1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 2 /* 3 * Copyright 2020-2021 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 * 23 */ 24 25 #ifndef KFD_SVM_H_ 26 #define KFD_SVM_H_ 27 28 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 29 30 #include <linux/rwsem.h> 31 #include <linux/list.h> 32 #include <linux/mutex.h> 33 #include <linux/sched/mm.h> 34 #include <linux/hmm.h> 35 #include "amdgpu.h" 36 #include "kfd_priv.h" 37 38 #define SVM_RANGE_VRAM_DOMAIN (1UL << 0) 39 #define SVM_ADEV_PGMAP_OWNER(adev)\ 40 ((adev)->hive ? (void *)(adev)->hive : (void *)(adev)) 41 42 struct svm_range_bo { 43 struct amdgpu_bo *bo; 44 struct kref kref; 45 struct list_head range_list; /* all svm ranges shared this bo */ 46 spinlock_t list_lock; 47 struct amdgpu_amdkfd_fence *eviction_fence; 48 struct work_struct eviction_work; 49 struct svm_range_list *svms; 50 uint32_t evicting; 51 struct work_struct release_work; 52 }; 53 54 enum svm_work_list_ops { 55 SVM_OP_NULL, 56 SVM_OP_UNMAP_RANGE, 57 SVM_OP_UPDATE_RANGE_NOTIFIER, 58 SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP, 59 SVM_OP_ADD_RANGE, 60 SVM_OP_ADD_RANGE_AND_MAP 61 }; 62 63 struct svm_work_list_item { 64 enum svm_work_list_ops op; 65 struct mm_struct *mm; 66 }; 67 68 /** 69 * struct svm_range - shared virtual memory range 70 * 71 * @svms: list of svm ranges, structure defined in kfd_process 72 * @migrate_mutex: to serialize range migration, validation and mapping update 73 * @start: range start address in pages 74 * @last: range last address in pages 75 * @it_node: node [start, last] stored in interval tree, start, last are page 76 * aligned, page size is (last - start + 1) 77 * @list: link list node, used to scan all ranges of svms 78 * @update_list:link list node used to add to update_list 79 * @remove_list:link list node used to add to remove list 80 * @insert_list:link list node used to add to insert list 81 * @mapping: bo_va mapping structure to create and update GPU page table 82 * @npages: number of pages 83 * @dma_addr: dma mapping address on each GPU for system memory physical page 84 * @ttm_res: vram ttm resource map 85 * @offset: range start offset within mm_nodes 86 * @svm_bo: struct to manage splited amdgpu_bo 87 * @svm_bo_list:link list node, to scan all ranges which share same svm_bo 88 * @lock: protect prange start, last, child_list, svm_bo_list 89 * @saved_flags:save/restore current PF_MEMALLOC flags 90 * @flags: flags defined as KFD_IOCTL_SVM_FLAG_* 91 * @perferred_loc: perferred location, 0 for CPU, or GPU id 92 * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id 93 * @actual_loc: the actual location, 0 for CPU, or GPU id 94 * @granularity:migration granularity, log2 num pages 95 * @invalid: not 0 means cpu page table is invalidated 96 * @validate_timestamp: system timestamp when range is validated 97 * @notifier: register mmu interval notifier 98 * @work_item: deferred work item information 99 * @deferred_list: list header used to add range to deferred list 100 * @child_list: list header for split ranges which are not added to svms yet 101 * @bitmap_access: index bitmap of GPUs which can access the range 102 * @bitmap_aip: index bitmap of GPUs which can access the range in place 103 * 104 * Data structure for virtual memory range shared by CPU and GPUs, it can be 105 * allocated from system memory ram or device vram, and migrate from ram to vram 106 * or from vram to ram. 107 */ 108 struct svm_range { 109 struct svm_range_list *svms; 110 struct mutex migrate_mutex; 111 unsigned long start; 112 unsigned long last; 113 struct interval_tree_node it_node; 114 struct list_head list; 115 struct list_head update_list; 116 struct list_head remove_list; 117 struct list_head insert_list; 118 uint64_t npages; 119 dma_addr_t *dma_addr[MAX_GPU_INSTANCE]; 120 struct ttm_resource *ttm_res; 121 uint64_t offset; 122 struct svm_range_bo *svm_bo; 123 struct list_head svm_bo_list; 124 struct mutex lock; 125 unsigned int saved_flags; 126 uint32_t flags; 127 uint32_t preferred_loc; 128 uint32_t prefetch_loc; 129 uint32_t actual_loc; 130 uint8_t granularity; 131 atomic_t invalid; 132 uint64_t validate_timestamp; 133 struct mmu_interval_notifier notifier; 134 struct svm_work_list_item work_item; 135 struct list_head deferred_list; 136 struct list_head child_list; 137 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 138 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 139 bool validated_once; 140 }; 141 142 static inline void svm_range_lock(struct svm_range *prange) 143 { 144 mutex_lock(&prange->lock); 145 prange->saved_flags = memalloc_noreclaim_save(); 146 147 } 148 static inline void svm_range_unlock(struct svm_range *prange) 149 { 150 memalloc_noreclaim_restore(prange->saved_flags); 151 mutex_unlock(&prange->lock); 152 } 153 154 static inline struct svm_range_bo *svm_range_bo_ref(struct svm_range_bo *svm_bo) 155 { 156 if (svm_bo) 157 kref_get(&svm_bo->kref); 158 159 return svm_bo; 160 } 161 162 int svm_range_list_init(struct kfd_process *p); 163 void svm_range_list_fini(struct kfd_process *p); 164 int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 165 uint64_t size, uint32_t nattrs, 166 struct kfd_ioctl_svm_attribute *attrs); 167 struct svm_range *svm_range_from_addr(struct svm_range_list *svms, 168 unsigned long addr, 169 struct svm_range **parent); 170 struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange, 171 uint32_t id); 172 int svm_range_vram_node_new(struct amdgpu_device *adev, 173 struct svm_range *prange, bool clear); 174 void svm_range_vram_node_free(struct svm_range *prange); 175 int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, 176 unsigned long addr, struct svm_range *parent, 177 struct svm_range *prange); 178 int svm_range_restore_pages(struct amdgpu_device *adev, 179 unsigned int pasid, uint64_t addr, bool write_fault); 180 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence); 181 void svm_range_add_list_work(struct svm_range_list *svms, 182 struct svm_range *prange, struct mm_struct *mm, 183 enum svm_work_list_ops op); 184 void schedule_deferred_list_work(struct svm_range_list *svms); 185 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, 186 unsigned long offset, unsigned long npages); 187 void svm_range_free_dma_mappings(struct svm_range *prange); 188 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm, 189 void *owner); 190 struct kfd_process_device * 191 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev); 192 void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm); 193 194 /* SVM API and HMM page migration work together, device memory type 195 * is initialized to not 0 when page migration register device memory. 196 */ 197 #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) 198 199 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); 200 #else 201 202 struct kfd_process; 203 204 static inline int svm_range_list_init(struct kfd_process *p) 205 { 206 return 0; 207 } 208 static inline void svm_range_list_fini(struct kfd_process *p) 209 { 210 /* empty */ 211 } 212 213 static inline int svm_range_restore_pages(struct amdgpu_device *adev, 214 unsigned int pasid, uint64_t addr, 215 bool write_fault) 216 { 217 return -EFAULT; 218 } 219 220 static inline int svm_range_schedule_evict_svm_bo( 221 struct amdgpu_amdkfd_fence *fence) 222 { 223 WARN_ONCE(1, "SVM eviction fence triggered, but SVM is disabled"); 224 return -EINVAL; 225 } 226 227 #define KFD_IS_SVM_API_SUPPORTED(dev) false 228 229 #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */ 230 231 #endif /* KFD_SVM_H_ */ 232