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"
35 #include "kfd_smi_events.h"
40 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
42 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
44 /* Long enough to ensure no retry fault comes after svm range is restored and
45 * page table is updated.
47 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC)
49 /* Giant svm range split into smaller ranges based on this, it is decided using
50 * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to
53 static uint64_t max_svm_range_pages;
55 struct criu_svm_metadata {
56 struct list_head list;
57 struct kfd_criu_svm_range_priv_data data;
60 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
62 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
63 const struct mmu_notifier_range *range,
64 unsigned long cur_seq);
66 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
67 uint64_t *bo_s, uint64_t *bo_l);
68 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
69 .invalidate = svm_range_cpu_invalidate_pagetables,
73 * svm_range_unlink - unlink svm_range from lists and interval tree
74 * @prange: svm range structure to be removed
76 * Remove the svm_range from the svms and svm_bo lists and the svms
79 * Context: The caller must hold svms->lock
81 static void svm_range_unlink(struct svm_range *prange)
83 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
84 prange, prange->start, prange->last);
87 spin_lock(&prange->svm_bo->list_lock);
88 list_del(&prange->svm_bo_list);
89 spin_unlock(&prange->svm_bo->list_lock);
92 list_del(&prange->list);
93 if (prange->it_node.start != 0 && prange->it_node.last != 0)
94 interval_tree_remove(&prange->it_node, &prange->svms->objects);
98 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
100 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
101 prange, prange->start, prange->last);
103 mmu_interval_notifier_insert_locked(&prange->notifier, mm,
104 prange->start << PAGE_SHIFT,
105 prange->npages << PAGE_SHIFT,
110 * svm_range_add_to_svms - add svm range to svms
111 * @prange: svm range structure to be added
113 * Add the svm range to svms interval tree and link list
115 * Context: The caller must hold svms->lock
117 static void svm_range_add_to_svms(struct svm_range *prange)
119 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
120 prange, prange->start, prange->last);
122 list_move_tail(&prange->list, &prange->svms->list);
123 prange->it_node.start = prange->start;
124 prange->it_node.last = prange->last;
125 interval_tree_insert(&prange->it_node, &prange->svms->objects);
128 static void svm_range_remove_notifier(struct svm_range *prange)
130 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
131 prange->svms, prange,
132 prange->notifier.interval_tree.start >> PAGE_SHIFT,
133 prange->notifier.interval_tree.last >> PAGE_SHIFT);
135 if (prange->notifier.interval_tree.start != 0 &&
136 prange->notifier.interval_tree.last != 0)
137 mmu_interval_notifier_remove(&prange->notifier);
141 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
143 return dma_addr && !dma_mapping_error(dev, dma_addr) &&
144 !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
148 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
149 unsigned long offset, unsigned long npages,
150 unsigned long *hmm_pfns, uint32_t gpuidx)
152 enum dma_data_direction dir = DMA_BIDIRECTIONAL;
153 dma_addr_t *addr = prange->dma_addr[gpuidx];
154 struct device *dev = adev->dev;
159 addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL);
162 prange->dma_addr[gpuidx] = addr;
166 for (i = 0; i < npages; i++) {
167 if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
168 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
170 page = hmm_pfn_to_page(hmm_pfns[i]);
171 if (is_zone_device_page(page)) {
172 struct amdgpu_device *bo_adev =
173 amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
175 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
176 bo_adev->vm_manager.vram_base_offset -
177 bo_adev->kfd.dev->pgmap.range.start;
178 addr[i] |= SVM_RANGE_VRAM_DOMAIN;
179 pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]);
182 addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
183 r = dma_mapping_error(dev, addr[i]);
185 dev_err(dev, "failed %d dma_map_page\n", r);
188 pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n",
189 addr[i] >> PAGE_SHIFT, page_to_pfn(page));
195 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
196 unsigned long offset, unsigned long npages,
197 unsigned long *hmm_pfns)
199 struct kfd_process *p;
203 p = container_of(prange->svms, struct kfd_process, svms);
205 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
206 struct kfd_process_device *pdd;
208 pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
209 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
211 pr_debug("failed to find device idx %d\n", gpuidx);
215 r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
224 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
225 unsigned long offset, unsigned long npages)
227 enum dma_data_direction dir = DMA_BIDIRECTIONAL;
233 for (i = offset; i < offset + npages; i++) {
234 if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i]))
236 pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT);
237 dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir);
242 void svm_range_free_dma_mappings(struct svm_range *prange)
244 struct kfd_process_device *pdd;
245 dma_addr_t *dma_addr;
247 struct kfd_process *p;
250 p = container_of(prange->svms, struct kfd_process, svms);
252 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) {
253 dma_addr = prange->dma_addr[gpuidx];
257 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
259 pr_debug("failed to find device idx %d\n", gpuidx);
262 dev = &pdd->dev->pdev->dev;
263 svm_range_dma_unmap(dev, dma_addr, 0, prange->npages);
265 prange->dma_addr[gpuidx] = NULL;
269 static void svm_range_free(struct svm_range *prange, bool update_mem_usage)
271 uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT;
272 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
274 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange,
275 prange->start, prange->last);
277 svm_range_vram_node_free(prange);
278 svm_range_free_dma_mappings(prange);
280 if (update_mem_usage && !p->xnack_enabled) {
281 pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size);
282 amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
283 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
285 mutex_destroy(&prange->lock);
286 mutex_destroy(&prange->migrate_mutex);
291 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc,
292 uint8_t *granularity, uint32_t *flags)
294 *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
295 *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
298 KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT;
302 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
303 uint64_t last, bool update_mem_usage)
305 uint64_t size = last - start + 1;
306 struct svm_range *prange;
307 struct kfd_process *p;
309 prange = kzalloc(sizeof(*prange), GFP_KERNEL);
313 p = container_of(svms, struct kfd_process, svms);
314 if (!p->xnack_enabled && update_mem_usage &&
315 amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT,
316 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) {
317 pr_info("SVM mapping failed, exceeds resident system memory limit\n");
321 prange->npages = size;
323 prange->start = start;
325 INIT_LIST_HEAD(&prange->list);
326 INIT_LIST_HEAD(&prange->update_list);
327 INIT_LIST_HEAD(&prange->svm_bo_list);
328 INIT_LIST_HEAD(&prange->deferred_list);
329 INIT_LIST_HEAD(&prange->child_list);
330 atomic_set(&prange->invalid, 0);
331 prange->validate_timestamp = 0;
332 mutex_init(&prange->migrate_mutex);
333 mutex_init(&prange->lock);
335 if (p->xnack_enabled)
336 bitmap_copy(prange->bitmap_access, svms->bitmap_supported,
339 svm_range_set_default_attributes(&prange->preferred_loc,
340 &prange->prefetch_loc,
341 &prange->granularity, &prange->flags);
343 pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last);
348 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo)
350 if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref))
356 static void svm_range_bo_release(struct kref *kref)
358 struct svm_range_bo *svm_bo;
360 svm_bo = container_of(kref, struct svm_range_bo, kref);
361 pr_debug("svm_bo 0x%p\n", svm_bo);
363 spin_lock(&svm_bo->list_lock);
364 while (!list_empty(&svm_bo->range_list)) {
365 struct svm_range *prange =
366 list_first_entry(&svm_bo->range_list,
367 struct svm_range, svm_bo_list);
368 /* list_del_init tells a concurrent svm_range_vram_node_new when
369 * it's safe to reuse the svm_bo pointer and svm_bo_list head.
371 list_del_init(&prange->svm_bo_list);
372 spin_unlock(&svm_bo->list_lock);
374 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
375 prange->start, prange->last);
376 mutex_lock(&prange->lock);
377 prange->svm_bo = NULL;
378 mutex_unlock(&prange->lock);
380 spin_lock(&svm_bo->list_lock);
382 spin_unlock(&svm_bo->list_lock);
383 if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) {
384 /* We're not in the eviction worker.
385 * Signal the fence and synchronize with any
386 * pending eviction work.
388 dma_fence_signal(&svm_bo->eviction_fence->base);
389 cancel_work_sync(&svm_bo->eviction_work);
391 dma_fence_put(&svm_bo->eviction_fence->base);
392 amdgpu_bo_unref(&svm_bo->bo);
396 static void svm_range_bo_wq_release(struct work_struct *work)
398 struct svm_range_bo *svm_bo;
400 svm_bo = container_of(work, struct svm_range_bo, release_work);
401 svm_range_bo_release(&svm_bo->kref);
404 static void svm_range_bo_release_async(struct kref *kref)
406 struct svm_range_bo *svm_bo;
408 svm_bo = container_of(kref, struct svm_range_bo, kref);
409 pr_debug("svm_bo 0x%p\n", svm_bo);
410 INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release);
411 schedule_work(&svm_bo->release_work);
414 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo)
416 kref_put(&svm_bo->kref, svm_range_bo_release_async);
419 static void svm_range_bo_unref(struct svm_range_bo *svm_bo)
422 kref_put(&svm_bo->kref, svm_range_bo_release);
426 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange)
428 struct amdgpu_device *bo_adev;
430 mutex_lock(&prange->lock);
431 if (!prange->svm_bo) {
432 mutex_unlock(&prange->lock);
435 if (prange->ttm_res) {
436 /* We still have a reference, all is well */
437 mutex_unlock(&prange->lock);
440 if (svm_bo_ref_unless_zero(prange->svm_bo)) {
442 * Migrate from GPU to GPU, remove range from source bo_adev
443 * svm_bo range list, and return false to allocate svm_bo from
446 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
447 if (bo_adev != adev) {
448 mutex_unlock(&prange->lock);
450 spin_lock(&prange->svm_bo->list_lock);
451 list_del_init(&prange->svm_bo_list);
452 spin_unlock(&prange->svm_bo->list_lock);
454 svm_range_bo_unref(prange->svm_bo);
457 if (READ_ONCE(prange->svm_bo->evicting)) {
459 struct svm_range_bo *svm_bo;
460 /* The BO is getting evicted,
461 * we need to get a new one
463 mutex_unlock(&prange->lock);
464 svm_bo = prange->svm_bo;
465 f = dma_fence_get(&svm_bo->eviction_fence->base);
466 svm_range_bo_unref(prange->svm_bo);
467 /* wait for the fence to avoid long spin-loop
468 * at list_empty_careful
470 dma_fence_wait(f, false);
473 /* The BO was still around and we got
474 * a new reference to it
476 mutex_unlock(&prange->lock);
477 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
478 prange->svms, prange->start, prange->last);
480 prange->ttm_res = prange->svm_bo->bo->tbo.resource;
485 mutex_unlock(&prange->lock);
488 /* We need a new svm_bo. Spin-loop to wait for concurrent
489 * svm_range_bo_release to finish removing this range from
490 * its range list. After this, it is safe to reuse the
491 * svm_bo pointer and svm_bo_list head.
493 while (!list_empty_careful(&prange->svm_bo_list))
499 static struct svm_range_bo *svm_range_bo_new(void)
501 struct svm_range_bo *svm_bo;
503 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
507 kref_init(&svm_bo->kref);
508 INIT_LIST_HEAD(&svm_bo->range_list);
509 spin_lock_init(&svm_bo->list_lock);
515 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
518 struct amdgpu_bo_param bp;
519 struct svm_range_bo *svm_bo;
520 struct amdgpu_bo_user *ubo;
521 struct amdgpu_bo *bo;
522 struct kfd_process *p;
523 struct mm_struct *mm;
526 p = container_of(prange->svms, struct kfd_process, svms);
527 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
528 prange->start, prange->last);
530 if (svm_range_validate_svm_bo(adev, prange))
533 svm_bo = svm_range_bo_new();
535 pr_debug("failed to alloc svm bo\n");
538 mm = get_task_mm(p->lead_thread);
540 pr_debug("failed to get mm\n");
544 svm_bo->eviction_fence =
545 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
549 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
550 svm_bo->evicting = 0;
551 memset(&bp, 0, sizeof(bp));
552 bp.size = prange->npages * PAGE_SIZE;
553 bp.byte_align = PAGE_SIZE;
554 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
555 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
556 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
557 bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
558 bp.type = ttm_bo_type_device;
561 r = amdgpu_bo_create_user(adev, &bp, &ubo);
563 pr_debug("failed %d to create bo\n", r);
564 goto create_bo_failed;
567 r = amdgpu_bo_reserve(bo, true);
569 pr_debug("failed %d to reserve bo\n", r);
570 goto reserve_bo_failed;
574 r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
576 pr_debug("failed %d to sync bo\n", r);
577 amdgpu_bo_unreserve(bo);
578 goto reserve_bo_failed;
582 r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
584 pr_debug("failed %d to reserve bo\n", r);
585 amdgpu_bo_unreserve(bo);
586 goto reserve_bo_failed;
588 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
590 amdgpu_bo_unreserve(bo);
593 prange->svm_bo = svm_bo;
594 prange->ttm_res = bo->tbo.resource;
597 spin_lock(&svm_bo->list_lock);
598 list_add(&prange->svm_bo_list, &svm_bo->range_list);
599 spin_unlock(&svm_bo->list_lock);
604 amdgpu_bo_unref(&bo);
606 dma_fence_put(&svm_bo->eviction_fence->base);
608 prange->ttm_res = NULL;
613 void svm_range_vram_node_free(struct svm_range *prange)
615 svm_range_bo_unref(prange->svm_bo);
616 prange->ttm_res = NULL;
619 struct amdgpu_device *
620 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
622 struct kfd_process_device *pdd;
623 struct kfd_process *p;
626 p = container_of(prange->svms, struct kfd_process, svms);
628 gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id);
630 pr_debug("failed to get device by id 0x%x\n", gpu_id);
633 pdd = kfd_process_device_from_gpuidx(p, gpu_idx);
635 pr_debug("failed to get device by idx 0x%x\n", gpu_idx);
639 return pdd->dev->adev;
642 struct kfd_process_device *
643 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
645 struct kfd_process *p;
646 int32_t gpu_idx, gpuid;
649 p = container_of(prange->svms, struct kfd_process, svms);
651 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx);
653 pr_debug("failed to get device id by adev %p\n", adev);
657 return kfd_process_device_from_gpuidx(p, gpu_idx);
660 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
662 struct ttm_operation_ctx ctx = { false, false };
664 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
666 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
670 svm_range_check_attr(struct kfd_process *p,
671 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
675 for (i = 0; i < nattr; i++) {
676 uint32_t val = attrs[i].value;
677 int gpuidx = MAX_GPU_INSTANCE;
679 switch (attrs[i].type) {
680 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
681 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
682 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
683 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
685 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
686 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
687 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
689 case KFD_IOCTL_SVM_ATTR_ACCESS:
690 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
691 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
692 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
694 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
696 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
698 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
701 pr_debug("unknown attr type 0x%x\n", attrs[i].type);
706 pr_debug("no GPU 0x%x found\n", val);
708 } else if (gpuidx < MAX_GPU_INSTANCE &&
709 !test_bit(gpuidx, p->svms.bitmap_supported)) {
710 pr_debug("GPU 0x%x not supported\n", val);
719 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
720 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
721 bool *update_mapping)
726 for (i = 0; i < nattr; i++) {
727 switch (attrs[i].type) {
728 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
729 prange->preferred_loc = attrs[i].value;
731 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
732 prange->prefetch_loc = attrs[i].value;
734 case KFD_IOCTL_SVM_ATTR_ACCESS:
735 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
736 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
737 *update_mapping = true;
738 gpuidx = kfd_process_gpuidx_from_gpuid(p,
740 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
741 bitmap_clear(prange->bitmap_access, gpuidx, 1);
742 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
743 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
744 bitmap_set(prange->bitmap_access, gpuidx, 1);
745 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
747 bitmap_clear(prange->bitmap_access, gpuidx, 1);
748 bitmap_set(prange->bitmap_aip, gpuidx, 1);
751 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
752 *update_mapping = true;
753 prange->flags |= attrs[i].value;
755 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
756 *update_mapping = true;
757 prange->flags &= ~attrs[i].value;
759 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
760 prange->granularity = attrs[i].value;
763 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
769 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
770 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
775 for (i = 0; i < nattr; i++) {
776 switch (attrs[i].type) {
777 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
778 if (prange->preferred_loc != attrs[i].value)
781 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
782 /* Prefetch should always trigger a migration even
783 * if the value of the attribute didn't change.
786 case KFD_IOCTL_SVM_ATTR_ACCESS:
787 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
788 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
789 gpuidx = kfd_process_gpuidx_from_gpuid(p,
791 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
792 if (test_bit(gpuidx, prange->bitmap_access) ||
793 test_bit(gpuidx, prange->bitmap_aip))
795 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
796 if (!test_bit(gpuidx, prange->bitmap_access))
799 if (!test_bit(gpuidx, prange->bitmap_aip))
803 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
804 if ((prange->flags & attrs[i].value) != attrs[i].value)
807 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
808 if ((prange->flags & attrs[i].value) != 0)
811 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
812 if (prange->granularity != attrs[i].value)
816 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
824 * svm_range_debug_dump - print all range information from svms
825 * @svms: svm range list header
827 * debug output svm range start, end, prefetch location from svms
828 * interval tree and link list
830 * Context: The caller must hold svms->lock
832 static void svm_range_debug_dump(struct svm_range_list *svms)
834 struct interval_tree_node *node;
835 struct svm_range *prange;
837 pr_debug("dump svms 0x%p list\n", svms);
838 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
840 list_for_each_entry(prange, &svms->list, list) {
841 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
842 prange, prange->start, prange->npages,
843 prange->start + prange->npages - 1,
847 pr_debug("dump svms 0x%p interval tree\n", svms);
848 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
849 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
851 prange = container_of(node, struct svm_range, it_node);
852 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
853 prange, prange->start, prange->npages,
854 prange->start + prange->npages - 1,
856 node = interval_tree_iter_next(node, 0, ~0ULL);
861 svm_range_split_array(void *ppnew, void *ppold, size_t size,
862 uint64_t old_start, uint64_t old_n,
863 uint64_t new_start, uint64_t new_n)
865 unsigned char *new, *old, *pold;
870 pold = *(unsigned char **)ppold;
874 new = kvmalloc_array(new_n, size, GFP_KERNEL);
878 d = (new_start - old_start) * size;
879 memcpy(new, pold + d, new_n * size);
881 old = kvmalloc_array(old_n, size, GFP_KERNEL);
887 d = (new_start == old_start) ? new_n * size : 0;
888 memcpy(old, pold + d, old_n * size);
891 *(void **)ppold = old;
892 *(void **)ppnew = new;
898 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
899 uint64_t start, uint64_t last)
901 uint64_t npages = last - start + 1;
904 for (i = 0; i < MAX_GPU_INSTANCE; i++) {
905 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
906 sizeof(*old->dma_addr[i]), old->start,
907 npages, new->start, new->npages);
916 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
917 uint64_t start, uint64_t last)
919 uint64_t npages = last - start + 1;
921 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
922 new->svms, new, new->start, start, last);
924 if (new->start == old->start) {
925 new->offset = old->offset;
926 old->offset += new->npages;
928 new->offset = old->offset + npages;
931 new->svm_bo = svm_range_bo_ref(old->svm_bo);
932 new->ttm_res = old->ttm_res;
934 spin_lock(&new->svm_bo->list_lock);
935 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
936 spin_unlock(&new->svm_bo->list_lock);
942 * svm_range_split_adjust - split range and adjust
945 * @old: the old range
946 * @start: the old range adjust to start address in pages
947 * @last: the old range adjust to last address in pages
949 * Copy system memory dma_addr or vram ttm_res in old range to new
950 * range from new_start up to size new->npages, the remaining old range is from
954 * 0 - OK, -ENOMEM - out of memory
957 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
958 uint64_t start, uint64_t last)
962 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
963 new->svms, new->start, old->start, old->last, start, last);
965 if (new->start < old->start ||
966 new->last > old->last) {
967 WARN_ONCE(1, "invalid new range start or last\n");
971 r = svm_range_split_pages(new, old, start, last);
975 if (old->actual_loc && old->ttm_res) {
976 r = svm_range_split_nodes(new, old, start, last);
981 old->npages = last - start + 1;
984 new->flags = old->flags;
985 new->preferred_loc = old->preferred_loc;
986 new->prefetch_loc = old->prefetch_loc;
987 new->actual_loc = old->actual_loc;
988 new->granularity = old->granularity;
989 new->mapped_to_gpu = old->mapped_to_gpu;
990 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
991 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
997 * svm_range_split - split a range in 2 ranges
999 * @prange: the svm range to split
1000 * @start: the remaining range start address in pages
1001 * @last: the remaining range last address in pages
1002 * @new: the result new range generated
1005 * case 1: if start == prange->start
1006 * prange ==> prange[start, last]
1007 * new range [last + 1, prange->last]
1009 * case 2: if last == prange->last
1010 * prange ==> prange[start, last]
1011 * new range [prange->start, start - 1]
1014 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1017 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1018 struct svm_range **new)
1020 uint64_t old_start = prange->start;
1021 uint64_t old_last = prange->last;
1022 struct svm_range_list *svms;
1025 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1026 old_start, old_last, start, last);
1028 if (old_start != start && old_last != last)
1030 if (start < old_start || last > old_last)
1033 svms = prange->svms;
1034 if (old_start == start)
1035 *new = svm_range_new(svms, last + 1, old_last, false);
1037 *new = svm_range_new(svms, old_start, start - 1, false);
1041 r = svm_range_split_adjust(*new, prange, start, last);
1043 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1044 r, old_start, old_last, start, last);
1045 svm_range_free(*new, false);
1053 svm_range_split_tail(struct svm_range *prange,
1054 uint64_t new_last, struct list_head *insert_list)
1056 struct svm_range *tail;
1057 int r = svm_range_split(prange, prange->start, new_last, &tail);
1060 list_add(&tail->list, insert_list);
1065 svm_range_split_head(struct svm_range *prange,
1066 uint64_t new_start, struct list_head *insert_list)
1068 struct svm_range *head;
1069 int r = svm_range_split(prange, new_start, prange->last, &head);
1072 list_add(&head->list, insert_list);
1077 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1078 struct svm_range *pchild, enum svm_work_list_ops op)
1080 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1081 pchild, pchild->start, pchild->last, prange, op);
1083 pchild->work_item.mm = mm;
1084 pchild->work_item.op = op;
1085 list_add_tail(&pchild->child_list, &prange->child_list);
1089 * svm_range_split_by_granularity - collect ranges within granularity boundary
1091 * @p: the process with svms list
1093 * @addr: the vm fault address in pages, to split the prange
1094 * @parent: parent range if prange is from child list
1095 * @prange: prange to split
1097 * Trims @prange to be a single aligned block of prange->granularity if
1098 * possible. The head and tail are added to the child_list in @parent.
1100 * Context: caller must hold mmap_read_lock and prange->lock
1103 * 0 - OK, otherwise error code
1106 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1107 unsigned long addr, struct svm_range *parent,
1108 struct svm_range *prange)
1110 struct svm_range *head, *tail;
1111 unsigned long start, last, size;
1114 /* Align splited range start and size to granularity size, then a single
1115 * PTE will be used for whole range, this reduces the number of PTE
1116 * updated and the L1 TLB space used for translation.
1118 size = 1UL << prange->granularity;
1119 start = ALIGN_DOWN(addr, size);
1120 last = ALIGN(addr + 1, size) - 1;
1122 pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1123 prange->svms, prange->start, prange->last, start, last, size);
1125 if (start > prange->start) {
1126 r = svm_range_split(prange, start, prange->last, &head);
1129 svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1132 if (last < prange->last) {
1133 r = svm_range_split(prange, prange->start, last, &tail);
1136 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1139 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1140 if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1141 prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1142 pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1143 prange, prange->start, prange->last,
1144 SVM_OP_ADD_RANGE_AND_MAP);
1150 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
1153 struct amdgpu_device *bo_adev;
1154 uint32_t flags = prange->flags;
1155 uint32_t mapping_flags = 0;
1157 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1158 bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1160 if (domain == SVM_RANGE_VRAM_DOMAIN)
1161 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1163 switch (KFD_GC_VERSION(adev->kfd.dev)) {
1164 case IP_VERSION(9, 4, 1):
1165 if (domain == SVM_RANGE_VRAM_DOMAIN) {
1166 if (bo_adev == adev) {
1167 mapping_flags |= coherent ?
1168 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1170 mapping_flags |= coherent ?
1171 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1172 if (amdgpu_xgmi_same_hive(adev, bo_adev))
1176 mapping_flags |= coherent ?
1177 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1180 case IP_VERSION(9, 4, 2):
1181 if (domain == SVM_RANGE_VRAM_DOMAIN) {
1182 if (bo_adev == adev) {
1183 mapping_flags |= coherent ?
1184 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1185 if (adev->gmc.xgmi.connected_to_cpu)
1188 mapping_flags |= coherent ?
1189 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1190 if (amdgpu_xgmi_same_hive(adev, bo_adev))
1194 mapping_flags |= coherent ?
1195 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1199 mapping_flags |= coherent ?
1200 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1203 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1205 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1206 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1207 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1208 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1210 pte_flags = AMDGPU_PTE_VALID;
1211 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1212 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1214 pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
1219 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1220 uint64_t start, uint64_t last,
1221 struct dma_fence **fence)
1223 uint64_t init_pte_value = 0;
1225 pr_debug("[0x%llx 0x%llx]\n", start, last);
1227 return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1228 last, init_pte_value, 0, 0, NULL, NULL,
1233 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1234 unsigned long last, uint32_t trigger)
1236 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1237 struct kfd_process_device *pdd;
1238 struct dma_fence *fence = NULL;
1239 struct kfd_process *p;
1243 if (!prange->mapped_to_gpu) {
1244 pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1245 prange, prange->start, prange->last);
1249 if (prange->start == start && prange->last == last) {
1250 pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1251 prange->mapped_to_gpu = false;
1254 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1256 p = container_of(prange->svms, struct kfd_process, svms);
1258 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1259 pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1260 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1262 pr_debug("failed to find device idx %d\n", gpuidx);
1266 kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1267 start, last, trigger);
1269 r = svm_range_unmap_from_gpu(pdd->dev->adev,
1270 drm_priv_to_vm(pdd->drm_priv),
1271 start, last, &fence);
1276 r = dma_fence_wait(fence, false);
1277 dma_fence_put(fence);
1282 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1289 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1290 unsigned long offset, unsigned long npages, bool readonly,
1291 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1292 struct dma_fence **fence, bool flush_tlb)
1294 struct amdgpu_device *adev = pdd->dev->adev;
1295 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1297 unsigned long last_start;
1302 last_start = prange->start + offset;
1304 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1305 last_start, last_start + npages - 1, readonly);
1307 for (i = offset; i < offset + npages; i++) {
1308 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1309 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1311 /* Collect all pages in the same address range and memory domain
1312 * that can be mapped with a single call to update mapping.
1314 if (i < offset + npages - 1 &&
1315 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1318 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1319 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1321 pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
1323 pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1325 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1326 prange->svms, last_start, prange->start + i,
1327 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1330 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL,
1331 last_start, prange->start + i,
1333 (last_start - prange->start) << PAGE_SHIFT,
1334 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1335 NULL, dma_addr, &vm->last_update);
1337 for (j = last_start - prange->start; j <= i; j++)
1338 dma_addr[j] |= last_domain;
1341 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1344 last_start = prange->start + i + 1;
1347 r = amdgpu_vm_update_pdes(adev, vm, false);
1349 pr_debug("failed %d to update directories 0x%lx\n", r,
1355 *fence = dma_fence_get(vm->last_update);
1362 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1363 unsigned long npages, bool readonly,
1364 unsigned long *bitmap, bool wait, bool flush_tlb)
1366 struct kfd_process_device *pdd;
1367 struct amdgpu_device *bo_adev;
1368 struct kfd_process *p;
1369 struct dma_fence *fence = NULL;
1373 if (prange->svm_bo && prange->ttm_res)
1374 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
1378 p = container_of(prange->svms, struct kfd_process, svms);
1379 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1380 pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1381 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1383 pr_debug("failed to find device idx %d\n", gpuidx);
1387 pdd = kfd_bind_process_to_device(pdd->dev, p);
1391 if (bo_adev && pdd->dev->adev != bo_adev &&
1392 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1393 pr_debug("cannot map to device idx %d\n", gpuidx);
1397 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1398 prange->dma_addr[gpuidx],
1399 bo_adev, wait ? &fence : NULL,
1405 r = dma_fence_wait(fence, false);
1406 dma_fence_put(fence);
1409 pr_debug("failed %d to dma fence wait\n", r);
1414 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1420 struct svm_validate_context {
1421 struct kfd_process *process;
1422 struct svm_range *prange;
1424 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1425 struct ttm_validate_buffer tv[MAX_GPU_INSTANCE];
1426 struct list_head validate_list;
1427 struct ww_acquire_ctx ticket;
1430 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1432 struct kfd_process_device *pdd;
1433 struct amdgpu_vm *vm;
1437 INIT_LIST_HEAD(&ctx->validate_list);
1438 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1439 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1441 pr_debug("failed to find device idx %d\n", gpuidx);
1444 vm = drm_priv_to_vm(pdd->drm_priv);
1446 ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1447 ctx->tv[gpuidx].num_shared = 4;
1448 list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1451 r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1454 pr_debug("failed %d to reserve bo\n", r);
1458 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1459 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1461 pr_debug("failed to find device idx %d\n", gpuidx);
1466 r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1467 drm_priv_to_vm(pdd->drm_priv),
1468 svm_range_bo_validate, NULL);
1470 pr_debug("failed %d validate pt bos\n", r);
1478 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1482 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1484 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1487 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1489 struct kfd_process_device *pdd;
1491 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1493 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1497 * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1499 * To prevent concurrent destruction or change of range attributes, the
1500 * svm_read_lock must be held. The caller must not hold the svm_write_lock
1501 * because that would block concurrent evictions and lead to deadlocks. To
1502 * serialize concurrent migrations or validations of the same range, the
1503 * prange->migrate_mutex must be held.
1505 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1508 * The following sequence ensures race-free validation and GPU mapping:
1510 * 1. Reserve page table (and SVM BO if range is in VRAM)
1511 * 2. hmm_range_fault to get page addresses (if system memory)
1512 * 3. DMA-map pages (if system memory)
1513 * 4-a. Take notifier lock
1514 * 4-b. Check that pages still valid (mmu_interval_read_retry)
1515 * 4-c. Check that the range was not split or otherwise invalidated
1516 * 4-d. Update GPU page table
1517 * 4.e. Release notifier lock
1518 * 5. Release page table (and SVM BO) reservation
1520 static int svm_range_validate_and_map(struct mm_struct *mm,
1521 struct svm_range *prange, int32_t gpuidx,
1522 bool intr, bool wait, bool flush_tlb)
1524 struct svm_validate_context ctx;
1525 unsigned long start, end, addr;
1526 struct kfd_process *p;
1531 ctx.process = container_of(prange->svms, struct kfd_process, svms);
1532 ctx.prange = prange;
1535 if (gpuidx < MAX_GPU_INSTANCE) {
1536 bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE);
1537 bitmap_set(ctx.bitmap, gpuidx, 1);
1538 } else if (ctx.process->xnack_enabled) {
1539 bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1541 /* If prefetch range to GPU, or GPU retry fault migrate range to
1542 * GPU, which has ACCESS attribute to the range, create mapping
1545 if (prange->actual_loc) {
1546 gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process,
1547 prange->actual_loc);
1549 WARN_ONCE(1, "failed get device by id 0x%x\n",
1550 prange->actual_loc);
1553 if (test_bit(gpuidx, prange->bitmap_access))
1554 bitmap_set(ctx.bitmap, gpuidx, 1);
1557 bitmap_or(ctx.bitmap, prange->bitmap_access,
1558 prange->bitmap_aip, MAX_GPU_INSTANCE);
1561 if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) {
1562 if (!prange->mapped_to_gpu)
1565 bitmap_copy(ctx.bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1568 if (prange->actual_loc && !prange->ttm_res) {
1569 /* This should never happen. actual_loc gets set by
1570 * svm_migrate_ram_to_vram after allocating a BO.
1572 WARN_ONCE(1, "VRAM BO missing during validation\n");
1576 svm_range_reserve_bos(&ctx);
1578 p = container_of(prange->svms, struct kfd_process, svms);
1579 owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap,
1581 for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) {
1582 if (kfd_svm_page_owner(p, idx) != owner) {
1588 start = prange->start << PAGE_SHIFT;
1589 end = (prange->last + 1) << PAGE_SHIFT;
1590 for (addr = start; addr < end && !r; ) {
1591 struct hmm_range *hmm_range;
1592 struct vm_area_struct *vma;
1594 unsigned long offset;
1595 unsigned long npages;
1598 vma = find_vma(mm, addr);
1599 if (!vma || addr < vma->vm_start) {
1603 readonly = !(vma->vm_flags & VM_WRITE);
1605 next = min(vma->vm_end, end);
1606 npages = (next - addr) >> PAGE_SHIFT;
1607 WRITE_ONCE(p->svms.faulting_task, current);
1608 r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
1609 addr, npages, &hmm_range,
1610 readonly, true, owner);
1611 WRITE_ONCE(p->svms.faulting_task, NULL);
1613 pr_debug("failed %d to get svm range pages\n", r);
1617 offset = (addr - start) >> PAGE_SHIFT;
1618 r = svm_range_dma_map(prange, ctx.bitmap, offset, npages,
1619 hmm_range->hmm_pfns);
1621 pr_debug("failed %d to dma map range\n", r);
1625 svm_range_lock(prange);
1626 if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1627 pr_debug("hmm update the range, need validate again\n");
1631 if (!list_empty(&prange->child_list)) {
1632 pr_debug("range split by unmap in parallel, validate again\n");
1637 r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1638 ctx.bitmap, wait, flush_tlb);
1641 svm_range_unlock(prange);
1647 prange->validated_once = true;
1648 prange->mapped_to_gpu = true;
1652 svm_range_unreserve_bos(&ctx);
1655 prange->validate_timestamp = ktime_get_boottime();
1661 * svm_range_list_lock_and_flush_work - flush pending deferred work
1663 * @svms: the svm range list
1664 * @mm: the mm structure
1666 * Context: Returns with mmap write lock held, pending deferred work flushed
1670 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1671 struct mm_struct *mm)
1674 flush_work(&svms->deferred_list_work);
1675 mmap_write_lock(mm);
1677 if (list_empty(&svms->deferred_range_list))
1679 mmap_write_unlock(mm);
1680 pr_debug("retry flush\n");
1681 goto retry_flush_work;
1684 static void svm_range_restore_work(struct work_struct *work)
1686 struct delayed_work *dwork = to_delayed_work(work);
1687 struct amdkfd_process_info *process_info;
1688 struct svm_range_list *svms;
1689 struct svm_range *prange;
1690 struct kfd_process *p;
1691 struct mm_struct *mm;
1696 svms = container_of(dwork, struct svm_range_list, restore_work);
1697 evicted_ranges = atomic_read(&svms->evicted_ranges);
1698 if (!evicted_ranges)
1701 pr_debug("restore svm ranges\n");
1703 p = container_of(svms, struct kfd_process, svms);
1704 process_info = p->kgd_process_info;
1706 /* Keep mm reference when svm_range_validate_and_map ranges */
1707 mm = get_task_mm(p->lead_thread);
1709 pr_debug("svms 0x%p process mm gone\n", svms);
1713 mutex_lock(&process_info->lock);
1714 svm_range_list_lock_and_flush_work(svms, mm);
1715 mutex_lock(&svms->lock);
1717 evicted_ranges = atomic_read(&svms->evicted_ranges);
1719 list_for_each_entry(prange, &svms->list, list) {
1720 invalid = atomic_read(&prange->invalid);
1724 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1725 prange->svms, prange, prange->start, prange->last,
1729 * If range is migrating, wait for migration is done.
1731 mutex_lock(&prange->migrate_mutex);
1733 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1734 false, true, false);
1736 pr_debug("failed %d to map 0x%lx to gpus\n", r,
1739 mutex_unlock(&prange->migrate_mutex);
1741 goto out_reschedule;
1743 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1744 goto out_reschedule;
1747 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1749 goto out_reschedule;
1753 r = kgd2kfd_resume_mm(mm);
1755 /* No recovery from this failure. Probably the CP is
1756 * hanging. No point trying again.
1758 pr_debug("failed %d to resume KFD\n", r);
1761 pr_debug("restore svm ranges successfully\n");
1764 mutex_unlock(&svms->lock);
1765 mmap_write_unlock(mm);
1766 mutex_unlock(&process_info->lock);
1768 /* If validation failed, reschedule another attempt */
1769 if (evicted_ranges) {
1770 pr_debug("reschedule to restore svm range\n");
1771 schedule_delayed_work(&svms->restore_work,
1772 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1774 kfd_smi_event_queue_restore_rescheduled(mm);
1780 * svm_range_evict - evict svm range
1781 * @prange: svm range structure
1782 * @mm: current process mm_struct
1783 * @start: starting process queue number
1784 * @last: last process queue number
1786 * Stop all queues of the process to ensure GPU doesn't access the memory, then
1787 * return to let CPU evict the buffer and proceed CPU pagetable update.
1789 * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1790 * If invalidation happens while restore work is running, restore work will
1791 * restart to ensure to get the latest CPU pages mapping to GPU, then start
1795 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1796 unsigned long start, unsigned long last,
1797 enum mmu_notifier_event event)
1799 struct svm_range_list *svms = prange->svms;
1800 struct svm_range *pchild;
1801 struct kfd_process *p;
1804 p = container_of(svms, struct kfd_process, svms);
1806 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1807 svms, prange->start, prange->last, start, last);
1809 if (!p->xnack_enabled ||
1810 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1812 bool mapped = prange->mapped_to_gpu;
1814 list_for_each_entry(pchild, &prange->child_list, child_list) {
1815 if (!pchild->mapped_to_gpu)
1818 mutex_lock_nested(&pchild->lock, 1);
1819 if (pchild->start <= last && pchild->last >= start) {
1820 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1821 pchild->start, pchild->last);
1822 atomic_inc(&pchild->invalid);
1824 mutex_unlock(&pchild->lock);
1830 if (prange->start <= last && prange->last >= start)
1831 atomic_inc(&prange->invalid);
1833 evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1834 if (evicted_ranges != 1)
1837 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1838 prange->svms, prange->start, prange->last);
1840 /* First eviction, stop the queues */
1841 r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1843 pr_debug("failed to quiesce KFD\n");
1845 pr_debug("schedule to restore svm %p ranges\n", svms);
1846 schedule_delayed_work(&svms->restore_work,
1847 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1852 if (event == MMU_NOTIFY_MIGRATE)
1853 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1855 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1857 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1858 prange->svms, start, last);
1859 list_for_each_entry(pchild, &prange->child_list, child_list) {
1860 mutex_lock_nested(&pchild->lock, 1);
1861 s = max(start, pchild->start);
1862 l = min(last, pchild->last);
1864 svm_range_unmap_from_gpus(pchild, s, l, trigger);
1865 mutex_unlock(&pchild->lock);
1867 s = max(start, prange->start);
1868 l = min(last, prange->last);
1870 svm_range_unmap_from_gpus(prange, s, l, trigger);
1876 static struct svm_range *svm_range_clone(struct svm_range *old)
1878 struct svm_range *new;
1880 new = svm_range_new(old->svms, old->start, old->last, false);
1885 new->ttm_res = old->ttm_res;
1886 new->offset = old->offset;
1887 new->svm_bo = svm_range_bo_ref(old->svm_bo);
1888 spin_lock(&new->svm_bo->list_lock);
1889 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1890 spin_unlock(&new->svm_bo->list_lock);
1892 new->flags = old->flags;
1893 new->preferred_loc = old->preferred_loc;
1894 new->prefetch_loc = old->prefetch_loc;
1895 new->actual_loc = old->actual_loc;
1896 new->granularity = old->granularity;
1897 new->mapped_to_gpu = old->mapped_to_gpu;
1898 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1899 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1904 void svm_range_set_max_pages(struct amdgpu_device *adev)
1907 uint64_t pages, _pages;
1909 /* 1/32 VRAM size in pages */
1910 pages = adev->gmc.real_vram_size >> 17;
1911 pages = clamp(pages, 1ULL << 9, 1ULL << 18);
1912 pages = rounddown_pow_of_two(pages);
1914 max_pages = READ_ONCE(max_svm_range_pages);
1915 _pages = min_not_zero(max_pages, pages);
1916 } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
1920 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
1921 uint64_t max_pages, struct list_head *insert_list,
1922 struct list_head *update_list)
1924 struct svm_range *prange;
1927 pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
1928 max_pages, start, last);
1930 while (last >= start) {
1931 l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
1933 prange = svm_range_new(svms, start, l, true);
1936 list_add(&prange->list, insert_list);
1937 list_add(&prange->update_list, update_list);
1945 * svm_range_add - add svm range and handle overlap
1946 * @p: the range add to this process svms
1947 * @start: page size aligned
1948 * @size: page size aligned
1949 * @nattr: number of attributes
1950 * @attrs: array of attributes
1951 * @update_list: output, the ranges need validate and update GPU mapping
1952 * @insert_list: output, the ranges need insert to svms
1953 * @remove_list: output, the ranges are replaced and need remove from svms
1955 * Check if the virtual address range has overlap with any existing ranges,
1956 * split partly overlapping ranges and add new ranges in the gaps. All changes
1957 * should be applied to the range_list and interval tree transactionally. If
1958 * any range split or allocation fails, the entire update fails. Therefore any
1959 * existing overlapping svm_ranges are cloned and the original svm_ranges left
1962 * If the transaction succeeds, the caller can update and insert clones and
1963 * new ranges, then free the originals.
1965 * Otherwise the caller can free the clones and new ranges, while the old
1966 * svm_ranges remain unchanged.
1968 * Context: Process context, caller must hold svms->lock
1971 * 0 - OK, otherwise error code
1974 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
1975 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
1976 struct list_head *update_list, struct list_head *insert_list,
1977 struct list_head *remove_list)
1979 unsigned long last = start + size - 1UL;
1980 struct svm_range_list *svms = &p->svms;
1981 struct interval_tree_node *node;
1982 struct svm_range *prange;
1983 struct svm_range *tmp;
1984 struct list_head new_list;
1987 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
1989 INIT_LIST_HEAD(update_list);
1990 INIT_LIST_HEAD(insert_list);
1991 INIT_LIST_HEAD(remove_list);
1992 INIT_LIST_HEAD(&new_list);
1994 node = interval_tree_iter_first(&svms->objects, start, last);
1996 struct interval_tree_node *next;
1997 unsigned long next_start;
1999 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2002 prange = container_of(node, struct svm_range, it_node);
2003 next = interval_tree_iter_next(node, start, last);
2004 next_start = min(node->last, last) + 1;
2006 if (svm_range_is_same_attrs(p, prange, nattr, attrs)) {
2008 } else if (node->start < start || node->last > last) {
2009 /* node intersects the update range and its attributes
2010 * will change. Clone and split it, apply updates only
2011 * to the overlapping part
2013 struct svm_range *old = prange;
2015 prange = svm_range_clone(old);
2021 list_add(&old->update_list, remove_list);
2022 list_add(&prange->list, insert_list);
2023 list_add(&prange->update_list, update_list);
2025 if (node->start < start) {
2026 pr_debug("change old range start\n");
2027 r = svm_range_split_head(prange, start,
2032 if (node->last > last) {
2033 pr_debug("change old range last\n");
2034 r = svm_range_split_tail(prange, last,
2040 /* The node is contained within start..last,
2043 list_add(&prange->update_list, update_list);
2046 /* insert a new node if needed */
2047 if (node->start > start) {
2048 r = svm_range_split_new(svms, start, node->start - 1,
2049 READ_ONCE(max_svm_range_pages),
2050 &new_list, update_list);
2059 /* add a final range at the end if needed */
2061 r = svm_range_split_new(svms, start, last,
2062 READ_ONCE(max_svm_range_pages),
2063 &new_list, update_list);
2067 list_for_each_entry_safe(prange, tmp, insert_list, list)
2068 svm_range_free(prange, false);
2069 list_for_each_entry_safe(prange, tmp, &new_list, list)
2070 svm_range_free(prange, true);
2072 list_splice(&new_list, insert_list);
2079 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2080 struct svm_range *prange)
2082 unsigned long start;
2085 start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2086 last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2088 if (prange->start == start && prange->last == last)
2091 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2092 prange->svms, prange, start, last, prange->start,
2095 if (start != 0 && last != 0) {
2096 interval_tree_remove(&prange->it_node, &prange->svms->objects);
2097 svm_range_remove_notifier(prange);
2099 prange->it_node.start = prange->start;
2100 prange->it_node.last = prange->last;
2102 interval_tree_insert(&prange->it_node, &prange->svms->objects);
2103 svm_range_add_notifier_locked(mm, prange);
2107 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2108 struct mm_struct *mm)
2110 switch (prange->work_item.op) {
2112 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2113 svms, prange, prange->start, prange->last);
2115 case SVM_OP_UNMAP_RANGE:
2116 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2117 svms, prange, prange->start, prange->last);
2118 svm_range_unlink(prange);
2119 svm_range_remove_notifier(prange);
2120 svm_range_free(prange, true);
2122 case SVM_OP_UPDATE_RANGE_NOTIFIER:
2123 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2124 svms, prange, prange->start, prange->last);
2125 svm_range_update_notifier_and_interval_tree(mm, prange);
2127 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2128 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2129 svms, prange, prange->start, prange->last);
2130 svm_range_update_notifier_and_interval_tree(mm, prange);
2131 /* TODO: implement deferred validation and mapping */
2133 case SVM_OP_ADD_RANGE:
2134 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2135 prange->start, prange->last);
2136 svm_range_add_to_svms(prange);
2137 svm_range_add_notifier_locked(mm, prange);
2139 case SVM_OP_ADD_RANGE_AND_MAP:
2140 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2141 prange, prange->start, prange->last);
2142 svm_range_add_to_svms(prange);
2143 svm_range_add_notifier_locked(mm, prange);
2144 /* TODO: implement deferred validation and mapping */
2147 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2148 prange->work_item.op);
2152 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2154 struct kfd_process_device *pdd;
2155 struct kfd_process *p;
2159 p = container_of(svms, struct kfd_process, svms);
2162 drain = atomic_read(&svms->drain_pagefaults);
2166 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2171 pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2173 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2174 &pdd->dev->adev->irq.ih1);
2175 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2177 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2181 static void svm_range_deferred_list_work(struct work_struct *work)
2183 struct svm_range_list *svms;
2184 struct svm_range *prange;
2185 struct mm_struct *mm;
2187 svms = container_of(work, struct svm_range_list, deferred_list_work);
2188 pr_debug("enter svms 0x%p\n", svms);
2190 spin_lock(&svms->deferred_list_lock);
2191 while (!list_empty(&svms->deferred_range_list)) {
2192 prange = list_first_entry(&svms->deferred_range_list,
2193 struct svm_range, deferred_list);
2194 spin_unlock(&svms->deferred_list_lock);
2196 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2197 prange->start, prange->last, prange->work_item.op);
2199 mm = prange->work_item.mm;
2201 mmap_write_lock(mm);
2203 /* Checking for the need to drain retry faults must be inside
2204 * mmap write lock to serialize with munmap notifiers.
2206 if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2207 mmap_write_unlock(mm);
2208 svm_range_drain_retry_fault(svms);
2212 /* Remove from deferred_list must be inside mmap write lock, for
2214 * 1. unmap_from_cpu may change work_item.op and add the range
2215 * to deferred_list again, cause use after free bug.
2216 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2217 * lock and continue because deferred_list is empty, but
2218 * deferred_list work is actually waiting for mmap lock.
2220 spin_lock(&svms->deferred_list_lock);
2221 list_del_init(&prange->deferred_list);
2222 spin_unlock(&svms->deferred_list_lock);
2224 mutex_lock(&svms->lock);
2225 mutex_lock(&prange->migrate_mutex);
2226 while (!list_empty(&prange->child_list)) {
2227 struct svm_range *pchild;
2229 pchild = list_first_entry(&prange->child_list,
2230 struct svm_range, child_list);
2231 pr_debug("child prange 0x%p op %d\n", pchild,
2232 pchild->work_item.op);
2233 list_del_init(&pchild->child_list);
2234 svm_range_handle_list_op(svms, pchild, mm);
2236 mutex_unlock(&prange->migrate_mutex);
2238 svm_range_handle_list_op(svms, prange, mm);
2239 mutex_unlock(&svms->lock);
2240 mmap_write_unlock(mm);
2242 /* Pairs with mmget in svm_range_add_list_work */
2245 spin_lock(&svms->deferred_list_lock);
2247 spin_unlock(&svms->deferred_list_lock);
2248 pr_debug("exit svms 0x%p\n", svms);
2252 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2253 struct mm_struct *mm, enum svm_work_list_ops op)
2255 spin_lock(&svms->deferred_list_lock);
2256 /* if prange is on the deferred list */
2257 if (!list_empty(&prange->deferred_list)) {
2258 pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2259 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2260 if (op != SVM_OP_NULL &&
2261 prange->work_item.op != SVM_OP_UNMAP_RANGE)
2262 prange->work_item.op = op;
2264 prange->work_item.op = op;
2266 /* Pairs with mmput in deferred_list_work */
2268 prange->work_item.mm = mm;
2269 list_add_tail(&prange->deferred_list,
2270 &prange->svms->deferred_range_list);
2271 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2272 prange, prange->start, prange->last, op);
2274 spin_unlock(&svms->deferred_list_lock);
2277 void schedule_deferred_list_work(struct svm_range_list *svms)
2279 spin_lock(&svms->deferred_list_lock);
2280 if (!list_empty(&svms->deferred_range_list))
2281 schedule_work(&svms->deferred_list_work);
2282 spin_unlock(&svms->deferred_list_lock);
2286 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2287 struct svm_range *prange, unsigned long start,
2290 struct svm_range *head;
2291 struct svm_range *tail;
2293 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2294 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2295 prange->start, prange->last);
2298 if (start > prange->last || last < prange->start)
2301 head = tail = prange;
2302 if (start > prange->start)
2303 svm_range_split(prange, prange->start, start - 1, &tail);
2304 if (last < tail->last)
2305 svm_range_split(tail, last + 1, tail->last, &head);
2307 if (head != prange && tail != prange) {
2308 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2309 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2310 } else if (tail != prange) {
2311 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2312 } else if (head != prange) {
2313 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2314 } else if (parent != prange) {
2315 prange->work_item.op = SVM_OP_UNMAP_RANGE;
2320 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2321 unsigned long start, unsigned long last)
2323 uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU;
2324 struct svm_range_list *svms;
2325 struct svm_range *pchild;
2326 struct kfd_process *p;
2330 p = kfd_lookup_process_by_mm(mm);
2335 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2336 prange, prange->start, prange->last, start, last);
2338 /* Make sure pending page faults are drained in the deferred worker
2339 * before the range is freed to avoid straggler interrupts on
2340 * unmapped memory causing "phantom faults".
2342 atomic_inc(&svms->drain_pagefaults);
2344 unmap_parent = start <= prange->start && last >= prange->last;
2346 list_for_each_entry(pchild, &prange->child_list, child_list) {
2347 mutex_lock_nested(&pchild->lock, 1);
2348 s = max(start, pchild->start);
2349 l = min(last, pchild->last);
2351 svm_range_unmap_from_gpus(pchild, s, l, trigger);
2352 svm_range_unmap_split(mm, prange, pchild, start, last);
2353 mutex_unlock(&pchild->lock);
2355 s = max(start, prange->start);
2356 l = min(last, prange->last);
2358 svm_range_unmap_from_gpus(prange, s, l, trigger);
2359 svm_range_unmap_split(mm, prange, prange, start, last);
2362 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2364 svm_range_add_list_work(svms, prange, mm,
2365 SVM_OP_UPDATE_RANGE_NOTIFIER);
2366 schedule_deferred_list_work(svms);
2368 kfd_unref_process(p);
2372 * svm_range_cpu_invalidate_pagetables - interval notifier callback
2373 * @mni: mmu_interval_notifier struct
2374 * @range: mmu_notifier_range struct
2375 * @cur_seq: value to pass to mmu_interval_set_seq()
2377 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2378 * is from migration, or CPU page invalidation callback.
2380 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2381 * work thread, and split prange if only part of prange is unmapped.
2383 * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2384 * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2385 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2386 * update GPU mapping to recover.
2388 * Context: mmap lock, notifier_invalidate_start lock are held
2389 * for invalidate event, prange lock is held if this is from migration
2392 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2393 const struct mmu_notifier_range *range,
2394 unsigned long cur_seq)
2396 struct svm_range *prange;
2397 unsigned long start;
2400 if (range->event == MMU_NOTIFY_RELEASE)
2402 if (!mmget_not_zero(mni->mm))
2405 start = mni->interval_tree.start;
2406 last = mni->interval_tree.last;
2407 start = max(start, range->start) >> PAGE_SHIFT;
2408 last = min(last, range->end - 1) >> PAGE_SHIFT;
2409 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2410 start, last, range->start >> PAGE_SHIFT,
2411 (range->end - 1) >> PAGE_SHIFT,
2412 mni->interval_tree.start >> PAGE_SHIFT,
2413 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2415 prange = container_of(mni, struct svm_range, notifier);
2417 svm_range_lock(prange);
2418 mmu_interval_set_seq(mni, cur_seq);
2420 switch (range->event) {
2421 case MMU_NOTIFY_UNMAP:
2422 svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2425 svm_range_evict(prange, mni->mm, start, last, range->event);
2429 svm_range_unlock(prange);
2436 * svm_range_from_addr - find svm range from fault address
2437 * @svms: svm range list header
2438 * @addr: address to search range interval tree, in pages
2439 * @parent: parent range if range is on child list
2441 * Context: The caller must hold svms->lock
2443 * Return: the svm_range found or NULL
2446 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2447 struct svm_range **parent)
2449 struct interval_tree_node *node;
2450 struct svm_range *prange;
2451 struct svm_range *pchild;
2453 node = interval_tree_iter_first(&svms->objects, addr, addr);
2457 prange = container_of(node, struct svm_range, it_node);
2458 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2459 addr, prange->start, prange->last, node->start, node->last);
2461 if (addr >= prange->start && addr <= prange->last) {
2466 list_for_each_entry(pchild, &prange->child_list, child_list)
2467 if (addr >= pchild->start && addr <= pchild->last) {
2468 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2469 addr, pchild->start, pchild->last);
2478 /* svm_range_best_restore_location - decide the best fault restore location
2479 * @prange: svm range structure
2480 * @adev: the GPU on which vm fault happened
2482 * This is only called when xnack is on, to decide the best location to restore
2483 * the range mapping after GPU vm fault. Caller uses the best location to do
2484 * migration if actual loc is not best location, then update GPU page table
2485 * mapping to the best location.
2487 * If the preferred loc is accessible by faulting GPU, use preferred loc.
2488 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2489 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2490 * if range actual loc is cpu, best_loc is cpu
2491 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2493 * Otherwise, GPU no access, best_loc is -1.
2496 * -1 means vm fault GPU no access
2497 * 0 for CPU or GPU id
2500 svm_range_best_restore_location(struct svm_range *prange,
2501 struct amdgpu_device *adev,
2504 struct amdgpu_device *bo_adev, *preferred_adev;
2505 struct kfd_process *p;
2509 p = container_of(prange->svms, struct kfd_process, svms);
2511 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx);
2513 pr_debug("failed to get gpuid from kgd\n");
2517 if (prange->preferred_loc == gpuid ||
2518 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2519 return prange->preferred_loc;
2520 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2521 preferred_adev = svm_range_get_adev_by_id(prange,
2522 prange->preferred_loc);
2523 if (amdgpu_xgmi_same_hive(adev, preferred_adev))
2524 return prange->preferred_loc;
2528 if (test_bit(*gpuidx, prange->bitmap_access))
2531 if (test_bit(*gpuidx, prange->bitmap_aip)) {
2532 if (!prange->actual_loc)
2535 bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
2536 if (amdgpu_xgmi_same_hive(adev, bo_adev))
2537 return prange->actual_loc;
2546 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2547 unsigned long *start, unsigned long *last,
2548 bool *is_heap_stack)
2550 struct vm_area_struct *vma;
2551 struct interval_tree_node *node;
2552 unsigned long start_limit, end_limit;
2554 vma = find_vma(p->mm, addr << PAGE_SHIFT);
2555 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2556 pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2560 *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
2561 vma->vm_end >= vma->vm_mm->start_brk) ||
2562 (vma->vm_start <= vma->vm_mm->start_stack &&
2563 vma->vm_end >= vma->vm_mm->start_stack);
2565 start_limit = max(vma->vm_start >> PAGE_SHIFT,
2566 (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2567 end_limit = min(vma->vm_end >> PAGE_SHIFT,
2568 (unsigned long)ALIGN(addr + 1, 2UL << 8));
2569 /* First range that starts after the fault address */
2570 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2572 end_limit = min(end_limit, node->start);
2573 /* Last range that ends before the fault address */
2574 node = container_of(rb_prev(&node->rb),
2575 struct interval_tree_node, rb);
2577 /* Last range must end before addr because
2578 * there was no range after addr
2580 node = container_of(rb_last(&p->svms.objects.rb_root),
2581 struct interval_tree_node, rb);
2584 if (node->last >= addr) {
2585 WARN(1, "Overlap with prev node and page fault addr\n");
2588 start_limit = max(start_limit, node->last + 1);
2591 *start = start_limit;
2592 *last = end_limit - 1;
2594 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2595 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2596 *start, *last, *is_heap_stack);
2602 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2603 uint64_t *bo_s, uint64_t *bo_l)
2605 struct amdgpu_bo_va_mapping *mapping;
2606 struct interval_tree_node *node;
2607 struct amdgpu_bo *bo = NULL;
2608 unsigned long userptr;
2612 for (i = 0; i < p->n_pdds; i++) {
2613 struct amdgpu_vm *vm;
2615 if (!p->pdds[i]->drm_priv)
2618 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2619 r = amdgpu_bo_reserve(vm->root.bo, false);
2623 /* Check userptr by searching entire vm->va interval tree */
2624 node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2626 mapping = container_of((struct rb_node *)node,
2627 struct amdgpu_bo_va_mapping, rb);
2628 bo = mapping->bo_va->base.bo;
2630 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2631 start << PAGE_SHIFT,
2634 node = interval_tree_iter_next(node, 0, ~0ULL);
2638 pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2641 *bo_s = userptr >> PAGE_SHIFT;
2642 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2644 amdgpu_bo_unreserve(vm->root.bo);
2647 amdgpu_bo_unreserve(vm->root.bo);
2653 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
2654 struct kfd_process *p,
2655 struct mm_struct *mm,
2658 struct svm_range *prange = NULL;
2659 unsigned long start, last;
2660 uint32_t gpuid, gpuidx;
2666 if (svm_range_get_range_boundaries(p, addr, &start, &last,
2670 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2671 if (r != -EADDRINUSE)
2672 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2674 if (r == -EADDRINUSE) {
2675 if (addr >= bo_s && addr <= bo_l)
2678 /* Create one page svm range if 2MB range overlapping */
2683 prange = svm_range_new(&p->svms, start, last, true);
2685 pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2688 if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) {
2689 pr_debug("failed to get gpuid from kgd\n");
2690 svm_range_free(prange, true);
2695 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2697 svm_range_add_to_svms(prange);
2698 svm_range_add_notifier_locked(mm, prange);
2703 /* svm_range_skip_recover - decide if prange can be recovered
2704 * @prange: svm range structure
2706 * GPU vm retry fault handle skip recover the range for cases:
2707 * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2708 * deferred list work will drain the stale fault before free the prange.
2709 * 2. prange is on deferred list to add interval notifier after split, or
2710 * 3. prange is child range, it is split from parent prange, recover later
2711 * after interval notifier is added.
2713 * Return: true to skip recover, false to recover
2715 static bool svm_range_skip_recover(struct svm_range *prange)
2717 struct svm_range_list *svms = prange->svms;
2719 spin_lock(&svms->deferred_list_lock);
2720 if (list_empty(&prange->deferred_list) &&
2721 list_empty(&prange->child_list)) {
2722 spin_unlock(&svms->deferred_list_lock);
2725 spin_unlock(&svms->deferred_list_lock);
2727 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2728 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2729 svms, prange, prange->start, prange->last);
2732 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2733 prange->work_item.op == SVM_OP_ADD_RANGE) {
2734 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2735 svms, prange, prange->start, prange->last);
2742 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
2745 struct kfd_process_device *pdd;
2747 /* fault is on different page of same range
2748 * or fault is skipped to recover later
2749 * or fault is on invalid virtual address
2751 if (gpuidx == MAX_GPU_INSTANCE) {
2755 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx);
2760 /* fault is recovered
2761 * or fault cannot recover because GPU no access on the range
2763 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2765 WRITE_ONCE(pdd->faults, pdd->faults + 1);
2769 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2771 unsigned long requested = VM_READ;
2774 requested |= VM_WRITE;
2776 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2778 return (vma->vm_flags & requested) == requested;
2782 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2783 uint64_t addr, bool write_fault)
2785 struct mm_struct *mm = NULL;
2786 struct svm_range_list *svms;
2787 struct svm_range *prange;
2788 struct kfd_process *p;
2789 ktime_t timestamp = ktime_get_boottime();
2791 int32_t gpuidx = MAX_GPU_INSTANCE;
2792 bool write_locked = false;
2793 struct vm_area_struct *vma;
2794 bool migration = false;
2797 if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
2798 pr_debug("device does not support SVM\n");
2802 p = kfd_lookup_process_by_pasid(pasid);
2804 pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2809 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2811 if (atomic_read(&svms->drain_pagefaults)) {
2812 pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2817 if (!p->xnack_enabled) {
2818 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2823 /* p->lead_thread is available as kfd_process_wq_release flush the work
2824 * before releasing task ref.
2826 mm = get_task_mm(p->lead_thread);
2828 pr_debug("svms 0x%p failed to get mm\n", svms);
2835 mutex_lock(&svms->lock);
2836 prange = svm_range_from_addr(svms, addr, NULL);
2838 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2840 if (!write_locked) {
2841 /* Need the write lock to create new range with MMU notifier.
2842 * Also flush pending deferred work to make sure the interval
2843 * tree is up to date before we add a new range
2845 mutex_unlock(&svms->lock);
2846 mmap_read_unlock(mm);
2847 mmap_write_lock(mm);
2848 write_locked = true;
2849 goto retry_write_locked;
2851 prange = svm_range_create_unregistered_range(adev, p, mm, addr);
2853 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2855 mmap_write_downgrade(mm);
2857 goto out_unlock_svms;
2861 mmap_write_downgrade(mm);
2863 mutex_lock(&prange->migrate_mutex);
2865 if (svm_range_skip_recover(prange)) {
2866 amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2868 goto out_unlock_range;
2871 /* skip duplicate vm fault on different pages of same range */
2872 if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
2873 AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
2874 pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2875 svms, prange->start, prange->last);
2877 goto out_unlock_range;
2880 /* __do_munmap removed VMA, return success as we are handling stale
2883 vma = find_vma(mm, addr << PAGE_SHIFT);
2884 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
2885 pr_debug("address 0x%llx VMA is removed\n", addr);
2887 goto out_unlock_range;
2890 if (!svm_fault_allowed(vma, write_fault)) {
2891 pr_debug("fault addr 0x%llx no %s permission\n", addr,
2892 write_fault ? "write" : "read");
2894 goto out_unlock_range;
2897 best_loc = svm_range_best_restore_location(prange, adev, &gpuidx);
2898 if (best_loc == -1) {
2899 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2900 svms, prange->start, prange->last);
2902 goto out_unlock_range;
2905 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2906 svms, prange->start, prange->last, best_loc,
2907 prange->actual_loc);
2909 kfd_smi_event_page_fault_start(adev->kfd.dev, p->lead_thread->pid, addr,
2910 write_fault, timestamp);
2912 if (prange->actual_loc != best_loc) {
2915 r = svm_migrate_to_vram(prange, best_loc, mm,
2916 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
2918 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2920 /* Fallback to system memory if migration to
2923 if (prange->actual_loc)
2924 r = svm_migrate_vram_to_ram(prange, mm,
2925 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
2931 r = svm_migrate_vram_to_ram(prange, mm,
2932 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
2936 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
2937 r, svms, prange->start, prange->last);
2938 goto out_unlock_range;
2942 r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
2944 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
2945 r, svms, prange->start, prange->last);
2947 kfd_smi_event_page_fault_end(adev->kfd.dev, p->lead_thread->pid, addr,
2951 mutex_unlock(&prange->migrate_mutex);
2953 mutex_unlock(&svms->lock);
2954 mmap_read_unlock(mm);
2956 svm_range_count_fault(adev, p, gpuidx);
2960 kfd_unref_process(p);
2963 pr_debug("recover vm fault later\n");
2964 amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
2971 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
2973 struct svm_range *prange, *pchild;
2974 uint64_t reserved_size = 0;
2978 pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
2980 mutex_lock(&p->svms.lock);
2982 list_for_each_entry(prange, &p->svms.list, list) {
2983 svm_range_lock(prange);
2984 list_for_each_entry(pchild, &prange->child_list, child_list) {
2985 size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
2986 if (xnack_enabled) {
2987 amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
2988 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
2990 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
2991 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
2994 reserved_size += size;
2998 size = (prange->last - prange->start + 1) << PAGE_SHIFT;
2999 if (xnack_enabled) {
3000 amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3001 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
3003 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3004 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
3007 reserved_size += size;
3010 svm_range_unlock(prange);
3016 amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3017 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR);
3019 /* Change xnack mode must be inside svms lock, to avoid race with
3020 * svm_range_deferred_list_work unreserve memory in parallel.
3022 p->xnack_enabled = xnack_enabled;
3024 mutex_unlock(&p->svms.lock);
3028 void svm_range_list_fini(struct kfd_process *p)
3030 struct svm_range *prange;
3031 struct svm_range *next;
3033 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3035 cancel_delayed_work_sync(&p->svms.restore_work);
3037 /* Ensure list work is finished before process is destroyed */
3038 flush_work(&p->svms.deferred_list_work);
3041 * Ensure no retry fault comes in afterwards, as page fault handler will
3042 * not find kfd process and take mm lock to recover fault.
3044 atomic_inc(&p->svms.drain_pagefaults);
3045 svm_range_drain_retry_fault(&p->svms);
3047 list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3048 svm_range_unlink(prange);
3049 svm_range_remove_notifier(prange);
3050 svm_range_free(prange, true);
3053 mutex_destroy(&p->svms.lock);
3055 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3058 int svm_range_list_init(struct kfd_process *p)
3060 struct svm_range_list *svms = &p->svms;
3063 svms->objects = RB_ROOT_CACHED;
3064 mutex_init(&svms->lock);
3065 INIT_LIST_HEAD(&svms->list);
3066 atomic_set(&svms->evicted_ranges, 0);
3067 atomic_set(&svms->drain_pagefaults, 0);
3068 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3069 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3070 INIT_LIST_HEAD(&svms->deferred_range_list);
3071 INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3072 spin_lock_init(&svms->deferred_list_lock);
3074 for (i = 0; i < p->n_pdds; i++)
3075 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev))
3076 bitmap_set(svms->bitmap_supported, i, 1);
3082 * svm_range_check_vm - check if virtual address range mapped already
3083 * @p: current kfd_process
3084 * @start: range start address, in pages
3085 * @last: range last address, in pages
3086 * @bo_s: mapping start address in pages if address range already mapped
3087 * @bo_l: mapping last address in pages if address range already mapped
3089 * The purpose is to avoid virtual address ranges already allocated by
3090 * kfd_ioctl_alloc_memory_of_gpu ioctl.
3091 * It looks for each pdd in the kfd_process.
3093 * Context: Process context
3095 * Return 0 - OK, if the range is not mapped.
3096 * Otherwise error code:
3097 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3098 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3099 * a signal. Release all buffer reservations and return to user-space.
3102 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3103 uint64_t *bo_s, uint64_t *bo_l)
3105 struct amdgpu_bo_va_mapping *mapping;
3106 struct interval_tree_node *node;
3110 for (i = 0; i < p->n_pdds; i++) {
3111 struct amdgpu_vm *vm;
3113 if (!p->pdds[i]->drm_priv)
3116 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3117 r = amdgpu_bo_reserve(vm->root.bo, false);
3121 node = interval_tree_iter_first(&vm->va, start, last);
3123 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3125 mapping = container_of((struct rb_node *)node,
3126 struct amdgpu_bo_va_mapping, rb);
3128 *bo_s = mapping->start;
3129 *bo_l = mapping->last;
3131 amdgpu_bo_unreserve(vm->root.bo);
3134 amdgpu_bo_unreserve(vm->root.bo);
3141 * svm_range_is_valid - check if virtual address range is valid
3142 * @p: current kfd_process
3143 * @start: range start address, in pages
3144 * @size: range size, in pages
3146 * Valid virtual address range means it belongs to one or more VMAs
3148 * Context: Process context
3151 * 0 - OK, otherwise error code
3154 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3156 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3157 struct vm_area_struct *vma;
3159 unsigned long start_unchg = start;
3161 start <<= PAGE_SHIFT;
3162 end = start + (size << PAGE_SHIFT);
3164 vma = find_vma(p->mm, start);
3165 if (!vma || start < vma->vm_start ||
3166 (vma->vm_flags & device_vma))
3168 start = min(end, vma->vm_end);
3169 } while (start < end);
3171 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3176 * svm_range_best_prefetch_location - decide the best prefetch location
3177 * @prange: svm range structure
3180 * If range map to single GPU, the best prefetch location is prefetch_loc, which
3181 * can be CPU or GPU.
3183 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3184 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3185 * the best prefetch location is always CPU, because GPU can not have coherent
3186 * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3189 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3190 * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3192 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3193 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3194 * prefetch location is always CPU.
3196 * Context: Process context
3199 * 0 for CPU or GPU id
3202 svm_range_best_prefetch_location(struct svm_range *prange)
3204 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3205 uint32_t best_loc = prange->prefetch_loc;
3206 struct kfd_process_device *pdd;
3207 struct amdgpu_device *bo_adev;
3208 struct kfd_process *p;
3211 p = container_of(prange->svms, struct kfd_process, svms);
3213 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3216 bo_adev = svm_range_get_adev_by_id(prange, best_loc);
3218 WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc);
3223 if (p->xnack_enabled)
3224 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3226 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3229 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3230 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3232 pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3236 if (pdd->dev->adev == bo_adev)
3239 if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
3246 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3247 p->xnack_enabled, &p->svms, prange->start, prange->last,
3253 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3254 * @mm: current process mm_struct
3255 * @prange: svm range structure
3256 * @migrated: output, true if migration is triggered
3258 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3260 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3263 * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3265 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3266 * stops all queues, schedule restore work
3267 * 2. svm_range_restore_work wait for migration is done by
3268 * a. svm_range_validate_vram takes prange->migrate_mutex
3269 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3270 * 3. restore work update mappings of GPU, resume all queues.
3272 * Context: Process context
3275 * 0 - OK, otherwise - error code of migration
3278 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3285 best_loc = svm_range_best_prefetch_location(prange);
3287 if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3288 best_loc == prange->actual_loc)
3292 r = svm_migrate_vram_to_ram(prange, mm,
3293 KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3298 r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3304 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3309 if (dma_fence_is_signaled(&fence->base))
3312 if (fence->svm_bo) {
3313 WRITE_ONCE(fence->svm_bo->evicting, 1);
3314 schedule_work(&fence->svm_bo->eviction_work);
3320 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3322 struct svm_range_bo *svm_bo;
3323 struct mm_struct *mm;
3326 svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3327 if (!svm_bo_ref_unless_zero(svm_bo))
3328 return; /* svm_bo was freed while eviction was pending */
3330 if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3331 mm = svm_bo->eviction_fence->mm;
3333 svm_range_bo_unref(svm_bo);
3338 spin_lock(&svm_bo->list_lock);
3339 while (!list_empty(&svm_bo->range_list) && !r) {
3340 struct svm_range *prange =
3341 list_first_entry(&svm_bo->range_list,
3342 struct svm_range, svm_bo_list);
3345 list_del_init(&prange->svm_bo_list);
3346 spin_unlock(&svm_bo->list_lock);
3348 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3349 prange->start, prange->last);
3351 mutex_lock(&prange->migrate_mutex);
3353 r = svm_migrate_vram_to_ram(prange, mm,
3354 KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3355 } while (!r && prange->actual_loc && --retries);
3357 if (!r && prange->actual_loc)
3358 pr_info_once("Migration failed during eviction");
3360 if (!prange->actual_loc) {
3361 mutex_lock(&prange->lock);
3362 prange->svm_bo = NULL;
3363 mutex_unlock(&prange->lock);
3365 mutex_unlock(&prange->migrate_mutex);
3367 spin_lock(&svm_bo->list_lock);
3369 spin_unlock(&svm_bo->list_lock);
3370 mmap_read_unlock(mm);
3373 dma_fence_signal(&svm_bo->eviction_fence->base);
3375 /* This is the last reference to svm_bo, after svm_range_vram_node_free
3376 * has been called in svm_migrate_vram_to_ram
3378 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3379 svm_range_bo_unref(svm_bo);
3383 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3384 uint64_t start, uint64_t size, uint32_t nattr,
3385 struct kfd_ioctl_svm_attribute *attrs)
3387 struct amdkfd_process_info *process_info = p->kgd_process_info;
3388 struct list_head update_list;
3389 struct list_head insert_list;
3390 struct list_head remove_list;
3391 struct svm_range_list *svms;
3392 struct svm_range *prange;
3393 struct svm_range *next;
3394 bool update_mapping = false;
3398 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3399 p->pasid, &p->svms, start, start + size - 1, size);
3401 r = svm_range_check_attr(p, nattr, attrs);
3407 mutex_lock(&process_info->lock);
3409 svm_range_list_lock_and_flush_work(svms, mm);
3411 r = svm_range_is_valid(p, start, size);
3413 pr_debug("invalid range r=%d\n", r);
3414 mmap_write_unlock(mm);
3418 mutex_lock(&svms->lock);
3420 /* Add new range and split existing ranges as needed */
3421 r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3422 &insert_list, &remove_list);
3424 mutex_unlock(&svms->lock);
3425 mmap_write_unlock(mm);
3428 /* Apply changes as a transaction */
3429 list_for_each_entry_safe(prange, next, &insert_list, list) {
3430 svm_range_add_to_svms(prange);
3431 svm_range_add_notifier_locked(mm, prange);
3433 list_for_each_entry(prange, &update_list, update_list) {
3434 svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3435 /* TODO: unmap ranges from GPU that lost access */
3437 list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3438 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3439 prange->svms, prange, prange->start,
3441 svm_range_unlink(prange);
3442 svm_range_remove_notifier(prange);
3443 svm_range_free(prange, false);
3446 mmap_write_downgrade(mm);
3447 /* Trigger migrations and revalidate and map to GPUs as needed. If
3448 * this fails we may be left with partially completed actions. There
3449 * is no clean way of rolling back to the previous state in such a
3450 * case because the rollback wouldn't be guaranteed to work either.
3452 list_for_each_entry(prange, &update_list, update_list) {
3455 mutex_lock(&prange->migrate_mutex);
3457 r = svm_range_trigger_migration(mm, prange, &migrated);
3459 goto out_unlock_range;
3461 if (migrated && (!p->xnack_enabled ||
3462 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3463 prange->mapped_to_gpu) {
3464 pr_debug("restore_work will update mappings of GPUs\n");
3465 mutex_unlock(&prange->migrate_mutex);
3469 if (!migrated && !update_mapping) {
3470 mutex_unlock(&prange->migrate_mutex);
3474 flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3476 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3477 true, true, flush_tlb);
3479 pr_debug("failed %d to map svm range\n", r);
3482 mutex_unlock(&prange->migrate_mutex);
3487 svm_range_debug_dump(svms);
3489 mutex_unlock(&svms->lock);
3490 mmap_read_unlock(mm);
3492 mutex_unlock(&process_info->lock);
3494 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3495 &p->svms, start, start + size - 1, r);
3501 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3502 uint64_t start, uint64_t size, uint32_t nattr,
3503 struct kfd_ioctl_svm_attribute *attrs)
3505 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3506 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3507 bool get_preferred_loc = false;
3508 bool get_prefetch_loc = false;
3509 bool get_granularity = false;
3510 bool get_accessible = false;
3511 bool get_flags = false;
3512 uint64_t last = start + size - 1UL;
3513 uint8_t granularity = 0xff;
3514 struct interval_tree_node *node;
3515 struct svm_range_list *svms;
3516 struct svm_range *prange;
3517 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3518 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3519 uint32_t flags_and = 0xffffffff;
3520 uint32_t flags_or = 0;
3525 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3526 start + size - 1, nattr);
3528 /* Flush pending deferred work to avoid racing with deferred actions from
3529 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3530 * can still race with get_attr because we don't hold the mmap lock. But that
3531 * would be a race condition in the application anyway, and undefined
3532 * behaviour is acceptable in that case.
3534 flush_work(&p->svms.deferred_list_work);
3537 r = svm_range_is_valid(p, start, size);
3538 mmap_read_unlock(mm);
3540 pr_debug("invalid range r=%d\n", r);
3544 for (i = 0; i < nattr; i++) {
3545 switch (attrs[i].type) {
3546 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3547 get_preferred_loc = true;
3549 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3550 get_prefetch_loc = true;
3552 case KFD_IOCTL_SVM_ATTR_ACCESS:
3553 get_accessible = true;
3555 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3556 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3559 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3560 get_granularity = true;
3562 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3563 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3566 pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3573 mutex_lock(&svms->lock);
3575 node = interval_tree_iter_first(&svms->objects, start, last);
3577 pr_debug("range attrs not found return default values\n");
3578 svm_range_set_default_attributes(&location, &prefetch_loc,
3579 &granularity, &flags_and);
3580 flags_or = flags_and;
3581 if (p->xnack_enabled)
3582 bitmap_copy(bitmap_access, svms->bitmap_supported,
3585 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3586 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3589 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3590 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3593 struct interval_tree_node *next;
3595 prange = container_of(node, struct svm_range, it_node);
3596 next = interval_tree_iter_next(node, start, last);
3598 if (get_preferred_loc) {
3599 if (prange->preferred_loc ==
3600 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3601 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3602 location != prange->preferred_loc)) {
3603 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3604 get_preferred_loc = false;
3606 location = prange->preferred_loc;
3609 if (get_prefetch_loc) {
3610 if (prange->prefetch_loc ==
3611 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3612 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3613 prefetch_loc != prange->prefetch_loc)) {
3614 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3615 get_prefetch_loc = false;
3617 prefetch_loc = prange->prefetch_loc;
3620 if (get_accessible) {
3621 bitmap_and(bitmap_access, bitmap_access,
3622 prange->bitmap_access, MAX_GPU_INSTANCE);
3623 bitmap_and(bitmap_aip, bitmap_aip,
3624 prange->bitmap_aip, MAX_GPU_INSTANCE);
3627 flags_and &= prange->flags;
3628 flags_or |= prange->flags;
3631 if (get_granularity && prange->granularity < granularity)
3632 granularity = prange->granularity;
3637 mutex_unlock(&svms->lock);
3639 for (i = 0; i < nattr; i++) {
3640 switch (attrs[i].type) {
3641 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3642 attrs[i].value = location;
3644 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3645 attrs[i].value = prefetch_loc;
3647 case KFD_IOCTL_SVM_ATTR_ACCESS:
3648 gpuidx = kfd_process_gpuidx_from_gpuid(p,
3651 pr_debug("invalid gpuid %x\n", attrs[i].value);
3654 if (test_bit(gpuidx, bitmap_access))
3655 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3656 else if (test_bit(gpuidx, bitmap_aip))
3658 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3660 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3662 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3663 attrs[i].value = flags_and;
3665 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3666 attrs[i].value = ~flags_or;
3668 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3669 attrs[i].value = (uint32_t)granularity;
3677 int kfd_criu_resume_svm(struct kfd_process *p)
3679 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3680 int nattr_common = 4, nattr_accessibility = 1;
3681 struct criu_svm_metadata *criu_svm_md = NULL;
3682 struct svm_range_list *svms = &p->svms;
3683 struct criu_svm_metadata *next = NULL;
3684 uint32_t set_flags = 0xffffffff;
3685 int i, j, num_attrs, ret = 0;
3686 uint64_t set_attr_size;
3687 struct mm_struct *mm;
3689 if (list_empty(&svms->criu_svm_metadata_list)) {
3690 pr_debug("No SVM data from CRIU restore stage 2\n");
3694 mm = get_task_mm(p->lead_thread);
3696 pr_err("failed to get mm for the target process\n");
3700 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3703 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3704 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3705 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3707 for (j = 0; j < num_attrs; j++) {
3708 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3709 i, j, criu_svm_md->data.attrs[j].type,
3710 i, j, criu_svm_md->data.attrs[j].value);
3711 switch (criu_svm_md->data.attrs[j].type) {
3712 /* During Checkpoint operation, the query for
3713 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3714 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3715 * not used by the range which was checkpointed. Care
3716 * must be taken to not restore with an invalid value
3717 * otherwise the gpuidx value will be invalid and
3718 * set_attr would eventually fail so just replace those
3719 * with another dummy attribute such as
3720 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3722 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3723 if (criu_svm_md->data.attrs[j].value ==
3724 KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3725 criu_svm_md->data.attrs[j].type =
3726 KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3727 criu_svm_md->data.attrs[j].value = 0;
3730 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3731 set_flags = criu_svm_md->data.attrs[j].value;
3738 /* CLR_FLAGS is not available via get_attr during checkpoint but
3739 * it needs to be inserted before restoring the ranges so
3740 * allocate extra space for it before calling set_attr
3742 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3744 set_attr_new = krealloc(set_attr, set_attr_size,
3746 if (!set_attr_new) {
3750 set_attr = set_attr_new;
3752 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3753 sizeof(struct kfd_ioctl_svm_attribute));
3754 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3755 set_attr[num_attrs].value = ~set_flags;
3757 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3758 criu_svm_md->data.size, num_attrs + 1,
3761 pr_err("CRIU: failed to set range attributes\n");
3769 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3770 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3771 criu_svm_md->data.start_addr);
3780 int kfd_criu_restore_svm(struct kfd_process *p,
3781 uint8_t __user *user_priv_ptr,
3782 uint64_t *priv_data_offset,
3783 uint64_t max_priv_data_size)
3785 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3786 int nattr_common = 4, nattr_accessibility = 1;
3787 struct criu_svm_metadata *criu_svm_md = NULL;
3788 struct svm_range_list *svms = &p->svms;
3789 uint32_t num_devices;
3792 num_devices = p->n_pdds;
3793 /* Handle one SVM range object at a time, also the number of gpus are
3794 * assumed to be same on the restore node, checking must be done while
3795 * evaluating the topology earlier
3798 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3799 (nattr_common + nattr_accessibility * num_devices);
3800 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3802 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3805 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3807 pr_err("failed to allocate memory to store svm metadata\n");
3810 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3815 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3816 svm_priv_data_size);
3821 *priv_data_offset += svm_priv_data_size;
3823 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3833 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3834 uint64_t *svm_priv_data_size)
3836 uint64_t total_size, accessibility_size, common_attr_size;
3837 int nattr_common = 4, nattr_accessibility = 1;
3838 int num_devices = p->n_pdds;
3839 struct svm_range_list *svms;
3840 struct svm_range *prange;
3843 *svm_priv_data_size = 0;
3849 mutex_lock(&svms->lock);
3850 list_for_each_entry(prange, &svms->list, list) {
3851 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3852 prange, prange->start, prange->npages,
3853 prange->start + prange->npages - 1);
3856 mutex_unlock(&svms->lock);
3858 *num_svm_ranges = count;
3859 /* Only the accessbility attributes need to be queried for all the gpus
3860 * individually, remaining ones are spanned across the entire process
3861 * regardless of the various gpu nodes. Of the remaining attributes,
3862 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3864 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3865 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3866 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3867 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3869 * ** ACCESSBILITY ATTRIBUTES **
3870 * (Considered as one, type is altered during query, value is gpuid)
3871 * KFD_IOCTL_SVM_ATTR_ACCESS
3872 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3873 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3875 if (*num_svm_ranges > 0) {
3876 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3878 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3879 nattr_accessibility * num_devices;
3881 total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3882 common_attr_size + accessibility_size;
3884 *svm_priv_data_size = *num_svm_ranges * total_size;
3887 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
3888 *svm_priv_data_size);
3892 int kfd_criu_checkpoint_svm(struct kfd_process *p,
3893 uint8_t __user *user_priv_data,
3894 uint64_t *priv_data_offset)
3896 struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
3897 struct kfd_ioctl_svm_attribute *query_attr = NULL;
3898 uint64_t svm_priv_data_size, query_attr_size = 0;
3899 int index, nattr_common = 4, ret = 0;
3900 struct svm_range_list *svms;
3901 int num_devices = p->n_pdds;
3902 struct svm_range *prange;
3903 struct mm_struct *mm;
3909 mm = get_task_mm(p->lead_thread);
3911 pr_err("failed to get mm for the target process\n");
3915 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3916 (nattr_common + num_devices);
3918 query_attr = kzalloc(query_attr_size, GFP_KERNEL);
3924 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
3925 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
3926 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3927 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
3929 for (index = 0; index < num_devices; index++) {
3930 struct kfd_process_device *pdd = p->pdds[index];
3932 query_attr[index + nattr_common].type =
3933 KFD_IOCTL_SVM_ATTR_ACCESS;
3934 query_attr[index + nattr_common].value = pdd->user_gpu_id;
3937 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
3939 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
3946 list_for_each_entry(prange, &svms->list, list) {
3948 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
3949 svm_priv->start_addr = prange->start;
3950 svm_priv->size = prange->npages;
3951 memcpy(&svm_priv->attrs, query_attr, query_attr_size);
3952 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
3953 prange, prange->start, prange->npages,
3954 prange->start + prange->npages - 1,
3955 prange->npages * PAGE_SIZE);
3957 ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
3959 (nattr_common + num_devices),
3962 pr_err("CRIU: failed to obtain range attributes\n");
3966 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
3967 svm_priv_data_size)) {
3968 pr_err("Failed to copy svm priv to user\n");
3973 *priv_data_offset += svm_priv_data_size;
3988 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
3989 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
3991 struct mm_struct *mm = current->mm;
3994 start >>= PAGE_SHIFT;
3995 size >>= PAGE_SHIFT;
3998 case KFD_IOCTL_SVM_OP_SET_ATTR:
3999 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4001 case KFD_IOCTL_SVM_OP_GET_ATTR:
4002 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);