xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_svm.c (revision 2984f26a)
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 	return 0;
202 }
203 
204 static int
205 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
206 		  unsigned long offset, unsigned long npages,
207 		  unsigned long *hmm_pfns)
208 {
209 	struct kfd_process *p;
210 	uint32_t gpuidx;
211 	int r;
212 
213 	p = container_of(prange->svms, struct kfd_process, svms);
214 
215 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
216 		struct kfd_process_device *pdd;
217 
218 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
219 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
220 		if (!pdd) {
221 			pr_debug("failed to find device idx %d\n", gpuidx);
222 			return -EINVAL;
223 		}
224 
225 		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
226 					  hmm_pfns, gpuidx);
227 		if (r)
228 			break;
229 	}
230 
231 	return r;
232 }
233 
234 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
235 			 unsigned long offset, unsigned long npages)
236 {
237 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
238 	int i;
239 
240 	if (!dma_addr)
241 		return;
242 
243 	for (i = offset; i < offset + npages; i++) {
244 		if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
245 			continue;
246 		pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
247 		dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
248 		dma_addr[i] = 0;
249 	}
250 }
251 
252 void svm_range_free_dma_mappings(struct svm_range *prange, bool unmap_dma)
253 {
254 	struct kfd_process_device *pdd;
255 	dma_addr_t *dma_addr;
256 	struct device *dev;
257 	struct kfd_process *p;
258 	uint32_t gpuidx;
259 
260 	p = container_of(prange->svms, struct kfd_process, svms);
261 
262 	for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
263 		dma_addr = prange->dma_addr[gpuidx];
264 		if (!dma_addr)
265 			continue;
266 
267 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
268 		if (!pdd) {
269 			pr_debug("failed to find device idx %d\n", gpuidx);
270 			continue;
271 		}
272 		dev = &pdd->dev->adev->pdev->dev;
273 		if (unmap_dma)
274 			svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
275 		kvfree(dma_addr);
276 		prange->dma_addr[gpuidx] = NULL;
277 	}
278 }
279 
280 static void svm_range_free(struct svm_range *prange, bool do_unmap)
281 {
282 	uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
283 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
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 	svm_range_free_dma_mappings(prange, do_unmap);
290 
291 	if (do_unmap && !p->xnack_enabled) {
292 		pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size);
293 		amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
294 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
295 	}
296 	mutex_destroy(&prange->lock);
297 	mutex_destroy(&prange->migrate_mutex);
298 	kfree(prange);
299 }
300 
301 static void
302 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
303 				 uint8_t *granularity, uint32_t *flags)
304 {
305 	*location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
306 	*prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
307 	*granularity = 9;
308 	*flags =
309 		KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
310 }
311 
312 static struct
313 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
314 			 uint64_t last, bool update_mem_usage)
315 {
316 	uint64_t size = last - start + 1;
317 	struct svm_range *prange;
318 	struct kfd_process *p;
319 
320 	prange = kzalloc(sizeof(*prange), GFP_KERNEL);
321 	if (!prange)
322 		return NULL;
323 
324 	p = container_of(svms, struct kfd_process, svms);
325 	if (!p->xnack_enabled && update_mem_usage &&
326 	    amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
327 				    KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) {
328 		pr_info("SVM mapping failed, exceeds resident system memory limit\n");
329 		kfree(prange);
330 		return NULL;
331 	}
332 	prange->npages = size;
333 	prange->svms = svms;
334 	prange->start = start;
335 	prange->last = last;
336 	INIT_LIST_HEAD(&prange->list);
337 	INIT_LIST_HEAD(&prange->update_list);
338 	INIT_LIST_HEAD(&prange->svm_bo_list);
339 	INIT_LIST_HEAD(&prange->deferred_list);
340 	INIT_LIST_HEAD(&prange->child_list);
341 	atomic_set(&prange->invalid, 0);
342 	prange->validate_timestamp = 0;
343 	mutex_init(&prange->migrate_mutex);
344 	mutex_init(&prange->lock);
345 
346 	if (p->xnack_enabled)
347 		bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
348 			    MAX_GPU_INSTANCE);
349 
350 	svm_range_set_default_attributes(&prange->preferred_loc,
351 					 &prange->prefetch_loc,
352 					 &prange->granularity, &prange->flags);
353 
354 	pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
355 
356 	return prange;
357 }
358 
359 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
360 {
361 	if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
362 		return false;
363 
364 	return true;
365 }
366 
367 static void svm_range_bo_release(struct kref *kref)
368 {
369 	struct svm_range_bo *svm_bo;
370 
371 	svm_bo = container_of(kref, struct svm_range_bo, kref);
372 	pr_debug("svm_bo 0x%p\n", svm_bo);
373 
374 	spin_lock(&svm_bo->list_lock);
375 	while (!list_empty(&svm_bo->range_list)) {
376 		struct svm_range *prange =
377 				list_first_entry(&svm_bo->range_list,
378 						struct svm_range, svm_bo_list);
379 		/* list_del_init tells a concurrent svm_range_vram_node_new when
380 		 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
381 		 */
382 		list_del_init(&prange->svm_bo_list);
383 		spin_unlock(&svm_bo->list_lock);
384 
385 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
386 			 prange->start, prange->last);
387 		mutex_lock(&prange->lock);
388 		prange->svm_bo = NULL;
389 		mutex_unlock(&prange->lock);
390 
391 		spin_lock(&svm_bo->list_lock);
392 	}
393 	spin_unlock(&svm_bo->list_lock);
394 	if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base))
395 		/* We're not in the eviction worker. Signal the fence. */
396 		dma_fence_signal(&svm_bo->eviction_fence->base);
397 	dma_fence_put(&svm_bo->eviction_fence->base);
398 	amdgpu_bo_unref(&svm_bo->bo);
399 	kfree(svm_bo);
400 }
401 
402 static void svm_range_bo_wq_release(struct work_struct *work)
403 {
404 	struct svm_range_bo *svm_bo;
405 
406 	svm_bo = container_of(work, struct svm_range_bo, release_work);
407 	svm_range_bo_release(&svm_bo->kref);
408 }
409 
410 static void svm_range_bo_release_async(struct kref *kref)
411 {
412 	struct svm_range_bo *svm_bo;
413 
414 	svm_bo = container_of(kref, struct svm_range_bo, kref);
415 	pr_debug("svm_bo 0x%p\n", svm_bo);
416 	INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
417 	schedule_work(&svm_bo->release_work);
418 }
419 
420 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
421 {
422 	kref_put(&svm_bo->kref, svm_range_bo_release_async);
423 }
424 
425 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
426 {
427 	if (svm_bo)
428 		kref_put(&svm_bo->kref, svm_range_bo_release);
429 }
430 
431 static bool
432 svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange)
433 {
434 	mutex_lock(&prange->lock);
435 	if (!prange->svm_bo) {
436 		mutex_unlock(&prange->lock);
437 		return false;
438 	}
439 	if (prange->ttm_res) {
440 		/* We still have a reference, all is well */
441 		mutex_unlock(&prange->lock);
442 		return true;
443 	}
444 	if (svm_bo_ref_unless_zero(prange->svm_bo)) {
445 		/*
446 		 * Migrate from GPU to GPU, remove range from source svm_bo->node
447 		 * range list, and return false to allocate svm_bo from destination
448 		 * node.
449 		 */
450 		if (prange->svm_bo->node != node) {
451 			mutex_unlock(&prange->lock);
452 
453 			spin_lock(&prange->svm_bo->list_lock);
454 			list_del_init(&prange->svm_bo_list);
455 			spin_unlock(&prange->svm_bo->list_lock);
456 
457 			svm_range_bo_unref(prange->svm_bo);
458 			return false;
459 		}
460 		if (READ_ONCE(prange->svm_bo->evicting)) {
461 			struct dma_fence *f;
462 			struct svm_range_bo *svm_bo;
463 			/* The BO is getting evicted,
464 			 * we need to get a new one
465 			 */
466 			mutex_unlock(&prange->lock);
467 			svm_bo = prange->svm_bo;
468 			f = dma_fence_get(&svm_bo->eviction_fence->base);
469 			svm_range_bo_unref(prange->svm_bo);
470 			/* wait for the fence to avoid long spin-loop
471 			 * at list_empty_careful
472 			 */
473 			dma_fence_wait(f, false);
474 			dma_fence_put(f);
475 		} else {
476 			/* The BO was still around and we got
477 			 * a new reference to it
478 			 */
479 			mutex_unlock(&prange->lock);
480 			pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
481 				 prange->svms, prange->start, prange->last);
482 
483 			prange->ttm_res = prange->svm_bo->bo->tbo.resource;
484 			return true;
485 		}
486 
487 	} else {
488 		mutex_unlock(&prange->lock);
489 	}
490 
491 	/* We need a new svm_bo. Spin-loop to wait for concurrent
492 	 * svm_range_bo_release to finish removing this range from
493 	 * its range list and set prange->svm_bo to null. After this,
494 	 * it is safe to reuse the svm_bo pointer and svm_bo_list head.
495 	 */
496 	while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo)
497 		cond_resched();
498 
499 	return false;
500 }
501 
502 static struct svm_range_bo *svm_range_bo_new(void)
503 {
504 	struct svm_range_bo *svm_bo;
505 
506 	svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
507 	if (!svm_bo)
508 		return NULL;
509 
510 	kref_init(&svm_bo->kref);
511 	INIT_LIST_HEAD(&svm_bo->range_list);
512 	spin_lock_init(&svm_bo->list_lock);
513 
514 	return svm_bo;
515 }
516 
517 int
518 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
519 			bool clear)
520 {
521 	struct amdgpu_bo_param bp;
522 	struct svm_range_bo *svm_bo;
523 	struct amdgpu_bo_user *ubo;
524 	struct amdgpu_bo *bo;
525 	struct kfd_process *p;
526 	struct mm_struct *mm;
527 	int r;
528 
529 	p = container_of(prange->svms, struct kfd_process, svms);
530 	pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
531 		 prange->start, prange->last);
532 
533 	if (svm_range_validate_svm_bo(node, prange))
534 		return 0;
535 
536 	svm_bo = svm_range_bo_new();
537 	if (!svm_bo) {
538 		pr_debug("failed to alloc svm bo\n");
539 		return -ENOMEM;
540 	}
541 	mm = get_task_mm(p->lead_thread);
542 	if (!mm) {
543 		pr_debug("failed to get mm\n");
544 		kfree(svm_bo);
545 		return -ESRCH;
546 	}
547 	svm_bo->node = node;
548 	svm_bo->eviction_fence =
549 		amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
550 					   mm,
551 					   svm_bo);
552 	mmput(mm);
553 	INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
554 	svm_bo->evicting = 0;
555 	memset(&bp, 0, sizeof(bp));
556 	bp.size = prange->npages * PAGE_SIZE;
557 	bp.byte_align = PAGE_SIZE;
558 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
559 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
560 	bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
561 	bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
562 	bp.type = ttm_bo_type_device;
563 	bp.resv = NULL;
564 	if (node->xcp)
565 		bp.xcp_id_plus1 = node->xcp->id + 1;
566 
567 	r = amdgpu_bo_create_user(node->adev, &bp, &ubo);
568 	if (r) {
569 		pr_debug("failed %d to create bo\n", r);
570 		goto create_bo_failed;
571 	}
572 	bo = &ubo->bo;
573 
574 	pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n",
575 		 bo->tbo.resource->start << PAGE_SHIFT, bp.size,
576 		 bp.xcp_id_plus1 - 1);
577 
578 	r = amdgpu_bo_reserve(bo, true);
579 	if (r) {
580 		pr_debug("failed %d to reserve bo\n", r);
581 		goto reserve_bo_failed;
582 	}
583 
584 	if (clear) {
585 		r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
586 		if (r) {
587 			pr_debug("failed %d to sync bo\n", r);
588 			amdgpu_bo_unreserve(bo);
589 			goto reserve_bo_failed;
590 		}
591 	}
592 
593 	r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
594 	if (r) {
595 		pr_debug("failed %d to reserve bo\n", r);
596 		amdgpu_bo_unreserve(bo);
597 		goto reserve_bo_failed;
598 	}
599 	amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
600 
601 	amdgpu_bo_unreserve(bo);
602 
603 	svm_bo->bo = bo;
604 	prange->svm_bo = svm_bo;
605 	prange->ttm_res = bo->tbo.resource;
606 	prange->offset = 0;
607 
608 	spin_lock(&svm_bo->list_lock);
609 	list_add(&prange->svm_bo_list, &svm_bo->range_list);
610 	spin_unlock(&svm_bo->list_lock);
611 
612 	return 0;
613 
614 reserve_bo_failed:
615 	amdgpu_bo_unref(&bo);
616 create_bo_failed:
617 	dma_fence_put(&svm_bo->eviction_fence->base);
618 	kfree(svm_bo);
619 	prange->ttm_res = NULL;
620 
621 	return r;
622 }
623 
624 void svm_range_vram_node_free(struct svm_range *prange)
625 {
626 	/* serialize prange->svm_bo unref */
627 	mutex_lock(&prange->lock);
628 	/* prange->svm_bo has not been unref */
629 	if (prange->ttm_res) {
630 		prange->ttm_res = NULL;
631 		mutex_unlock(&prange->lock);
632 		svm_range_bo_unref(prange->svm_bo);
633 	} else
634 		mutex_unlock(&prange->lock);
635 }
636 
637 struct kfd_node *
638 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id)
639 {
640 	struct kfd_process *p;
641 	struct kfd_process_device *pdd;
642 
643 	p = container_of(prange->svms, struct kfd_process, svms);
644 	pdd = kfd_process_device_data_by_id(p, gpu_id);
645 	if (!pdd) {
646 		pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id);
647 		return NULL;
648 	}
649 
650 	return pdd->dev;
651 }
652 
653 struct kfd_process_device *
654 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node)
655 {
656 	struct kfd_process *p;
657 
658 	p = container_of(prange->svms, struct kfd_process, svms);
659 
660 	return kfd_get_process_device_data(node, p);
661 }
662 
663 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
664 {
665 	struct ttm_operation_ctx ctx = { false, false };
666 
667 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
668 
669 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
670 }
671 
672 static int
673 svm_range_check_attr(struct kfd_process *p,
674 		     uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
675 {
676 	uint32_t i;
677 
678 	for (i = 0; i < nattr; i++) {
679 		uint32_t val = attrs[i].value;
680 		int gpuidx = MAX_GPU_INSTANCE;
681 
682 		switch (attrs[i].type) {
683 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
684 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
685 			    val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
686 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
687 			break;
688 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
689 			if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
690 				gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
691 			break;
692 		case KFD_IOCTL_SVM_ATTR_ACCESS:
693 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
694 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
695 			gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
696 			break;
697 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
698 			break;
699 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
700 			break;
701 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
702 			break;
703 		default:
704 			pr_debug("unknown attr type 0x%x\n", attrs[i].type);
705 			return -EINVAL;
706 		}
707 
708 		if (gpuidx < 0) {
709 			pr_debug("no GPU 0x%x found\n", val);
710 			return -EINVAL;
711 		} else if (gpuidx < MAX_GPU_INSTANCE &&
712 			   !test_bit(gpuidx, p->svms.bitmap_supported)) {
713 			pr_debug("GPU 0x%x not supported\n", val);
714 			return -EINVAL;
715 		}
716 	}
717 
718 	return 0;
719 }
720 
721 static void
722 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
723 		      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
724 		      bool *update_mapping)
725 {
726 	uint32_t i;
727 	int gpuidx;
728 
729 	for (i = 0; i < nattr; i++) {
730 		switch (attrs[i].type) {
731 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
732 			prange->preferred_loc = attrs[i].value;
733 			break;
734 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
735 			prange->prefetch_loc = attrs[i].value;
736 			break;
737 		case KFD_IOCTL_SVM_ATTR_ACCESS:
738 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
739 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
740 			if (!p->xnack_enabled)
741 				*update_mapping = true;
742 
743 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
744 							       attrs[i].value);
745 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
746 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
747 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
748 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
749 				bitmap_set(prange->bitmap_access, gpuidx, 1);
750 				bitmap_clear(prange->bitmap_aip, gpuidx, 1);
751 			} else {
752 				bitmap_clear(prange->bitmap_access, gpuidx, 1);
753 				bitmap_set(prange->bitmap_aip, gpuidx, 1);
754 			}
755 			break;
756 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
757 			*update_mapping = true;
758 			prange->flags |= attrs[i].value;
759 			break;
760 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
761 			*update_mapping = true;
762 			prange->flags &= ~attrs[i].value;
763 			break;
764 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
765 			prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F);
766 			break;
767 		default:
768 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
769 		}
770 	}
771 }
772 
773 static bool
774 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
775 			uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
776 {
777 	uint32_t i;
778 	int gpuidx;
779 
780 	for (i = 0; i < nattr; i++) {
781 		switch (attrs[i].type) {
782 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
783 			if (prange->preferred_loc != attrs[i].value)
784 				return false;
785 			break;
786 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
787 			/* Prefetch should always trigger a migration even
788 			 * if the value of the attribute didn't change.
789 			 */
790 			return false;
791 		case KFD_IOCTL_SVM_ATTR_ACCESS:
792 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
793 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
794 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
795 							       attrs[i].value);
796 			if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
797 				if (test_bit(gpuidx, prange->bitmap_access) ||
798 				    test_bit(gpuidx, prange->bitmap_aip))
799 					return false;
800 			} else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
801 				if (!test_bit(gpuidx, prange->bitmap_access))
802 					return false;
803 			} else {
804 				if (!test_bit(gpuidx, prange->bitmap_aip))
805 					return false;
806 			}
807 			break;
808 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
809 			if ((prange->flags & attrs[i].value) != attrs[i].value)
810 				return false;
811 			break;
812 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
813 			if ((prange->flags & attrs[i].value) != 0)
814 				return false;
815 			break;
816 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
817 			if (prange->granularity != attrs[i].value)
818 				return false;
819 			break;
820 		default:
821 			WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
822 		}
823 	}
824 
825 	return true;
826 }
827 
828 /**
829  * svm_range_debug_dump - print all range information from svms
830  * @svms: svm range list header
831  *
832  * debug output svm range start, end, prefetch location from svms
833  * interval tree and link list
834  *
835  * Context: The caller must hold svms->lock
836  */
837 static void svm_range_debug_dump(struct svm_range_list *svms)
838 {
839 	struct interval_tree_node *node;
840 	struct svm_range *prange;
841 
842 	pr_debug("dump svms 0x%p list\n", svms);
843 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
844 
845 	list_for_each_entry(prange, &svms->list, list) {
846 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
847 			 prange, prange->start, prange->npages,
848 			 prange->start + prange->npages - 1,
849 			 prange->actual_loc);
850 	}
851 
852 	pr_debug("dump svms 0x%p interval tree\n", svms);
853 	pr_debug("range\tstart\tpage\tend\t\tlocation\n");
854 	node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
855 	while (node) {
856 		prange = container_of(node, struct svm_range, it_node);
857 		pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
858 			 prange, prange->start, prange->npages,
859 			 prange->start + prange->npages - 1,
860 			 prange->actual_loc);
861 		node = interval_tree_iter_next(node, 0, ~0ULL);
862 	}
863 }
864 
865 static void *
866 svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements,
867 		     uint64_t offset)
868 {
869 	unsigned char *dst;
870 
871 	dst = kvmalloc_array(num_elements, size, GFP_KERNEL);
872 	if (!dst)
873 		return NULL;
874 	memcpy(dst, (unsigned char *)psrc + offset, num_elements * size);
875 
876 	return (void *)dst;
877 }
878 
879 static int
880 svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src)
881 {
882 	int i;
883 
884 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
885 		if (!src->dma_addr[i])
886 			continue;
887 		dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i],
888 					sizeof(*src->dma_addr[i]), src->npages, 0);
889 		if (!dst->dma_addr[i])
890 			return -ENOMEM;
891 	}
892 
893 	return 0;
894 }
895 
896 static int
897 svm_range_split_array(void *ppnew, void *ppold, size_t size,
898 		      uint64_t old_start, uint64_t old_n,
899 		      uint64_t new_start, uint64_t new_n)
900 {
901 	unsigned char *new, *old, *pold;
902 	uint64_t d;
903 
904 	if (!ppold)
905 		return 0;
906 	pold = *(unsigned char **)ppold;
907 	if (!pold)
908 		return 0;
909 
910 	d = (new_start - old_start) * size;
911 	new = svm_range_copy_array(pold, size, new_n, d);
912 	if (!new)
913 		return -ENOMEM;
914 	d = (new_start == old_start) ? new_n * size : 0;
915 	old = svm_range_copy_array(pold, size, old_n, d);
916 	if (!old) {
917 		kvfree(new);
918 		return -ENOMEM;
919 	}
920 	kvfree(pold);
921 	*(void **)ppold = old;
922 	*(void **)ppnew = new;
923 
924 	return 0;
925 }
926 
927 static int
928 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
929 		      uint64_t start, uint64_t last)
930 {
931 	uint64_t npages = last - start + 1;
932 	int i, r;
933 
934 	for (i = 0; i < MAX_GPU_INSTANCE; i++) {
935 		r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
936 					  sizeof(*old->dma_addr[i]), old->start,
937 					  npages, new->start, new->npages);
938 		if (r)
939 			return r;
940 	}
941 
942 	return 0;
943 }
944 
945 static int
946 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
947 		      uint64_t start, uint64_t last)
948 {
949 	uint64_t npages = last - start + 1;
950 
951 	pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
952 		 new->svms, new, new->start, start, last);
953 
954 	if (new->start == old->start) {
955 		new->offset = old->offset;
956 		old->offset += new->npages;
957 	} else {
958 		new->offset = old->offset + npages;
959 	}
960 
961 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
962 	new->ttm_res = old->ttm_res;
963 
964 	spin_lock(&new->svm_bo->list_lock);
965 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
966 	spin_unlock(&new->svm_bo->list_lock);
967 
968 	return 0;
969 }
970 
971 /**
972  * svm_range_split_adjust - split range and adjust
973  *
974  * @new: new range
975  * @old: the old range
976  * @start: the old range adjust to start address in pages
977  * @last: the old range adjust to last address in pages
978  *
979  * Copy system memory dma_addr or vram ttm_res in old range to new
980  * range from new_start up to size new->npages, the remaining old range is from
981  * start to last
982  *
983  * Return:
984  * 0 - OK, -ENOMEM - out of memory
985  */
986 static int
987 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
988 		      uint64_t start, uint64_t last)
989 {
990 	int r;
991 
992 	pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
993 		 new->svms, new->start, old->start, old->last, start, last);
994 
995 	if (new->start < old->start ||
996 	    new->last > old->last) {
997 		WARN_ONCE(1, "invalid new range start or last\n");
998 		return -EINVAL;
999 	}
1000 
1001 	r = svm_range_split_pages(new, old, start, last);
1002 	if (r)
1003 		return r;
1004 
1005 	if (old->actual_loc && old->ttm_res) {
1006 		r = svm_range_split_nodes(new, old, start, last);
1007 		if (r)
1008 			return r;
1009 	}
1010 
1011 	old->npages = last - start + 1;
1012 	old->start = start;
1013 	old->last = last;
1014 	new->flags = old->flags;
1015 	new->preferred_loc = old->preferred_loc;
1016 	new->prefetch_loc = old->prefetch_loc;
1017 	new->actual_loc = old->actual_loc;
1018 	new->granularity = old->granularity;
1019 	new->mapped_to_gpu = old->mapped_to_gpu;
1020 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1021 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1022 
1023 	return 0;
1024 }
1025 
1026 /**
1027  * svm_range_split - split a range in 2 ranges
1028  *
1029  * @prange: the svm range to split
1030  * @start: the remaining range start address in pages
1031  * @last: the remaining range last address in pages
1032  * @new: the result new range generated
1033  *
1034  * Two cases only:
1035  * case 1: if start == prange->start
1036  *         prange ==> prange[start, last]
1037  *         new range [last + 1, prange->last]
1038  *
1039  * case 2: if last == prange->last
1040  *         prange ==> prange[start, last]
1041  *         new range [prange->start, start - 1]
1042  *
1043  * Return:
1044  * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1045  */
1046 static int
1047 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1048 		struct svm_range **new)
1049 {
1050 	uint64_t old_start = prange->start;
1051 	uint64_t old_last = prange->last;
1052 	struct svm_range_list *svms;
1053 	int r = 0;
1054 
1055 	pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1056 		 old_start, old_last, start, last);
1057 
1058 	if (old_start != start && old_last != last)
1059 		return -EINVAL;
1060 	if (start < old_start || last > old_last)
1061 		return -EINVAL;
1062 
1063 	svms = prange->svms;
1064 	if (old_start == start)
1065 		*new = svm_range_new(svms, last + 1, old_last, false);
1066 	else
1067 		*new = svm_range_new(svms, old_start, start - 1, false);
1068 	if (!*new)
1069 		return -ENOMEM;
1070 
1071 	r = svm_range_split_adjust(*new, prange, start, last);
1072 	if (r) {
1073 		pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1074 			 r, old_start, old_last, start, last);
1075 		svm_range_free(*new, false);
1076 		*new = NULL;
1077 	}
1078 
1079 	return r;
1080 }
1081 
1082 static int
1083 svm_range_split_tail(struct svm_range *prange,
1084 		     uint64_t new_last, struct list_head *insert_list)
1085 {
1086 	struct svm_range *tail;
1087 	int r = svm_range_split(prange, prange->start, new_last, &tail);
1088 
1089 	if (!r)
1090 		list_add(&tail->list, insert_list);
1091 	return r;
1092 }
1093 
1094 static int
1095 svm_range_split_head(struct svm_range *prange,
1096 		     uint64_t new_start, struct list_head *insert_list)
1097 {
1098 	struct svm_range *head;
1099 	int r = svm_range_split(prange, new_start, prange->last, &head);
1100 
1101 	if (!r)
1102 		list_add(&head->list, insert_list);
1103 	return r;
1104 }
1105 
1106 static void
1107 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1108 		    struct svm_range *pchild, enum svm_work_list_ops op)
1109 {
1110 	pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1111 		 pchild, pchild->start, pchild->last, prange, op);
1112 
1113 	pchild->work_item.mm = mm;
1114 	pchild->work_item.op = op;
1115 	list_add_tail(&pchild->child_list, &prange->child_list);
1116 }
1117 
1118 /**
1119  * svm_range_split_by_granularity - collect ranges within granularity boundary
1120  *
1121  * @p: the process with svms list
1122  * @mm: mm structure
1123  * @addr: the vm fault address in pages, to split the prange
1124  * @parent: parent range if prange is from child list
1125  * @prange: prange to split
1126  *
1127  * Trims @prange to be a single aligned block of prange->granularity if
1128  * possible. The head and tail are added to the child_list in @parent.
1129  *
1130  * Context: caller must hold mmap_read_lock and prange->lock
1131  *
1132  * Return:
1133  * 0 - OK, otherwise error code
1134  */
1135 int
1136 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1137 			       unsigned long addr, struct svm_range *parent,
1138 			       struct svm_range *prange)
1139 {
1140 	struct svm_range *head, *tail;
1141 	unsigned long start, last, size;
1142 	int r;
1143 
1144 	/* Align splited range start and size to granularity size, then a single
1145 	 * PTE will be used for whole range, this reduces the number of PTE
1146 	 * updated and the L1 TLB space used for translation.
1147 	 */
1148 	size = 1UL << prange->granularity;
1149 	start = ALIGN_DOWN(addr, size);
1150 	last = ALIGN(addr + 1, size) - 1;
1151 
1152 	pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1153 		 prange->svms, prange->start, prange->last, start, last, size);
1154 
1155 	if (start > prange->start) {
1156 		r = svm_range_split(prange, start, prange->last, &head);
1157 		if (r)
1158 			return r;
1159 		svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1160 	}
1161 
1162 	if (last < prange->last) {
1163 		r = svm_range_split(prange, prange->start, last, &tail);
1164 		if (r)
1165 			return r;
1166 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1167 	}
1168 
1169 	/* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1170 	if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1171 		prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1172 		pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1173 			 prange, prange->start, prange->last,
1174 			 SVM_OP_ADD_RANGE_AND_MAP);
1175 	}
1176 	return 0;
1177 }
1178 static bool
1179 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
1180 {
1181 	return (node_a->adev == node_b->adev ||
1182 		amdgpu_xgmi_same_hive(node_a->adev, node_b->adev));
1183 }
1184 
1185 static uint64_t
1186 svm_range_get_pte_flags(struct kfd_node *node,
1187 			struct svm_range *prange, int domain)
1188 {
1189 	struct kfd_node *bo_node;
1190 	uint32_t flags = prange->flags;
1191 	uint32_t mapping_flags = 0;
1192 	uint64_t pte_flags;
1193 	bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1194 	bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1195 	bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/
1196 	unsigned int mtype_local;
1197 
1198 	if (domain == SVM_RANGE_VRAM_DOMAIN)
1199 		bo_node = prange->svm_bo->node;
1200 
1201 	switch (node->adev->ip_versions[GC_HWIP][0]) {
1202 	case IP_VERSION(9, 4, 1):
1203 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1204 			if (bo_node == node) {
1205 				mapping_flags |= coherent ?
1206 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1207 			} else {
1208 				mapping_flags |= coherent ?
1209 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1210 				if (svm_nodes_in_same_hive(node, bo_node))
1211 					snoop = true;
1212 			}
1213 		} else {
1214 			mapping_flags |= coherent ?
1215 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1216 		}
1217 		break;
1218 	case IP_VERSION(9, 4, 2):
1219 		if (domain == SVM_RANGE_VRAM_DOMAIN) {
1220 			if (bo_node == node) {
1221 				mapping_flags |= coherent ?
1222 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1223 				if (node->adev->gmc.xgmi.connected_to_cpu)
1224 					snoop = true;
1225 			} else {
1226 				mapping_flags |= coherent ?
1227 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1228 				if (svm_nodes_in_same_hive(node, bo_node))
1229 					snoop = true;
1230 			}
1231 		} else {
1232 			mapping_flags |= coherent ?
1233 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1234 		}
1235 		break;
1236 	case IP_VERSION(9, 4, 3):
1237 		mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC :
1238 			     (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW);
1239 		snoop = true;
1240 		if (uncached) {
1241 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1242 		} else if (domain == SVM_RANGE_VRAM_DOMAIN) {
1243 			/* local HBM region close to partition */
1244 			if (bo_node->adev == node->adev &&
1245 			    (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
1246 				mapping_flags |= mtype_local;
1247 			/* local HBM region far from partition or remote XGMI GPU */
1248 			else if (svm_nodes_in_same_hive(bo_node, node))
1249 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1250 			/* PCIe P2P */
1251 			else
1252 				mapping_flags |= AMDGPU_VM_MTYPE_UC;
1253 		/* system memory accessed by the APU */
1254 		} else if (node->adev->flags & AMD_IS_APU) {
1255 			/* On NUMA systems, locality is determined per-page
1256 			 * in amdgpu_gmc_override_vm_pte_flags
1257 			 */
1258 			if (num_possible_nodes() <= 1)
1259 				mapping_flags |= mtype_local;
1260 			else
1261 				mapping_flags |= AMDGPU_VM_MTYPE_NC;
1262 		/* system memory accessed by the dGPU */
1263 		} else {
1264 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
1265 		}
1266 		break;
1267 	default:
1268 		mapping_flags |= coherent ?
1269 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1270 	}
1271 
1272 	mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1273 
1274 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1275 		mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1276 	if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1277 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1278 
1279 	pte_flags = AMDGPU_PTE_VALID;
1280 	pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1281 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1282 
1283 	pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags);
1284 	return pte_flags;
1285 }
1286 
1287 static int
1288 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1289 			 uint64_t start, uint64_t last,
1290 			 struct dma_fence **fence)
1291 {
1292 	uint64_t init_pte_value = 0;
1293 
1294 	pr_debug("[0x%llx 0x%llx]\n", start, last);
1295 
1296 	return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1297 				      last, init_pte_value, 0, 0, NULL, NULL,
1298 				      fence);
1299 }
1300 
1301 static int
1302 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1303 			  unsigned long last, uint32_t trigger)
1304 {
1305 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1306 	struct kfd_process_device *pdd;
1307 	struct dma_fence *fence = NULL;
1308 	struct kfd_process *p;
1309 	uint32_t gpuidx;
1310 	int r = 0;
1311 
1312 	if (!prange->mapped_to_gpu) {
1313 		pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1314 			 prange, prange->start, prange->last);
1315 		return 0;
1316 	}
1317 
1318 	if (prange->start == start && prange->last == last) {
1319 		pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1320 		prange->mapped_to_gpu = false;
1321 	}
1322 
1323 	bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1324 		  MAX_GPU_INSTANCE);
1325 	p = container_of(prange->svms, struct kfd_process, svms);
1326 
1327 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1328 		pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1329 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1330 		if (!pdd) {
1331 			pr_debug("failed to find device idx %d\n", gpuidx);
1332 			return -EINVAL;
1333 		}
1334 
1335 		kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1336 					     start, last, trigger);
1337 
1338 		r = svm_range_unmap_from_gpu(pdd->dev->adev,
1339 					     drm_priv_to_vm(pdd->drm_priv),
1340 					     start, last, &fence);
1341 		if (r)
1342 			break;
1343 
1344 		if (fence) {
1345 			r = dma_fence_wait(fence, false);
1346 			dma_fence_put(fence);
1347 			fence = NULL;
1348 			if (r)
1349 				break;
1350 		}
1351 		kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1352 	}
1353 
1354 	return r;
1355 }
1356 
1357 static int
1358 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1359 		     unsigned long offset, unsigned long npages, bool readonly,
1360 		     dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1361 		     struct dma_fence **fence, bool flush_tlb)
1362 {
1363 	struct amdgpu_device *adev = pdd->dev->adev;
1364 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1365 	uint64_t pte_flags;
1366 	unsigned long last_start;
1367 	int last_domain;
1368 	int r = 0;
1369 	int64_t i, j;
1370 
1371 	last_start = prange->start + offset;
1372 
1373 	pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1374 		 last_start, last_start + npages - 1, readonly);
1375 
1376 	for (i = offset; i < offset + npages; i++) {
1377 		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1378 		dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1379 
1380 		/* Collect all pages in the same address range and memory domain
1381 		 * that can be mapped with a single call to update mapping.
1382 		 */
1383 		if (i < offset + npages - 1 &&
1384 		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1385 			continue;
1386 
1387 		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1388 			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1389 
1390 		pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain);
1391 		if (readonly)
1392 			pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1393 
1394 		pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1395 			 prange->svms, last_start, prange->start + i,
1396 			 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1397 			 pte_flags);
1398 
1399 		/* For dGPU mode, we use same vm_manager to allocate VRAM for
1400 		 * different memory partition based on fpfn/lpfn, we should use
1401 		 * same vm_manager.vram_base_offset regardless memory partition.
1402 		 */
1403 		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL,
1404 					   last_start, prange->start + i,
1405 					   pte_flags,
1406 					   (last_start - prange->start) << PAGE_SHIFT,
1407 					   bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1408 					   NULL, dma_addr, &vm->last_update);
1409 
1410 		for (j = last_start - prange->start; j <= i; j++)
1411 			dma_addr[j] |= last_domain;
1412 
1413 		if (r) {
1414 			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1415 			goto out;
1416 		}
1417 		last_start = prange->start + i + 1;
1418 	}
1419 
1420 	r = amdgpu_vm_update_pdes(adev, vm, false);
1421 	if (r) {
1422 		pr_debug("failed %d to update directories 0x%lx\n", r,
1423 			 prange->start);
1424 		goto out;
1425 	}
1426 
1427 	if (fence)
1428 		*fence = dma_fence_get(vm->last_update);
1429 
1430 out:
1431 	return r;
1432 }
1433 
1434 static int
1435 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1436 		      unsigned long npages, bool readonly,
1437 		      unsigned long *bitmap, bool wait, bool flush_tlb)
1438 {
1439 	struct kfd_process_device *pdd;
1440 	struct amdgpu_device *bo_adev = NULL;
1441 	struct kfd_process *p;
1442 	struct dma_fence *fence = NULL;
1443 	uint32_t gpuidx;
1444 	int r = 0;
1445 
1446 	if (prange->svm_bo && prange->ttm_res)
1447 		bo_adev = prange->svm_bo->node->adev;
1448 
1449 	p = container_of(prange->svms, struct kfd_process, svms);
1450 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1451 		pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1452 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1453 		if (!pdd) {
1454 			pr_debug("failed to find device idx %d\n", gpuidx);
1455 			return -EINVAL;
1456 		}
1457 
1458 		pdd = kfd_bind_process_to_device(pdd->dev, p);
1459 		if (IS_ERR(pdd))
1460 			return -EINVAL;
1461 
1462 		if (bo_adev && pdd->dev->adev != bo_adev &&
1463 		    !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1464 			pr_debug("cannot map to device idx %d\n", gpuidx);
1465 			continue;
1466 		}
1467 
1468 		r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1469 					 prange->dma_addr[gpuidx],
1470 					 bo_adev, wait ? &fence : NULL,
1471 					 flush_tlb);
1472 		if (r)
1473 			break;
1474 
1475 		if (fence) {
1476 			r = dma_fence_wait(fence, false);
1477 			dma_fence_put(fence);
1478 			fence = NULL;
1479 			if (r) {
1480 				pr_debug("failed %d to dma fence wait\n", r);
1481 				break;
1482 			}
1483 		}
1484 
1485 		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1486 	}
1487 
1488 	return r;
1489 }
1490 
1491 struct svm_validate_context {
1492 	struct kfd_process *process;
1493 	struct svm_range *prange;
1494 	bool intr;
1495 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1496 	struct drm_exec exec;
1497 };
1498 
1499 static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr)
1500 {
1501 	struct kfd_process_device *pdd;
1502 	struct amdgpu_vm *vm;
1503 	uint32_t gpuidx;
1504 	int r;
1505 
1506 	drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0);
1507 	drm_exec_until_all_locked(&ctx->exec) {
1508 		for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1509 			pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1510 			if (!pdd) {
1511 				pr_debug("failed to find device idx %d\n", gpuidx);
1512 				r = -EINVAL;
1513 				goto unreserve_out;
1514 			}
1515 			vm = drm_priv_to_vm(pdd->drm_priv);
1516 
1517 			r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1518 			drm_exec_retry_on_contention(&ctx->exec);
1519 			if (unlikely(r)) {
1520 				pr_debug("failed %d to reserve bo\n", r);
1521 				goto unreserve_out;
1522 			}
1523 		}
1524 	}
1525 
1526 	for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1527 		pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1528 		if (!pdd) {
1529 			pr_debug("failed to find device idx %d\n", gpuidx);
1530 			r = -EINVAL;
1531 			goto unreserve_out;
1532 		}
1533 
1534 		r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1535 					      drm_priv_to_vm(pdd->drm_priv),
1536 					      svm_range_bo_validate, NULL);
1537 		if (r) {
1538 			pr_debug("failed %d validate pt bos\n", r);
1539 			goto unreserve_out;
1540 		}
1541 	}
1542 
1543 	return 0;
1544 
1545 unreserve_out:
1546 	drm_exec_fini(&ctx->exec);
1547 	return r;
1548 }
1549 
1550 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1551 {
1552 	drm_exec_fini(&ctx->exec);
1553 }
1554 
1555 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1556 {
1557 	struct kfd_process_device *pdd;
1558 
1559 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1560 	if (!pdd)
1561 		return NULL;
1562 
1563 	return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1564 }
1565 
1566 /*
1567  * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1568  *
1569  * To prevent concurrent destruction or change of range attributes, the
1570  * svm_read_lock must be held. The caller must not hold the svm_write_lock
1571  * because that would block concurrent evictions and lead to deadlocks. To
1572  * serialize concurrent migrations or validations of the same range, the
1573  * prange->migrate_mutex must be held.
1574  *
1575  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1576  * eviction fence.
1577  *
1578  * The following sequence ensures race-free validation and GPU mapping:
1579  *
1580  * 1. Reserve page table (and SVM BO if range is in VRAM)
1581  * 2. hmm_range_fault to get page addresses (if system memory)
1582  * 3. DMA-map pages (if system memory)
1583  * 4-a. Take notifier lock
1584  * 4-b. Check that pages still valid (mmu_interval_read_retry)
1585  * 4-c. Check that the range was not split or otherwise invalidated
1586  * 4-d. Update GPU page table
1587  * 4.e. Release notifier lock
1588  * 5. Release page table (and SVM BO) reservation
1589  */
1590 static int svm_range_validate_and_map(struct mm_struct *mm,
1591 				      struct svm_range *prange, int32_t gpuidx,
1592 				      bool intr, bool wait, bool flush_tlb)
1593 {
1594 	struct svm_validate_context *ctx;
1595 	unsigned long start, end, addr;
1596 	struct kfd_process *p;
1597 	void *owner;
1598 	int32_t idx;
1599 	int r = 0;
1600 
1601 	ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
1602 	if (!ctx)
1603 		return -ENOMEM;
1604 	ctx->process = container_of(prange->svms, struct kfd_process, svms);
1605 	ctx->prange = prange;
1606 	ctx->intr = intr;
1607 
1608 	if (gpuidx < MAX_GPU_INSTANCE) {
1609 		bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE);
1610 		bitmap_set(ctx->bitmap, gpuidx, 1);
1611 	} else if (ctx->process->xnack_enabled) {
1612 		bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1613 
1614 		/* If prefetch range to GPU, or GPU retry fault migrate range to
1615 		 * GPU, which has ACCESS attribute to the range, create mapping
1616 		 * on that GPU.
1617 		 */
1618 		if (prange->actual_loc) {
1619 			gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process,
1620 							prange->actual_loc);
1621 			if (gpuidx < 0) {
1622 				WARN_ONCE(1, "failed get device by id 0x%x\n",
1623 					 prange->actual_loc);
1624 				r = -EINVAL;
1625 				goto free_ctx;
1626 			}
1627 			if (test_bit(gpuidx, prange->bitmap_access))
1628 				bitmap_set(ctx->bitmap, gpuidx, 1);
1629 		}
1630 
1631 		/*
1632 		 * If prange is already mapped or with always mapped flag,
1633 		 * update mapping on GPUs with ACCESS attribute
1634 		 */
1635 		if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1636 			if (prange->mapped_to_gpu ||
1637 			    prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)
1638 				bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1639 		}
1640 	} else {
1641 		bitmap_or(ctx->bitmap, prange->bitmap_access,
1642 			  prange->bitmap_aip, MAX_GPU_INSTANCE);
1643 	}
1644 
1645 	if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1646 		r = 0;
1647 		goto free_ctx;
1648 	}
1649 
1650 	if (prange->actual_loc && !prange->ttm_res) {
1651 		/* This should never happen. actual_loc gets set by
1652 		 * svm_migrate_ram_to_vram after allocating a BO.
1653 		 */
1654 		WARN_ONCE(1, "VRAM BO missing during validation\n");
1655 		r = -EINVAL;
1656 		goto free_ctx;
1657 	}
1658 
1659 	svm_range_reserve_bos(ctx, intr);
1660 
1661 	p = container_of(prange->svms, struct kfd_process, svms);
1662 	owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
1663 						MAX_GPU_INSTANCE));
1664 	for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) {
1665 		if (kfd_svm_page_owner(p, idx) != owner) {
1666 			owner = NULL;
1667 			break;
1668 		}
1669 	}
1670 
1671 	start = prange->start << PAGE_SHIFT;
1672 	end = (prange->last + 1) << PAGE_SHIFT;
1673 	for (addr = start; !r && addr < end; ) {
1674 		struct hmm_range *hmm_range;
1675 		struct vm_area_struct *vma;
1676 		unsigned long next = 0;
1677 		unsigned long offset;
1678 		unsigned long npages;
1679 		bool readonly;
1680 
1681 		vma = vma_lookup(mm, addr);
1682 		if (vma) {
1683 			readonly = !(vma->vm_flags & VM_WRITE);
1684 
1685 			next = min(vma->vm_end, end);
1686 			npages = (next - addr) >> PAGE_SHIFT;
1687 			WRITE_ONCE(p->svms.faulting_task, current);
1688 			r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages,
1689 						       readonly, owner, NULL,
1690 						       &hmm_range);
1691 			WRITE_ONCE(p->svms.faulting_task, NULL);
1692 			if (r) {
1693 				pr_debug("failed %d to get svm range pages\n", r);
1694 				if (r == -EBUSY)
1695 					r = -EAGAIN;
1696 			}
1697 		} else {
1698 			r = -EFAULT;
1699 		}
1700 
1701 		if (!r) {
1702 			offset = (addr - start) >> PAGE_SHIFT;
1703 			r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
1704 					      hmm_range->hmm_pfns);
1705 			if (r)
1706 				pr_debug("failed %d to dma map range\n", r);
1707 		}
1708 
1709 		svm_range_lock(prange);
1710 		if (!r && amdgpu_hmm_range_get_pages_done(hmm_range)) {
1711 			pr_debug("hmm update the range, need validate again\n");
1712 			r = -EAGAIN;
1713 		}
1714 
1715 		if (!r && !list_empty(&prange->child_list)) {
1716 			pr_debug("range split by unmap in parallel, validate again\n");
1717 			r = -EAGAIN;
1718 		}
1719 
1720 		if (!r)
1721 			r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1722 						  ctx->bitmap, wait, flush_tlb);
1723 
1724 		if (!r && next == end)
1725 			prange->mapped_to_gpu = true;
1726 
1727 		svm_range_unlock(prange);
1728 
1729 		addr = next;
1730 	}
1731 
1732 	svm_range_unreserve_bos(ctx);
1733 	if (!r)
1734 		prange->validate_timestamp = ktime_get_boottime();
1735 
1736 free_ctx:
1737 	kfree(ctx);
1738 
1739 	return r;
1740 }
1741 
1742 /**
1743  * svm_range_list_lock_and_flush_work - flush pending deferred work
1744  *
1745  * @svms: the svm range list
1746  * @mm: the mm structure
1747  *
1748  * Context: Returns with mmap write lock held, pending deferred work flushed
1749  *
1750  */
1751 void
1752 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1753 				   struct mm_struct *mm)
1754 {
1755 retry_flush_work:
1756 	flush_work(&svms->deferred_list_work);
1757 	mmap_write_lock(mm);
1758 
1759 	if (list_empty(&svms->deferred_range_list))
1760 		return;
1761 	mmap_write_unlock(mm);
1762 	pr_debug("retry flush\n");
1763 	goto retry_flush_work;
1764 }
1765 
1766 static void svm_range_restore_work(struct work_struct *work)
1767 {
1768 	struct delayed_work *dwork = to_delayed_work(work);
1769 	struct amdkfd_process_info *process_info;
1770 	struct svm_range_list *svms;
1771 	struct svm_range *prange;
1772 	struct kfd_process *p;
1773 	struct mm_struct *mm;
1774 	int evicted_ranges;
1775 	int invalid;
1776 	int r;
1777 
1778 	svms = container_of(dwork, struct svm_range_list, restore_work);
1779 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1780 	if (!evicted_ranges)
1781 		return;
1782 
1783 	pr_debug("restore svm ranges\n");
1784 
1785 	p = container_of(svms, struct kfd_process, svms);
1786 	process_info = p->kgd_process_info;
1787 
1788 	/* Keep mm reference when svm_range_validate_and_map ranges */
1789 	mm = get_task_mm(p->lead_thread);
1790 	if (!mm) {
1791 		pr_debug("svms 0x%p process mm gone\n", svms);
1792 		return;
1793 	}
1794 
1795 	mutex_lock(&process_info->lock);
1796 	svm_range_list_lock_and_flush_work(svms, mm);
1797 	mutex_lock(&svms->lock);
1798 
1799 	evicted_ranges = atomic_read(&svms->evicted_ranges);
1800 
1801 	list_for_each_entry(prange, &svms->list, list) {
1802 		invalid = atomic_read(&prange->invalid);
1803 		if (!invalid)
1804 			continue;
1805 
1806 		pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1807 			 prange->svms, prange, prange->start, prange->last,
1808 			 invalid);
1809 
1810 		/*
1811 		 * If range is migrating, wait for migration is done.
1812 		 */
1813 		mutex_lock(&prange->migrate_mutex);
1814 
1815 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1816 					       false, true, false);
1817 		if (r)
1818 			pr_debug("failed %d to map 0x%lx to gpus\n", r,
1819 				 prange->start);
1820 
1821 		mutex_unlock(&prange->migrate_mutex);
1822 		if (r)
1823 			goto out_reschedule;
1824 
1825 		if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1826 			goto out_reschedule;
1827 	}
1828 
1829 	if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1830 	    evicted_ranges)
1831 		goto out_reschedule;
1832 
1833 	evicted_ranges = 0;
1834 
1835 	r = kgd2kfd_resume_mm(mm);
1836 	if (r) {
1837 		/* No recovery from this failure. Probably the CP is
1838 		 * hanging. No point trying again.
1839 		 */
1840 		pr_debug("failed %d to resume KFD\n", r);
1841 	}
1842 
1843 	pr_debug("restore svm ranges successfully\n");
1844 
1845 out_reschedule:
1846 	mutex_unlock(&svms->lock);
1847 	mmap_write_unlock(mm);
1848 	mutex_unlock(&process_info->lock);
1849 
1850 	/* If validation failed, reschedule another attempt */
1851 	if (evicted_ranges) {
1852 		pr_debug("reschedule to restore svm range\n");
1853 		schedule_delayed_work(&svms->restore_work,
1854 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1855 
1856 		kfd_smi_event_queue_restore_rescheduled(mm);
1857 	}
1858 	mmput(mm);
1859 }
1860 
1861 /**
1862  * svm_range_evict - evict svm range
1863  * @prange: svm range structure
1864  * @mm: current process mm_struct
1865  * @start: starting process queue number
1866  * @last: last process queue number
1867  * @event: mmu notifier event when range is evicted or migrated
1868  *
1869  * Stop all queues of the process to ensure GPU doesn't access the memory, then
1870  * return to let CPU evict the buffer and proceed CPU pagetable update.
1871  *
1872  * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1873  * If invalidation happens while restore work is running, restore work will
1874  * restart to ensure to get the latest CPU pages mapping to GPU, then start
1875  * the queues.
1876  */
1877 static int
1878 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1879 		unsigned long start, unsigned long last,
1880 		enum mmu_notifier_event event)
1881 {
1882 	struct svm_range_list *svms = prange->svms;
1883 	struct svm_range *pchild;
1884 	struct kfd_process *p;
1885 	int r = 0;
1886 
1887 	p = container_of(svms, struct kfd_process, svms);
1888 
1889 	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1890 		 svms, prange->start, prange->last, start, last);
1891 
1892 	if (!p->xnack_enabled ||
1893 	    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1894 		int evicted_ranges;
1895 		bool mapped = prange->mapped_to_gpu;
1896 
1897 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1898 			if (!pchild->mapped_to_gpu)
1899 				continue;
1900 			mapped = true;
1901 			mutex_lock_nested(&pchild->lock, 1);
1902 			if (pchild->start <= last && pchild->last >= start) {
1903 				pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1904 					 pchild->start, pchild->last);
1905 				atomic_inc(&pchild->invalid);
1906 			}
1907 			mutex_unlock(&pchild->lock);
1908 		}
1909 
1910 		if (!mapped)
1911 			return r;
1912 
1913 		if (prange->start <= last && prange->last >= start)
1914 			atomic_inc(&prange->invalid);
1915 
1916 		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1917 		if (evicted_ranges != 1)
1918 			return r;
1919 
1920 		pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1921 			 prange->svms, prange->start, prange->last);
1922 
1923 		/* First eviction, stop the queues */
1924 		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1925 		if (r)
1926 			pr_debug("failed to quiesce KFD\n");
1927 
1928 		pr_debug("schedule to restore svm %p ranges\n", svms);
1929 		schedule_delayed_work(&svms->restore_work,
1930 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1931 	} else {
1932 		unsigned long s, l;
1933 		uint32_t trigger;
1934 
1935 		if (event == MMU_NOTIFY_MIGRATE)
1936 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1937 		else
1938 			trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1939 
1940 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1941 			 prange->svms, start, last);
1942 		list_for_each_entry(pchild, &prange->child_list, child_list) {
1943 			mutex_lock_nested(&pchild->lock, 1);
1944 			s = max(start, pchild->start);
1945 			l = min(last, pchild->last);
1946 			if (l >= s)
1947 				svm_range_unmap_from_gpus(pchild, s, l, trigger);
1948 			mutex_unlock(&pchild->lock);
1949 		}
1950 		s = max(start, prange->start);
1951 		l = min(last, prange->last);
1952 		if (l >= s)
1953 			svm_range_unmap_from_gpus(prange, s, l, trigger);
1954 	}
1955 
1956 	return r;
1957 }
1958 
1959 static struct svm_range *svm_range_clone(struct svm_range *old)
1960 {
1961 	struct svm_range *new;
1962 
1963 	new = svm_range_new(old->svms, old->start, old->last, false);
1964 	if (!new)
1965 		return NULL;
1966 	if (svm_range_copy_dma_addrs(new, old)) {
1967 		svm_range_free(new, false);
1968 		return NULL;
1969 	}
1970 	if (old->svm_bo) {
1971 		new->ttm_res = old->ttm_res;
1972 		new->offset = old->offset;
1973 		new->svm_bo = svm_range_bo_ref(old->svm_bo);
1974 		spin_lock(&new->svm_bo->list_lock);
1975 		list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1976 		spin_unlock(&new->svm_bo->list_lock);
1977 	}
1978 	new->flags = old->flags;
1979 	new->preferred_loc = old->preferred_loc;
1980 	new->prefetch_loc = old->prefetch_loc;
1981 	new->actual_loc = old->actual_loc;
1982 	new->granularity = old->granularity;
1983 	new->mapped_to_gpu = old->mapped_to_gpu;
1984 	bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1985 	bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1986 
1987 	return new;
1988 }
1989 
1990 void svm_range_set_max_pages(struct amdgpu_device *adev)
1991 {
1992 	uint64_t max_pages;
1993 	uint64_t pages, _pages;
1994 	uint64_t min_pages = 0;
1995 	int i, id;
1996 
1997 	for (i = 0; i < adev->kfd.dev->num_nodes; i++) {
1998 		if (adev->kfd.dev->nodes[i]->xcp)
1999 			id = adev->kfd.dev->nodes[i]->xcp->id;
2000 		else
2001 			id = -1;
2002 		pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17;
2003 		pages = clamp(pages, 1ULL << 9, 1ULL << 18);
2004 		pages = rounddown_pow_of_two(pages);
2005 		min_pages = min_not_zero(min_pages, pages);
2006 	}
2007 
2008 	do {
2009 		max_pages = READ_ONCE(max_svm_range_pages);
2010 		_pages = min_not_zero(max_pages, min_pages);
2011 	} while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
2012 }
2013 
2014 static int
2015 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
2016 		    uint64_t max_pages, struct list_head *insert_list,
2017 		    struct list_head *update_list)
2018 {
2019 	struct svm_range *prange;
2020 	uint64_t l;
2021 
2022 	pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
2023 		 max_pages, start, last);
2024 
2025 	while (last >= start) {
2026 		l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
2027 
2028 		prange = svm_range_new(svms, start, l, true);
2029 		if (!prange)
2030 			return -ENOMEM;
2031 		list_add(&prange->list, insert_list);
2032 		list_add(&prange->update_list, update_list);
2033 
2034 		start = l + 1;
2035 	}
2036 	return 0;
2037 }
2038 
2039 /**
2040  * svm_range_add - add svm range and handle overlap
2041  * @p: the range add to this process svms
2042  * @start: page size aligned
2043  * @size: page size aligned
2044  * @nattr: number of attributes
2045  * @attrs: array of attributes
2046  * @update_list: output, the ranges need validate and update GPU mapping
2047  * @insert_list: output, the ranges need insert to svms
2048  * @remove_list: output, the ranges are replaced and need remove from svms
2049  *
2050  * Check if the virtual address range has overlap with any existing ranges,
2051  * split partly overlapping ranges and add new ranges in the gaps. All changes
2052  * should be applied to the range_list and interval tree transactionally. If
2053  * any range split or allocation fails, the entire update fails. Therefore any
2054  * existing overlapping svm_ranges are cloned and the original svm_ranges left
2055  * unchanged.
2056  *
2057  * If the transaction succeeds, the caller can update and insert clones and
2058  * new ranges, then free the originals.
2059  *
2060  * Otherwise the caller can free the clones and new ranges, while the old
2061  * svm_ranges remain unchanged.
2062  *
2063  * Context: Process context, caller must hold svms->lock
2064  *
2065  * Return:
2066  * 0 - OK, otherwise error code
2067  */
2068 static int
2069 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2070 	      uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2071 	      struct list_head *update_list, struct list_head *insert_list,
2072 	      struct list_head *remove_list)
2073 {
2074 	unsigned long last = start + size - 1UL;
2075 	struct svm_range_list *svms = &p->svms;
2076 	struct interval_tree_node *node;
2077 	struct svm_range *prange;
2078 	struct svm_range *tmp;
2079 	struct list_head new_list;
2080 	int r = 0;
2081 
2082 	pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
2083 
2084 	INIT_LIST_HEAD(update_list);
2085 	INIT_LIST_HEAD(insert_list);
2086 	INIT_LIST_HEAD(remove_list);
2087 	INIT_LIST_HEAD(&new_list);
2088 
2089 	node = interval_tree_iter_first(&svms->objects, start, last);
2090 	while (node) {
2091 		struct interval_tree_node *next;
2092 		unsigned long next_start;
2093 
2094 		pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2095 			 node->last);
2096 
2097 		prange = container_of(node, struct svm_range, it_node);
2098 		next = interval_tree_iter_next(node, start, last);
2099 		next_start = min(node->last, last) + 1;
2100 
2101 		if (svm_range_is_same_attrs(p, prange, nattr, attrs) &&
2102 		    prange->mapped_to_gpu) {
2103 			/* nothing to do */
2104 		} else if (node->start < start || node->last > last) {
2105 			/* node intersects the update range and its attributes
2106 			 * will change. Clone and split it, apply updates only
2107 			 * to the overlapping part
2108 			 */
2109 			struct svm_range *old = prange;
2110 
2111 			prange = svm_range_clone(old);
2112 			if (!prange) {
2113 				r = -ENOMEM;
2114 				goto out;
2115 			}
2116 
2117 			list_add(&old->update_list, remove_list);
2118 			list_add(&prange->list, insert_list);
2119 			list_add(&prange->update_list, update_list);
2120 
2121 			if (node->start < start) {
2122 				pr_debug("change old range start\n");
2123 				r = svm_range_split_head(prange, start,
2124 							 insert_list);
2125 				if (r)
2126 					goto out;
2127 			}
2128 			if (node->last > last) {
2129 				pr_debug("change old range last\n");
2130 				r = svm_range_split_tail(prange, last,
2131 							 insert_list);
2132 				if (r)
2133 					goto out;
2134 			}
2135 		} else {
2136 			/* The node is contained within start..last,
2137 			 * just update it
2138 			 */
2139 			list_add(&prange->update_list, update_list);
2140 		}
2141 
2142 		/* insert a new node if needed */
2143 		if (node->start > start) {
2144 			r = svm_range_split_new(svms, start, node->start - 1,
2145 						READ_ONCE(max_svm_range_pages),
2146 						&new_list, update_list);
2147 			if (r)
2148 				goto out;
2149 		}
2150 
2151 		node = next;
2152 		start = next_start;
2153 	}
2154 
2155 	/* add a final range at the end if needed */
2156 	if (start <= last)
2157 		r = svm_range_split_new(svms, start, last,
2158 					READ_ONCE(max_svm_range_pages),
2159 					&new_list, update_list);
2160 
2161 out:
2162 	if (r) {
2163 		list_for_each_entry_safe(prange, tmp, insert_list, list)
2164 			svm_range_free(prange, false);
2165 		list_for_each_entry_safe(prange, tmp, &new_list, list)
2166 			svm_range_free(prange, true);
2167 	} else {
2168 		list_splice(&new_list, insert_list);
2169 	}
2170 
2171 	return r;
2172 }
2173 
2174 static void
2175 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2176 					    struct svm_range *prange)
2177 {
2178 	unsigned long start;
2179 	unsigned long last;
2180 
2181 	start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2182 	last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2183 
2184 	if (prange->start == start && prange->last == last)
2185 		return;
2186 
2187 	pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2188 		  prange->svms, prange, start, last, prange->start,
2189 		  prange->last);
2190 
2191 	if (start != 0 && last != 0) {
2192 		interval_tree_remove(&prange->it_node, &prange->svms->objects);
2193 		svm_range_remove_notifier(prange);
2194 	}
2195 	prange->it_node.start = prange->start;
2196 	prange->it_node.last = prange->last;
2197 
2198 	interval_tree_insert(&prange->it_node, &prange->svms->objects);
2199 	svm_range_add_notifier_locked(mm, prange);
2200 }
2201 
2202 static void
2203 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2204 			 struct mm_struct *mm)
2205 {
2206 	switch (prange->work_item.op) {
2207 	case SVM_OP_NULL:
2208 		pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2209 			 svms, prange, prange->start, prange->last);
2210 		break;
2211 	case SVM_OP_UNMAP_RANGE:
2212 		pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2213 			 svms, prange, prange->start, prange->last);
2214 		svm_range_unlink(prange);
2215 		svm_range_remove_notifier(prange);
2216 		svm_range_free(prange, true);
2217 		break;
2218 	case SVM_OP_UPDATE_RANGE_NOTIFIER:
2219 		pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2220 			 svms, prange, prange->start, prange->last);
2221 		svm_range_update_notifier_and_interval_tree(mm, prange);
2222 		break;
2223 	case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2224 		pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2225 			 svms, prange, prange->start, prange->last);
2226 		svm_range_update_notifier_and_interval_tree(mm, prange);
2227 		/* TODO: implement deferred validation and mapping */
2228 		break;
2229 	case SVM_OP_ADD_RANGE:
2230 		pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2231 			 prange->start, prange->last);
2232 		svm_range_add_to_svms(prange);
2233 		svm_range_add_notifier_locked(mm, prange);
2234 		break;
2235 	case SVM_OP_ADD_RANGE_AND_MAP:
2236 		pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2237 			 prange, prange->start, prange->last);
2238 		svm_range_add_to_svms(prange);
2239 		svm_range_add_notifier_locked(mm, prange);
2240 		/* TODO: implement deferred validation and mapping */
2241 		break;
2242 	default:
2243 		WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2244 			 prange->work_item.op);
2245 	}
2246 }
2247 
2248 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2249 {
2250 	struct kfd_process_device *pdd;
2251 	struct kfd_process *p;
2252 	int drain;
2253 	uint32_t i;
2254 
2255 	p = container_of(svms, struct kfd_process, svms);
2256 
2257 restart:
2258 	drain = atomic_read(&svms->drain_pagefaults);
2259 	if (!drain)
2260 		return;
2261 
2262 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2263 		pdd = p->pdds[i];
2264 		if (!pdd)
2265 			continue;
2266 
2267 		pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2268 
2269 		amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2270 				pdd->dev->adev->irq.retry_cam_enabled ?
2271 				&pdd->dev->adev->irq.ih :
2272 				&pdd->dev->adev->irq.ih1);
2273 
2274 		if (pdd->dev->adev->irq.retry_cam_enabled)
2275 			amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2276 				&pdd->dev->adev->irq.ih_soft);
2277 
2278 
2279 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2280 	}
2281 	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2282 		goto restart;
2283 }
2284 
2285 static void svm_range_deferred_list_work(struct work_struct *work)
2286 {
2287 	struct svm_range_list *svms;
2288 	struct svm_range *prange;
2289 	struct mm_struct *mm;
2290 
2291 	svms = container_of(work, struct svm_range_list, deferred_list_work);
2292 	pr_debug("enter svms 0x%p\n", svms);
2293 
2294 	spin_lock(&svms->deferred_list_lock);
2295 	while (!list_empty(&svms->deferred_range_list)) {
2296 		prange = list_first_entry(&svms->deferred_range_list,
2297 					  struct svm_range, deferred_list);
2298 		spin_unlock(&svms->deferred_list_lock);
2299 
2300 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2301 			 prange->start, prange->last, prange->work_item.op);
2302 
2303 		mm = prange->work_item.mm;
2304 retry:
2305 		mmap_write_lock(mm);
2306 
2307 		/* Checking for the need to drain retry faults must be inside
2308 		 * mmap write lock to serialize with munmap notifiers.
2309 		 */
2310 		if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2311 			mmap_write_unlock(mm);
2312 			svm_range_drain_retry_fault(svms);
2313 			goto retry;
2314 		}
2315 
2316 		/* Remove from deferred_list must be inside mmap write lock, for
2317 		 * two race cases:
2318 		 * 1. unmap_from_cpu may change work_item.op and add the range
2319 		 *    to deferred_list again, cause use after free bug.
2320 		 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2321 		 *    lock and continue because deferred_list is empty, but
2322 		 *    deferred_list work is actually waiting for mmap lock.
2323 		 */
2324 		spin_lock(&svms->deferred_list_lock);
2325 		list_del_init(&prange->deferred_list);
2326 		spin_unlock(&svms->deferred_list_lock);
2327 
2328 		mutex_lock(&svms->lock);
2329 		mutex_lock(&prange->migrate_mutex);
2330 		while (!list_empty(&prange->child_list)) {
2331 			struct svm_range *pchild;
2332 
2333 			pchild = list_first_entry(&prange->child_list,
2334 						struct svm_range, child_list);
2335 			pr_debug("child prange 0x%p op %d\n", pchild,
2336 				 pchild->work_item.op);
2337 			list_del_init(&pchild->child_list);
2338 			svm_range_handle_list_op(svms, pchild, mm);
2339 		}
2340 		mutex_unlock(&prange->migrate_mutex);
2341 
2342 		svm_range_handle_list_op(svms, prange, mm);
2343 		mutex_unlock(&svms->lock);
2344 		mmap_write_unlock(mm);
2345 
2346 		/* Pairs with mmget in svm_range_add_list_work. If dropping the
2347 		 * last mm refcount, schedule release work to avoid circular locking
2348 		 */
2349 		mmput_async(mm);
2350 
2351 		spin_lock(&svms->deferred_list_lock);
2352 	}
2353 	spin_unlock(&svms->deferred_list_lock);
2354 	pr_debug("exit svms 0x%p\n", svms);
2355 }
2356 
2357 void
2358 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2359 			struct mm_struct *mm, enum svm_work_list_ops op)
2360 {
2361 	spin_lock(&svms->deferred_list_lock);
2362 	/* if prange is on the deferred list */
2363 	if (!list_empty(&prange->deferred_list)) {
2364 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2365 		WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2366 		if (op != SVM_OP_NULL &&
2367 		    prange->work_item.op != SVM_OP_UNMAP_RANGE)
2368 			prange->work_item.op = op;
2369 	} else {
2370 		prange->work_item.op = op;
2371 
2372 		/* Pairs with mmput in deferred_list_work */
2373 		mmget(mm);
2374 		prange->work_item.mm = mm;
2375 		list_add_tail(&prange->deferred_list,
2376 			      &prange->svms->deferred_range_list);
2377 		pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2378 			 prange, prange->start, prange->last, op);
2379 	}
2380 	spin_unlock(&svms->deferred_list_lock);
2381 }
2382 
2383 void schedule_deferred_list_work(struct svm_range_list *svms)
2384 {
2385 	spin_lock(&svms->deferred_list_lock);
2386 	if (!list_empty(&svms->deferred_range_list))
2387 		schedule_work(&svms->deferred_list_work);
2388 	spin_unlock(&svms->deferred_list_lock);
2389 }
2390 
2391 static void
2392 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2393 		      struct svm_range *prange, unsigned long start,
2394 		      unsigned long last)
2395 {
2396 	struct svm_range *head;
2397 	struct svm_range *tail;
2398 
2399 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2400 		pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2401 			 prange->start, prange->last);
2402 		return;
2403 	}
2404 	if (start > prange->last || last < prange->start)
2405 		return;
2406 
2407 	head = tail = prange;
2408 	if (start > prange->start)
2409 		svm_range_split(prange, prange->start, start - 1, &tail);
2410 	if (last < tail->last)
2411 		svm_range_split(tail, last + 1, tail->last, &head);
2412 
2413 	if (head != prange && tail != prange) {
2414 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2415 		svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2416 	} else if (tail != prange) {
2417 		svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2418 	} else if (head != prange) {
2419 		svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2420 	} else if (parent != prange) {
2421 		prange->work_item.op = SVM_OP_UNMAP_RANGE;
2422 	}
2423 }
2424 
2425 static void
2426 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2427 			 unsigned long start, unsigned long last)
2428 {
2429 	uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU;
2430 	struct svm_range_list *svms;
2431 	struct svm_range *pchild;
2432 	struct kfd_process *p;
2433 	unsigned long s, l;
2434 	bool unmap_parent;
2435 
2436 	p = kfd_lookup_process_by_mm(mm);
2437 	if (!p)
2438 		return;
2439 	svms = &p->svms;
2440 
2441 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2442 		 prange, prange->start, prange->last, start, last);
2443 
2444 	/* Make sure pending page faults are drained in the deferred worker
2445 	 * before the range is freed to avoid straggler interrupts on
2446 	 * unmapped memory causing "phantom faults".
2447 	 */
2448 	atomic_inc(&svms->drain_pagefaults);
2449 
2450 	unmap_parent = start <= prange->start && last >= prange->last;
2451 
2452 	list_for_each_entry(pchild, &prange->child_list, child_list) {
2453 		mutex_lock_nested(&pchild->lock, 1);
2454 		s = max(start, pchild->start);
2455 		l = min(last, pchild->last);
2456 		if (l >= s)
2457 			svm_range_unmap_from_gpus(pchild, s, l, trigger);
2458 		svm_range_unmap_split(mm, prange, pchild, start, last);
2459 		mutex_unlock(&pchild->lock);
2460 	}
2461 	s = max(start, prange->start);
2462 	l = min(last, prange->last);
2463 	if (l >= s)
2464 		svm_range_unmap_from_gpus(prange, s, l, trigger);
2465 	svm_range_unmap_split(mm, prange, prange, start, last);
2466 
2467 	if (unmap_parent)
2468 		svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2469 	else
2470 		svm_range_add_list_work(svms, prange, mm,
2471 					SVM_OP_UPDATE_RANGE_NOTIFIER);
2472 	schedule_deferred_list_work(svms);
2473 
2474 	kfd_unref_process(p);
2475 }
2476 
2477 /**
2478  * svm_range_cpu_invalidate_pagetables - interval notifier callback
2479  * @mni: mmu_interval_notifier struct
2480  * @range: mmu_notifier_range struct
2481  * @cur_seq: value to pass to mmu_interval_set_seq()
2482  *
2483  * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2484  * is from migration, or CPU page invalidation callback.
2485  *
2486  * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2487  * work thread, and split prange if only part of prange is unmapped.
2488  *
2489  * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2490  * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2491  * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2492  * update GPU mapping to recover.
2493  *
2494  * Context: mmap lock, notifier_invalidate_start lock are held
2495  *          for invalidate event, prange lock is held if this is from migration
2496  */
2497 static bool
2498 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2499 				    const struct mmu_notifier_range *range,
2500 				    unsigned long cur_seq)
2501 {
2502 	struct svm_range *prange;
2503 	unsigned long start;
2504 	unsigned long last;
2505 
2506 	if (range->event == MMU_NOTIFY_RELEASE)
2507 		return true;
2508 	if (!mmget_not_zero(mni->mm))
2509 		return true;
2510 
2511 	start = mni->interval_tree.start;
2512 	last = mni->interval_tree.last;
2513 	start = max(start, range->start) >> PAGE_SHIFT;
2514 	last = min(last, range->end - 1) >> PAGE_SHIFT;
2515 	pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2516 		 start, last, range->start >> PAGE_SHIFT,
2517 		 (range->end - 1) >> PAGE_SHIFT,
2518 		 mni->interval_tree.start >> PAGE_SHIFT,
2519 		 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2520 
2521 	prange = container_of(mni, struct svm_range, notifier);
2522 
2523 	svm_range_lock(prange);
2524 	mmu_interval_set_seq(mni, cur_seq);
2525 
2526 	switch (range->event) {
2527 	case MMU_NOTIFY_UNMAP:
2528 		svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2529 		break;
2530 	default:
2531 		svm_range_evict(prange, mni->mm, start, last, range->event);
2532 		break;
2533 	}
2534 
2535 	svm_range_unlock(prange);
2536 	mmput(mni->mm);
2537 
2538 	return true;
2539 }
2540 
2541 /**
2542  * svm_range_from_addr - find svm range from fault address
2543  * @svms: svm range list header
2544  * @addr: address to search range interval tree, in pages
2545  * @parent: parent range if range is on child list
2546  *
2547  * Context: The caller must hold svms->lock
2548  *
2549  * Return: the svm_range found or NULL
2550  */
2551 struct svm_range *
2552 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2553 		    struct svm_range **parent)
2554 {
2555 	struct interval_tree_node *node;
2556 	struct svm_range *prange;
2557 	struct svm_range *pchild;
2558 
2559 	node = interval_tree_iter_first(&svms->objects, addr, addr);
2560 	if (!node)
2561 		return NULL;
2562 
2563 	prange = container_of(node, struct svm_range, it_node);
2564 	pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2565 		 addr, prange->start, prange->last, node->start, node->last);
2566 
2567 	if (addr >= prange->start && addr <= prange->last) {
2568 		if (parent)
2569 			*parent = prange;
2570 		return prange;
2571 	}
2572 	list_for_each_entry(pchild, &prange->child_list, child_list)
2573 		if (addr >= pchild->start && addr <= pchild->last) {
2574 			pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2575 				 addr, pchild->start, pchild->last);
2576 			if (parent)
2577 				*parent = prange;
2578 			return pchild;
2579 		}
2580 
2581 	return NULL;
2582 }
2583 
2584 /* svm_range_best_restore_location - decide the best fault restore location
2585  * @prange: svm range structure
2586  * @adev: the GPU on which vm fault happened
2587  *
2588  * This is only called when xnack is on, to decide the best location to restore
2589  * the range mapping after GPU vm fault. Caller uses the best location to do
2590  * migration if actual loc is not best location, then update GPU page table
2591  * mapping to the best location.
2592  *
2593  * If the preferred loc is accessible by faulting GPU, use preferred loc.
2594  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2595  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2596  *    if range actual loc is cpu, best_loc is cpu
2597  *    if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2598  *    range actual loc.
2599  * Otherwise, GPU no access, best_loc is -1.
2600  *
2601  * Return:
2602  * -1 means vm fault GPU no access
2603  * 0 for CPU or GPU id
2604  */
2605 static int32_t
2606 svm_range_best_restore_location(struct svm_range *prange,
2607 				struct kfd_node *node,
2608 				int32_t *gpuidx)
2609 {
2610 	struct kfd_node *bo_node, *preferred_node;
2611 	struct kfd_process *p;
2612 	uint32_t gpuid;
2613 	int r;
2614 
2615 	p = container_of(prange->svms, struct kfd_process, svms);
2616 
2617 	r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx);
2618 	if (r < 0) {
2619 		pr_debug("failed to get gpuid from kgd\n");
2620 		return -1;
2621 	}
2622 
2623 	if (node->adev->gmc.is_app_apu)
2624 		return 0;
2625 
2626 	if (prange->preferred_loc == gpuid ||
2627 	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2628 		return prange->preferred_loc;
2629 	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2630 		preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc);
2631 		if (preferred_node && svm_nodes_in_same_hive(node, preferred_node))
2632 			return prange->preferred_loc;
2633 		/* fall through */
2634 	}
2635 
2636 	if (test_bit(*gpuidx, prange->bitmap_access))
2637 		return gpuid;
2638 
2639 	if (test_bit(*gpuidx, prange->bitmap_aip)) {
2640 		if (!prange->actual_loc)
2641 			return 0;
2642 
2643 		bo_node = svm_range_get_node_by_id(prange, prange->actual_loc);
2644 		if (bo_node && svm_nodes_in_same_hive(node, bo_node))
2645 			return prange->actual_loc;
2646 		else
2647 			return 0;
2648 	}
2649 
2650 	return -1;
2651 }
2652 
2653 static int
2654 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2655 			       unsigned long *start, unsigned long *last,
2656 			       bool *is_heap_stack)
2657 {
2658 	struct vm_area_struct *vma;
2659 	struct interval_tree_node *node;
2660 	struct rb_node *rb_node;
2661 	unsigned long start_limit, end_limit;
2662 
2663 	vma = vma_lookup(p->mm, addr << PAGE_SHIFT);
2664 	if (!vma) {
2665 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2666 		return -EFAULT;
2667 	}
2668 
2669 	*is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma);
2670 
2671 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
2672 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2673 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
2674 		    (unsigned long)ALIGN(addr + 1, 2UL << 8));
2675 	/* First range that starts after the fault address */
2676 	node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2677 	if (node) {
2678 		end_limit = min(end_limit, node->start);
2679 		/* Last range that ends before the fault address */
2680 		rb_node = rb_prev(&node->rb);
2681 	} else {
2682 		/* Last range must end before addr because
2683 		 * there was no range after addr
2684 		 */
2685 		rb_node = rb_last(&p->svms.objects.rb_root);
2686 	}
2687 	if (rb_node) {
2688 		node = container_of(rb_node, struct interval_tree_node, rb);
2689 		if (node->last >= addr) {
2690 			WARN(1, "Overlap with prev node and page fault addr\n");
2691 			return -EFAULT;
2692 		}
2693 		start_limit = max(start_limit, node->last + 1);
2694 	}
2695 
2696 	*start = start_limit;
2697 	*last = end_limit - 1;
2698 
2699 	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2700 		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2701 		 *start, *last, *is_heap_stack);
2702 
2703 	return 0;
2704 }
2705 
2706 static int
2707 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2708 			   uint64_t *bo_s, uint64_t *bo_l)
2709 {
2710 	struct amdgpu_bo_va_mapping *mapping;
2711 	struct interval_tree_node *node;
2712 	struct amdgpu_bo *bo = NULL;
2713 	unsigned long userptr;
2714 	uint32_t i;
2715 	int r;
2716 
2717 	for (i = 0; i < p->n_pdds; i++) {
2718 		struct amdgpu_vm *vm;
2719 
2720 		if (!p->pdds[i]->drm_priv)
2721 			continue;
2722 
2723 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2724 		r = amdgpu_bo_reserve(vm->root.bo, false);
2725 		if (r)
2726 			return r;
2727 
2728 		/* Check userptr by searching entire vm->va interval tree */
2729 		node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2730 		while (node) {
2731 			mapping = container_of((struct rb_node *)node,
2732 					       struct amdgpu_bo_va_mapping, rb);
2733 			bo = mapping->bo_va->base.bo;
2734 
2735 			if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2736 							 start << PAGE_SHIFT,
2737 							 last << PAGE_SHIFT,
2738 							 &userptr)) {
2739 				node = interval_tree_iter_next(node, 0, ~0ULL);
2740 				continue;
2741 			}
2742 
2743 			pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2744 				 start, last);
2745 			if (bo_s && bo_l) {
2746 				*bo_s = userptr >> PAGE_SHIFT;
2747 				*bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2748 			}
2749 			amdgpu_bo_unreserve(vm->root.bo);
2750 			return -EADDRINUSE;
2751 		}
2752 		amdgpu_bo_unreserve(vm->root.bo);
2753 	}
2754 	return 0;
2755 }
2756 
2757 static struct
2758 svm_range *svm_range_create_unregistered_range(struct kfd_node *node,
2759 						struct kfd_process *p,
2760 						struct mm_struct *mm,
2761 						int64_t addr)
2762 {
2763 	struct svm_range *prange = NULL;
2764 	unsigned long start, last;
2765 	uint32_t gpuid, gpuidx;
2766 	bool is_heap_stack;
2767 	uint64_t bo_s = 0;
2768 	uint64_t bo_l = 0;
2769 	int r;
2770 
2771 	if (svm_range_get_range_boundaries(p, addr, &start, &last,
2772 					   &is_heap_stack))
2773 		return NULL;
2774 
2775 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2776 	if (r != -EADDRINUSE)
2777 		r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2778 
2779 	if (r == -EADDRINUSE) {
2780 		if (addr >= bo_s && addr <= bo_l)
2781 			return NULL;
2782 
2783 		/* Create one page svm range if 2MB range overlapping */
2784 		start = addr;
2785 		last = addr;
2786 	}
2787 
2788 	prange = svm_range_new(&p->svms, start, last, true);
2789 	if (!prange) {
2790 		pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2791 		return NULL;
2792 	}
2793 	if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) {
2794 		pr_debug("failed to get gpuid from kgd\n");
2795 		svm_range_free(prange, true);
2796 		return NULL;
2797 	}
2798 
2799 	if (is_heap_stack)
2800 		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2801 
2802 	svm_range_add_to_svms(prange);
2803 	svm_range_add_notifier_locked(mm, prange);
2804 
2805 	return prange;
2806 }
2807 
2808 /* svm_range_skip_recover - decide if prange can be recovered
2809  * @prange: svm range structure
2810  *
2811  * GPU vm retry fault handle skip recover the range for cases:
2812  * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2813  *    deferred list work will drain the stale fault before free the prange.
2814  * 2. prange is on deferred list to add interval notifier after split, or
2815  * 3. prange is child range, it is split from parent prange, recover later
2816  *    after interval notifier is added.
2817  *
2818  * Return: true to skip recover, false to recover
2819  */
2820 static bool svm_range_skip_recover(struct svm_range *prange)
2821 {
2822 	struct svm_range_list *svms = prange->svms;
2823 
2824 	spin_lock(&svms->deferred_list_lock);
2825 	if (list_empty(&prange->deferred_list) &&
2826 	    list_empty(&prange->child_list)) {
2827 		spin_unlock(&svms->deferred_list_lock);
2828 		return false;
2829 	}
2830 	spin_unlock(&svms->deferred_list_lock);
2831 
2832 	if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2833 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2834 			 svms, prange, prange->start, prange->last);
2835 		return true;
2836 	}
2837 	if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2838 	    prange->work_item.op == SVM_OP_ADD_RANGE) {
2839 		pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2840 			 svms, prange, prange->start, prange->last);
2841 		return true;
2842 	}
2843 	return false;
2844 }
2845 
2846 static void
2847 svm_range_count_fault(struct kfd_node *node, struct kfd_process *p,
2848 		      int32_t gpuidx)
2849 {
2850 	struct kfd_process_device *pdd;
2851 
2852 	/* fault is on different page of same range
2853 	 * or fault is skipped to recover later
2854 	 * or fault is on invalid virtual address
2855 	 */
2856 	if (gpuidx == MAX_GPU_INSTANCE) {
2857 		uint32_t gpuid;
2858 		int r;
2859 
2860 		r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx);
2861 		if (r < 0)
2862 			return;
2863 	}
2864 
2865 	/* fault is recovered
2866 	 * or fault cannot recover because GPU no access on the range
2867 	 */
2868 	pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2869 	if (pdd)
2870 		WRITE_ONCE(pdd->faults, pdd->faults + 1);
2871 }
2872 
2873 static bool
2874 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2875 {
2876 	unsigned long requested = VM_READ;
2877 
2878 	if (write_fault)
2879 		requested |= VM_WRITE;
2880 
2881 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2882 		vma->vm_flags);
2883 	return (vma->vm_flags & requested) == requested;
2884 }
2885 
2886 int
2887 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2888 			uint32_t vmid, uint32_t node_id,
2889 			uint64_t addr, bool write_fault)
2890 {
2891 	struct mm_struct *mm = NULL;
2892 	struct svm_range_list *svms;
2893 	struct svm_range *prange;
2894 	struct kfd_process *p;
2895 	ktime_t timestamp = ktime_get_boottime();
2896 	struct kfd_node *node;
2897 	int32_t best_loc;
2898 	int32_t gpuidx = MAX_GPU_INSTANCE;
2899 	bool write_locked = false;
2900 	struct vm_area_struct *vma;
2901 	bool migration = false;
2902 	int r = 0;
2903 
2904 	if (!KFD_IS_SVM_API_SUPPORTED(adev)) {
2905 		pr_debug("device does not support SVM\n");
2906 		return -EFAULT;
2907 	}
2908 
2909 	p = kfd_lookup_process_by_pasid(pasid);
2910 	if (!p) {
2911 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2912 		return 0;
2913 	}
2914 	svms = &p->svms;
2915 
2916 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2917 
2918 	if (atomic_read(&svms->drain_pagefaults)) {
2919 		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2920 		r = 0;
2921 		goto out;
2922 	}
2923 
2924 	if (!p->xnack_enabled) {
2925 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2926 		r = -EFAULT;
2927 		goto out;
2928 	}
2929 
2930 	/* p->lead_thread is available as kfd_process_wq_release flush the work
2931 	 * before releasing task ref.
2932 	 */
2933 	mm = get_task_mm(p->lead_thread);
2934 	if (!mm) {
2935 		pr_debug("svms 0x%p failed to get mm\n", svms);
2936 		r = 0;
2937 		goto out;
2938 	}
2939 
2940 	node = kfd_node_by_irq_ids(adev, node_id, vmid);
2941 	if (!node) {
2942 		pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id,
2943 			 vmid);
2944 		r = -EFAULT;
2945 		goto out;
2946 	}
2947 	mmap_read_lock(mm);
2948 retry_write_locked:
2949 	mutex_lock(&svms->lock);
2950 	prange = svm_range_from_addr(svms, addr, NULL);
2951 	if (!prange) {
2952 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2953 			 svms, addr);
2954 		if (!write_locked) {
2955 			/* Need the write lock to create new range with MMU notifier.
2956 			 * Also flush pending deferred work to make sure the interval
2957 			 * tree is up to date before we add a new range
2958 			 */
2959 			mutex_unlock(&svms->lock);
2960 			mmap_read_unlock(mm);
2961 			mmap_write_lock(mm);
2962 			write_locked = true;
2963 			goto retry_write_locked;
2964 		}
2965 		prange = svm_range_create_unregistered_range(node, p, mm, addr);
2966 		if (!prange) {
2967 			pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2968 				 svms, addr);
2969 			mmap_write_downgrade(mm);
2970 			r = -EFAULT;
2971 			goto out_unlock_svms;
2972 		}
2973 	}
2974 	if (write_locked)
2975 		mmap_write_downgrade(mm);
2976 
2977 	mutex_lock(&prange->migrate_mutex);
2978 
2979 	if (svm_range_skip_recover(prange)) {
2980 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
2981 		r = 0;
2982 		goto out_unlock_range;
2983 	}
2984 
2985 	/* skip duplicate vm fault on different pages of same range */
2986 	if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
2987 				AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
2988 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2989 			 svms, prange->start, prange->last);
2990 		r = 0;
2991 		goto out_unlock_range;
2992 	}
2993 
2994 	/* __do_munmap removed VMA, return success as we are handling stale
2995 	 * retry fault.
2996 	 */
2997 	vma = vma_lookup(mm, addr << PAGE_SHIFT);
2998 	if (!vma) {
2999 		pr_debug("address 0x%llx VMA is removed\n", addr);
3000 		r = 0;
3001 		goto out_unlock_range;
3002 	}
3003 
3004 	if (!svm_fault_allowed(vma, write_fault)) {
3005 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
3006 			write_fault ? "write" : "read");
3007 		r = -EPERM;
3008 		goto out_unlock_range;
3009 	}
3010 
3011 	best_loc = svm_range_best_restore_location(prange, node, &gpuidx);
3012 	if (best_loc == -1) {
3013 		pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
3014 			 svms, prange->start, prange->last);
3015 		r = -EACCES;
3016 		goto out_unlock_range;
3017 	}
3018 
3019 	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
3020 		 svms, prange->start, prange->last, best_loc,
3021 		 prange->actual_loc);
3022 
3023 	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
3024 				       write_fault, timestamp);
3025 
3026 	if (prange->actual_loc != best_loc) {
3027 		migration = true;
3028 		if (best_loc) {
3029 			r = svm_migrate_to_vram(prange, best_loc, mm,
3030 					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
3031 			if (r) {
3032 				pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
3033 					 r, addr);
3034 				/* Fallback to system memory if migration to
3035 				 * VRAM failed
3036 				 */
3037 				if (prange->actual_loc)
3038 					r = svm_migrate_vram_to_ram(prange, mm,
3039 					   KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3040 					   NULL);
3041 				else
3042 					r = 0;
3043 			}
3044 		} else {
3045 			r = svm_migrate_vram_to_ram(prange, mm,
3046 					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3047 					NULL);
3048 		}
3049 		if (r) {
3050 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
3051 				 r, svms, prange->start, prange->last);
3052 			goto out_unlock_range;
3053 		}
3054 	}
3055 
3056 	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
3057 	if (r)
3058 		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
3059 			 r, svms, prange->start, prange->last);
3060 
3061 	kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
3062 				     migration);
3063 
3064 out_unlock_range:
3065 	mutex_unlock(&prange->migrate_mutex);
3066 out_unlock_svms:
3067 	mutex_unlock(&svms->lock);
3068 	mmap_read_unlock(mm);
3069 
3070 	svm_range_count_fault(node, p, gpuidx);
3071 
3072 	mmput(mm);
3073 out:
3074 	kfd_unref_process(p);
3075 
3076 	if (r == -EAGAIN) {
3077 		pr_debug("recover vm fault later\n");
3078 		amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3079 		r = 0;
3080 	}
3081 	return r;
3082 }
3083 
3084 int
3085 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
3086 {
3087 	struct svm_range *prange, *pchild;
3088 	uint64_t reserved_size = 0;
3089 	uint64_t size;
3090 	int r = 0;
3091 
3092 	pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
3093 
3094 	mutex_lock(&p->svms.lock);
3095 
3096 	list_for_each_entry(prange, &p->svms.list, list) {
3097 		svm_range_lock(prange);
3098 		list_for_each_entry(pchild, &prange->child_list, child_list) {
3099 			size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
3100 			if (xnack_enabled) {
3101 				amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3102 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3103 			} else {
3104 				r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3105 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3106 				if (r)
3107 					goto out_unlock;
3108 				reserved_size += size;
3109 			}
3110 		}
3111 
3112 		size = (prange->last - prange->start + 1) << PAGE_SHIFT;
3113 		if (xnack_enabled) {
3114 			amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3115 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3116 		} else {
3117 			r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3118 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3119 			if (r)
3120 				goto out_unlock;
3121 			reserved_size += size;
3122 		}
3123 out_unlock:
3124 		svm_range_unlock(prange);
3125 		if (r)
3126 			break;
3127 	}
3128 
3129 	if (r)
3130 		amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3131 					KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3132 	else
3133 		/* Change xnack mode must be inside svms lock, to avoid race with
3134 		 * svm_range_deferred_list_work unreserve memory in parallel.
3135 		 */
3136 		p->xnack_enabled = xnack_enabled;
3137 
3138 	mutex_unlock(&p->svms.lock);
3139 	return r;
3140 }
3141 
3142 void svm_range_list_fini(struct kfd_process *p)
3143 {
3144 	struct svm_range *prange;
3145 	struct svm_range *next;
3146 
3147 	pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3148 
3149 	cancel_delayed_work_sync(&p->svms.restore_work);
3150 
3151 	/* Ensure list work is finished before process is destroyed */
3152 	flush_work(&p->svms.deferred_list_work);
3153 
3154 	/*
3155 	 * Ensure no retry fault comes in afterwards, as page fault handler will
3156 	 * not find kfd process and take mm lock to recover fault.
3157 	 */
3158 	atomic_inc(&p->svms.drain_pagefaults);
3159 	svm_range_drain_retry_fault(&p->svms);
3160 
3161 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3162 		svm_range_unlink(prange);
3163 		svm_range_remove_notifier(prange);
3164 		svm_range_free(prange, true);
3165 	}
3166 
3167 	mutex_destroy(&p->svms.lock);
3168 
3169 	pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3170 }
3171 
3172 int svm_range_list_init(struct kfd_process *p)
3173 {
3174 	struct svm_range_list *svms = &p->svms;
3175 	int i;
3176 
3177 	svms->objects = RB_ROOT_CACHED;
3178 	mutex_init(&svms->lock);
3179 	INIT_LIST_HEAD(&svms->list);
3180 	atomic_set(&svms->evicted_ranges, 0);
3181 	atomic_set(&svms->drain_pagefaults, 0);
3182 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3183 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3184 	INIT_LIST_HEAD(&svms->deferred_range_list);
3185 	INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3186 	spin_lock_init(&svms->deferred_list_lock);
3187 
3188 	for (i = 0; i < p->n_pdds; i++)
3189 		if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev))
3190 			bitmap_set(svms->bitmap_supported, i, 1);
3191 
3192 	return 0;
3193 }
3194 
3195 /**
3196  * svm_range_check_vm - check if virtual address range mapped already
3197  * @p: current kfd_process
3198  * @start: range start address, in pages
3199  * @last: range last address, in pages
3200  * @bo_s: mapping start address in pages if address range already mapped
3201  * @bo_l: mapping last address in pages if address range already mapped
3202  *
3203  * The purpose is to avoid virtual address ranges already allocated by
3204  * kfd_ioctl_alloc_memory_of_gpu ioctl.
3205  * It looks for each pdd in the kfd_process.
3206  *
3207  * Context: Process context
3208  *
3209  * Return 0 - OK, if the range is not mapped.
3210  * Otherwise error code:
3211  * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3212  * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3213  * a signal. Release all buffer reservations and return to user-space.
3214  */
3215 static int
3216 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3217 		   uint64_t *bo_s, uint64_t *bo_l)
3218 {
3219 	struct amdgpu_bo_va_mapping *mapping;
3220 	struct interval_tree_node *node;
3221 	uint32_t i;
3222 	int r;
3223 
3224 	for (i = 0; i < p->n_pdds; i++) {
3225 		struct amdgpu_vm *vm;
3226 
3227 		if (!p->pdds[i]->drm_priv)
3228 			continue;
3229 
3230 		vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3231 		r = amdgpu_bo_reserve(vm->root.bo, false);
3232 		if (r)
3233 			return r;
3234 
3235 		node = interval_tree_iter_first(&vm->va, start, last);
3236 		if (node) {
3237 			pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3238 				 start, last);
3239 			mapping = container_of((struct rb_node *)node,
3240 					       struct amdgpu_bo_va_mapping, rb);
3241 			if (bo_s && bo_l) {
3242 				*bo_s = mapping->start;
3243 				*bo_l = mapping->last;
3244 			}
3245 			amdgpu_bo_unreserve(vm->root.bo);
3246 			return -EADDRINUSE;
3247 		}
3248 		amdgpu_bo_unreserve(vm->root.bo);
3249 	}
3250 
3251 	return 0;
3252 }
3253 
3254 /**
3255  * svm_range_is_valid - check if virtual address range is valid
3256  * @p: current kfd_process
3257  * @start: range start address, in pages
3258  * @size: range size, in pages
3259  *
3260  * Valid virtual address range means it belongs to one or more VMAs
3261  *
3262  * Context: Process context
3263  *
3264  * Return:
3265  *  0 - OK, otherwise error code
3266  */
3267 static int
3268 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3269 {
3270 	const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3271 	struct vm_area_struct *vma;
3272 	unsigned long end;
3273 	unsigned long start_unchg = start;
3274 
3275 	start <<= PAGE_SHIFT;
3276 	end = start + (size << PAGE_SHIFT);
3277 	do {
3278 		vma = vma_lookup(p->mm, start);
3279 		if (!vma || (vma->vm_flags & device_vma))
3280 			return -EFAULT;
3281 		start = min(end, vma->vm_end);
3282 	} while (start < end);
3283 
3284 	return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3285 				  NULL);
3286 }
3287 
3288 /**
3289  * svm_range_best_prefetch_location - decide the best prefetch location
3290  * @prange: svm range structure
3291  *
3292  * For xnack off:
3293  * If range map to single GPU, the best prefetch location is prefetch_loc, which
3294  * can be CPU or GPU.
3295  *
3296  * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3297  * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3298  * the best prefetch location is always CPU, because GPU can not have coherent
3299  * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3300  *
3301  * For xnack on:
3302  * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3303  * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3304  *
3305  * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3306  * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3307  * prefetch location is always CPU.
3308  *
3309  * Context: Process context
3310  *
3311  * Return:
3312  * 0 for CPU or GPU id
3313  */
3314 static uint32_t
3315 svm_range_best_prefetch_location(struct svm_range *prange)
3316 {
3317 	DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3318 	uint32_t best_loc = prange->prefetch_loc;
3319 	struct kfd_process_device *pdd;
3320 	struct kfd_node *bo_node;
3321 	struct kfd_process *p;
3322 	uint32_t gpuidx;
3323 
3324 	p = container_of(prange->svms, struct kfd_process, svms);
3325 
3326 	if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3327 		goto out;
3328 
3329 	bo_node = svm_range_get_node_by_id(prange, best_loc);
3330 	if (!bo_node) {
3331 		WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc);
3332 		best_loc = 0;
3333 		goto out;
3334 	}
3335 
3336 	if (bo_node->adev->gmc.is_app_apu) {
3337 		best_loc = 0;
3338 		goto out;
3339 	}
3340 
3341 	if (p->xnack_enabled)
3342 		bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3343 	else
3344 		bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3345 			  MAX_GPU_INSTANCE);
3346 
3347 	for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3348 		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3349 		if (!pdd) {
3350 			pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3351 			continue;
3352 		}
3353 
3354 		if (pdd->dev->adev == bo_node->adev)
3355 			continue;
3356 
3357 		if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) {
3358 			best_loc = 0;
3359 			break;
3360 		}
3361 	}
3362 
3363 out:
3364 	pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3365 		 p->xnack_enabled, &p->svms, prange->start, prange->last,
3366 		 best_loc);
3367 
3368 	return best_loc;
3369 }
3370 
3371 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3372  * @mm: current process mm_struct
3373  * @prange: svm range structure
3374  * @migrated: output, true if migration is triggered
3375  *
3376  * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3377  * from ram to vram.
3378  * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3379  * from vram to ram.
3380  *
3381  * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3382  * and restore work:
3383  * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3384  *    stops all queues, schedule restore work
3385  * 2. svm_range_restore_work wait for migration is done by
3386  *    a. svm_range_validate_vram takes prange->migrate_mutex
3387  *    b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3388  * 3. restore work update mappings of GPU, resume all queues.
3389  *
3390  * Context: Process context
3391  *
3392  * Return:
3393  * 0 - OK, otherwise - error code of migration
3394  */
3395 static int
3396 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3397 			    bool *migrated)
3398 {
3399 	uint32_t best_loc;
3400 	int r = 0;
3401 
3402 	*migrated = false;
3403 	best_loc = svm_range_best_prefetch_location(prange);
3404 
3405 	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3406 	    best_loc == prange->actual_loc)
3407 		return 0;
3408 
3409 	if (!best_loc) {
3410 		r = svm_migrate_vram_to_ram(prange, mm,
3411 					KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3412 		*migrated = !r;
3413 		return r;
3414 	}
3415 
3416 	r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3417 	*migrated = !r;
3418 
3419 	return r;
3420 }
3421 
3422 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3423 {
3424 	/* Dereferencing fence->svm_bo is safe here because the fence hasn't
3425 	 * signaled yet and we're under the protection of the fence->lock.
3426 	 * After the fence is signaled in svm_range_bo_release, we cannot get
3427 	 * here any more.
3428 	 *
3429 	 * Reference is dropped in svm_range_evict_svm_bo_worker.
3430 	 */
3431 	if (svm_bo_ref_unless_zero(fence->svm_bo)) {
3432 		WRITE_ONCE(fence->svm_bo->evicting, 1);
3433 		schedule_work(&fence->svm_bo->eviction_work);
3434 	}
3435 
3436 	return 0;
3437 }
3438 
3439 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3440 {
3441 	struct svm_range_bo *svm_bo;
3442 	struct mm_struct *mm;
3443 	int r = 0;
3444 
3445 	svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3446 
3447 	if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3448 		mm = svm_bo->eviction_fence->mm;
3449 	} else {
3450 		svm_range_bo_unref(svm_bo);
3451 		return;
3452 	}
3453 
3454 	mmap_read_lock(mm);
3455 	spin_lock(&svm_bo->list_lock);
3456 	while (!list_empty(&svm_bo->range_list) && !r) {
3457 		struct svm_range *prange =
3458 				list_first_entry(&svm_bo->range_list,
3459 						struct svm_range, svm_bo_list);
3460 		int retries = 3;
3461 
3462 		list_del_init(&prange->svm_bo_list);
3463 		spin_unlock(&svm_bo->list_lock);
3464 
3465 		pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3466 			 prange->start, prange->last);
3467 
3468 		mutex_lock(&prange->migrate_mutex);
3469 		do {
3470 			r = svm_migrate_vram_to_ram(prange, mm,
3471 					KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3472 		} while (!r && prange->actual_loc && --retries);
3473 
3474 		if (!r && prange->actual_loc)
3475 			pr_info_once("Migration failed during eviction");
3476 
3477 		if (!prange->actual_loc) {
3478 			mutex_lock(&prange->lock);
3479 			prange->svm_bo = NULL;
3480 			mutex_unlock(&prange->lock);
3481 		}
3482 		mutex_unlock(&prange->migrate_mutex);
3483 
3484 		spin_lock(&svm_bo->list_lock);
3485 	}
3486 	spin_unlock(&svm_bo->list_lock);
3487 	mmap_read_unlock(mm);
3488 	mmput(mm);
3489 
3490 	dma_fence_signal(&svm_bo->eviction_fence->base);
3491 
3492 	/* This is the last reference to svm_bo, after svm_range_vram_node_free
3493 	 * has been called in svm_migrate_vram_to_ram
3494 	 */
3495 	WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3496 	svm_range_bo_unref(svm_bo);
3497 }
3498 
3499 static int
3500 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3501 		   uint64_t start, uint64_t size, uint32_t nattr,
3502 		   struct kfd_ioctl_svm_attribute *attrs)
3503 {
3504 	struct amdkfd_process_info *process_info = p->kgd_process_info;
3505 	struct list_head update_list;
3506 	struct list_head insert_list;
3507 	struct list_head remove_list;
3508 	struct svm_range_list *svms;
3509 	struct svm_range *prange;
3510 	struct svm_range *next;
3511 	bool update_mapping = false;
3512 	bool flush_tlb;
3513 	int r, ret = 0;
3514 
3515 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3516 		 p->pasid, &p->svms, start, start + size - 1, size);
3517 
3518 	r = svm_range_check_attr(p, nattr, attrs);
3519 	if (r)
3520 		return r;
3521 
3522 	svms = &p->svms;
3523 
3524 	mutex_lock(&process_info->lock);
3525 
3526 	svm_range_list_lock_and_flush_work(svms, mm);
3527 
3528 	r = svm_range_is_valid(p, start, size);
3529 	if (r) {
3530 		pr_debug("invalid range r=%d\n", r);
3531 		mmap_write_unlock(mm);
3532 		goto out;
3533 	}
3534 
3535 	mutex_lock(&svms->lock);
3536 
3537 	/* Add new range and split existing ranges as needed */
3538 	r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3539 			  &insert_list, &remove_list);
3540 	if (r) {
3541 		mutex_unlock(&svms->lock);
3542 		mmap_write_unlock(mm);
3543 		goto out;
3544 	}
3545 	/* Apply changes as a transaction */
3546 	list_for_each_entry_safe(prange, next, &insert_list, list) {
3547 		svm_range_add_to_svms(prange);
3548 		svm_range_add_notifier_locked(mm, prange);
3549 	}
3550 	list_for_each_entry(prange, &update_list, update_list) {
3551 		svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3552 		/* TODO: unmap ranges from GPU that lost access */
3553 	}
3554 	list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3555 		pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3556 			 prange->svms, prange, prange->start,
3557 			 prange->last);
3558 		svm_range_unlink(prange);
3559 		svm_range_remove_notifier(prange);
3560 		svm_range_free(prange, false);
3561 	}
3562 
3563 	mmap_write_downgrade(mm);
3564 	/* Trigger migrations and revalidate and map to GPUs as needed. If
3565 	 * this fails we may be left with partially completed actions. There
3566 	 * is no clean way of rolling back to the previous state in such a
3567 	 * case because the rollback wouldn't be guaranteed to work either.
3568 	 */
3569 	list_for_each_entry(prange, &update_list, update_list) {
3570 		bool migrated;
3571 
3572 		mutex_lock(&prange->migrate_mutex);
3573 
3574 		r = svm_range_trigger_migration(mm, prange, &migrated);
3575 		if (r)
3576 			goto out_unlock_range;
3577 
3578 		if (migrated && (!p->xnack_enabled ||
3579 		    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3580 		    prange->mapped_to_gpu) {
3581 			pr_debug("restore_work will update mappings of GPUs\n");
3582 			mutex_unlock(&prange->migrate_mutex);
3583 			continue;
3584 		}
3585 
3586 		if (!migrated && !update_mapping) {
3587 			mutex_unlock(&prange->migrate_mutex);
3588 			continue;
3589 		}
3590 
3591 		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3592 
3593 		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3594 					       true, true, flush_tlb);
3595 		if (r)
3596 			pr_debug("failed %d to map svm range\n", r);
3597 
3598 out_unlock_range:
3599 		mutex_unlock(&prange->migrate_mutex);
3600 		if (r)
3601 			ret = r;
3602 	}
3603 
3604 	dynamic_svm_range_dump(svms);
3605 
3606 	mutex_unlock(&svms->lock);
3607 	mmap_read_unlock(mm);
3608 out:
3609 	mutex_unlock(&process_info->lock);
3610 
3611 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3612 		 &p->svms, start, start + size - 1, r);
3613 
3614 	return ret ? ret : r;
3615 }
3616 
3617 static int
3618 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3619 		   uint64_t start, uint64_t size, uint32_t nattr,
3620 		   struct kfd_ioctl_svm_attribute *attrs)
3621 {
3622 	DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3623 	DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3624 	bool get_preferred_loc = false;
3625 	bool get_prefetch_loc = false;
3626 	bool get_granularity = false;
3627 	bool get_accessible = false;
3628 	bool get_flags = false;
3629 	uint64_t last = start + size - 1UL;
3630 	uint8_t granularity = 0xff;
3631 	struct interval_tree_node *node;
3632 	struct svm_range_list *svms;
3633 	struct svm_range *prange;
3634 	uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3635 	uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3636 	uint32_t flags_and = 0xffffffff;
3637 	uint32_t flags_or = 0;
3638 	int gpuidx;
3639 	uint32_t i;
3640 	int r = 0;
3641 
3642 	pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3643 		 start + size - 1, nattr);
3644 
3645 	/* Flush pending deferred work to avoid racing with deferred actions from
3646 	 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3647 	 * can still race with get_attr because we don't hold the mmap lock. But that
3648 	 * would be a race condition in the application anyway, and undefined
3649 	 * behaviour is acceptable in that case.
3650 	 */
3651 	flush_work(&p->svms.deferred_list_work);
3652 
3653 	mmap_read_lock(mm);
3654 	r = svm_range_is_valid(p, start, size);
3655 	mmap_read_unlock(mm);
3656 	if (r) {
3657 		pr_debug("invalid range r=%d\n", r);
3658 		return r;
3659 	}
3660 
3661 	for (i = 0; i < nattr; i++) {
3662 		switch (attrs[i].type) {
3663 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3664 			get_preferred_loc = true;
3665 			break;
3666 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3667 			get_prefetch_loc = true;
3668 			break;
3669 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3670 			get_accessible = true;
3671 			break;
3672 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3673 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3674 			get_flags = true;
3675 			break;
3676 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3677 			get_granularity = true;
3678 			break;
3679 		case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3680 		case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3681 			fallthrough;
3682 		default:
3683 			pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3684 			return -EINVAL;
3685 		}
3686 	}
3687 
3688 	svms = &p->svms;
3689 
3690 	mutex_lock(&svms->lock);
3691 
3692 	node = interval_tree_iter_first(&svms->objects, start, last);
3693 	if (!node) {
3694 		pr_debug("range attrs not found return default values\n");
3695 		svm_range_set_default_attributes(&location, &prefetch_loc,
3696 						 &granularity, &flags_and);
3697 		flags_or = flags_and;
3698 		if (p->xnack_enabled)
3699 			bitmap_copy(bitmap_access, svms->bitmap_supported,
3700 				    MAX_GPU_INSTANCE);
3701 		else
3702 			bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3703 		bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3704 		goto fill_values;
3705 	}
3706 	bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3707 	bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3708 
3709 	while (node) {
3710 		struct interval_tree_node *next;
3711 
3712 		prange = container_of(node, struct svm_range, it_node);
3713 		next = interval_tree_iter_next(node, start, last);
3714 
3715 		if (get_preferred_loc) {
3716 			if (prange->preferred_loc ==
3717 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3718 			    (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3719 			     location != prange->preferred_loc)) {
3720 				location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3721 				get_preferred_loc = false;
3722 			} else {
3723 				location = prange->preferred_loc;
3724 			}
3725 		}
3726 		if (get_prefetch_loc) {
3727 			if (prange->prefetch_loc ==
3728 					KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3729 			    (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3730 			     prefetch_loc != prange->prefetch_loc)) {
3731 				prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3732 				get_prefetch_loc = false;
3733 			} else {
3734 				prefetch_loc = prange->prefetch_loc;
3735 			}
3736 		}
3737 		if (get_accessible) {
3738 			bitmap_and(bitmap_access, bitmap_access,
3739 				   prange->bitmap_access, MAX_GPU_INSTANCE);
3740 			bitmap_and(bitmap_aip, bitmap_aip,
3741 				   prange->bitmap_aip, MAX_GPU_INSTANCE);
3742 		}
3743 		if (get_flags) {
3744 			flags_and &= prange->flags;
3745 			flags_or |= prange->flags;
3746 		}
3747 
3748 		if (get_granularity && prange->granularity < granularity)
3749 			granularity = prange->granularity;
3750 
3751 		node = next;
3752 	}
3753 fill_values:
3754 	mutex_unlock(&svms->lock);
3755 
3756 	for (i = 0; i < nattr; i++) {
3757 		switch (attrs[i].type) {
3758 		case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3759 			attrs[i].value = location;
3760 			break;
3761 		case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3762 			attrs[i].value = prefetch_loc;
3763 			break;
3764 		case KFD_IOCTL_SVM_ATTR_ACCESS:
3765 			gpuidx = kfd_process_gpuidx_from_gpuid(p,
3766 							       attrs[i].value);
3767 			if (gpuidx < 0) {
3768 				pr_debug("invalid gpuid %x\n", attrs[i].value);
3769 				return -EINVAL;
3770 			}
3771 			if (test_bit(gpuidx, bitmap_access))
3772 				attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3773 			else if (test_bit(gpuidx, bitmap_aip))
3774 				attrs[i].type =
3775 					KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3776 			else
3777 				attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3778 			break;
3779 		case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3780 			attrs[i].value = flags_and;
3781 			break;
3782 		case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3783 			attrs[i].value = ~flags_or;
3784 			break;
3785 		case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3786 			attrs[i].value = (uint32_t)granularity;
3787 			break;
3788 		}
3789 	}
3790 
3791 	return 0;
3792 }
3793 
3794 int kfd_criu_resume_svm(struct kfd_process *p)
3795 {
3796 	struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3797 	int nattr_common = 4, nattr_accessibility = 1;
3798 	struct criu_svm_metadata *criu_svm_md = NULL;
3799 	struct svm_range_list *svms = &p->svms;
3800 	struct criu_svm_metadata *next = NULL;
3801 	uint32_t set_flags = 0xffffffff;
3802 	int i, j, num_attrs, ret = 0;
3803 	uint64_t set_attr_size;
3804 	struct mm_struct *mm;
3805 
3806 	if (list_empty(&svms->criu_svm_metadata_list)) {
3807 		pr_debug("No SVM data from CRIU restore stage 2\n");
3808 		return ret;
3809 	}
3810 
3811 	mm = get_task_mm(p->lead_thread);
3812 	if (!mm) {
3813 		pr_err("failed to get mm for the target process\n");
3814 		return -ESRCH;
3815 	}
3816 
3817 	num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3818 
3819 	i = j = 0;
3820 	list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3821 		pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3822 			 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3823 
3824 		for (j = 0; j < num_attrs; j++) {
3825 			pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3826 				 i, j, criu_svm_md->data.attrs[j].type,
3827 				 i, j, criu_svm_md->data.attrs[j].value);
3828 			switch (criu_svm_md->data.attrs[j].type) {
3829 			/* During Checkpoint operation, the query for
3830 			 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3831 			 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3832 			 * not used by the range which was checkpointed. Care
3833 			 * must be taken to not restore with an invalid value
3834 			 * otherwise the gpuidx value will be invalid and
3835 			 * set_attr would eventually fail so just replace those
3836 			 * with another dummy attribute such as
3837 			 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3838 			 */
3839 			case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3840 				if (criu_svm_md->data.attrs[j].value ==
3841 				    KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3842 					criu_svm_md->data.attrs[j].type =
3843 						KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3844 					criu_svm_md->data.attrs[j].value = 0;
3845 				}
3846 				break;
3847 			case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3848 				set_flags = criu_svm_md->data.attrs[j].value;
3849 				break;
3850 			default:
3851 				break;
3852 			}
3853 		}
3854 
3855 		/* CLR_FLAGS is not available via get_attr during checkpoint but
3856 		 * it needs to be inserted before restoring the ranges so
3857 		 * allocate extra space for it before calling set_attr
3858 		 */
3859 		set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3860 						(num_attrs + 1);
3861 		set_attr_new = krealloc(set_attr, set_attr_size,
3862 					    GFP_KERNEL);
3863 		if (!set_attr_new) {
3864 			ret = -ENOMEM;
3865 			goto exit;
3866 		}
3867 		set_attr = set_attr_new;
3868 
3869 		memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3870 					sizeof(struct kfd_ioctl_svm_attribute));
3871 		set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3872 		set_attr[num_attrs].value = ~set_flags;
3873 
3874 		ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3875 					 criu_svm_md->data.size, num_attrs + 1,
3876 					 set_attr);
3877 		if (ret) {
3878 			pr_err("CRIU: failed to set range attributes\n");
3879 			goto exit;
3880 		}
3881 
3882 		i++;
3883 	}
3884 exit:
3885 	kfree(set_attr);
3886 	list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3887 		pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3888 						criu_svm_md->data.start_addr);
3889 		kfree(criu_svm_md);
3890 	}
3891 
3892 	mmput(mm);
3893 	return ret;
3894 
3895 }
3896 
3897 int kfd_criu_restore_svm(struct kfd_process *p,
3898 			 uint8_t __user *user_priv_ptr,
3899 			 uint64_t *priv_data_offset,
3900 			 uint64_t max_priv_data_size)
3901 {
3902 	uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3903 	int nattr_common = 4, nattr_accessibility = 1;
3904 	struct criu_svm_metadata *criu_svm_md = NULL;
3905 	struct svm_range_list *svms = &p->svms;
3906 	uint32_t num_devices;
3907 	int ret = 0;
3908 
3909 	num_devices = p->n_pdds;
3910 	/* Handle one SVM range object at a time, also the number of gpus are
3911 	 * assumed to be same on the restore node, checking must be done while
3912 	 * evaluating the topology earlier
3913 	 */
3914 
3915 	svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3916 		(nattr_common + nattr_accessibility * num_devices);
3917 	svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3918 
3919 	svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3920 								svm_attrs_size;
3921 
3922 	criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3923 	if (!criu_svm_md) {
3924 		pr_err("failed to allocate memory to store svm metadata\n");
3925 		return -ENOMEM;
3926 	}
3927 	if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3928 		ret = -EINVAL;
3929 		goto exit;
3930 	}
3931 
3932 	ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3933 			     svm_priv_data_size);
3934 	if (ret) {
3935 		ret = -EFAULT;
3936 		goto exit;
3937 	}
3938 	*priv_data_offset += svm_priv_data_size;
3939 
3940 	list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3941 
3942 	return 0;
3943 
3944 
3945 exit:
3946 	kfree(criu_svm_md);
3947 	return ret;
3948 }
3949 
3950 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3951 		       uint64_t *svm_priv_data_size)
3952 {
3953 	uint64_t total_size, accessibility_size, common_attr_size;
3954 	int nattr_common = 4, nattr_accessibility = 1;
3955 	int num_devices = p->n_pdds;
3956 	struct svm_range_list *svms;
3957 	struct svm_range *prange;
3958 	uint32_t count = 0;
3959 
3960 	*svm_priv_data_size = 0;
3961 
3962 	svms = &p->svms;
3963 	if (!svms)
3964 		return -EINVAL;
3965 
3966 	mutex_lock(&svms->lock);
3967 	list_for_each_entry(prange, &svms->list, list) {
3968 		pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3969 			 prange, prange->start, prange->npages,
3970 			 prange->start + prange->npages - 1);
3971 		count++;
3972 	}
3973 	mutex_unlock(&svms->lock);
3974 
3975 	*num_svm_ranges = count;
3976 	/* Only the accessbility attributes need to be queried for all the gpus
3977 	 * individually, remaining ones are spanned across the entire process
3978 	 * regardless of the various gpu nodes. Of the remaining attributes,
3979 	 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3980 	 *
3981 	 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3982 	 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3983 	 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3984 	 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3985 	 *
3986 	 * ** ACCESSBILITY ATTRIBUTES **
3987 	 * (Considered as one, type is altered during query, value is gpuid)
3988 	 * KFD_IOCTL_SVM_ATTR_ACCESS
3989 	 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3990 	 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3991 	 */
3992 	if (*num_svm_ranges > 0) {
3993 		common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3994 			nattr_common;
3995 		accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3996 			nattr_accessibility * num_devices;
3997 
3998 		total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3999 			common_attr_size + accessibility_size;
4000 
4001 		*svm_priv_data_size = *num_svm_ranges * total_size;
4002 	}
4003 
4004 	pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
4005 		 *svm_priv_data_size);
4006 	return 0;
4007 }
4008 
4009 int kfd_criu_checkpoint_svm(struct kfd_process *p,
4010 			    uint8_t __user *user_priv_data,
4011 			    uint64_t *priv_data_offset)
4012 {
4013 	struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
4014 	struct kfd_ioctl_svm_attribute *query_attr = NULL;
4015 	uint64_t svm_priv_data_size, query_attr_size = 0;
4016 	int index, nattr_common = 4, ret = 0;
4017 	struct svm_range_list *svms;
4018 	int num_devices = p->n_pdds;
4019 	struct svm_range *prange;
4020 	struct mm_struct *mm;
4021 
4022 	svms = &p->svms;
4023 	if (!svms)
4024 		return -EINVAL;
4025 
4026 	mm = get_task_mm(p->lead_thread);
4027 	if (!mm) {
4028 		pr_err("failed to get mm for the target process\n");
4029 		return -ESRCH;
4030 	}
4031 
4032 	query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
4033 				(nattr_common + num_devices);
4034 
4035 	query_attr = kzalloc(query_attr_size, GFP_KERNEL);
4036 	if (!query_attr) {
4037 		ret = -ENOMEM;
4038 		goto exit;
4039 	}
4040 
4041 	query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
4042 	query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
4043 	query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
4044 	query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
4045 
4046 	for (index = 0; index < num_devices; index++) {
4047 		struct kfd_process_device *pdd = p->pdds[index];
4048 
4049 		query_attr[index + nattr_common].type =
4050 			KFD_IOCTL_SVM_ATTR_ACCESS;
4051 		query_attr[index + nattr_common].value = pdd->user_gpu_id;
4052 	}
4053 
4054 	svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
4055 
4056 	svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
4057 	if (!svm_priv) {
4058 		ret = -ENOMEM;
4059 		goto exit_query;
4060 	}
4061 
4062 	index = 0;
4063 	list_for_each_entry(prange, &svms->list, list) {
4064 
4065 		svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
4066 		svm_priv->start_addr = prange->start;
4067 		svm_priv->size = prange->npages;
4068 		memcpy(&svm_priv->attrs, query_attr, query_attr_size);
4069 		pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
4070 			 prange, prange->start, prange->npages,
4071 			 prange->start + prange->npages - 1,
4072 			 prange->npages * PAGE_SIZE);
4073 
4074 		ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
4075 					 svm_priv->size,
4076 					 (nattr_common + num_devices),
4077 					 svm_priv->attrs);
4078 		if (ret) {
4079 			pr_err("CRIU: failed to obtain range attributes\n");
4080 			goto exit_priv;
4081 		}
4082 
4083 		if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
4084 				 svm_priv_data_size)) {
4085 			pr_err("Failed to copy svm priv to user\n");
4086 			ret = -EFAULT;
4087 			goto exit_priv;
4088 		}
4089 
4090 		*priv_data_offset += svm_priv_data_size;
4091 
4092 	}
4093 
4094 
4095 exit_priv:
4096 	kfree(svm_priv);
4097 exit_query:
4098 	kfree(query_attr);
4099 exit:
4100 	mmput(mm);
4101 	return ret;
4102 }
4103 
4104 int
4105 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
4106 	  uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
4107 {
4108 	struct mm_struct *mm = current->mm;
4109 	int r;
4110 
4111 	start >>= PAGE_SHIFT;
4112 	size >>= PAGE_SHIFT;
4113 
4114 	switch (op) {
4115 	case KFD_IOCTL_SVM_OP_SET_ATTR:
4116 		r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4117 		break;
4118 	case KFD_IOCTL_SVM_OP_GET_ATTR:
4119 		r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);
4120 		break;
4121 	default:
4122 		r = EINVAL;
4123 		break;
4124 	}
4125 
4126 	return r;
4127 }
4128