs[i].value;
729 			break;
730 		default:
731 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
732 		}
733 	}
734 }
735 
736 static bool
737 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
738 			uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
739 {
740 	uint32_t i;
741 	int gpuidx;
742 
743 	for (i = 0; i < nattr; i++) {
744 		switch (attrs[i].type) {
745 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
746 			if (prange->preferred_loc != attrs[i].value)
747 				return false;
748 			break;
749 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
750 			/* Prefetch should always trigger a migration even
751 			 * if the value of the attribute didn't change.
752 			 */
753 			return false;
754 		case KFD_IOCTL_SVM_ATTR_ACCESS:
755 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
756 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
757 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
758 							       attrs[i].value);
759 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
760 				if (test_bit(gpuidx, prange->bitmap_access) ||
761 				    test_bit(gpuidx, prange->bitmap_aip))
762 					return false;
763 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
764 				if (!test_bit(gpuidx, prange->bitmap_access))
765 					return false;
766 			} else {
767 				if (!test_bit(gpuidx, prange->bitmap_aip))
768 					return false;
769 			}
770 			break;
771 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
772 			if ((prange->flags & attrs[i].value) != attrs[i].value)
773 				return false;
774 			break;
775 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
776 			if ((prange->flags & attrs[i].value) != 0)
777 				return false;
778 			break;
779 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
780 			if (prange->granularity != attrs[i].value)
781 				return false;
782 			break;
783 		default:
784 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
785 		}
786 	}
787 
788 	return true;
789 }
790 
791 /**
792  * svm_range_debug_dump - print all range information from svms
793  * @svms: svm range list header
794  *
795  * debug output svm range start, end, prefetch location from svms
796  * interval tree and link list
797  *
798  * Context: The caller must hold svms->lock
799  */
800 static void svm_range_debug_dump(struct svm_range_list *svms)
801 {
802 	struct interval_tree_node *node;
803 	struct svm_range *prange;
804 
805 	pr_debug("dump svms 0x%p list\n", svms);
806 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
807 
808 	list_for_each_entry(prange, &svms->list, list) {
809 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
810 			 prange, prange->start, prange->npages,
811 			 prange->start + prange->npages - 1,
812 			 prange->actual_loc);
813 	}
814 
815 	pr_debug("dump svms 0x%p interval tree\n", svms);
816 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
817 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
818 	while (node) {
819 		prange = container_of(node, struct svm_range, it_node);
820 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
821 			 prange, prange->start, prange->npages,
822 			 prange->start + prange->npages - 1,
823 			 prange->actual_loc);
824 		node = interval_tree_iter_next(node, 0, ~0ULL);
825 	}
826 }
827 
828 static int
829 svm_range_split_array(void *ppnew, void *ppold, size_t size,
830 		      uint64_t old_start, uint64_t old_n,
831 		      uint64_t new_start, uint64_t new_n)
832 {
833 	unsigned char *new, *old, *pold;
834 	uint64_t d;
835 
836 	if (!ppold)
837 		return 0;
838 	pold = *(unsigned char **)ppold;
839 	if (!pold)
840 		return 0;
841 
842 	new = kvmalloc_array(new_n, size, GFP_KERNEL);
843 	if (!new)
844 		return -ENOMEM;
845 
846 	d = (new_start - old_start) * size;
847 	memcpy(new, pold + d, new_n * size);
848 
849 	old = kvmalloc_array(old_n, size, GFP_KERNEL);
850 	if (!old) {
851 		kvfree(new);
852 		return -ENOMEM;
853 	}
854 
855 	d = (new_start == old_start) ? new_n * size : 0;
856 	memcpy(old, pold + d, old_n * size);
857 
858 	kvfree(pold);
859 	*(void **)ppold = old;
860 	*(void **)ppnew = new;
861 
862 	return 0;
863 }
864 
865 static int
866 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
867 		      uint64_t start, uint64_t last)
868 {
869 	uint64_t npages = last - start + 1;
870 	int i, r;
871 
872 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
873 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
874 					  sizeof(*old->dma_addr[i]), old->start,
875 					  npages, new->start, new->npages);
876 		if (r)
877 			return r;
878 	}
879 
880 	return 0;
881 }
882 
883 static int
884 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
885 		      uint64_t start, uint64_t last)
886 {
887 	uint64_t npages = last - start + 1;
888 
889 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
890 		 new->svms, new, new->start, start, last);
891 
892 	if (new->start == old->start) {
893 		new->offset = old->offset;
894 		old->offset += new->npages;
895 	} else {
896 		new->offset = old->offset + npages;
897 	}
898 
899 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
900 	new->ttm_res = old->ttm_res;
901 
902 	spin_lock(&new->svm_bo->list_lock);
903 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
904 	spin_unlock(&new->svm_bo->list_lock);
905 
906 	return 0;
907 }
908 
909 /**
910  * svm_range_split_adjust - split range and adjust
911  *
912  * @new: new range
913  * @old: the old range
914  * @start: the old range adjust to start address in pages
915  * @last: the old range adjust to last address in pages
916  *
917  * Copy system memory dma_addr or vram ttm_res in old range to new
918  * range from new_start up to size new->npages, the remaining old range is from
919  * start to last
920  *
921  * Return:
922  * 0 - OK, -ENOMEM - out of memory
923  */
924 static int
925 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
926 		      uint64_t start, uint64_t last)
927 {
928 	int r;
929 
930 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
931 		 new->svms, new->start, old->start, old->last, start, last);
932 
933 	if (new->start < old->start ||
934 	    new->last > old->last) {
935 		WARN_ONCE(1, "invalid new range start or last\n");
936 		return -EINVAL;
937 	}
938 
939 	r = svm_range_split_pages(new, old, start, last);
940 	if (r)
941 		return r;
942 
943 	if (old->actual_loc && old->ttm_res) {
944 		r = svm_range_split_nodes(new, old, start, last);
945 		if (r)
946 			return r;
947 	}
948 
949 	old->npages = last - start + 1;
950 	old->start = start;
951 	old->last = last;
952 	new->flags = old->flags;
953 	new->preferred_loc = old->preferred_loc;
954 	new->prefetch_loc = old->prefetch_loc;
955 	new->actual_loc = old->actual_loc;
956 	new->granularity = old->granularity;
957 	new->mapped_to_gpu = old->mapped_to_gpu;
958 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
959 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
960 
961 	return 0;
962 }
963 
964 /**
965  * svm_range_split - split a range in 2 ranges
966  *
967  * @prange: the svm range to split
968  * @start: the remaining range start address in pages
969  * @last: the remaining range last address in pages
970  * @new: the result new range generated
971  *
972  * Two cases only:
973  * case 1: if start == prange->start
974  *         prange ==> prange[start, last]
975  *         new range [last + 1, prange->last]
976  *
977  * case 2: if last == prange->last
978  *         prange ==> prange[start, last]
979  *         new range [prange->start, start - 1]
980  *
981  * Return:
982  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
983  */
984 static int
985 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
986 		struct svm_range **new)
987 {
988 	uint64_t old_start = prange->start;
989 	uint64_t old_last = prange->last;
990 	struct svm_range_list *svms;
991 	int r = 0;
992 
993 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
994 		 old_start, old_last, start, last);
995 
996 	if (old_start != start && old_last != last)
997 		return -EINVAL;
998 	if (start < old_start || last > old_last)
999 		return -EINVAL;
1000 
1001 	svms = prange->svms;
1002 	if (old_start == start)
1003 		*new = svm_range_new(svms, last + 1, old_last);
1004 	else
1005 		*new = svm_range_new(svms, old_start, start - 1);
1006 	if (!*new)
1007 		return -ENOMEM;
1008 
1009 	r = svm_range_split_adjust(*new, prange, start, last);
1010 	if (r) {
1011 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1012 			 r, old_start, old_last, start, last);
1013 		svm_range_free(*new);
1014 		*new = NULL;
1015 	}
1016 
1017 	return r;
1018 }
1019 
1020 static int
1021 svm_range_split_tail(struct svm_range *prange,
1022 		     uint64_t new_last, struct list_head *insert_list)
1023 {
1024 	struct svm_range *tail;
1025 	int r = svm_range_split(prange, prange->start, new_last, &tail);
1026 
1027 	if (!r)
1028 		list_add(&tail->list, insert_list);
1029 	return r;
1030 }
1031 
1032 static int
1033 svm_range_split_head(struct svm_range *prange,
1034 		     uint64_t new_start, struct list_head *insert_list)
1035 {
1036 	struct svm_range *head;
1037 	int r = svm_range_split(prange, new_start, prange->last, &head);
1038 
1039 	if (!r)
1040 		list_add(&head->list, insert_list);
1041 	return r;
1042 }
1043 
1044 static void
1045 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1046 		    struct svm_range *pchild, enum svm_work_list_ops op)
1047 {
1048 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1049 		 pchild, pchild->start, pchild->last, prange, op);
1050 
1051 	pchild->work_item.mm = mm;
1052 	pchild->work_item.op = op;
1053 	list_add_tail(&pchild->child_list, &prange->child_list);
1054 }
1055 
1056 /**
1057  * svm_range_split_by_granularity - collect ranges within granularity boundary
1058  *
1059  * @p: the process with svms list
1060  * @mm: mm structure
1061  * @addr: the vm fault address in pages, to split the prange
1062  * @parent: parent range if prange is from child list
1063  * @prange: prange to split
1064  *
1065  * Trims @prange to be a single aligned block of prange->granularity if
1066  * possible. The head and tail are added to the child_list in @parent.
1067  *
1068  * Context: caller must hold mmap_read_lock and prange->lock
1069  *
1070  * Return:
1071  * 0 - OK, otherwise error code
1072  */
1073 int
1074 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1075 			       unsigned long addr, struct svm_range *parent,
1076 			       struct svm_range *prange)
1077 {
1078 	struct svm_range *head, *tail;
1079 	unsigned long start, last, size;
1080 	int r;
1081 
1082 	/* Align splited range start and size to granularity size, then a single
1083 	 * PTE will be used for whole range, this reduces the number of PTE
1084 	 * updated and the L1 TLB space used for translation.
1085 	 */
1086 	size = 1UL << prange->granularity;
1087 	start = ALIGN_DOWN(addr, size);
1088 	last = ALIGN(addr + 1, size) - 1;
1089 
1090 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1091 		 prange->svms, prange->start, prange->last, start, last, size);
1092 
1093 	if (start > prange->start) {
1094 		r = svm_range_split(prange, start, prange->last, &head);
1095 		if (r)
1096 			return r;
1097 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1098 	}
1099 
1100 	if (last < prange->last) {
1101 		r = svm_range_split(prange, prange->start, last, &tail);
1102 		if (r)
1103 			return r;
1104 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1105 	}
1106 
1107 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1108 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1109 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1110 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1111 			 prange, prange->start, prange->last,
1112 			 SVM_OP_ADD_RANGE_AND_MAP);
1113 	}
1114 	return 0;
1115 }
1116 
1117 static uint64_t
1118 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1119 			int domain)
1120 {
1121 	struct amdgpu_device *bo_adev;
1122 	uint32_t flags = prange->flags;
1123 	uint32_t mapping_flags = 0;
1124 	uint64_t pte_flags;
1125 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1126 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1127 
1128 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1129 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1130 
1131 	switch (KFD_GC_VERSION(adev->kfd.dev)) {
1132 	case IP_VERSION(9, 4, 1):
1133 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1134 			if (bo_adev == adev) {
1135 				mapping_flags |= coherent ?
1136 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1137 			} else {
1138 				mapping_flags |= coherent ?
1139 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1140 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1141 					snoop = true;
1142 			}
1143 		} else {
1144 			mapping_flags |= coherent ?
1145 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1146 		}
1147 		break;
1148 	case IP_VERSION(9, 4, 2):
1149 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1150 			if (bo_adev == adev) {
1151 				mapping_flags |= coherent ?
1152 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1153 				if (adev->gmc.xgmi.connected_to_cpu)
1154 					snoop = true;
1155 			} else {
1156 				mapping_flags |= coherent ?
1157 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1158 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
1159 					snoop = true;
1160 			}
1161 		} else {
1162 			mapping_flags |= coherent ?
1163 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1164 		}
1165 		break;
1166 	default:
1167 		mapping_flags |= coherent ?
1168 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1169 	}
1170 
1171 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1172 
1173 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1174 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1175 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1176 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1177 
1178 	pte_flags = AMDGPU_PTE_VALID;
1179 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1180 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1181 
1182 	pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1183 	return pte_flags;
1184 }
1185 
1186 static int
1187 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1188 			 uint64_t start, uint64_t last,
1189 			 struct dma_fence **fence)
1190 {
1191 	uint64_t init_pte_value = 0;
1192 
1193 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1194 
1195 	return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1196 				      last, init_pte_value, 0, 0, NULL, NULL,
1197 				      fence);
1198 }
1199 
1200 static int
1201 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1202 			  unsigned long last)
1203 {
1204 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1205 	struct kfd_process_device *pdd;
1206 	struct dma_fence *fence = NULL;
1207 	struct kfd_process *p;
1208 	uint32_t gpuidx;
1209 	int r = 0;
1210 
1211 	if (!prange->mapped_to_gpu) {
1212 		pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1213 			 prange, prange->start, prange->last);
1214 		return 0;
1215 	}
1216 
1217 	if (prange->start == start && prange->last == last) {
1218 		pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1219 		prange->mapped_to_gpu = false;
1220 	}
1221 
1222 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1223 		  MAX_GPU_INSTANCE);
1224 	p = container_of(prange->svms, struct kfd_process, svms);
1225 
1226 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1227 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1228 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1229 		if (!pdd) {
1230 			pr_debug("failed to find device idx %d\n", gpuidx);
1231 			return -EINVAL;
1232 		}
1233 
1234 		r = svm_range_unmap_from_gpu(pdd->dev->adev,
1235 					     drm_priv_to_vm(pdd->drm_priv),
1236 					     start, last, &fence);
1237 		if (r)
1238 			break;
1239 
1240 		if (fence) {
1241 			r = dma_fence_wait(fence, false);
1242 			dma_fence_put(fence);
1243 			fence = NULL;
1244 			if (r)
1245 				break;
1246 		}
1247 		kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1248 	}
1249 
1250 	return r;
1251 }
1252 
1253 static int
1254 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1255 		     unsigned long offset, unsigned long npages, bool readonly,
1256 		     dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1257 		     struct dma_fence **fence, bool flush_tlb)
1258 {
1259 	struct amdgpu_device *adev = pdd->dev->adev;
1260 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1261 	uint64_t pte_flags;
1262 	unsigned long last_start;
1263 	int last_domain;
1264 	int r = 0;
1265 	int64_t i, j;
1266 
1267 	last_start = prange->start + offset;
1268 
1269 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1270 		 last_start, last_start + npages - 1, readonly);
1271 
1272 	for (i = offset; i < offset + npages; i++) {
1273 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1274 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1275 
1276 		/* Collect all pages in the same address range and memory domain
1277 		 * that can be mapped with a single call to update mapping.
1278 		 */
1279 		if (i < offset + npages - 1 &&
1280 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1281 			continue;
1282 
1283 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1284 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1285 
1286 		pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1287 		if (readonly)
1288 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1289 
1290 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1291 			 prange->svms, last_start, prange->start + i,
1292 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1293 			 pte_flags);
1294 
1295 		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL,
1296 					   last_start, prange->start + i,
1297 					   pte_flags,
1298 					   (last_start - prange->start) << PAGE_SHIFT,
1299 					   bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1300 					   NULL, dma_addr, &vm->last_update);
1301 
1302 		for (j = last_start - prange->start; j <= i; j++)
1303 			dma_addr[j] |= last_domain;
1304 
1305 		if (r) {
1306 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1307 			goto out;
1308 		}
1309 		last_start = prange->start + i + 1;
1310 	}
1311 
1312 	r = amdgpu_vm_update_pdes(adev, vm, false);
1313 	if (r) {
1314 		pr_debug("failed %d to update directories 0x%lx\n", r,
1315 			 prange->start);
1316 		goto out;
1317 	}
1318 
1319 	if (fence)
1320 		*fence = dma_fence_get(vm->last_update);
1321 
1322 out:
1323 	return r;
1324 }
1325 
1326 static int
1327 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1328 		      unsigned long npages, bool readonly,
1329 		      unsigned long *bitmap, bool wait, bool flush_tlb)
1330 {
1331 	struct kfd_process_device *pdd;
1332 	struct amdgpu_device *bo_adev;
1333 	struct kfd_process *p;
1334 	struct dma_fence *fence = NULL;
1335 	uint32_t gpuidx;
1336 	int r = 0;
1337 
1338 	if (prange->svm_bo && prange->ttm_res)
1339 		bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1340 	else
1341 		bo_adev = NULL;
1342 
1343 	p = container_of(prange->svms, struct kfd_process, svms);
1344 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1345 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1346 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1347 		if (!pdd) {
1348 			pr_debug("failed to find device idx %d\n", gpuidx);
1349 			return -EINVAL;
1350 		}
1351 
1352 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1353 		if (IS_ERR(pdd))
1354 			return -EINVAL;
1355 
1356 		if (bo_adev && pdd->dev->adev != bo_adev &&
1357 		    !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1358 			pr_debug("cannot map to device idx %d\n", gpuidx);
1359 			continue;
1360 		}
1361 
1362 		r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1363 					 prange->dma_addr[gpuidx],
1364 					 bo_adev, wait ? &fence : NULL,
1365 					 flush_tlb);
1366 		if (r)
1367 			break;
1368 
1369 		if (fence) {
1370 			r = dma_fence_wait(fence, false);
1371 			dma_fence_put(fence);
1372 			fence = NULL;
1373 			if (r) {
1374 				pr_debug("failed %d to dma fence wait\n", r);
1375 				break;
1376 			}
1377 		}
1378 
1379 		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1380 	}
1381 
1382 	return r;
1383 }
1384 
1385 struct svm_validate_context {
1386 	struct kfd_process *process;
1387 	struct svm_range *prange;
1388 	bool intr;
1389 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1390 	struct ttm_validate_buffer tv[MAX_GPU_INSTANCE];
1391 	struct list_head validate_list;
1392 	struct ww_acquire_ctx ticket;
1393 };
1394 
1395 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1396 {
1397 	struct kfd_process_device *pdd;
1398 	struct amdgpu_vm *vm;
1399 	uint32_t gpuidx;
1400 	int r;
1401 
1402 	INIT_LIST_HEAD(&ctx->validate_list);
1403 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1404 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1405 		if (!pdd) {
1406 			pr_debug("failed to find device idx %d\n", gpuidx);
1407 			return -EINVAL;
1408 		}
1409 		vm = drm_priv_to_vm(pdd->drm_priv);
1410 
1411 		ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1412 		ctx->tv[gpuidx].num_shared = 4;
1413 		list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1414 	}
1415 
1416 	r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1417 				   ctx->intr, NULL);
1418 	if (r) {
1419 		pr_debug("failed %d to reserve bo\n", r);
1420 		return r;
1421 	}
1422 
1423 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1424 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1425 		if (!pdd) {
1426 			pr_debug("failed to find device idx %d\n", gpuidx);
1427 			r = -EINVAL;
1428 			goto unreserve_out;
1429 		}
1430 
1431 		r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1432 					      drm_priv_to_vm(pdd->drm_priv),
1433 					      svm_range_bo_validate, NULL);
1434 		if (r) {
1435 			pr_debug("failed %d validate pt bos\n", r);
1436 			goto unreserve_out;
1437 		}
1438 	}
1439 
1440 	return 0;
1441 
1442 unreserve_out:
1443 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1444 	return r;
1445 }
1446 
1447 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1448 {
1449 	ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1450 }
1451 
1452 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1453 {
1454 	struct kfd_process_device *pdd;
1455 
1456 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1457 
1458 	return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1459 }
1460 
1461 /*
1462  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1463  *
1464  * To prevent concurrent destruction or change of range attributes, the
1465  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1466  * because that would block concurrent evictions and lead to deadlocks. To
1467  * serialize concurrent migrations or validations of the same range, the
1468  * prange->migrate_mutex must be held.
1469  *
1470  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1471  * eviction fence.
1472  *
1473  * The following sequence ensures race-free validation and GPU mapping:
1474  *
1475  * 1. Reserve page table (and SVM BO if range is in VRAM)
1476  * 2. hmm_range_fault to get page addresses (if system memory)
1477  * 3. DMA-map pages (if system memory)
1478  * 4-a. Take notifier lock
1479  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1480  * 4-c. Check that the range was not split or otherwise invalidated
1481  * 4-d. Update GPU page table
1482  * 4.e. Release notifier lock
1483  * 5. Release page table (and SVM BO) reservation
1484  */
1485 static int svm_range_validate_and_map(struct mm_struct *mm,
1486 				      struct svm_range *prange, int32_t gpuidx,
1487 				      bool intr, bool wait, bool flush_tlb)
1488 {
1489 	struct svm_validate_context ctx;
1490 	unsigned long start, end, addr;
1491 	struct kfd_process *p;
1492 	void *owner;
1493 	int32_t idx;
1494 	int r = 0;
1495 
1496 	ctx.process = container_of(prange->svms, struct kfd_process, svms);
1497 	ctx.prange = prange;
1498 	ctx.intr = intr;
1499 
1500 	if (gpuidx < MAX_GPU_INSTANCE) {
1501 		bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1502 		bitmap_set(ctx.bitmap, gpuidx, 1);
1503 	} else if (ctx.process->xnack_enabled) {
1504 		bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1505 
1506 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1507 		 * GPU, which has ACCESS attribute to the range, create mapping
1508 		 * on that GPU.
1509 		 */
1510 		if (prange->actual_loc) {
1511 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1512 							prange->actual_loc);
1513 			if (gpuidx < 0) {
1514 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1515 					 prange->actual_loc);
1516 				return -EINVAL;
1517 			}
1518 			if (test_bit(gpuidx, prange->bitmap_access))
1519 				bitmap_set(ctx.bitmap, gpuidx, 1);
1520 		}
1521 	} else {
1522 		bitmap_or(ctx.bitmap, prange->bitmap_access,
1523 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1524 	}
1525 
1526 	if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) {
1527 		if (!prange->mapped_to_gpu)
1528 			return 0;
1529 
1530 		bitmap_copy(ctx.bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1531 	}
1532 
1533 	if (prange->actual_loc && !prange->ttm_res) {
1534 		/* This should never happen. actual_loc gets set by
1535 		 * svm_migrate_ram_to_vram after allocating a BO.
1536 		 */
1537 		WARN_ONCE(1, "VRAM BO missing during validation\n");
1538 		return -EINVAL;
1539 	}
1540 
1541 	svm_range_reserve_bos(&ctx);
1542 
1543 	p = container_of(prange->svms, struct kfd_process, svms);
1544 	owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1545 						MAX_GPU_INSTANCE));
1546 	for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1547 		if (kfd_svm_page_owner(p, idx) != owner) {
1548 			owner = NULL;
1549 			break;
1550 		}
1551 	}
1552 
1553 	start = prange->start << PAGE_SHIFT;
1554 	end = (prange->last + 1) << PAGE_SHIFT;
1555 	for (addr = start; addr < end && !r; ) {
1556 		struct hmm_range *hmm_range;
1557 		struct vm_area_struct *vma;
1558 		unsigned long next;
1559 		unsigned long offset;
1560 		unsigned long npages;
1561 		bool readonly;
1562 
1563 		vma = find_vma(mm, addr);
1564 		if (!vma || addr < vma->vm_start) {
1565 			r = -EFAULT;
1566 			goto unreserve_out;
1567 		}
1568 		readonly = !(vma->vm_flags & VM_WRITE);
1569 
1570 		next = min(vma->vm_end, end);
1571 		npages = (next - addr) >> PAGE_SHIFT;
1572 		WRITE_ONCE(p->svms.faulting_task, current);
1573 		r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1574 					       addr, npages, &hmm_range,
1575 					       readonly, true, owner);
1576 		WRITE_ONCE(p->svms.faulting_task, NULL);
1577 		if (r) {
1578 			pr_debug("failed %d to get svm range pages\n", r);
1579 			goto unreserve_out;
1580 		}
1581 
1582 		offset = (addr - start) >> PAGE_SHIFT;
1583 		r = svm_range_dma_map(prange, ctx.bitmap, offset, npages,
1584 				      hmm_range->hmm_pfns);
1585 		if (r) {
1586 			pr_debug("failed %d to dma map range\n", r);
1587 			goto unreserve_out;
1588 		}
1589 
1590 		svm_range_lock(prange);
1591 		if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1592 			pr_debug("hmm update the range, need validate again\n");
1593 			r = -EAGAIN;
1594 			goto unlock_out;
1595 		}
1596 		if (!list_empty(&prange->child_list)) {
1597 			pr_debug("range split by unmap in parallel, validate again\n");
1598 			r = -EAGAIN;
1599 			goto unlock_out;
1600 		}
1601 
1602 		r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1603 					  ctx.bitmap, wait, flush_tlb);
1604 
1605 unlock_out:
1606 		svm_range_unlock(prange);
1607 
1608 		addr = next;
1609 	}
1610 
1611 	if (addr == end) {
1612 		prange->validated_once = true;
1613 		prange->mapped_to_gpu = true;
1614 	}
1615 
1616 unreserve_out:
1617 	svm_range_unreserve_bos(&ctx);
1618 
1619 	if (!r)
1620 		prange->validate_timestamp = ktime_to_us(ktime_get());
1621 
1622 	return r;
1623 }
1624 
1625 /**
1626  * svm_range_list_lock_and_flush_work - flush pending deferred work
1627  *
1628  * @svms: the svm range list
1629  * @mm: the mm structure
1630  *
1631  * Context: Returns with mmap write lock held, pending deferred work flushed
1632  *
1633  */
1634 void
1635 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1636 				   struct mm_struct *mm)
1637 {
1638 retry_flush_work:
1639 	flush_work(&svms->deferred_list_work);
1640 	mmap_write_lock(mm);
1641 
1642 	if (list_empty(&svms->deferred_range_list))
1643 		return;
1644 	mmap_write_unlock(mm);
1645 	pr_debug("retry flush\n");
1646 	goto retry_flush_work;
1647 }
1648 
1649 static void svm_range_restore_work(struct work_struct *work)
1650 {
1651 	struct delayed_work *dwork = to_delayed_work(work);
1652 	struct amdkfd_process_info *process_info;
1653 	struct svm_range_list *svms;
1654 	struct svm_range *prange;
1655 	struct kfd_process *p;
1656 	struct mm_struct *mm;
1657 	int evicted_ranges;
1658 	int invalid;
1659 	int r;
1660 
1661 	svms = container_of(dwork, struct svm_range_list, restore_work);
1662 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1663 	if (!evicted_ranges)
1664 		return;
1665 
1666 	pr_debug("restore svm ranges\n");
1667 
1668 	p = container_of(svms, struct kfd_process, svms);
1669 	process_info = p->kgd_process_info;
1670 
1671 	/* Keep mm reference when svm_range_validate_and_map ranges */
1672 	mm = get_task_mm(p->lead_thread);
1673 	if (!mm) {
1674 		pr_debug("svms 0x%p process mm gone\n", svms);
1675 		return;
1676 	}
1677 
1678 	mutex_lock(&process_info->lock);
1679 	svm_range_list_lock_and_flush_work(svms, mm);
1680 	mutex_lock(&svms->lock);
1681 
1682 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1683 
1684 	list_for_each_entry(prange, &svms->list, list) {
1685 		invalid = atomic_read(&prange->invalid);
1686 		if (!invalid)
1687 			continue;
1688 
1689 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1690 			 prange->svms, prange, prange->start, prange->last,
1691 			 invalid);
1692 
1693 		/*
1694 		 * If range is migrating, wait for migration is done.
1695 		 */
1696 		mutex_lock(&prange->migrate_mutex);
1697 
1698 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1699 					       false, true, false);
1700 		if (r)
1701 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1702 				 prange->start);
1703 
1704 		mutex_unlock(&prange->migrate_mutex);
1705 		if (r)
1706 			goto out_reschedule;
1707 
1708 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1709 			goto out_reschedule;
1710 	}
1711 
1712 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1713 	    evicted_ranges)
1714 		goto out_reschedule;
1715 
1716 	evicted_ranges = 0;
1717 
1718 	r = kgd2kfd_resume_mm(mm);
1719 	if (r) {
1720 		/* No recovery from this failure. Probably the CP is
1721 		 * hanging. No point trying again.
1722 		 */
1723 		pr_debug("failed %d to resume KFD\n", r);
1724 	}
1725 
1726 	pr_debug("restore svm ranges successfully\n");
1727 
1728 out_reschedule:
1729 	mutex_unlock(&svms->lock);
1730 	mmap_write_unlock(mm);
1731 	mutex_unlock(&process_info->lock);
1732 	mmput(mm);
1733 
1734 	/* If validation failed, reschedule another attempt */
1735 	if (evicted_ranges) {
1736 		pr_debug("reschedule to restore svm range\n");
1737 		schedule_delayed_work(&svms->restore_work,
1738 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1739 	}
1740 }
1741 
1742 /**
1743  * svm_range_evict - evict svm range
1744  * @prange: svm range structure
1745  * @mm: current process mm_struct
1746  * @start: starting process queue number
1747  * @last: last process queue number
1748  *
1749  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1750  * return to let CPU evict the buffer and proceed CPU pagetable update.
1751  *
1752  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1753  * If invalidation happens while restore work is running, restore work will
1754  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1755  * the queues.
1756  */
1757 static int
1758 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1759 		unsigned long start, unsigned long last)
1760 {
1761 	struct svm_range_list *svms = prange->svms;
1762 	struct svm_range *pchild;
1763 	struct kfd_process *p;
1764 	int r = 0;
1765 
1766 	p = container_of(svms, struct kfd_process, svms);
1767 
1768 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1769 		 svms, prange->start, prange->last, start, last);
1770 
1771 	if (!p->xnack_enabled) {
1772 		int evicted_ranges;
1773 
1774 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1775 			mutex_lock_nested(&pchild->lock, 1);
1776 			if (pchild->start <= last && pchild->last >= start) {
1777 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1778 					 pchild->start, pchild->last);
1779 				atomic_inc(&pchild->invalid);
1780 			}
1781 			mutex_unlock(&pchild->lock);
1782 		}
1783 
1784 		if (prange->start <= last && prange->last >= start)
1785 			atomic_inc(&prange->invalid);
1786 
1787 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1788 		if (evicted_ranges != 1)
1789 			return r;
1790 
1791 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1792 			 prange->svms, prange->start, prange->last);
1793 
1794 		/* First eviction, stop the queues */
1795 		r = kgd2kfd_quiesce_mm(mm);
1796 		if (r)
1797 			pr_debug("failed to quiesce KFD\n");
1798 
1799 		pr_debug("schedule to restore svm %p ranges\n", svms);
1800 		schedule_delayed_work(&svms->restore_work,
1801 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1802 	} else {
1803 		unsigned long s, l;
1804 
1805 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1806 			 prange->svms, start, last);
1807 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1808 			mutex_lock_nested(&pchild->lock, 1);
1809 			s = max(start, pchild->start);
1810 			l = min(last, pchild->last);
1811 			if (l >= s)
1812 				svm_range_unmap_from_gpus(pchild, s, l);
1813 			mutex_unlock(&pchild->lock);
1814 		}
1815 		s = max(start, prange->start);
1816 		l = min(last, prange->last);
1817 		if (l >= s)
1818 			svm_range_unmap_from_gpus(prange, s, l);
1819 	}
1820 
1821 	return r;
1822 }
1823 
1824 static struct svm_range *svm_range_clone(struct svm_range *old)
1825 {
1826 	struct svm_range *new;
1827 
1828 	new = svm_range_new(old->svms, old->start, old->last);
1829 	if (!new)
1830 		return NULL;
1831 
1832 	if (old->svm_bo) {
1833 		new->ttm_res = old->ttm_res;
1834 		new->offset = old->offset;
1835 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1836 		spin_lock(&new->svm_bo->list_lock);
1837 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1838 		spin_unlock(&new->svm_bo->list_lock);
1839 	}
1840 	new->flags = old->flags;
1841 	new->preferred_loc = old->preferred_loc;
1842 	new->prefetch_loc = old->prefetch_loc;
1843 	new->actual_loc = old->actual_loc;
1844 	new->granularity = old->granularity;
1845 	new->mapped_to_gpu = old->mapped_to_gpu;
1846 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1847 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1848 
1849 	return new;
1850 }
1851 
1852 /**
1853  * svm_range_add - add svm range and handle overlap
1854  * @p: the range add to this process svms
1855  * @start: page size aligned
1856  * @size: page size aligned
1857  * @nattr: number of attributes
1858  * @attrs: array of attributes
1859  * @update_list: output, the ranges need validate and update GPU mapping
1860  * @insert_list: output, the ranges need insert to svms
1861  * @remove_list: output, the ranges are replaced and need remove from svms
1862  *
1863  * Check if the virtual address range has overlap with any existing ranges,
1864  * split partly overlapping ranges and add new ranges in the gaps. All changes
1865  * should be applied to the range_list and interval tree transactionally. If
1866  * any range split or allocation fails, the entire update fails. Therefore any
1867  * existing overlapping svm_ranges are cloned and the original svm_ranges left
1868  * unchanged.
1869  *
1870  * If the transaction succeeds, the caller can update and insert clones and
1871  * new ranges, then free the originals.
1872  *
1873  * Otherwise the caller can free the clones and new ranges, while the old
1874  * svm_ranges remain unchanged.
1875  *
1876  * Context: Process context, caller must hold svms->lock
1877  *
1878  * Return:
1879  * 0 - OK, otherwise error code
1880  */
1881 static int
1882 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
1883 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
1884 	      struct list_head *update_list, struct list_head *insert_list,
1885 	      struct list_head *remove_list)
1886 {
1887 	unsigned long last = start + size - 1UL;
1888 	struct svm_range_list *svms = &p->svms;
1889 	struct interval_tree_node *node;
1890 	struct svm_range *prange;
1891 	struct svm_range *tmp;
1892 	int r = 0;
1893 
1894 	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
1895 
1896 	INIT_LIST_HEAD(update_list);
1897 	INIT_LIST_HEAD(insert_list);
1898 	INIT_LIST_HEAD(remove_list);
1899 
1900 	node = interval_tree_iter_first(&svms->objects, start, last);
1901 	while (node) {
1902 		struct interval_tree_node *next;
1903 		unsigned long next_start;
1904 
1905 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
1906 			 node->last);
1907 
1908 		prange = container_of(node, struct svm_range, it_node);
1909 		next = interval_tree_iter_next(node, start, last);
1910 		next_start = min(node->last, last) + 1;
1911 
1912 		if (svm_range_is_same_attrs(p, prange, nattr, attrs)) {
1913 			/* nothing to do */
1914 		} else if (node->start < start || node->last > last) {
1915 			/* node intersects the update range and its attributes
1916 			 * will change. Clone and split it, apply updates only
1917 			 * to the overlapping part
1918 			 */
1919 			struct svm_range *old = prange;
1920 
1921 			prange = svm_range_clone(old);
1922 			if (!prange) {
1923 				r = -ENOMEM;
1924 				goto out;
1925 			}
1926 
1927 			list_add(&old->update_list, remove_list);
1928 			list_add(&prange->list, insert_list);
1929 			list_add(&prange->update_list, update_list);
1930 
1931 			if (node->start < start) {
1932 				pr_debug("change old range start\n");
1933 				r = svm_range_split_head(prange, start,
1934 							 insert_list);
1935 				if (r)
1936 					goto out;
1937 			}
1938 			if (node->last > last) {
1939 				pr_debug("change old range last\n");
1940 				r = svm_range_split_tail(prange, last,
1941 							 insert_list);
1942 				if (r)
1943 					goto out;
1944 			}
1945 		} else {
1946 			/* The node is contained within start..last,
1947 			 * just update it
1948 			 */
1949 			list_add(&prange->update_list, update_list);
1950 		}
1951 
1952 		/* insert a new node if needed */
1953 		if (node->start > start) {
1954 			prange = svm_range_new(svms, start, node->start - 1);
1955 			if (!prange) {
1956 				r = -ENOMEM;
1957 				goto out;
1958 			}
1959 
1960 			list_add(&prange->list, insert_list);
1961 			list_add(&prange->update_list, update_list);
1962 		}
1963 
1964 		node = next;
1965 		start = next_start;
1966 	}
1967 
1968 	/* add a final range at the end if needed */
1969 	if (start <= last) {
1970 		prange = svm_range_new(svms, start, last);
1971 		if (!prange) {
1972 			r = -ENOMEM;
1973 			goto out;
1974 		}
1975 		list_add(&prange->list, insert_list);
1976 		list_add(&prange->update_list, update_list);
1977 	}
1978 
1979 out:
1980 	if (r)
1981 		list_for_each_entry_safe(prange, tmp, insert_list, list)
1982 			svm_range_free(prange);
1983 
1984 	return r;
1985 }
1986 
1987 static void
1988 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
1989 					    struct svm_range *prange)
1990 {
1991 	unsigned long start;
1992 	unsigned long last;
1993 
1994 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
1995 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
1996 
1997 	if (prange->start == start && prange->last == last)
1998 		return;
1999 
2000 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2001 		  prange->svms, prange, start, last, prange->start,
2002 		  prange->last);
2003 
2004 	if (start != 0 && last != 0) {
2005 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
2006 		svm_range_remove_notifier(prange);
2007 	}
2008 	prange->it_node.start = prange->start;
2009 	prange->it_node.last = prange->last;
2010 
2011 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
2012 	svm_range_add_notifier_locked(mm, prange);
2013 }
2014 
2015 static void
2016 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2017 			 struct mm_struct *mm)
2018 {
2019 	switch (prange->work_item.op) {
2020 	case SVM_OP_NULL:
2021 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2022 			 svms, prange, prange->start, prange->last);
2023 		break;
2024 	case SVM_OP_UNMAP_RANGE:
2025 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2026 			 svms, prange, prange->start, prange->last);
2027 		svm_range_unlink(prange);
2028 		svm_range_remove_notifier(prange);
2029 		svm_range_free(prange);
2030 		break;
2031 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
2032 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2033 			 svms, prange, prange->start, prange->last);
2034 		svm_range_update_notifier_and_interval_tree(mm, prange);
2035 		break;
2036 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2037 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2038 			 svms, prange, prange->start, prange->last);
2039 		svm_range_update_notifier_and_interval_tree(mm, prange);
2040 		/* TODO: implement deferred validation and mapping */
2041 		break;
2042 	case SVM_OP_ADD_RANGE:
2043 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2044 			 prange->start, prange->last);
2045 		svm_range_add_to_svms(prange);
2046 		svm_range_add_notifier_locked(mm, prange);
2047 		break;
2048 	case SVM_OP_ADD_RANGE_AND_MAP:
2049 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2050 			 prange, prange->start, prange->last);
2051 		svm_range_add_to_svms(prange);
2052 		svm_range_add_notifier_locked(mm, prange);
2053 		/* TODO: implement deferred validation and mapping */
2054 		break;
2055 	default:
2056 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2057 			 prange->work_item.op);
2058 	}
2059 }
2060 
2061 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2062 {
2063 	struct kfd_process_device *pdd;
2064 	struct kfd_process *p;
2065 	int drain;
2066 	uint32_t i;
2067 
2068 	p = container_of(svms, struct kfd_process, svms);
2069 
2070 restart:
2071 	drain = atomic_read(&svms->drain_pagefaults);
2072 	if (!drain)
2073 		return;
2074 
2075 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2076 		pdd = p->pdds[i];
2077 		if (!pdd)
2078 			continue;
2079 
2080 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2081 
2082 		amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2083 						     &pdd->dev->adev->irq.ih1);
2084 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2085 	}
2086 	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2087 		goto restart;
2088 }
2089 
2090 static void svm_range_deferred_list_work(struct work_struct *work)
2091 {
2092 	struct svm_range_list *svms;
2093 	struct svm_range *prange;
2094 	struct mm_struct *mm;
2095 
2096 	svms = container_of(work, struct svm_range_list, deferred_list_work);
2097 	pr_debug("enter svms 0x%p\n", svms);
2098 
2099 	spin_lock(&svms->deferred_list_lock);
2100 	while (!list_empty(&svms->deferred_range_list)) {
2101 		prange = list_first_entry(&svms->deferred_range_list,
2102 					  struct svm_range, deferred_list);
2103 		spin_unlock(&svms->deferred_list_lock);
2104 
2105 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2106 			 prange->start, prange->last, prange->work_item.op);
2107 
2108 		mm = prange->work_item.mm;
2109 retry:
2110 		mmap_write_lock(mm);
2111 
2112 		/* Checking for the need to drain retry faults must be inside
2113 		 * mmap write lock to serialize with munmap notifiers.
2114 		 */
2115 		if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2116 			mmap_write_unlock(mm);
2117 			svm_range_drain_retry_fault(svms);
2118 			goto retry;
2119 		}
2120 
2121 		/* Remove from deferred_list must be inside mmap write lock, for
2122 		 * two race cases:
2123 		 * 1. unmap_from_cpu may change work_item.op and add the range
2124 		 *    to deferred_list again, cause use after free bug.
2125 		 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2126 		 *    lock and continue because deferred_list is empty, but
2127 		 *    deferred_list work is actually waiting for mmap lock.
2128 		 */
2129 		spin_lock(&svms->deferred_list_lock);
2130 		list_del_init(&prange->deferred_list);
2131 		spin_unlock(&svms->deferred_list_lock);
2132 
2133 		mutex_lock(&svms->lock);
2134 		mutex_lock(&prange->migrate_mutex);
2135 		while (!list_empty(&prange->child_list)) {
2136 			struct svm_range *pchild;
2137 
2138 			pchild = list_first_entry(&prange->child_list,
2139 						struct svm_range, child_list);
2140 			pr_debug("child prange 0x%p op %d\n", pchild,
2141 				 pchild->work_item.op);
2142 			list_del_init(&pchild->child_list);
2143 			svm_range_handle_list_op(svms, pchild, mm);
2144 		}
2145 		mutex_unlock(&prange->migrate_mutex);
2146 
2147 		svm_range_handle_list_op(svms, prange, mm);
2148 		mutex_unlock(&svms->lock);
2149 		mmap_write_unlock(mm);
2150 
2151 		/* Pairs with mmget in svm_range_add_list_work */
2152 		mmput(mm);
2153 
2154 		spin_lock(&svms->deferred_list_lock);
2155 	}
2156 	spin_unlock(&svms->deferred_list_lock);
2157 	pr_debug("exit svms 0x%p\n", svms);
2158 }
2159 
2160 void
2161 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2162 			struct mm_struct *mm, enum svm_work_list_ops op)
2163 {
2164 	spin_lock(&svms->deferred_list_lock);
2165 	/* if prange is on the deferred list */
2166 	if (!list_empty(&prange->deferred_list)) {
2167 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2168 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2169 		if (op != SVM_OP_NULL &&
2170 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2171 			prange->work_item.op = op;
2172 	} else {
2173 		prange->work_item.op = op;
2174 
2175 		/* Pairs with mmput in deferred_list_work */
2176 		mmget(mm);
2177 		prange->work_item.mm = mm;
2178 		list_add_tail(&prange->deferred_list,
2179 			      &prange->svms->deferred_range_list);
2180 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2181 			 prange, prange->start, prange->last, op);
2182 	}
2183 	spin_unlock(&svms->deferred_list_lock);
2184 }
2185 
2186 void schedule_deferred_list_work(struct svm_range_list *svms)
2187 {
2188 	spin_lock(&svms->deferred_list_lock);
2189 	if (!list_empty(&svms->deferred_range_list))
2190 		schedule_work(&svms->deferred_list_work);
2191 	spin_unlock(&svms->deferred_list_lock);
2192 }
2193 
2194 static void
2195 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2196 		      struct svm_range *prange, unsigned long start,
2197 		      unsigned long last)
2198 {
2199 	struct svm_range *head;
2200 	struct svm_range *tail;
2201 
2202 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2203 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2204 			 prange->start, prange->last);
2205 		return;
2206 	}
2207 	if (start > prange->last || last < prange->start)
2208 		return;
2209 
2210 	head = tail = prange;
2211 	if (start > prange->start)
2212 		svm_range_split(prange, prange->start, start - 1, &tail);
2213 	if (last < tail->last)
2214 		svm_range_split(tail, last + 1, tail->last, &head);
2215 
2216 	if (head != prange && tail != prange) {
2217 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2218 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2219 	} else if (tail != prange) {
2220 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2221 	} else if (head != prange) {
2222 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2223 	} else if (parent != prange) {
2224 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2225 	}
2226 }
2227 
2228 static void
2229 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2230 			 unsigned long start, unsigned long last)
2231 {
2232 	struct svm_range_list *svms;
2233 	struct svm_range *pchild;
2234 	struct kfd_process *p;
2235 	unsigned long s, l;
2236 	bool unmap_parent;
2237 
2238 	p = kfd_lookup_process_by_mm(mm);
2239 	if (!p)
2240 		return;
2241 	svms = &p->svms;
2242 
2243 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2244 		 prange, prange->start, prange->last, start, last);
2245 
2246 	/* Make sure pending page faults are drained in the deferred worker
2247 	 * before the range is freed to avoid straggler interrupts on
2248 	 * unmapped memory causing "phantom faults".
2249 	 */
2250 	atomic_inc(&svms->drain_pagefaults);
2251 
2252 	unmap_parent = start <= prange->start && last >= prange->last;
2253 
2254 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2255 		mutex_lock_nested(&pchild->lock, 1);
2256 		s = max(start, pchild->start);
2257 		l = min(last, pchild->last);
2258 		if (l >= s)
2259 			svm_range_unmap_from_gpus(pchild, s, l);
2260 		svm_range_unmap_split(mm, prange, pchild, start, last);
2261 		mutex_unlock(&pchild->lock);
2262 	}
2263 	s = max(start, prange->start);
2264 	l = min(last, prange->last);
2265 	if (l >= s)
2266 		svm_range_unmap_from_gpus(prange, s, l);
2267 	svm_range_unmap_split(mm, prange, prange, start, last);
2268 
2269 	if (unmap_parent)
2270 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2271 	else
2272 		svm_range_add_list_work(svms, prange, mm,
2273 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2274 	schedule_deferred_list_work(svms);
2275 
2276 	kfd_unref_process(p);
2277 }
2278 
2279 /**
2280  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2281  * @mni: mmu_interval_notifier struct
2282  * @range: mmu_notifier_range struct
2283  * @cur_seq: value to pass to mmu_interval_set_seq()
2284  *
2285  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2286  * is from migration, or CPU page invalidation callback.
2287  *
2288  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2289  * work thread, and split prange if only part of prange is unmapped.
2290  *
2291  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2292  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2293  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2294  * update GPU mapping to recover.
2295  *
2296  * Context: mmap lock, notifier_invalidate_start lock are held
2297  *          for invalidate event, prange lock is held if this is from migration
2298  */
2299 static bool
2300 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2301 				    const struct mmu_notifier_range *range,
2302 				    unsigned long cur_seq)
2303 {
2304 	struct svm_range *prange;
2305 	unsigned long start;
2306 	unsigned long last;
2307 
2308 	if (range->event == MMU_NOTIFY_RELEASE)
2309 		return true;
2310 	if (!mmget_not_zero(mni->mm))
2311 		return true;
2312 
2313 	start = mni->interval_tree.start;
2314 	last = mni->interval_tree.last;
2315 	start = max(start, range->start) >> PAGE_SHIFT;
2316 	last = min(last, range->end - 1) >> PAGE_SHIFT;
2317 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2318 		 start, last, range->start >> PAGE_SHIFT,
2319 		 (range->end - 1) >> PAGE_SHIFT,
2320 		 mni->interval_tree.start >> PAGE_SHIFT,
2321 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2322 
2323 	prange = container_of(mni, struct svm_range, notifier);
2324 
2325 	svm_range_lock(prange);
2326 	mmu_interval_set_seq(mni, cur_seq);
2327 
2328 	switch (range->event) {
2329 	case MMU_NOTIFY_UNMAP:
2330 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2331 		break;
2332 	default:
2333 		svm_range_evict(prange, mni->mm, start, last);
2334 		break;
2335 	}
2336 
2337 	svm_range_unlock(prange);
2338 	mmput(mni->mm);
2339 
2340 	return true;
2341 }
2342 
2343 /**
2344  * svm_range_from_addr - find svm range from fault address
2345  * @svms: svm range list header
2346  * @addr: address to search range interval tree, in pages
2347  * @parent: parent range if range is on child list
2348  *
2349  * Context: The caller must hold svms->lock
2350  *
2351  * Return: the svm_range found or NULL
2352  */
2353 struct svm_range *
2354 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2355 		    struct svm_range **parent)
2356 {
2357 	struct interval_tree_node *node;
2358 	struct svm_range *prange;
2359 	struct svm_range *pchild;
2360 
2361 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2362 	if (!node)
2363 		return NULL;
2364 
2365 	prange = container_of(node, struct svm_range, it_node);
2366 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2367 		 addr, prange->start, prange->last, node->start, node->last);
2368 
2369 	if (addr >= prange->start && addr <= prange->last) {
2370 		if (parent)
2371 			*parent = prange;
2372 		return prange;
2373 	}
2374 	list_for_each_entry(pchild, &prange->child_list, child_list)
2375 		if (addr >= pchild->start && addr <= pchild->last) {
2376 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2377 				 addr, pchild->start, pchild->last);
2378 			if (parent)
2379 				*parent = prange;
2380 			return pchild;
2381 		}
2382 
2383 	return NULL;
2384 }
2385 
2386 /* svm_range_best_restore_location - decide the best fault restore location
2387  * @prange: svm range structure
2388  * @adev: the GPU on which vm fault happened
2389  *
2390  * This is only called when xnack is on, to decide the best location to restore
2391  * the range mapping after GPU vm fault. Caller uses the best location to do
2392  * migration if actual loc is not best location, then update GPU page table
2393  * mapping to the best location.
2394  *
2395  * If the preferred loc is accessible by faulting GPU, use preferred loc.
2396  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2397  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2398  *    if range actual loc is cpu, best_loc is cpu
2399  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2400  *    range actual loc.
2401  * Otherwise, GPU no access, best_loc is -1.
2402  *
2403  * Return:
2404  * -1 means vm fault GPU no access
2405  * 0 for CPU or GPU id
2406  */
2407 static int32_t
2408 svm_range_best_restore_location(struct svm_range *prange,
2409 				struct amdgpu_device *adev,
2410 				int32_t *gpuidx)
2411 {
2412 	struct amdgpu_device *bo_adev, *preferred_adev;
2413 	struct kfd_process *p;
2414 	uint32_t gpuid;
2415 	int r;
2416 
2417 	p = container_of(prange->svms, struct kfd_process, svms);
2418 
2419 	r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx);
2420 	if (r < 0) {
2421 		pr_debug("failed to get gpuid from kgd\n");
2422 		return -1;
2423 	}
2424 
2425 	if (prange->preferred_loc == gpuid ||
2426 	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2427 		return prange->preferred_loc;
2428 	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2429 		preferred_adev = svm_range_get_adev_by_id(prange,
2430 							prange->preferred_loc);
2431 		if (amdgpu_xgmi_same_hive(adev, preferred_adev))
2432 			return prange->preferred_loc;
2433 		/* fall through */
2434 	}
2435 
2436 	if (test_bit(*gpuidx, prange->bitmap_access))
2437 		return gpuid;
2438 
2439 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2440 		if (!prange->actual_loc)
2441 			return 0;
2442 
2443 		bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2444 		if (amdgpu_xgmi_same_hive(adev, bo_adev))
2445 			return prange->actual_loc;
2446 		else
2447 			return 0;
2448 	}
2449 
2450 	return -1;
2451 }
2452 
2453 static int
2454 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2455 			       unsigned long *start, unsigned long *last,
2456 			       bool *is_heap_stack)
2457 {
2458 	struct vm_area_struct *vma;
2459 	struct interval_tree_node *node;
2460 	unsigned long start_limit, end_limit;
2461 
2462 	vma = find_vma(p->mm, addr << PAGE_SHIFT);
2463 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2464 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2465 		return -EFAULT;
2466 	}
2467 
2468 	*is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
2469 			  vma->vm_end >= vma->vm_mm->start_brk) ||
2470 			 (vma->vm_start <= vma->vm_mm->start_stack &&
2471 			  vma->vm_end >= vma->vm_mm->start_stack);
2472 
2473 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2474 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2475 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2476 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2477 	/* First range that starts after the fault address */
2478 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2479 	if (node) {
2480 		end_limit = min(end_limit, node->start);
2481 		/* Last range that ends before the fault address */
2482 		node = container_of(rb_prev(&node->rb),
2483 				    struct interval_tree_node, rb);
2484 	} else {
2485 		/* Last range must end before addr because
2486 		 * there was no range after addr
2487 		 */
2488 		node = container_of(rb_last(&p->svms.objects.rb_root),
2489 				    struct interval_tree_node, rb);
2490 	}
2491 	if (node) {
2492 		if (node->last >= addr) {
2493 			WARN(1, "Overlap with prev node and page fault addr\n");
2494 			return -EFAULT;
2495 		}
2496 		start_limit = max(start_limit, node->last + 1);
2497 	}
2498 
2499 	*start = start_limit;
2500 	*last = end_limit - 1;
2501 
2502 	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2503 		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2504 		 *start, *last, *is_heap_stack);
2505 
2506 	return 0;
2507 }
2508 
2509 static int
2510 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2511 			   uint64_t *bo_s, uint64_t *bo_l)
2512 {
2513 	struct amdgpu_bo_va_mapping *mapping;
2514 	struct interval_tree_node *node;
2515 	struct amdgpu_bo *bo = NULL;
2516 	unsigned long userptr;
2517 	uint32_t i;
2518 	int r;
2519 
2520 	for (i = 0; i < p->n_pdds; i++) {
2521 		struct amdgpu_vm *vm;
2522 
2523 		if (!p->pdds[i]->drm_priv)
2524 			continue;
2525 
2526 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2527 		r = amdgpu_bo_reserve(vm->root.bo, false);
2528 		if (r)
2529 			return r;
2530 
2531 		/* Check userptr by searching entire vm->va interval tree */
2532 		node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2533 		while (node) {
2534 			mapping = container_of((struct rb_node *)node,
2535 					       struct amdgpu_bo_va_mapping, rb);
2536 			bo = mapping->bo_va->base.bo;
2537 
2538 			if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2539 							 start << PAGE_SHIFT,
2540 							 last << PAGE_SHIFT,
2541 							 &userptr)) {
2542 				node = interval_tree_iter_next(node, 0, ~0ULL);
2543 				continue;
2544 			}
2545 
2546 			pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2547 				 start, last);
2548 			if (bo_s && bo_l) {
2549 				*bo_s = userptr >> PAGE_SHIFT;
2550 				*bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2551 			}
2552 			amdgpu_bo_unreserve(vm->root.bo);
2553 			return -EADDRINUSE;
2554 		}
2555 		amdgpu_bo_unreserve(vm->root.bo);
2556 	}
2557 	return 0;
2558 }
2559 
2560 static struct
2561 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2562 						struct kfd_process *p,
2563 						struct mm_struct *mm,
2564 						int64_t addr)
2565 {
2566 	struct svm_range *prange = NULL;
2567 	unsigned long start, last;
2568 	uint32_t gpuid, gpuidx;
2569 	bool is_heap_stack;
2570 	uint64_t bo_s = 0;
2571 	uint64_t bo_l = 0;
2572 	int r;
2573 
2574 	if (svm_range_get_range_boundaries(p, addr, &start, &last,
2575 					   &is_heap_stack))
2576 		return NULL;
2577 
2578 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2579 	if (r != -EADDRINUSE)
2580 		r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2581 
2582 	if (r == -EADDRINUSE) {
2583 		if (addr >= bo_s && addr <= bo_l)
2584 			return NULL;
2585 
2586 		/* Create one page svm range if 2MB range overlapping */
2587 		start = addr;
2588 		last = addr;
2589 	}
2590 
2591 	prange = svm_range_new(&p->svms, start, last);
2592 	if (!prange) {
2593 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2594 		return NULL;
2595 	}
2596 	if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) {
2597 		pr_debug("failed to get gpuid from kgd\n");
2598 		svm_range_free(prange);
2599 		return NULL;
2600 	}
2601 
2602 	if (is_heap_stack)
2603 		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2604 
2605 	svm_range_add_to_svms(prange);
2606 	svm_range_add_notifier_locked(mm, prange);
2607 
2608 	return prange;
2609 }
2610 
2611 /* svm_range_skip_recover - decide if prange can be recovered
2612  * @prange: svm range structure
2613  *
2614  * GPU vm retry fault handle skip recover the range for cases:
2615  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2616  *    deferred list work will drain the stale fault before free the prange.
2617  * 2. prange is on deferred list to add interval notifier after split, or
2618  * 3. prange is child range, it is split from parent prange, recover later
2619  *    after interval notifier is added.
2620  *
2621  * Return: true to skip recover, false to recover
2622  */
2623 static bool svm_range_skip_recover(struct svm_range *prange)
2624 {
2625 	struct svm_range_list *svms = prange->svms;
2626 
2627 	spin_lock(&svms->deferred_list_lock);
2628 	if (list_empty(&prange->deferred_list) &&
2629 	    list_empty(&prange->child_list)) {
2630 		spin_unlock(&svms->deferred_list_lock);
2631 		return false;
2632 	}
2633 	spin_unlock(&svms->deferred_list_lock);
2634 
2635 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2636 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2637 			 svms, prange, prange->start, prange->last);
2638 		return true;
2639 	}
2640 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2641 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2642 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2643 			 svms, prange, prange->start, prange->last);
2644 		return true;
2645 	}
2646 	return false;
2647 }
2648 
2649 static void
2650 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2651 		      int32_t gpuidx)
2652 {
2653 	struct kfd_process_device *pdd;
2654 
2655 	/* fault is on different page of same range
2656 	 * or fault is skipped to recover later
2657 	 * or fault is on invalid virtual address
2658 	 */
2659 	if (gpuidx == MAX_GPU_INSTANCE) {
2660 		uint32_t gpuid;
2661 		int r;
2662 
2663 		r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx);
2664 		if (r < 0)
2665 			return;
2666 	}
2667 
2668 	/* fault is recovered
2669 	 * or fault cannot recover because GPU no access on the range
2670 	 */
2671 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2672 	if (pdd)
2673 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2674 }
2675 
2676 static bool
2677 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2678 {
2679 	unsigned long requested = VM_READ;
2680 
2681 	if (write_fault)
2682 		requested |= VM_WRITE;
2683 
2684 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2685 		vma->vm_flags);
2686 	return (vma->vm_flags & requested) == requested;
2687 }
2688 
2689 int
2690 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2691 			uint64_t addr, bool write_fault)
2692 {
2693 	struct mm_struct *mm = NULL;
2694 	struct svm_range_list *svms;
2695 	struct svm_range *prange;
2696 	struct kfd_process *p;
2697 	uint64_t timestamp;
2698 	int32_t best_loc;
2699 	int32_t gpuidx = MAX_GPU_INSTANCE;
2700 	bool write_locked = false;
2701 	struct vm_area_struct *vma;
2702 	int r = 0;
2703 
2704 	if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2705 		pr_debug("device does not support SVM\n");
2706 		return -EFAULT;
2707 	}
2708 
2709 	p = kfd_lookup_process_by_pasid(pasid);
2710 	if (!p) {
2711 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2712 		return 0;
2713 	}
2714 	svms = &p->svms;
2715 
2716 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2717 
2718 	if (atomic_read(&svms->drain_pagefaults)) {
2719 		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2720 		r = 0;
2721 		goto out;
2722 	}
2723 
2724 	if (!p->xnack_enabled) {
2725 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2726 		r = -EFAULT;
2727 		goto out;
2728 	}
2729 
2730 	/* p->lead_thread is available as kfd_process_wq_release flush the work
2731 	 * before releasing task ref.
2732 	 */
2733 	mm = get_task_mm(p->lead_thread);
2734 	if (!mm) {
2735 		pr_debug("svms 0x%p failed to get mm\n", svms);
2736 		r = 0;
2737 		goto out;
2738 	}
2739 
2740 	mmap_read_lock(mm);
2741 retry_write_locked:
2742 	mutex_lock(&svms->lock);
2743 	prange = svm_range_from_addr(svms, addr, NULL);
2744 	if (!prange) {
2745 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2746 			 svms, addr);
2747 		if (!write_locked) {
2748 			/* Need the write lock to create new range with MMU notifier.
2749 			 * Also flush pending deferred work to make sure the interval
2750 			 * tree is up to date before we add a new range
2751 			 */
2752 			mutex_unlock(&svms->lock);
2753 			mmap_read_unlock(mm);
2754 			mmap_write_lock(mm);
2755 			write_locked = true;
2756 			goto retry_write_locked;
2757 		}
2758 		prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2759 		if (!prange) {
2760 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2761 				 svms, addr);
2762 			mmap_write_downgrade(mm);
2763 			r = -EFAULT;
2764 			goto out_unlock_svms;
2765 		}
2766 	}
2767 	if (write_locked)
2768 		mmap_write_downgrade(mm);
2769 
2770 	mutex_lock(&prange->migrate_mutex);
2771 
2772 	if (svm_range_skip_recover(prange)) {
2773 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2774 		r = 0;
2775 		goto out_unlock_range;
2776 	}
2777 
2778 	timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp;
2779 	/* skip duplicate vm fault on different pages of same range */
2780 	if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
2781 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2782 			 svms, prange->start, prange->last);
2783 		r = 0;
2784 		goto out_unlock_range;
2785 	}
2786 
2787 	/* __do_munmap removed VMA, return success as we are handling stale
2788 	 * retry fault.
2789 	 */
2790 	vma = find_vma(mm, addr << PAGE_SHIFT);
2791 	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2792 		pr_debug("address 0x%llx VMA is removed\n", addr);
2793 		r = 0;
2794 		goto out_unlock_range;
2795 	}
2796 
2797 	if (!svm_fault_allowed(vma, write_fault)) {
2798 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
2799 			write_fault ? "write" : "read");
2800 		r = -EPERM;
2801 		goto out_unlock_range;
2802 	}
2803 
2804 	best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2805 	if (best_loc == -1) {
2806 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2807 			 svms, prange->start, prange->last);
2808 		r = -EACCES;
2809 		goto out_unlock_range;
2810 	}
2811 
2812 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2813 		 svms, prange->start, prange->last, best_loc,
2814 		 prange->actual_loc);
2815 
2816 	if (prange->actual_loc != best_loc) {
2817 		if (best_loc) {
2818 			r = svm_migrate_to_vram(prange, best_loc, mm);
2819 			if (r) {
2820 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2821 					 r, addr);
2822 				/* Fallback to system memory if migration to
2823 				 * VRAM failed
2824 				 */
2825 				if (prange->actual_loc)
2826 					r = svm_migrate_vram_to_ram(prange, mm);
2827 				else
2828 					r = 0;
2829 			}
2830 		} else {
2831 			r = svm_migrate_vram_to_ram(prange, mm);
2832 		}
2833 		if (r) {
2834 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2835 				 r, svms, prange->start, prange->last);
2836 			goto out_unlock_range;
2837 		}
2838 	}
2839 
2840 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
2841 	if (r)
2842 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2843 			 r, svms, prange->start, prange->last);
2844 
2845 out_unlock_range:
2846 	mutex_unlock(&prange->migrate_mutex);
2847 out_unlock_svms:
2848 	mutex_unlock(&svms->lock);
2849 	mmap_read_unlock(mm);
2850 
2851 	svm_range_count_fault(adev, p, gpuidx);
2852 
2853 	mmput(mm);
2854 out:
2855 	kfd_unref_process(p);
2856 
2857 	if (r == -EAGAIN) {
2858 		pr_debug("recover vm fault later\n");
2859 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2860 		r = 0;
2861 	}
2862 	return r;
2863 }
2864 
2865 void svm_range_list_fini(struct kfd_process *p)
2866 {
2867 	struct svm_range *prange;
2868 	struct svm_range *next;
2869 
2870 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
2871 
2872 	cancel_delayed_work_sync(&p->svms.restore_work);
2873 
2874 	/* Ensure list work is finished before process is destroyed */
2875 	flush_work(&p->svms.deferred_list_work);
2876 
2877 	/*
2878 	 * Ensure no retry fault comes in afterwards, as page fault handler will
2879 	 * not find kfd process and take mm lock to recover fault.
2880 	 */
2881 	atomic_inc(&p->svms.drain_pagefaults);
2882 	svm_range_drain_retry_fault(&p->svms);
2883 
2884 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
2885 		svm_range_unlink(prange);
2886 		svm_range_remove_notifier(prange);
2887 		svm_range_free(prange);
2888 	}
2889 
2890 	mutex_destroy(&p->svms.lock);
2891 
2892 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
2893 }
2894 
2895 int svm_range_list_init(struct kfd_process *p)
2896 {
2897 	struct svm_range_list *svms = &p->svms;
2898 	int i;
2899 
2900 	svms->objects = RB_ROOT_CACHED;
2901 	mutex_init(&svms->lock);
2902 	INIT_LIST_HEAD(&svms->list);
2903 	atomic_set(&svms->evicted_ranges, 0);
2904 	atomic_set(&svms->drain_pagefaults, 0);
2905 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
2906 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
2907 	INIT_LIST_HEAD(&svms->deferred_range_list);
2908 	INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
2909 	spin_lock_init(&svms->deferred_list_lock);
2910 
2911 	for (i = 0; i < p->n_pdds; i++)
2912 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
2913 			bitmap_set(svms->bitmap_supported, i, 1);
2914 
2915 	return 0;
2916 }
2917 
2918 /**
2919  * svm_range_check_vm - check if virtual address range mapped already
2920  * @p: current kfd_process
2921  * @start: range start address, in pages
2922  * @last: range last address, in pages
2923  * @bo_s: mapping start address in pages if address range already mapped
2924  * @bo_l: mapping last address in pages if address range already mapped
2925  *
2926  * The purpose is to avoid virtual address ranges already allocated by
2927  * kfd_ioctl_alloc_memory_of_gpu ioctl.
2928  * It looks for each pdd in the kfd_process.
2929  *
2930  * Context: Process context
2931  *
2932  * Return 0 - OK, if the range is not mapped.
2933  * Otherwise error code:
2934  * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
2935  * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
2936  * a signal. Release all buffer reservations and return to user-space.
2937  */
2938 static int
2939 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
2940 		   uint64_t *bo_s, uint64_t *bo_l)
2941 {
2942 	struct amdgpu_bo_va_mapping *mapping;
2943 	struct interval_tree_node *node;
2944 	uint32_t i;
2945 	int r;
2946 
2947 	for (i = 0; i < p->n_pdds; i++) {
2948 		struct amdgpu_vm *vm;
2949 
2950 		if (!p->pdds[i]->drm_priv)
2951 			continue;
2952 
2953 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2954 		r = amdgpu_bo_reserve(vm->root.bo, false);
2955 		if (r)
2956 			return r;
2957 
2958 		node = interval_tree_iter_first(&vm->va, start, last);
2959 		if (node) {
2960 			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
2961 				 start, last);
2962 			mapping = container_of((struct rb_node *)node,
2963 					       struct amdgpu_bo_va_mapping, rb);
2964 			if (bo_s && bo_l) {
2965 				*bo_s = mapping->start;
2966 				*bo_l = mapping->last;
2967 			}
2968 			amdgpu_bo_unreserve(vm->root.bo);
2969 			return -EADDRINUSE;
2970 		}
2971 		amdgpu_bo_unreserve(vm->root.bo);
2972 	}
2973 
2974 	return 0;
2975 }
2976 
2977 /**
2978  * svm_range_is_valid - check if virtual address range is valid
2979  * @p: current kfd_process
2980  * @start: range start address, in pages
2981  * @size: range size, in pages
2982  *
2983  * Valid virtual address range means it belongs to one or more VMAs
2984  *
2985  * Context: Process context
2986  *
2987  * Return:
2988  *  0 - OK, otherwise error code
2989  */
2990 static int
2991 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
2992 {
2993 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
2994 	struct vm_area_struct *vma;
2995 	unsigned long end;
2996 	unsigned long start_unchg = start;
2997 
2998 	start <<= PAGE_SHIFT;
2999 	end = start + (size << PAGE_SHIFT);
3000 	do {
3001 		vma = find_vma(p->mm, start);
3002 		if (!vma || start < vma->vm_start ||
3003 		    (vma->vm_flags & device_vma))
3004 			return -EFAULT;
3005 		start = min(end, vma->vm_end);
3006 	} while (start < end);
3007 
3008 	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3009 				  NULL);
3010 }
3011 
3012 /**
3013  * svm_range_best_prefetch_location - decide the best prefetch location
3014  * @prange: svm range structure
3015  *
3016  * For xnack off:
3017  * If range map to single GPU, the best prefetch location is prefetch_loc, which
3018  * can be CPU or GPU.
3019  *
3020  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3021  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3022  * the best prefetch location is always CPU, because GPU can not have coherent
3023  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3024  *
3025  * For xnack on:
3026  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3027  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3028  *
3029  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3030  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3031  * prefetch location is always CPU.
3032  *
3033  * Context: Process context
3034  *
3035  * Return:
3036  * 0 for CPU or GPU id
3037  */
3038 static uint32_t
3039 svm_range_best_prefetch_location(struct svm_range *prange)
3040 {
3041 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3042 	uint32_t best_loc = prange->prefetch_loc;
3043 	struct kfd_process_device *pdd;
3044 	struct amdgpu_device *bo_adev;
3045 	struct kfd_process *p;
3046 	uint32_t gpuidx;
3047 
3048 	p = container_of(prange->svms, struct kfd_process, svms);
3049 
3050 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3051 		goto out;
3052 
3053 	bo_adev = svm_range_get_adev_by_id(prange, best_loc);
3054 	if (!bo_adev) {
3055 		WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
3056 		best_loc = 0;
3057 		goto out;
3058 	}
3059 
3060 	if (p->xnack_enabled)
3061 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3062 	else
3063 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3064 			  MAX_GPU_INSTANCE);
3065 
3066 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3067 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3068 		if (!pdd) {
3069 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3070 			continue;
3071 		}
3072 
3073 		if (pdd->dev->adev == bo_adev)
3074 			continue;
3075 
3076 		if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
3077 			best_loc = 0;
3078 			break;
3079 		}
3080 	}
3081 
3082 out:
3083 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3084 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3085 		 best_loc);
3086 
3087 	return best_loc;
3088 }
3089 
3090 /* FIXME: This is a workaround for page locking bug when some pages are
3091  * invalid during migration to VRAM
3092  */
3093 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
3094 			void *owner)
3095 {
3096 	struct hmm_range *hmm_range;
3097 	int r;
3098 
3099 	if (prange->validated_once)
3100 		return;
3101 
3102 	r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
3103 				       prange->start << PAGE_SHIFT,
3104 				       prange->npages, &hmm_range,
3105 				       false, true, owner);
3106 	if (!r) {
3107 		amdgpu_hmm_range_get_pages_done(hmm_range);
3108 		prange->validated_once = true;
3109 	}
3110 }
3111 
3112 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3113  * @mm: current process mm_struct
3114  * @prange: svm range structure
3115  * @migrated: output, true if migration is triggered
3116  *
3117  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3118  * from ram to vram.
3119  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3120  * from vram to ram.
3121  *
3122  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3123  * and restore work:
3124  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3125  *    stops all queues, schedule restore work
3126  * 2. svm_range_restore_work wait for migration is done by
3127  *    a. svm_range_validate_vram takes prange->migrate_mutex
3128  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3129  * 3. restore work update mappings of GPU, resume all queues.
3130  *
3131  * Context: Process context
3132  *
3133  * Return:
3134  * 0 - OK, otherwise - error code of migration
3135  */
3136 static int
3137 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3138 			    bool *migrated)
3139 {
3140 	uint32_t best_loc;
3141 	int r = 0;
3142 
3143 	*migrated = false;
3144 	best_loc = svm_range_best_prefetch_location(prange);
3145 
3146 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3147 	    best_loc == prange->actual_loc)
3148 		return 0;
3149 
3150 	if (!best_loc) {
3151 		r = svm_migrate_vram_to_ram(prange, mm);
3152 		*migrated = !r;
3153 		return r;
3154 	}
3155 
3156 	r = svm_migrate_to_vram(prange, best_loc, mm);
3157 	*migrated = !r;
3158 
3159 	return r;
3160 }
3161 
3162 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3163 {
3164 	if (!fence)
3165 		return -EINVAL;
3166 
3167 	if (dma_fence_is_signaled(&fence->base))
3168 		return 0;
3169 
3170 	if (fence->svm_bo) {
3171 		WRITE_ONCE(fence->svm_bo->evicting, 1);
3172 		schedule_work(&fence->svm_bo->eviction_work);
3173 	}
3174 
3175 	return 0;
3176 }
3177 
3178 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3179 {
3180 	struct svm_range_bo *svm_bo;
3181 	struct kfd_process *p;
3182 	struct mm_struct *mm;
3183 	int r = 0;
3184 
3185 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3186 	if (!svm_bo_ref_unless_zero(svm_bo))
3187 		return; /* svm_bo was freed while eviction was pending */
3188 
3189 	/* svm_range_bo_release destroys this worker thread. So during
3190 	 * the lifetime of this thread, kfd_process and mm will be valid.
3191 	 */
3192 	p = container_of(svm_bo->svms, struct kfd_process, svms);
3193 	mm = p->mm;
3194 	if (!mm)
3195 		return;
3196 
3197 	mmap_read_lock(mm);
3198 	spin_lock(&svm_bo->list_lock);
3199 	while (!list_empty(&svm_bo->range_list) && !r) {
3200 		struct svm_range *prange =
3201 				list_first_entry(&svm_bo->range_list,
3202 						struct svm_range, svm_bo_list);
3203 		int retries = 3;
3204 
3205 		list_del_init(&prange->svm_bo_list);
3206 		spin_unlock(&svm_bo->list_lock);
3207 
3208 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3209 			 prange->start, prange->last);
3210 
3211 		mutex_lock(&prange->migrate_mutex);
3212 		do {
3213 			r = svm_migrate_vram_to_ram(prange,
3214 						svm_bo->eviction_fence->mm);
3215 		} while (!r && prange->actual_loc && --retries);
3216 
3217 		if (!r && prange->actual_loc)
3218 			pr_info_once("Migration failed during eviction");
3219 
3220 		if (!prange->actual_loc) {
3221 			mutex_lock(&prange->lock);
3222 			prange->svm_bo = NULL;
3223 			mutex_unlock(&prange->lock);
3224 		}
3225 		mutex_unlock(&prange->migrate_mutex);
3226 
3227 		spin_lock(&svm_bo->list_lock);
3228 	}
3229 	spin_unlock(&svm_bo->list_lock);
3230 	mmap_read_unlock(mm);
3231 
3232 	dma_fence_signal(&svm_bo->eviction_fence->base);
3233 
3234 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3235 	 * has been called in svm_migrate_vram_to_ram
3236 	 */
3237 	WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3238 	svm_range_bo_unref(svm_bo);
3239 }
3240 
3241 static int
3242 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3243 		   uint64_t start, uint64_t size, uint32_t nattr,
3244 		   struct kfd_ioctl_svm_attribute *attrs)
3245 {
3246 	struct amdkfd_process_info *process_info = p->kgd_process_info;
3247 	struct list_head update_list;
3248 	struct list_head insert_list;
3249 	struct list_head remove_list;
3250 	struct svm_range_list *svms;
3251 	struct svm_range *prange;
3252 	struct svm_range *next;
3253 	bool update_mapping = false;
3254 	bool flush_tlb;
3255 	int r = 0;
3256 
3257 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3258 		 p->pasid, &p->svms, start, start + size - 1, size);
3259 
3260 	r = svm_range_check_attr(p, nattr, attrs);
3261 	if (r)
3262 		return r;
3263 
3264 	svms = &p->svms;
3265 
3266 	mutex_lock(&process_info->lock);
3267 
3268 	svm_range_list_lock_and_flush_work(svms, mm);
3269 
3270 	r = svm_range_is_valid(p, start, size);
3271 	if (r) {
3272 		pr_debug("invalid range r=%d\n", r);
3273 		mmap_write_unlock(mm);
3274 		goto out;
3275 	}
3276 
3277 	mutex_lock(&svms->lock);
3278 
3279 	/* Add new range and split existing ranges as needed */
3280 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3281 			  &insert_list, &remove_list);
3282 	if (r) {
3283 		mutex_unlock(&svms->lock);
3284 		mmap_write_unlock(mm);
3285 		goto out;
3286 	}
3287 	/* Apply changes as a transaction */
3288 	list_for_each_entry_safe(prange, next, &insert_list, list) {
3289 		svm_range_add_to_svms(prange);
3290 		svm_range_add_notifier_locked(mm, prange);
3291 	}
3292 	list_for_each_entry(prange, &update_list, update_list) {
3293 		svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3294 		/* TODO: unmap ranges from GPU that lost access */
3295 	}
3296 	list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3297 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3298 			 prange->svms, prange, prange->start,
3299 			 prange->last);
3300 		svm_range_unlink(prange);
3301 		svm_range_remove_notifier(prange);
3302 		svm_range_free(prange);
3303 	}
3304 
3305 	mmap_write_downgrade(mm);
3306 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3307 	 * this fails we may be left with partially completed actions. There
3308 	 * is no clean way of rolling back to the previous state in such a
3309 	 * case because the rollback wouldn't be guaranteed to work either.
3310 	 */
3311 	list_for_each_entry(prange, &update_list, update_list) {
3312 		bool migrated;
3313 
3314 		mutex_lock(&prange->migrate_mutex);
3315 
3316 		r = svm_range_trigger_migration(mm, prange, &migrated);
3317 		if (r)
3318 			goto out_unlock_range;
3319 
3320 		if (migrated && !p->xnack_enabled) {
3321 			pr_debug("restore_work will update mappings of GPUs\n");
3322 			mutex_unlock(&prange->migrate_mutex);
3323 			continue;
3324 		}
3325 
3326 		if (!migrated && !update_mapping) {
3327 			mutex_unlock(&prange->migrate_mutex);
3328 			continue;
3329 		}
3330 
3331 		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3332 
3333 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3334 					       true, true, flush_tlb);
3335 		if (r)
3336 			pr_debug("failed %d to map svm range\n", r);
3337 
3338 out_unlock_range:
3339 		mutex_unlock(&prange->migrate_mutex);
3340 		if (r)
3341 			break;
3342 	}
3343 
3344 	svm_range_debug_dump(svms);
3345 
3346 	mutex_unlock(&svms->lock);
3347 	mmap_read_unlock(mm);
3348 out:
3349 	mutex_unlock(&process_info->lock);
3350 
3351 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3352 		 &p->svms, start, start + size - 1, r);
3353 
3354 	return r;
3355 }
3356 
3357 static int
3358 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3359 		   uint64_t start, uint64_t size, uint32_t nattr,
3360 		   struct kfd_ioctl_svm_attribute *attrs)
3361 {
3362 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3363 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3364 	bool get_preferred_loc = false;
3365 	bool get_prefetch_loc = false;
3366 	bool get_granularity = false;
3367 	bool get_accessible = false;
3368 	bool get_flags = false;
3369 	uint64_t last = start + size - 1UL;
3370 	uint8_t granularity = 0xff;
3371 	struct interval_tree_node *node;
3372 	struct svm_range_list *svms;
3373 	struct svm_range *prange;
3374 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3375 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3376 	uint32_t flags_and = 0xffffffff;
3377 	uint32_t flags_or = 0;
3378 	int gpuidx;
3379 	uint32_t i;
3380 	int r = 0;
3381 
3382 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3383 		 start + size - 1, nattr);
3384 
3385 	/* Flush pending deferred work to avoid racing with deferred actions from
3386 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3387 	 * can still race with get_attr because we don't hold the mmap lock. But that
3388 	 * would be a race condition in the application anyway, and undefined
3389 	 * behaviour is acceptable in that case.
3390 	 */
3391 	flush_work(&p->svms.deferred_list_work);
3392 
3393 	mmap_read_lock(mm);
3394 	r = svm_range_is_valid(p, start, size);
3395 	mmap_read_unlock(mm);
3396 	if (r) {
3397 		pr_debug("invalid range r=%d\n", r);
3398 		return r;
3399 	}
3400 
3401 	for (i = 0; i < nattr; i++) {
3402 		switch (attrs[i].type) {
3403 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3404 			get_preferred_loc = true;
3405 			break;
3406 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3407 			get_prefetch_loc = true;
3408 			break;
3409 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3410 			get_accessible = true;
3411 			break;
3412 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3413 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3414 			get_flags = true;
3415 			break;
3416 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3417 			get_granularity = true;
3418 			break;
3419 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3420 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3421 			fallthrough;
3422 		default:
3423 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3424 			return -EINVAL;
3425 		}
3426 	}
3427 
3428 	svms = &p->svms;
3429 
3430 	mutex_lock(&svms->lock);
3431 
3432 	node = interval_tree_iter_first(&svms->objects, start, last);
3433 	if (!node) {
3434 		pr_debug("range attrs not found return default values\n");
3435 		svm_range_set_default_attributes(&location, &prefetch_loc,
3436 						 &granularity, &flags_and);
3437 		flags_or = flags_and;
3438 		if (p->xnack_enabled)
3439 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3440 				    MAX_GPU_INSTANCE);
3441 		else
3442 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3443 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3444 		goto fill_values;
3445 	}
3446 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3447 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3448 
3449 	while (node) {
3450 		struct interval_tree_node *next;
3451 
3452 		prange = container_of(node, struct svm_range, it_node);
3453 		next = interval_tree_iter_next(node, start, last);
3454 
3455 		if (get_preferred_loc) {
3456 			if (prange->preferred_loc ==
3457 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3458 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3459 			     location != prange->preferred_loc)) {
3460 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3461 				get_preferred_loc = false;
3462 			} else {
3463 				location = prange->preferred_loc;
3464 			}
3465 		}
3466 		if (get_prefetch_loc) {
3467 			if (prange->prefetch_loc ==
3468 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3469 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3470 			     prefetch_loc != prange->prefetch_loc)) {
3471 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3472 				get_prefetch_loc = false;
3473 			} else {
3474 				prefetch_loc = prange->prefetch_loc;
3475 			}
3476 		}
3477 		if (get_accessible) {
3478 			bitmap_and(bitmap_access, bitmap_access,
3479 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3480 			bitmap_and(bitmap_aip, bitmap_aip,
3481 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3482 		}
3483 		if (get_flags) {
3484 			flags_and &= prange->flags;
3485 			flags_or |= prange->flags;
3486 		}
3487 
3488 		if (get_granularity && prange->granularity < granularity)
3489 			granularity = prange->granularity;
3490 
3491 		node = next;
3492 	}
3493 fill_values:
3494 	mutex_unlock(&svms->lock);
3495 
3496 	for (i = 0; i < nattr; i++) {
3497 		switch (attrs[i].type) {
3498 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3499 			attrs[i].value = location;
3500 			break;
3501 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3502 			attrs[i].value = prefetch_loc;
3503 			break;
3504 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3505 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3506 							       attrs[i].value);
3507 			if (gpuidx < 0) {
3508 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3509 				return -EINVAL;
3510 			}
3511 			if (test_bit(gpuidx, bitmap_access))
3512 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3513 			else if (test_bit(gpuidx, bitmap_aip))
3514 				attrs[i].type =
3515 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3516 			else
3517 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3518 			break;
3519 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3520 			attrs[i].value = flags_and;
3521 			break;
3522 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3523 			attrs[i].value = ~flags_or;
3524 			break;
3525 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3526 			attrs[i].value = (uint32_t)granularity;
3527 			break;
3528 		}
3529 	}
3530 
3531 	return 0;
3532 }
3533 
3534 int kfd_criu_resume_svm(struct kfd_process *p)
3535 {
3536 	struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3537 	int nattr_common = 4, nattr_accessibility = 1;
3538 	struct criu_svm_metadata *criu_svm_md = NULL;
3539 	struct svm_range_list *svms = &p->svms;
3540 	struct criu_svm_metadata *next = NULL;
3541 	uint32_t set_flags = 0xffffffff;
3542 	int i, j, num_attrs, ret = 0;
3543 	uint64_t set_attr_size;
3544 	struct mm_struct *mm;
3545 
3546 	if (list_empty(&svms->criu_svm_metadata_list)) {
3547 		pr_debug("No SVM data from CRIU restore stage 2\n");
3548 		return ret;
3549 	}
3550 
3551 	mm = get_task_mm(p->lead_thread);
3552 	if (!mm) {
3553 		pr_err("failed to get mm for the target process\n");
3554 		return -ESRCH;
3555 	}
3556 
3557 	num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3558 
3559 	i = j = 0;
3560 	list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3561 		pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3562 			 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3563 
3564 		for (j = 0; j < num_attrs; j++) {
3565 			pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3566 				 i, j, criu_svm_md->data.attrs[j].type,
3567 				 i, j, criu_svm_md->data.attrs[j].value);
3568 			switch (criu_svm_md->data.attrs[j].type) {
3569 			/* During Checkpoint operation, the query for
3570 			 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3571 			 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3572 			 * not used by the range which was checkpointed. Care
3573 			 * must be taken to not restore with an invalid value
3574 			 * otherwise the gpuidx value will be invalid and
3575 			 * set_attr would eventually fail so just replace those
3576 			 * with another dummy attribute such as
3577 			 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3578 			 */
3579 			case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3580 				if (criu_svm_md->data.attrs[j].value ==
3581 				    KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3582 					criu_svm_md->data.attrs[j].type =
3583 						KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3584 					criu_svm_md->data.attrs[j].value = 0;
3585 				}
3586 				break;
3587 			case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3588 				set_flags = criu_svm_md->data.attrs[j].value;
3589 				break;
3590 			default:
3591 				break;
3592 			}
3593 		}
3594 
3595 		/* CLR_FLAGS is not available via get_attr during checkpoint but
3596 		 * it needs to be inserted before restoring the ranges so
3597 		 * allocate extra space for it before calling set_attr
3598 		 */
3599 		set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3600 						(num_attrs + 1);
3601 		set_attr_new = krealloc(set_attr, set_attr_size,
3602 					    GFP_KERNEL);
3603 		if (!set_attr_new) {
3604 			ret = -ENOMEM;
3605 			goto exit;
3606 		}
3607 		set_attr = set_attr_new;
3608 
3609 		memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3610 					sizeof(struct kfd_ioctl_svm_attribute));
3611 		set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3612 		set_attr[num_attrs].value = ~set_flags;
3613 
3614 		ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3615 					 criu_svm_md->data.size, num_attrs + 1,
3616 					 set_attr);
3617 		if (ret) {
3618 			pr_err("CRIU: failed to set range attributes\n");
3619 			goto exit;
3620 		}
3621 
3622 		i++;
3623 	}
3624 exit:
3625 	kfree(set_attr);
3626 	list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3627 		pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3628 						criu_svm_md->data.start_addr);
3629 		kfree(criu_svm_md);
3630 	}
3631 
3632 	mmput(mm);
3633 	return ret;
3634 
3635 }
3636 
3637 int kfd_criu_restore_svm(struct kfd_process *p,
3638 			 uint8_t __user *user_priv_ptr,
3639 			 uint64_t *priv_data_offset,
3640 			 uint64_t max_priv_data_size)
3641 {
3642 	uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3643 	int nattr_common = 4, nattr_accessibility = 1;
3644 	struct criu_svm_metadata *criu_svm_md = NULL;
3645 	struct svm_range_list *svms = &p->svms;
3646 	uint32_t num_devices;
3647 	int ret = 0;
3648 
3649 	num_devices = p->n_pdds;
3650 	/* Handle one SVM range object at a time, also the number of gpus are
3651 	 * assumed to be same on the restore node, checking must be done while
3652 	 * evaluating the topology earlier
3653 	 */
3654 
3655 	svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3656 		(nattr_common + nattr_accessibility * num_devices);
3657 	svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3658 
3659 	svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3660 								svm_attrs_size;
3661 
3662 	criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3663 	if (!criu_svm_md) {
3664 		pr_err("failed to allocate memory to store svm metadata\n");
3665 		return -ENOMEM;
3666 	}
3667 	if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3668 		ret = -EINVAL;
3669 		goto exit;
3670 	}
3671 
3672 	ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3673 			     svm_priv_data_size);
3674 	if (ret) {
3675 		ret = -EFAULT;
3676 		goto exit;
3677 	}
3678 	*priv_data_offset += svm_priv_data_size;
3679 
3680 	list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3681 
3682 	return 0;
3683 
3684 
3685 exit:
3686 	kfree(criu_svm_md);
3687 	return ret;
3688 }
3689 
3690 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3691 		       uint64_t *svm_priv_data_size)
3692 {
3693 	uint64_t total_size, accessibility_size, common_attr_size;
3694 	int nattr_common = 4, nattr_accessibility = 1;
3695 	int num_devices = p->n_pdds;
3696 	struct svm_range_list *svms;
3697 	struct svm_range *prange;
3698 	uint32_t count = 0;
3699 
3700 	*svm_priv_data_size = 0;
3701 
3702 	svms = &p->svms;
3703 	if (!svms)
3704 		return -EINVAL;
3705 
3706 	mutex_lock(&svms->lock);
3707 	list_for_each_entry(prange, &svms->list, list) {
3708 		pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3709 			 prange, prange->start, prange->npages,
3710 			 prange->start + prange->npages - 1);
3711 		count++;
3712 	}
3713 	mutex_unlock(&svms->lock);
3714 
3715 	*num_svm_ranges = count;
3716 	/* Only the accessbility attributes need to be queried for all the gpus
3717 	 * individually, remaining ones are spanned across the entire process
3718 	 * regardless of the various gpu nodes. Of the remaining attributes,
3719 	 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3720 	 *
3721 	 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3722 	 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3723 	 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3724 	 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3725 	 *
3726 	 * ** ACCESSBILITY ATTRIBUTES **
3727 	 * (Considered as one, type is altered during query, value is gpuid)
3728 	 * KFD_IOCTL_SVM_ATTR_ACCESS
3729 	 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3730 	 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3731 	 */
3732 	if (*num_svm_ranges > 0) {
3733 		common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3734 			nattr_common;
3735 		accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3736 			nattr_accessibility * num_devices;
3737 
3738 		total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3739 			common_attr_size + accessibility_size;
3740 
3741 		*svm_priv_data_size = *num_svm_ranges * total_size;
3742 	}
3743 
3744 	pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
3745 		 *svm_priv_data_size);
3746 	return 0;
3747 }
3748 
3749 int kfd_criu_checkpoint_svm(struct kfd_process *p,
3750 			    uint8_t __user *user_priv_data,
3751 			    uint64_t *priv_data_offset)
3752 {
3753 	struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
3754 	struct kfd_ioctl_svm_attribute *query_attr = NULL;
3755 	uint64_t svm_priv_data_size, query_attr_size = 0;
3756 	int index, nattr_common = 4, ret = 0;
3757 	struct svm_range_list *svms;
3758 	int num_devices = p->n_pdds;
3759 	struct svm_range *prange;
3760 	struct mm_struct *mm;
3761 
3762 	svms = &p->svms;
3763 	if (!svms)
3764 		return -EINVAL;
3765 
3766 	mm = get_task_mm(p->lead_thread);
3767 	if (!mm) {
3768 		pr_err("failed to get mm for the target process\n");
3769 		return -ESRCH;
3770 	}
3771 
3772 	query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3773 				(nattr_common + num_devices);
3774 
3775 	query_attr = kzalloc(query_attr_size, GFP_KERNEL);
3776 	if (!query_attr) {
3777 		ret = -ENOMEM;
3778 		goto exit;
3779 	}
3780 
3781 	query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
3782 	query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
3783 	query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3784 	query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
3785 
3786 	for (index = 0; index < num_devices; index++) {
3787 		struct kfd_process_device *pdd = p->pdds[index];
3788 
3789 		query_attr[index + nattr_common].type =
3790 			KFD_IOCTL_SVM_ATTR_ACCESS;
3791 		query_attr[index + nattr_common].value = pdd->user_gpu_id;
3792 	}
3793 
3794 	svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
3795 
3796 	svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
3797 	if (!svm_priv) {
3798 		ret = -ENOMEM;
3799 		goto exit_query;
3800 	}
3801 
3802 	index = 0;
3803 	list_for_each_entry(prange, &svms->list, list) {
3804 
3805 		svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
3806 		svm_priv->start_addr = prange->start;
3807 		svm_priv->size = prange->npages;
3808 		memcpy(&svm_priv->attrs, query_attr, query_attr_size);
3809 		pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
3810 			 prange, prange->start, prange->npages,
3811 			 prange->start + prange->npages - 1,
3812 			 prange->npages * PAGE_SIZE);
3813 
3814 		ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
3815 					 svm_priv->size,
3816 					 (nattr_common + num_devices),
3817 					 svm_priv->attrs);
3818 		if (ret) {
3819 			pr_err("CRIU: failed to obtain range attributes\n");
3820 			goto exit_priv;
3821 		}
3822 
3823 		if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
3824 				 svm_priv_data_size)) {
3825 			pr_err("Failed to copy svm priv to user\n");
3826 			ret = -EFAULT;
3827 			goto exit_priv;
3828 		}
3829 
3830 		*priv_data_offset += svm_priv_data_size;
3831 
3832 	}
3833 
3834 
3835 exit_priv:
3836 	kfree(svm_priv);
3837 exit_query:
3838 	kfree(query_attr);
3839 exit:
3840 	mmput(mm);
3841 	return ret;
3842 }
3843 
3844 int
3845 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3846 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3847 {
3848 	struct mm_struct *mm = current->mm;
3849 	int r;
3850 
3851 	start >>= PAGE_SHIFT;
3852 	size >>= PAGE_SHIFT;
3853 
3854 	switch (op) {
3855 	case KFD_IOCTL_SVM_OP_SET_ATTR:
3856 		r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
3857 		break;
3858 	case KFD_IOCTL_SVM_OP_GET_ATTR:
3859 		r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);
3860 		break;
3861 	default:
3862 		r = EINVAL;
3863 		break;
3864 	}
3865 
3866 	return r;
3867 }
3868