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 struct svm_range_bo { 39 struct amdgpu_bo *bo; 40 struct kref kref; 41 struct list_head range_list; /* all svm ranges shared this bo */ 42 spinlock_t list_lock; 43 struct amdgpu_amdkfd_fence *eviction_fence; 44 struct work_struct eviction_work; 45 struct svm_range_list *svms; 46 uint32_t evicting; 47 }; 48 49 enum svm_work_list_ops { 50 SVM_OP_NULL, 51 SVM_OP_UNMAP_RANGE, 52 SVM_OP_UPDATE_RANGE_NOTIFIER, 53 SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP, 54 SVM_OP_ADD_RANGE, 55 SVM_OP_ADD_RANGE_AND_MAP 56 }; 57 58 struct svm_work_list_item { 59 enum svm_work_list_ops op; 60 struct mm_struct *mm; 61 }; 62 63 /** 64 * struct svm_range - shared virtual memory range 65 * 66 * @svms: list of svm ranges, structure defined in kfd_process 67 * @migrate_mutex: to serialize range migration, validation and mapping update 68 * @start: range start address in pages 69 * @last: range last address in pages 70 * @it_node: node [start, last] stored in interval tree, start, last are page 71 * aligned, page size is (last - start + 1) 72 * @list: link list node, used to scan all ranges of svms 73 * @update_list:link list node used to add to update_list 74 * @remove_list:link list node used to add to remove list 75 * @insert_list:link list node used to add to insert list 76 * @mapping: bo_va mapping structure to create and update GPU page table 77 * @npages: number of pages 78 * @dma_addr: dma mapping address on each GPU for system memory physical page 79 * @ttm_res: vram ttm resource map 80 * @offset: range start offset within mm_nodes 81 * @svm_bo: struct to manage splited amdgpu_bo 82 * @svm_bo_list:link list node, to scan all ranges which share same svm_bo 83 * @lock: protect prange start, last, child_list, svm_bo_list 84 * @saved_flags:save/restore current PF_MEMALLOC flags 85 * @flags: flags defined as KFD_IOCTL_SVM_FLAG_* 86 * @perferred_loc: perferred location, 0 for CPU, or GPU id 87 * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id 88 * @actual_loc: the actual location, 0 for CPU, or GPU id 89 * @granularity:migration granularity, log2 num pages 90 * @invalid: not 0 means cpu page table is invalidated 91 * @validate_timestamp: system timestamp when range is validated 92 * @notifier: register mmu interval notifier 93 * @work_item: deferred work item information 94 * @deferred_list: list header used to add range to deferred list 95 * @child_list: list header for split ranges which are not added to svms yet 96 * @bitmap_access: index bitmap of GPUs which can access the range 97 * @bitmap_aip: index bitmap of GPUs which can access the range in place 98 * 99 * Data structure for virtual memory range shared by CPU and GPUs, it can be 100 * allocated from system memory ram or device vram, and migrate from ram to vram 101 * or from vram to ram. 102 */ 103 struct svm_range { 104 struct svm_range_list *svms; 105 struct mutex migrate_mutex; 106 unsigned long start; 107 unsigned long last; 108 struct interval_tree_node it_node; 109 struct list_head list; 110 struct list_head update_list; 111 struct list_head remove_list; 112 struct list_head insert_list; 113 struct amdgpu_bo_va_mapping mapping; 114 uint64_t npages; 115 dma_addr_t *dma_addr[MAX_GPU_INSTANCE]; 116 struct ttm_resource *ttm_res; 117 uint64_t offset; 118 struct svm_range_bo *svm_bo; 119 struct list_head svm_bo_list; 120 struct mutex lock; 121 unsigned int saved_flags; 122 uint32_t flags; 123 uint32_t preferred_loc; 124 uint32_t prefetch_loc; 125 uint32_t actual_loc; 126 uint8_t granularity; 127 atomic_t invalid; 128 uint64_t validate_timestamp; 129 struct mmu_interval_notifier notifier; 130 struct svm_work_list_item work_item; 131 struct list_head deferred_list; 132 struct list_head child_list; 133 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 134 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 135 bool validated_once; 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 int svm_range_list_init(struct kfd_process *p); 151 void svm_range_list_fini(struct kfd_process *p); 152 int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 153 uint64_t size, uint32_t nattrs, 154 struct kfd_ioctl_svm_attribute *attrs); 155 struct svm_range *svm_range_from_addr(struct svm_range_list *svms, 156 unsigned long addr, 157 struct svm_range **parent); 158 struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange, 159 uint32_t id); 160 int svm_range_vram_node_new(struct amdgpu_device *adev, 161 struct svm_range *prange, bool clear); 162 void svm_range_vram_node_free(struct svm_range *prange); 163 int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, 164 unsigned long addr, struct svm_range *parent, 165 struct svm_range *prange); 166 int svm_range_restore_pages(struct amdgpu_device *adev, 167 unsigned int pasid, uint64_t addr); 168 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence); 169 void svm_range_add_list_work(struct svm_range_list *svms, 170 struct svm_range *prange, struct mm_struct *mm, 171 enum svm_work_list_ops op); 172 void schedule_deferred_list_work(struct svm_range_list *svms); 173 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, 174 unsigned long offset, unsigned long npages); 175 void svm_range_free_dma_mappings(struct svm_range *prange); 176 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm); 177 178 /* SVM API and HMM page migration work together, device memory type 179 * is initialized to not 0 when page migration register device memory. 180 */ 181 #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) 182 183 #else 184 185 struct kfd_process; 186 187 static inline int svm_range_list_init(struct kfd_process *p) 188 { 189 return 0; 190 } 191 static inline void svm_range_list_fini(struct kfd_process *p) 192 { 193 /* empty */ 194 } 195 196 static inline int svm_range_restore_pages(struct amdgpu_device *adev, 197 unsigned int pasid, uint64_t addr) 198 { 199 return -EFAULT; 200 } 201 202 static inline int svm_range_schedule_evict_svm_bo( 203 struct amdgpu_amdkfd_fence *fence) 204 { 205 WARN_ONCE(1, "SVM eviction fence triggered, but SVM is disabled"); 206 return -EINVAL; 207 } 208 209 #define KFD_IS_SVM_API_SUPPORTED(dev) false 210 211 #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */ 212 213 #endif /* KFD_SVM_H_ */ 214