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 <drm/ttm/ttm_tt.h>
27 #include "amdgpu_sync.h"
28 #include "amdgpu_object.h"
29 #include "amdgpu_vm.h"
30 #include "amdgpu_hmm.h"
32 #include "amdgpu_xgmi.h"
35 #include "kfd_migrate.h"
36 #include "kfd_smi_events.h"
41 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__
43 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
45 /* Long enough to ensure no retry fault comes after svm range is restored and
46 * page table is updated.
48 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC)
50 /* Giant svm range split into smaller ranges based on this, it is decided using
51 * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to
54 static uint64_t max_svm_range_pages;
56 struct criu_svm_metadata {
57 struct list_head list;
58 struct kfd_criu_svm_range_priv_data data;
61 static void svm_range_evict_svm_bo_worker(struct work_struct *work);
63 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
64 const struct mmu_notifier_range *range,
65 unsigned long cur_seq);
67 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
68 uint64_t *bo_s, uint64_t *bo_l);
69 static const struct mmu_interval_notifier_ops svm_range_mn_ops = {
70 .invalidate = svm_range_cpu_invalidate_pagetables,
74 * svm_range_unlink - unlink svm_range from lists and interval tree
75 * @prange: svm range structure to be removed
77 * Remove the svm_range from the svms and svm_bo lists and the svms
80 * Context: The caller must hold svms->lock
82 static void svm_range_unlink(struct svm_range *prange)
84 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
85 prange, prange->start, prange->last);
88 spin_lock(&prange->svm_bo->list_lock);
89 list_del(&prange->svm_bo_list);
90 spin_unlock(&prange->svm_bo->list_lock);
93 list_del(&prange->list);
94 if (prange->it_node.start != 0 && prange->it_node.last != 0)
95 interval_tree_remove(&prange->it_node, &prange->svms->objects);
99 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange)
101 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
102 prange, prange->start, prange->last);
104 mmu_interval_notifier_insert_locked(&prange->notifier, mm,
105 prange->start << PAGE_SHIFT,
106 prange->npages << PAGE_SHIFT,
111 * svm_range_add_to_svms - add svm range to svms
112 * @prange: svm range structure to be added
114 * Add the svm range to svms interval tree and link list
116 * Context: The caller must hold svms->lock
118 static void svm_range_add_to_svms(struct svm_range *prange)
120 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms,
121 prange, prange->start, prange->last);
123 list_move_tail(&prange->list, &prange->svms->list);
124 prange->it_node.start = prange->start;
125 prange->it_node.last = prange->last;
126 interval_tree_insert(&prange->it_node, &prange->svms->objects);
129 static void svm_range_remove_notifier(struct svm_range *prange)
131 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n",
132 prange->svms, prange,
133 prange->notifier.interval_tree.start >> PAGE_SHIFT,
134 prange->notifier.interval_tree.last >> PAGE_SHIFT);
136 if (prange->notifier.interval_tree.start != 0 &&
137 prange->notifier.interval_tree.last != 0)
138 mmu_interval_notifier_remove(&prange->notifier);
142 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
144 return dma_addr && !dma_mapping_error(dev, dma_addr) &&
145 !(dma_addr & SVM_RANGE_VRAM_DOMAIN);
149 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
150 unsigned long offset, unsigned long npages,
151 unsigned long *hmm_pfns, uint32_t gpuidx)
153 enum dma_data_direction dir = DMA_BIDIRECTIONAL;
154 dma_addr_t *addr = prange->dma_addr[gpuidx];
155 struct device *dev = adev->dev;
160 addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL);
163 prange->dma_addr[gpuidx] = addr;
167 for (i = 0; i < npages; i++) {
168 if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
169 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
171 page = hmm_pfn_to_page(hmm_pfns[i]);
172 if (is_zone_device_page(page)) {
173 struct amdgpu_device *bo_adev = prange->svm_bo->node->adev;
175 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
176 bo_adev->vm_manager.vram_base_offset -
177 bo_adev->kfd.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->adev->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, 0);
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, 0)) {
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 kfd_node *node, struct svm_range *prange)
428 mutex_lock(&prange->lock);
429 if (!prange->svm_bo) {
430 mutex_unlock(&prange->lock);
433 if (prange->ttm_res) {
434 /* We still have a reference, all is well */
435 mutex_unlock(&prange->lock);
438 if (svm_bo_ref_unless_zero(prange->svm_bo)) {
440 * Migrate from GPU to GPU, remove range from source svm_bo->node
441 * range list, and return false to allocate svm_bo from destination
444 if (prange->svm_bo->node != node) {
445 mutex_unlock(&prange->lock);
447 spin_lock(&prange->svm_bo->list_lock);
448 list_del_init(&prange->svm_bo_list);
449 spin_unlock(&prange->svm_bo->list_lock);
451 svm_range_bo_unref(prange->svm_bo);
454 if (READ_ONCE(prange->svm_bo->evicting)) {
456 struct svm_range_bo *svm_bo;
457 /* The BO is getting evicted,
458 * we need to get a new one
460 mutex_unlock(&prange->lock);
461 svm_bo = prange->svm_bo;
462 f = dma_fence_get(&svm_bo->eviction_fence->base);
463 svm_range_bo_unref(prange->svm_bo);
464 /* wait for the fence to avoid long spin-loop
465 * at list_empty_careful
467 dma_fence_wait(f, false);
470 /* The BO was still around and we got
471 * a new reference to it
473 mutex_unlock(&prange->lock);
474 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n",
475 prange->svms, prange->start, prange->last);
477 prange->ttm_res = prange->svm_bo->bo->tbo.resource;
482 mutex_unlock(&prange->lock);
485 /* We need a new svm_bo. Spin-loop to wait for concurrent
486 * svm_range_bo_release to finish removing this range from
487 * its range list. After this, it is safe to reuse the
488 * svm_bo pointer and svm_bo_list head.
490 while (!list_empty_careful(&prange->svm_bo_list))
496 static struct svm_range_bo *svm_range_bo_new(void)
498 struct svm_range_bo *svm_bo;
500 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL);
504 kref_init(&svm_bo->kref);
505 INIT_LIST_HEAD(&svm_bo->range_list);
506 spin_lock_init(&svm_bo->list_lock);
512 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
515 struct amdgpu_bo_param bp;
516 struct svm_range_bo *svm_bo;
517 struct amdgpu_bo_user *ubo;
518 struct amdgpu_bo *bo;
519 struct kfd_process *p;
520 struct mm_struct *mm;
523 p = container_of(prange->svms, struct kfd_process, svms);
524 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms,
525 prange->start, prange->last);
527 if (svm_range_validate_svm_bo(node, prange))
530 svm_bo = svm_range_bo_new();
532 pr_debug("failed to alloc svm bo\n");
535 mm = get_task_mm(p->lead_thread);
537 pr_debug("failed to get mm\n");
542 svm_bo->eviction_fence =
543 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
547 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
548 svm_bo->evicting = 0;
549 memset(&bp, 0, sizeof(bp));
550 bp.size = prange->npages * PAGE_SIZE;
551 bp.byte_align = PAGE_SIZE;
552 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
553 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
554 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
555 bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
556 bp.type = ttm_bo_type_device;
559 bp.xcp_id_plus1 = node->xcp->id + 1;
561 r = amdgpu_bo_create_user(node->adev, &bp, &ubo);
563 pr_debug("failed %d to create bo\n", r);
564 goto create_bo_failed;
568 pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n",
569 bo->tbo.resource->start << PAGE_SHIFT, bp.size,
570 bp.xcp_id_plus1 - 1);
572 r = amdgpu_bo_reserve(bo, true);
574 pr_debug("failed %d to reserve bo\n", r);
575 goto reserve_bo_failed;
579 r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
581 pr_debug("failed %d to sync bo\n", r);
582 amdgpu_bo_unreserve(bo);
583 goto reserve_bo_failed;
587 r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
589 pr_debug("failed %d to reserve bo\n", r);
590 amdgpu_bo_unreserve(bo);
591 goto reserve_bo_failed;
593 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true);
595 amdgpu_bo_unreserve(bo);
598 prange->svm_bo = svm_bo;
599 prange->ttm_res = bo->tbo.resource;
602 spin_lock(&svm_bo->list_lock);
603 list_add(&prange->svm_bo_list, &svm_bo->range_list);
604 spin_unlock(&svm_bo->list_lock);
609 amdgpu_bo_unref(&bo);
611 dma_fence_put(&svm_bo->eviction_fence->base);
613 prange->ttm_res = NULL;
618 void svm_range_vram_node_free(struct svm_range *prange)
620 svm_range_bo_unref(prange->svm_bo);
621 prange->ttm_res = NULL;
625 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id)
627 struct kfd_process *p;
628 struct kfd_process_device *pdd;
630 p = container_of(prange->svms, struct kfd_process, svms);
631 pdd = kfd_process_device_data_by_id(p, gpu_id);
633 pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id);
640 struct kfd_process_device *
641 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node)
643 struct kfd_process *p;
645 p = container_of(prange->svms, struct kfd_process, svms);
647 return kfd_get_process_device_data(node, p);
650 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
652 struct ttm_operation_ctx ctx = { false, false };
654 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
656 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
660 svm_range_check_attr(struct kfd_process *p,
661 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
665 for (i = 0; i < nattr; i++) {
666 uint32_t val = attrs[i].value;
667 int gpuidx = MAX_GPU_INSTANCE;
669 switch (attrs[i].type) {
670 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
671 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM &&
672 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED)
673 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
675 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
676 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM)
677 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
679 case KFD_IOCTL_SVM_ATTR_ACCESS:
680 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
681 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
682 gpuidx = kfd_process_gpuidx_from_gpuid(p, val);
684 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
686 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
688 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
691 pr_debug("unknown attr type 0x%x\n", attrs[i].type);
696 pr_debug("no GPU 0x%x found\n", val);
698 } else if (gpuidx < MAX_GPU_INSTANCE &&
699 !test_bit(gpuidx, p->svms.bitmap_supported)) {
700 pr_debug("GPU 0x%x not supported\n", val);
709 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange,
710 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
711 bool *update_mapping)
716 for (i = 0; i < nattr; i++) {
717 switch (attrs[i].type) {
718 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
719 prange->preferred_loc = attrs[i].value;
721 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
722 prange->prefetch_loc = attrs[i].value;
724 case KFD_IOCTL_SVM_ATTR_ACCESS:
725 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
726 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
727 if (!p->xnack_enabled)
728 *update_mapping = true;
730 gpuidx = kfd_process_gpuidx_from_gpuid(p,
732 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
733 bitmap_clear(prange->bitmap_access, gpuidx, 1);
734 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
735 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
736 bitmap_set(prange->bitmap_access, gpuidx, 1);
737 bitmap_clear(prange->bitmap_aip, gpuidx, 1);
739 bitmap_clear(prange->bitmap_access, gpuidx, 1);
740 bitmap_set(prange->bitmap_aip, gpuidx, 1);
743 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
744 *update_mapping = true;
745 prange->flags |= attrs[i].value;
747 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
748 *update_mapping = true;
749 prange->flags &= ~attrs[i].value;
751 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
752 prange->granularity = attrs[i].value;
755 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
761 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
762 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
767 for (i = 0; i < nattr; i++) {
768 switch (attrs[i].type) {
769 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
770 if (prange->preferred_loc != attrs[i].value)
773 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
774 /* Prefetch should always trigger a migration even
775 * if the value of the attribute didn't change.
778 case KFD_IOCTL_SVM_ATTR_ACCESS:
779 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
780 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
781 gpuidx = kfd_process_gpuidx_from_gpuid(p,
783 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) {
784 if (test_bit(gpuidx, prange->bitmap_access) ||
785 test_bit(gpuidx, prange->bitmap_aip))
787 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) {
788 if (!test_bit(gpuidx, prange->bitmap_access))
791 if (!test_bit(gpuidx, prange->bitmap_aip))
795 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
796 if ((prange->flags & attrs[i].value) != attrs[i].value)
799 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
800 if ((prange->flags & attrs[i].value) != 0)
803 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
804 if (prange->granularity != attrs[i].value)
808 WARN_ONCE(1, "svm_range_check_attrs wasn't called?");
816 * svm_range_debug_dump - print all range information from svms
817 * @svms: svm range list header
819 * debug output svm range start, end, prefetch location from svms
820 * interval tree and link list
822 * Context: The caller must hold svms->lock
824 static void svm_range_debug_dump(struct svm_range_list *svms)
826 struct interval_tree_node *node;
827 struct svm_range *prange;
829 pr_debug("dump svms 0x%p list\n", svms);
830 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
832 list_for_each_entry(prange, &svms->list, list) {
833 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
834 prange, prange->start, prange->npages,
835 prange->start + prange->npages - 1,
839 pr_debug("dump svms 0x%p interval tree\n", svms);
840 pr_debug("range\tstart\tpage\tend\t\tlocation\n");
841 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL);
843 prange = container_of(node, struct svm_range, it_node);
844 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n",
845 prange, prange->start, prange->npages,
846 prange->start + prange->npages - 1,
848 node = interval_tree_iter_next(node, 0, ~0ULL);
853 svm_range_split_array(void *ppnew, void *ppold, size_t size,
854 uint64_t old_start, uint64_t old_n,
855 uint64_t new_start, uint64_t new_n)
857 unsigned char *new, *old, *pold;
862 pold = *(unsigned char **)ppold;
866 new = kvmalloc_array(new_n, size, GFP_KERNEL);
870 d = (new_start - old_start) * size;
871 memcpy(new, pold + d, new_n * size);
873 old = kvmalloc_array(old_n, size, GFP_KERNEL);
879 d = (new_start == old_start) ? new_n * size : 0;
880 memcpy(old, pold + d, old_n * size);
883 *(void **)ppold = old;
884 *(void **)ppnew = new;
890 svm_range_split_pages(struct svm_range *new, struct svm_range *old,
891 uint64_t start, uint64_t last)
893 uint64_t npages = last - start + 1;
896 for (i = 0; i < MAX_GPU_INSTANCE; i++) {
897 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i],
898 sizeof(*old->dma_addr[i]), old->start,
899 npages, new->start, new->npages);
908 svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
909 uint64_t start, uint64_t last)
911 uint64_t npages = last - start + 1;
913 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n",
914 new->svms, new, new->start, start, last);
916 if (new->start == old->start) {
917 new->offset = old->offset;
918 old->offset += new->npages;
920 new->offset = old->offset + npages;
923 new->svm_bo = svm_range_bo_ref(old->svm_bo);
924 new->ttm_res = old->ttm_res;
926 spin_lock(&new->svm_bo->list_lock);
927 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
928 spin_unlock(&new->svm_bo->list_lock);
934 * svm_range_split_adjust - split range and adjust
937 * @old: the old range
938 * @start: the old range adjust to start address in pages
939 * @last: the old range adjust to last address in pages
941 * Copy system memory dma_addr or vram ttm_res in old range to new
942 * range from new_start up to size new->npages, the remaining old range is from
946 * 0 - OK, -ENOMEM - out of memory
949 svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
950 uint64_t start, uint64_t last)
954 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
955 new->svms, new->start, old->start, old->last, start, last);
957 if (new->start < old->start ||
958 new->last > old->last) {
959 WARN_ONCE(1, "invalid new range start or last\n");
963 r = svm_range_split_pages(new, old, start, last);
967 if (old->actual_loc && old->ttm_res) {
968 r = svm_range_split_nodes(new, old, start, last);
973 old->npages = last - start + 1;
976 new->flags = old->flags;
977 new->preferred_loc = old->preferred_loc;
978 new->prefetch_loc = old->prefetch_loc;
979 new->actual_loc = old->actual_loc;
980 new->granularity = old->granularity;
981 new->mapped_to_gpu = old->mapped_to_gpu;
982 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
983 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
989 * svm_range_split - split a range in 2 ranges
991 * @prange: the svm range to split
992 * @start: the remaining range start address in pages
993 * @last: the remaining range last address in pages
994 * @new: the result new range generated
997 * case 1: if start == prange->start
998 * prange ==> prange[start, last]
999 * new range [last + 1, prange->last]
1001 * case 2: if last == prange->last
1002 * prange ==> prange[start, last]
1003 * new range [prange->start, start - 1]
1006 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
1009 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
1010 struct svm_range **new)
1012 uint64_t old_start = prange->start;
1013 uint64_t old_last = prange->last;
1014 struct svm_range_list *svms;
1017 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
1018 old_start, old_last, start, last);
1020 if (old_start != start && old_last != last)
1022 if (start < old_start || last > old_last)
1025 svms = prange->svms;
1026 if (old_start == start)
1027 *new = svm_range_new(svms, last + 1, old_last, false);
1029 *new = svm_range_new(svms, old_start, start - 1, false);
1033 r = svm_range_split_adjust(*new, prange, start, last);
1035 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n",
1036 r, old_start, old_last, start, last);
1037 svm_range_free(*new, false);
1045 svm_range_split_tail(struct svm_range *prange,
1046 uint64_t new_last, struct list_head *insert_list)
1048 struct svm_range *tail;
1049 int r = svm_range_split(prange, prange->start, new_last, &tail);
1052 list_add(&tail->list, insert_list);
1057 svm_range_split_head(struct svm_range *prange,
1058 uint64_t new_start, struct list_head *insert_list)
1060 struct svm_range *head;
1061 int r = svm_range_split(prange, new_start, prange->last, &head);
1064 list_add(&head->list, insert_list);
1069 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
1070 struct svm_range *pchild, enum svm_work_list_ops op)
1072 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n",
1073 pchild, pchild->start, pchild->last, prange, op);
1075 pchild->work_item.mm = mm;
1076 pchild->work_item.op = op;
1077 list_add_tail(&pchild->child_list, &prange->child_list);
1081 * svm_range_split_by_granularity - collect ranges within granularity boundary
1083 * @p: the process with svms list
1085 * @addr: the vm fault address in pages, to split the prange
1086 * @parent: parent range if prange is from child list
1087 * @prange: prange to split
1089 * Trims @prange to be a single aligned block of prange->granularity if
1090 * possible. The head and tail are added to the child_list in @parent.
1092 * Context: caller must hold mmap_read_lock and prange->lock
1095 * 0 - OK, otherwise error code
1098 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
1099 unsigned long addr, struct svm_range *parent,
1100 struct svm_range *prange)
1102 struct svm_range *head, *tail;
1103 unsigned long start, last, size;
1106 /* Align splited range start and size to granularity size, then a single
1107 * PTE will be used for whole range, this reduces the number of PTE
1108 * updated and the L1 TLB space used for translation.
1110 size = 1UL << prange->granularity;
1111 start = ALIGN_DOWN(addr, size);
1112 last = ALIGN(addr + 1, size) - 1;
1114 pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
1115 prange->svms, prange->start, prange->last, start, last, size);
1117 if (start > prange->start) {
1118 r = svm_range_split(prange, start, prange->last, &head);
1121 svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
1124 if (last < prange->last) {
1125 r = svm_range_split(prange, prange->start, last, &tail);
1128 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
1131 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
1132 if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
1133 prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
1134 pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
1135 prange, prange->start, prange->last,
1136 SVM_OP_ADD_RANGE_AND_MAP);
1141 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
1143 return (node_a->adev == node_b->adev ||
1144 amdgpu_xgmi_same_hive(node_a->adev, node_b->adev));
1148 svm_range_get_pte_flags(struct kfd_node *node,
1149 struct svm_range *prange, int domain)
1151 struct kfd_node *bo_node;
1152 uint32_t flags = prange->flags;
1153 uint32_t mapping_flags = 0;
1155 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
1156 bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
1157 bool uncached = flags & KFD_IOCTL_SVM_FLAG_UNCACHED;
1158 unsigned int mtype_local;
1160 if (domain == SVM_RANGE_VRAM_DOMAIN)
1161 bo_node = prange->svm_bo->node;
1163 switch (node->adev->ip_versions[GC_HWIP][0]) {
1164 case IP_VERSION(9, 4, 1):
1165 if (domain == SVM_RANGE_VRAM_DOMAIN) {
1166 if (bo_node == node) {
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 (svm_nodes_in_same_hive(node, bo_node))
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_node == node) {
1183 mapping_flags |= coherent ?
1184 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
1185 if (node->adev->gmc.xgmi.connected_to_cpu)
1188 mapping_flags |= coherent ?
1189 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1190 if (svm_nodes_in_same_hive(node, bo_node))
1194 mapping_flags |= coherent ?
1195 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1198 case IP_VERSION(9, 4, 3):
1199 mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC :
1200 (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW);
1203 mapping_flags |= AMDGPU_VM_MTYPE_UC;
1204 } else if (domain == SVM_RANGE_VRAM_DOMAIN) {
1205 /* local HBM region close to partition */
1206 if (bo_node->adev == node->adev &&
1207 (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
1208 mapping_flags |= mtype_local;
1209 /* local HBM region far from partition or remote XGMI GPU */
1210 else if (svm_nodes_in_same_hive(bo_node, node))
1211 mapping_flags |= AMDGPU_VM_MTYPE_NC;
1214 mapping_flags |= AMDGPU_VM_MTYPE_UC;
1215 /* system memory accessed by the APU */
1216 } else if (node->adev->flags & AMD_IS_APU) {
1217 /* On NUMA systems, locality is determined per-page
1218 * in amdgpu_gmc_override_vm_pte_flags
1220 if (num_possible_nodes() <= 1)
1221 mapping_flags |= mtype_local;
1223 mapping_flags |= AMDGPU_VM_MTYPE_NC;
1224 /* system memory accessed by the dGPU */
1226 mapping_flags |= AMDGPU_VM_MTYPE_UC;
1230 mapping_flags |= coherent ?
1231 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
1234 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE;
1236 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO)
1237 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE;
1238 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC)
1239 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
1241 pte_flags = AMDGPU_PTE_VALID;
1242 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
1243 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1245 pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags);
1250 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1251 uint64_t start, uint64_t last,
1252 struct dma_fence **fence)
1254 uint64_t init_pte_value = 0;
1256 pr_debug("[0x%llx 0x%llx]\n", start, last);
1258 return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start,
1259 last, init_pte_value, 0, 0, NULL, NULL,
1264 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,
1265 unsigned long last, uint32_t trigger)
1267 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1268 struct kfd_process_device *pdd;
1269 struct dma_fence *fence = NULL;
1270 struct kfd_process *p;
1274 if (!prange->mapped_to_gpu) {
1275 pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n",
1276 prange, prange->start, prange->last);
1280 if (prange->start == start && prange->last == last) {
1281 pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange);
1282 prange->mapped_to_gpu = false;
1285 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
1287 p = container_of(prange->svms, struct kfd_process, svms);
1289 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1290 pr_debug("unmap from gpu idx 0x%x\n", gpuidx);
1291 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1293 pr_debug("failed to find device idx %d\n", gpuidx);
1297 kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid,
1298 start, last, trigger);
1300 r = svm_range_unmap_from_gpu(pdd->dev->adev,
1301 drm_priv_to_vm(pdd->drm_priv),
1302 start, last, &fence);
1307 r = dma_fence_wait(fence, false);
1308 dma_fence_put(fence);
1313 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT);
1320 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange,
1321 unsigned long offset, unsigned long npages, bool readonly,
1322 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev,
1323 struct dma_fence **fence, bool flush_tlb)
1325 struct amdgpu_device *adev = pdd->dev->adev;
1326 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1328 unsigned long last_start;
1333 last_start = prange->start + offset;
1335 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms,
1336 last_start, last_start + npages - 1, readonly);
1338 for (i = offset; i < offset + npages; i++) {
1339 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
1340 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN;
1342 /* Collect all pages in the same address range and memory domain
1343 * that can be mapped with a single call to update mapping.
1345 if (i < offset + npages - 1 &&
1346 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
1349 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
1350 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
1352 pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain);
1354 pte_flags &= ~AMDGPU_PTE_WRITEABLE;
1356 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n",
1357 prange->svms, last_start, prange->start + i,
1358 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0,
1361 /* For dGPU mode, we use same vm_manager to allocate VRAM for
1362 * different memory partition based on fpfn/lpfn, we should use
1363 * same vm_manager.vram_base_offset regardless memory partition.
1365 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL,
1366 last_start, prange->start + i,
1368 (last_start - prange->start) << PAGE_SHIFT,
1369 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0,
1370 NULL, dma_addr, &vm->last_update);
1372 for (j = last_start - prange->start; j <= i; j++)
1373 dma_addr[j] |= last_domain;
1376 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
1379 last_start = prange->start + i + 1;
1382 r = amdgpu_vm_update_pdes(adev, vm, false);
1384 pr_debug("failed %d to update directories 0x%lx\n", r,
1390 *fence = dma_fence_get(vm->last_update);
1397 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset,
1398 unsigned long npages, bool readonly,
1399 unsigned long *bitmap, bool wait, bool flush_tlb)
1401 struct kfd_process_device *pdd;
1402 struct amdgpu_device *bo_adev = NULL;
1403 struct kfd_process *p;
1404 struct dma_fence *fence = NULL;
1408 if (prange->svm_bo && prange->ttm_res)
1409 bo_adev = prange->svm_bo->node->adev;
1411 p = container_of(prange->svms, struct kfd_process, svms);
1412 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
1413 pr_debug("mapping to gpu idx 0x%x\n", gpuidx);
1414 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1416 pr_debug("failed to find device idx %d\n", gpuidx);
1420 pdd = kfd_bind_process_to_device(pdd->dev, p);
1424 if (bo_adev && pdd->dev->adev != bo_adev &&
1425 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) {
1426 pr_debug("cannot map to device idx %d\n", gpuidx);
1430 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly,
1431 prange->dma_addr[gpuidx],
1432 bo_adev, wait ? &fence : NULL,
1438 r = dma_fence_wait(fence, false);
1439 dma_fence_put(fence);
1442 pr_debug("failed %d to dma fence wait\n", r);
1447 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1453 struct svm_validate_context {
1454 struct kfd_process *process;
1455 struct svm_range *prange;
1457 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
1458 struct ttm_validate_buffer tv[MAX_GPU_INSTANCE];
1459 struct list_head validate_list;
1460 struct ww_acquire_ctx ticket;
1463 static int svm_range_reserve_bos(struct svm_validate_context *ctx)
1465 struct kfd_process_device *pdd;
1466 struct amdgpu_vm *vm;
1470 INIT_LIST_HEAD(&ctx->validate_list);
1471 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1472 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1474 pr_debug("failed to find device idx %d\n", gpuidx);
1477 vm = drm_priv_to_vm(pdd->drm_priv);
1479 ctx->tv[gpuidx].bo = &vm->root.bo->tbo;
1480 ctx->tv[gpuidx].num_shared = 4;
1481 list_add(&ctx->tv[gpuidx].head, &ctx->validate_list);
1484 r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list,
1487 pr_debug("failed %d to reserve bo\n", r);
1491 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) {
1492 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx);
1494 pr_debug("failed to find device idx %d\n", gpuidx);
1499 r = amdgpu_vm_validate_pt_bos(pdd->dev->adev,
1500 drm_priv_to_vm(pdd->drm_priv),
1501 svm_range_bo_validate, NULL);
1503 pr_debug("failed %d validate pt bos\n", r);
1511 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1515 static void svm_range_unreserve_bos(struct svm_validate_context *ctx)
1517 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list);
1520 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
1522 struct kfd_process_device *pdd;
1524 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
1526 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev);
1530 * Validation+GPU mapping with concurrent invalidation (MMU notifiers)
1532 * To prevent concurrent destruction or change of range attributes, the
1533 * svm_read_lock must be held. The caller must not hold the svm_write_lock
1534 * because that would block concurrent evictions and lead to deadlocks. To
1535 * serialize concurrent migrations or validations of the same range, the
1536 * prange->migrate_mutex must be held.
1538 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
1541 * The following sequence ensures race-free validation and GPU mapping:
1543 * 1. Reserve page table (and SVM BO if range is in VRAM)
1544 * 2. hmm_range_fault to get page addresses (if system memory)
1545 * 3. DMA-map pages (if system memory)
1546 * 4-a. Take notifier lock
1547 * 4-b. Check that pages still valid (mmu_interval_read_retry)
1548 * 4-c. Check that the range was not split or otherwise invalidated
1549 * 4-d. Update GPU page table
1550 * 4.e. Release notifier lock
1551 * 5. Release page table (and SVM BO) reservation
1553 static int svm_range_validate_and_map(struct mm_struct *mm,
1554 struct svm_range *prange, int32_t gpuidx,
1555 bool intr, bool wait, bool flush_tlb)
1557 struct svm_validate_context *ctx;
1558 unsigned long start, end, addr;
1559 struct kfd_process *p;
1564 ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
1567 ctx->process = container_of(prange->svms, struct kfd_process, svms);
1568 ctx->prange = prange;
1571 if (gpuidx < MAX_GPU_INSTANCE) {
1572 bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE);
1573 bitmap_set(ctx->bitmap, gpuidx, 1);
1574 } else if (ctx->process->xnack_enabled) {
1575 bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
1577 /* If prefetch range to GPU, or GPU retry fault migrate range to
1578 * GPU, which has ACCESS attribute to the range, create mapping
1581 if (prange->actual_loc) {
1582 gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process,
1583 prange->actual_loc);
1585 WARN_ONCE(1, "failed get device by id 0x%x\n",
1586 prange->actual_loc);
1590 if (test_bit(gpuidx, prange->bitmap_access))
1591 bitmap_set(ctx->bitmap, gpuidx, 1);
1594 bitmap_or(ctx->bitmap, prange->bitmap_access,
1595 prange->bitmap_aip, MAX_GPU_INSTANCE);
1598 if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) {
1599 if (!prange->mapped_to_gpu) {
1604 bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE);
1607 if (prange->actual_loc && !prange->ttm_res) {
1608 /* This should never happen. actual_loc gets set by
1609 * svm_migrate_ram_to_vram after allocating a BO.
1611 WARN_ONCE(1, "VRAM BO missing during validation\n");
1616 svm_range_reserve_bos(ctx);
1618 p = container_of(prange->svms, struct kfd_process, svms);
1619 owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap,
1621 for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) {
1622 if (kfd_svm_page_owner(p, idx) != owner) {
1628 start = prange->start << PAGE_SHIFT;
1629 end = (prange->last + 1) << PAGE_SHIFT;
1630 for (addr = start; addr < end && !r; ) {
1631 struct hmm_range *hmm_range;
1632 struct vm_area_struct *vma;
1634 unsigned long offset;
1635 unsigned long npages;
1638 vma = vma_lookup(mm, addr);
1643 readonly = !(vma->vm_flags & VM_WRITE);
1645 next = min(vma->vm_end, end);
1646 npages = (next - addr) >> PAGE_SHIFT;
1647 WRITE_ONCE(p->svms.faulting_task, current);
1648 r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages,
1649 readonly, owner, NULL,
1651 WRITE_ONCE(p->svms.faulting_task, NULL);
1653 pr_debug("failed %d to get svm range pages\n", r);
1657 offset = (addr - start) >> PAGE_SHIFT;
1658 r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
1659 hmm_range->hmm_pfns);
1661 pr_debug("failed %d to dma map range\n", r);
1665 svm_range_lock(prange);
1666 if (amdgpu_hmm_range_get_pages_done(hmm_range)) {
1667 pr_debug("hmm update the range, need validate again\n");
1671 if (!list_empty(&prange->child_list)) {
1672 pr_debug("range split by unmap in parallel, validate again\n");
1677 r = svm_range_map_to_gpus(prange, offset, npages, readonly,
1678 ctx->bitmap, wait, flush_tlb);
1681 svm_range_unlock(prange);
1687 prange->validated_once = true;
1688 prange->mapped_to_gpu = true;
1692 svm_range_unreserve_bos(ctx);
1695 prange->validate_timestamp = ktime_get_boottime();
1704 * svm_range_list_lock_and_flush_work - flush pending deferred work
1706 * @svms: the svm range list
1707 * @mm: the mm structure
1709 * Context: Returns with mmap write lock held, pending deferred work flushed
1713 svm_range_list_lock_and_flush_work(struct svm_range_list *svms,
1714 struct mm_struct *mm)
1717 flush_work(&svms->deferred_list_work);
1718 mmap_write_lock(mm);
1720 if (list_empty(&svms->deferred_range_list))
1722 mmap_write_unlock(mm);
1723 pr_debug("retry flush\n");
1724 goto retry_flush_work;
1727 static void svm_range_restore_work(struct work_struct *work)
1729 struct delayed_work *dwork = to_delayed_work(work);
1730 struct amdkfd_process_info *process_info;
1731 struct svm_range_list *svms;
1732 struct svm_range *prange;
1733 struct kfd_process *p;
1734 struct mm_struct *mm;
1739 svms = container_of(dwork, struct svm_range_list, restore_work);
1740 evicted_ranges = atomic_read(&svms->evicted_ranges);
1741 if (!evicted_ranges)
1744 pr_debug("restore svm ranges\n");
1746 p = container_of(svms, struct kfd_process, svms);
1747 process_info = p->kgd_process_info;
1749 /* Keep mm reference when svm_range_validate_and_map ranges */
1750 mm = get_task_mm(p->lead_thread);
1752 pr_debug("svms 0x%p process mm gone\n", svms);
1756 mutex_lock(&process_info->lock);
1757 svm_range_list_lock_and_flush_work(svms, mm);
1758 mutex_lock(&svms->lock);
1760 evicted_ranges = atomic_read(&svms->evicted_ranges);
1762 list_for_each_entry(prange, &svms->list, list) {
1763 invalid = atomic_read(&prange->invalid);
1767 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n",
1768 prange->svms, prange, prange->start, prange->last,
1772 * If range is migrating, wait for migration is done.
1774 mutex_lock(&prange->migrate_mutex);
1776 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
1777 false, true, false);
1779 pr_debug("failed %d to map 0x%lx to gpus\n", r,
1782 mutex_unlock(&prange->migrate_mutex);
1784 goto out_reschedule;
1786 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid)
1787 goto out_reschedule;
1790 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) !=
1792 goto out_reschedule;
1796 r = kgd2kfd_resume_mm(mm);
1798 /* No recovery from this failure. Probably the CP is
1799 * hanging. No point trying again.
1801 pr_debug("failed %d to resume KFD\n", r);
1804 pr_debug("restore svm ranges successfully\n");
1807 mutex_unlock(&svms->lock);
1808 mmap_write_unlock(mm);
1809 mutex_unlock(&process_info->lock);
1811 /* If validation failed, reschedule another attempt */
1812 if (evicted_ranges) {
1813 pr_debug("reschedule to restore svm range\n");
1814 schedule_delayed_work(&svms->restore_work,
1815 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1817 kfd_smi_event_queue_restore_rescheduled(mm);
1823 * svm_range_evict - evict svm range
1824 * @prange: svm range structure
1825 * @mm: current process mm_struct
1826 * @start: starting process queue number
1827 * @last: last process queue number
1829 * Stop all queues of the process to ensure GPU doesn't access the memory, then
1830 * return to let CPU evict the buffer and proceed CPU pagetable update.
1832 * Don't need use lock to sync cpu pagetable invalidation with GPU execution.
1833 * If invalidation happens while restore work is running, restore work will
1834 * restart to ensure to get the latest CPU pages mapping to GPU, then start
1838 svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
1839 unsigned long start, unsigned long last,
1840 enum mmu_notifier_event event)
1842 struct svm_range_list *svms = prange->svms;
1843 struct svm_range *pchild;
1844 struct kfd_process *p;
1847 p = container_of(svms, struct kfd_process, svms);
1849 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
1850 svms, prange->start, prange->last, start, last);
1852 if (!p->xnack_enabled ||
1853 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
1855 bool mapped = prange->mapped_to_gpu;
1857 list_for_each_entry(pchild, &prange->child_list, child_list) {
1858 if (!pchild->mapped_to_gpu)
1861 mutex_lock_nested(&pchild->lock, 1);
1862 if (pchild->start <= last && pchild->last >= start) {
1863 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n",
1864 pchild->start, pchild->last);
1865 atomic_inc(&pchild->invalid);
1867 mutex_unlock(&pchild->lock);
1873 if (prange->start <= last && prange->last >= start)
1874 atomic_inc(&prange->invalid);
1876 evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
1877 if (evicted_ranges != 1)
1880 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n",
1881 prange->svms, prange->start, prange->last);
1883 /* First eviction, stop the queues */
1884 r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM);
1886 pr_debug("failed to quiesce KFD\n");
1888 pr_debug("schedule to restore svm %p ranges\n", svms);
1889 schedule_delayed_work(&svms->restore_work,
1890 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
1895 if (event == MMU_NOTIFY_MIGRATE)
1896 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE;
1898 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY;
1900 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
1901 prange->svms, start, last);
1902 list_for_each_entry(pchild, &prange->child_list, child_list) {
1903 mutex_lock_nested(&pchild->lock, 1);
1904 s = max(start, pchild->start);
1905 l = min(last, pchild->last);
1907 svm_range_unmap_from_gpus(pchild, s, l, trigger);
1908 mutex_unlock(&pchild->lock);
1910 s = max(start, prange->start);
1911 l = min(last, prange->last);
1913 svm_range_unmap_from_gpus(prange, s, l, trigger);
1919 static struct svm_range *svm_range_clone(struct svm_range *old)
1921 struct svm_range *new;
1923 new = svm_range_new(old->svms, old->start, old->last, false);
1928 new->ttm_res = old->ttm_res;
1929 new->offset = old->offset;
1930 new->svm_bo = svm_range_bo_ref(old->svm_bo);
1931 spin_lock(&new->svm_bo->list_lock);
1932 list_add(&new->svm_bo_list, &new->svm_bo->range_list);
1933 spin_unlock(&new->svm_bo->list_lock);
1935 new->flags = old->flags;
1936 new->preferred_loc = old->preferred_loc;
1937 new->prefetch_loc = old->prefetch_loc;
1938 new->actual_loc = old->actual_loc;
1939 new->granularity = old->granularity;
1940 new->mapped_to_gpu = old->mapped_to_gpu;
1941 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
1942 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
1947 void svm_range_set_max_pages(struct amdgpu_device *adev)
1950 uint64_t pages, _pages;
1951 uint64_t min_pages = 0;
1954 for (i = 0; i < adev->kfd.dev->num_nodes; i++) {
1955 if (adev->kfd.dev->nodes[i]->xcp)
1956 id = adev->kfd.dev->nodes[i]->xcp->id;
1959 pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17;
1960 pages = clamp(pages, 1ULL << 9, 1ULL << 18);
1961 pages = rounddown_pow_of_two(pages);
1962 min_pages = min_not_zero(min_pages, pages);
1966 max_pages = READ_ONCE(max_svm_range_pages);
1967 _pages = min_not_zero(max_pages, min_pages);
1968 } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages);
1972 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last,
1973 uint64_t max_pages, struct list_head *insert_list,
1974 struct list_head *update_list)
1976 struct svm_range *prange;
1979 pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
1980 max_pages, start, last);
1982 while (last >= start) {
1983 l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1);
1985 prange = svm_range_new(svms, start, l, true);
1988 list_add(&prange->list, insert_list);
1989 list_add(&prange->update_list, update_list);
1997 * svm_range_add - add svm range and handle overlap
1998 * @p: the range add to this process svms
1999 * @start: page size aligned
2000 * @size: page size aligned
2001 * @nattr: number of attributes
2002 * @attrs: array of attributes
2003 * @update_list: output, the ranges need validate and update GPU mapping
2004 * @insert_list: output, the ranges need insert to svms
2005 * @remove_list: output, the ranges are replaced and need remove from svms
2007 * Check if the virtual address range has overlap with any existing ranges,
2008 * split partly overlapping ranges and add new ranges in the gaps. All changes
2009 * should be applied to the range_list and interval tree transactionally. If
2010 * any range split or allocation fails, the entire update fails. Therefore any
2011 * existing overlapping svm_ranges are cloned and the original svm_ranges left
2014 * If the transaction succeeds, the caller can update and insert clones and
2015 * new ranges, then free the originals.
2017 * Otherwise the caller can free the clones and new ranges, while the old
2018 * svm_ranges remain unchanged.
2020 * Context: Process context, caller must hold svms->lock
2023 * 0 - OK, otherwise error code
2026 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
2027 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs,
2028 struct list_head *update_list, struct list_head *insert_list,
2029 struct list_head *remove_list)
2031 unsigned long last = start + size - 1UL;
2032 struct svm_range_list *svms = &p->svms;
2033 struct interval_tree_node *node;
2034 struct svm_range *prange;
2035 struct svm_range *tmp;
2036 struct list_head new_list;
2039 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last);
2041 INIT_LIST_HEAD(update_list);
2042 INIT_LIST_HEAD(insert_list);
2043 INIT_LIST_HEAD(remove_list);
2044 INIT_LIST_HEAD(&new_list);
2046 node = interval_tree_iter_first(&svms->objects, start, last);
2048 struct interval_tree_node *next;
2049 unsigned long next_start;
2051 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
2054 prange = container_of(node, struct svm_range, it_node);
2055 next = interval_tree_iter_next(node, start, last);
2056 next_start = min(node->last, last) + 1;
2058 if (svm_range_is_same_attrs(p, prange, nattr, attrs)) {
2060 } else if (node->start < start || node->last > last) {
2061 /* node intersects the update range and its attributes
2062 * will change. Clone and split it, apply updates only
2063 * to the overlapping part
2065 struct svm_range *old = prange;
2067 prange = svm_range_clone(old);
2073 list_add(&old->update_list, remove_list);
2074 list_add(&prange->list, insert_list);
2075 list_add(&prange->update_list, update_list);
2077 if (node->start < start) {
2078 pr_debug("change old range start\n");
2079 r = svm_range_split_head(prange, start,
2084 if (node->last > last) {
2085 pr_debug("change old range last\n");
2086 r = svm_range_split_tail(prange, last,
2092 /* The node is contained within start..last,
2095 list_add(&prange->update_list, update_list);
2098 /* insert a new node if needed */
2099 if (node->start > start) {
2100 r = svm_range_split_new(svms, start, node->start - 1,
2101 READ_ONCE(max_svm_range_pages),
2102 &new_list, update_list);
2111 /* add a final range at the end if needed */
2113 r = svm_range_split_new(svms, start, last,
2114 READ_ONCE(max_svm_range_pages),
2115 &new_list, update_list);
2119 list_for_each_entry_safe(prange, tmp, insert_list, list)
2120 svm_range_free(prange, false);
2121 list_for_each_entry_safe(prange, tmp, &new_list, list)
2122 svm_range_free(prange, true);
2124 list_splice(&new_list, insert_list);
2131 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm,
2132 struct svm_range *prange)
2134 unsigned long start;
2137 start = prange->notifier.interval_tree.start >> PAGE_SHIFT;
2138 last = prange->notifier.interval_tree.last >> PAGE_SHIFT;
2140 if (prange->start == start && prange->last == last)
2143 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
2144 prange->svms, prange, start, last, prange->start,
2147 if (start != 0 && last != 0) {
2148 interval_tree_remove(&prange->it_node, &prange->svms->objects);
2149 svm_range_remove_notifier(prange);
2151 prange->it_node.start = prange->start;
2152 prange->it_node.last = prange->last;
2154 interval_tree_insert(&prange->it_node, &prange->svms->objects);
2155 svm_range_add_notifier_locked(mm, prange);
2159 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange,
2160 struct mm_struct *mm)
2162 switch (prange->work_item.op) {
2164 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2165 svms, prange, prange->start, prange->last);
2167 case SVM_OP_UNMAP_RANGE:
2168 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2169 svms, prange, prange->start, prange->last);
2170 svm_range_unlink(prange);
2171 svm_range_remove_notifier(prange);
2172 svm_range_free(prange, true);
2174 case SVM_OP_UPDATE_RANGE_NOTIFIER:
2175 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2176 svms, prange, prange->start, prange->last);
2177 svm_range_update_notifier_and_interval_tree(mm, prange);
2179 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP:
2180 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n",
2181 svms, prange, prange->start, prange->last);
2182 svm_range_update_notifier_and_interval_tree(mm, prange);
2183 /* TODO: implement deferred validation and mapping */
2185 case SVM_OP_ADD_RANGE:
2186 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange,
2187 prange->start, prange->last);
2188 svm_range_add_to_svms(prange);
2189 svm_range_add_notifier_locked(mm, prange);
2191 case SVM_OP_ADD_RANGE_AND_MAP:
2192 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms,
2193 prange, prange->start, prange->last);
2194 svm_range_add_to_svms(prange);
2195 svm_range_add_notifier_locked(mm, prange);
2196 /* TODO: implement deferred validation and mapping */
2199 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange,
2200 prange->work_item.op);
2204 static void svm_range_drain_retry_fault(struct svm_range_list *svms)
2206 struct kfd_process_device *pdd;
2207 struct kfd_process *p;
2211 p = container_of(svms, struct kfd_process, svms);
2214 drain = atomic_read(&svms->drain_pagefaults);
2218 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
2223 pr_debug("drain retry fault gpu %d svms %p\n", i, svms);
2225 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2226 pdd->dev->adev->irq.retry_cam_enabled ?
2227 &pdd->dev->adev->irq.ih :
2228 &pdd->dev->adev->irq.ih1);
2230 if (pdd->dev->adev->irq.retry_cam_enabled)
2231 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev,
2232 &pdd->dev->adev->irq.ih_soft);
2235 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
2237 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
2241 static void svm_range_deferred_list_work(struct work_struct *work)
2243 struct svm_range_list *svms;
2244 struct svm_range *prange;
2245 struct mm_struct *mm;
2247 svms = container_of(work, struct svm_range_list, deferred_list_work);
2248 pr_debug("enter svms 0x%p\n", svms);
2250 spin_lock(&svms->deferred_list_lock);
2251 while (!list_empty(&svms->deferred_range_list)) {
2252 prange = list_first_entry(&svms->deferred_range_list,
2253 struct svm_range, deferred_list);
2254 spin_unlock(&svms->deferred_list_lock);
2256 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
2257 prange->start, prange->last, prange->work_item.op);
2259 mm = prange->work_item.mm;
2261 mmap_write_lock(mm);
2263 /* Checking for the need to drain retry faults must be inside
2264 * mmap write lock to serialize with munmap notifiers.
2266 if (unlikely(atomic_read(&svms->drain_pagefaults))) {
2267 mmap_write_unlock(mm);
2268 svm_range_drain_retry_fault(svms);
2272 /* Remove from deferred_list must be inside mmap write lock, for
2274 * 1. unmap_from_cpu may change work_item.op and add the range
2275 * to deferred_list again, cause use after free bug.
2276 * 2. svm_range_list_lock_and_flush_work may hold mmap write
2277 * lock and continue because deferred_list is empty, but
2278 * deferred_list work is actually waiting for mmap lock.
2280 spin_lock(&svms->deferred_list_lock);
2281 list_del_init(&prange->deferred_list);
2282 spin_unlock(&svms->deferred_list_lock);
2284 mutex_lock(&svms->lock);
2285 mutex_lock(&prange->migrate_mutex);
2286 while (!list_empty(&prange->child_list)) {
2287 struct svm_range *pchild;
2289 pchild = list_first_entry(&prange->child_list,
2290 struct svm_range, child_list);
2291 pr_debug("child prange 0x%p op %d\n", pchild,
2292 pchild->work_item.op);
2293 list_del_init(&pchild->child_list);
2294 svm_range_handle_list_op(svms, pchild, mm);
2296 mutex_unlock(&prange->migrate_mutex);
2298 svm_range_handle_list_op(svms, prange, mm);
2299 mutex_unlock(&svms->lock);
2300 mmap_write_unlock(mm);
2302 /* Pairs with mmget in svm_range_add_list_work */
2305 spin_lock(&svms->deferred_list_lock);
2307 spin_unlock(&svms->deferred_list_lock);
2308 pr_debug("exit svms 0x%p\n", svms);
2312 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
2313 struct mm_struct *mm, enum svm_work_list_ops op)
2315 spin_lock(&svms->deferred_list_lock);
2316 /* if prange is on the deferred list */
2317 if (!list_empty(&prange->deferred_list)) {
2318 pr_debug("update exist prange 0x%p work op %d\n", prange, op);
2319 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n");
2320 if (op != SVM_OP_NULL &&
2321 prange->work_item.op != SVM_OP_UNMAP_RANGE)
2322 prange->work_item.op = op;
2324 prange->work_item.op = op;
2326 /* Pairs with mmput in deferred_list_work */
2328 prange->work_item.mm = mm;
2329 list_add_tail(&prange->deferred_list,
2330 &prange->svms->deferred_range_list);
2331 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n",
2332 prange, prange->start, prange->last, op);
2334 spin_unlock(&svms->deferred_list_lock);
2337 void schedule_deferred_list_work(struct svm_range_list *svms)
2339 spin_lock(&svms->deferred_list_lock);
2340 if (!list_empty(&svms->deferred_range_list))
2341 schedule_work(&svms->deferred_list_work);
2342 spin_unlock(&svms->deferred_list_lock);
2346 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent,
2347 struct svm_range *prange, unsigned long start,
2350 struct svm_range *head;
2351 struct svm_range *tail;
2353 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2354 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange,
2355 prange->start, prange->last);
2358 if (start > prange->last || last < prange->start)
2361 head = tail = prange;
2362 if (start > prange->start)
2363 svm_range_split(prange, prange->start, start - 1, &tail);
2364 if (last < tail->last)
2365 svm_range_split(tail, last + 1, tail->last, &head);
2367 if (head != prange && tail != prange) {
2368 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2369 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
2370 } else if (tail != prange) {
2371 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE);
2372 } else if (head != prange) {
2373 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE);
2374 } else if (parent != prange) {
2375 prange->work_item.op = SVM_OP_UNMAP_RANGE;
2380 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
2381 unsigned long start, unsigned long last)
2383 uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU;
2384 struct svm_range_list *svms;
2385 struct svm_range *pchild;
2386 struct kfd_process *p;
2390 p = kfd_lookup_process_by_mm(mm);
2395 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
2396 prange, prange->start, prange->last, start, last);
2398 /* Make sure pending page faults are drained in the deferred worker
2399 * before the range is freed to avoid straggler interrupts on
2400 * unmapped memory causing "phantom faults".
2402 atomic_inc(&svms->drain_pagefaults);
2404 unmap_parent = start <= prange->start && last >= prange->last;
2406 list_for_each_entry(pchild, &prange->child_list, child_list) {
2407 mutex_lock_nested(&pchild->lock, 1);
2408 s = max(start, pchild->start);
2409 l = min(last, pchild->last);
2411 svm_range_unmap_from_gpus(pchild, s, l, trigger);
2412 svm_range_unmap_split(mm, prange, pchild, start, last);
2413 mutex_unlock(&pchild->lock);
2415 s = max(start, prange->start);
2416 l = min(last, prange->last);
2418 svm_range_unmap_from_gpus(prange, s, l, trigger);
2419 svm_range_unmap_split(mm, prange, prange, start, last);
2422 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE);
2424 svm_range_add_list_work(svms, prange, mm,
2425 SVM_OP_UPDATE_RANGE_NOTIFIER);
2426 schedule_deferred_list_work(svms);
2428 kfd_unref_process(p);
2432 * svm_range_cpu_invalidate_pagetables - interval notifier callback
2433 * @mni: mmu_interval_notifier struct
2434 * @range: mmu_notifier_range struct
2435 * @cur_seq: value to pass to mmu_interval_set_seq()
2437 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it
2438 * is from migration, or CPU page invalidation callback.
2440 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed
2441 * work thread, and split prange if only part of prange is unmapped.
2443 * For invalidation event, if GPU retry fault is not enabled, evict the queues,
2444 * then schedule svm_range_restore_work to update GPU mapping and resume queues.
2445 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will
2446 * update GPU mapping to recover.
2448 * Context: mmap lock, notifier_invalidate_start lock are held
2449 * for invalidate event, prange lock is held if this is from migration
2452 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,
2453 const struct mmu_notifier_range *range,
2454 unsigned long cur_seq)
2456 struct svm_range *prange;
2457 unsigned long start;
2460 if (range->event == MMU_NOTIFY_RELEASE)
2462 if (!mmget_not_zero(mni->mm))
2465 start = mni->interval_tree.start;
2466 last = mni->interval_tree.last;
2467 start = max(start, range->start) >> PAGE_SHIFT;
2468 last = min(last, range->end - 1) >> PAGE_SHIFT;
2469 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n",
2470 start, last, range->start >> PAGE_SHIFT,
2471 (range->end - 1) >> PAGE_SHIFT,
2472 mni->interval_tree.start >> PAGE_SHIFT,
2473 mni->interval_tree.last >> PAGE_SHIFT, range->event);
2475 prange = container_of(mni, struct svm_range, notifier);
2477 svm_range_lock(prange);
2478 mmu_interval_set_seq(mni, cur_seq);
2480 switch (range->event) {
2481 case MMU_NOTIFY_UNMAP:
2482 svm_range_unmap_from_cpu(mni->mm, prange, start, last);
2485 svm_range_evict(prange, mni->mm, start, last, range->event);
2489 svm_range_unlock(prange);
2496 * svm_range_from_addr - find svm range from fault address
2497 * @svms: svm range list header
2498 * @addr: address to search range interval tree, in pages
2499 * @parent: parent range if range is on child list
2501 * Context: The caller must hold svms->lock
2503 * Return: the svm_range found or NULL
2506 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
2507 struct svm_range **parent)
2509 struct interval_tree_node *node;
2510 struct svm_range *prange;
2511 struct svm_range *pchild;
2513 node = interval_tree_iter_first(&svms->objects, addr, addr);
2517 prange = container_of(node, struct svm_range, it_node);
2518 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n",
2519 addr, prange->start, prange->last, node->start, node->last);
2521 if (addr >= prange->start && addr <= prange->last) {
2526 list_for_each_entry(pchild, &prange->child_list, child_list)
2527 if (addr >= pchild->start && addr <= pchild->last) {
2528 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n",
2529 addr, pchild->start, pchild->last);
2538 /* svm_range_best_restore_location - decide the best fault restore location
2539 * @prange: svm range structure
2540 * @adev: the GPU on which vm fault happened
2542 * This is only called when xnack is on, to decide the best location to restore
2543 * the range mapping after GPU vm fault. Caller uses the best location to do
2544 * migration if actual loc is not best location, then update GPU page table
2545 * mapping to the best location.
2547 * If the preferred loc is accessible by faulting GPU, use preferred loc.
2548 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
2549 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
2550 * if range actual loc is cpu, best_loc is cpu
2551 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is
2553 * Otherwise, GPU no access, best_loc is -1.
2556 * -1 means vm fault GPU no access
2557 * 0 for CPU or GPU id
2560 svm_range_best_restore_location(struct svm_range *prange,
2561 struct kfd_node *node,
2564 struct kfd_node *bo_node, *preferred_node;
2565 struct kfd_process *p;
2569 p = container_of(prange->svms, struct kfd_process, svms);
2571 r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx);
2573 pr_debug("failed to get gpuid from kgd\n");
2577 if (node->adev->gmc.is_app_apu)
2580 if (prange->preferred_loc == gpuid ||
2581 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
2582 return prange->preferred_loc;
2583 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
2584 preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc);
2585 if (preferred_node && svm_nodes_in_same_hive(node, preferred_node))
2586 return prange->preferred_loc;
2590 if (test_bit(*gpuidx, prange->bitmap_access))
2593 if (test_bit(*gpuidx, prange->bitmap_aip)) {
2594 if (!prange->actual_loc)
2597 bo_node = svm_range_get_node_by_id(prange, prange->actual_loc);
2598 if (bo_node && svm_nodes_in_same_hive(node, bo_node))
2599 return prange->actual_loc;
2608 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
2609 unsigned long *start, unsigned long *last,
2610 bool *is_heap_stack)
2612 struct vm_area_struct *vma;
2613 struct interval_tree_node *node;
2614 unsigned long start_limit, end_limit;
2616 vma = vma_lookup(p->mm, addr << PAGE_SHIFT);
2618 pr_debug("VMA does not exist in address [0x%llx]\n", addr);
2622 *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
2623 vma->vm_end >= vma->vm_mm->start_brk) ||
2624 (vma->vm_start <= vma->vm_mm->start_stack &&
2625 vma->vm_end >= vma->vm_mm->start_stack);
2627 start_limit = max(vma->vm_start >> PAGE_SHIFT,
2628 (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
2629 end_limit = min(vma->vm_end >> PAGE_SHIFT,
2630 (unsigned long)ALIGN(addr + 1, 2UL << 8));
2631 /* First range that starts after the fault address */
2632 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX);
2634 end_limit = min(end_limit, node->start);
2635 /* Last range that ends before the fault address */
2636 node = container_of(rb_prev(&node->rb),
2637 struct interval_tree_node, rb);
2639 /* Last range must end before addr because
2640 * there was no range after addr
2642 node = container_of(rb_last(&p->svms.objects.rb_root),
2643 struct interval_tree_node, rb);
2646 if (node->last >= addr) {
2647 WARN(1, "Overlap with prev node and page fault addr\n");
2650 start_limit = max(start_limit, node->last + 1);
2653 *start = start_limit;
2654 *last = end_limit - 1;
2656 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
2657 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
2658 *start, *last, *is_heap_stack);
2664 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last,
2665 uint64_t *bo_s, uint64_t *bo_l)
2667 struct amdgpu_bo_va_mapping *mapping;
2668 struct interval_tree_node *node;
2669 struct amdgpu_bo *bo = NULL;
2670 unsigned long userptr;
2674 for (i = 0; i < p->n_pdds; i++) {
2675 struct amdgpu_vm *vm;
2677 if (!p->pdds[i]->drm_priv)
2680 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
2681 r = amdgpu_bo_reserve(vm->root.bo, false);
2685 /* Check userptr by searching entire vm->va interval tree */
2686 node = interval_tree_iter_first(&vm->va, 0, ~0ULL);
2688 mapping = container_of((struct rb_node *)node,
2689 struct amdgpu_bo_va_mapping, rb);
2690 bo = mapping->bo_va->base.bo;
2692 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
2693 start << PAGE_SHIFT,
2696 node = interval_tree_iter_next(node, 0, ~0ULL);
2700 pr_debug("[0x%llx 0x%llx] already userptr mapped\n",
2703 *bo_s = userptr >> PAGE_SHIFT;
2704 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1;
2706 amdgpu_bo_unreserve(vm->root.bo);
2709 amdgpu_bo_unreserve(vm->root.bo);
2715 svm_range *svm_range_create_unregistered_range(struct kfd_node *node,
2716 struct kfd_process *p,
2717 struct mm_struct *mm,
2720 struct svm_range *prange = NULL;
2721 unsigned long start, last;
2722 uint32_t gpuid, gpuidx;
2728 if (svm_range_get_range_boundaries(p, addr, &start, &last,
2732 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
2733 if (r != -EADDRINUSE)
2734 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l);
2736 if (r == -EADDRINUSE) {
2737 if (addr >= bo_s && addr <= bo_l)
2740 /* Create one page svm range if 2MB range overlapping */
2745 prange = svm_range_new(&p->svms, start, last, true);
2747 pr_debug("Failed to create prange in address [0x%llx]\n", addr);
2750 if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) {
2751 pr_debug("failed to get gpuid from kgd\n");
2752 svm_range_free(prange, true);
2757 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
2759 svm_range_add_to_svms(prange);
2760 svm_range_add_notifier_locked(mm, prange);
2765 /* svm_range_skip_recover - decide if prange can be recovered
2766 * @prange: svm range structure
2768 * GPU vm retry fault handle skip recover the range for cases:
2769 * 1. prange is on deferred list to be removed after unmap, it is stale fault,
2770 * deferred list work will drain the stale fault before free the prange.
2771 * 2. prange is on deferred list to add interval notifier after split, or
2772 * 3. prange is child range, it is split from parent prange, recover later
2773 * after interval notifier is added.
2775 * Return: true to skip recover, false to recover
2777 static bool svm_range_skip_recover(struct svm_range *prange)
2779 struct svm_range_list *svms = prange->svms;
2781 spin_lock(&svms->deferred_list_lock);
2782 if (list_empty(&prange->deferred_list) &&
2783 list_empty(&prange->child_list)) {
2784 spin_unlock(&svms->deferred_list_lock);
2787 spin_unlock(&svms->deferred_list_lock);
2789 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) {
2790 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n",
2791 svms, prange, prange->start, prange->last);
2794 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP ||
2795 prange->work_item.op == SVM_OP_ADD_RANGE) {
2796 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n",
2797 svms, prange, prange->start, prange->last);
2804 svm_range_count_fault(struct kfd_node *node, struct kfd_process *p,
2807 struct kfd_process_device *pdd;
2809 /* fault is on different page of same range
2810 * or fault is skipped to recover later
2811 * or fault is on invalid virtual address
2813 if (gpuidx == MAX_GPU_INSTANCE) {
2817 r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx);
2822 /* fault is recovered
2823 * or fault cannot recover because GPU no access on the range
2825 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
2827 WRITE_ONCE(pdd->faults, pdd->faults + 1);
2831 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
2833 unsigned long requested = VM_READ;
2836 requested |= VM_WRITE;
2838 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
2840 return (vma->vm_flags & requested) == requested;
2844 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
2845 uint32_t vmid, uint32_t node_id,
2846 uint64_t addr, bool write_fault)
2848 struct mm_struct *mm = NULL;
2849 struct svm_range_list *svms;
2850 struct svm_range *prange;
2851 struct kfd_process *p;
2852 ktime_t timestamp = ktime_get_boottime();
2853 struct kfd_node *node;
2855 int32_t gpuidx = MAX_GPU_INSTANCE;
2856 bool write_locked = false;
2857 struct vm_area_struct *vma;
2858 bool migration = false;
2861 if (!KFD_IS_SVM_API_SUPPORTED(adev)) {
2862 pr_debug("device does not support SVM\n");
2866 p = kfd_lookup_process_by_pasid(pasid);
2868 pr_debug("kfd process not founded pasid 0x%x\n", pasid);
2873 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
2875 if (atomic_read(&svms->drain_pagefaults)) {
2876 pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
2881 if (!p->xnack_enabled) {
2882 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
2887 /* p->lead_thread is available as kfd_process_wq_release flush the work
2888 * before releasing task ref.
2890 mm = get_task_mm(p->lead_thread);
2892 pr_debug("svms 0x%p failed to get mm\n", svms);
2897 node = kfd_node_by_irq_ids(adev, node_id, vmid);
2899 pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id,
2906 mutex_lock(&svms->lock);
2907 prange = svm_range_from_addr(svms, addr, NULL);
2909 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
2911 if (!write_locked) {
2912 /* Need the write lock to create new range with MMU notifier.
2913 * Also flush pending deferred work to make sure the interval
2914 * tree is up to date before we add a new range
2916 mutex_unlock(&svms->lock);
2917 mmap_read_unlock(mm);
2918 mmap_write_lock(mm);
2919 write_locked = true;
2920 goto retry_write_locked;
2922 prange = svm_range_create_unregistered_range(node, p, mm, addr);
2924 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n",
2926 mmap_write_downgrade(mm);
2928 goto out_unlock_svms;
2932 mmap_write_downgrade(mm);
2934 mutex_lock(&prange->migrate_mutex);
2936 if (svm_range_skip_recover(prange)) {
2937 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
2939 goto out_unlock_range;
2942 /* skip duplicate vm fault on different pages of same range */
2943 if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
2944 AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
2945 pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
2946 svms, prange->start, prange->last);
2948 goto out_unlock_range;
2951 /* __do_munmap removed VMA, return success as we are handling stale
2954 vma = vma_lookup(mm, addr << PAGE_SHIFT);
2956 pr_debug("address 0x%llx VMA is removed\n", addr);
2958 goto out_unlock_range;
2961 if (!svm_fault_allowed(vma, write_fault)) {
2962 pr_debug("fault addr 0x%llx no %s permission\n", addr,
2963 write_fault ? "write" : "read");
2965 goto out_unlock_range;
2968 best_loc = svm_range_best_restore_location(prange, node, &gpuidx);
2969 if (best_loc == -1) {
2970 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n",
2971 svms, prange->start, prange->last);
2973 goto out_unlock_range;
2976 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
2977 svms, prange->start, prange->last, best_loc,
2978 prange->actual_loc);
2980 kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
2981 write_fault, timestamp);
2983 if (prange->actual_loc != best_loc) {
2986 r = svm_migrate_to_vram(prange, best_loc, mm,
2987 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
2989 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n",
2991 /* Fallback to system memory if migration to
2994 if (prange->actual_loc)
2995 r = svm_migrate_vram_to_ram(prange, mm,
2996 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3002 r = svm_migrate_vram_to_ram(prange, mm,
3003 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
3007 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
3008 r, svms, prange->start, prange->last);
3009 goto out_unlock_range;
3013 r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
3015 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
3016 r, svms, prange->start, prange->last);
3018 kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
3022 mutex_unlock(&prange->migrate_mutex);
3024 mutex_unlock(&svms->lock);
3025 mmap_read_unlock(mm);
3027 svm_range_count_fault(node, p, gpuidx);
3031 kfd_unref_process(p);
3034 pr_debug("recover vm fault later\n");
3035 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid);
3042 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled)
3044 struct svm_range *prange, *pchild;
3045 uint64_t reserved_size = 0;
3049 pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled);
3051 mutex_lock(&p->svms.lock);
3053 list_for_each_entry(prange, &p->svms.list, list) {
3054 svm_range_lock(prange);
3055 list_for_each_entry(pchild, &prange->child_list, child_list) {
3056 size = (pchild->last - pchild->start + 1) << PAGE_SHIFT;
3057 if (xnack_enabled) {
3058 amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3059 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3061 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3062 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3065 reserved_size += size;
3069 size = (prange->last - prange->start + 1) << PAGE_SHIFT;
3070 if (xnack_enabled) {
3071 amdgpu_amdkfd_unreserve_mem_limit(NULL, size,
3072 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3074 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size,
3075 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3078 reserved_size += size;
3081 svm_range_unlock(prange);
3087 amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size,
3088 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0);
3090 /* Change xnack mode must be inside svms lock, to avoid race with
3091 * svm_range_deferred_list_work unreserve memory in parallel.
3093 p->xnack_enabled = xnack_enabled;
3095 mutex_unlock(&p->svms.lock);
3099 void svm_range_list_fini(struct kfd_process *p)
3101 struct svm_range *prange;
3102 struct svm_range *next;
3104 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms);
3106 cancel_delayed_work_sync(&p->svms.restore_work);
3108 /* Ensure list work is finished before process is destroyed */
3109 flush_work(&p->svms.deferred_list_work);
3112 * Ensure no retry fault comes in afterwards, as page fault handler will
3113 * not find kfd process and take mm lock to recover fault.
3115 atomic_inc(&p->svms.drain_pagefaults);
3116 svm_range_drain_retry_fault(&p->svms);
3118 list_for_each_entry_safe(prange, next, &p->svms.list, list) {
3119 svm_range_unlink(prange);
3120 svm_range_remove_notifier(prange);
3121 svm_range_free(prange, true);
3124 mutex_destroy(&p->svms.lock);
3126 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms);
3129 int svm_range_list_init(struct kfd_process *p)
3131 struct svm_range_list *svms = &p->svms;
3134 svms->objects = RB_ROOT_CACHED;
3135 mutex_init(&svms->lock);
3136 INIT_LIST_HEAD(&svms->list);
3137 atomic_set(&svms->evicted_ranges, 0);
3138 atomic_set(&svms->drain_pagefaults, 0);
3139 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
3140 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
3141 INIT_LIST_HEAD(&svms->deferred_range_list);
3142 INIT_LIST_HEAD(&svms->criu_svm_metadata_list);
3143 spin_lock_init(&svms->deferred_list_lock);
3145 for (i = 0; i < p->n_pdds; i++)
3146 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev))
3147 bitmap_set(svms->bitmap_supported, i, 1);
3153 * svm_range_check_vm - check if virtual address range mapped already
3154 * @p: current kfd_process
3155 * @start: range start address, in pages
3156 * @last: range last address, in pages
3157 * @bo_s: mapping start address in pages if address range already mapped
3158 * @bo_l: mapping last address in pages if address range already mapped
3160 * The purpose is to avoid virtual address ranges already allocated by
3161 * kfd_ioctl_alloc_memory_of_gpu ioctl.
3162 * It looks for each pdd in the kfd_process.
3164 * Context: Process context
3166 * Return 0 - OK, if the range is not mapped.
3167 * Otherwise error code:
3168 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu
3169 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by
3170 * a signal. Release all buffer reservations and return to user-space.
3173 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last,
3174 uint64_t *bo_s, uint64_t *bo_l)
3176 struct amdgpu_bo_va_mapping *mapping;
3177 struct interval_tree_node *node;
3181 for (i = 0; i < p->n_pdds; i++) {
3182 struct amdgpu_vm *vm;
3184 if (!p->pdds[i]->drm_priv)
3187 vm = drm_priv_to_vm(p->pdds[i]->drm_priv);
3188 r = amdgpu_bo_reserve(vm->root.bo, false);
3192 node = interval_tree_iter_first(&vm->va, start, last);
3194 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n",
3196 mapping = container_of((struct rb_node *)node,
3197 struct amdgpu_bo_va_mapping, rb);
3199 *bo_s = mapping->start;
3200 *bo_l = mapping->last;
3202 amdgpu_bo_unreserve(vm->root.bo);
3205 amdgpu_bo_unreserve(vm->root.bo);
3212 * svm_range_is_valid - check if virtual address range is valid
3213 * @p: current kfd_process
3214 * @start: range start address, in pages
3215 * @size: range size, in pages
3217 * Valid virtual address range means it belongs to one or more VMAs
3219 * Context: Process context
3222 * 0 - OK, otherwise error code
3225 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size)
3227 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
3228 struct vm_area_struct *vma;
3230 unsigned long start_unchg = start;
3232 start <<= PAGE_SHIFT;
3233 end = start + (size << PAGE_SHIFT);
3235 vma = vma_lookup(p->mm, start);
3236 if (!vma || (vma->vm_flags & device_vma))
3238 start = min(end, vma->vm_end);
3239 } while (start < end);
3241 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL,
3246 * svm_range_best_prefetch_location - decide the best prefetch location
3247 * @prange: svm range structure
3250 * If range map to single GPU, the best prefetch location is prefetch_loc, which
3251 * can be CPU or GPU.
3253 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on
3254 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise
3255 * the best prefetch location is always CPU, because GPU can not have coherent
3256 * mapping VRAM of other GPUs even with large-BAR PCIe connection.
3259 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is
3260 * prefetch_loc, other GPU access will generate vm fault and trigger migration.
3262 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same
3263 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best
3264 * prefetch location is always CPU.
3266 * Context: Process context
3269 * 0 for CPU or GPU id
3272 svm_range_best_prefetch_location(struct svm_range *prange)
3274 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE);
3275 uint32_t best_loc = prange->prefetch_loc;
3276 struct kfd_process_device *pdd;
3277 struct kfd_node *bo_node;
3278 struct kfd_process *p;
3281 p = container_of(prange->svms, struct kfd_process, svms);
3283 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED)
3286 bo_node = svm_range_get_node_by_id(prange, best_loc);
3288 WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc);
3293 if (bo_node->adev->gmc.is_app_apu) {
3298 if (p->xnack_enabled)
3299 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE);
3301 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip,
3304 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) {
3305 pdd = kfd_process_device_from_gpuidx(p, gpuidx);
3307 pr_debug("failed to get device by idx 0x%x\n", gpuidx);
3311 if (pdd->dev->adev == bo_node->adev)
3314 if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) {
3321 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n",
3322 p->xnack_enabled, &p->svms, prange->start, prange->last,
3328 /* svm_range_trigger_migration - start page migration if prefetch loc changed
3329 * @mm: current process mm_struct
3330 * @prange: svm range structure
3331 * @migrated: output, true if migration is triggered
3333 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range
3335 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range
3338 * If GPU vm fault retry is not enabled, migration interact with MMU notifier
3340 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict
3341 * stops all queues, schedule restore work
3342 * 2. svm_range_restore_work wait for migration is done by
3343 * a. svm_range_validate_vram takes prange->migrate_mutex
3344 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns
3345 * 3. restore work update mappings of GPU, resume all queues.
3347 * Context: Process context
3350 * 0 - OK, otherwise - error code of migration
3353 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
3360 best_loc = svm_range_best_prefetch_location(prange);
3362 if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3363 best_loc == prange->actual_loc)
3367 r = svm_migrate_vram_to_ram(prange, mm,
3368 KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
3373 r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH);
3379 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence)
3384 if (dma_fence_is_signaled(&fence->base))
3387 if (fence->svm_bo) {
3388 WRITE_ONCE(fence->svm_bo->evicting, 1);
3389 schedule_work(&fence->svm_bo->eviction_work);
3395 static void svm_range_evict_svm_bo_worker(struct work_struct *work)
3397 struct svm_range_bo *svm_bo;
3398 struct mm_struct *mm;
3401 svm_bo = container_of(work, struct svm_range_bo, eviction_work);
3402 if (!svm_bo_ref_unless_zero(svm_bo))
3403 return; /* svm_bo was freed while eviction was pending */
3405 if (mmget_not_zero(svm_bo->eviction_fence->mm)) {
3406 mm = svm_bo->eviction_fence->mm;
3408 svm_range_bo_unref(svm_bo);
3413 spin_lock(&svm_bo->list_lock);
3414 while (!list_empty(&svm_bo->range_list) && !r) {
3415 struct svm_range *prange =
3416 list_first_entry(&svm_bo->range_list,
3417 struct svm_range, svm_bo_list);
3420 list_del_init(&prange->svm_bo_list);
3421 spin_unlock(&svm_bo->list_lock);
3423 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
3424 prange->start, prange->last);
3426 mutex_lock(&prange->migrate_mutex);
3428 r = svm_migrate_vram_to_ram(prange, mm,
3429 KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL);
3430 } while (!r && prange->actual_loc && --retries);
3432 if (!r && prange->actual_loc)
3433 pr_info_once("Migration failed during eviction");
3435 if (!prange->actual_loc) {
3436 mutex_lock(&prange->lock);
3437 prange->svm_bo = NULL;
3438 mutex_unlock(&prange->lock);
3440 mutex_unlock(&prange->migrate_mutex);
3442 spin_lock(&svm_bo->list_lock);
3444 spin_unlock(&svm_bo->list_lock);
3445 mmap_read_unlock(mm);
3448 dma_fence_signal(&svm_bo->eviction_fence->base);
3450 /* This is the last reference to svm_bo, after svm_range_vram_node_free
3451 * has been called in svm_migrate_vram_to_ram
3453 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n");
3454 svm_range_bo_unref(svm_bo);
3458 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
3459 uint64_t start, uint64_t size, uint32_t nattr,
3460 struct kfd_ioctl_svm_attribute *attrs)
3462 struct amdkfd_process_info *process_info = p->kgd_process_info;
3463 struct list_head update_list;
3464 struct list_head insert_list;
3465 struct list_head remove_list;
3466 struct svm_range_list *svms;
3467 struct svm_range *prange;
3468 struct svm_range *next;
3469 bool update_mapping = false;
3473 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n",
3474 p->pasid, &p->svms, start, start + size - 1, size);
3476 r = svm_range_check_attr(p, nattr, attrs);
3482 mutex_lock(&process_info->lock);
3484 svm_range_list_lock_and_flush_work(svms, mm);
3486 r = svm_range_is_valid(p, start, size);
3488 pr_debug("invalid range r=%d\n", r);
3489 mmap_write_unlock(mm);
3493 mutex_lock(&svms->lock);
3495 /* Add new range and split existing ranges as needed */
3496 r = svm_range_add(p, start, size, nattr, attrs, &update_list,
3497 &insert_list, &remove_list);
3499 mutex_unlock(&svms->lock);
3500 mmap_write_unlock(mm);
3503 /* Apply changes as a transaction */
3504 list_for_each_entry_safe(prange, next, &insert_list, list) {
3505 svm_range_add_to_svms(prange);
3506 svm_range_add_notifier_locked(mm, prange);
3508 list_for_each_entry(prange, &update_list, update_list) {
3509 svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping);
3510 /* TODO: unmap ranges from GPU that lost access */
3512 list_for_each_entry_safe(prange, next, &remove_list, update_list) {
3513 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n",
3514 prange->svms, prange, prange->start,
3516 svm_range_unlink(prange);
3517 svm_range_remove_notifier(prange);
3518 svm_range_free(prange, false);
3521 mmap_write_downgrade(mm);
3522 /* Trigger migrations and revalidate and map to GPUs as needed. If
3523 * this fails we may be left with partially completed actions. There
3524 * is no clean way of rolling back to the previous state in such a
3525 * case because the rollback wouldn't be guaranteed to work either.
3527 list_for_each_entry(prange, &update_list, update_list) {
3530 mutex_lock(&prange->migrate_mutex);
3532 r = svm_range_trigger_migration(mm, prange, &migrated);
3534 goto out_unlock_range;
3536 if (migrated && (!p->xnack_enabled ||
3537 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
3538 prange->mapped_to_gpu) {
3539 pr_debug("restore_work will update mappings of GPUs\n");
3540 mutex_unlock(&prange->migrate_mutex);
3544 if (!migrated && !update_mapping) {
3545 mutex_unlock(&prange->migrate_mutex);
3549 flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
3551 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
3552 true, true, flush_tlb);
3554 pr_debug("failed %d to map svm range\n", r);
3557 mutex_unlock(&prange->migrate_mutex);
3562 svm_range_debug_dump(svms);
3564 mutex_unlock(&svms->lock);
3565 mmap_read_unlock(mm);
3567 mutex_unlock(&process_info->lock);
3569 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
3570 &p->svms, start, start + size - 1, r);
3576 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm,
3577 uint64_t start, uint64_t size, uint32_t nattr,
3578 struct kfd_ioctl_svm_attribute *attrs)
3580 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE);
3581 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE);
3582 bool get_preferred_loc = false;
3583 bool get_prefetch_loc = false;
3584 bool get_granularity = false;
3585 bool get_accessible = false;
3586 bool get_flags = false;
3587 uint64_t last = start + size - 1UL;
3588 uint8_t granularity = 0xff;
3589 struct interval_tree_node *node;
3590 struct svm_range_list *svms;
3591 struct svm_range *prange;
3592 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3593 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3594 uint32_t flags_and = 0xffffffff;
3595 uint32_t flags_or = 0;
3600 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start,
3601 start + size - 1, nattr);
3603 /* Flush pending deferred work to avoid racing with deferred actions from
3604 * previous memory map changes (e.g. munmap). Concurrent memory map changes
3605 * can still race with get_attr because we don't hold the mmap lock. But that
3606 * would be a race condition in the application anyway, and undefined
3607 * behaviour is acceptable in that case.
3609 flush_work(&p->svms.deferred_list_work);
3612 r = svm_range_is_valid(p, start, size);
3613 mmap_read_unlock(mm);
3615 pr_debug("invalid range r=%d\n", r);
3619 for (i = 0; i < nattr; i++) {
3620 switch (attrs[i].type) {
3621 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3622 get_preferred_loc = true;
3624 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3625 get_prefetch_loc = true;
3627 case KFD_IOCTL_SVM_ATTR_ACCESS:
3628 get_accessible = true;
3630 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3631 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3634 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3635 get_granularity = true;
3637 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE:
3638 case KFD_IOCTL_SVM_ATTR_NO_ACCESS:
3641 pr_debug("get invalid attr type 0x%x\n", attrs[i].type);
3648 mutex_lock(&svms->lock);
3650 node = interval_tree_iter_first(&svms->objects, start, last);
3652 pr_debug("range attrs not found return default values\n");
3653 svm_range_set_default_attributes(&location, &prefetch_loc,
3654 &granularity, &flags_and);
3655 flags_or = flags_and;
3656 if (p->xnack_enabled)
3657 bitmap_copy(bitmap_access, svms->bitmap_supported,
3660 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE);
3661 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE);
3664 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE);
3665 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE);
3668 struct interval_tree_node *next;
3670 prange = container_of(node, struct svm_range, it_node);
3671 next = interval_tree_iter_next(node, start, last);
3673 if (get_preferred_loc) {
3674 if (prange->preferred_loc ==
3675 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3676 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3677 location != prange->preferred_loc)) {
3678 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3679 get_preferred_loc = false;
3681 location = prange->preferred_loc;
3684 if (get_prefetch_loc) {
3685 if (prange->prefetch_loc ==
3686 KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
3687 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED &&
3688 prefetch_loc != prange->prefetch_loc)) {
3689 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED;
3690 get_prefetch_loc = false;
3692 prefetch_loc = prange->prefetch_loc;
3695 if (get_accessible) {
3696 bitmap_and(bitmap_access, bitmap_access,
3697 prange->bitmap_access, MAX_GPU_INSTANCE);
3698 bitmap_and(bitmap_aip, bitmap_aip,
3699 prange->bitmap_aip, MAX_GPU_INSTANCE);
3702 flags_and &= prange->flags;
3703 flags_or |= prange->flags;
3706 if (get_granularity && prange->granularity < granularity)
3707 granularity = prange->granularity;
3712 mutex_unlock(&svms->lock);
3714 for (i = 0; i < nattr; i++) {
3715 switch (attrs[i].type) {
3716 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC:
3717 attrs[i].value = location;
3719 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3720 attrs[i].value = prefetch_loc;
3722 case KFD_IOCTL_SVM_ATTR_ACCESS:
3723 gpuidx = kfd_process_gpuidx_from_gpuid(p,
3726 pr_debug("invalid gpuid %x\n", attrs[i].value);
3729 if (test_bit(gpuidx, bitmap_access))
3730 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS;
3731 else if (test_bit(gpuidx, bitmap_aip))
3733 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE;
3735 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS;
3737 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3738 attrs[i].value = flags_and;
3740 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS:
3741 attrs[i].value = ~flags_or;
3743 case KFD_IOCTL_SVM_ATTR_GRANULARITY:
3744 attrs[i].value = (uint32_t)granularity;
3752 int kfd_criu_resume_svm(struct kfd_process *p)
3754 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL;
3755 int nattr_common = 4, nattr_accessibility = 1;
3756 struct criu_svm_metadata *criu_svm_md = NULL;
3757 struct svm_range_list *svms = &p->svms;
3758 struct criu_svm_metadata *next = NULL;
3759 uint32_t set_flags = 0xffffffff;
3760 int i, j, num_attrs, ret = 0;
3761 uint64_t set_attr_size;
3762 struct mm_struct *mm;
3764 if (list_empty(&svms->criu_svm_metadata_list)) {
3765 pr_debug("No SVM data from CRIU restore stage 2\n");
3769 mm = get_task_mm(p->lead_thread);
3771 pr_err("failed to get mm for the target process\n");
3775 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds);
3778 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) {
3779 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n",
3780 i, criu_svm_md->data.start_addr, criu_svm_md->data.size);
3782 for (j = 0; j < num_attrs; j++) {
3783 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n",
3784 i, j, criu_svm_md->data.attrs[j].type,
3785 i, j, criu_svm_md->data.attrs[j].value);
3786 switch (criu_svm_md->data.attrs[j].type) {
3787 /* During Checkpoint operation, the query for
3788 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might
3789 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were
3790 * not used by the range which was checkpointed. Care
3791 * must be taken to not restore with an invalid value
3792 * otherwise the gpuidx value will be invalid and
3793 * set_attr would eventually fail so just replace those
3794 * with another dummy attribute such as
3795 * KFD_IOCTL_SVM_ATTR_SET_FLAGS.
3797 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC:
3798 if (criu_svm_md->data.attrs[j].value ==
3799 KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
3800 criu_svm_md->data.attrs[j].type =
3801 KFD_IOCTL_SVM_ATTR_SET_FLAGS;
3802 criu_svm_md->data.attrs[j].value = 0;
3805 case KFD_IOCTL_SVM_ATTR_SET_FLAGS:
3806 set_flags = criu_svm_md->data.attrs[j].value;
3813 /* CLR_FLAGS is not available via get_attr during checkpoint but
3814 * it needs to be inserted before restoring the ranges so
3815 * allocate extra space for it before calling set_attr
3817 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3819 set_attr_new = krealloc(set_attr, set_attr_size,
3821 if (!set_attr_new) {
3825 set_attr = set_attr_new;
3827 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs *
3828 sizeof(struct kfd_ioctl_svm_attribute));
3829 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS;
3830 set_attr[num_attrs].value = ~set_flags;
3832 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr,
3833 criu_svm_md->data.size, num_attrs + 1,
3836 pr_err("CRIU: failed to set range attributes\n");
3844 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) {
3845 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n",
3846 criu_svm_md->data.start_addr);
3855 int kfd_criu_restore_svm(struct kfd_process *p,
3856 uint8_t __user *user_priv_ptr,
3857 uint64_t *priv_data_offset,
3858 uint64_t max_priv_data_size)
3860 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size;
3861 int nattr_common = 4, nattr_accessibility = 1;
3862 struct criu_svm_metadata *criu_svm_md = NULL;
3863 struct svm_range_list *svms = &p->svms;
3864 uint32_t num_devices;
3867 num_devices = p->n_pdds;
3868 /* Handle one SVM range object at a time, also the number of gpus are
3869 * assumed to be same on the restore node, checking must be done while
3870 * evaluating the topology earlier
3873 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) *
3874 (nattr_common + nattr_accessibility * num_devices);
3875 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size;
3877 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3880 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL);
3882 pr_err("failed to allocate memory to store svm metadata\n");
3885 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) {
3890 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset,
3891 svm_priv_data_size);
3896 *priv_data_offset += svm_priv_data_size;
3898 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list);
3908 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges,
3909 uint64_t *svm_priv_data_size)
3911 uint64_t total_size, accessibility_size, common_attr_size;
3912 int nattr_common = 4, nattr_accessibility = 1;
3913 int num_devices = p->n_pdds;
3914 struct svm_range_list *svms;
3915 struct svm_range *prange;
3918 *svm_priv_data_size = 0;
3924 mutex_lock(&svms->lock);
3925 list_for_each_entry(prange, &svms->list, list) {
3926 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n",
3927 prange, prange->start, prange->npages,
3928 prange->start + prange->npages - 1);
3931 mutex_unlock(&svms->lock);
3933 *num_svm_ranges = count;
3934 /* Only the accessbility attributes need to be queried for all the gpus
3935 * individually, remaining ones are spanned across the entire process
3936 * regardless of the various gpu nodes. Of the remaining attributes,
3937 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved.
3939 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC
3940 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC
3941 * KFD_IOCTL_SVM_ATTR_SET_FLAGS
3942 * KFD_IOCTL_SVM_ATTR_GRANULARITY
3944 * ** ACCESSBILITY ATTRIBUTES **
3945 * (Considered as one, type is altered during query, value is gpuid)
3946 * KFD_IOCTL_SVM_ATTR_ACCESS
3947 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE
3948 * KFD_IOCTL_SVM_ATTR_NO_ACCESS
3950 if (*num_svm_ranges > 0) {
3951 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3953 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) *
3954 nattr_accessibility * num_devices;
3956 total_size = sizeof(struct kfd_criu_svm_range_priv_data) +
3957 common_attr_size + accessibility_size;
3959 *svm_priv_data_size = *num_svm_ranges * total_size;
3962 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges,
3963 *svm_priv_data_size);
3967 int kfd_criu_checkpoint_svm(struct kfd_process *p,
3968 uint8_t __user *user_priv_data,
3969 uint64_t *priv_data_offset)
3971 struct kfd_criu_svm_range_priv_data *svm_priv = NULL;
3972 struct kfd_ioctl_svm_attribute *query_attr = NULL;
3973 uint64_t svm_priv_data_size, query_attr_size = 0;
3974 int index, nattr_common = 4, ret = 0;
3975 struct svm_range_list *svms;
3976 int num_devices = p->n_pdds;
3977 struct svm_range *prange;
3978 struct mm_struct *mm;
3984 mm = get_task_mm(p->lead_thread);
3986 pr_err("failed to get mm for the target process\n");
3990 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) *
3991 (nattr_common + num_devices);
3993 query_attr = kzalloc(query_attr_size, GFP_KERNEL);
3999 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC;
4000 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC;
4001 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS;
4002 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY;
4004 for (index = 0; index < num_devices; index++) {
4005 struct kfd_process_device *pdd = p->pdds[index];
4007 query_attr[index + nattr_common].type =
4008 KFD_IOCTL_SVM_ATTR_ACCESS;
4009 query_attr[index + nattr_common].value = pdd->user_gpu_id;
4012 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size;
4014 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL);
4021 list_for_each_entry(prange, &svms->list, list) {
4023 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE;
4024 svm_priv->start_addr = prange->start;
4025 svm_priv->size = prange->npages;
4026 memcpy(&svm_priv->attrs, query_attr, query_attr_size);
4027 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n",
4028 prange, prange->start, prange->npages,
4029 prange->start + prange->npages - 1,
4030 prange->npages * PAGE_SIZE);
4032 ret = svm_range_get_attr(p, mm, svm_priv->start_addr,
4034 (nattr_common + num_devices),
4037 pr_err("CRIU: failed to obtain range attributes\n");
4041 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv,
4042 svm_priv_data_size)) {
4043 pr_err("Failed to copy svm priv to user\n");
4048 *priv_data_offset += svm_priv_data_size;
4063 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start,
4064 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs)
4066 struct mm_struct *mm = current->mm;
4069 start >>= PAGE_SHIFT;
4070 size >>= PAGE_SHIFT;
4073 case KFD_IOCTL_SVM_OP_SET_ATTR:
4074 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs);
4076 case KFD_IOCTL_SVM_OP_GET_ATTR:
4077 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs);