2 * Copyright 2014-2018 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 #include <linux/dma-buf.h>
23 #include <linux/list.h>
24 #include <linux/pagemap.h>
25 #include <linux/sched/mm.h>
26 #include <linux/sched/task.h>
28 #include "amdgpu_object.h"
29 #include "amdgpu_gem.h"
30 #include "amdgpu_vm.h"
31 #include "amdgpu_amdkfd.h"
32 #include "amdgpu_dma_buf.h"
33 #include <uapi/linux/kfd_ioctl.h>
34 #include "amdgpu_xgmi.h"
36 /* BO flag to indicate a KFD userptr BO */
37 #define AMDGPU_AMDKFD_USERPTR_BO (1ULL << 63)
39 /* Userptr restore delay, just long enough to allow consecutive VM
40 * changes to accumulate
42 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
44 /* Impose limit on how much memory KFD can use */
46 uint64_t max_system_mem_limit;
47 uint64_t max_ttm_mem_limit;
48 int64_t system_mem_used;
50 spinlock_t mem_limit_lock;
53 /* Struct used for amdgpu_amdkfd_bo_validate */
54 struct amdgpu_vm_parser {
59 static const char * const domain_bit_to_string[] = {
68 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
70 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
73 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
75 return (struct amdgpu_device *)kgd;
78 static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm,
81 struct kfd_bo_va_list *entry;
83 list_for_each_entry(entry, &mem->bo_va_list, bo_list)
84 if (entry->bo_va->base.vm == avm)
90 /* Set memory usage limits. Current, limits are
91 * System (TTM + userptr) memory - 15/16th System RAM
92 * TTM memory - 3/8th System RAM
94 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
100 mem = si.freeram - si.freehigh;
103 spin_lock_init(&kfd_mem_limit.mem_limit_lock);
104 kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
105 kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3);
106 pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
107 (kfd_mem_limit.max_system_mem_limit >> 20),
108 (kfd_mem_limit.max_ttm_mem_limit >> 20));
111 /* Estimate page table size needed to represent a given memory size
113 * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
114 * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
115 * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
116 * for 2MB pages for TLB efficiency. However, small allocations and
117 * fragmented system memory still need some 4KB pages. We choose a
118 * compromise that should work in most cases without reserving too
119 * much memory for page tables unnecessarily (factor 16K, >> 14).
121 #define ESTIMATE_PT_SIZE(mem_size) ((mem_size) >> 14)
123 static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
124 uint64_t size, u32 domain, bool sg)
126 uint64_t reserved_for_pt =
127 ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
128 size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed;
131 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
132 sizeof(struct amdgpu_bo));
135 if (domain == AMDGPU_GEM_DOMAIN_GTT) {
137 system_mem_needed = acc_size + size;
138 ttm_mem_needed = acc_size + size;
139 } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
141 system_mem_needed = acc_size + size;
142 ttm_mem_needed = acc_size;
145 system_mem_needed = acc_size;
146 ttm_mem_needed = acc_size;
147 if (domain == AMDGPU_GEM_DOMAIN_VRAM)
151 spin_lock(&kfd_mem_limit.mem_limit_lock);
153 if (kfd_mem_limit.system_mem_used + system_mem_needed >
154 kfd_mem_limit.max_system_mem_limit)
155 pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
157 if ((kfd_mem_limit.system_mem_used + system_mem_needed >
158 kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
159 (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
160 kfd_mem_limit.max_ttm_mem_limit) ||
161 (adev->kfd.vram_used + vram_needed >
162 adev->gmc.real_vram_size - reserved_for_pt)) {
165 kfd_mem_limit.system_mem_used += system_mem_needed;
166 kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
167 adev->kfd.vram_used += vram_needed;
170 spin_unlock(&kfd_mem_limit.mem_limit_lock);
174 static void unreserve_mem_limit(struct amdgpu_device *adev,
175 uint64_t size, u32 domain, bool sg)
179 acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
180 sizeof(struct amdgpu_bo));
182 spin_lock(&kfd_mem_limit.mem_limit_lock);
183 if (domain == AMDGPU_GEM_DOMAIN_GTT) {
184 kfd_mem_limit.system_mem_used -= (acc_size + size);
185 kfd_mem_limit.ttm_mem_used -= (acc_size + size);
186 } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) {
187 kfd_mem_limit.system_mem_used -= (acc_size + size);
188 kfd_mem_limit.ttm_mem_used -= acc_size;
190 kfd_mem_limit.system_mem_used -= acc_size;
191 kfd_mem_limit.ttm_mem_used -= acc_size;
192 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
193 adev->kfd.vram_used -= size;
194 WARN_ONCE(adev->kfd.vram_used < 0,
195 "kfd VRAM memory accounting unbalanced");
198 WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
199 "kfd system memory accounting unbalanced");
200 WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
201 "kfd TTM memory accounting unbalanced");
203 spin_unlock(&kfd_mem_limit.mem_limit_lock);
206 void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
208 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
209 u32 domain = bo->preferred_domains;
210 bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU);
212 if (bo->flags & AMDGPU_AMDKFD_USERPTR_BO) {
213 domain = AMDGPU_GEM_DOMAIN_CPU;
217 unreserve_mem_limit(adev, amdgpu_bo_size(bo), domain, sg);
221 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
222 * reservation object.
224 * @bo: [IN] Remove eviction fence(s) from this BO
225 * @ef: [IN] This eviction fence is removed if it
226 * is present in the shared list.
228 * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
230 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
231 struct amdgpu_amdkfd_fence *ef)
233 struct dma_resv *resv = bo->tbo.base.resv;
234 struct dma_resv_list *old, *new;
235 unsigned int i, j, k;
240 old = dma_resv_get_list(resv);
244 new = kmalloc(struct_size(new, shared, old->shared_max), GFP_KERNEL);
248 /* Go through all the shared fences in the resevation object and sort
249 * the interesting ones to the end of the list.
251 for (i = 0, j = old->shared_count, k = 0; i < old->shared_count; ++i) {
254 f = rcu_dereference_protected(old->shared[i],
255 dma_resv_held(resv));
257 if (f->context == ef->base.context)
258 RCU_INIT_POINTER(new->shared[--j], f);
260 RCU_INIT_POINTER(new->shared[k++], f);
262 new->shared_max = old->shared_max;
263 new->shared_count = k;
265 /* Install the new fence list, seqcount provides the barriers */
266 write_seqcount_begin(&resv->seq);
267 RCU_INIT_POINTER(resv->fence, new);
268 write_seqcount_end(&resv->seq);
270 /* Drop the references to the removed fences or move them to ef_list */
271 for (i = j, k = 0; i < old->shared_count; ++i) {
274 f = rcu_dereference_protected(new->shared[i],
275 dma_resv_held(resv));
283 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
285 struct amdgpu_bo *root = bo;
286 struct amdgpu_vm_bo_base *vm_bo;
287 struct amdgpu_vm *vm;
288 struct amdkfd_process_info *info;
289 struct amdgpu_amdkfd_fence *ef;
292 /* we can always get vm_bo from root PD bo.*/
304 info = vm->process_info;
305 if (!info || !info->eviction_fence)
308 ef = container_of(dma_fence_get(&info->eviction_fence->base),
309 struct amdgpu_amdkfd_fence, base);
311 BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
312 ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
313 dma_resv_unlock(bo->tbo.base.resv);
315 dma_fence_put(&ef->base);
319 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
322 struct ttm_operation_ctx ctx = { false, false };
325 if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
326 "Called with userptr BO"))
329 amdgpu_bo_placement_from_domain(bo, domain);
331 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
335 amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
341 static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo)
343 struct amdgpu_vm_parser *p = param;
345 return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait);
348 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
350 * Page directories are not updated here because huge page handling
351 * during page table updates can invalidate page directory entries
352 * again. Page directories are only updated after updating page
355 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
357 struct amdgpu_bo *pd = vm->root.base.bo;
358 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
359 struct amdgpu_vm_parser param;
362 param.domain = AMDGPU_GEM_DOMAIN_VRAM;
365 ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate,
368 pr_err("failed to validate PT BOs\n");
372 ret = amdgpu_amdkfd_validate(¶m, pd);
374 pr_err("failed to validate PD\n");
378 vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
380 if (vm->use_cpu_for_update) {
381 ret = amdgpu_bo_kmap(pd, NULL);
383 pr_err("failed to kmap PD, ret=%d\n", ret);
391 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
393 struct amdgpu_bo *pd = vm->root.base.bo;
394 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
397 ret = amdgpu_vm_update_pdes(adev, vm, false);
401 return amdgpu_sync_fence(sync, vm->last_update);
404 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
406 struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
407 bool coherent = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
408 bool uncached = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
409 uint32_t mapping_flags;
413 mapping_flags = AMDGPU_VM_PAGE_READABLE;
414 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
415 mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
416 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
417 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
419 switch (adev->asic_type) {
421 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
423 mapping_flags |= coherent ?
424 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
426 mapping_flags |= AMDGPU_VM_MTYPE_UC;
428 mapping_flags |= coherent ?
429 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
433 if (coherent && uncached) {
434 if (adev->gmc.xgmi.connected_to_cpu ||
435 !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM))
437 mapping_flags |= AMDGPU_VM_MTYPE_UC;
438 } else if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
439 if (bo_adev == adev) {
440 mapping_flags |= AMDGPU_VM_MTYPE_RW;
441 if (adev->gmc.xgmi.connected_to_cpu)
444 mapping_flags |= AMDGPU_VM_MTYPE_NC;
445 if (amdgpu_xgmi_same_hive(adev, bo_adev))
450 if (adev->gmc.xgmi.connected_to_cpu)
451 /* system memory uses NC on A+A */
452 mapping_flags |= AMDGPU_VM_MTYPE_NC;
454 mapping_flags |= coherent ?
455 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
459 mapping_flags |= coherent ?
460 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
463 pte_flags = amdgpu_gem_va_map_flags(adev, mapping_flags);
464 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
469 /* add_bo_to_vm - Add a BO to a VM
471 * Everything that needs to bo done only once when a BO is first added
472 * to a VM. It can later be mapped and unmapped many times without
473 * repeating these steps.
475 * 1. Allocate and initialize BO VA entry data structure
476 * 2. Add BO to the VM
477 * 3. Determine ASIC-specific PTE flags
478 * 4. Alloc page tables and directories if needed
479 * 4a. Validate new page tables and directories
481 static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem,
482 struct amdgpu_vm *vm, bool is_aql,
483 struct kfd_bo_va_list **p_bo_va_entry)
486 struct kfd_bo_va_list *bo_va_entry;
487 struct amdgpu_bo *bo = mem->bo;
488 uint64_t va = mem->va;
489 struct list_head *list_bo_va = &mem->bo_va_list;
490 unsigned long bo_size = bo->tbo.base.size;
493 pr_err("Invalid VA when adding BO to VM\n");
500 bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL);
504 pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
507 /* Add BO to VM internal data structures*/
508 bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, vm, bo);
509 if (!bo_va_entry->bo_va) {
511 pr_err("Failed to add BO object to VM. ret == %d\n",
516 bo_va_entry->va = va;
517 bo_va_entry->pte_flags = get_pte_flags(adev, mem);
518 bo_va_entry->kgd_dev = (void *)adev;
519 list_add(&bo_va_entry->bo_list, list_bo_va);
522 *p_bo_va_entry = bo_va_entry;
524 /* Allocate validate page tables if needed */
525 ret = vm_validate_pt_pd_bos(vm);
527 pr_err("validate_pt_pd_bos() failed\n");
534 amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va);
535 list_del(&bo_va_entry->bo_list);
541 static void remove_bo_from_vm(struct amdgpu_device *adev,
542 struct kfd_bo_va_list *entry, unsigned long size)
544 pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n",
546 entry->va + size, entry);
547 amdgpu_vm_bo_rmv(adev, entry->bo_va);
548 list_del(&entry->bo_list);
552 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
553 struct amdkfd_process_info *process_info,
556 struct ttm_validate_buffer *entry = &mem->validate_list;
557 struct amdgpu_bo *bo = mem->bo;
559 INIT_LIST_HEAD(&entry->head);
560 entry->num_shared = 1;
561 entry->bo = &bo->tbo;
562 mutex_lock(&process_info->lock);
564 list_add_tail(&entry->head, &process_info->userptr_valid_list);
566 list_add_tail(&entry->head, &process_info->kfd_bo_list);
567 mutex_unlock(&process_info->lock);
570 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
571 struct amdkfd_process_info *process_info)
573 struct ttm_validate_buffer *bo_list_entry;
575 bo_list_entry = &mem->validate_list;
576 mutex_lock(&process_info->lock);
577 list_del(&bo_list_entry->head);
578 mutex_unlock(&process_info->lock);
581 /* Initializes user pages. It registers the MMU notifier and validates
582 * the userptr BO in the GTT domain.
584 * The BO must already be on the userptr_valid_list. Otherwise an
585 * eviction and restore may happen that leaves the new BO unmapped
586 * with the user mode queues running.
588 * Takes the process_info->lock to protect against concurrent restore
591 * Returns 0 for success, negative errno for errors.
593 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr)
595 struct amdkfd_process_info *process_info = mem->process_info;
596 struct amdgpu_bo *bo = mem->bo;
597 struct ttm_operation_ctx ctx = { true, false };
600 mutex_lock(&process_info->lock);
602 ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
604 pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
608 ret = amdgpu_mn_register(bo, user_addr);
610 pr_err("%s: Failed to register MMU notifier: %d\n",
615 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
617 pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
621 ret = amdgpu_bo_reserve(bo, true);
623 pr_err("%s: Failed to reserve BO\n", __func__);
626 amdgpu_bo_placement_from_domain(bo, mem->domain);
627 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
629 pr_err("%s: failed to validate BO\n", __func__);
630 amdgpu_bo_unreserve(bo);
633 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
636 amdgpu_mn_unregister(bo);
638 mutex_unlock(&process_info->lock);
642 /* Reserving a BO and its page table BOs must happen atomically to
643 * avoid deadlocks. Some operations update multiple VMs at once. Track
644 * all the reservation info in a context structure. Optionally a sync
645 * object can track VM updates.
647 struct bo_vm_reservation_context {
648 struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
649 unsigned int n_vms; /* Number of VMs reserved */
650 struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries */
651 struct ww_acquire_ctx ticket; /* Reservation ticket */
652 struct list_head list, duplicates; /* BO lists */
653 struct amdgpu_sync *sync; /* Pointer to sync object */
654 bool reserved; /* Whether BOs are reserved */
658 BO_VM_NOT_MAPPED = 0, /* Match VMs where a BO is not mapped */
659 BO_VM_MAPPED, /* Match VMs where a BO is mapped */
660 BO_VM_ALL, /* Match all VMs a BO was added to */
664 * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
665 * @mem: KFD BO structure.
666 * @vm: the VM to reserve.
667 * @ctx: the struct that will be used in unreserve_bo_and_vms().
669 static int reserve_bo_and_vm(struct kgd_mem *mem,
670 struct amdgpu_vm *vm,
671 struct bo_vm_reservation_context *ctx)
673 struct amdgpu_bo *bo = mem->bo;
678 ctx->reserved = false;
680 ctx->sync = &mem->sync;
682 INIT_LIST_HEAD(&ctx->list);
683 INIT_LIST_HEAD(&ctx->duplicates);
685 ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
689 ctx->kfd_bo.priority = 0;
690 ctx->kfd_bo.tv.bo = &bo->tbo;
691 ctx->kfd_bo.tv.num_shared = 1;
692 list_add(&ctx->kfd_bo.tv.head, &ctx->list);
694 amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
696 ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
697 false, &ctx->duplicates);
699 pr_err("Failed to reserve buffers in ttm.\n");
705 ctx->reserved = true;
710 * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
711 * @mem: KFD BO structure.
712 * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
713 * is used. Otherwise, a single VM associated with the BO.
714 * @map_type: the mapping status that will be used to filter the VMs.
715 * @ctx: the struct that will be used in unreserve_bo_and_vms().
717 * Returns 0 for success, negative for failure.
719 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
720 struct amdgpu_vm *vm, enum bo_vm_match map_type,
721 struct bo_vm_reservation_context *ctx)
723 struct amdgpu_bo *bo = mem->bo;
724 struct kfd_bo_va_list *entry;
728 ctx->reserved = false;
731 ctx->sync = &mem->sync;
733 INIT_LIST_HEAD(&ctx->list);
734 INIT_LIST_HEAD(&ctx->duplicates);
736 list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
737 if ((vm && vm != entry->bo_va->base.vm) ||
738 (entry->is_mapped != map_type
739 && map_type != BO_VM_ALL))
745 if (ctx->n_vms != 0) {
746 ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
752 ctx->kfd_bo.priority = 0;
753 ctx->kfd_bo.tv.bo = &bo->tbo;
754 ctx->kfd_bo.tv.num_shared = 1;
755 list_add(&ctx->kfd_bo.tv.head, &ctx->list);
758 list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
759 if ((vm && vm != entry->bo_va->base.vm) ||
760 (entry->is_mapped != map_type
761 && map_type != BO_VM_ALL))
764 amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
769 ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
770 false, &ctx->duplicates);
772 pr_err("Failed to reserve buffers in ttm.\n");
778 ctx->reserved = true;
783 * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
784 * @ctx: Reservation context to unreserve
785 * @wait: Optionally wait for a sync object representing pending VM updates
786 * @intr: Whether the wait is interruptible
788 * Also frees any resources allocated in
789 * reserve_bo_and_(cond_)vm(s). Returns the status from
792 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
793 bool wait, bool intr)
798 ret = amdgpu_sync_wait(ctx->sync, intr);
801 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
806 ctx->reserved = false;
812 static int unmap_bo_from_gpuvm(struct amdgpu_device *adev,
813 struct kfd_bo_va_list *entry,
814 struct amdgpu_sync *sync)
816 struct amdgpu_bo_va *bo_va = entry->bo_va;
817 struct amdgpu_vm *vm = bo_va->base.vm;
819 amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
821 amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
823 amdgpu_sync_fence(sync, bo_va->last_pt_update);
828 static int update_gpuvm_pte(struct amdgpu_device *adev,
829 struct kfd_bo_va_list *entry,
830 struct amdgpu_sync *sync)
833 struct amdgpu_bo_va *bo_va = entry->bo_va;
835 /* Update the page tables */
836 ret = amdgpu_vm_bo_update(adev, bo_va, false);
838 pr_err("amdgpu_vm_bo_update failed\n");
842 return amdgpu_sync_fence(sync, bo_va->last_pt_update);
845 static int map_bo_to_gpuvm(struct amdgpu_device *adev,
846 struct kfd_bo_va_list *entry, struct amdgpu_sync *sync,
851 /* Set virtual address for the allocation */
852 ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0,
853 amdgpu_bo_size(entry->bo_va->base.bo),
856 pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
864 ret = update_gpuvm_pte(adev, entry, sync);
866 pr_err("update_gpuvm_pte() failed\n");
867 goto update_gpuvm_pte_failed;
872 update_gpuvm_pte_failed:
873 unmap_bo_from_gpuvm(adev, entry, sync);
877 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
879 struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
883 if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
887 sg->sgl->dma_address = addr;
888 sg->sgl->length = size;
889 #ifdef CONFIG_NEED_SG_DMA_LENGTH
890 sg->sgl->dma_length = size;
895 static int process_validate_vms(struct amdkfd_process_info *process_info)
897 struct amdgpu_vm *peer_vm;
900 list_for_each_entry(peer_vm, &process_info->vm_list_head,
902 ret = vm_validate_pt_pd_bos(peer_vm);
910 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
911 struct amdgpu_sync *sync)
913 struct amdgpu_vm *peer_vm;
916 list_for_each_entry(peer_vm, &process_info->vm_list_head,
918 struct amdgpu_bo *pd = peer_vm->root.base.bo;
920 ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
921 AMDGPU_SYNC_NE_OWNER,
922 AMDGPU_FENCE_OWNER_KFD);
930 static int process_update_pds(struct amdkfd_process_info *process_info,
931 struct amdgpu_sync *sync)
933 struct amdgpu_vm *peer_vm;
936 list_for_each_entry(peer_vm, &process_info->vm_list_head,
938 ret = vm_update_pds(peer_vm, sync);
946 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
947 struct dma_fence **ef)
949 struct amdkfd_process_info *info = NULL;
952 if (!*process_info) {
953 info = kzalloc(sizeof(*info), GFP_KERNEL);
957 mutex_init(&info->lock);
958 INIT_LIST_HEAD(&info->vm_list_head);
959 INIT_LIST_HEAD(&info->kfd_bo_list);
960 INIT_LIST_HEAD(&info->userptr_valid_list);
961 INIT_LIST_HEAD(&info->userptr_inval_list);
963 info->eviction_fence =
964 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
966 if (!info->eviction_fence) {
967 pr_err("Failed to create eviction fence\n");
969 goto create_evict_fence_fail;
972 info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
973 atomic_set(&info->evicted_bos, 0);
974 INIT_DELAYED_WORK(&info->restore_userptr_work,
975 amdgpu_amdkfd_restore_userptr_worker);
977 *process_info = info;
978 *ef = dma_fence_get(&info->eviction_fence->base);
981 vm->process_info = *process_info;
983 /* Validate page directory and attach eviction fence */
984 ret = amdgpu_bo_reserve(vm->root.base.bo, true);
986 goto reserve_pd_fail;
987 ret = vm_validate_pt_pd_bos(vm);
989 pr_err("validate_pt_pd_bos() failed\n");
990 goto validate_pd_fail;
992 ret = amdgpu_bo_sync_wait(vm->root.base.bo,
993 AMDGPU_FENCE_OWNER_KFD, false);
996 ret = dma_resv_reserve_shared(vm->root.base.bo->tbo.base.resv, 1);
998 goto reserve_shared_fail;
999 amdgpu_bo_fence(vm->root.base.bo,
1000 &vm->process_info->eviction_fence->base, true);
1001 amdgpu_bo_unreserve(vm->root.base.bo);
1003 /* Update process info */
1004 mutex_lock(&vm->process_info->lock);
1005 list_add_tail(&vm->vm_list_node,
1006 &(vm->process_info->vm_list_head));
1007 vm->process_info->n_vms++;
1008 mutex_unlock(&vm->process_info->lock);
1012 reserve_shared_fail:
1015 amdgpu_bo_unreserve(vm->root.base.bo);
1017 vm->process_info = NULL;
1019 /* Two fence references: one in info and one in *ef */
1020 dma_fence_put(&info->eviction_fence->base);
1023 *process_info = NULL;
1025 create_evict_fence_fail:
1026 mutex_destroy(&info->lock);
1032 int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, u32 pasid,
1033 void **vm, void **process_info,
1034 struct dma_fence **ef)
1036 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1037 struct amdgpu_vm *new_vm;
1040 new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL);
1044 /* Initialize AMDGPU part of the VM */
1045 ret = amdgpu_vm_init(adev, new_vm, AMDGPU_VM_CONTEXT_COMPUTE, pasid);
1047 pr_err("Failed init vm ret %d\n", ret);
1048 goto amdgpu_vm_init_fail;
1051 /* Initialize KFD part of the VM and process info */
1052 ret = init_kfd_vm(new_vm, process_info, ef);
1054 goto init_kfd_vm_fail;
1056 *vm = (void *) new_vm;
1061 amdgpu_vm_fini(adev, new_vm);
1062 amdgpu_vm_init_fail:
1067 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd,
1068 struct file *filp, u32 pasid,
1069 void **vm, void **process_info,
1070 struct dma_fence **ef)
1072 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1073 struct drm_file *drm_priv = filp->private_data;
1074 struct amdgpu_fpriv *drv_priv = drm_priv->driver_priv;
1075 struct amdgpu_vm *avm = &drv_priv->vm;
1078 /* Already a compute VM? */
1079 if (avm->process_info)
1082 /* Convert VM into a compute VM */
1083 ret = amdgpu_vm_make_compute(adev, avm, pasid);
1087 /* Initialize KFD part of the VM and process info */
1088 ret = init_kfd_vm(avm, process_info, ef);
1097 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1098 struct amdgpu_vm *vm)
1100 struct amdkfd_process_info *process_info = vm->process_info;
1101 struct amdgpu_bo *pd = vm->root.base.bo;
1106 /* Release eviction fence from PD */
1107 amdgpu_bo_reserve(pd, false);
1108 amdgpu_bo_fence(pd, NULL, false);
1109 amdgpu_bo_unreserve(pd);
1111 /* Update process info */
1112 mutex_lock(&process_info->lock);
1113 process_info->n_vms--;
1114 list_del(&vm->vm_list_node);
1115 mutex_unlock(&process_info->lock);
1117 vm->process_info = NULL;
1119 /* Release per-process resources when last compute VM is destroyed */
1120 if (!process_info->n_vms) {
1121 WARN_ON(!list_empty(&process_info->kfd_bo_list));
1122 WARN_ON(!list_empty(&process_info->userptr_valid_list));
1123 WARN_ON(!list_empty(&process_info->userptr_inval_list));
1125 dma_fence_put(&process_info->eviction_fence->base);
1126 cancel_delayed_work_sync(&process_info->restore_userptr_work);
1127 put_pid(process_info->pid);
1128 mutex_destroy(&process_info->lock);
1129 kfree(process_info);
1133 void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm)
1135 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1136 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1138 if (WARN_ON(!kgd || !vm))
1141 pr_debug("Destroying process vm %p\n", vm);
1143 /* Release the VM context */
1144 amdgpu_vm_fini(adev, avm);
1148 void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm)
1150 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1151 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1153 if (WARN_ON(!kgd || !vm))
1156 pr_debug("Releasing process vm %p\n", vm);
1158 /* The original pasid of amdgpu vm has already been
1159 * released during making a amdgpu vm to a compute vm
1160 * The current pasid is managed by kfd and will be
1161 * released on kfd process destroy. Set amdgpu pasid
1162 * to 0 to avoid duplicate release.
1164 amdgpu_vm_release_compute(adev, avm);
1167 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm)
1169 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1170 struct amdgpu_bo *pd = avm->root.base.bo;
1171 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1173 if (adev->asic_type < CHIP_VEGA10)
1174 return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1175 return avm->pd_phys_addr;
1178 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1179 struct kgd_dev *kgd, uint64_t va, uint64_t size,
1180 void *vm, struct kgd_mem **mem,
1181 uint64_t *offset, uint32_t flags)
1183 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1184 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1185 enum ttm_bo_type bo_type = ttm_bo_type_device;
1186 struct sg_table *sg = NULL;
1187 uint64_t user_addr = 0;
1188 struct amdgpu_bo *bo;
1189 struct drm_gem_object *gobj;
1190 u32 domain, alloc_domain;
1195 * Check on which domain to allocate BO
1197 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1198 domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1199 alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1200 alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1201 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED :
1202 AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
1203 } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1204 domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1206 } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1207 domain = AMDGPU_GEM_DOMAIN_GTT;
1208 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1210 if (!offset || !*offset)
1212 user_addr = untagged_addr(*offset);
1213 } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1214 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1215 domain = AMDGPU_GEM_DOMAIN_GTT;
1216 alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1217 bo_type = ttm_bo_type_sg;
1219 if (size > UINT_MAX)
1221 sg = create_doorbell_sg(*offset, size);
1228 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1233 INIT_LIST_HEAD(&(*mem)->bo_va_list);
1234 mutex_init(&(*mem)->lock);
1235 (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1237 /* Workaround for AQL queue wraparound bug. Map the same
1238 * memory twice. That means we only actually allocate half
1241 if ((*mem)->aql_queue)
1244 (*mem)->alloc_flags = flags;
1246 amdgpu_sync_create(&(*mem)->sync);
1248 ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, !!sg);
1250 pr_debug("Insufficient memory\n");
1251 goto err_reserve_limit;
1254 pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1255 va, size, domain_string(alloc_domain));
1257 ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
1258 bo_type, NULL, &gobj);
1260 pr_debug("Failed to create BO on domain %s. ret %d\n",
1261 domain_string(alloc_domain), ret);
1264 bo = gem_to_amdgpu_bo(gobj);
1265 if (bo_type == ttm_bo_type_sg) {
1267 bo->tbo.ttm->sg = sg;
1272 bo->flags |= AMDGPU_AMDKFD_USERPTR_BO;
1275 (*mem)->domain = domain;
1276 (*mem)->mapped_to_gpu_memory = 0;
1277 (*mem)->process_info = avm->process_info;
1278 add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1281 ret = init_user_pages(*mem, user_addr);
1283 goto allocate_init_user_pages_failed;
1287 *offset = amdgpu_bo_mmap_offset(bo);
1291 allocate_init_user_pages_failed:
1292 remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1293 amdgpu_bo_unref(&bo);
1294 /* Don't unreserve system mem limit twice */
1295 goto err_reserve_limit;
1297 unreserve_mem_limit(adev, size, alloc_domain, !!sg);
1299 mutex_destroy(&(*mem)->lock);
1309 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1310 struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *size)
1312 struct amdkfd_process_info *process_info = mem->process_info;
1313 unsigned long bo_size = mem->bo->tbo.base.size;
1314 struct kfd_bo_va_list *entry, *tmp;
1315 struct bo_vm_reservation_context ctx;
1316 struct ttm_validate_buffer *bo_list_entry;
1317 unsigned int mapped_to_gpu_memory;
1319 bool is_imported = false;
1321 mutex_lock(&mem->lock);
1322 mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1323 is_imported = mem->is_imported;
1324 mutex_unlock(&mem->lock);
1325 /* lock is not needed after this, since mem is unused and will
1329 if (mapped_to_gpu_memory > 0) {
1330 pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1335 /* Make sure restore workers don't access the BO any more */
1336 bo_list_entry = &mem->validate_list;
1337 mutex_lock(&process_info->lock);
1338 list_del(&bo_list_entry->head);
1339 mutex_unlock(&process_info->lock);
1341 /* No more MMU notifiers */
1342 amdgpu_mn_unregister(mem->bo);
1344 ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1348 /* The eviction fence should be removed by the last unmap.
1349 * TODO: Log an error condition if the bo still has the eviction fence
1352 amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1353 process_info->eviction_fence);
1354 pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1355 mem->va + bo_size * (1 + mem->aql_queue));
1357 /* Remove from VM internal data structures */
1358 list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list)
1359 remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev,
1362 ret = unreserve_bo_and_vms(&ctx, false, false);
1364 /* Free the sync object */
1365 amdgpu_sync_free(&mem->sync);
1367 /* If the SG is not NULL, it's one we created for a doorbell or mmio
1368 * remap BO. We need to free it.
1370 if (mem->bo->tbo.sg) {
1371 sg_free_table(mem->bo->tbo.sg);
1372 kfree(mem->bo->tbo.sg);
1375 /* Update the size of the BO being freed if it was allocated from
1376 * VRAM and is not imported.
1379 if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1387 drm_gem_object_put(&mem->bo->tbo.base);
1388 mutex_destroy(&mem->lock);
1394 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1395 struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1397 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1398 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1400 struct amdgpu_bo *bo;
1402 struct kfd_bo_va_list *entry;
1403 struct bo_vm_reservation_context ctx;
1404 struct kfd_bo_va_list *bo_va_entry = NULL;
1405 struct kfd_bo_va_list *bo_va_entry_aql = NULL;
1406 unsigned long bo_size;
1407 bool is_invalid_userptr = false;
1411 pr_err("Invalid BO when mapping memory to GPU\n");
1415 /* Make sure restore is not running concurrently. Since we
1416 * don't map invalid userptr BOs, we rely on the next restore
1417 * worker to do the mapping
1419 mutex_lock(&mem->process_info->lock);
1421 /* Lock mmap-sem. If we find an invalid userptr BO, we can be
1422 * sure that the MMU notifier is no longer running
1423 * concurrently and the queues are actually stopped
1425 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1426 mmap_write_lock(current->mm);
1427 is_invalid_userptr = atomic_read(&mem->invalid);
1428 mmap_write_unlock(current->mm);
1431 mutex_lock(&mem->lock);
1433 domain = mem->domain;
1434 bo_size = bo->tbo.base.size;
1436 pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1438 mem->va + bo_size * (1 + mem->aql_queue),
1439 vm, domain_string(domain));
1441 ret = reserve_bo_and_vm(mem, vm, &ctx);
1445 /* Userptr can be marked as "not invalid", but not actually be
1446 * validated yet (still in the system domain). In that case
1447 * the queues are still stopped and we can leave mapping for
1448 * the next restore worker
1450 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1451 bo->tbo.mem.mem_type == TTM_PL_SYSTEM)
1452 is_invalid_userptr = true;
1454 if (check_if_add_bo_to_vm(avm, mem)) {
1455 ret = add_bo_to_vm(adev, mem, avm, false,
1458 goto add_bo_to_vm_failed;
1459 if (mem->aql_queue) {
1460 ret = add_bo_to_vm(adev, mem, avm,
1461 true, &bo_va_entry_aql);
1463 goto add_bo_to_vm_failed_aql;
1466 ret = vm_validate_pt_pd_bos(avm);
1468 goto add_bo_to_vm_failed;
1471 if (mem->mapped_to_gpu_memory == 0 &&
1472 !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1473 /* Validate BO only once. The eviction fence gets added to BO
1474 * the first time it is mapped. Validate will wait for all
1475 * background evictions to complete.
1477 ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1479 pr_debug("Validate failed\n");
1480 goto map_bo_to_gpuvm_failed;
1484 list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1485 if (entry->bo_va->base.vm == vm && !entry->is_mapped) {
1486 pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1487 entry->va, entry->va + bo_size,
1490 ret = map_bo_to_gpuvm(adev, entry, ctx.sync,
1491 is_invalid_userptr);
1493 pr_err("Failed to map bo to gpuvm\n");
1494 goto map_bo_to_gpuvm_failed;
1497 ret = vm_update_pds(vm, ctx.sync);
1499 pr_err("Failed to update page directories\n");
1500 goto map_bo_to_gpuvm_failed;
1503 entry->is_mapped = true;
1504 mem->mapped_to_gpu_memory++;
1505 pr_debug("\t INC mapping count %d\n",
1506 mem->mapped_to_gpu_memory);
1510 if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
1512 &avm->process_info->eviction_fence->base,
1514 ret = unreserve_bo_and_vms(&ctx, false, false);
1518 map_bo_to_gpuvm_failed:
1519 if (bo_va_entry_aql)
1520 remove_bo_from_vm(adev, bo_va_entry_aql, bo_size);
1521 add_bo_to_vm_failed_aql:
1523 remove_bo_from_vm(adev, bo_va_entry, bo_size);
1524 add_bo_to_vm_failed:
1525 unreserve_bo_and_vms(&ctx, false, false);
1527 mutex_unlock(&mem->process_info->lock);
1528 mutex_unlock(&mem->lock);
1532 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1533 struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
1535 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1536 struct amdkfd_process_info *process_info =
1537 ((struct amdgpu_vm *)vm)->process_info;
1538 unsigned long bo_size = mem->bo->tbo.base.size;
1539 struct kfd_bo_va_list *entry;
1540 struct bo_vm_reservation_context ctx;
1543 mutex_lock(&mem->lock);
1545 ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx);
1548 /* If no VMs were reserved, it means the BO wasn't actually mapped */
1549 if (ctx.n_vms == 0) {
1554 ret = vm_validate_pt_pd_bos((struct amdgpu_vm *)vm);
1558 pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1560 mem->va + bo_size * (1 + mem->aql_queue),
1563 list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
1564 if (entry->bo_va->base.vm == vm && entry->is_mapped) {
1565 pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1567 entry->va + bo_size,
1570 ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync);
1572 entry->is_mapped = false;
1574 pr_err("failed to unmap VA 0x%llx\n",
1579 mem->mapped_to_gpu_memory--;
1580 pr_debug("\t DEC mapping count %d\n",
1581 mem->mapped_to_gpu_memory);
1585 /* If BO is unmapped from all VMs, unfence it. It can be evicted if
1588 if (mem->mapped_to_gpu_memory == 0 &&
1589 !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
1590 !mem->bo->tbo.pin_count)
1591 amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1592 process_info->eviction_fence);
1595 unreserve_bo_and_vms(&ctx, false, false);
1597 mutex_unlock(&mem->lock);
1601 int amdgpu_amdkfd_gpuvm_sync_memory(
1602 struct kgd_dev *kgd, struct kgd_mem *mem, bool intr)
1604 struct amdgpu_sync sync;
1607 amdgpu_sync_create(&sync);
1609 mutex_lock(&mem->lock);
1610 amdgpu_sync_clone(&mem->sync, &sync);
1611 mutex_unlock(&mem->lock);
1613 ret = amdgpu_sync_wait(&sync, intr);
1614 amdgpu_sync_free(&sync);
1618 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
1619 struct kgd_mem *mem, void **kptr, uint64_t *size)
1622 struct amdgpu_bo *bo = mem->bo;
1624 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1625 pr_err("userptr can't be mapped to kernel\n");
1629 /* delete kgd_mem from kfd_bo_list to avoid re-validating
1630 * this BO in BO's restoring after eviction.
1632 mutex_lock(&mem->process_info->lock);
1634 ret = amdgpu_bo_reserve(bo, true);
1636 pr_err("Failed to reserve bo. ret %d\n", ret);
1637 goto bo_reserve_failed;
1640 ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1642 pr_err("Failed to pin bo. ret %d\n", ret);
1646 ret = amdgpu_bo_kmap(bo, kptr);
1648 pr_err("Failed to map bo to kernel. ret %d\n", ret);
1652 amdgpu_amdkfd_remove_eviction_fence(
1653 bo, mem->process_info->eviction_fence);
1654 list_del_init(&mem->validate_list.head);
1657 *size = amdgpu_bo_size(bo);
1659 amdgpu_bo_unreserve(bo);
1661 mutex_unlock(&mem->process_info->lock);
1665 amdgpu_bo_unpin(bo);
1667 amdgpu_bo_unreserve(bo);
1669 mutex_unlock(&mem->process_info->lock);
1674 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
1675 struct kfd_vm_fault_info *mem)
1677 struct amdgpu_device *adev;
1679 adev = (struct amdgpu_device *)kgd;
1680 if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
1681 *mem = *adev->gmc.vm_fault_info;
1683 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1688 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,
1689 struct dma_buf *dma_buf,
1690 uint64_t va, void *vm,
1691 struct kgd_mem **mem, uint64_t *size,
1692 uint64_t *mmap_offset)
1694 struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
1695 struct drm_gem_object *obj;
1696 struct amdgpu_bo *bo;
1697 struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;
1699 if (dma_buf->ops != &amdgpu_dmabuf_ops)
1700 /* Can't handle non-graphics buffers */
1703 obj = dma_buf->priv;
1704 if (drm_to_adev(obj->dev) != adev)
1705 /* Can't handle buffers from other devices */
1708 bo = gem_to_amdgpu_bo(obj);
1709 if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
1710 AMDGPU_GEM_DOMAIN_GTT)))
1711 /* Only VRAM and GTT BOs are supported */
1714 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1719 *size = amdgpu_bo_size(bo);
1722 *mmap_offset = amdgpu_bo_mmap_offset(bo);
1724 INIT_LIST_HEAD(&(*mem)->bo_va_list);
1725 mutex_init(&(*mem)->lock);
1727 (*mem)->alloc_flags =
1728 ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
1729 KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
1730 | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
1731 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1733 drm_gem_object_get(&bo->tbo.base);
1736 (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
1737 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
1738 (*mem)->mapped_to_gpu_memory = 0;
1739 (*mem)->process_info = avm->process_info;
1740 add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
1741 amdgpu_sync_create(&(*mem)->sync);
1742 (*mem)->is_imported = true;
1747 /* Evict a userptr BO by stopping the queues if necessary
1749 * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
1750 * cannot do any memory allocations, and cannot take any locks that
1751 * are held elsewhere while allocating memory. Therefore this is as
1752 * simple as possible, using atomic counters.
1754 * It doesn't do anything to the BO itself. The real work happens in
1755 * restore, where we get updated page addresses. This function only
1756 * ensures that GPU access to the BO is stopped.
1758 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
1759 struct mm_struct *mm)
1761 struct amdkfd_process_info *process_info = mem->process_info;
1765 atomic_inc(&mem->invalid);
1766 evicted_bos = atomic_inc_return(&process_info->evicted_bos);
1767 if (evicted_bos == 1) {
1768 /* First eviction, stop the queues */
1769 r = kgd2kfd_quiesce_mm(mm);
1771 pr_err("Failed to quiesce KFD\n");
1772 schedule_delayed_work(&process_info->restore_userptr_work,
1773 msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
1779 /* Update invalid userptr BOs
1781 * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
1782 * userptr_inval_list and updates user pages for all BOs that have
1783 * been invalidated since their last update.
1785 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
1786 struct mm_struct *mm)
1788 struct kgd_mem *mem, *tmp_mem;
1789 struct amdgpu_bo *bo;
1790 struct ttm_operation_ctx ctx = { false, false };
1793 /* Move all invalidated BOs to the userptr_inval_list and
1794 * release their user pages by migration to the CPU domain
1796 list_for_each_entry_safe(mem, tmp_mem,
1797 &process_info->userptr_valid_list,
1798 validate_list.head) {
1799 if (!atomic_read(&mem->invalid))
1800 continue; /* BO is still valid */
1804 if (amdgpu_bo_reserve(bo, true))
1806 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
1807 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1808 amdgpu_bo_unreserve(bo);
1810 pr_err("%s: Failed to invalidate userptr BO\n",
1815 list_move_tail(&mem->validate_list.head,
1816 &process_info->userptr_inval_list);
1819 if (list_empty(&process_info->userptr_inval_list))
1820 return 0; /* All evicted userptr BOs were freed */
1822 /* Go through userptr_inval_list and update any invalid user_pages */
1823 list_for_each_entry(mem, &process_info->userptr_inval_list,
1824 validate_list.head) {
1825 invalid = atomic_read(&mem->invalid);
1827 /* BO hasn't been invalidated since the last
1828 * revalidation attempt. Keep its BO list.
1834 /* Get updated user pages */
1835 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
1837 pr_debug("%s: Failed to get user pages: %d\n",
1840 /* Return error -EBUSY or -ENOMEM, retry restore */
1845 * FIXME: Cannot ignore the return code, must hold
1848 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1850 /* Mark the BO as valid unless it was invalidated
1851 * again concurrently.
1853 if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
1860 /* Validate invalid userptr BOs
1862 * Validates BOs on the userptr_inval_list, and moves them back to the
1863 * userptr_valid_list. Also updates GPUVM page tables with new page
1864 * addresses and waits for the page table updates to complete.
1866 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
1868 struct amdgpu_bo_list_entry *pd_bo_list_entries;
1869 struct list_head resv_list, duplicates;
1870 struct ww_acquire_ctx ticket;
1871 struct amdgpu_sync sync;
1873 struct amdgpu_vm *peer_vm;
1874 struct kgd_mem *mem, *tmp_mem;
1875 struct amdgpu_bo *bo;
1876 struct ttm_operation_ctx ctx = { false, false };
1879 pd_bo_list_entries = kcalloc(process_info->n_vms,
1880 sizeof(struct amdgpu_bo_list_entry),
1882 if (!pd_bo_list_entries) {
1883 pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
1888 INIT_LIST_HEAD(&resv_list);
1889 INIT_LIST_HEAD(&duplicates);
1891 /* Get all the page directory BOs that need to be reserved */
1893 list_for_each_entry(peer_vm, &process_info->vm_list_head,
1895 amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
1896 &pd_bo_list_entries[i++]);
1897 /* Add the userptr_inval_list entries to resv_list */
1898 list_for_each_entry(mem, &process_info->userptr_inval_list,
1899 validate_list.head) {
1900 list_add_tail(&mem->resv_list.head, &resv_list);
1901 mem->resv_list.bo = mem->validate_list.bo;
1902 mem->resv_list.num_shared = mem->validate_list.num_shared;
1905 /* Reserve all BOs and page tables for validation */
1906 ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
1907 WARN(!list_empty(&duplicates), "Duplicates should be empty");
1911 amdgpu_sync_create(&sync);
1913 ret = process_validate_vms(process_info);
1917 /* Validate BOs and update GPUVM page tables */
1918 list_for_each_entry_safe(mem, tmp_mem,
1919 &process_info->userptr_inval_list,
1920 validate_list.head) {
1921 struct kfd_bo_va_list *bo_va_entry;
1925 /* Validate the BO if we got user pages */
1926 if (bo->tbo.ttm->pages[0]) {
1927 amdgpu_bo_placement_from_domain(bo, mem->domain);
1928 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1930 pr_err("%s: failed to validate BO\n", __func__);
1935 list_move_tail(&mem->validate_list.head,
1936 &process_info->userptr_valid_list);
1938 /* Update mapping. If the BO was not validated
1939 * (because we couldn't get user pages), this will
1940 * clear the page table entries, which will result in
1941 * VM faults if the GPU tries to access the invalid
1944 list_for_each_entry(bo_va_entry, &mem->bo_va_list, bo_list) {
1945 if (!bo_va_entry->is_mapped)
1948 ret = update_gpuvm_pte((struct amdgpu_device *)
1949 bo_va_entry->kgd_dev,
1950 bo_va_entry, &sync);
1952 pr_err("%s: update PTE failed\n", __func__);
1953 /* make sure this gets validated again */
1954 atomic_inc(&mem->invalid);
1960 /* Update page directories */
1961 ret = process_update_pds(process_info, &sync);
1964 ttm_eu_backoff_reservation(&ticket, &resv_list);
1965 amdgpu_sync_wait(&sync, false);
1966 amdgpu_sync_free(&sync);
1968 kfree(pd_bo_list_entries);
1974 /* Worker callback to restore evicted userptr BOs
1976 * Tries to update and validate all userptr BOs. If successful and no
1977 * concurrent evictions happened, the queues are restarted. Otherwise,
1978 * reschedule for another attempt later.
1980 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
1982 struct delayed_work *dwork = to_delayed_work(work);
1983 struct amdkfd_process_info *process_info =
1984 container_of(dwork, struct amdkfd_process_info,
1985 restore_userptr_work);
1986 struct task_struct *usertask;
1987 struct mm_struct *mm;
1990 evicted_bos = atomic_read(&process_info->evicted_bos);
1994 /* Reference task and mm in case of concurrent process termination */
1995 usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
1998 mm = get_task_mm(usertask);
2000 put_task_struct(usertask);
2004 mutex_lock(&process_info->lock);
2006 if (update_invalid_user_pages(process_info, mm))
2008 /* userptr_inval_list can be empty if all evicted userptr BOs
2009 * have been freed. In that case there is nothing to validate
2010 * and we can just restart the queues.
2012 if (!list_empty(&process_info->userptr_inval_list)) {
2013 if (atomic_read(&process_info->evicted_bos) != evicted_bos)
2014 goto unlock_out; /* Concurrent eviction, try again */
2016 if (validate_invalid_user_pages(process_info))
2019 /* Final check for concurrent evicton and atomic update. If
2020 * another eviction happens after successful update, it will
2021 * be a first eviction that calls quiesce_mm. The eviction
2022 * reference counting inside KFD will handle this case.
2024 if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
2028 if (kgd2kfd_resume_mm(mm)) {
2029 pr_err("%s: Failed to resume KFD\n", __func__);
2030 /* No recovery from this failure. Probably the CP is
2031 * hanging. No point trying again.
2036 mutex_unlock(&process_info->lock);
2038 put_task_struct(usertask);
2040 /* If validation failed, reschedule another attempt */
2042 schedule_delayed_work(&process_info->restore_userptr_work,
2043 msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2046 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2047 * KFD process identified by process_info
2049 * @process_info: amdkfd_process_info of the KFD process
2051 * After memory eviction, restore thread calls this function. The function
2052 * should be called when the Process is still valid. BO restore involves -
2054 * 1. Release old eviction fence and create new one
2055 * 2. Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2056 * 3 Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2057 * BOs that need to be reserved.
2058 * 4. Reserve all the BOs
2059 * 5. Validate of PD and PT BOs.
2060 * 6. Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2061 * 7. Add fence to all PD and PT BOs.
2062 * 8. Unreserve all BOs
2064 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2066 struct amdgpu_bo_list_entry *pd_bo_list;
2067 struct amdkfd_process_info *process_info = info;
2068 struct amdgpu_vm *peer_vm;
2069 struct kgd_mem *mem;
2070 struct bo_vm_reservation_context ctx;
2071 struct amdgpu_amdkfd_fence *new_fence;
2073 struct list_head duplicate_save;
2074 struct amdgpu_sync sync_obj;
2075 unsigned long failed_size = 0;
2076 unsigned long total_size = 0;
2078 INIT_LIST_HEAD(&duplicate_save);
2079 INIT_LIST_HEAD(&ctx.list);
2080 INIT_LIST_HEAD(&ctx.duplicates);
2082 pd_bo_list = kcalloc(process_info->n_vms,
2083 sizeof(struct amdgpu_bo_list_entry),
2089 mutex_lock(&process_info->lock);
2090 list_for_each_entry(peer_vm, &process_info->vm_list_head,
2092 amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2094 /* Reserve all BOs and page tables/directory. Add all BOs from
2095 * kfd_bo_list to ctx.list
2097 list_for_each_entry(mem, &process_info->kfd_bo_list,
2098 validate_list.head) {
2100 list_add_tail(&mem->resv_list.head, &ctx.list);
2101 mem->resv_list.bo = mem->validate_list.bo;
2102 mem->resv_list.num_shared = mem->validate_list.num_shared;
2105 ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2106 false, &duplicate_save);
2108 pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2109 goto ttm_reserve_fail;
2112 amdgpu_sync_create(&sync_obj);
2114 /* Validate PDs and PTs */
2115 ret = process_validate_vms(process_info);
2117 goto validate_map_fail;
2119 ret = process_sync_pds_resv(process_info, &sync_obj);
2121 pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2122 goto validate_map_fail;
2125 /* Validate BOs and map them to GPUVM (update VM page tables). */
2126 list_for_each_entry(mem, &process_info->kfd_bo_list,
2127 validate_list.head) {
2129 struct amdgpu_bo *bo = mem->bo;
2130 uint32_t domain = mem->domain;
2131 struct kfd_bo_va_list *bo_va_entry;
2133 total_size += amdgpu_bo_size(bo);
2135 ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2137 pr_debug("Memory eviction: Validate BOs failed\n");
2138 failed_size += amdgpu_bo_size(bo);
2139 ret = amdgpu_amdkfd_bo_validate(bo,
2140 AMDGPU_GEM_DOMAIN_GTT, false);
2142 pr_debug("Memory eviction: Try again\n");
2143 goto validate_map_fail;
2146 ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving);
2148 pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2149 goto validate_map_fail;
2151 list_for_each_entry(bo_va_entry, &mem->bo_va_list,
2153 ret = update_gpuvm_pte((struct amdgpu_device *)
2154 bo_va_entry->kgd_dev,
2158 pr_debug("Memory eviction: update PTE failed. Try again\n");
2159 goto validate_map_fail;
2165 pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2167 /* Update page directories */
2168 ret = process_update_pds(process_info, &sync_obj);
2170 pr_debug("Memory eviction: update PDs failed. Try again\n");
2171 goto validate_map_fail;
2174 /* Wait for validate and PT updates to finish */
2175 amdgpu_sync_wait(&sync_obj, false);
2177 /* Release old eviction fence and create new one, because fence only
2178 * goes from unsignaled to signaled, fence cannot be reused.
2179 * Use context and mm from the old fence.
2181 new_fence = amdgpu_amdkfd_fence_create(
2182 process_info->eviction_fence->base.context,
2183 process_info->eviction_fence->mm);
2185 pr_err("Failed to create eviction fence\n");
2187 goto validate_map_fail;
2189 dma_fence_put(&process_info->eviction_fence->base);
2190 process_info->eviction_fence = new_fence;
2191 *ef = dma_fence_get(&new_fence->base);
2193 /* Attach new eviction fence to all BOs */
2194 list_for_each_entry(mem, &process_info->kfd_bo_list,
2196 amdgpu_bo_fence(mem->bo,
2197 &process_info->eviction_fence->base, true);
2199 /* Attach eviction fence to PD / PT BOs */
2200 list_for_each_entry(peer_vm, &process_info->vm_list_head,
2202 struct amdgpu_bo *bo = peer_vm->root.base.bo;
2204 amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2208 ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2209 amdgpu_sync_free(&sync_obj);
2211 mutex_unlock(&process_info->lock);
2216 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2218 struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2219 struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2225 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2229 mutex_init(&(*mem)->lock);
2230 INIT_LIST_HEAD(&(*mem)->bo_va_list);
2231 (*mem)->bo = amdgpu_bo_ref(gws_bo);
2232 (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2233 (*mem)->process_info = process_info;
2234 add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2235 amdgpu_sync_create(&(*mem)->sync);
2238 /* Validate gws bo the first time it is added to process */
2239 mutex_lock(&(*mem)->process_info->lock);
2240 ret = amdgpu_bo_reserve(gws_bo, false);
2241 if (unlikely(ret)) {
2242 pr_err("Reserve gws bo failed %d\n", ret);
2243 goto bo_reservation_failure;
2246 ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2248 pr_err("GWS BO validate failed %d\n", ret);
2249 goto bo_validation_failure;
2251 /* GWS resource is shared b/t amdgpu and amdkfd
2252 * Add process eviction fence to bo so they can
2255 ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2257 goto reserve_shared_fail;
2258 amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2259 amdgpu_bo_unreserve(gws_bo);
2260 mutex_unlock(&(*mem)->process_info->lock);
2264 reserve_shared_fail:
2265 bo_validation_failure:
2266 amdgpu_bo_unreserve(gws_bo);
2267 bo_reservation_failure:
2268 mutex_unlock(&(*mem)->process_info->lock);
2269 amdgpu_sync_free(&(*mem)->sync);
2270 remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2271 amdgpu_bo_unref(&gws_bo);
2272 mutex_destroy(&(*mem)->lock);
2278 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2281 struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2282 struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2283 struct amdgpu_bo *gws_bo = kgd_mem->bo;
2285 /* Remove BO from process's validate list so restore worker won't touch
2288 remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2290 ret = amdgpu_bo_reserve(gws_bo, false);
2291 if (unlikely(ret)) {
2292 pr_err("Reserve gws bo failed %d\n", ret);
2293 //TODO add BO back to validate_list?
2296 amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2297 process_info->eviction_fence);
2298 amdgpu_bo_unreserve(gws_bo);
2299 amdgpu_sync_free(&kgd_mem->sync);
2300 amdgpu_bo_unref(&gws_bo);
2301 mutex_destroy(&kgd_mem->lock);
2306 /* Returns GPU-specific tiling mode information */
2307 int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
2308 struct tile_config *config)
2310 struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
2312 config->gb_addr_config = adev->gfx.config.gb_addr_config;
2313 config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2314 config->num_tile_configs =
2315 ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2316 config->macro_tile_config_ptr =
2317 adev->gfx.config.macrotile_mode_array;
2318 config->num_macro_tile_configs =
2319 ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2321 /* Those values are not set from GFX9 onwards */
2322 config->num_banks = adev->gfx.config.num_banks;
2323 config->num_ranks = adev->gfx.config.num_ranks;