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 uint32_t evicting; 50 struct work_struct release_work; 51 }; 52 53 enum svm_work_list_ops { 54 SVM_OP_NULL, 55 SVM_OP_UNMAP_RANGE, 56 SVM_OP_UPDATE_RANGE_NOTIFIER, 57 SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP, 58 SVM_OP_ADD_RANGE, 59 SVM_OP_ADD_RANGE_AND_MAP 60 }; 61 62 struct svm_work_list_item { 63 enum svm_work_list_ops op; 64 struct mm_struct *mm; 65 }; 66 67 /** 68 * struct svm_range - shared virtual memory range 69 * 70 * @svms: list of svm ranges, structure defined in kfd_process 71 * @migrate_mutex: to serialize range migration, validation and mapping update 72 * @start: range start address in pages 73 * @last: range last address in pages 74 * @it_node: node [start, last] stored in interval tree, start, last are page 75 * aligned, page size is (last - start + 1) 76 * @list: link list node, used to scan all ranges of svms 77 * @update_list:link list node used to add to update_list 78 * @mapping: bo_va mapping structure to create and update GPU page table 79 * @npages: number of pages 80 * @dma_addr: dma mapping address on each GPU for system memory physical page 81 * @ttm_res: vram ttm resource map 82 * @offset: range start offset within mm_nodes 83 * @svm_bo: struct to manage splited amdgpu_bo 84 * @svm_bo_list:link list node, to scan all ranges which share same svm_bo 85 * @lock: protect prange start, last, child_list, svm_bo_list 86 * @saved_flags:save/restore current PF_MEMALLOC flags 87 * @flags: flags defined as KFD_IOCTL_SVM_FLAG_* 88 * @perferred_loc: perferred location, 0 for CPU, or GPU id 89 * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id 90 * @actual_loc: the actual location, 0 for CPU, or GPU id 91 * @granularity:migration granularity, log2 num pages 92 * @invalid: not 0 means cpu page table is invalidated 93 * @validate_timestamp: system timestamp when range is validated 94 * @notifier: register mmu interval notifier 95 * @work_item: deferred work item information 96 * @deferred_list: list header used to add range to deferred list 97 * @child_list: list header for split ranges which are not added to svms yet 98 * @bitmap_access: index bitmap of GPUs which can access the range 99 * @bitmap_aip: index bitmap of GPUs which can access the range in place 100 * 101 * Data structure for virtual memory range shared by CPU and GPUs, it can be 102 * allocated from system memory ram or device vram, and migrate from ram to vram 103 * or from vram to ram. 104 */ 105 struct svm_range { 106 struct svm_range_list *svms; 107 struct mutex migrate_mutex; 108 unsigned long start; 109 unsigned long last; 110 struct interval_tree_node it_node; 111 struct list_head list; 112 struct list_head update_list; 113 uint64_t npages; 114 dma_addr_t *dma_addr[MAX_GPU_INSTANCE]; 115 struct ttm_resource *ttm_res; 116 uint64_t offset; 117 struct svm_range_bo *svm_bo; 118 struct list_head svm_bo_list; 119 struct mutex lock; 120 unsigned int saved_flags; 121 uint32_t flags; 122 uint32_t preferred_loc; 123 uint32_t prefetch_loc; 124 uint32_t actual_loc; 125 uint8_t granularity; 126 atomic_t invalid; 127 ktime_t validate_timestamp; 128 struct mmu_interval_notifier notifier; 129 struct svm_work_list_item work_item; 130 struct list_head deferred_list; 131 struct list_head child_list; 132 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 133 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 134 bool validated_once; 135 bool mapped_to_gpu; 136 }; 137 138 static inline void svm_range_lock(struct svm_range *prange) 139 { 140 mutex_lock(&prange->lock); 141 prange->saved_flags = memalloc_noreclaim_save(); 142 143 } 144 static inline void svm_range_unlock(struct svm_range *prange) 145 { 146 memalloc_noreclaim_restore(prange->saved_flags); 147 mutex_unlock(&prange->lock); 148 } 149 150 static inline struct svm_range_bo *svm_range_bo_ref(struct svm_range_bo *svm_bo) 151 { 152 if (svm_bo) 153 kref_get(&svm_bo->kref); 154 155 return svm_bo; 156 } 157 158 int svm_range_list_init(struct kfd_process *p); 159 void svm_range_list_fini(struct kfd_process *p); 160 int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 161 uint64_t size, uint32_t nattrs, 162 struct kfd_ioctl_svm_attribute *attrs); 163 struct svm_range *svm_range_from_addr(struct svm_range_list *svms, 164 unsigned long addr, 165 struct svm_range **parent); 166 struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange, 167 uint32_t id); 168 int svm_range_vram_node_new(struct amdgpu_device *adev, 169 struct svm_range *prange, bool clear); 170 void svm_range_vram_node_free(struct svm_range *prange); 171 int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, 172 unsigned long addr, struct svm_range *parent, 173 struct svm_range *prange); 174 int svm_range_restore_pages(struct amdgpu_device *adev, 175 unsigned int pasid, uint64_t addr, bool write_fault); 176 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence); 177 void svm_range_add_list_work(struct svm_range_list *svms, 178 struct svm_range *prange, struct mm_struct *mm, 179 enum svm_work_list_ops op); 180 void schedule_deferred_list_work(struct svm_range_list *svms); 181 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, 182 unsigned long offset, unsigned long npages); 183 void svm_range_free_dma_mappings(struct svm_range *prange); 184 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, 185 uint64_t *svm_priv_data_size); 186 int kfd_criu_checkpoint_svm(struct kfd_process *p, 187 uint8_t __user *user_priv_data, 188 uint64_t *priv_offset); 189 int kfd_criu_restore_svm(struct kfd_process *p, 190 uint8_t __user *user_priv_ptr, 191 uint64_t *priv_data_offset, 192 uint64_t max_priv_data_size); 193 int kfd_criu_resume_svm(struct kfd_process *p); 194 struct kfd_process_device * 195 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev); 196 void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm); 197 198 /* SVM API and HMM page migration work together, device memory type 199 * is initialized to not 0 when page migration register device memory. 200 */ 201 #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) 202 203 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); 204 205 void svm_range_set_max_pages(struct amdgpu_device *adev); 206 int svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled); 207 208 #else 209 210 struct kfd_process; 211 212 static inline int svm_range_list_init(struct kfd_process *p) 213 { 214 return 0; 215 } 216 static inline void svm_range_list_fini(struct kfd_process *p) 217 { 218 /* empty */ 219 } 220 221 static inline int svm_range_restore_pages(struct amdgpu_device *adev, 222 unsigned int pasid, uint64_t addr, 223 bool write_fault) 224 { 225 return -EFAULT; 226 } 227 228 static inline int svm_range_schedule_evict_svm_bo( 229 struct amdgpu_amdkfd_fence *fence) 230 { 231 WARN_ONCE(1, "SVM eviction fence triggered, but SVM is disabled"); 232 return -EINVAL; 233 } 234 235 static inline int svm_range_get_info(struct kfd_process *p, 236 uint32_t *num_svm_ranges, 237 uint64_t *svm_priv_data_size) 238 { 239 *num_svm_ranges = 0; 240 *svm_priv_data_size = 0; 241 return 0; 242 } 243 244 static inline int kfd_criu_checkpoint_svm(struct kfd_process *p, 245 uint8_t __user *user_priv_data, 246 uint64_t *priv_offset) 247 { 248 return 0; 249 } 250 251 static inline int kfd_criu_restore_svm(struct kfd_process *p, 252 uint8_t __user *user_priv_ptr, 253 uint64_t *priv_data_offset, 254 uint64_t max_priv_data_size) 255 { 256 return -EINVAL; 257 } 258 259 static inline int kfd_criu_resume_svm(struct kfd_process *p) 260 { 261 return 0; 262 } 263 264 #define KFD_IS_SVM_API_SUPPORTED(dev) false 265 266 #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */ 267 268 #endif /* KFD_SVM_H_ */ 269