xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2020-2021 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/types.h>
25 #include <linux/sched/task.h>
26 #include <linux/dynamic_debug.h>
27 #include <drm/ttm/ttm_tt.h>
28 #include <drm/drm_exec.h>
29 
30 #include "amdgpu_sync.h"
31 #include "amdgpu_object.h"
32 #include "amdgpu_vm.h"
33 #include "amdgpu_hmm.h"
34 #include "amdgpu.h"
35 #include "amdgpu_xgmi.h"
36 #include "kfd_priv.h"
37 #include "kfd_svm.h"
38 #include "kfd_migrate.h"
39 #include "kfd_smi_events.h"
40 
41 #ifdef dev_fmt
42 #undef dev_fmt
43 #endif
44 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
45 
46 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
47 
48 /* Long enough to ensure no retry fault comes after svm range is restored and
49  * page table is updated.
50  */
51 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING	(2UL * NSEC_PER_MSEC)
52 #if IS_ENABLED(CONFIG_DYNAMIC_DEBUG)
53 #define dynamic_svm_range_dump(svms) \
54 	_dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms)
55 #else
56 #define dynamic_svm_range_dump(svms) \
57 	do { if (0) svm_range_debug_dump(svms); } while (0)
58 #endif
59 
60 /* Giant svm range split into smaller ranges based on this, it is decided using
61  * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to
62  * power of 2MB.
63  */
64 static uint64_t max_svm_range_pages;
65 
66 struct criu_svm_metadata {
67 	struct list_head list;
68 	struct kfd_criu_svm_range_priv_data data;
69 };
70 
71 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
72 static bool
73 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
74 				    const struct mmu_notifier_range *range,
75 				    unsigned long cur_seq);
76 static int
77 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
78 		   uint64_t *bo_s, uint64_t *bo_l);
79 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
80 	.invalidate = svm_range_cpu_invalidate_pagetables,
81 };
82 
83 /**
84  * svm_range_unlink - unlink svm_range from lists and interval tree
85  * @prange: svm range structure to be removed
86  *
87  * Remove the svm_range from the svms and svm_bo lists and the svms
88  * interval tree.
89  *
90  * Context: The caller must hold svms->lock
91  */
92 static void svm_range_unlink(struct svm_range *prange)
93 {
94 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
95 		 prange, prange->start, prange->last);
96 
97 	if (prange->svm_bo) {
98 		spin_lock(&prange->svm_bo->list_lock);
99 		list_del(&prange->svm_bo_list);
100 		spin_unlock(&prange->svm_bo->list_lock);
101 	}
102 
103 	list_del(&prange->list);
104 	if (prange->it_node.start != 0 && prange->it_node.last != 0)
105 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
106 }
107 
108 static void
109 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
110 {
111 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
112 		 prange, prange->start, prange->last);
113 
114 	mmu_interval_notifier_insert_locked(&prange->notifier, mm,
115 				     prange->start << PAGE_SHIFT,
116 				     prange->npages << PAGE_SHIFT,
117 				     &svm_range_mn_ops);
118 }
119 
120 /**
121  * svm_range_add_to_svms - add svm range to svms
122  * @prange: svm range structure to be added
123  *
124  * Add the svm range to svms interval tree and link list
125  *
126  * Context: The caller must hold svms->lock
127  */
128 static void svm_range_add_to_svms(struct svm_range *prange)
129 {
130 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
131 		 prange, prange->start, prange->last);
132 
133 	list_move_tail(&prange->list, &prange->svms->list);
134 	prange->it_node.start = prange->start;
135 	prange->it_node.last = prange->last;
136 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
137 }
138 
139 static void svm_range_remove_notifier(struct svm_range *prange)
140 {
141 	pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
142 		 prange->svms, prange,
143 		 prange->notifier.interval_tree.start >> PAGE_SHIFT,
144 		 prange->notifier.interval_tree.last >> PAGE_SHIFT);
145 
146 	if (prange->notifier.interval_tree.start != 0 &&
147 	    prange->notifier.interval_tree.last != 0)
148 		mmu_interval_notifier_remove(&prange->notifier);
149 }
150 
151 static bool
152 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
153 {
154 	return dma_addr && !dma_mapping_error(dev, dma_addr) &&
155 	       !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
156 }
157 
158 static int
159 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
160 		      unsigned long offset, unsigned long npages,
161 		      unsigned long *hmm_pfns, uint32_t gpuidx)
162 {
163 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
164 	dma_addr_t *addr = prange->dma_addr[gpuidx];
165 	struct device *dev = adev->dev;
166 	struct page *page;
167 	int i, r;
168 
169 	if (!addr) {
170 		addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL);
171 		if (!addr)
172 			return -ENOMEM;
173 		prange->dma_addr[gpuidx] = addr;
174 	}
175 
176 	addr += offset;
177 	for (i = 0; i < npages; i++) {
178 		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
179 			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
180 
181 		page = hmm_pfn_to_page(hmm_pfns[i]);
182 		if (is_zone_device_page(page)) {
183 			struct amdgpu_device *bo_adev = prange->svm_bo->node->adev;
184 
185 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
186 				   bo_adev->vm_manager.vram_base_offset -
187 				   bo_adev->kfd.pgmap.range.start;
188 			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
189 			pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
190 			continue;
191 		}
192 		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
193 		r = dma_mapping_error(dev, addr[i]);
194 		if (r) {
195 			dev_err(dev, "failed %d dma_map_page\n", r);
196 			return r;
197 		}
198 		pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
199 				     addr[i] >> PAGE_SHIFT, page_to_pfn(page));
200 	}
201 
202 	return 0;
203 }
204 
205 static int
206 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
207 		  unsigned long offset, unsigned long npages,
208 		  unsigned long *hmm_pfns)
209 {
210 	struct kfd_process *p;
211 	uint32_t gpuidx;
212 	int r;
213 
214 	p = container_of(prange->svms, struct kfd_process, svms);
215 
216 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
217 		struct kfd_process_device *pdd;
218 
219 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
220 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
221 		if (!pdd) {
222 			pr_debug("failed to find device idx %d\n", gpuidx);
223 			return -EINVAL;
224 		}
225 
226 		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
227 					  hmm_pfns, gpuidx);
228 		if (r)
229 			break;
230 	}
231 
232 	return r;
233 }
234 
235 void svm_range_dma_unmap_dev(struct device *dev, dma_addr_t *dma_addr,
236 			 unsigned long offset, unsigned long npages)
237 {
238 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
239 	int i;
240 
241 	if (!dma_addr)
242 		return;
243 
244 	for (i = offset; i < offset + npages; i++) {
245 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
246 			continue;
247 		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
248 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
249 		dma_addr[i] = 0;
250 	}
251 }
252 
253 void svm_range_dma_unmap(struct svm_range *prange)
254 {
255 	struct kfd_process_device *pdd;
256 	dma_addr_t *dma_addr;
257 	struct device *dev;
258 	struct kfd_process *p;
259 	uint32_t gpuidx;
260 
261 	p = container_of(prange->svms, struct kfd_process, svms);
262 
263 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
264 		dma_addr = prange->dma_addr[gpuidx];
265 		if (!dma_addr)
266 			continue;
267 
268 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
269 		if (!pdd) {
270 			pr_debug("failed to find device idx %d\n", gpuidx);
271 			continue;
272 		}
273 		dev = &pdd->dev->adev->pdev->dev;
274 
275 		svm_range_dma_unmap_dev(dev, dma_addr, 0, prange->npages);
276 	}
277 }
278 
279 static void svm_range_free(struct svm_range *prange, bool do_unmap)
280 {
281 	uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
282 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
283 	uint32_t gpuidx;
284 
285 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
286 		 prange->start, prange->last);
287 
288 	svm_range_vram_node_free(prange);
289 	if (do_unmap)
290 		svm_range_dma_unmap(prange);
291 
292 	if (do_unmap && !p->xnack_enabled) {
293 		pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size);
294 		amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
295 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
296 	}
297 
298 	/* free dma_addr array for each gpu */
299 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
300 		if (prange->dma_addr[gpuidx]) {
301 			kvfree(prange->dma_addr[gpuidx]);
302 			prange->dma_addr[gpuidx] = NULL;
303 		}
304 	}
305 
306 	mutex_destroy(&prange->lock);
307 	mutex_destroy(&prange->migrate_mutex);
308 	kfree(prange);
309 }
310 
311 static void
312 svm_range_set_default_attributes(struct svm_range_list *svms, int32_t *location,
313 				 int32_t *prefetch_loc, uint8_t *granularity,
314 				 uint32_t *flags)
315 {
316 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
317 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
318 	*granularity = svms->default_granularity;
319 	*flags =
320 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
321 }
322 
323 static struct
324 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
325 			 uint64_t last, bool update_mem_usage)
326 {
327 	uint64_t size = last - start + 1;
328 	struct svm_range *prange;
329 	struct kfd_process *p;
330 
331 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
332 	if (!prange)
333 		return NULL;
334 
335 	p = container_of(svms, struct kfd_process, svms);
336 	if (!p->xnack_enabled && update_mem_usage &&
337 	    amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
338 				    KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) {
339 		pr_info("SVM mapping failed, exceeds resident system memory limit\n");
340 		kfree(prange);
341 		return NULL;
342 	}
343 	prange->npages = size;
344 	prange->svms = svms;
345 	prange->start = start;
346 	prange->last = last;
347 	INIT_LIST_HEAD(&prange->list);
348 	INIT_LIST_HEAD(&prange->update_list);
349 	INIT_LIST_HEAD(&prange->svm_bo_list);
350 	INIT_LIST_HEAD(&prange->deferred_list);
351 	INIT_LIST_HEAD(&prange->child_list);
352 	atomic_set(&prange->invalid, 0);
353 	prange->validate_timestamp = 0;
354 	prange->vram_pages = 0;
355 	mutex_init(&prange->migrate_mutex);
356 	mutex_init(&prange->lock);
357 
358 	if (p->xnack_enabled)
359 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
360 			    MAX_GPU_INSTANCE);
361 
362 	svm_range_set_default_attributes(svms, &prange->preferred_loc,
363 					 &prange->prefetch_loc,
364 					 &prange->granularity, &prange->flags);
365 
366 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
367 
368 	return prange;
369 }
370 
371 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
372 {
373 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
374 		return false;
375 
376 	return true;
377 }
378 
379 static void svm_range_bo_release(struct kref *kref)
380 {
381 	struct svm_range_bo *svm_bo;
382 
383 	svm_bo = container_of(kref, struct svm_range_bo, kref);
384 	pr_debug("svm_bo 0x%p\n", svm_bo);
385 
386 	spin_lock(&svm_bo->list_lock);
387 	while (!list_empty(&svm_bo->range_list)) {
388 		struct svm_range *prange =
389 				list_first_entry(&svm_bo->range_list,
390 						struct svm_range, svm_bo_list);
391 		/* list_del_init tells a concurrent svm_range_vram_node_new when
392 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
393 		 */
394 		list_del_init(&prange->svm_bo_list);
395 		spin_unlock(&svm_bo->list_lock);
396 
397 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
398 			 prange->start, prange->last);
399 		mutex_lock(&prange->lock);
400 		prange->svm_bo = NULL;
401 		/* prange should not hold vram page now */
402 		WARN_ONCE(prange->actual_loc, "prange should not hold vram page");
403 		mutex_unlock(&prange->lock);
404 
405 		spin_lock(&svm_bo->list_lock);
406 	}
407 	spin_unlock(&svm_bo->list_lock);
408 
409 	if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
410 		struct kfd_process_device *pdd;
411 		struct kfd_process *p;
412 		struct mm_struct *mm;
413 
414 		mm = svm_bo->eviction_fence->mm;
415 		/*
416 		 * The forked child process takes svm_bo device pages ref, svm_bo could be
417 		 * released after parent process is gone.
418 		 */
419 		p = kfd_lookup_process_by_mm(mm);
420 		if (p) {
421 			pdd = kfd_get_process_device_data(svm_bo->node, p);
422 			if (pdd)
423 				atomic64_sub(amdgpu_bo_size(svm_bo->bo), &pdd->vram_usage);
424 			kfd_unref_process(p);
425 		}
426 		mmput(mm);
427 	}
428 
429 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base))
430 		/* We're not in the eviction worker. Signal the fence. */
431 		dma_fence_signal(&svm_bo->eviction_fence->base);
432 	dma_fence_put(&svm_bo->eviction_fence->base);
433 	amdgpu_bo_unref(&svm_bo->bo);
434 	kfree(svm_bo);
435 }
436 
437 static void svm_range_bo_wq_release(struct work_struct *work)
438 {
439 	struct svm_range_bo *svm_bo;
440 
441 	svm_bo = container_of(work, struct svm_range_bo, release_work);
442 	svm_range_bo_release(&svm_bo->kref);
443 }
444 
445 static void svm_range_bo_release_async(struct kref *kref)
446 {
447 	struct svm_range_bo *svm_bo;
448 
449 	svm_bo = container_of(kref, struct svm_range_bo, kref);
450 	pr_debug("svm_bo 0x%p\n", svm_bo);
451 	INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
452 	schedule_work(&svm_bo->release_work);
453 }
454 
455 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
456 {
457 	kref_put(&svm_bo->kref, svm_range_bo_release_async);
458 }
459 
460 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
461 {
462 	if (svm_bo)
463 		kref_put(&svm_bo->kref, svm_range_bo_release);
464 }
465 
466 static bool
467 svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange)
468 {
469 	mutex_lock(&prange->lock);
470 	if (!prange->svm_bo) {
471 		mutex_unlock(&prange->lock);
472 		return false;
473 	}
474 	if (prange->ttm_res) {
475 		/* We still have a reference, all is well */
476 		mutex_unlock(&prange->lock);
477 		return true;
478 	}
479 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
480 		/*
481 		 * Migrate from GPU to GPU, remove range from source svm_bo->node
482 		 * range list, and return false to allocate svm_bo from destination
483 		 * node.
484 		 */
485 		if (prange->svm_bo->node != node) {
486 			mutex_unlock(&prange->lock);
487 
488 			spin_lock(&prange->svm_bo->list_lock);
489 			list_del_init(&prange->svm_bo_list);
490 			spin_unlock(&prange->svm_bo->list_lock);
491 
492 			svm_range_bo_unref(prange->svm_bo);
493 			return false;
494 		}
495 		if (READ_ONCE(prange->svm_bo->evicting)) {
496 			struct dma_fence *f;
497 			struct svm_range_bo *svm_bo;
498 			/* The BO is getting evicted,
499 			 * we need to get a new one
500 			 */
501 			mutex_unlock(&prange->lock);
502 			svm_bo = prange->svm_bo;
503 			f = dma_fence_get(&svm_bo->eviction_fence->base);
504 			svm_range_bo_unref(prange->svm_bo);
505 			/* wait for the fence to avoid long spin-loop
506 			 * at list_empty_careful
507 			 */
508 			dma_fence_wait(f, false);
509 			dma_fence_put(f);
510 		} else {
511 			/* The BO was still around and we got
512 			 * a new reference to it
513 			 */
514 			mutex_unlock(&prange->lock);
515 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
516 				 prange->svms, prange->start, prange->last);
517 
518 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
519 			return true;
520 		}
521 
522 	} else {
523 		mutex_unlock(&prange->lock);
524 	}
525 
526 	/* We need a new svm_bo. Spin-loop to wait for concurrent
527 	 * svm_range_bo_release to finish removing this range from
528 	 * its range list and set prange->svm_bo to null. After this,
529 	 * it is safe to reuse the svm_bo pointer and svm_bo_list head.
530 	 */
531 	while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo)
532 		cond_resched();
533 
534 	return false;
535 }
536 
537 static struct svm_range_bo *svm_range_bo_new(void)
538 {
539 	struct svm_range_bo *svm_bo;
540 
541 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
542 	if (!svm_bo)
543 		return NULL;
544 
545 	kref_init(&svm_bo->kref);
546 	INIT_LIST_HEAD(&svm_bo->range_list);
547 	spin_lock_init(&svm_bo->list_lock);
548 
549 	return svm_bo;
550 }
551 
552 int
553 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
554 			bool clear)
555 {
556 	struct kfd_process_device *pdd;
557 	struct amdgpu_bo_param bp;
558 	struct svm_range_bo *svm_bo;
559 	struct amdgpu_bo_user *ubo;
560 	struct amdgpu_bo *bo;
561 	struct kfd_process *p;
562 	struct mm_struct *mm;
563 	int r;
564 
565 	p = container_of(prange->svms, struct kfd_process, svms);
566 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
567 		 prange->start, prange->last);
568 
569 	if (svm_range_validate_svm_bo(node, prange))
570 		return 0;
571 
572 	svm_bo = svm_range_bo_new();
573 	if (!svm_bo) {
574 		pr_debug("failed to alloc svm bo\n");
575 		return -ENOMEM;
576 	}
577 	mm = get_task_mm(p->lead_thread);
578 	if (!mm) {
579 		pr_debug("failed to get mm\n");
580 		kfree(svm_bo);
581 		return -ESRCH;
582 	}
583 	svm_bo->node = node;
584 	svm_bo->eviction_fence =
585 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
586 					   mm,
587 					   svm_bo);
588 	mmput(mm);
589 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
590 	svm_bo->evicting = 0;
591 	memset(&bp, 0, sizeof(bp));
592 	bp.size = prange->npages * PAGE_SIZE;
593 	bp.byte_align = PAGE_SIZE;
594 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
595 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
596 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
597 	bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
598 	bp.type = ttm_bo_type_device;
599 	bp.resv = NULL;
600 	if (node->xcp)
601 		bp.xcp_id_plus1 = node->xcp->id + 1;
602 
603 	r = amdgpu_bo_create_user(node->adev, &bp, &ubo);
604 	if (r) {
605 		pr_debug("failed %d to create bo\n", r);
606 		goto create_bo_failed;
607 	}
608 	bo = &ubo->bo;
609 
610 	pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n",
611 		 bo->tbo.resource->start << PAGE_SHIFT, bp.size,
612 		 bp.xcp_id_plus1 - 1);
613 
614 	r = amdgpu_bo_reserve(bo, true);
615 	if (r) {
616 		pr_debug("failed %d to reserve bo\n", r);
617 		goto reserve_bo_failed;
618 	}
619 
620 	if (clear) {
621 		r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
622 		if (r) {
623 			pr_debug("failed %d to sync bo\n", r);
624 			amdgpu_bo_unreserve(bo);
625 			goto reserve_bo_failed;
626 		}
627 	}
628 
629 	r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
630 	if (r) {
631 		pr_debug("failed %d to reserve bo\n", r);
632 		amdgpu_bo_unreserve(bo);
633 		goto reserve_bo_failed;
634 	}
635 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
636 
637 	amdgpu_bo_unreserve(bo);
638 
639 	svm_bo->bo = bo;
640 	prange->svm_bo = svm_bo;
641 	prange->ttm_res = bo->tbo.resource;
642 	prange->offset = 0;
643 
644 	spin_lock(&svm_bo->list_lock);
645 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
646 	spin_unlock(&svm_bo->list_lock);
647 
648 	pdd = svm_range_get_pdd_by_node(prange, node);
649 	if (pdd)
650 		atomic64_add(amdgpu_bo_size(bo), &pdd->vram_usage);
651 
652 	return 0;
653 
654 reserve_bo_failed:
655 	amdgpu_bo_unref(&bo);
656 create_bo_failed:
657 	dma_fence_put(&svm_bo->eviction_fence->base);
658 	kfree(svm_bo);
659 	prange->ttm_res = NULL;
660 
661 	return r;
662 }
663 
664 void svm_range_vram_node_free(struct svm_range *prange)
665 {
666 	/* serialize prange->svm_bo unref */
667 	mutex_lock(&prange->lock);
668 	/* prange->svm_bo has not been unref */
669 	if (prange->ttm_res) {
670 		prange->ttm_res = NULL;
671 		mutex_unlock(&prange->lock);
672 		svm_range_bo_unref(prange->svm_bo);
673 	} else
674 		mutex_unlock(&prange->lock);
675 }
676 
677 struct kfd_node *
678 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id)
679 {
680 	struct kfd_process *p;
681 	struct kfd_process_device *pdd;
682 
683 	p = container_of(prange->svms, struct kfd_process, svms);
684 	pdd = kfd_process_device_data_by_id(p, gpu_id);
685 	if (!pdd) {
686 		pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id);
687 		return NULL;
688 	}
689 
690 	return pdd->dev;
691 }
692 
693 struct kfd_process_device *
694 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node)
695 {
696 	struct kfd_process *p;
697 
698 	p = container_of(prange->svms, struct kfd_process, svms);
699 
700 	return kfd_get_process_device_data(node, p);
701 }
702 
703 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
704 {
705 	struct ttm_operation_ctx ctx = { false, false };
706 
707 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
708 
709 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
710 }
711 
712 static int
713 svm_range_check_attr(struct kfd_process *p,
714 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
715 {
716 	uint32_t i;
717 
718 	for (i = 0; i < nattr; i++) {
719 		uint32_t val = attrs[i].value;
720 		int gpuidx = MAX_GPU_INSTANCE;
721 
722 		switch (attrs[i].type) {
723 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
724 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
725 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
726 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
727 			break;
728 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
729 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
730 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
731 			break;
732 		case KFD_IOCTL_SVM_ATTR_ACCESS:
733 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
734 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
735 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
736 			break;
737 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
738 			break;
739 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
740 			break;
741 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
742 			break;
743 		default:
744 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
745 			return -EINVAL;
746 		}
747 
748 		if (gpuidx < 0) {
749 			pr_debug("no GPU 0x%x found\n", val);
750 			return -EINVAL;
751 		} else if (gpuidx < MAX_GPU_INSTANCE &&
752 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
753 			pr_debug("GPU 0x%x not supported\n", val);
754 			return -EINVAL;
755 		}
756 	}
757 
758 	return 0;
759 }
760 
761 static void
762 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
763 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
764 		      bool *update_mapping)
765 {
766 	uint32_t i;
767 	int gpuidx;
768 
769 	for (i = 0; i < nattr; i++) {
770 		switch (attrs[i].type) {
771 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
772 			prange->preferred_loc = attrs[i].value;
773 			break;
774 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
775 			prange->prefetch_loc = attrs[i].value;
776 			break;
777 		case KFD_IOCTL_SVM_ATTR_ACCESS:
778 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
779 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
780 			if (!p->xnack_enabled)
781 				*update_mapping = true;
782 
783 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
784 							       attrs[i].value);
785 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
786 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
787 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
788 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
789 				bitmap_set(prange->bitmap_access, gpuidx, 1);
790 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
791 			} else {
792 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
793 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
794 			}
795 			break;
796 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
797 			*update_mapping = true;
798 			prange->flags |= attrs[i].value;
799 			break;
800 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
801 			*update_mapping = true;
802 			prange->flags &= ~attrs[i].value;
803 			break;
804 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
805 			prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F);
806 			break;
807 		default:
808 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
809 		}
810 	}
811 }
812 
813 static bool
814 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
815 			uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
816 {
817 	uint32_t i;
818 	int gpuidx;
819 
820 	for (i = 0; i < nattr; i++) {
821 		switch (attrs[i].type) {
822 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
823 			if (prange->preferred_loc != attrs[i].value)
824 				return false;
825 			break;
826 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
827 			/* Prefetch should always trigger a migration even
828 			 * if the value of the attribute didn't change.
829 			 */
830 			return false;
831 		case KFD_IOCTL_SVM_ATTR_ACCESS:
832 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
833 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
834 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
835 							       attrs[i].value);
836 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
837 				if (test_bit(gpuidx, prange->bitmap_access) ||
838 				    test_bit(gpuidx, prange->bitmap_aip))
839 					return false;
840 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
841 				if (!test_bit(gpuidx, prange->bitmap_access))
842 					return false;
843 			} else {
844 				if (!test_bit(gpuidx, prange->bitmap_aip))
845 					return false;
846 			}
847 			break;
848 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
849 			if ((prange->flags & attrs[i].value) != attrs[i].value)
850 				return false;
851 			break;
852 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
853 			if ((prange->flags & attrs[i].value) != 0)
854 				return false;
855 			break;
856 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
857 			if (prange->granularity != attrs[i].value)
858 				return false;
859 			break;
860 		default:
861 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
862 		}
863 	}
864 
865 	return true;
866 }
867 
868 /**
869  * svm_range_debug_dump - print all range information from svms
870  * @svms: svm range list header
871  *
872  * debug output svm range start, end, prefetch location from svms
873  * interval tree and link list
874  *
875  * Context: The caller must hold svms->lock
876  */
877 static void svm_range_debug_dump(struct svm_range_list *svms)
878 {
879 	struct interval_tree_node *node;
880 	struct svm_range *prange;
881 
882 	pr_debug("dump svms 0x%p list\n", svms);
883 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
884 
885 	list_for_each_entry(prange, &svms->list, list) {
886 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
887 			 prange, prange->start, prange->npages,
888 			 prange->start + prange->npages - 1,
889 			 prange->actual_loc);
890 	}
891 
892 	pr_debug("dump svms 0x%p interval tree\n", svms);
893 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
894 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
895 	while (node) {
896 		prange = container_of(node, struct svm_range, it_node);
897 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
898 			 prange, prange->start, prange->npages,
899 			 prange->start + prange->npages - 1,
900 			 prange->actual_loc);
901 		node = interval_tree_iter_next(node, 0, ~0ULL);
902 	}
903 }
904 
905 static void *
906 svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements,
907 		     uint64_t offset, uint64_t *vram_pages)
908 {
909 	unsigned char *src = (unsigned char *)psrc + offset;
910 	unsigned char *dst;
911 	uint64_t i;
912 
913 	dst = kvmalloc_array(num_elements, size, GFP_KERNEL);
914 	if (!dst)
915 		return NULL;
916 
917 	if (!vram_pages) {
918 		memcpy(dst, src, num_elements * size);
919 		return (void *)dst;
920 	}
921 
922 	*vram_pages = 0;
923 	for (i = 0; i < num_elements; i++) {
924 		dma_addr_t *temp;
925 		temp = (dma_addr_t *)dst + i;
926 		*temp = *((dma_addr_t *)src + i);
927 		if (*temp&SVM_RANGE_VRAM_DOMAIN)
928 			(*vram_pages)++;
929 	}
930 
931 	return (void *)dst;
932 }
933 
934 static int
935 svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src)
936 {
937 	int i;
938 
939 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
940 		if (!src->dma_addr[i])
941 			continue;
942 		dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i],
943 					sizeof(*src->dma_addr[i]), src->npages, 0, NULL);
944 		if (!dst->dma_addr[i])
945 			return -ENOMEM;
946 	}
947 
948 	return 0;
949 }
950 
951 static int
952 svm_range_split_array(void *ppnew, void *ppold, size_t size,
953 		      uint64_t old_start, uint64_t old_n,
954 		      uint64_t new_start, uint64_t new_n, uint64_t *new_vram_pages)
955 {
956 	unsigned char *new, *old, *pold;
957 	uint64_t d;
958 
959 	if (!ppold)
960 		return 0;
961 	pold = *(unsigned char **)ppold;
962 	if (!pold)
963 		return 0;
964 
965 	d = (new_start - old_start) * size;
966 	/* get dma addr array for new range and calculte its vram page number */
967 	new = svm_range_copy_array(pold, size, new_n, d, new_vram_pages);
968 	if (!new)
969 		return -ENOMEM;
970 	d = (new_start == old_start) ? new_n * size : 0;
971 	old = svm_range_copy_array(pold, size, old_n, d, NULL);
972 	if (!old) {
973 		kvfree(new);
974 		return -ENOMEM;
975 	}
976 	kvfree(pold);
977 	*(void **)ppold = old;
978 	*(void **)ppnew = new;
979 
980 	return 0;
981 }
982 
983 static int
984 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
985 		      uint64_t start, uint64_t last)
986 {
987 	uint64_t npages = last - start + 1;
988 	int i, r;
989 
990 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
991 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
992 					  sizeof(*old->dma_addr[i]), old->start,
993 					  npages, new->start, new->npages,
994 					  old->actual_loc ? &new->vram_pages : NULL);
995 		if (r)
996 			return r;
997 	}
998 	if (old->actual_loc)
999 		old->vram_pages -= new->vram_pages;
1000 
1001 	return 0;
1002 }
1003 
1004 static int
1005 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
1006 		      uint64_t start, uint64_t last)
1007 {
1008 	uint64_t npages = last - start + 1;
1009 
1010 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
1011 		 new->svms, new, new->start, start, last);
1012 
1013 	if (new->start == old->start) {
1014 		new->offset = old->offset;
1015 		old->offset += new->npages;
1016 	} else {
1017 		new->offset = old->offset + npages;
1018 	}
1019 
1020 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
1021 	new->ttm_res = old->ttm_res;
1022 
1023 	spin_lock(&new->svm_bo->list_lock);
1024 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1025 	spin_unlock(&new->svm_bo->list_lock);
1026 
1027 	return 0;
1028 }
1029 
1030 /**
1031  * svm_range_split_adjust - split range and adjust
1032  *
1033  * @new: new range
1034  * @old: the old range
1035  * @start: the old range adjust to start address in pages
1036  * @last: the old range adjust to last address in pages
1037  *
1038  * Copy system memory dma_addr or vram ttm_res in old range to new
1039  * range from new_start up to size new->npages, the remaining old range is from
1040  * start to last
1041  *
1042  * Return:
1043  * 0 - OK, -ENOMEM - out of memory
1044  */
1045 static int
1046 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
1047 		      uint64_t start, uint64_t last)
1048 {
1049 	int r;
1050 
1051 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
1052 		 new->svms, new->start, old->start, old->last, start, last);
1053 
1054 	if (new->start < old->start ||
1055 	    new->last > old->last) {
1056 		WARN_ONCE(1, "invalid new range start or last\n");
1057 		return -EINVAL;
1058 	}
1059 
1060 	r = svm_range_split_pages(new, old, start, last);
1061 	if (r)
1062 		return r;
1063 
1064 	if (old->actual_loc && old->ttm_res) {
1065 		r = svm_range_split_nodes(new, old, start, last);
1066 		if (r)
1067 			return r;
1068 	}
1069 
1070 	old->npages = last - start + 1;
1071 	old->start = start;
1072 	old->last = last;
1073 	new->flags = old->flags;
1074 	new->preferred_loc = old->preferred_loc;
1075 	new->prefetch_loc = old->prefetch_loc;
1076 	new->actual_loc = old->actual_loc;
1077 	new->granularity = old->granularity;
1078 	new->mapped_to_gpu = old->mapped_to_gpu;
1079 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1080 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1081 	atomic_set(&new->queue_refcount, atomic_read(&old->queue_refcount));
1082 
1083 	return 0;
1084 }
1085 
1086 /**
1087  * svm_range_split - split a range in 2 ranges
1088  *
1089  * @prange: the svm range to split
1090  * @start: the remaining range start address in pages
1091  * @last: the remaining range last address in pages
1092  * @new: the result new range generated
1093  *
1094  * Two cases only:
1095  * case 1: if start == prange->start
1096  *         prange ==> prange[start, last]
1097  *         new range [last + 1, prange->last]
1098  *
1099  * case 2: if last == prange->last
1100  *         prange ==> prange[start, last]
1101  *         new range [prange->start, start - 1]
1102  *
1103  * Return:
1104  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1105  */
1106 static int
1107 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1108 		struct svm_range **new)
1109 {
1110 	uint64_t old_start = prange->start;
1111 	uint64_t old_last = prange->last;
1112 	struct svm_range_list *svms;
1113 	int r = 0;
1114 
1115 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1116 		 old_start, old_last, start, last);
1117 
1118 	if (old_start != start && old_last != last)
1119 		return -EINVAL;
1120 	if (start < old_start || last > old_last)
1121 		return -EINVAL;
1122 
1123 	svms = prange->svms;
1124 	if (old_start == start)
1125 		*new = svm_range_new(svms, last + 1, old_last, false);
1126 	else
1127 		*new = svm_range_new(svms, old_start, start - 1, false);
1128 	if (!*new)
1129 		return -ENOMEM;
1130 
1131 	r = svm_range_split_adjust(*new, prange, start, last);
1132 	if (r) {
1133 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1134 			 r, old_start, old_last, start, last);
1135 		svm_range_free(*new, false);
1136 		*new = NULL;
1137 	}
1138 
1139 	return r;
1140 }
1141 
1142 static int
1143 svm_range_split_tail(struct svm_range *prange, uint64_t new_last,
1144 		     struct list_head *insert_list, struct list_head *remap_list)
1145 {
1146 	struct svm_range *tail = NULL;
1147 	int r = svm_range_split(prange, prange->start, new_last, &tail);
1148 
1149 	if (!r) {
1150 		list_add(&tail->list, insert_list);
1151 		if (!IS_ALIGNED(new_last + 1, 1UL << prange->granularity))
1152 			list_add(&tail->update_list, remap_list);
1153 	}
1154 	return r;
1155 }
1156 
1157 static int
1158 svm_range_split_head(struct svm_range *prange, uint64_t new_start,
1159 		     struct list_head *insert_list, struct list_head *remap_list)
1160 {
1161 	struct svm_range *head = NULL;
1162 	int r = svm_range_split(prange, new_start, prange->last, &head);
1163 
1164 	if (!r) {
1165 		list_add(&head->list, insert_list);
1166 		if (!IS_ALIGNED(new_start, 1UL << prange->granularity))
1167 			list_add(&head->update_list, remap_list);
1168 	}
1169 	return r;
1170 }
1171 
1172 static void
1173 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1174 		    struct svm_range *pchild, enum svm_work_list_ops op)
1175 {
1176 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1177 		 pchild, pchild->start, pchild->last, prange, op);
1178 
1179 	pchild->work_item.mm = mm;
1180 	pchild->work_item.op = op;
1181 	list_add_tail(&pchild->child_list, &prange->child_list);
1182 }
1183 
1184 static bool
1185 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
1186 {
1187 	return (node_a->adev == node_b->adev ||
1188 		amdgpu_xgmi_same_hive(node_a->adev, node_b->adev));
1189 }
1190 
1191 static uint64_t
1192 svm_range_get_pte_flags(struct kfd_node *node,
1193 			struct svm_range *prange, int domain)
1194 {
1195 	struct kfd_node *bo_node;
1196 	uint32_t flags = prange->flags;
1197 	uint32_t mapping_flags = 0;
1198 	uint64_t pte_flags;
1199 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1200 	bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENT);
1201 	bool ext_coherent = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENT;
1202 	unsigned int mtype_local;
1203 
1204 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1205 		bo_node = prange->svm_bo->node;
1206 
1207 	switch (amdgpu_ip_version(node->adev, GC_HWIP, 0)) {
1208 	case IP_VERSION(9, 4, 1):
1209 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1210 			if (bo_node == node) {
1211 				mapping_flags |= coherent ?
1212 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1213 			} else {
1214 				mapping_flags |= coherent ?
1215 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1216 				if (svm_nodes_in_same_hive(node, bo_node))
1217 					snoop = true;
1218 			}
1219 		} else {
1220 			mapping_flags |= coherent ?
1221 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1222 		}
1223 		break;
1224 	case IP_VERSION(9, 4, 2):
1225 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1226 			if (bo_node == node) {
1227 				mapping_flags |= coherent ?
1228 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1229 				if (node->adev->gmc.xgmi.connected_to_cpu)
1230 					snoop = true;
1231 			} else {
1232 				mapping_flags |= coherent ?
1233 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1234 				if (svm_nodes_in_same_hive(node, bo_node))
1235 					snoop = true;
1236 			}
1237 		} else {
1238 			mapping_flags |= coherent ?
1239 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1240 		}
1241 		break;
1242 	case IP_VERSION(9, 4, 3):
1243 	case IP_VERSION(9, 4, 4):
1244 		if (ext_coherent)
1245 			mtype_local = node->adev->rev_id ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_UC;
1246 		else
1247 			mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC :
1248 				amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1249 		snoop = true;
1250 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1251 			/* local HBM region close to partition */
1252 			if (bo_node->adev == node->adev &&
1253 			    (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
1254 				mapping_flags |= mtype_local;
1255 			/* local HBM region far from partition or remote XGMI GPU
1256 			 * with regular system scope coherence
1257 			 */
1258 			else if (svm_nodes_in_same_hive(bo_node, node) && !ext_coherent)
1259 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1260 			/* PCIe P2P or extended system scope coherence */
1261 			else
1262 				mapping_flags |= AMDGPU_VM_MTYPE_UC;
1263 		/* system memory accessed by the APU */
1264 		} else if (node->adev->flags & AMD_IS_APU) {
1265 			/* On NUMA systems, locality is determined per-page
1266 			 * in amdgpu_gmc_override_vm_pte_flags
1267 			 */
1268 			if (num_possible_nodes() <= 1)
1269 				mapping_flags |= mtype_local;
1270 			else
1271 				mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1272 		/* system memory accessed by the dGPU */
1273 		} else {
1274 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1275 		}
1276 		break;
1277 	case IP_VERSION(12, 0, 0):
1278 	case IP_VERSION(12, 0, 1):
1279 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1280 			if (bo_node != node)
1281 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1282 		} else {
1283 			mapping_flags |= coherent ?
1284 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1285 		}
1286 		break;
1287 	default:
1288 		mapping_flags |= coherent ?
1289 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1290 	}
1291 
1292 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1293 
1294 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1295 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1296 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1297 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1298 
1299 	pte_flags = AMDGPU_PTE_VALID;
1300 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1301 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1302 	if (KFD_GC_VERSION(node) >= IP_VERSION(12, 0, 0))
1303 		pte_flags |= AMDGPU_PTE_IS_PTE;
1304 
1305 	pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags);
1306 	return pte_flags;
1307 }
1308 
1309 static int
1310 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1311 			 uint64_t start, uint64_t last,
1312 			 struct dma_fence **fence)
1313 {
1314 	uint64_t init_pte_value = 0;
1315 
1316 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1317 
1318 	return amdgpu_vm_update_range(adev, vm, false, true, true, false, NULL, start,
1319 				      last, init_pte_value, 0, 0, NULL, NULL,
1320 				      fence);
1321 }
1322 
1323 static int
1324 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1325 			  unsigned long last, uint32_t trigger)
1326 {
1327 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1328 	struct kfd_process_device *pdd;
1329 	struct dma_fence *fence = NULL;
1330 	struct kfd_process *p;
1331 	uint32_t gpuidx;
1332 	int r = 0;
1333 
1334 	if (!prange->mapped_to_gpu) {
1335 		pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1336 			 prange, prange->start, prange->last);
1337 		return 0;
1338 	}
1339 
1340 	if (prange->start == start && prange->last == last) {
1341 		pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1342 		prange->mapped_to_gpu = false;
1343 	}
1344 
1345 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1346 		  MAX_GPU_INSTANCE);
1347 	p = container_of(prange->svms, struct kfd_process, svms);
1348 
1349 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1350 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1351 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1352 		if (!pdd) {
1353 			pr_debug("failed to find device idx %d\n", gpuidx);
1354 			return -EINVAL;
1355 		}
1356 
1357 		kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1358 					     start, last, trigger);
1359 
1360 		r = svm_range_unmap_from_gpu(pdd->dev->adev,
1361 					     drm_priv_to_vm(pdd->drm_priv),
1362 					     start, last, &fence);
1363 		if (r)
1364 			break;
1365 
1366 		if (fence) {
1367 			r = dma_fence_wait(fence, false);
1368 			dma_fence_put(fence);
1369 			fence = NULL;
1370 			if (r)
1371 				break;
1372 		}
1373 		kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1374 	}
1375 
1376 	return r;
1377 }
1378 
1379 static int
1380 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1381 		     unsigned long offset, unsigned long npages, bool readonly,
1382 		     dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1383 		     struct dma_fence **fence, bool flush_tlb)
1384 {
1385 	struct amdgpu_device *adev = pdd->dev->adev;
1386 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1387 	uint64_t pte_flags;
1388 	unsigned long last_start;
1389 	int last_domain;
1390 	int r = 0;
1391 	int64_t i, j;
1392 
1393 	last_start = prange->start + offset;
1394 
1395 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1396 		 last_start, last_start + npages - 1, readonly);
1397 
1398 	for (i = offset; i < offset + npages; i++) {
1399 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1400 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1401 
1402 		/* Collect all pages in the same address range and memory domain
1403 		 * that can be mapped with a single call to update mapping.
1404 		 */
1405 		if (i < offset + npages - 1 &&
1406 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1407 			continue;
1408 
1409 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1410 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1411 
1412 		pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain);
1413 		if (readonly)
1414 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1415 
1416 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1417 			 prange->svms, last_start, prange->start + i,
1418 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1419 			 pte_flags);
1420 
1421 		/* For dGPU mode, we use same vm_manager to allocate VRAM for
1422 		 * different memory partition based on fpfn/lpfn, we should use
1423 		 * same vm_manager.vram_base_offset regardless memory partition.
1424 		 */
1425 		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, true,
1426 					   NULL, last_start, prange->start + i,
1427 					   pte_flags,
1428 					   (last_start - prange->start) << PAGE_SHIFT,
1429 					   bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1430 					   NULL, dma_addr, &vm->last_update);
1431 
1432 		for (j = last_start - prange->start; j <= i; j++)
1433 			dma_addr[j] |= last_domain;
1434 
1435 		if (r) {
1436 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1437 			goto out;
1438 		}
1439 		last_start = prange->start + i + 1;
1440 	}
1441 
1442 	r = amdgpu_vm_update_pdes(adev, vm, false);
1443 	if (r) {
1444 		pr_debug("failed %d to update directories 0x%lx\n", r,
1445 			 prange->start);
1446 		goto out;
1447 	}
1448 
1449 	if (fence)
1450 		*fence = dma_fence_get(vm->last_update);
1451 
1452 out:
1453 	return r;
1454 }
1455 
1456 static int
1457 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1458 		      unsigned long npages, bool readonly,
1459 		      unsigned long *bitmap, bool wait, bool flush_tlb)
1460 {
1461 	struct kfd_process_device *pdd;
1462 	struct amdgpu_device *bo_adev = NULL;
1463 	struct kfd_process *p;
1464 	struct dma_fence *fence = NULL;
1465 	uint32_t gpuidx;
1466 	int r = 0;
1467 
1468 	if (prange->svm_bo && prange->ttm_res)
1469 		bo_adev = prange->svm_bo->node->adev;
1470 
1471 	p = container_of(prange->svms, struct kfd_process, svms);
1472 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1473 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1474 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1475 		if (!pdd) {
1476 			pr_debug("failed to find device idx %d\n", gpuidx);
1477 			return -EINVAL;
1478 		}
1479 
1480 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1481 		if (IS_ERR(pdd))
1482 			return -EINVAL;
1483 
1484 		if (bo_adev && pdd->dev->adev != bo_adev &&
1485 		    !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1486 			pr_debug("cannot map to device idx %d\n", gpuidx);
1487 			continue;
1488 		}
1489 
1490 		r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1491 					 prange->dma_addr[gpuidx],
1492 					 bo_adev, wait ? &fence : NULL,
1493 					 flush_tlb);
1494 		if (r)
1495 			break;
1496 
1497 		if (fence) {
1498 			r = dma_fence_wait(fence, false);
1499 			dma_fence_put(fence);
1500 			fence = NULL;
1501 			if (r) {
1502 				pr_debug("failed %d to dma fence wait\n", r);
1503 				break;
1504 			}
1505 		}
1506 
1507 		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1508 	}
1509 
1510 	return r;
1511 }
1512 
1513 struct svm_validate_context {
1514 	struct kfd_process *process;
1515 	struct svm_range *prange;
1516 	bool intr;
1517 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1518 	struct drm_exec exec;
1519 };
1520 
1521 static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr)
1522 {
1523 	struct kfd_process_device *pdd;
1524 	struct amdgpu_vm *vm;
1525 	uint32_t gpuidx;
1526 	int r;
1527 
1528 	drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0, 0);
1529 	drm_exec_until_all_locked(&ctx->exec) {
1530 		for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1531 			pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1532 			if (!pdd) {
1533 				pr_debug("failed to find device idx %d\n", gpuidx);
1534 				r = -EINVAL;
1535 				goto unreserve_out;
1536 			}
1537 			vm = drm_priv_to_vm(pdd->drm_priv);
1538 
1539 			r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1540 			drm_exec_retry_on_contention(&ctx->exec);
1541 			if (unlikely(r)) {
1542 				pr_debug("failed %d to reserve bo\n", r);
1543 				goto unreserve_out;
1544 			}
1545 		}
1546 	}
1547 
1548 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1549 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1550 		if (!pdd) {
1551 			pr_debug("failed to find device idx %d\n", gpuidx);
1552 			r = -EINVAL;
1553 			goto unreserve_out;
1554 		}
1555 
1556 		r = amdgpu_vm_validate(pdd->dev->adev,
1557 				       drm_priv_to_vm(pdd->drm_priv), NULL,
1558 				       svm_range_bo_validate, NULL);
1559 		if (r) {
1560 			pr_debug("failed %d validate pt bos\n", r);
1561 			goto unreserve_out;
1562 		}
1563 	}
1564 
1565 	return 0;
1566 
1567 unreserve_out:
1568 	drm_exec_fini(&ctx->exec);
1569 	return r;
1570 }
1571 
1572 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1573 {
1574 	drm_exec_fini(&ctx->exec);
1575 }
1576 
1577 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1578 {
1579 	struct kfd_process_device *pdd;
1580 
1581 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1582 	if (!pdd)
1583 		return NULL;
1584 
1585 	return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1586 }
1587 
1588 /*
1589  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1590  *
1591  * To prevent concurrent destruction or change of range attributes, the
1592  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1593  * because that would block concurrent evictions and lead to deadlocks. To
1594  * serialize concurrent migrations or validations of the same range, the
1595  * prange->migrate_mutex must be held.
1596  *
1597  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1598  * eviction fence.
1599  *
1600  * The following sequence ensures race-free validation and GPU mapping:
1601  *
1602  * 1. Reserve page table (and SVM BO if range is in VRAM)
1603  * 2. hmm_range_fault to get page addresses (if system memory)
1604  * 3. DMA-map pages (if system memory)
1605  * 4-a. Take notifier lock
1606  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1607  * 4-c. Check that the range was not split or otherwise invalidated
1608  * 4-d. Update GPU page table
1609  * 4.e. Release notifier lock
1610  * 5. Release page table (and SVM BO) reservation
1611  */
1612 static int svm_range_validate_and_map(struct mm_struct *mm,
1613 				      unsigned long map_start, unsigned long map_last,
1614 				      struct svm_range *prange, int32_t gpuidx,
1615 				      bool intr, bool wait, bool flush_tlb)
1616 {
1617 	struct svm_validate_context *ctx;
1618 	unsigned long start, end, addr;
1619 	struct kfd_process *p;
1620 	void *owner;
1621 	int32_t idx;
1622 	int r = 0;
1623 
1624 	ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
1625 	if (!ctx)
1626 		return -ENOMEM;
1627 	ctx->process = container_of(prange->svms, struct kfd_process, svms);
1628 	ctx->prange = prange;
1629 	ctx->intr = intr;
1630 
1631 	if (gpuidx < MAX_GPU_INSTANCE) {
1632 		bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE);
1633 		bitmap_set(ctx->bitmap, gpuidx, 1);
1634 	} else if (ctx->process->xnack_enabled) {
1635 		bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1636 
1637 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1638 		 * GPU, which has ACCESS attribute to the range, create mapping
1639 		 * on that GPU.
1640 		 */
1641 		if (prange->actual_loc) {
1642 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process,
1643 							prange->actual_loc);
1644 			if (gpuidx < 0) {
1645 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1646 					 prange->actual_loc);
1647 				r = -EINVAL;
1648 				goto free_ctx;
1649 			}
1650 			if (test_bit(gpuidx, prange->bitmap_access))
1651 				bitmap_set(ctx->bitmap, gpuidx, 1);
1652 		}
1653 
1654 		/*
1655 		 * If prange is already mapped or with always mapped flag,
1656 		 * update mapping on GPUs with ACCESS attribute
1657 		 */
1658 		if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1659 			if (prange->mapped_to_gpu ||
1660 			    prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)
1661 				bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1662 		}
1663 	} else {
1664 		bitmap_or(ctx->bitmap, prange->bitmap_access,
1665 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1666 	}
1667 
1668 	if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1669 		r = 0;
1670 		goto free_ctx;
1671 	}
1672 
1673 	if (prange->actual_loc && !prange->ttm_res) {
1674 		/* This should never happen. actual_loc gets set by
1675 		 * svm_migrate_ram_to_vram after allocating a BO.
1676 		 */
1677 		WARN_ONCE(1, "VRAM BO missing during validation\n");
1678 		r = -EINVAL;
1679 		goto free_ctx;
1680 	}
1681 
1682 	r = svm_range_reserve_bos(ctx, intr);
1683 	if (r)
1684 		goto free_ctx;
1685 
1686 	p = container_of(prange->svms, struct kfd_process, svms);
1687 	owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
1688 						MAX_GPU_INSTANCE));
1689 	for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) {
1690 		if (kfd_svm_page_owner(p, idx) != owner) {
1691 			owner = NULL;
1692 			break;
1693 		}
1694 	}
1695 
1696 	start = map_start << PAGE_SHIFT;
1697 	end = (map_last + 1) << PAGE_SHIFT;
1698 	for (addr = start; !r && addr < end; ) {
1699 		struct hmm_range *hmm_range = NULL;
1700 		unsigned long map_start_vma;
1701 		unsigned long map_last_vma;
1702 		struct vm_area_struct *vma;
1703 		unsigned long next = 0;
1704 		unsigned long offset;
1705 		unsigned long npages;
1706 		bool readonly;
1707 
1708 		vma = vma_lookup(mm, addr);
1709 		if (vma) {
1710 			readonly = !(vma->vm_flags & VM_WRITE);
1711 
1712 			next = min(vma->vm_end, end);
1713 			npages = (next - addr) >> PAGE_SHIFT;
1714 			WRITE_ONCE(p->svms.faulting_task, current);
1715 			r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages,
1716 						       readonly, owner, NULL,
1717 						       &hmm_range);
1718 			WRITE_ONCE(p->svms.faulting_task, NULL);
1719 			if (r)
1720 				pr_debug("failed %d to get svm range pages\n", r);
1721 		} else {
1722 			r = -EFAULT;
1723 		}
1724 
1725 		if (!r) {
1726 			offset = (addr >> PAGE_SHIFT) - prange->start;
1727 			r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
1728 					      hmm_range->hmm_pfns);
1729 			if (r)
1730 				pr_debug("failed %d to dma map range\n", r);
1731 		}
1732 
1733 		svm_range_lock(prange);
1734 
1735 		/* Free backing memory of hmm_range if it was initialized
1736 		 * Overrride return value to TRY AGAIN only if prior returns
1737 		 * were successful
1738 		 */
1739 		if (hmm_range && amdgpu_hmm_range_get_pages_done(hmm_range) && !r) {
1740 			pr_debug("hmm update the range, need validate again\n");
1741 			r = -EAGAIN;
1742 		}
1743 
1744 		if (!r && !list_empty(&prange->child_list)) {
1745 			pr_debug("range split by unmap in parallel, validate again\n");
1746 			r = -EAGAIN;
1747 		}
1748 
1749 		if (!r) {
1750 			map_start_vma = max(map_start, prange->start + offset);
1751 			map_last_vma = min(map_last, prange->start + offset + npages - 1);
1752 			if (map_start_vma <= map_last_vma) {
1753 				offset = map_start_vma - prange->start;
1754 				npages = map_last_vma - map_start_vma + 1;
1755 				r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1756 							  ctx->bitmap, wait, flush_tlb);
1757 			}
1758 		}
1759 
1760 		if (!r && next == end)
1761 			prange->mapped_to_gpu = true;
1762 
1763 		svm_range_unlock(prange);
1764 
1765 		addr = next;
1766 	}
1767 
1768 	svm_range_unreserve_bos(ctx);
1769 	if (!r)
1770 		prange->validate_timestamp = ktime_get_boottime();
1771 
1772 free_ctx:
1773 	kfree(ctx);
1774 
1775 	return r;
1776 }
1777 
1778 /**
1779  * svm_range_list_lock_and_flush_work - flush pending deferred work
1780  *
1781  * @svms: the svm range list
1782  * @mm: the mm structure
1783  *
1784  * Context: Returns with mmap write lock held, pending deferred work flushed
1785  *
1786  */
1787 void
1788 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1789 				   struct mm_struct *mm)
1790 {
1791 retry_flush_work:
1792 	flush_work(&svms->deferred_list_work);
1793 	mmap_write_lock(mm);
1794 
1795 	if (list_empty(&svms->deferred_range_list))
1796 		return;
1797 	mmap_write_unlock(mm);
1798 	pr_debug("retry flush\n");
1799 	goto retry_flush_work;
1800 }
1801 
1802 static void svm_range_restore_work(struct work_struct *work)
1803 {
1804 	struct delayed_work *dwork = to_delayed_work(work);
1805 	struct amdkfd_process_info *process_info;
1806 	struct svm_range_list *svms;
1807 	struct svm_range *prange;
1808 	struct kfd_process *p;
1809 	struct mm_struct *mm;
1810 	int evicted_ranges;
1811 	int invalid;
1812 	int r;
1813 
1814 	svms = container_of(dwork, struct svm_range_list, restore_work);
1815 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1816 	if (!evicted_ranges)
1817 		return;
1818 
1819 	pr_debug("restore svm ranges\n");
1820 
1821 	p = container_of(svms, struct kfd_process, svms);
1822 	process_info = p->kgd_process_info;
1823 
1824 	/* Keep mm reference when svm_range_validate_and_map ranges */
1825 	mm = get_task_mm(p->lead_thread);
1826 	if (!mm) {
1827 		pr_debug("svms 0x%p process mm gone\n", svms);
1828 		return;
1829 	}
1830 
1831 	mutex_lock(&process_info->lock);
1832 	svm_range_list_lock_and_flush_work(svms, mm);
1833 	mutex_lock(&svms->lock);
1834 
1835 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1836 
1837 	list_for_each_entry(prange, &svms->list, list) {
1838 		invalid = atomic_read(&prange->invalid);
1839 		if (!invalid)
1840 			continue;
1841 
1842 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1843 			 prange->svms, prange, prange->start, prange->last,
1844 			 invalid);
1845 
1846 		/*
1847 		 * If range is migrating, wait for migration is done.
1848 		 */
1849 		mutex_lock(&prange->migrate_mutex);
1850 
1851 		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
1852 					       MAX_GPU_INSTANCE, false, true, false);
1853 		if (r)
1854 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1855 				 prange->start);
1856 
1857 		mutex_unlock(&prange->migrate_mutex);
1858 		if (r)
1859 			goto out_reschedule;
1860 
1861 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1862 			goto out_reschedule;
1863 	}
1864 
1865 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1866 	    evicted_ranges)
1867 		goto out_reschedule;
1868 
1869 	evicted_ranges = 0;
1870 
1871 	r = kgd2kfd_resume_mm(mm);
1872 	if (r) {
1873 		/* No recovery from this failure. Probably the CP is
1874 		 * hanging. No point trying again.
1875 		 */
1876 		pr_debug("failed %d to resume KFD\n", r);
1877 	}
1878 
1879 	pr_debug("restore svm ranges successfully\n");
1880 
1881 out_reschedule:
1882 	mutex_unlock(&svms->lock);
1883 	mmap_write_unlock(mm);
1884 	mutex_unlock(&process_info->lock);
1885 
1886 	/* If validation failed, reschedule another attempt */
1887 	if (evicted_ranges) {
1888 		pr_debug("reschedule to restore svm range\n");
1889 		queue_delayed_work(system_freezable_wq, &svms->restore_work,
1890 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1891 
1892 		kfd_smi_event_queue_restore_rescheduled(mm);
1893 	}
1894 	mmput(mm);
1895 }
1896 
1897 /**
1898  * svm_range_evict - evict svm range
1899  * @prange: svm range structure
1900  * @mm: current process mm_struct
1901  * @start: starting process queue number
1902  * @last: last process queue number
1903  * @event: mmu notifier event when range is evicted or migrated
1904  *
1905  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1906  * return to let CPU evict the buffer and proceed CPU pagetable update.
1907  *
1908  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1909  * If invalidation happens while restore work is running, restore work will
1910  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1911  * the queues.
1912  */
1913 static int
1914 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1915 		unsigned long start, unsigned long last,
1916 		enum mmu_notifier_event event)
1917 {
1918 	struct svm_range_list *svms = prange->svms;
1919 	struct svm_range *pchild;
1920 	struct kfd_process *p;
1921 	int r = 0;
1922 
1923 	p = container_of(svms, struct kfd_process, svms);
1924 
1925 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1926 		 svms, prange->start, prange->last, start, last);
1927 
1928 	if (!p->xnack_enabled ||
1929 	    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1930 		int evicted_ranges;
1931 		bool mapped = prange->mapped_to_gpu;
1932 
1933 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1934 			if (!pchild->mapped_to_gpu)
1935 				continue;
1936 			mapped = true;
1937 			mutex_lock_nested(&pchild->lock, 1);
1938 			if (pchild->start <= last && pchild->last >= start) {
1939 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1940 					 pchild->start, pchild->last);
1941 				atomic_inc(&pchild->invalid);
1942 			}
1943 			mutex_unlock(&pchild->lock);
1944 		}
1945 
1946 		if (!mapped)
1947 			return r;
1948 
1949 		if (prange->start <= last && prange->last >= start)
1950 			atomic_inc(&prange->invalid);
1951 
1952 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1953 		if (evicted_ranges != 1)
1954 			return r;
1955 
1956 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1957 			 prange->svms, prange->start, prange->last);
1958 
1959 		/* First eviction, stop the queues */
1960 		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1961 		if (r)
1962 			pr_debug("failed to quiesce KFD\n");
1963 
1964 		pr_debug("schedule to restore svm %p ranges\n", svms);
1965 		queue_delayed_work(system_freezable_wq, &svms->restore_work,
1966 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1967 	} else {
1968 		unsigned long s, l;
1969 		uint32_t trigger;
1970 
1971 		if (event == MMU_NOTIFY_MIGRATE)
1972 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1973 		else
1974 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1975 
1976 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1977 			 prange->svms, start, last);
1978 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1979 			mutex_lock_nested(&pchild->lock, 1);
1980 			s = max(start, pchild->start);
1981 			l = min(last, pchild->last);
1982 			if (l >= s)
1983 				svm_range_unmap_from_gpus(pchild, s, l, trigger);
1984 			mutex_unlock(&pchild->lock);
1985 		}
1986 		s = max(start, prange->start);
1987 		l = min(last, prange->last);
1988 		if (l >= s)
1989 			svm_range_unmap_from_gpus(prange, s, l, trigger);
1990 	}
1991 
1992 	return r;
1993 }
1994 
1995 static struct svm_range *svm_range_clone(struct svm_range *old)
1996 {
1997 	struct svm_range *new;
1998 
1999 	new = svm_range_new(old->svms, old->start, old->last, false);
2000 	if (!new)
2001 		return NULL;
2002 	if (svm_range_copy_dma_addrs(new, old)) {
2003 		svm_range_free(new, false);
2004 		return NULL;
2005 	}
2006 	if (old->svm_bo) {
2007 		new->ttm_res = old->ttm_res;
2008 		new->offset = old->offset;
2009 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
2010 		spin_lock(&new->svm_bo->list_lock);
2011 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
2012 		spin_unlock(&new->svm_bo->list_lock);
2013 	}
2014 	new->flags = old->flags;
2015 	new->preferred_loc = old->preferred_loc;
2016 	new->prefetch_loc = old->prefetch_loc;
2017 	new->actual_loc = old->actual_loc;
2018 	new->granularity = old->granularity;
2019 	new->mapped_to_gpu = old->mapped_to_gpu;
2020 	new->vram_pages = old->vram_pages;
2021 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
2022 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
2023 	atomic_set(&new->queue_refcount, atomic_read(&old->queue_refcount));
2024 
2025 	return new;
2026 }
2027 
2028 void svm_range_set_max_pages(struct amdgpu_device *adev)
2029 {
2030 	uint64_t max_pages;
2031 	uint64_t pages, _pages;
2032 	uint64_t min_pages = 0;
2033 	int i, id;
2034 
2035 	for (i = 0; i < adev->kfd.dev->num_nodes; i++) {
2036 		if (adev->kfd.dev->nodes[i]->xcp)
2037 			id = adev->kfd.dev->nodes[i]->xcp->id;
2038 		else
2039 			id = -1;
2040 		pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17;
2041 		pages = clamp(pages, 1ULL << 9, 1ULL << 18);
2042 		pages = rounddown_pow_of_two(pages);
2043 		min_pages = min_not_zero(min_pages, pages);
2044 	}
2045 
2046 	do {
2047 		max_pages = READ_ONCE(max_svm_range_pages);
2048 		_pages = min_not_zero(max_pages, min_pages);
2049 	} while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
2050 }
2051 
2052 static int
2053 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
2054 		    uint64_t max_pages, struct list_head *insert_list,
2055 		    struct list_head *update_list)
2056 {
2057 	struct svm_range *prange;
2058 	uint64_t l;
2059 
2060 	pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
2061 		 max_pages, start, last);
2062 
2063 	while (last >= start) {
2064 		l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
2065 
2066 		prange = svm_range_new(svms, start, l, true);
2067 		if (!prange)
2068 			return -ENOMEM;
2069 		list_add(&prange->list, insert_list);
2070 		list_add(&prange->update_list, update_list);
2071 
2072 		start = l + 1;
2073 	}
2074 	return 0;
2075 }
2076 
2077 /**
2078  * svm_range_add - add svm range and handle overlap
2079  * @p: the range add to this process svms
2080  * @start: page size aligned
2081  * @size: page size aligned
2082  * @nattr: number of attributes
2083  * @attrs: array of attributes
2084  * @update_list: output, the ranges need validate and update GPU mapping
2085  * @insert_list: output, the ranges need insert to svms
2086  * @remove_list: output, the ranges are replaced and need remove from svms
2087  * @remap_list: output, remap unaligned svm ranges
2088  *
2089  * Check if the virtual address range has overlap with any existing ranges,
2090  * split partly overlapping ranges and add new ranges in the gaps. All changes
2091  * should be applied to the range_list and interval tree transactionally. If
2092  * any range split or allocation fails, the entire update fails. Therefore any
2093  * existing overlapping svm_ranges are cloned and the original svm_ranges left
2094  * unchanged.
2095  *
2096  * If the transaction succeeds, the caller can update and insert clones and
2097  * new ranges, then free the originals.
2098  *
2099  * Otherwise the caller can free the clones and new ranges, while the old
2100  * svm_ranges remain unchanged.
2101  *
2102  * Context: Process context, caller must hold svms->lock
2103  *
2104  * Return:
2105  * 0 - OK, otherwise error code
2106  */
2107 static int
2108 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2109 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2110 	      struct list_head *update_list, struct list_head *insert_list,
2111 	      struct list_head *remove_list, struct list_head *remap_list)
2112 {
2113 	unsigned long last = start + size - 1UL;
2114 	struct svm_range_list *svms = &p->svms;
2115 	struct interval_tree_node *node;
2116 	struct svm_range *prange;
2117 	struct svm_range *tmp;
2118 	struct list_head new_list;
2119 	int r = 0;
2120 
2121 	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
2122 
2123 	INIT_LIST_HEAD(update_list);
2124 	INIT_LIST_HEAD(insert_list);
2125 	INIT_LIST_HEAD(remove_list);
2126 	INIT_LIST_HEAD(&new_list);
2127 	INIT_LIST_HEAD(remap_list);
2128 
2129 	node = interval_tree_iter_first(&svms->objects, start, last);
2130 	while (node) {
2131 		struct interval_tree_node *next;
2132 		unsigned long next_start;
2133 
2134 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2135 			 node->last);
2136 
2137 		prange = container_of(node, struct svm_range, it_node);
2138 		next = interval_tree_iter_next(node, start, last);
2139 		next_start = min(node->last, last) + 1;
2140 
2141 		if (svm_range_is_same_attrs(p, prange, nattr, attrs) &&
2142 		    prange->mapped_to_gpu) {
2143 			/* nothing to do */
2144 		} else if (node->start < start || node->last > last) {
2145 			/* node intersects the update range and its attributes
2146 			 * will change. Clone and split it, apply updates only
2147 			 * to the overlapping part
2148 			 */
2149 			struct svm_range *old = prange;
2150 
2151 			prange = svm_range_clone(old);
2152 			if (!prange) {
2153 				r = -ENOMEM;
2154 				goto out;
2155 			}
2156 
2157 			list_add(&old->update_list, remove_list);
2158 			list_add(&prange->list, insert_list);
2159 			list_add(&prange->update_list, update_list);
2160 
2161 			if (node->start < start) {
2162 				pr_debug("change old range start\n");
2163 				r = svm_range_split_head(prange, start,
2164 							 insert_list, remap_list);
2165 				if (r)
2166 					goto out;
2167 			}
2168 			if (node->last > last) {
2169 				pr_debug("change old range last\n");
2170 				r = svm_range_split_tail(prange, last,
2171 							 insert_list, remap_list);
2172 				if (r)
2173 					goto out;
2174 			}
2175 		} else {
2176 			/* The node is contained within start..last,
2177 			 * just update it
2178 			 */
2179 			list_add(&prange->update_list, update_list);
2180 		}
2181 
2182 		/* insert a new node if needed */
2183 		if (node->start > start) {
2184 			r = svm_range_split_new(svms, start, node->start - 1,
2185 						READ_ONCE(max_svm_range_pages),
2186 						&new_list, update_list);
2187 			if (r)
2188 				goto out;
2189 		}
2190 
2191 		node = next;
2192 		start = next_start;
2193 	}
2194 
2195 	/* add a final range at the end if needed */
2196 	if (start <= last)
2197 		r = svm_range_split_new(svms, start, last,
2198 					READ_ONCE(max_svm_range_pages),
2199 					&new_list, update_list);
2200 
2201 out:
2202 	if (r) {
2203 		list_for_each_entry_safe(prange, tmp, insert_list, list)
2204 			svm_range_free(prange, false);
2205 		list_for_each_entry_safe(prange, tmp, &new_list, list)
2206 			svm_range_free(prange, true);
2207 	} else {
2208 		list_splice(&new_list, insert_list);
2209 	}
2210 
2211 	return r;
2212 }
2213 
2214 static void
2215 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2216 					    struct svm_range *prange)
2217 {
2218 	unsigned long start;
2219 	unsigned long last;
2220 
2221 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2222 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2223 
2224 	if (prange->start == start && prange->last == last)
2225 		return;
2226 
2227 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2228 		  prange->svms, prange, start, last, prange->start,
2229 		  prange->last);
2230 
2231 	if (start != 0 && last != 0) {
2232 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
2233 		svm_range_remove_notifier(prange);
2234 	}
2235 	prange->it_node.start = prange->start;
2236 	prange->it_node.last = prange->last;
2237 
2238 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
2239 	svm_range_add_notifier_locked(mm, prange);
2240 }
2241 
2242 static void
2243 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2244 			 struct mm_struct *mm)
2245 {
2246 	switch (prange->work_item.op) {
2247 	case SVM_OP_NULL:
2248 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2249 			 svms, prange, prange->start, prange->last);
2250 		break;
2251 	case SVM_OP_UNMAP_RANGE:
2252 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2253 			 svms, prange, prange->start, prange->last);
2254 		svm_range_unlink(prange);
2255 		svm_range_remove_notifier(prange);
2256 		svm_range_free(prange, true);
2257 		break;
2258 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
2259 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2260 			 svms, prange, prange->start, prange->last);
2261 		svm_range_update_notifier_and_interval_tree(mm, prange);
2262 		break;
2263 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2264 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2265 			 svms, prange, prange->start, prange->last);
2266 		svm_range_update_notifier_and_interval_tree(mm, prange);
2267 		/* TODO: implement deferred validation and mapping */
2268 		break;
2269 	case SVM_OP_ADD_RANGE:
2270 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2271 			 prange->start, prange->last);
2272 		svm_range_add_to_svms(prange);
2273 		svm_range_add_notifier_locked(mm, prange);
2274 		break;
2275 	case SVM_OP_ADD_RANGE_AND_MAP:
2276 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2277 			 prange, prange->start, prange->last);
2278 		svm_range_add_to_svms(prange);
2279 		svm_range_add_notifier_locked(mm, prange);
2280 		/* TODO: implement deferred validation and mapping */
2281 		break;
2282 	default:
2283 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2284 			 prange->work_item.op);
2285 	}
2286 }
2287 
2288 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2289 {
2290 	struct kfd_process_device *pdd;
2291 	struct kfd_process *p;
2292 	uint32_t i;
2293 
2294 	p = container_of(svms, struct kfd_process, svms);
2295 
2296 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2297 		pdd = p->pdds[i];
2298 		if (!pdd)
2299 			continue;
2300 
2301 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2302 
2303 		amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2304 				pdd->dev->adev->irq.retry_cam_enabled ?
2305 				&pdd->dev->adev->irq.ih :
2306 				&pdd->dev->adev->irq.ih1);
2307 
2308 		if (pdd->dev->adev->irq.retry_cam_enabled)
2309 			amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2310 				&pdd->dev->adev->irq.ih_soft);
2311 
2312 
2313 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2314 	}
2315 }
2316 
2317 static void svm_range_deferred_list_work(struct work_struct *work)
2318 {
2319 	struct svm_range_list *svms;
2320 	struct svm_range *prange;
2321 	struct mm_struct *mm;
2322 
2323 	svms = container_of(work, struct svm_range_list, deferred_list_work);
2324 	pr_debug("enter svms 0x%p\n", svms);
2325 
2326 	spin_lock(&svms->deferred_list_lock);
2327 	while (!list_empty(&svms->deferred_range_list)) {
2328 		prange = list_first_entry(&svms->deferred_range_list,
2329 					  struct svm_range, deferred_list);
2330 		spin_unlock(&svms->deferred_list_lock);
2331 
2332 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2333 			 prange->start, prange->last, prange->work_item.op);
2334 
2335 		mm = prange->work_item.mm;
2336 
2337 		mmap_write_lock(mm);
2338 
2339 		/* Remove from deferred_list must be inside mmap write lock, for
2340 		 * two race cases:
2341 		 * 1. unmap_from_cpu may change work_item.op and add the range
2342 		 *    to deferred_list again, cause use after free bug.
2343 		 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2344 		 *    lock and continue because deferred_list is empty, but
2345 		 *    deferred_list work is actually waiting for mmap lock.
2346 		 */
2347 		spin_lock(&svms->deferred_list_lock);
2348 		list_del_init(&prange->deferred_list);
2349 		spin_unlock(&svms->deferred_list_lock);
2350 
2351 		mutex_lock(&svms->lock);
2352 		mutex_lock(&prange->migrate_mutex);
2353 		while (!list_empty(&prange->child_list)) {
2354 			struct svm_range *pchild;
2355 
2356 			pchild = list_first_entry(&prange->child_list,
2357 						struct svm_range, child_list);
2358 			pr_debug("child prange 0x%p op %d\n", pchild,
2359 				 pchild->work_item.op);
2360 			list_del_init(&pchild->child_list);
2361 			svm_range_handle_list_op(svms, pchild, mm);
2362 		}
2363 		mutex_unlock(&prange->migrate_mutex);
2364 
2365 		svm_range_handle_list_op(svms, prange, mm);
2366 		mutex_unlock(&svms->lock);
2367 		mmap_write_unlock(mm);
2368 
2369 		/* Pairs with mmget in svm_range_add_list_work. If dropping the
2370 		 * last mm refcount, schedule release work to avoid circular locking
2371 		 */
2372 		mmput_async(mm);
2373 
2374 		spin_lock(&svms->deferred_list_lock);
2375 	}
2376 	spin_unlock(&svms->deferred_list_lock);
2377 	pr_debug("exit svms 0x%p\n", svms);
2378 }
2379 
2380 void
2381 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2382 			struct mm_struct *mm, enum svm_work_list_ops op)
2383 {
2384 	spin_lock(&svms->deferred_list_lock);
2385 	/* if prange is on the deferred list */
2386 	if (!list_empty(&prange->deferred_list)) {
2387 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2388 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2389 		if (op != SVM_OP_NULL &&
2390 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2391 			prange->work_item.op = op;
2392 	} else {
2393 		prange->work_item.op = op;
2394 
2395 		/* Pairs with mmput in deferred_list_work */
2396 		mmget(mm);
2397 		prange->work_item.mm = mm;
2398 		list_add_tail(&prange->deferred_list,
2399 			      &prange->svms->deferred_range_list);
2400 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2401 			 prange, prange->start, prange->last, op);
2402 	}
2403 	spin_unlock(&svms->deferred_list_lock);
2404 }
2405 
2406 void schedule_deferred_list_work(struct svm_range_list *svms)
2407 {
2408 	spin_lock(&svms->deferred_list_lock);
2409 	if (!list_empty(&svms->deferred_range_list))
2410 		schedule_work(&svms->deferred_list_work);
2411 	spin_unlock(&svms->deferred_list_lock);
2412 }
2413 
2414 static void
2415 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2416 		      struct svm_range *prange, unsigned long start,
2417 		      unsigned long last)
2418 {
2419 	struct svm_range *head;
2420 	struct svm_range *tail;
2421 
2422 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2423 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2424 			 prange->start, prange->last);
2425 		return;
2426 	}
2427 	if (start > prange->last || last < prange->start)
2428 		return;
2429 
2430 	head = tail = prange;
2431 	if (start > prange->start)
2432 		svm_range_split(prange, prange->start, start - 1, &tail);
2433 	if (last < tail->last)
2434 		svm_range_split(tail, last + 1, tail->last, &head);
2435 
2436 	if (head != prange && tail != prange) {
2437 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2438 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2439 	} else if (tail != prange) {
2440 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2441 	} else if (head != prange) {
2442 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2443 	} else if (parent != prange) {
2444 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2445 	}
2446 }
2447 
2448 static void
2449 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2450 			 unsigned long start, unsigned long last)
2451 {
2452 	uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU;
2453 	struct svm_range_list *svms;
2454 	struct svm_range *pchild;
2455 	struct kfd_process *p;
2456 	unsigned long s, l;
2457 	bool unmap_parent;
2458 	uint32_t i;
2459 
2460 	if (atomic_read(&prange->queue_refcount)) {
2461 		int r;
2462 
2463 		pr_warn("Freeing queue vital buffer 0x%lx, queue evicted\n",
2464 			prange->start << PAGE_SHIFT);
2465 		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
2466 		if (r)
2467 			pr_debug("failed %d to quiesce KFD queues\n", r);
2468 	}
2469 
2470 	p = kfd_lookup_process_by_mm(mm);
2471 	if (!p)
2472 		return;
2473 	svms = &p->svms;
2474 
2475 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2476 		 prange, prange->start, prange->last, start, last);
2477 
2478 	/* calculate time stamps that are used to decide which page faults need be
2479 	 * dropped or handled before unmap pages from gpu vm
2480 	 */
2481 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2482 		struct kfd_process_device *pdd;
2483 		struct amdgpu_device *adev;
2484 		struct amdgpu_ih_ring *ih;
2485 		uint32_t checkpoint_wptr;
2486 
2487 		pdd = p->pdds[i];
2488 		if (!pdd)
2489 			continue;
2490 
2491 		adev = pdd->dev->adev;
2492 
2493 		/* Check and drain ih1 ring if cam not available */
2494 		if (adev->irq.ih1.ring_size) {
2495 			ih = &adev->irq.ih1;
2496 			checkpoint_wptr = amdgpu_ih_get_wptr(adev, ih);
2497 			if (ih->rptr != checkpoint_wptr) {
2498 				svms->checkpoint_ts[i] =
2499 					amdgpu_ih_decode_iv_ts(adev, ih, checkpoint_wptr, -1);
2500 				continue;
2501 			}
2502 		}
2503 
2504 		/* check if dev->irq.ih_soft is not empty */
2505 		ih = &adev->irq.ih_soft;
2506 		checkpoint_wptr = amdgpu_ih_get_wptr(adev, ih);
2507 		if (ih->rptr != checkpoint_wptr)
2508 			svms->checkpoint_ts[i] = amdgpu_ih_decode_iv_ts(adev, ih, checkpoint_wptr, -1);
2509 	}
2510 
2511 	unmap_parent = start <= prange->start && last >= prange->last;
2512 
2513 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2514 		mutex_lock_nested(&pchild->lock, 1);
2515 		s = max(start, pchild->start);
2516 		l = min(last, pchild->last);
2517 		if (l >= s)
2518 			svm_range_unmap_from_gpus(pchild, s, l, trigger);
2519 		svm_range_unmap_split(mm, prange, pchild, start, last);
2520 		mutex_unlock(&pchild->lock);
2521 	}
2522 	s = max(start, prange->start);
2523 	l = min(last, prange->last);
2524 	if (l >= s)
2525 		svm_range_unmap_from_gpus(prange, s, l, trigger);
2526 	svm_range_unmap_split(mm, prange, prange, start, last);
2527 
2528 	if (unmap_parent)
2529 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2530 	else
2531 		svm_range_add_list_work(svms, prange, mm,
2532 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2533 	schedule_deferred_list_work(svms);
2534 
2535 	kfd_unref_process(p);
2536 }
2537 
2538 /**
2539  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2540  * @mni: mmu_interval_notifier struct
2541  * @range: mmu_notifier_range struct
2542  * @cur_seq: value to pass to mmu_interval_set_seq()
2543  *
2544  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2545  * is from migration, or CPU page invalidation callback.
2546  *
2547  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2548  * work thread, and split prange if only part of prange is unmapped.
2549  *
2550  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2551  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2552  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2553  * update GPU mapping to recover.
2554  *
2555  * Context: mmap lock, notifier_invalidate_start lock are held
2556  *          for invalidate event, prange lock is held if this is from migration
2557  */
2558 static bool
2559 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2560 				    const struct mmu_notifier_range *range,
2561 				    unsigned long cur_seq)
2562 {
2563 	struct svm_range *prange;
2564 	unsigned long start;
2565 	unsigned long last;
2566 
2567 	if (range->event == MMU_NOTIFY_RELEASE)
2568 		return true;
2569 	if (!mmget_not_zero(mni->mm))
2570 		return true;
2571 
2572 	start = mni->interval_tree.start;
2573 	last = mni->interval_tree.last;
2574 	start = max(start, range->start) >> PAGE_SHIFT;
2575 	last = min(last, range->end - 1) >> PAGE_SHIFT;
2576 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2577 		 start, last, range->start >> PAGE_SHIFT,
2578 		 (range->end - 1) >> PAGE_SHIFT,
2579 		 mni->interval_tree.start >> PAGE_SHIFT,
2580 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2581 
2582 	prange = container_of(mni, struct svm_range, notifier);
2583 
2584 	svm_range_lock(prange);
2585 	mmu_interval_set_seq(mni, cur_seq);
2586 
2587 	switch (range->event) {
2588 	case MMU_NOTIFY_UNMAP:
2589 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2590 		break;
2591 	default:
2592 		svm_range_evict(prange, mni->mm, start, last, range->event);
2593 		break;
2594 	}
2595 
2596 	svm_range_unlock(prange);
2597 	mmput(mni->mm);
2598 
2599 	return true;
2600 }
2601 
2602 /**
2603  * svm_range_from_addr - find svm range from fault address
2604  * @svms: svm range list header
2605  * @addr: address to search range interval tree, in pages
2606  * @parent: parent range if range is on child list
2607  *
2608  * Context: The caller must hold svms->lock
2609  *
2610  * Return: the svm_range found or NULL
2611  */
2612 struct svm_range *
2613 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2614 		    struct svm_range **parent)
2615 {
2616 	struct interval_tree_node *node;
2617 	struct svm_range *prange;
2618 	struct svm_range *pchild;
2619 
2620 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2621 	if (!node)
2622 		return NULL;
2623 
2624 	prange = container_of(node, struct svm_range, it_node);
2625 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2626 		 addr, prange->start, prange->last, node->start, node->last);
2627 
2628 	if (addr >= prange->start && addr <= prange->last) {
2629 		if (parent)
2630 			*parent = prange;
2631 		return prange;
2632 	}
2633 	list_for_each_entry(pchild, &prange->child_list, child_list)
2634 		if (addr >= pchild->start && addr <= pchild->last) {
2635 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2636 				 addr, pchild->start, pchild->last);
2637 			if (parent)
2638 				*parent = prange;
2639 			return pchild;
2640 		}
2641 
2642 	return NULL;
2643 }
2644 
2645 /* svm_range_best_restore_location - decide the best fault restore location
2646  * @prange: svm range structure
2647  * @adev: the GPU on which vm fault happened
2648  *
2649  * This is only called when xnack is on, to decide the best location to restore
2650  * the range mapping after GPU vm fault. Caller uses the best location to do
2651  * migration if actual loc is not best location, then update GPU page table
2652  * mapping to the best location.
2653  *
2654  * If the preferred loc is accessible by faulting GPU, use preferred loc.
2655  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2656  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2657  *    if range actual loc is cpu, best_loc is cpu
2658  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2659  *    range actual loc.
2660  * Otherwise, GPU no access, best_loc is -1.
2661  *
2662  * Return:
2663  * -1 means vm fault GPU no access
2664  * 0 for CPU or GPU id
2665  */
2666 static int32_t
2667 svm_range_best_restore_location(struct svm_range *prange,
2668 				struct kfd_node *node,
2669 				int32_t *gpuidx)
2670 {
2671 	struct kfd_node *bo_node, *preferred_node;
2672 	struct kfd_process *p;
2673 	uint32_t gpuid;
2674 	int r;
2675 
2676 	p = container_of(prange->svms, struct kfd_process, svms);
2677 
2678 	r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx);
2679 	if (r < 0) {
2680 		pr_debug("failed to get gpuid from kgd\n");
2681 		return -1;
2682 	}
2683 
2684 	if (node->adev->flags & AMD_IS_APU)
2685 		return 0;
2686 
2687 	if (prange->preferred_loc == gpuid ||
2688 	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2689 		return prange->preferred_loc;
2690 	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2691 		preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc);
2692 		if (preferred_node && svm_nodes_in_same_hive(node, preferred_node))
2693 			return prange->preferred_loc;
2694 		/* fall through */
2695 	}
2696 
2697 	if (test_bit(*gpuidx, prange->bitmap_access))
2698 		return gpuid;
2699 
2700 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2701 		if (!prange->actual_loc)
2702 			return 0;
2703 
2704 		bo_node = svm_range_get_node_by_id(prange, prange->actual_loc);
2705 		if (bo_node && svm_nodes_in_same_hive(node, bo_node))
2706 			return prange->actual_loc;
2707 		else
2708 			return 0;
2709 	}
2710 
2711 	return -1;
2712 }
2713 
2714 static int
2715 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2716 			       unsigned long *start, unsigned long *last,
2717 			       bool *is_heap_stack)
2718 {
2719 	struct vm_area_struct *vma;
2720 	struct interval_tree_node *node;
2721 	struct rb_node *rb_node;
2722 	unsigned long start_limit, end_limit;
2723 
2724 	vma = vma_lookup(p->mm, addr << PAGE_SHIFT);
2725 	if (!vma) {
2726 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2727 		return -EFAULT;
2728 	}
2729 
2730 	*is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma);
2731 
2732 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2733 		      (unsigned long)ALIGN_DOWN(addr, 1UL << p->svms.default_granularity));
2734 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2735 		    (unsigned long)ALIGN(addr + 1, 1UL << p->svms.default_granularity));
2736 
2737 	/* First range that starts after the fault address */
2738 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2739 	if (node) {
2740 		end_limit = min(end_limit, node->start);
2741 		/* Last range that ends before the fault address */
2742 		rb_node = rb_prev(&node->rb);
2743 	} else {
2744 		/* Last range must end before addr because
2745 		 * there was no range after addr
2746 		 */
2747 		rb_node = rb_last(&p->svms.objects.rb_root);
2748 	}
2749 	if (rb_node) {
2750 		node = container_of(rb_node, struct interval_tree_node, rb);
2751 		if (node->last >= addr) {
2752 			WARN(1, "Overlap with prev node and page fault addr\n");
2753 			return -EFAULT;
2754 		}
2755 		start_limit = max(start_limit, node->last + 1);
2756 	}
2757 
2758 	*start = start_limit;
2759 	*last = end_limit - 1;
2760 
2761 	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2762 		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2763 		 *start, *last, *is_heap_stack);
2764 
2765 	return 0;
2766 }
2767 
2768 static int
2769 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2770 			   uint64_t *bo_s, uint64_t *bo_l)
2771 {
2772 	struct amdgpu_bo_va_mapping *mapping;
2773 	struct interval_tree_node *node;
2774 	struct amdgpu_bo *bo = NULL;
2775 	unsigned long userptr;
2776 	uint32_t i;
2777 	int r;
2778 
2779 	for (i = 0; i < p->n_pdds; i++) {
2780 		struct amdgpu_vm *vm;
2781 
2782 		if (!p->pdds[i]->drm_priv)
2783 			continue;
2784 
2785 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2786 		r = amdgpu_bo_reserve(vm->root.bo, false);
2787 		if (r)
2788 			return r;
2789 
2790 		/* Check userptr by searching entire vm->va interval tree */
2791 		node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2792 		while (node) {
2793 			mapping = container_of((struct rb_node *)node,
2794 					       struct amdgpu_bo_va_mapping, rb);
2795 			bo = mapping->bo_va->base.bo;
2796 
2797 			if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2798 							 start << PAGE_SHIFT,
2799 							 last << PAGE_SHIFT,
2800 							 &userptr)) {
2801 				node = interval_tree_iter_next(node, 0, ~0ULL);
2802 				continue;
2803 			}
2804 
2805 			pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2806 				 start, last);
2807 			if (bo_s && bo_l) {
2808 				*bo_s = userptr >> PAGE_SHIFT;
2809 				*bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2810 			}
2811 			amdgpu_bo_unreserve(vm->root.bo);
2812 			return -EADDRINUSE;
2813 		}
2814 		amdgpu_bo_unreserve(vm->root.bo);
2815 	}
2816 	return 0;
2817 }
2818 
2819 static struct
2820 svm_range *svm_range_create_unregistered_range(struct kfd_node *node,
2821 						struct kfd_process *p,
2822 						struct mm_struct *mm,
2823 						int64_t addr)
2824 {
2825 	struct svm_range *prange = NULL;
2826 	unsigned long start, last;
2827 	uint32_t gpuid, gpuidx;
2828 	bool is_heap_stack;
2829 	uint64_t bo_s = 0;
2830 	uint64_t bo_l = 0;
2831 	int r;
2832 
2833 	if (svm_range_get_range_boundaries(p, addr, &start, &last,
2834 					   &is_heap_stack))
2835 		return NULL;
2836 
2837 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2838 	if (r != -EADDRINUSE)
2839 		r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2840 
2841 	if (r == -EADDRINUSE) {
2842 		if (addr >= bo_s && addr <= bo_l)
2843 			return NULL;
2844 
2845 		/* Create one page svm range if 2MB range overlapping */
2846 		start = addr;
2847 		last = addr;
2848 	}
2849 
2850 	prange = svm_range_new(&p->svms, start, last, true);
2851 	if (!prange) {
2852 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2853 		return NULL;
2854 	}
2855 	if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) {
2856 		pr_debug("failed to get gpuid from kgd\n");
2857 		svm_range_free(prange, true);
2858 		return NULL;
2859 	}
2860 
2861 	if (is_heap_stack)
2862 		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2863 
2864 	svm_range_add_to_svms(prange);
2865 	svm_range_add_notifier_locked(mm, prange);
2866 
2867 	return prange;
2868 }
2869 
2870 /* svm_range_skip_recover - decide if prange can be recovered
2871  * @prange: svm range structure
2872  *
2873  * GPU vm retry fault handle skip recover the range for cases:
2874  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2875  *    deferred list work will drain the stale fault before free the prange.
2876  * 2. prange is on deferred list to add interval notifier after split, or
2877  * 3. prange is child range, it is split from parent prange, recover later
2878  *    after interval notifier is added.
2879  *
2880  * Return: true to skip recover, false to recover
2881  */
2882 static bool svm_range_skip_recover(struct svm_range *prange)
2883 {
2884 	struct svm_range_list *svms = prange->svms;
2885 
2886 	spin_lock(&svms->deferred_list_lock);
2887 	if (list_empty(&prange->deferred_list) &&
2888 	    list_empty(&prange->child_list)) {
2889 		spin_unlock(&svms->deferred_list_lock);
2890 		return false;
2891 	}
2892 	spin_unlock(&svms->deferred_list_lock);
2893 
2894 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2895 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2896 			 svms, prange, prange->start, prange->last);
2897 		return true;
2898 	}
2899 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2900 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2901 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2902 			 svms, prange, prange->start, prange->last);
2903 		return true;
2904 	}
2905 	return false;
2906 }
2907 
2908 static void
2909 svm_range_count_fault(struct kfd_node *node, struct kfd_process *p,
2910 		      int32_t gpuidx)
2911 {
2912 	struct kfd_process_device *pdd;
2913 
2914 	/* fault is on different page of same range
2915 	 * or fault is skipped to recover later
2916 	 * or fault is on invalid virtual address
2917 	 */
2918 	if (gpuidx == MAX_GPU_INSTANCE) {
2919 		uint32_t gpuid;
2920 		int r;
2921 
2922 		r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx);
2923 		if (r < 0)
2924 			return;
2925 	}
2926 
2927 	/* fault is recovered
2928 	 * or fault cannot recover because GPU no access on the range
2929 	 */
2930 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2931 	if (pdd)
2932 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2933 }
2934 
2935 static bool
2936 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2937 {
2938 	unsigned long requested = VM_READ;
2939 
2940 	if (write_fault)
2941 		requested |= VM_WRITE;
2942 
2943 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2944 		vma->vm_flags);
2945 	return (vma->vm_flags & requested) == requested;
2946 }
2947 
2948 int
2949 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2950 			uint32_t vmid, uint32_t node_id,
2951 			uint64_t addr, uint64_t ts, bool write_fault)
2952 {
2953 	unsigned long start, last, size;
2954 	struct mm_struct *mm = NULL;
2955 	struct svm_range_list *svms;
2956 	struct svm_range *prange;
2957 	struct kfd_process *p;
2958 	ktime_t timestamp = ktime_get_boottime();
2959 	struct kfd_node *node;
2960 	int32_t best_loc;
2961 	int32_t gpuid, gpuidx = MAX_GPU_INSTANCE;
2962 	bool write_locked = false;
2963 	struct vm_area_struct *vma;
2964 	bool migration = false;
2965 	int r = 0;
2966 
2967 	if (!KFD_IS_SVM_API_SUPPORTED(adev)) {
2968 		pr_debug("device does not support SVM\n");
2969 		return -EFAULT;
2970 	}
2971 
2972 	p = kfd_lookup_process_by_pasid(pasid);
2973 	if (!p) {
2974 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2975 		return 0;
2976 	}
2977 	svms = &p->svms;
2978 
2979 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2980 
2981 	if (atomic_read(&svms->drain_pagefaults)) {
2982 		pr_debug("page fault handling disabled, drop fault 0x%llx\n", addr);
2983 		r = 0;
2984 		goto out;
2985 	}
2986 
2987 	node = kfd_node_by_irq_ids(adev, node_id, vmid);
2988 	if (!node) {
2989 		pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id,
2990 			 vmid);
2991 		r = -EFAULT;
2992 		goto out;
2993 	}
2994 
2995 	if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) {
2996 		pr_debug("failed to get gpuid/gpuidex for node_id: %d\n", node_id);
2997 		r = -EFAULT;
2998 		goto out;
2999 	}
3000 
3001 	/* check if this page fault time stamp is before svms->checkpoint_ts */
3002 	if (svms->checkpoint_ts[gpuidx] != 0) {
3003 		if (amdgpu_ih_ts_after(ts,  svms->checkpoint_ts[gpuidx])) {
3004 			pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
3005 			r = 0;
3006 			goto out;
3007 		} else
3008 			/* ts is after svms->checkpoint_ts now, reset svms->checkpoint_ts
3009 			 * to zero to avoid following ts wrap around give wrong comparing
3010 			 */
3011 			svms->checkpoint_ts[gpuidx] = 0;
3012 	}
3013 
3014 	if (!p->xnack_enabled) {
3015 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
3016 		r = -EFAULT;
3017 		goto out;
3018 	}
3019 
3020 	/* p->lead_thread is available as kfd_process_wq_release flush the work
3021 	 * before releasing task ref.
3022 	 */
3023 	mm = get_task_mm(p->lead_thread);
3024 	if (!mm) {
3025 		pr_debug("svms 0x%p failed to get mm\n", svms);
3026 		r = 0;
3027 		goto out;
3028 	}
3029 
3030 	mmap_read_lock(mm);
3031 retry_write_locked:
3032 	mutex_lock(&svms->lock);
3033 	prange = svm_range_from_addr(svms, addr, NULL);
3034 	if (!prange) {
3035 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
3036 			 svms, addr);
3037 		if (!write_locked) {
3038 			/* Need the write lock to create new range with MMU notifier.
3039 			 * Also flush pending deferred work to make sure the interval
3040 			 * tree is up to date before we add a new range
3041 			 */
3042 			mutex_unlock(&svms->lock);
3043 			mmap_read_unlock(mm);
3044 			mmap_write_lock(mm);
3045 			write_locked = true;
3046 			goto retry_write_locked;
3047 		}
3048 		prange = svm_range_create_unregistered_range(node, p, mm, addr);
3049 		if (!prange) {
3050 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
3051 				 svms, addr);
3052 			mmap_write_downgrade(mm);
3053 			r = -EFAULT;
3054 			goto out_unlock_svms;
3055 		}
3056 	}
3057 	if (write_locked)
3058 		mmap_write_downgrade(mm);
3059 
3060 	mutex_lock(&prange->migrate_mutex);
3061 
3062 	if (svm_range_skip_recover(prange)) {
3063 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3064 		r = 0;
3065 		goto out_unlock_range;
3066 	}
3067 
3068 	/* skip duplicate vm fault on different pages of same range */
3069 	if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
3070 				AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
3071 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
3072 			 svms, prange->start, prange->last);
3073 		r = 0;
3074 		goto out_unlock_range;
3075 	}
3076 
3077 	/* __do_munmap removed VMA, return success as we are handling stale
3078 	 * retry fault.
3079 	 */
3080 	vma = vma_lookup(mm, addr << PAGE_SHIFT);
3081 	if (!vma) {
3082 		pr_debug("address 0x%llx VMA is removed\n", addr);
3083 		r = 0;
3084 		goto out_unlock_range;
3085 	}
3086 
3087 	if (!svm_fault_allowed(vma, write_fault)) {
3088 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
3089 			write_fault ? "write" : "read");
3090 		r = -EPERM;
3091 		goto out_unlock_range;
3092 	}
3093 
3094 	best_loc = svm_range_best_restore_location(prange, node, &gpuidx);
3095 	if (best_loc == -1) {
3096 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
3097 			 svms, prange->start, prange->last);
3098 		r = -EACCES;
3099 		goto out_unlock_range;
3100 	}
3101 
3102 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
3103 		 svms, prange->start, prange->last, best_loc,
3104 		 prange->actual_loc);
3105 
3106 	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
3107 				       write_fault, timestamp);
3108 
3109 	/* Align migration range start and size to granularity size */
3110 	size = 1UL << prange->granularity;
3111 	start = max_t(unsigned long, ALIGN_DOWN(addr, size), prange->start);
3112 	last = min_t(unsigned long, ALIGN(addr + 1, size) - 1, prange->last);
3113 	if (prange->actual_loc != 0 || best_loc != 0) {
3114 		if (best_loc) {
3115 			r = svm_migrate_to_vram(prange, best_loc, start, last,
3116 					mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
3117 			if (r) {
3118 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
3119 					 r, addr);
3120 				/* Fallback to system memory if migration to
3121 				 * VRAM failed
3122 				 */
3123 				if (prange->actual_loc && prange->actual_loc != best_loc)
3124 					r = svm_migrate_vram_to_ram(prange, mm, start, last,
3125 						KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
3126 				else
3127 					r = 0;
3128 			}
3129 		} else {
3130 			r = svm_migrate_vram_to_ram(prange, mm, start, last,
3131 					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
3132 		}
3133 		if (r) {
3134 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
3135 				 r, svms, start, last);
3136 			goto out_migrate_fail;
3137 		} else {
3138 			migration = true;
3139 		}
3140 	}
3141 
3142 	r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
3143 				       false, false);
3144 	if (r)
3145 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
3146 			 r, svms, start, last);
3147 
3148 out_migrate_fail:
3149 	kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
3150 				     migration);
3151 
3152 out_unlock_range:
3153 	mutex_unlock(&prange->migrate_mutex);
3154 out_unlock_svms:
3155 	mutex_unlock(&svms->lock);
3156 	mmap_read_unlock(mm);
3157 
3158 	svm_range_count_fault(node, p, gpuidx);
3159 
3160 	mmput(mm);
3161 out:
3162 	kfd_unref_process(p);
3163 
3164 	if (r == -EAGAIN) {
3165 		pr_debug("recover vm fault later\n");
3166 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3167 		r = 0;
3168 	}
3169 	return r;
3170 }
3171 
3172 int
3173 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
3174 {
3175 	struct svm_range *prange, *pchild;
3176 	uint64_t reserved_size = 0;
3177 	uint64_t size;
3178 	int r = 0;
3179 
3180 	pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
3181 
3182 	mutex_lock(&p->svms.lock);
3183 
3184 	list_for_each_entry(prange, &p->svms.list, list) {
3185 		svm_range_lock(prange);
3186 		list_for_each_entry(pchild, &prange->child_list, child_list) {
3187 			size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
3188 			if (xnack_enabled) {
3189 				amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3190 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3191 			} else {
3192 				r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3193 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3194 				if (r)
3195 					goto out_unlock;
3196 				reserved_size += size;
3197 			}
3198 		}
3199 
3200 		size = (prange->last - prange->start + 1) << PAGE_SHIFT;
3201 		if (xnack_enabled) {
3202 			amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3203 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3204 		} else {
3205 			r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3206 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3207 			if (r)
3208 				goto out_unlock;
3209 			reserved_size += size;
3210 		}
3211 out_unlock:
3212 		svm_range_unlock(prange);
3213 		if (r)
3214 			break;
3215 	}
3216 
3217 	if (r)
3218 		amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3219 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3220 	else
3221 		/* Change xnack mode must be inside svms lock, to avoid race with
3222 		 * svm_range_deferred_list_work unreserve memory in parallel.
3223 		 */
3224 		p->xnack_enabled = xnack_enabled;
3225 
3226 	mutex_unlock(&p->svms.lock);
3227 	return r;
3228 }
3229 
3230 void svm_range_list_fini(struct kfd_process *p)
3231 {
3232 	struct svm_range *prange;
3233 	struct svm_range *next;
3234 
3235 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3236 
3237 	cancel_delayed_work_sync(&p->svms.restore_work);
3238 
3239 	/* Ensure list work is finished before process is destroyed */
3240 	flush_work(&p->svms.deferred_list_work);
3241 
3242 	/*
3243 	 * Ensure no retry fault comes in afterwards, as page fault handler will
3244 	 * not find kfd process and take mm lock to recover fault.
3245 	 * stop kfd page fault handing, then wait pending page faults got drained
3246 	 */
3247 	atomic_set(&p->svms.drain_pagefaults, 1);
3248 	svm_range_drain_retry_fault(&p->svms);
3249 
3250 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3251 		svm_range_unlink(prange);
3252 		svm_range_remove_notifier(prange);
3253 		svm_range_free(prange, true);
3254 	}
3255 
3256 	mutex_destroy(&p->svms.lock);
3257 
3258 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3259 }
3260 
3261 int svm_range_list_init(struct kfd_process *p)
3262 {
3263 	struct svm_range_list *svms = &p->svms;
3264 	int i;
3265 
3266 	svms->objects = RB_ROOT_CACHED;
3267 	mutex_init(&svms->lock);
3268 	INIT_LIST_HEAD(&svms->list);
3269 	atomic_set(&svms->evicted_ranges, 0);
3270 	atomic_set(&svms->drain_pagefaults, 0);
3271 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3272 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3273 	INIT_LIST_HEAD(&svms->deferred_range_list);
3274 	INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3275 	spin_lock_init(&svms->deferred_list_lock);
3276 
3277 	for (i = 0; i < p->n_pdds; i++)
3278 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev))
3279 			bitmap_set(svms->bitmap_supported, i, 1);
3280 
3281 	 /* Value of default granularity cannot exceed 0x1B, the
3282 	  * number of pages supported by a 4-level paging table
3283 	  */
3284 	svms->default_granularity = min_t(u8, amdgpu_svm_default_granularity, 0x1B);
3285 	pr_debug("Default SVM Granularity to use: %d\n", svms->default_granularity);
3286 
3287 	return 0;
3288 }
3289 
3290 /**
3291  * svm_range_check_vm - check if virtual address range mapped already
3292  * @p: current kfd_process
3293  * @start: range start address, in pages
3294  * @last: range last address, in pages
3295  * @bo_s: mapping start address in pages if address range already mapped
3296  * @bo_l: mapping last address in pages if address range already mapped
3297  *
3298  * The purpose is to avoid virtual address ranges already allocated by
3299  * kfd_ioctl_alloc_memory_of_gpu ioctl.
3300  * It looks for each pdd in the kfd_process.
3301  *
3302  * Context: Process context
3303  *
3304  * Return 0 - OK, if the range is not mapped.
3305  * Otherwise error code:
3306  * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3307  * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3308  * a signal. Release all buffer reservations and return to user-space.
3309  */
3310 static int
3311 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3312 		   uint64_t *bo_s, uint64_t *bo_l)
3313 {
3314 	struct amdgpu_bo_va_mapping *mapping;
3315 	struct interval_tree_node *node;
3316 	uint32_t i;
3317 	int r;
3318 
3319 	for (i = 0; i < p->n_pdds; i++) {
3320 		struct amdgpu_vm *vm;
3321 
3322 		if (!p->pdds[i]->drm_priv)
3323 			continue;
3324 
3325 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3326 		r = amdgpu_bo_reserve(vm->root.bo, false);
3327 		if (r)
3328 			return r;
3329 
3330 		node = interval_tree_iter_first(&vm->va, start, last);
3331 		if (node) {
3332 			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3333 				 start, last);
3334 			mapping = container_of((struct rb_node *)node,
3335 					       struct amdgpu_bo_va_mapping, rb);
3336 			if (bo_s && bo_l) {
3337 				*bo_s = mapping->start;
3338 				*bo_l = mapping->last;
3339 			}
3340 			amdgpu_bo_unreserve(vm->root.bo);
3341 			return -EADDRINUSE;
3342 		}
3343 		amdgpu_bo_unreserve(vm->root.bo);
3344 	}
3345 
3346 	return 0;
3347 }
3348 
3349 /**
3350  * svm_range_is_valid - check if virtual address range is valid
3351  * @p: current kfd_process
3352  * @start: range start address, in pages
3353  * @size: range size, in pages
3354  *
3355  * Valid virtual address range means it belongs to one or more VMAs
3356  *
3357  * Context: Process context
3358  *
3359  * Return:
3360  *  0 - OK, otherwise error code
3361  */
3362 static int
3363 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3364 {
3365 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3366 	struct vm_area_struct *vma;
3367 	unsigned long end;
3368 	unsigned long start_unchg = start;
3369 
3370 	start <<= PAGE_SHIFT;
3371 	end = start + (size << PAGE_SHIFT);
3372 	do {
3373 		vma = vma_lookup(p->mm, start);
3374 		if (!vma || (vma->vm_flags & device_vma))
3375 			return -EFAULT;
3376 		start = min(end, vma->vm_end);
3377 	} while (start < end);
3378 
3379 	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3380 				  NULL);
3381 }
3382 
3383 /**
3384  * svm_range_best_prefetch_location - decide the best prefetch location
3385  * @prange: svm range structure
3386  *
3387  * For xnack off:
3388  * If range map to single GPU, the best prefetch location is prefetch_loc, which
3389  * can be CPU or GPU.
3390  *
3391  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3392  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3393  * the best prefetch location is always CPU, because GPU can not have coherent
3394  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3395  *
3396  * For xnack on:
3397  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3398  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3399  *
3400  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3401  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3402  * prefetch location is always CPU.
3403  *
3404  * Context: Process context
3405  *
3406  * Return:
3407  * 0 for CPU or GPU id
3408  */
3409 static uint32_t
3410 svm_range_best_prefetch_location(struct svm_range *prange)
3411 {
3412 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3413 	uint32_t best_loc = prange->prefetch_loc;
3414 	struct kfd_process_device *pdd;
3415 	struct kfd_node *bo_node;
3416 	struct kfd_process *p;
3417 	uint32_t gpuidx;
3418 
3419 	p = container_of(prange->svms, struct kfd_process, svms);
3420 
3421 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3422 		goto out;
3423 
3424 	bo_node = svm_range_get_node_by_id(prange, best_loc);
3425 	if (!bo_node) {
3426 		WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc);
3427 		best_loc = 0;
3428 		goto out;
3429 	}
3430 
3431 	if (bo_node->adev->flags & AMD_IS_APU) {
3432 		best_loc = 0;
3433 		goto out;
3434 	}
3435 
3436 	if (p->xnack_enabled)
3437 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3438 	else
3439 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3440 			  MAX_GPU_INSTANCE);
3441 
3442 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3443 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3444 		if (!pdd) {
3445 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3446 			continue;
3447 		}
3448 
3449 		if (pdd->dev->adev == bo_node->adev)
3450 			continue;
3451 
3452 		if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) {
3453 			best_loc = 0;
3454 			break;
3455 		}
3456 	}
3457 
3458 out:
3459 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3460 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3461 		 best_loc);
3462 
3463 	return best_loc;
3464 }
3465 
3466 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3467  * @mm: current process mm_struct
3468  * @prange: svm range structure
3469  * @migrated: output, true if migration is triggered
3470  *
3471  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3472  * from ram to vram.
3473  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3474  * from vram to ram.
3475  *
3476  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3477  * and restore work:
3478  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3479  *    stops all queues, schedule restore work
3480  * 2. svm_range_restore_work wait for migration is done by
3481  *    a. svm_range_validate_vram takes prange->migrate_mutex
3482  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3483  * 3. restore work update mappings of GPU, resume all queues.
3484  *
3485  * Context: Process context
3486  *
3487  * Return:
3488  * 0 - OK, otherwise - error code of migration
3489  */
3490 static int
3491 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3492 			    bool *migrated)
3493 {
3494 	uint32_t best_loc;
3495 	int r = 0;
3496 
3497 	*migrated = false;
3498 	best_loc = svm_range_best_prefetch_location(prange);
3499 
3500 	/* when best_loc is a gpu node and same as prange->actual_loc
3501 	 * we still need do migration as prange->actual_loc !=0 does
3502 	 * not mean all pages in prange are vram. hmm migrate will pick
3503 	 * up right pages during migration.
3504 	 */
3505 	if ((best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) ||
3506 	    (best_loc == 0 && prange->actual_loc == 0))
3507 		return 0;
3508 
3509 	if (!best_loc) {
3510 		r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last,
3511 					KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3512 		*migrated = !r;
3513 		return r;
3514 	}
3515 
3516 	r = svm_migrate_to_vram(prange, best_loc, prange->start, prange->last,
3517 				mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3518 	*migrated = !r;
3519 
3520 	return 0;
3521 }
3522 
3523 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3524 {
3525 	/* Dereferencing fence->svm_bo is safe here because the fence hasn't
3526 	 * signaled yet and we're under the protection of the fence->lock.
3527 	 * After the fence is signaled in svm_range_bo_release, we cannot get
3528 	 * here any more.
3529 	 *
3530 	 * Reference is dropped in svm_range_evict_svm_bo_worker.
3531 	 */
3532 	if (svm_bo_ref_unless_zero(fence->svm_bo)) {
3533 		WRITE_ONCE(fence->svm_bo->evicting, 1);
3534 		schedule_work(&fence->svm_bo->eviction_work);
3535 	}
3536 
3537 	return 0;
3538 }
3539 
3540 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3541 {
3542 	struct svm_range_bo *svm_bo;
3543 	struct mm_struct *mm;
3544 	int r = 0;
3545 
3546 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3547 
3548 	if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3549 		mm = svm_bo->eviction_fence->mm;
3550 	} else {
3551 		svm_range_bo_unref(svm_bo);
3552 		return;
3553 	}
3554 
3555 	mmap_read_lock(mm);
3556 	spin_lock(&svm_bo->list_lock);
3557 	while (!list_empty(&svm_bo->range_list) && !r) {
3558 		struct svm_range *prange =
3559 				list_first_entry(&svm_bo->range_list,
3560 						struct svm_range, svm_bo_list);
3561 		int retries = 3;
3562 
3563 		list_del_init(&prange->svm_bo_list);
3564 		spin_unlock(&svm_bo->list_lock);
3565 
3566 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3567 			 prange->start, prange->last);
3568 
3569 		mutex_lock(&prange->migrate_mutex);
3570 		do {
3571 			/* migrate all vram pages in this prange to sys ram
3572 			 * after that prange->actual_loc should be zero
3573 			 */
3574 			r = svm_migrate_vram_to_ram(prange, mm,
3575 					prange->start, prange->last,
3576 					KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3577 		} while (!r && prange->actual_loc && --retries);
3578 
3579 		if (!r && prange->actual_loc)
3580 			pr_info_once("Migration failed during eviction");
3581 
3582 		if (!prange->actual_loc) {
3583 			mutex_lock(&prange->lock);
3584 			prange->svm_bo = NULL;
3585 			mutex_unlock(&prange->lock);
3586 		}
3587 		mutex_unlock(&prange->migrate_mutex);
3588 
3589 		spin_lock(&svm_bo->list_lock);
3590 	}
3591 	spin_unlock(&svm_bo->list_lock);
3592 	mmap_read_unlock(mm);
3593 	mmput(mm);
3594 
3595 	dma_fence_signal(&svm_bo->eviction_fence->base);
3596 
3597 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3598 	 * has been called in svm_migrate_vram_to_ram
3599 	 */
3600 	WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3601 	svm_range_bo_unref(svm_bo);
3602 }
3603 
3604 static int
3605 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3606 		   uint64_t start, uint64_t size, uint32_t nattr,
3607 		   struct kfd_ioctl_svm_attribute *attrs)
3608 {
3609 	struct amdkfd_process_info *process_info = p->kgd_process_info;
3610 	struct list_head update_list;
3611 	struct list_head insert_list;
3612 	struct list_head remove_list;
3613 	struct list_head remap_list;
3614 	struct svm_range_list *svms;
3615 	struct svm_range *prange;
3616 	struct svm_range *next;
3617 	bool update_mapping = false;
3618 	bool flush_tlb;
3619 	int r, ret = 0;
3620 
3621 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3622 		 p->pasid, &p->svms, start, start + size - 1, size);
3623 
3624 	r = svm_range_check_attr(p, nattr, attrs);
3625 	if (r)
3626 		return r;
3627 
3628 	svms = &p->svms;
3629 
3630 	mutex_lock(&process_info->lock);
3631 
3632 	svm_range_list_lock_and_flush_work(svms, mm);
3633 
3634 	r = svm_range_is_valid(p, start, size);
3635 	if (r) {
3636 		pr_debug("invalid range r=%d\n", r);
3637 		mmap_write_unlock(mm);
3638 		goto out;
3639 	}
3640 
3641 	mutex_lock(&svms->lock);
3642 
3643 	/* Add new range and split existing ranges as needed */
3644 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3645 			  &insert_list, &remove_list, &remap_list);
3646 	if (r) {
3647 		mutex_unlock(&svms->lock);
3648 		mmap_write_unlock(mm);
3649 		goto out;
3650 	}
3651 	/* Apply changes as a transaction */
3652 	list_for_each_entry_safe(prange, next, &insert_list, list) {
3653 		svm_range_add_to_svms(prange);
3654 		svm_range_add_notifier_locked(mm, prange);
3655 	}
3656 	list_for_each_entry(prange, &update_list, update_list) {
3657 		svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3658 		/* TODO: unmap ranges from GPU that lost access */
3659 	}
3660 	list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3661 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3662 			 prange->svms, prange, prange->start,
3663 			 prange->last);
3664 		svm_range_unlink(prange);
3665 		svm_range_remove_notifier(prange);
3666 		svm_range_free(prange, false);
3667 	}
3668 
3669 	mmap_write_downgrade(mm);
3670 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3671 	 * this fails we may be left with partially completed actions. There
3672 	 * is no clean way of rolling back to the previous state in such a
3673 	 * case because the rollback wouldn't be guaranteed to work either.
3674 	 */
3675 	list_for_each_entry(prange, &update_list, update_list) {
3676 		bool migrated;
3677 
3678 		mutex_lock(&prange->migrate_mutex);
3679 
3680 		r = svm_range_trigger_migration(mm, prange, &migrated);
3681 		if (r)
3682 			goto out_unlock_range;
3683 
3684 		if (migrated && (!p->xnack_enabled ||
3685 		    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3686 		    prange->mapped_to_gpu) {
3687 			pr_debug("restore_work will update mappings of GPUs\n");
3688 			mutex_unlock(&prange->migrate_mutex);
3689 			continue;
3690 		}
3691 
3692 		if (!migrated && !update_mapping) {
3693 			mutex_unlock(&prange->migrate_mutex);
3694 			continue;
3695 		}
3696 
3697 		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3698 
3699 		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
3700 					       MAX_GPU_INSTANCE, true, true, flush_tlb);
3701 		if (r)
3702 			pr_debug("failed %d to map svm range\n", r);
3703 
3704 out_unlock_range:
3705 		mutex_unlock(&prange->migrate_mutex);
3706 		if (r)
3707 			ret = r;
3708 	}
3709 
3710 	list_for_each_entry(prange, &remap_list, update_list) {
3711 		pr_debug("Remapping prange 0x%p [0x%lx 0x%lx]\n",
3712 			 prange, prange->start, prange->last);
3713 		mutex_lock(&prange->migrate_mutex);
3714 		r = svm_range_validate_and_map(mm,  prange->start, prange->last, prange,
3715 					       MAX_GPU_INSTANCE, true, true, prange->mapped_to_gpu);
3716 		if (r)
3717 			pr_debug("failed %d on remap svm range\n", r);
3718 		mutex_unlock(&prange->migrate_mutex);
3719 		if (r)
3720 			ret = r;
3721 	}
3722 
3723 	dynamic_svm_range_dump(svms);
3724 
3725 	mutex_unlock(&svms->lock);
3726 	mmap_read_unlock(mm);
3727 out:
3728 	mutex_unlock(&process_info->lock);
3729 
3730 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3731 		 &p->svms, start, start + size - 1, r);
3732 
3733 	return ret ? ret : r;
3734 }
3735 
3736 static int
3737 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3738 		   uint64_t start, uint64_t size, uint32_t nattr,
3739 		   struct kfd_ioctl_svm_attribute *attrs)
3740 {
3741 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3742 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3743 	bool get_preferred_loc = false;
3744 	bool get_prefetch_loc = false;
3745 	bool get_granularity = false;
3746 	bool get_accessible = false;
3747 	bool get_flags = false;
3748 	uint64_t last = start + size - 1UL;
3749 	uint8_t granularity = 0xff;
3750 	struct interval_tree_node *node;
3751 	struct svm_range_list *svms;
3752 	struct svm_range *prange;
3753 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3754 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3755 	uint32_t flags_and = 0xffffffff;
3756 	uint32_t flags_or = 0;
3757 	int gpuidx;
3758 	uint32_t i;
3759 	int r = 0;
3760 
3761 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3762 		 start + size - 1, nattr);
3763 
3764 	/* Flush pending deferred work to avoid racing with deferred actions from
3765 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3766 	 * can still race with get_attr because we don't hold the mmap lock. But that
3767 	 * would be a race condition in the application anyway, and undefined
3768 	 * behaviour is acceptable in that case.
3769 	 */
3770 	flush_work(&p->svms.deferred_list_work);
3771 
3772 	mmap_read_lock(mm);
3773 	r = svm_range_is_valid(p, start, size);
3774 	mmap_read_unlock(mm);
3775 	if (r) {
3776 		pr_debug("invalid range r=%d\n", r);
3777 		return r;
3778 	}
3779 
3780 	for (i = 0; i < nattr; i++) {
3781 		switch (attrs[i].type) {
3782 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3783 			get_preferred_loc = true;
3784 			break;
3785 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3786 			get_prefetch_loc = true;
3787 			break;
3788 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3789 			get_accessible = true;
3790 			break;
3791 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3792 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3793 			get_flags = true;
3794 			break;
3795 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3796 			get_granularity = true;
3797 			break;
3798 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3799 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3800 			fallthrough;
3801 		default:
3802 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3803 			return -EINVAL;
3804 		}
3805 	}
3806 
3807 	svms = &p->svms;
3808 
3809 	mutex_lock(&svms->lock);
3810 
3811 	node = interval_tree_iter_first(&svms->objects, start, last);
3812 	if (!node) {
3813 		pr_debug("range attrs not found return default values\n");
3814 		svm_range_set_default_attributes(svms, &location, &prefetch_loc,
3815 						 &granularity, &flags_and);
3816 		flags_or = flags_and;
3817 		if (p->xnack_enabled)
3818 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3819 				    MAX_GPU_INSTANCE);
3820 		else
3821 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3822 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3823 		goto fill_values;
3824 	}
3825 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3826 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3827 
3828 	while (node) {
3829 		struct interval_tree_node *next;
3830 
3831 		prange = container_of(node, struct svm_range, it_node);
3832 		next = interval_tree_iter_next(node, start, last);
3833 
3834 		if (get_preferred_loc) {
3835 			if (prange->preferred_loc ==
3836 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3837 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3838 			     location != prange->preferred_loc)) {
3839 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3840 				get_preferred_loc = false;
3841 			} else {
3842 				location = prange->preferred_loc;
3843 			}
3844 		}
3845 		if (get_prefetch_loc) {
3846 			if (prange->prefetch_loc ==
3847 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3848 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3849 			     prefetch_loc != prange->prefetch_loc)) {
3850 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3851 				get_prefetch_loc = false;
3852 			} else {
3853 				prefetch_loc = prange->prefetch_loc;
3854 			}
3855 		}
3856 		if (get_accessible) {
3857 			bitmap_and(bitmap_access, bitmap_access,
3858 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3859 			bitmap_and(bitmap_aip, bitmap_aip,
3860 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3861 		}
3862 		if (get_flags) {
3863 			flags_and &= prange->flags;
3864 			flags_or |= prange->flags;
3865 		}
3866 
3867 		if (get_granularity && prange->granularity < granularity)
3868 			granularity = prange->granularity;
3869 
3870 		node = next;
3871 	}
3872 fill_values:
3873 	mutex_unlock(&svms->lock);
3874 
3875 	for (i = 0; i < nattr; i++) {
3876 		switch (attrs[i].type) {
3877 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3878 			attrs[i].value = location;
3879 			break;
3880 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3881 			attrs[i].value = prefetch_loc;
3882 			break;
3883 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3884 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3885 							       attrs[i].value);
3886 			if (gpuidx < 0) {
3887 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3888 				return -EINVAL;
3889 			}
3890 			if (test_bit(gpuidx, bitmap_access))
3891 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3892 			else if (test_bit(gpuidx, bitmap_aip))
3893 				attrs[i].type =
3894 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3895 			else
3896 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3897 			break;
3898 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3899 			attrs[i].value = flags_and;
3900 			break;
3901 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3902 			attrs[i].value = ~flags_or;
3903 			break;
3904 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3905 			attrs[i].value = (uint32_t)granularity;
3906 			break;
3907 		}
3908 	}
3909 
3910 	return 0;
3911 }
3912 
3913 int kfd_criu_resume_svm(struct kfd_process *p)
3914 {
3915 	struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3916 	int nattr_common = 4, nattr_accessibility = 1;
3917 	struct criu_svm_metadata *criu_svm_md = NULL;
3918 	struct svm_range_list *svms = &p->svms;
3919 	struct criu_svm_metadata *next = NULL;
3920 	uint32_t set_flags = 0xffffffff;
3921 	int i, j, num_attrs, ret = 0;
3922 	uint64_t set_attr_size;
3923 	struct mm_struct *mm;
3924 
3925 	if (list_empty(&svms->criu_svm_metadata_list)) {
3926 		pr_debug("No SVM data from CRIU restore stage 2\n");
3927 		return ret;
3928 	}
3929 
3930 	mm = get_task_mm(p->lead_thread);
3931 	if (!mm) {
3932 		pr_err("failed to get mm for the target process\n");
3933 		return -ESRCH;
3934 	}
3935 
3936 	num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3937 
3938 	i = j = 0;
3939 	list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3940 		pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3941 			 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3942 
3943 		for (j = 0; j < num_attrs; j++) {
3944 			pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3945 				 i, j, criu_svm_md->data.attrs[j].type,
3946 				 i, j, criu_svm_md->data.attrs[j].value);
3947 			switch (criu_svm_md->data.attrs[j].type) {
3948 			/* During Checkpoint operation, the query for
3949 			 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3950 			 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3951 			 * not used by the range which was checkpointed. Care
3952 			 * must be taken to not restore with an invalid value
3953 			 * otherwise the gpuidx value will be invalid and
3954 			 * set_attr would eventually fail so just replace those
3955 			 * with another dummy attribute such as
3956 			 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3957 			 */
3958 			case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3959 				if (criu_svm_md->data.attrs[j].value ==
3960 				    KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3961 					criu_svm_md->data.attrs[j].type =
3962 						KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3963 					criu_svm_md->data.attrs[j].value = 0;
3964 				}
3965 				break;
3966 			case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3967 				set_flags = criu_svm_md->data.attrs[j].value;
3968 				break;
3969 			default:
3970 				break;
3971 			}
3972 		}
3973 
3974 		/* CLR_FLAGS is not available via get_attr during checkpoint but
3975 		 * it needs to be inserted before restoring the ranges so
3976 		 * allocate extra space for it before calling set_attr
3977 		 */
3978 		set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3979 						(num_attrs + 1);
3980 		set_attr_new = krealloc(set_attr, set_attr_size,
3981 					    GFP_KERNEL);
3982 		if (!set_attr_new) {
3983 			ret = -ENOMEM;
3984 			goto exit;
3985 		}
3986 		set_attr = set_attr_new;
3987 
3988 		memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3989 					sizeof(struct kfd_ioctl_svm_attribute));
3990 		set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3991 		set_attr[num_attrs].value = ~set_flags;
3992 
3993 		ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3994 					 criu_svm_md->data.size, num_attrs + 1,
3995 					 set_attr);
3996 		if (ret) {
3997 			pr_err("CRIU: failed to set range attributes\n");
3998 			goto exit;
3999 		}
4000 
4001 		i++;
4002 	}
4003 exit:
4004 	kfree(set_attr);
4005 	list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
4006 		pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
4007 						criu_svm_md->data.start_addr);
4008 		kfree(criu_svm_md);
4009 	}
4010 
4011 	mmput(mm);
4012 	return ret;
4013 
4014 }
4015 
4016 int kfd_criu_restore_svm(struct kfd_process *p,
4017 			 uint8_t __user *user_priv_ptr,
4018 			 uint64_t *priv_data_offset,
4019 			 uint64_t max_priv_data_size)
4020 {
4021 	uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
4022 	int nattr_common = 4, nattr_accessibility = 1;
4023 	struct criu_svm_metadata *criu_svm_md = NULL;
4024 	struct svm_range_list *svms = &p->svms;
4025 	uint32_t num_devices;
4026 	int ret = 0;
4027 
4028 	num_devices = p->n_pdds;
4029 	/* Handle one SVM range object at a time, also the number of gpus are
4030 	 * assumed to be same on the restore node, checking must be done while
4031 	 * evaluating the topology earlier
4032 	 */
4033 
4034 	svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
4035 		(nattr_common + nattr_accessibility * num_devices);
4036 	svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
4037 
4038 	svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
4039 								svm_attrs_size;
4040 
4041 	criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
4042 	if (!criu_svm_md) {
4043 		pr_err("failed to allocate memory to store svm metadata\n");
4044 		return -ENOMEM;
4045 	}
4046 	if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
4047 		ret = -EINVAL;
4048 		goto exit;
4049 	}
4050 
4051 	ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
4052 			     svm_priv_data_size);
4053 	if (ret) {
4054 		ret = -EFAULT;
4055 		goto exit;
4056 	}
4057 	*priv_data_offset += svm_priv_data_size;
4058 
4059 	list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
4060 
4061 	return 0;
4062 
4063 
4064 exit:
4065 	kfree(criu_svm_md);
4066 	return ret;
4067 }
4068 
4069 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
4070 		       uint64_t *svm_priv_data_size)
4071 {
4072 	uint64_t total_size, accessibility_size, common_attr_size;
4073 	int nattr_common = 4, nattr_accessibility = 1;
4074 	int num_devices = p->n_pdds;
4075 	struct svm_range_list *svms;
4076 	struct svm_range *prange;
4077 	uint32_t count = 0;
4078 
4079 	*svm_priv_data_size = 0;
4080 
4081 	svms = &p->svms;
4082 	if (!svms)
4083 		return -EINVAL;
4084 
4085 	mutex_lock(&svms->lock);
4086 	list_for_each_entry(prange, &svms->list, list) {
4087 		pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
4088 			 prange, prange->start, prange->npages,
4089 			 prange->start + prange->npages - 1);
4090 		count++;
4091 	}
4092 	mutex_unlock(&svms->lock);
4093 
4094 	*num_svm_ranges = count;
4095 	/* Only the accessbility attributes need to be queried for all the gpus
4096 	 * individually, remaining ones are spanned across the entire process
4097 	 * regardless of the various gpu nodes. Of the remaining attributes,
4098 	 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
4099 	 *
4100 	 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
4101 	 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
4102 	 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
4103 	 * KFD_IOCTL_SVM_ATTR_GRANULARITY
4104 	 *
4105 	 * ** ACCESSBILITY ATTRIBUTES **
4106 	 * (Considered as one, type is altered during query, value is gpuid)
4107 	 * KFD_IOCTL_SVM_ATTR_ACCESS
4108 	 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
4109 	 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
4110 	 */
4111 	if (*num_svm_ranges > 0) {
4112 		common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4113 			nattr_common;
4114 		accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
4115 			nattr_accessibility * num_devices;
4116 
4117 		total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
4118 			common_attr_size + accessibility_size;
4119 
4120 		*svm_priv_data_size = *num_svm_ranges * total_size;
4121 	}
4122 
4123 	pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
4124 		 *svm_priv_data_size);
4125 	return 0;
4126 }
4127 
4128 int kfd_criu_checkpoint_svm(struct kfd_process *p,
4129 			    uint8_t __user *user_priv_data,
4130 			    uint64_t *priv_data_offset)
4131 {
4132 	struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
4133 	struct kfd_ioctl_svm_attribute *query_attr = NULL;
4134 	uint64_t svm_priv_data_size, query_attr_size = 0;
4135 	int index, nattr_common = 4, ret = 0;
4136 	struct svm_range_list *svms;
4137 	int num_devices = p->n_pdds;
4138 	struct svm_range *prange;
4139 	struct mm_struct *mm;
4140 
4141 	svms = &p->svms;
4142 	if (!svms)
4143 		return -EINVAL;
4144 
4145 	mm = get_task_mm(p->lead_thread);
4146 	if (!mm) {
4147 		pr_err("failed to get mm for the target process\n");
4148 		return -ESRCH;
4149 	}
4150 
4151 	query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4152 				(nattr_common + num_devices);
4153 
4154 	query_attr = kzalloc(query_attr_size, GFP_KERNEL);
4155 	if (!query_attr) {
4156 		ret = -ENOMEM;
4157 		goto exit;
4158 	}
4159 
4160 	query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
4161 	query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
4162 	query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
4163 	query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
4164 
4165 	for (index = 0; index < num_devices; index++) {
4166 		struct kfd_process_device *pdd = p->pdds[index];
4167 
4168 		query_attr[index + nattr_common].type =
4169 			KFD_IOCTL_SVM_ATTR_ACCESS;
4170 		query_attr[index + nattr_common].value = pdd->user_gpu_id;
4171 	}
4172 
4173 	svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
4174 
4175 	svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
4176 	if (!svm_priv) {
4177 		ret = -ENOMEM;
4178 		goto exit_query;
4179 	}
4180 
4181 	index = 0;
4182 	list_for_each_entry(prange, &svms->list, list) {
4183 
4184 		svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
4185 		svm_priv->start_addr = prange->start;
4186 		svm_priv->size = prange->npages;
4187 		memcpy(&svm_priv->attrs, query_attr, query_attr_size);
4188 		pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
4189 			 prange, prange->start, prange->npages,
4190 			 prange->start + prange->npages - 1,
4191 			 prange->npages * PAGE_SIZE);
4192 
4193 		ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
4194 					 svm_priv->size,
4195 					 (nattr_common + num_devices),
4196 					 svm_priv->attrs);
4197 		if (ret) {
4198 			pr_err("CRIU: failed to obtain range attributes\n");
4199 			goto exit_priv;
4200 		}
4201 
4202 		if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
4203 				 svm_priv_data_size)) {
4204 			pr_err("Failed to copy svm priv to user\n");
4205 			ret = -EFAULT;
4206 			goto exit_priv;
4207 		}
4208 
4209 		*priv_data_offset += svm_priv_data_size;
4210 
4211 	}
4212 
4213 
4214 exit_priv:
4215 	kfree(svm_priv);
4216 exit_query:
4217 	kfree(query_attr);
4218 exit:
4219 	mmput(mm);
4220 	return ret;
4221 }
4222 
4223 int
4224 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
4225 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
4226 {
4227 	struct mm_struct *mm = current->mm;
4228 	int r;
4229 
4230 	start >>= PAGE_SHIFT;
4231 	size >>= PAGE_SHIFT;
4232 
4233 	switch (op) {
4234 	case KFD_IOCTL_SVM_OP_SET_ATTR:
4235 		r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4236 		break;
4237 	case KFD_IOCTL_SVM_OP_GET_ATTR:
4238 		r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);
4239 		break;
4240 	default:
4241 		r = EINVAL;
4242 		break;
4243 	}
4244 
4245 	return r;
4246 }
4247