1 // SPDX-License-Identifier: GPL-2.0 OR MIT
3 * Copyright 2020-2021 Advanced Micro Devices, Inc.
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:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
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.
24 #include <linux/types.h>
25 #include <linux/sched/task.h>
26 #include "amdgpu_sync.h"
27 #include "amdgpu_object.h"
28 #include "amdgpu_vm.h"
29 #include "amdgpu_mn.h"
31 #include "amdgpu_xgmi.h"
34 #include "kfd_migrate.h"
39 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
41 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
43 /* Long enough to ensure no retry fault comes after svm range is restored and
44 * page table is updated.
46 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING 2000
48 struct criu_svm_metadata {
49 struct list_head list;
50 struct kfd_criu_svm_range_priv_data data;
53 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
55 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
56 const struct mmu_notifier_range *range,
57 unsigned long cur_seq);
59 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
60 uint64_t *bo_s, uint64_t *bo_l);
61 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
62 .invalidate = svm_range_cpu_invalidate_pagetables,
66 * svm_range_unlink - unlink svm_range from lists and interval tree
67 * @prange: svm range structure to be removed
69 * Remove the svm_range from the svms and svm_bo lists and the svms
72 * Context: The caller must hold svms->lock
74 static void svm_range_unlink(struct svm_range *prange)
76 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
77 prange, prange->start, prange->last);
80 spin_lock(&prange->svm_bo->list_lock);
81 list_del(&prange->svm_bo_list);
82 spin_unlock(&prange->svm_bo->list_lock);
85 list_del(&prange->list);
86 if (prange->it_node.start != 0 && prange->it_node.last != 0)
87 interval_tree_remove(&prange->it_node, &prange->svms->objects);
91 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
93 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
94 prange, prange->start, prange->last);
96 mmu_interval_notifier_insert_locked(&prange->notifier, mm,
97 prange->start << PAGE_SHIFT,
98 prange->npages << PAGE_SHIFT,
103 * svm_range_add_to_svms - add svm range to svms
104 * @prange: svm range structure to be added
106 * Add the svm range to svms interval tree and link list
108 * Context: The caller must hold svms->lock
110 static void svm_range_add_to_svms(struct svm_range *prange)
112 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
113 prange, prange->start, prange->last);
115 list_move_tail(&prange->list, &prange->svms->list);
116 prange->it_node.start = prange->start;
117 prange->it_node.last = prange->last;
118 interval_tree_insert(&prange->it_node, &prange->svms->objects);
121 static void svm_range_remove_notifier(struct svm_range *prange)
123 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
124 prange->svms, prange,
125 prange->notifier.interval_tree.start >> PAGE_SHIFT,
126 prange->notifier.interval_tree.last >> PAGE_SHIFT);
128 if (prange->notifier.interval_tree.start != 0 &&
129 prange->notifier.interval_tree.last != 0)
130 mmu_interval_notifier_remove(&prange->notifier);
134 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
136 return dma_addr && !dma_mapping_error(dev, dma_addr) &&
137 !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
141 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
142 unsigned long offset, unsigned long npages,
143 unsigned long *hmm_pfns, uint32_t gpuidx)
145 enum dma_data_direction dir = DMA_BIDIRECTIONAL;
146 dma_addr_t *addr = prange->dma_addr[gpuidx];
147 struct device *dev = adev->dev;
152 addr = kvmalloc_array(prange->npages, sizeof(*addr),
153 GFP_KERNEL | __GFP_ZERO);
156 prange->dma_addr[gpuidx] = addr;
160 for (i = 0; i < npages; i++) {
161 if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
162 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
164 page = hmm_pfn_to_page(hmm_pfns[i]);
165 if (is_zone_device_page(page)) {
166 struct amdgpu_device *bo_adev =
167 amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
169 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
170 bo_adev->vm_manager.vram_base_offset -
171 bo_adev->kfd.dev->pgmap.range.start;
172 addr[i] |= SVM_RANGE_VRAM_DOMAIN;
173 pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
176 addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
177 r = dma_mapping_error(dev, addr[i]);
179 dev_err(dev, "failed %d dma_map_page\n", r);
182 pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
183 addr[i] >> PAGE_SHIFT, page_to_pfn(page));
189 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
190 unsigned long offset, unsigned long npages,
191 unsigned long *hmm_pfns)
193 struct kfd_process *p;
197 p = container_of(prange->svms, struct kfd_process, svms);
199 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
200 struct kfd_process_device *pdd;
202 pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
203 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
205 pr_debug("failed to find device idx %d\n", gpuidx);
209 r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
218 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
219 unsigned long offset, unsigned long npages)
221 enum dma_data_direction dir = DMA_BIDIRECTIONAL;
227 for (i = offset; i < offset + npages; i++) {
228 if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
230 pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
231 dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
236 void svm_range_free_dma_mappings(struct svm_range *prange)
238 struct kfd_process_device *pdd;
239 dma_addr_t *dma_addr;
241 struct kfd_process *p;
244 p = container_of(prange->svms, struct kfd_process, svms);
246 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
247 dma_addr = prange->dma_addr[gpuidx];
251 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
253 pr_debug("failed to find device idx %d\n", gpuidx);
256 dev = &pdd->dev->pdev->dev;
257 svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
259 prange->dma_addr[gpuidx] = NULL;
263 static void svm_range_free(struct svm_range *prange)
265 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
266 prange->start, prange->last);
268 svm_range_vram_node_free(prange);
269 svm_range_free_dma_mappings(prange);
270 mutex_destroy(&prange->lock);
271 mutex_destroy(&prange->migrate_mutex);
276 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
277 uint8_t *granularity, uint32_t *flags)
279 *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
280 *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
283 KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
287 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
290 uint64_t size = last - start + 1;
291 struct svm_range *prange;
292 struct kfd_process *p;
294 prange = kzalloc(sizeof(*prange), GFP_KERNEL);
297 prange->npages = size;
299 prange->start = start;
301 INIT_LIST_HEAD(&prange->list);
302 INIT_LIST_HEAD(&prange->update_list);
303 INIT_LIST_HEAD(&prange->svm_bo_list);
304 INIT_LIST_HEAD(&prange->deferred_list);
305 INIT_LIST_HEAD(&prange->child_list);
306 atomic_set(&prange->invalid, 0);
307 prange->validate_timestamp = 0;
308 mutex_init(&prange->migrate_mutex);
309 mutex_init(&prange->lock);
311 p = container_of(svms, struct kfd_process, svms);
312 if (p->xnack_enabled)
313 bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
316 svm_range_set_default_attributes(&prange->preferred_loc,
317 &prange->prefetch_loc,
318 &prange->granularity, &prange->flags);
320 pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
325 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
327 if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
333 static void svm_range_bo_release(struct kref *kref)
335 struct svm_range_bo *svm_bo;
337 svm_bo = container_of(kref, struct svm_range_bo, kref);
338 pr_debug("svm_bo 0x%p\n", svm_bo);
340 spin_lock(&svm_bo->list_lock);
341 while (!list_empty(&svm_bo->range_list)) {
342 struct svm_range *prange =
343 list_first_entry(&svm_bo->range_list,
344 struct svm_range, svm_bo_list);
345 /* list_del_init tells a concurrent svm_range_vram_node_new when
346 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
348 list_del_init(&prange->svm_bo_list);
349 spin_unlock(&svm_bo->list_lock);
351 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
352 prange->start, prange->last);
353 mutex_lock(&prange->lock);
354 prange->svm_bo = NULL;
355 mutex_unlock(&prange->lock);
357 spin_lock(&svm_bo->list_lock);
359 spin_unlock(&svm_bo->list_lock);
360 if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
361 /* We're not in the eviction worker.
362 * Signal the fence and synchronize with any
363 * pending eviction work.
365 dma_fence_signal(&svm_bo->eviction_fence->base);
366 cancel_work_sync(&svm_bo->eviction_work);
368 dma_fence_put(&svm_bo->eviction_fence->base);
369 amdgpu_bo_unref(&svm_bo->bo);
373 static void svm_range_bo_wq_release(struct work_struct *work)
375 struct svm_range_bo *svm_bo;
377 svm_bo = container_of(work, struct svm_range_bo, release_work);
378 svm_range_bo_release(&svm_bo->kref);
381 static void svm_range_bo_release_async(struct kref *kref)
383 struct svm_range_bo *svm_bo;
385 svm_bo = container_of(kref, struct svm_range_bo, kref);
386 pr_debug("svm_bo 0x%p\n", svm_bo);
387 INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
388 schedule_work(&svm_bo->release_work);
391 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
393 kref_put(&svm_bo->kref, svm_range_bo_release_async);
396 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
399 kref_put(&svm_bo->kref, svm_range_bo_release);
403 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
405 struct amdgpu_device *bo_adev;
407 mutex_lock(&prange->lock);
408 if (!prange->svm_bo) {
409 mutex_unlock(&prange->lock);
412 if (prange->ttm_res) {
413 /* We still have a reference, all is well */
414 mutex_unlock(&prange->lock);
417 if (svm_bo_ref_unless_zero(prange->svm_bo)) {
419 * Migrate from GPU to GPU, remove range from source bo_adev
420 * svm_bo range list, and return false to allocate svm_bo from
423 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
424 if (bo_adev != adev) {
425 mutex_unlock(&prange->lock);
427 spin_lock(&prange->svm_bo->list_lock);
428 list_del_init(&prange->svm_bo_list);
429 spin_unlock(&prange->svm_bo->list_lock);
431 svm_range_bo_unref(prange->svm_bo);
434 if (READ_ONCE(prange->svm_bo->evicting)) {
436 struct svm_range_bo *svm_bo;
437 /* The BO is getting evicted,
438 * we need to get a new one
440 mutex_unlock(&prange->lock);
441 svm_bo = prange->svm_bo;
442 f = dma_fence_get(&svm_bo->eviction_fence->base);
443 svm_range_bo_unref(prange->svm_bo);
444 /* wait for the fence to avoid long spin-loop
445 * at list_empty_careful
447 dma_fence_wait(f, false);
450 /* The BO was still around and we got
451 * a new reference to it
453 mutex_unlock(&prange->lock);
454 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
455 prange->svms, prange->start, prange->last);
457 prange->ttm_res = prange->svm_bo->bo->tbo.resource;
462 mutex_unlock(&prange->lock);
465 /* We need a new svm_bo. Spin-loop to wait for concurrent
466 * svm_range_bo_release to finish removing this range from
467 * its range list. After this, it is safe to reuse the
468 * svm_bo pointer and svm_bo_list head.
470 while (!list_empty_careful(&prange->svm_bo_list))
476 static struct svm_range_bo *svm_range_bo_new(void)
478 struct svm_range_bo *svm_bo;
480 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
484 kref_init(&svm_bo->kref);
485 INIT_LIST_HEAD(&svm_bo->range_list);
486 spin_lock_init(&svm_bo->list_lock);
492 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
495 struct amdgpu_bo_param bp;
496 struct svm_range_bo *svm_bo;
497 struct amdgpu_bo_user *ubo;
498 struct amdgpu_bo *bo;
499 struct kfd_process *p;
500 struct mm_struct *mm;
503 p = container_of(prange->svms, struct kfd_process, svms);
504 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
505 prange->start, prange->last);
507 if (svm_range_validate_svm_bo(adev, prange))
510 svm_bo = svm_range_bo_new();
512 pr_debug("failed to alloc svm bo\n");
515 mm = get_task_mm(p->lead_thread);
517 pr_debug("failed to get mm\n");
521 svm_bo->svms = prange->svms;
522 svm_bo->eviction_fence =
523 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
527 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
528 svm_bo->evicting = 0;
529 memset(&bp, 0, sizeof(bp));
530 bp.size = prange->npages * PAGE_SIZE;
531 bp.byte_align = PAGE_SIZE;
532 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
533 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
534 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
535 bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO;
536 bp.type = ttm_bo_type_device;
539 r = amdgpu_bo_create_user(adev, &bp, &ubo);
541 pr_debug("failed %d to create bo\n", r);
542 goto create_bo_failed;
545 r = amdgpu_bo_reserve(bo, true);
547 pr_debug("failed %d to reserve bo\n", r);
548 goto reserve_bo_failed;
551 r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
553 pr_debug("failed %d to reserve bo\n", r);
554 amdgpu_bo_unreserve(bo);
555 goto reserve_bo_failed;
557 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
559 amdgpu_bo_unreserve(bo);
562 prange->svm_bo = svm_bo;
563 prange->ttm_res = bo->tbo.resource;
566 spin_lock(&svm_bo->list_lock);
567 list_add(&prange->svm_bo_list, &svm_bo->range_list);
568 spin_unlock(&svm_bo->list_lock);
573 amdgpu_bo_unref(&bo);
575 dma_fence_put(&svm_bo->eviction_fence->base);
577 prange->ttm_res = NULL;
582 void svm_range_vram_node_free(struct svm_range *prange)
584 svm_range_bo_unref(prange->svm_bo);
585 prange->ttm_res = NULL;
588 struct amdgpu_device *
589 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
591 struct kfd_process_device *pdd;
592 struct kfd_process *p;
595 p = container_of(prange->svms, struct kfd_process, svms);
597 gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
599 pr_debug("failed to get device by id 0x%x\n", gpu_id);
602 pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
604 pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
608 return pdd->dev->adev;
611 struct kfd_process_device *
612 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
614 struct kfd_process *p;
615 int32_t gpu_idx, gpuid;
618 p = container_of(prange->svms, struct kfd_process, svms);
620 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx);
622 pr_debug("failed to get device id by adev %p\n", adev);
626 return kfd_process_device_from_gpuidx(p, gpu_idx);
629 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
631 struct ttm_operation_ctx ctx = { false, false };
633 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
635 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
639 svm_range_check_attr(struct kfd_process *p,
640 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
644 for (i = 0; i < nattr; i++) {
645 uint32_t val = attrs[i].value;
646 int gpuidx = MAX_GPU_INSTANCE;
648 switch (attrs[i].type) {
649 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
650 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
651 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
652 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
654 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
655 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
656 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
658 case KFD_IOCTL_SVM_ATTR_ACCESS:
659 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
660 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
661 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
663 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
665 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
667 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
670 pr_debug("unknown attr type 0x%x\n", attrs[i].type);
675 pr_debug("no GPU 0x%x found\n", val);
677 } else if (gpuidx < MAX_GPU_INSTANCE &&
678 !test_bit(gpuidx, p->svms.bitmap_supported)) {
679 pr_debug("GPU 0x%x not supported\n", val);
688 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
689 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
694 for (i = 0; i < nattr; i++) {
695 switch (attrs[i].type) {
696 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
697 prange->preferred_loc = attrs[i].value;
699 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
700 prange->prefetch_loc = attrs[i].value;
702 case KFD_IOCTL_SVM_ATTR_ACCESS:
703 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
704 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
705 gpuidx = kfd_process_gpuidx_from_gpuid(p,
707 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
708 bitmap_clear(prange->bitmap_access, gpuidx, 1);
709 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
710 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
711 bitmap_set(prange->bitmap_access, gpuidx, 1);
712 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
714 bitmap_clear(prange->bitmap_access, gpuidx, 1);
715 bitmap_set(prange->bitmap_aip, gpuidx, 1);
718 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
719 prange->flags |= attrs[i].value;
721 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
722 prange->flags &= ~attrs[i].value;
724 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
725 prange->granularity = attrs[i].value;
728 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
734 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
735 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
740 for (i = 0; i < nattr; i++) {
741 switch (attrs[i].type) {
742 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
743 if (prange->preferred_loc != attrs[i].value)
746 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
747 /* Prefetch should always trigger a migration even
748 * if the value of the attribute didn't change.
751 case KFD_IOCTL_SVM_ATTR_ACCESS:
752 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
753 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
754 gpuidx = kfd_process_gpuidx_from_gpuid(p,
756 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
757 if (test_bit(gpuidx, prange->bitmap_access) ||
758 test_bit(gpuidx, prange->bitmap_aip))
760 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
761 if (!test_bit(gpuidx, prange->bitmap_access))
764 if (!test_bit(gpuidx, prange->bitmap_aip))
768 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
769 if ((prange->flags & attrs[i].value) != attrs[i].value)
772 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
773 if ((prange->flags & attrs[i].value) != 0)
776 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
777 if (prange->granularity != attrs[i].value)
781 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
789 * svm_range_debug_dump - print all range information from svms
790 * @svms: svm range list header
792 * debug output svm range start, end, prefetch location from svms
793 * interval tree and link list
795 * Context: The caller must hold svms->lock
797 static void svm_range_debug_dump(struct svm_range_list *svms)
799 struct interval_tree_node *node;
800 struct svm_range *prange;
802 pr_debug("dump svms 0x%p list\n", svms);
803 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
805 list_for_each_entry(prange, &svms->list, list) {
806 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
807 prange, prange->start, prange->npages,
808 prange->start + prange->npages - 1,
812 pr_debug("dump svms 0x%p interval tree\n", svms);
813 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
814 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
816 prange = container_of(node, struct svm_range, it_node);
817 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
818 prange, prange->start, prange->npages,
819 prange->start + prange->npages - 1,
821 node = interval_tree_iter_next(node, 0, ~0ULL);
826 svm_range_split_array(void *ppnew, void *ppold, size_t size,
827 uint64_t old_start, uint64_t old_n,
828 uint64_t new_start, uint64_t new_n)
830 unsigned char *new, *old, *pold;
835 pold = *(unsigned char **)ppold;
839 new = kvmalloc_array(new_n, size, GFP_KERNEL);
843 d = (new_start - old_start) * size;
844 memcpy(new, pold + d, new_n * size);
846 old = kvmalloc_array(old_n, size, GFP_KERNEL);
852 d = (new_start == old_start) ? new_n * size : 0;
853 memcpy(old, pold + d, old_n * size);
856 *(void **)ppold = old;
857 *(void **)ppnew = new;
863 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
864 uint64_t start, uint64_t last)
866 uint64_t npages = last - start + 1;
869 for (i = 0; i < MAX_GPU_INSTANCE; i++) {
870 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
871 sizeof(*old->dma_addr[i]), old->start,
872 npages, new->start, new->npages);
881 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
882 uint64_t start, uint64_t last)
884 uint64_t npages = last - start + 1;
886 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
887 new->svms, new, new->start, start, last);
889 if (new->start == old->start) {
890 new->offset = old->offset;
891 old->offset += new->npages;
893 new->offset = old->offset + npages;
896 new->svm_bo = svm_range_bo_ref(old->svm_bo);
897 new->ttm_res = old->ttm_res;
899 spin_lock(&new->svm_bo->list_lock);
900 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
901 spin_unlock(&new->svm_bo->list_lock);
907 * svm_range_split_adjust - split range and adjust
910 * @old: the old range
911 * @start: the old range adjust to start address in pages
912 * @last: the old range adjust to last address in pages
914 * Copy system memory dma_addr or vram ttm_res in old range to new
915 * range from new_start up to size new->npages, the remaining old range is from
919 * 0 - OK, -ENOMEM - out of memory
922 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
923 uint64_t start, uint64_t last)
927 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
928 new->svms, new->start, old->start, old->last, start, last);
930 if (new->start < old->start ||
931 new->last > old->last) {
932 WARN_ONCE(1, "invalid new range start or last\n");
936 r = svm_range_split_pages(new, old, start, last);
940 if (old->actual_loc && old->ttm_res) {
941 r = svm_range_split_nodes(new, old, start, last);
946 old->npages = last - start + 1;
949 new->flags = old->flags;
950 new->preferred_loc = old->preferred_loc;
951 new->prefetch_loc = old->prefetch_loc;
952 new->actual_loc = old->actual_loc;
953 new->granularity = old->granularity;
954 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
955 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
961 * svm_range_split - split a range in 2 ranges
963 * @prange: the svm range to split
964 * @start: the remaining range start address in pages
965 * @last: the remaining range last address in pages
966 * @new: the result new range generated
969 * case 1: if start == prange->start
970 * prange ==> prange[start, last]
971 * new range [last + 1, prange->last]
973 * case 2: if last == prange->last
974 * prange ==> prange[start, last]
975 * new range [prange->start, start - 1]
978 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
981 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
982 struct svm_range **new)
984 uint64_t old_start = prange->start;
985 uint64_t old_last = prange->last;
986 struct svm_range_list *svms;
989 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
990 old_start, old_last, start, last);
992 if (old_start != start && old_last != last)
994 if (start < old_start || last > old_last)
998 if (old_start == start)
999 *new = svm_range_new(svms, last + 1, old_last);
1001 *new = svm_range_new(svms, old_start, start - 1);
1005 r = svm_range_split_adjust(*new, prange, start, last);
1007 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1008 r, old_start, old_last, start, last);
1009 svm_range_free(*new);
1017 svm_range_split_tail(struct svm_range *prange,
1018 uint64_t new_last, struct list_head *insert_list)
1020 struct svm_range *tail;
1021 int r = svm_range_split(prange, prange->start, new_last, &tail);
1024 list_add(&tail->list, insert_list);
1029 svm_range_split_head(struct svm_range *prange,
1030 uint64_t new_start, struct list_head *insert_list)
1032 struct svm_range *head;
1033 int r = svm_range_split(prange, new_start, prange->last, &head);
1036 list_add(&head->list, insert_list);
1041 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1042 struct svm_range *pchild, enum svm_work_list_ops op)
1044 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1045 pchild, pchild->start, pchild->last, prange, op);
1047 pchild->work_item.mm = mm;
1048 pchild->work_item.op = op;
1049 list_add_tail(&pchild->child_list, &prange->child_list);
1053 * svm_range_split_by_granularity - collect ranges within granularity boundary
1055 * @p: the process with svms list
1057 * @addr: the vm fault address in pages, to split the prange
1058 * @parent: parent range if prange is from child list
1059 * @prange: prange to split
1061 * Trims @prange to be a single aligned block of prange->granularity if
1062 * possible. The head and tail are added to the child_list in @parent.
1064 * Context: caller must hold mmap_read_lock and prange->lock
1067 * 0 - OK, otherwise error code
1070 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1071 unsigned long addr, struct svm_range *parent,
1072 struct svm_range *prange)
1074 struct svm_range *head, *tail;
1075 unsigned long start, last, size;
1078 /* Align splited range start and size to granularity size, then a single
1079 * PTE will be used for whole range, this reduces the number of PTE
1080 * updated and the L1 TLB space used for translation.
1082 size = 1UL << prange->granularity;
1083 start = ALIGN_DOWN(addr, size);
1084 last = ALIGN(addr + 1, size) - 1;
1086 pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1087 prange->svms, prange->start, prange->last, start, last, size);
1089 if (start > prange->start) {
1090 r = svm_range_split(prange, start, prange->last, &head);
1093 svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1096 if (last < prange->last) {
1097 r = svm_range_split(prange, prange->start, last, &tail);
1100 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1103 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1104 if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1105 prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1106 pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1107 prange, prange->start, prange->last,
1108 SVM_OP_ADD_RANGE_AND_MAP);
1114 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1117 struct amdgpu_device *bo_adev;
1118 uint32_t flags = prange->flags;
1119 uint32_t mapping_flags = 0;
1121 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1122 bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1124 if (domain == SVM_RANGE_VRAM_DOMAIN)
1125 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1127 switch (KFD_GC_VERSION(adev->kfd.dev)) {
1128 case IP_VERSION(9, 4, 1):
1129 if (domain == SVM_RANGE_VRAM_DOMAIN) {
1130 if (bo_adev == adev) {
1131 mapping_flags |= coherent ?
1132 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1134 mapping_flags |= coherent ?
1135 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1136 if (amdgpu_xgmi_same_hive(adev, bo_adev))
1140 mapping_flags |= coherent ?
1141 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1144 case IP_VERSION(9, 4, 2):
1145 if (domain == SVM_RANGE_VRAM_DOMAIN) {
1146 if (bo_adev == adev) {
1147 mapping_flags |= coherent ?
1148 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1149 if (adev->gmc.xgmi.connected_to_cpu)
1152 mapping_flags |= coherent ?
1153 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1154 if (amdgpu_xgmi_same_hive(adev, bo_adev))
1158 mapping_flags |= coherent ?
1159 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1163 mapping_flags |= coherent ?
1164 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1167 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1169 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1170 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1171 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1172 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1174 pte_flags = AMDGPU_PTE_VALID;
1175 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1176 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1178 pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1183 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1184 uint64_t start, uint64_t last,
1185 struct dma_fence **fence)
1187 uint64_t init_pte_value = 0;
1189 pr_debug("[0x%llx 0x%llx]\n", start, last);
1191 return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1192 last, init_pte_value, 0, 0, NULL, NULL,
1197 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1200 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1201 struct kfd_process_device *pdd;
1202 struct dma_fence *fence = NULL;
1203 struct kfd_process *p;
1207 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1209 p = container_of(prange->svms, struct kfd_process, svms);
1211 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1212 pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1213 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1215 pr_debug("failed to find device idx %d\n", gpuidx);
1219 r = svm_range_unmap_from_gpu(pdd->dev->adev,
1220 drm_priv_to_vm(pdd->drm_priv),
1221 start, last, &fence);
1226 r = dma_fence_wait(fence, false);
1227 dma_fence_put(fence);
1232 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1239 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1240 unsigned long offset, unsigned long npages, bool readonly,
1241 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1242 struct dma_fence **fence)
1244 struct amdgpu_device *adev = pdd->dev->adev;
1245 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1247 unsigned long last_start;
1252 last_start = prange->start + offset;
1254 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1255 last_start, last_start + npages - 1, readonly);
1257 for (i = offset; i < offset + npages; i++) {
1258 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1259 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1261 /* Collect all pages in the same address range and memory domain
1262 * that can be mapped with a single call to update mapping.
1264 if (i < offset + npages - 1 &&
1265 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1268 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1269 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1271 pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1273 pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1275 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1276 prange->svms, last_start, prange->start + i,
1277 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1280 r = amdgpu_vm_update_range(adev, vm, false, false, false, NULL,
1281 last_start, prange->start + i,
1283 last_start - prange->start,
1284 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1285 NULL, dma_addr, &vm->last_update);
1287 for (j = last_start - prange->start; j <= i; j++)
1288 dma_addr[j] |= last_domain;
1291 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1294 last_start = prange->start + i + 1;
1297 r = amdgpu_vm_update_pdes(adev, vm, false);
1299 pr_debug("failed %d to update directories 0x%lx\n", r,
1305 *fence = dma_fence_get(vm->last_update);
1312 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1313 unsigned long npages, bool readonly,
1314 unsigned long *bitmap, bool wait)
1316 struct kfd_process_device *pdd;
1317 struct amdgpu_device *bo_adev;
1318 struct kfd_process *p;
1319 struct dma_fence *fence = NULL;
1323 if (prange->svm_bo && prange->ttm_res)
1324 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1328 p = container_of(prange->svms, struct kfd_process, svms);
1329 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1330 pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1331 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1333 pr_debug("failed to find device idx %d\n", gpuidx);
1337 pdd = kfd_bind_process_to_device(pdd->dev, p);
1341 if (bo_adev && pdd->dev->adev != bo_adev &&
1342 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1343 pr_debug("cannot map to device idx %d\n", gpuidx);
1347 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1348 prange->dma_addr[gpuidx],
1349 bo_adev, wait ? &fence : NULL);
1354 r = dma_fence_wait(fence, false);
1355 dma_fence_put(fence);
1358 pr_debug("failed %d to dma fence wait\n", r);
1363 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1369 struct svm_validate_context {
1370 struct kfd_process *process;
1371 struct svm_range *prange;
1373 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1374 struct ttm_validate_buffer tv[MAX_GPU_INSTANCE];
1375 struct list_head validate_list;
1376 struct ww_acquire_ctx ticket;
1379 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1381 struct kfd_process_device *pdd;
1382 struct amdgpu_vm *vm;
1386 INIT_LIST_HEAD(&ctx->validate_list);
1387 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1388 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1390 pr_debug("failed to find device idx %d\n", gpuidx);
1393 vm = drm_priv_to_vm(pdd->drm_priv);
1395 ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1396 ctx->tv[gpuidx].num_shared = 4;
1397 list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1400 r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1403 pr_debug("failed %d to reserve bo\n", r);
1407 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1408 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1410 pr_debug("failed to find device idx %d\n", gpuidx);
1415 r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1416 drm_priv_to_vm(pdd->drm_priv),
1417 svm_range_bo_validate, NULL);
1419 pr_debug("failed %d validate pt bos\n", r);
1427 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1431 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1433 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1436 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1438 struct kfd_process_device *pdd;
1440 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1442 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1446 * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1448 * To prevent concurrent destruction or change of range attributes, the
1449 * svm_read_lock must be held. The caller must not hold the svm_write_lock
1450 * because that would block concurrent evictions and lead to deadlocks. To
1451 * serialize concurrent migrations or validations of the same range, the
1452 * prange->migrate_mutex must be held.
1454 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1457 * The following sequence ensures race-free validation and GPU mapping:
1459 * 1. Reserve page table (and SVM BO if range is in VRAM)
1460 * 2. hmm_range_fault to get page addresses (if system memory)
1461 * 3. DMA-map pages (if system memory)
1462 * 4-a. Take notifier lock
1463 * 4-b. Check that pages still valid (mmu_interval_read_retry)
1464 * 4-c. Check that the range was not split or otherwise invalidated
1465 * 4-d. Update GPU page table
1466 * 4.e. Release notifier lock
1467 * 5. Release page table (and SVM BO) reservation
1469 static int svm_range_validate_and_map(struct mm_struct *mm,
1470 struct svm_range *prange,
1471 int32_t gpuidx, bool intr, bool wait)
1473 struct svm_validate_context ctx;
1474 unsigned long start, end, addr;
1475 struct kfd_process *p;
1480 ctx.process = container_of(prange->svms, struct kfd_process, svms);
1481 ctx.prange = prange;
1484 if (gpuidx < MAX_GPU_INSTANCE) {
1485 bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1486 bitmap_set(ctx.bitmap, gpuidx, 1);
1487 } else if (ctx.process->xnack_enabled) {
1488 bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1490 /* If prefetch range to GPU, or GPU retry fault migrate range to
1491 * GPU, which has ACCESS attribute to the range, create mapping
1494 if (prange->actual_loc) {
1495 gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1496 prange->actual_loc);
1498 WARN_ONCE(1, "failed get device by id 0x%x\n",
1499 prange->actual_loc);
1502 if (test_bit(gpuidx, prange->bitmap_access))
1503 bitmap_set(ctx.bitmap, gpuidx, 1);
1506 bitmap_or(ctx.bitmap, prange->bitmap_access,
1507 prange->bitmap_aip, MAX_GPU_INSTANCE);
1510 if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE))
1513 if (prange->actual_loc && !prange->ttm_res) {
1514 /* This should never happen. actual_loc gets set by
1515 * svm_migrate_ram_to_vram after allocating a BO.
1517 WARN_ONCE(1, "VRAM BO missing during validation\n");
1521 svm_range_reserve_bos(&ctx);
1523 p = container_of(prange->svms, struct kfd_process, svms);
1524 owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1526 for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1527 if (kfd_svm_page_owner(p, idx) != owner) {
1533 start = prange->start << PAGE_SHIFT;
1534 end = (prange->last + 1) << PAGE_SHIFT;
1535 for (addr = start; addr < end && !r; ) {
1536 struct hmm_range *hmm_range;
1537 struct vm_area_struct *vma;
1539 unsigned long offset;
1540 unsigned long npages;
1543 vma = find_vma(mm, addr);
1544 if (!vma || addr < vma->vm_start) {
1548 readonly = !(vma->vm_flags & VM_WRITE);
1550 next = min(vma->vm_end, end);
1551 npages = (next - addr) >> PAGE_SHIFT;
1552 WRITE_ONCE(p->svms.faulting_task, current);
1553 r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1554 addr, npages, &hmm_range,
1555 readonly, true, owner);
1556 WRITE_ONCE(p->svms.faulting_task, NULL);
1558 pr_debug("failed %d to get svm range pages\n", r);
1562 offset = (addr - start) >> PAGE_SHIFT;
1563 r = svm_range_dma_map(prange, ctx.bitmap, offset, npages,
1564 hmm_range->hmm_pfns);
1566 pr_debug("failed %d to dma map range\n", r);
1570 svm_range_lock(prange);
1571 if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1572 pr_debug("hmm update the range, need validate again\n");
1576 if (!list_empty(&prange->child_list)) {
1577 pr_debug("range split by unmap in parallel, validate again\n");
1582 r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1586 svm_range_unlock(prange);
1592 prange->validated_once = true;
1595 svm_range_unreserve_bos(&ctx);
1598 prange->validate_timestamp = ktime_to_us(ktime_get());
1604 * svm_range_list_lock_and_flush_work - flush pending deferred work
1606 * @svms: the svm range list
1607 * @mm: the mm structure
1609 * Context: Returns with mmap write lock held, pending deferred work flushed
1613 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1614 struct mm_struct *mm)
1617 flush_work(&svms->deferred_list_work);
1618 mmap_write_lock(mm);
1620 if (list_empty(&svms->deferred_range_list))
1622 mmap_write_unlock(mm);
1623 pr_debug("retry flush\n");
1624 goto retry_flush_work;
1627 static void svm_range_restore_work(struct work_struct *work)
1629 struct delayed_work *dwork = to_delayed_work(work);
1630 struct amdkfd_process_info *process_info;
1631 struct svm_range_list *svms;
1632 struct svm_range *prange;
1633 struct kfd_process *p;
1634 struct mm_struct *mm;
1639 svms = container_of(dwork, struct svm_range_list, restore_work);
1640 evicted_ranges = atomic_read(&svms->evicted_ranges);
1641 if (!evicted_ranges)
1644 pr_debug("restore svm ranges\n");
1646 p = container_of(svms, struct kfd_process, svms);
1647 process_info = p->kgd_process_info;
1649 /* Keep mm reference when svm_range_validate_and_map ranges */
1650 mm = get_task_mm(p->lead_thread);
1652 pr_debug("svms 0x%p process mm gone\n", svms);
1656 mutex_lock(&process_info->lock);
1657 svm_range_list_lock_and_flush_work(svms, mm);
1658 mutex_lock(&svms->lock);
1660 evicted_ranges = atomic_read(&svms->evicted_ranges);
1662 list_for_each_entry(prange, &svms->list, list) {
1663 invalid = atomic_read(&prange->invalid);
1667 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1668 prange->svms, prange, prange->start, prange->last,
1672 * If range is migrating, wait for migration is done.
1674 mutex_lock(&prange->migrate_mutex);
1676 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1679 pr_debug("failed %d to map 0x%lx to gpus\n", r,
1682 mutex_unlock(&prange->migrate_mutex);
1684 goto out_reschedule;
1686 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1687 goto out_reschedule;
1690 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1692 goto out_reschedule;
1696 r = kgd2kfd_resume_mm(mm);
1698 /* No recovery from this failure. Probably the CP is
1699 * hanging. No point trying again.
1701 pr_debug("failed %d to resume KFD\n", r);
1704 pr_debug("restore svm ranges successfully\n");
1707 mutex_unlock(&svms->lock);
1708 mmap_write_unlock(mm);
1709 mutex_unlock(&process_info->lock);
1712 /* If validation failed, reschedule another attempt */
1713 if (evicted_ranges) {
1714 pr_debug("reschedule to restore svm range\n");
1715 schedule_delayed_work(&svms->restore_work,
1716 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1721 * svm_range_evict - evict svm range
1722 * @prange: svm range structure
1723 * @mm: current process mm_struct
1724 * @start: starting process queue number
1725 * @last: last process queue number
1727 * Stop all queues of the process to ensure GPU doesn't access the memory, then
1728 * return to let CPU evict the buffer and proceed CPU pagetable update.
1730 * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1731 * If invalidation happens while restore work is running, restore work will
1732 * restart to ensure to get the latest CPU pages mapping to GPU, then start
1736 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1737 unsigned long start, unsigned long last)
1739 struct svm_range_list *svms = prange->svms;
1740 struct svm_range *pchild;
1741 struct kfd_process *p;
1744 p = container_of(svms, struct kfd_process, svms);
1746 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1747 svms, prange->start, prange->last, start, last);
1749 if (!p->xnack_enabled) {
1752 list_for_each_entry(pchild, &prange->child_list, child_list) {
1753 mutex_lock_nested(&pchild->lock, 1);
1754 if (pchild->start <= last && pchild->last >= start) {
1755 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1756 pchild->start, pchild->last);
1757 atomic_inc(&pchild->invalid);
1759 mutex_unlock(&pchild->lock);
1762 if (prange->start <= last && prange->last >= start)
1763 atomic_inc(&prange->invalid);
1765 evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1766 if (evicted_ranges != 1)
1769 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1770 prange->svms, prange->start, prange->last);
1772 /* First eviction, stop the queues */
1773 r = kgd2kfd_quiesce_mm(mm);
1775 pr_debug("failed to quiesce KFD\n");
1777 pr_debug("schedule to restore svm %p ranges\n", svms);
1778 schedule_delayed_work(&svms->restore_work,
1779 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1783 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1784 prange->svms, start, last);
1785 list_for_each_entry(pchild, &prange->child_list, child_list) {
1786 mutex_lock_nested(&pchild->lock, 1);
1787 s = max(start, pchild->start);
1788 l = min(last, pchild->last);
1790 svm_range_unmap_from_gpus(pchild, s, l);
1791 mutex_unlock(&pchild->lock);
1793 s = max(start, prange->start);
1794 l = min(last, prange->last);
1796 svm_range_unmap_from_gpus(prange, s, l);
1802 static struct svm_range *svm_range_clone(struct svm_range *old)
1804 struct svm_range *new;
1806 new = svm_range_new(old->svms, old->start, old->last);
1811 new->ttm_res = old->ttm_res;
1812 new->offset = old->offset;
1813 new->svm_bo = svm_range_bo_ref(old->svm_bo);
1814 spin_lock(&new->svm_bo->list_lock);
1815 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1816 spin_unlock(&new->svm_bo->list_lock);
1818 new->flags = old->flags;
1819 new->preferred_loc = old->preferred_loc;
1820 new->prefetch_loc = old->prefetch_loc;
1821 new->actual_loc = old->actual_loc;
1822 new->granularity = old->granularity;
1823 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1824 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1830 * svm_range_add - add svm range and handle overlap
1831 * @p: the range add to this process svms
1832 * @start: page size aligned
1833 * @size: page size aligned
1834 * @nattr: number of attributes
1835 * @attrs: array of attributes
1836 * @update_list: output, the ranges need validate and update GPU mapping
1837 * @insert_list: output, the ranges need insert to svms
1838 * @remove_list: output, the ranges are replaced and need remove from svms
1840 * Check if the virtual address range has overlap with any existing ranges,
1841 * split partly overlapping ranges and add new ranges in the gaps. All changes
1842 * should be applied to the range_list and interval tree transactionally. If
1843 * any range split or allocation fails, the entire update fails. Therefore any
1844 * existing overlapping svm_ranges are cloned and the original svm_ranges left
1847 * If the transaction succeeds, the caller can update and insert clones and
1848 * new ranges, then free the originals.
1850 * Otherwise the caller can free the clones and new ranges, while the old
1851 * svm_ranges remain unchanged.
1853 * Context: Process context, caller must hold svms->lock
1856 * 0 - OK, otherwise error code
1859 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
1860 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
1861 struct list_head *update_list, struct list_head *insert_list,
1862 struct list_head *remove_list)
1864 unsigned long last = start + size - 1UL;
1865 struct svm_range_list *svms = &p->svms;
1866 struct interval_tree_node *node;
1867 struct svm_range *prange;
1868 struct svm_range *tmp;
1871 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
1873 INIT_LIST_HEAD(update_list);
1874 INIT_LIST_HEAD(insert_list);
1875 INIT_LIST_HEAD(remove_list);
1877 node = interval_tree_iter_first(&svms->objects, start, last);
1879 struct interval_tree_node *next;
1880 unsigned long next_start;
1882 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
1885 prange = container_of(node, struct svm_range, it_node);
1886 next = interval_tree_iter_next(node, start, last);
1887 next_start = min(node->last, last) + 1;
1889 if (svm_range_is_same_attrs(p, prange, nattr, attrs)) {
1891 } else if (node->start < start || node->last > last) {
1892 /* node intersects the update range and its attributes
1893 * will change. Clone and split it, apply updates only
1894 * to the overlapping part
1896 struct svm_range *old = prange;
1898 prange = svm_range_clone(old);
1904 list_add(&old->update_list, remove_list);
1905 list_add(&prange->list, insert_list);
1906 list_add(&prange->update_list, update_list);
1908 if (node->start < start) {
1909 pr_debug("change old range start\n");
1910 r = svm_range_split_head(prange, start,
1915 if (node->last > last) {
1916 pr_debug("change old range last\n");
1917 r = svm_range_split_tail(prange, last,
1923 /* The node is contained within start..last,
1926 list_add(&prange->update_list, update_list);
1929 /* insert a new node if needed */
1930 if (node->start > start) {
1931 prange = svm_range_new(svms, start, node->start - 1);
1937 list_add(&prange->list, insert_list);
1938 list_add(&prange->update_list, update_list);
1945 /* add a final range at the end if needed */
1946 if (start <= last) {
1947 prange = svm_range_new(svms, start, last);
1952 list_add(&prange->list, insert_list);
1953 list_add(&prange->update_list, update_list);
1958 list_for_each_entry_safe(prange, tmp, insert_list, list)
1959 svm_range_free(prange);
1965 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
1966 struct svm_range *prange)
1968 unsigned long start;
1971 start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
1972 last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
1974 if (prange->start == start && prange->last == last)
1977 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1978 prange->svms, prange, start, last, prange->start,
1981 if (start != 0 && last != 0) {
1982 interval_tree_remove(&prange->it_node, &prange->svms->objects);
1983 svm_range_remove_notifier(prange);
1985 prange->it_node.start = prange->start;
1986 prange->it_node.last = prange->last;
1988 interval_tree_insert(&prange->it_node, &prange->svms->objects);
1989 svm_range_add_notifier_locked(mm, prange);
1993 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
1994 struct mm_struct *mm)
1996 switch (prange->work_item.op) {
1998 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
1999 svms, prange, prange->start, prange->last);
2001 case SVM_OP_UNMAP_RANGE:
2002 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2003 svms, prange, prange->start, prange->last);
2004 svm_range_unlink(prange);
2005 svm_range_remove_notifier(prange);
2006 svm_range_free(prange);
2008 case SVM_OP_UPDATE_RANGE_NOTIFIER:
2009 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2010 svms, prange, prange->start, prange->last);
2011 svm_range_update_notifier_and_interval_tree(mm, prange);
2013 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2014 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2015 svms, prange, prange->start, prange->last);
2016 svm_range_update_notifier_and_interval_tree(mm, prange);
2017 /* TODO: implement deferred validation and mapping */
2019 case SVM_OP_ADD_RANGE:
2020 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2021 prange->start, prange->last);
2022 svm_range_add_to_svms(prange);
2023 svm_range_add_notifier_locked(mm, prange);
2025 case SVM_OP_ADD_RANGE_AND_MAP:
2026 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2027 prange, prange->start, prange->last);
2028 svm_range_add_to_svms(prange);
2029 svm_range_add_notifier_locked(mm, prange);
2030 /* TODO: implement deferred validation and mapping */
2033 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2034 prange->work_item.op);
2038 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2040 struct kfd_process_device *pdd;
2041 struct kfd_process *p;
2045 p = container_of(svms, struct kfd_process, svms);
2048 drain = atomic_read(&svms->drain_pagefaults);
2052 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2057 pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2059 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2060 &pdd->dev->adev->irq.ih1);
2061 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2063 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2067 static void svm_range_deferred_list_work(struct work_struct *work)
2069 struct svm_range_list *svms;
2070 struct svm_range *prange;
2071 struct mm_struct *mm;
2073 svms = container_of(work, struct svm_range_list, deferred_list_work);
2074 pr_debug("enter svms 0x%p\n", svms);
2076 spin_lock(&svms->deferred_list_lock);
2077 while (!list_empty(&svms->deferred_range_list)) {
2078 prange = list_first_entry(&svms->deferred_range_list,
2079 struct svm_range, deferred_list);
2080 spin_unlock(&svms->deferred_list_lock);
2082 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2083 prange->start, prange->last, prange->work_item.op);
2085 mm = prange->work_item.mm;
2087 mmap_write_lock(mm);
2089 /* Checking for the need to drain retry faults must be inside
2090 * mmap write lock to serialize with munmap notifiers.
2092 if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2093 mmap_write_unlock(mm);
2094 svm_range_drain_retry_fault(svms);
2098 /* Remove from deferred_list must be inside mmap write lock, for
2100 * 1. unmap_from_cpu may change work_item.op and add the range
2101 * to deferred_list again, cause use after free bug.
2102 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2103 * lock and continue because deferred_list is empty, but
2104 * deferred_list work is actually waiting for mmap lock.
2106 spin_lock(&svms->deferred_list_lock);
2107 list_del_init(&prange->deferred_list);
2108 spin_unlock(&svms->deferred_list_lock);
2110 mutex_lock(&svms->lock);
2111 mutex_lock(&prange->migrate_mutex);
2112 while (!list_empty(&prange->child_list)) {
2113 struct svm_range *pchild;
2115 pchild = list_first_entry(&prange->child_list,
2116 struct svm_range, child_list);
2117 pr_debug("child prange 0x%p op %d\n", pchild,
2118 pchild->work_item.op);
2119 list_del_init(&pchild->child_list);
2120 svm_range_handle_list_op(svms, pchild, mm);
2122 mutex_unlock(&prange->migrate_mutex);
2124 svm_range_handle_list_op(svms, prange, mm);
2125 mutex_unlock(&svms->lock);
2126 mmap_write_unlock(mm);
2128 /* Pairs with mmget in svm_range_add_list_work */
2131 spin_lock(&svms->deferred_list_lock);
2133 spin_unlock(&svms->deferred_list_lock);
2134 pr_debug("exit svms 0x%p\n", svms);
2138 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2139 struct mm_struct *mm, enum svm_work_list_ops op)
2141 spin_lock(&svms->deferred_list_lock);
2142 /* if prange is on the deferred list */
2143 if (!list_empty(&prange->deferred_list)) {
2144 pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2145 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2146 if (op != SVM_OP_NULL &&
2147 prange->work_item.op != SVM_OP_UNMAP_RANGE)
2148 prange->work_item.op = op;
2150 prange->work_item.op = op;
2152 /* Pairs with mmput in deferred_list_work */
2154 prange->work_item.mm = mm;
2155 list_add_tail(&prange->deferred_list,
2156 &prange->svms->deferred_range_list);
2157 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2158 prange, prange->start, prange->last, op);
2160 spin_unlock(&svms->deferred_list_lock);
2163 void schedule_deferred_list_work(struct svm_range_list *svms)
2165 spin_lock(&svms->deferred_list_lock);
2166 if (!list_empty(&svms->deferred_range_list))
2167 schedule_work(&svms->deferred_list_work);
2168 spin_unlock(&svms->deferred_list_lock);
2172 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2173 struct svm_range *prange, unsigned long start,
2176 struct svm_range *head;
2177 struct svm_range *tail;
2179 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2180 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2181 prange->start, prange->last);
2184 if (start > prange->last || last < prange->start)
2187 head = tail = prange;
2188 if (start > prange->start)
2189 svm_range_split(prange, prange->start, start - 1, &tail);
2190 if (last < tail->last)
2191 svm_range_split(tail, last + 1, tail->last, &head);
2193 if (head != prange && tail != prange) {
2194 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2195 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2196 } else if (tail != prange) {
2197 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2198 } else if (head != prange) {
2199 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2200 } else if (parent != prange) {
2201 prange->work_item.op = SVM_OP_UNMAP_RANGE;
2206 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2207 unsigned long start, unsigned long last)
2209 struct svm_range_list *svms;
2210 struct svm_range *pchild;
2211 struct kfd_process *p;
2215 p = kfd_lookup_process_by_mm(mm);
2220 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2221 prange, prange->start, prange->last, start, last);
2223 /* Make sure pending page faults are drained in the deferred worker
2224 * before the range is freed to avoid straggler interrupts on
2225 * unmapped memory causing "phantom faults".
2227 atomic_inc(&svms->drain_pagefaults);
2229 unmap_parent = start <= prange->start && last >= prange->last;
2231 list_for_each_entry(pchild, &prange->child_list, child_list) {
2232 mutex_lock_nested(&pchild->lock, 1);
2233 s = max(start, pchild->start);
2234 l = min(last, pchild->last);
2236 svm_range_unmap_from_gpus(pchild, s, l);
2237 svm_range_unmap_split(mm, prange, pchild, start, last);
2238 mutex_unlock(&pchild->lock);
2240 s = max(start, prange->start);
2241 l = min(last, prange->last);
2243 svm_range_unmap_from_gpus(prange, s, l);
2244 svm_range_unmap_split(mm, prange, prange, start, last);
2247 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2249 svm_range_add_list_work(svms, prange, mm,
2250 SVM_OP_UPDATE_RANGE_NOTIFIER);
2251 schedule_deferred_list_work(svms);
2253 kfd_unref_process(p);
2257 * svm_range_cpu_invalidate_pagetables - interval notifier callback
2258 * @mni: mmu_interval_notifier struct
2259 * @range: mmu_notifier_range struct
2260 * @cur_seq: value to pass to mmu_interval_set_seq()
2262 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2263 * is from migration, or CPU page invalidation callback.
2265 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2266 * work thread, and split prange if only part of prange is unmapped.
2268 * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2269 * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2270 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2271 * update GPU mapping to recover.
2273 * Context: mmap lock, notifier_invalidate_start lock are held
2274 * for invalidate event, prange lock is held if this is from migration
2277 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2278 const struct mmu_notifier_range *range,
2279 unsigned long cur_seq)
2281 struct svm_range *prange;
2282 unsigned long start;
2285 if (range->event == MMU_NOTIFY_RELEASE)
2288 start = mni->interval_tree.start;
2289 last = mni->interval_tree.last;
2290 start = max(start, range->start) >> PAGE_SHIFT;
2291 last = min(last, range->end - 1) >> PAGE_SHIFT;
2292 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2293 start, last, range->start >> PAGE_SHIFT,
2294 (range->end - 1) >> PAGE_SHIFT,
2295 mni->interval_tree.start >> PAGE_SHIFT,
2296 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2298 prange = container_of(mni, struct svm_range, notifier);
2300 svm_range_lock(prange);
2301 mmu_interval_set_seq(mni, cur_seq);
2303 switch (range->event) {
2304 case MMU_NOTIFY_UNMAP:
2305 svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2308 svm_range_evict(prange, mni->mm, start, last);
2312 svm_range_unlock(prange);
2318 * svm_range_from_addr - find svm range from fault address
2319 * @svms: svm range list header
2320 * @addr: address to search range interval tree, in pages
2321 * @parent: parent range if range is on child list
2323 * Context: The caller must hold svms->lock
2325 * Return: the svm_range found or NULL
2328 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2329 struct svm_range **parent)
2331 struct interval_tree_node *node;
2332 struct svm_range *prange;
2333 struct svm_range *pchild;
2335 node = interval_tree_iter_first(&svms->objects, addr, addr);
2339 prange = container_of(node, struct svm_range, it_node);
2340 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2341 addr, prange->start, prange->last, node->start, node->last);
2343 if (addr >= prange->start && addr <= prange->last) {
2348 list_for_each_entry(pchild, &prange->child_list, child_list)
2349 if (addr >= pchild->start && addr <= pchild->last) {
2350 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2351 addr, pchild->start, pchild->last);
2360 /* svm_range_best_restore_location - decide the best fault restore location
2361 * @prange: svm range structure
2362 * @adev: the GPU on which vm fault happened
2364 * This is only called when xnack is on, to decide the best location to restore
2365 * the range mapping after GPU vm fault. Caller uses the best location to do
2366 * migration if actual loc is not best location, then update GPU page table
2367 * mapping to the best location.
2369 * If the preferred loc is accessible by faulting GPU, use preferred loc.
2370 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2371 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2372 * if range actual loc is cpu, best_loc is cpu
2373 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2375 * Otherwise, GPU no access, best_loc is -1.
2378 * -1 means vm fault GPU no access
2379 * 0 for CPU or GPU id
2382 svm_range_best_restore_location(struct svm_range *prange,
2383 struct amdgpu_device *adev,
2386 struct amdgpu_device *bo_adev, *preferred_adev;
2387 struct kfd_process *p;
2391 p = container_of(prange->svms, struct kfd_process, svms);
2393 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx);
2395 pr_debug("failed to get gpuid from kgd\n");
2399 if (prange->preferred_loc == gpuid ||
2400 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2401 return prange->preferred_loc;
2402 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2403 preferred_adev = svm_range_get_adev_by_id(prange,
2404 prange->preferred_loc);
2405 if (amdgpu_xgmi_same_hive(adev, preferred_adev))
2406 return prange->preferred_loc;
2410 if (test_bit(*gpuidx, prange->bitmap_access))
2413 if (test_bit(*gpuidx, prange->bitmap_aip)) {
2414 if (!prange->actual_loc)
2417 bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2418 if (amdgpu_xgmi_same_hive(adev, bo_adev))
2419 return prange->actual_loc;
2428 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2429 unsigned long *start, unsigned long *last,
2430 bool *is_heap_stack)
2432 struct vm_area_struct *vma;
2433 struct interval_tree_node *node;
2434 unsigned long start_limit, end_limit;
2436 vma = find_vma(p->mm, addr << PAGE_SHIFT);
2437 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2438 pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2442 *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
2443 vma->vm_end >= vma->vm_mm->start_brk) ||
2444 (vma->vm_start <= vma->vm_mm->start_stack &&
2445 vma->vm_end >= vma->vm_mm->start_stack);
2447 start_limit = max(vma->vm_start >> PAGE_SHIFT,
2448 (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2449 end_limit = min(vma->vm_end >> PAGE_SHIFT,
2450 (unsigned long)ALIGN(addr + 1, 2UL << 8));
2451 /* First range that starts after the fault address */
2452 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2454 end_limit = min(end_limit, node->start);
2455 /* Last range that ends before the fault address */
2456 node = container_of(rb_prev(&node->rb),
2457 struct interval_tree_node, rb);
2459 /* Last range must end before addr because
2460 * there was no range after addr
2462 node = container_of(rb_last(&p->svms.objects.rb_root),
2463 struct interval_tree_node, rb);
2466 if (node->last >= addr) {
2467 WARN(1, "Overlap with prev node and page fault addr\n");
2470 start_limit = max(start_limit, node->last + 1);
2473 *start = start_limit;
2474 *last = end_limit - 1;
2476 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2477 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2478 *start, *last, *is_heap_stack);
2484 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2485 uint64_t *bo_s, uint64_t *bo_l)
2487 struct amdgpu_bo_va_mapping *mapping;
2488 struct interval_tree_node *node;
2489 struct amdgpu_bo *bo = NULL;
2490 unsigned long userptr;
2494 for (i = 0; i < p->n_pdds; i++) {
2495 struct amdgpu_vm *vm;
2497 if (!p->pdds[i]->drm_priv)
2500 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2501 r = amdgpu_bo_reserve(vm->root.bo, false);
2505 /* Check userptr by searching entire vm->va interval tree */
2506 node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2508 mapping = container_of((struct rb_node *)node,
2509 struct amdgpu_bo_va_mapping, rb);
2510 bo = mapping->bo_va->base.bo;
2512 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2513 start << PAGE_SHIFT,
2516 node = interval_tree_iter_next(node, 0, ~0ULL);
2520 pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2523 *bo_s = userptr >> PAGE_SHIFT;
2524 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2526 amdgpu_bo_unreserve(vm->root.bo);
2529 amdgpu_bo_unreserve(vm->root.bo);
2535 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2536 struct kfd_process *p,
2537 struct mm_struct *mm,
2540 struct svm_range *prange = NULL;
2541 unsigned long start, last;
2542 uint32_t gpuid, gpuidx;
2548 if (svm_range_get_range_boundaries(p, addr, &start, &last,
2552 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2553 if (r != -EADDRINUSE)
2554 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2556 if (r == -EADDRINUSE) {
2557 if (addr >= bo_s && addr <= bo_l)
2560 /* Create one page svm range if 2MB range overlapping */
2565 prange = svm_range_new(&p->svms, start, last);
2567 pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2570 if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) {
2571 pr_debug("failed to get gpuid from kgd\n");
2572 svm_range_free(prange);
2577 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2579 svm_range_add_to_svms(prange);
2580 svm_range_add_notifier_locked(mm, prange);
2585 /* svm_range_skip_recover - decide if prange can be recovered
2586 * @prange: svm range structure
2588 * GPU vm retry fault handle skip recover the range for cases:
2589 * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2590 * deferred list work will drain the stale fault before free the prange.
2591 * 2. prange is on deferred list to add interval notifier after split, or
2592 * 3. prange is child range, it is split from parent prange, recover later
2593 * after interval notifier is added.
2595 * Return: true to skip recover, false to recover
2597 static bool svm_range_skip_recover(struct svm_range *prange)
2599 struct svm_range_list *svms = prange->svms;
2601 spin_lock(&svms->deferred_list_lock);
2602 if (list_empty(&prange->deferred_list) &&
2603 list_empty(&prange->child_list)) {
2604 spin_unlock(&svms->deferred_list_lock);
2607 spin_unlock(&svms->deferred_list_lock);
2609 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2610 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2611 svms, prange, prange->start, prange->last);
2614 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2615 prange->work_item.op == SVM_OP_ADD_RANGE) {
2616 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2617 svms, prange, prange->start, prange->last);
2624 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2627 struct kfd_process_device *pdd;
2629 /* fault is on different page of same range
2630 * or fault is skipped to recover later
2631 * or fault is on invalid virtual address
2633 if (gpuidx == MAX_GPU_INSTANCE) {
2637 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx);
2642 /* fault is recovered
2643 * or fault cannot recover because GPU no access on the range
2645 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2647 WRITE_ONCE(pdd->faults, pdd->faults + 1);
2651 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2653 unsigned long requested = VM_READ;
2656 requested |= VM_WRITE;
2658 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2660 return (vma->vm_flags & requested) == requested;
2664 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2665 uint64_t addr, bool write_fault)
2667 struct mm_struct *mm = NULL;
2668 struct svm_range_list *svms;
2669 struct svm_range *prange;
2670 struct kfd_process *p;
2673 int32_t gpuidx = MAX_GPU_INSTANCE;
2674 bool write_locked = false;
2675 struct vm_area_struct *vma;
2678 if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2679 pr_debug("device does not support SVM\n");
2683 p = kfd_lookup_process_by_pasid(pasid);
2685 pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2690 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2692 if (atomic_read(&svms->drain_pagefaults)) {
2693 pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2698 if (!p->xnack_enabled) {
2699 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2704 /* p->lead_thread is available as kfd_process_wq_release flush the work
2705 * before releasing task ref.
2707 mm = get_task_mm(p->lead_thread);
2709 pr_debug("svms 0x%p failed to get mm\n", svms);
2716 mutex_lock(&svms->lock);
2717 prange = svm_range_from_addr(svms, addr, NULL);
2719 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2721 if (!write_locked) {
2722 /* Need the write lock to create new range with MMU notifier.
2723 * Also flush pending deferred work to make sure the interval
2724 * tree is up to date before we add a new range
2726 mutex_unlock(&svms->lock);
2727 mmap_read_unlock(mm);
2728 mmap_write_lock(mm);
2729 write_locked = true;
2730 goto retry_write_locked;
2732 prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2734 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2736 mmap_write_downgrade(mm);
2738 goto out_unlock_svms;
2742 mmap_write_downgrade(mm);
2744 mutex_lock(&prange->migrate_mutex);
2746 if (svm_range_skip_recover(prange)) {
2747 amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2749 goto out_unlock_range;
2752 timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp;
2753 /* skip duplicate vm fault on different pages of same range */
2754 if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
2755 pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2756 svms, prange->start, prange->last);
2758 goto out_unlock_range;
2761 /* __do_munmap removed VMA, return success as we are handling stale
2764 vma = find_vma(mm, addr << PAGE_SHIFT);
2765 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2766 pr_debug("address 0x%llx VMA is removed\n", addr);
2768 goto out_unlock_range;
2771 if (!svm_fault_allowed(vma, write_fault)) {
2772 pr_debug("fault addr 0x%llx no %s permission\n", addr,
2773 write_fault ? "write" : "read");
2775 goto out_unlock_range;
2778 best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2779 if (best_loc == -1) {
2780 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2781 svms, prange->start, prange->last);
2783 goto out_unlock_range;
2786 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2787 svms, prange->start, prange->last, best_loc,
2788 prange->actual_loc);
2790 if (prange->actual_loc != best_loc) {
2792 r = svm_migrate_to_vram(prange, best_loc, mm);
2794 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2796 /* Fallback to system memory if migration to
2799 if (prange->actual_loc)
2800 r = svm_migrate_vram_to_ram(prange, mm);
2805 r = svm_migrate_vram_to_ram(prange, mm);
2808 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2809 r, svms, prange->start, prange->last);
2810 goto out_unlock_range;
2814 r = svm_range_validate_and_map(mm, prange, gpuidx, false, false);
2816 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2817 r, svms, prange->start, prange->last);
2820 mutex_unlock(&prange->migrate_mutex);
2822 mutex_unlock(&svms->lock);
2823 mmap_read_unlock(mm);
2825 svm_range_count_fault(adev, p, gpuidx);
2829 kfd_unref_process(p);
2832 pr_debug("recover vm fault later\n");
2833 amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2839 void svm_range_list_fini(struct kfd_process *p)
2841 struct svm_range *prange;
2842 struct svm_range *next;
2844 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
2846 cancel_delayed_work_sync(&p->svms.restore_work);
2848 /* Ensure list work is finished before process is destroyed */
2849 flush_work(&p->svms.deferred_list_work);
2852 * Ensure no retry fault comes in afterwards, as page fault handler will
2853 * not find kfd process and take mm lock to recover fault.
2855 atomic_inc(&p->svms.drain_pagefaults);
2856 svm_range_drain_retry_fault(&p->svms);
2858 list_for_each_entry_safe(prange, next, &p->svms.list, list) {
2859 svm_range_unlink(prange);
2860 svm_range_remove_notifier(prange);
2861 svm_range_free(prange);
2864 mutex_destroy(&p->svms.lock);
2866 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
2869 int svm_range_list_init(struct kfd_process *p)
2871 struct svm_range_list *svms = &p->svms;
2874 svms->objects = RB_ROOT_CACHED;
2875 mutex_init(&svms->lock);
2876 INIT_LIST_HEAD(&svms->list);
2877 atomic_set(&svms->evicted_ranges, 0);
2878 atomic_set(&svms->drain_pagefaults, 0);
2879 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
2880 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
2881 INIT_LIST_HEAD(&svms->deferred_range_list);
2882 INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
2883 spin_lock_init(&svms->deferred_list_lock);
2885 for (i = 0; i < p->n_pdds; i++)
2886 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
2887 bitmap_set(svms->bitmap_supported, i, 1);
2893 * svm_range_check_vm - check if virtual address range mapped already
2894 * @p: current kfd_process
2895 * @start: range start address, in pages
2896 * @last: range last address, in pages
2897 * @bo_s: mapping start address in pages if address range already mapped
2898 * @bo_l: mapping last address in pages if address range already mapped
2900 * The purpose is to avoid virtual address ranges already allocated by
2901 * kfd_ioctl_alloc_memory_of_gpu ioctl.
2902 * It looks for each pdd in the kfd_process.
2904 * Context: Process context
2906 * Return 0 - OK, if the range is not mapped.
2907 * Otherwise error code:
2908 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
2909 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
2910 * a signal. Release all buffer reservations and return to user-space.
2913 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
2914 uint64_t *bo_s, uint64_t *bo_l)
2916 struct amdgpu_bo_va_mapping *mapping;
2917 struct interval_tree_node *node;
2921 for (i = 0; i < p->n_pdds; i++) {
2922 struct amdgpu_vm *vm;
2924 if (!p->pdds[i]->drm_priv)
2927 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2928 r = amdgpu_bo_reserve(vm->root.bo, false);
2932 node = interval_tree_iter_first(&vm->va, start, last);
2934 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
2936 mapping = container_of((struct rb_node *)node,
2937 struct amdgpu_bo_va_mapping, rb);
2939 *bo_s = mapping->start;
2940 *bo_l = mapping->last;
2942 amdgpu_bo_unreserve(vm->root.bo);
2945 amdgpu_bo_unreserve(vm->root.bo);
2952 * svm_range_is_valid - check if virtual address range is valid
2953 * @p: current kfd_process
2954 * @start: range start address, in pages
2955 * @size: range size, in pages
2957 * Valid virtual address range means it belongs to one or more VMAs
2959 * Context: Process context
2962 * 0 - OK, otherwise error code
2965 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
2967 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
2968 struct vm_area_struct *vma;
2970 unsigned long start_unchg = start;
2972 start <<= PAGE_SHIFT;
2973 end = start + (size << PAGE_SHIFT);
2975 vma = find_vma(p->mm, start);
2976 if (!vma || start < vma->vm_start ||
2977 (vma->vm_flags & device_vma))
2979 start = min(end, vma->vm_end);
2980 } while (start < end);
2982 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
2987 * svm_range_best_prefetch_location - decide the best prefetch location
2988 * @prange: svm range structure
2991 * If range map to single GPU, the best prefetch location is prefetch_loc, which
2992 * can be CPU or GPU.
2994 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
2995 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
2996 * the best prefetch location is always CPU, because GPU can not have coherent
2997 * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3000 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3001 * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3003 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3004 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3005 * prefetch location is always CPU.
3007 * Context: Process context
3010 * 0 for CPU or GPU id
3013 svm_range_best_prefetch_location(struct svm_range *prange)
3015 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3016 uint32_t best_loc = prange->prefetch_loc;
3017 struct kfd_process_device *pdd;
3018 struct amdgpu_device *bo_adev;
3019 struct kfd_process *p;
3022 p = container_of(prange->svms, struct kfd_process, svms);
3024 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3027 bo_adev = svm_range_get_adev_by_id(prange, best_loc);
3029 WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
3034 if (p->xnack_enabled)
3035 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3037 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3040 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3041 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3043 pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3047 if (pdd->dev->adev == bo_adev)
3050 if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
3057 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3058 p->xnack_enabled, &p->svms, prange->start, prange->last,
3064 /* FIXME: This is a workaround for page locking bug when some pages are
3065 * invalid during migration to VRAM
3067 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm,
3070 struct hmm_range *hmm_range;
3073 if (prange->validated_once)
3076 r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
3077 prange->start << PAGE_SHIFT,
3078 prange->npages, &hmm_range,
3079 false, true, owner);
3081 amdgpu_hmm_range_get_pages_done(hmm_range);
3082 prange->validated_once = true;
3086 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3087 * @mm: current process mm_struct
3088 * @prange: svm range structure
3089 * @migrated: output, true if migration is triggered
3091 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3093 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3096 * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3098 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3099 * stops all queues, schedule restore work
3100 * 2. svm_range_restore_work wait for migration is done by
3101 * a. svm_range_validate_vram takes prange->migrate_mutex
3102 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3103 * 3. restore work update mappings of GPU, resume all queues.
3105 * Context: Process context
3108 * 0 - OK, otherwise - error code of migration
3111 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3118 best_loc = svm_range_best_prefetch_location(prange);
3120 if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3121 best_loc == prange->actual_loc)
3125 r = svm_migrate_vram_to_ram(prange, mm);
3130 r = svm_migrate_to_vram(prange, best_loc, mm);
3136 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3141 if (dma_fence_is_signaled(&fence->base))
3144 if (fence->svm_bo) {
3145 WRITE_ONCE(fence->svm_bo->evicting, 1);
3146 schedule_work(&fence->svm_bo->eviction_work);
3152 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3154 struct svm_range_bo *svm_bo;
3155 struct kfd_process *p;
3156 struct mm_struct *mm;
3159 svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3160 if (!svm_bo_ref_unless_zero(svm_bo))
3161 return; /* svm_bo was freed while eviction was pending */
3163 /* svm_range_bo_release destroys this worker thread. So during
3164 * the lifetime of this thread, kfd_process and mm will be valid.
3166 p = container_of(svm_bo->svms, struct kfd_process, svms);
3172 spin_lock(&svm_bo->list_lock);
3173 while (!list_empty(&svm_bo->range_list) && !r) {
3174 struct svm_range *prange =
3175 list_first_entry(&svm_bo->range_list,
3176 struct svm_range, svm_bo_list);
3179 list_del_init(&prange->svm_bo_list);
3180 spin_unlock(&svm_bo->list_lock);
3182 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3183 prange->start, prange->last);
3185 mutex_lock(&prange->migrate_mutex);
3187 r = svm_migrate_vram_to_ram(prange,
3188 svm_bo->eviction_fence->mm);
3189 } while (!r && prange->actual_loc && --retries);
3191 if (!r && prange->actual_loc)
3192 pr_info_once("Migration failed during eviction");
3194 if (!prange->actual_loc) {
3195 mutex_lock(&prange->lock);
3196 prange->svm_bo = NULL;
3197 mutex_unlock(&prange->lock);
3199 mutex_unlock(&prange->migrate_mutex);
3201 spin_lock(&svm_bo->list_lock);
3203 spin_unlock(&svm_bo->list_lock);
3204 mmap_read_unlock(mm);
3206 dma_fence_signal(&svm_bo->eviction_fence->base);
3208 /* This is the last reference to svm_bo, after svm_range_vram_node_free
3209 * has been called in svm_migrate_vram_to_ram
3211 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3212 svm_range_bo_unref(svm_bo);
3216 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3217 uint64_t start, uint64_t size, uint32_t nattr,
3218 struct kfd_ioctl_svm_attribute *attrs)
3220 struct amdkfd_process_info *process_info = p->kgd_process_info;
3221 struct list_head update_list;
3222 struct list_head insert_list;
3223 struct list_head remove_list;
3224 struct svm_range_list *svms;
3225 struct svm_range *prange;
3226 struct svm_range *next;
3229 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3230 p->pasid, &p->svms, start, start + size - 1, size);
3232 r = svm_range_check_attr(p, nattr, attrs);
3238 mutex_lock(&process_info->lock);
3240 svm_range_list_lock_and_flush_work(svms, mm);
3242 r = svm_range_is_valid(p, start, size);
3244 pr_debug("invalid range r=%d\n", r);
3245 mmap_write_unlock(mm);
3249 mutex_lock(&svms->lock);
3251 /* Add new range and split existing ranges as needed */
3252 r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3253 &insert_list, &remove_list);
3255 mutex_unlock(&svms->lock);
3256 mmap_write_unlock(mm);
3259 /* Apply changes as a transaction */
3260 list_for_each_entry_safe(prange, next, &insert_list, list) {
3261 svm_range_add_to_svms(prange);
3262 svm_range_add_notifier_locked(mm, prange);
3264 list_for_each_entry(prange, &update_list, update_list) {
3265 svm_range_apply_attrs(p, prange, nattr, attrs);
3266 /* TODO: unmap ranges from GPU that lost access */
3268 list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3269 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3270 prange->svms, prange, prange->start,
3272 svm_range_unlink(prange);
3273 svm_range_remove_notifier(prange);
3274 svm_range_free(prange);
3277 mmap_write_downgrade(mm);
3278 /* Trigger migrations and revalidate and map to GPUs as needed. If
3279 * this fails we may be left with partially completed actions. There
3280 * is no clean way of rolling back to the previous state in such a
3281 * case because the rollback wouldn't be guaranteed to work either.
3283 list_for_each_entry(prange, &update_list, update_list) {
3286 mutex_lock(&prange->migrate_mutex);
3288 r = svm_range_trigger_migration(mm, prange, &migrated);
3290 goto out_unlock_range;
3292 if (migrated && !p->xnack_enabled) {
3293 pr_debug("restore_work will update mappings of GPUs\n");
3294 mutex_unlock(&prange->migrate_mutex);
3298 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3301 pr_debug("failed %d to map svm range\n", r);
3304 mutex_unlock(&prange->migrate_mutex);
3309 svm_range_debug_dump(svms);
3311 mutex_unlock(&svms->lock);
3312 mmap_read_unlock(mm);
3314 mutex_unlock(&process_info->lock);
3316 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3317 &p->svms, start, start + size - 1, r);
3323 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3324 uint64_t start, uint64_t size, uint32_t nattr,
3325 struct kfd_ioctl_svm_attribute *attrs)
3327 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3328 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3329 bool get_preferred_loc = false;
3330 bool get_prefetch_loc = false;
3331 bool get_granularity = false;
3332 bool get_accessible = false;
3333 bool get_flags = false;
3334 uint64_t last = start + size - 1UL;
3335 uint8_t granularity = 0xff;
3336 struct interval_tree_node *node;
3337 struct svm_range_list *svms;
3338 struct svm_range *prange;
3339 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3340 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3341 uint32_t flags_and = 0xffffffff;
3342 uint32_t flags_or = 0;
3347 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3348 start + size - 1, nattr);
3350 /* Flush pending deferred work to avoid racing with deferred actions from
3351 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3352 * can still race with get_attr because we don't hold the mmap lock. But that
3353 * would be a race condition in the application anyway, and undefined
3354 * behaviour is acceptable in that case.
3356 flush_work(&p->svms.deferred_list_work);
3359 r = svm_range_is_valid(p, start, size);
3360 mmap_read_unlock(mm);
3362 pr_debug("invalid range r=%d\n", r);
3366 for (i = 0; i < nattr; i++) {
3367 switch (attrs[i].type) {
3368 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3369 get_preferred_loc = true;
3371 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3372 get_prefetch_loc = true;
3374 case KFD_IOCTL_SVM_ATTR_ACCESS:
3375 get_accessible = true;
3377 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3378 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3381 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3382 get_granularity = true;
3384 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3385 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3388 pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3395 mutex_lock(&svms->lock);
3397 node = interval_tree_iter_first(&svms->objects, start, last);
3399 pr_debug("range attrs not found return default values\n");
3400 svm_range_set_default_attributes(&location, &prefetch_loc,
3401 &granularity, &flags_and);
3402 flags_or = flags_and;
3403 if (p->xnack_enabled)
3404 bitmap_copy(bitmap_access, svms->bitmap_supported,
3407 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3408 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3411 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3412 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3415 struct interval_tree_node *next;
3417 prange = container_of(node, struct svm_range, it_node);
3418 next = interval_tree_iter_next(node, start, last);
3420 if (get_preferred_loc) {
3421 if (prange->preferred_loc ==
3422 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3423 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3424 location != prange->preferred_loc)) {
3425 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3426 get_preferred_loc = false;
3428 location = prange->preferred_loc;
3431 if (get_prefetch_loc) {
3432 if (prange->prefetch_loc ==
3433 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3434 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3435 prefetch_loc != prange->prefetch_loc)) {
3436 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3437 get_prefetch_loc = false;
3439 prefetch_loc = prange->prefetch_loc;
3442 if (get_accessible) {
3443 bitmap_and(bitmap_access, bitmap_access,
3444 prange->bitmap_access, MAX_GPU_INSTANCE);
3445 bitmap_and(bitmap_aip, bitmap_aip,
3446 prange->bitmap_aip, MAX_GPU_INSTANCE);
3449 flags_and &= prange->flags;
3450 flags_or |= prange->flags;
3453 if (get_granularity && prange->granularity < granularity)
3454 granularity = prange->granularity;
3459 mutex_unlock(&svms->lock);
3461 for (i = 0; i < nattr; i++) {
3462 switch (attrs[i].type) {
3463 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3464 attrs[i].value = location;
3466 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3467 attrs[i].value = prefetch_loc;
3469 case KFD_IOCTL_SVM_ATTR_ACCESS:
3470 gpuidx = kfd_process_gpuidx_from_gpuid(p,
3473 pr_debug("invalid gpuid %x\n", attrs[i].value);
3476 if (test_bit(gpuidx, bitmap_access))
3477 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3478 else if (test_bit(gpuidx, bitmap_aip))
3480 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3482 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3484 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3485 attrs[i].value = flags_and;
3487 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3488 attrs[i].value = ~flags_or;
3490 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3491 attrs[i].value = (uint32_t)granularity;
3499 int kfd_criu_resume_svm(struct kfd_process *p)
3501 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3502 int nattr_common = 4, nattr_accessibility = 1;
3503 struct criu_svm_metadata *criu_svm_md = NULL;
3504 struct svm_range_list *svms = &p->svms;
3505 struct criu_svm_metadata *next = NULL;
3506 uint32_t set_flags = 0xffffffff;
3507 int i, j, num_attrs, ret = 0;
3508 uint64_t set_attr_size;
3509 struct mm_struct *mm;
3511 if (list_empty(&svms->criu_svm_metadata_list)) {
3512 pr_debug("No SVM data from CRIU restore stage 2\n");
3516 mm = get_task_mm(p->lead_thread);
3518 pr_err("failed to get mm for the target process\n");
3522 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3525 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3526 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3527 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3529 for (j = 0; j < num_attrs; j++) {
3530 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3531 i, j, criu_svm_md->data.attrs[j].type,
3532 i, j, criu_svm_md->data.attrs[j].value);
3533 switch (criu_svm_md->data.attrs[j].type) {
3534 /* During Checkpoint operation, the query for
3535 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3536 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3537 * not used by the range which was checkpointed. Care
3538 * must be taken to not restore with an invalid value
3539 * otherwise the gpuidx value will be invalid and
3540 * set_attr would eventually fail so just replace those
3541 * with another dummy attribute such as
3542 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3544 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3545 if (criu_svm_md->data.attrs[j].value ==
3546 KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3547 criu_svm_md->data.attrs[j].type =
3548 KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3549 criu_svm_md->data.attrs[j].value = 0;
3552 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3553 set_flags = criu_svm_md->data.attrs[j].value;
3560 /* CLR_FLAGS is not available via get_attr during checkpoint but
3561 * it needs to be inserted before restoring the ranges so
3562 * allocate extra space for it before calling set_attr
3564 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3566 set_attr_new = krealloc(set_attr, set_attr_size,
3568 if (!set_attr_new) {
3572 set_attr = set_attr_new;
3574 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3575 sizeof(struct kfd_ioctl_svm_attribute));
3576 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3577 set_attr[num_attrs].value = ~set_flags;
3579 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3580 criu_svm_md->data.size, num_attrs + 1,
3583 pr_err("CRIU: failed to set range attributes\n");
3591 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3592 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3593 criu_svm_md->data.start_addr);
3602 int kfd_criu_restore_svm(struct kfd_process *p,
3603 uint8_t __user *user_priv_ptr,
3604 uint64_t *priv_data_offset,
3605 uint64_t max_priv_data_size)
3607 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3608 int nattr_common = 4, nattr_accessibility = 1;
3609 struct criu_svm_metadata *criu_svm_md = NULL;
3610 struct svm_range_list *svms = &p->svms;
3611 uint32_t num_devices;
3614 num_devices = p->n_pdds;
3615 /* Handle one SVM range object at a time, also the number of gpus are
3616 * assumed to be same on the restore node, checking must be done while
3617 * evaluating the topology earlier
3620 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3621 (nattr_common + nattr_accessibility * num_devices);
3622 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3624 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3627 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3629 pr_err("failed to allocate memory to store svm metadata\n");
3632 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3637 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3638 svm_priv_data_size);
3643 *priv_data_offset += svm_priv_data_size;
3645 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3655 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3656 uint64_t *svm_priv_data_size)
3658 uint64_t total_size, accessibility_size, common_attr_size;
3659 int nattr_common = 4, nattr_accessibility = 1;
3660 int num_devices = p->n_pdds;
3661 struct svm_range_list *svms;
3662 struct svm_range *prange;
3665 *svm_priv_data_size = 0;
3671 mutex_lock(&svms->lock);
3672 list_for_each_entry(prange, &svms->list, list) {
3673 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3674 prange, prange->start, prange->npages,
3675 prange->start + prange->npages - 1);
3678 mutex_unlock(&svms->lock);
3680 *num_svm_ranges = count;
3681 /* Only the accessbility attributes need to be queried for all the gpus
3682 * individually, remaining ones are spanned across the entire process
3683 * regardless of the various gpu nodes. Of the remaining attributes,
3684 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3686 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3687 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3688 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3689 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3691 * ** ACCESSBILITY ATTRIBUTES **
3692 * (Considered as one, type is altered during query, value is gpuid)
3693 * KFD_IOCTL_SVM_ATTR_ACCESS
3694 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3695 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3697 if (*num_svm_ranges > 0) {
3698 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3700 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3701 nattr_accessibility * num_devices;
3703 total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3704 common_attr_size + accessibility_size;
3706 *svm_priv_data_size = *num_svm_ranges * total_size;
3709 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
3710 *svm_priv_data_size);
3714 int kfd_criu_checkpoint_svm(struct kfd_process *p,
3715 uint8_t __user *user_priv_data,
3716 uint64_t *priv_data_offset)
3718 struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
3719 struct kfd_ioctl_svm_attribute *query_attr = NULL;
3720 uint64_t svm_priv_data_size, query_attr_size = 0;
3721 int index, nattr_common = 4, ret = 0;
3722 struct svm_range_list *svms;
3723 int num_devices = p->n_pdds;
3724 struct svm_range *prange;
3725 struct mm_struct *mm;
3731 mm = get_task_mm(p->lead_thread);
3733 pr_err("failed to get mm for the target process\n");
3737 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3738 (nattr_common + num_devices);
3740 query_attr = kzalloc(query_attr_size, GFP_KERNEL);
3746 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
3747 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
3748 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3749 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
3751 for (index = 0; index < num_devices; index++) {
3752 struct kfd_process_device *pdd = p->pdds[index];
3754 query_attr[index + nattr_common].type =
3755 KFD_IOCTL_SVM_ATTR_ACCESS;
3756 query_attr[index + nattr_common].value = pdd->user_gpu_id;
3759 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
3761 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
3768 list_for_each_entry(prange, &svms->list, list) {
3770 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
3771 svm_priv->start_addr = prange->start;
3772 svm_priv->size = prange->npages;
3773 memcpy(&svm_priv->attrs, query_attr, query_attr_size);
3774 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
3775 prange, prange->start, prange->npages,
3776 prange->start + prange->npages - 1,
3777 prange->npages * PAGE_SIZE);
3779 ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
3781 (nattr_common + num_devices),
3784 pr_err("CRIU: failed to obtain range attributes\n");
3788 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
3789 svm_priv_data_size)) {
3790 pr_err("Failed to copy svm priv to user\n");
3795 *priv_data_offset += svm_priv_data_size;
3810 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3811 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3813 struct mm_struct *mm = current->mm;
3816 start >>= PAGE_SHIFT;
3817 size >>= PAGE_SHIFT;
3820 case KFD_IOCTL_SVM_OP_SET_ATTR:
3821 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
3823 case KFD_IOCTL_SVM_OP_GET_ATTR:
3824 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);