1 /* 2 * Copyright 2016 Advanced Micro Devices, Inc. 3 * 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: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 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. 21 * 22 * Authors: Christian König 23 */ 24 #ifndef __AMDGPU_VM_H__ 25 #define __AMDGPU_VM_H__ 26 27 #include <linux/rbtree.h> 28 29 #include "gpu_scheduler.h" 30 #include "amdgpu_sync.h" 31 #include "amdgpu_ring.h" 32 33 struct amdgpu_bo_va; 34 struct amdgpu_job; 35 struct amdgpu_bo_list_entry; 36 37 /* 38 * GPUVM handling 39 */ 40 41 /* maximum number of VMIDs */ 42 #define AMDGPU_NUM_VM 16 43 44 /* Maximum number of PTEs the hardware can write with one command */ 45 #define AMDGPU_VM_MAX_UPDATE_SIZE 0x3FFFF 46 47 /* number of entries in page table */ 48 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size) 49 50 /* PTBs (Page Table Blocks) need to be aligned to 32K */ 51 #define AMDGPU_VM_PTB_ALIGN_SIZE 32768 52 53 /* LOG2 number of continuous pages for the fragment field */ 54 #define AMDGPU_LOG2_PAGES_PER_FRAG 4 55 56 #define AMDGPU_PTE_VALID (1ULL << 0) 57 #define AMDGPU_PTE_SYSTEM (1ULL << 1) 58 #define AMDGPU_PTE_SNOOPED (1ULL << 2) 59 60 /* VI only */ 61 #define AMDGPU_PTE_EXECUTABLE (1ULL << 4) 62 63 #define AMDGPU_PTE_READABLE (1ULL << 5) 64 #define AMDGPU_PTE_WRITEABLE (1ULL << 6) 65 66 #define AMDGPU_PTE_FRAG(x) ((x & 0x1fULL) << 7) 67 68 /* TILED for VEGA10, reserved for older ASICs */ 69 #define AMDGPU_PTE_PRT (1ULL << 51) 70 71 /* VEGA10 only */ 72 #define AMDGPU_PTE_MTYPE(a) ((uint64_t)a << 57) 73 #define AMDGPU_PTE_MTYPE_MASK AMDGPU_PTE_MTYPE(3ULL) 74 75 /* How to programm VM fault handling */ 76 #define AMDGPU_VM_FAULT_STOP_NEVER 0 77 #define AMDGPU_VM_FAULT_STOP_FIRST 1 78 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2 79 80 /* max number of VMHUB */ 81 #define AMDGPU_MAX_VMHUBS 2 82 #define AMDGPU_GFXHUB 0 83 #define AMDGPU_MMHUB 1 84 85 /* hardcode that limit for now */ 86 #define AMDGPU_VA_RESERVED_SIZE (8 << 20) 87 /* max vmids dedicated for process */ 88 #define AMDGPU_VM_MAX_RESERVED_VMID 1 89 90 struct amdgpu_vm_pt { 91 struct amdgpu_bo *bo; 92 uint64_t addr; 93 94 /* array of page tables, one for each directory entry */ 95 struct amdgpu_vm_pt *entries; 96 unsigned last_entry_used; 97 }; 98 99 struct amdgpu_vm { 100 /* tree of virtual addresses mapped */ 101 struct rb_root va; 102 103 /* protecting invalidated */ 104 spinlock_t status_lock; 105 106 /* BOs moved, but not yet updated in the PT */ 107 struct list_head invalidated; 108 109 /* BOs cleared in the PT because of a move */ 110 struct list_head cleared; 111 112 /* BO mappings freed, but not yet updated in the PT */ 113 struct list_head freed; 114 115 /* contains the page directory */ 116 struct amdgpu_vm_pt root; 117 struct dma_fence *last_dir_update; 118 uint64_t last_eviction_counter; 119 120 /* protecting freed */ 121 spinlock_t freed_lock; 122 123 /* Scheduler entity for page table updates */ 124 struct amd_sched_entity entity; 125 126 /* client id */ 127 u64 client_id; 128 /* dedicated to vm */ 129 struct amdgpu_vm_id *reserved_vmid[AMDGPU_MAX_VMHUBS]; 130 /* each VM will map on CSA */ 131 struct amdgpu_bo_va *csa_bo_va; 132 }; 133 134 struct amdgpu_vm_id { 135 struct list_head list; 136 struct amdgpu_sync active; 137 struct dma_fence *last_flush; 138 atomic64_t owner; 139 140 uint64_t pd_gpu_addr; 141 /* last flushed PD/PT update */ 142 struct dma_fence *flushed_updates; 143 144 uint32_t current_gpu_reset_count; 145 146 uint32_t gds_base; 147 uint32_t gds_size; 148 uint32_t gws_base; 149 uint32_t gws_size; 150 uint32_t oa_base; 151 uint32_t oa_size; 152 }; 153 154 struct amdgpu_vm_id_manager { 155 struct mutex lock; 156 unsigned num_ids; 157 struct list_head ids_lru; 158 struct amdgpu_vm_id ids[AMDGPU_NUM_VM]; 159 atomic_t reserved_vmid_num; 160 }; 161 162 struct amdgpu_vm_manager { 163 /* Handling of VMIDs */ 164 struct amdgpu_vm_id_manager id_mgr[AMDGPU_MAX_VMHUBS]; 165 166 /* Handling of VM fences */ 167 u64 fence_context; 168 unsigned seqno[AMDGPU_MAX_RINGS]; 169 170 uint64_t max_pfn; 171 uint32_t num_level; 172 uint64_t vm_size; 173 uint32_t block_size; 174 /* vram base address for page table entry */ 175 u64 vram_base_offset; 176 /* is vm enabled? */ 177 bool enabled; 178 /* vm pte handling */ 179 const struct amdgpu_vm_pte_funcs *vm_pte_funcs; 180 struct amdgpu_ring *vm_pte_rings[AMDGPU_MAX_RINGS]; 181 unsigned vm_pte_num_rings; 182 atomic_t vm_pte_next_ring; 183 /* client id counter */ 184 atomic64_t client_counter; 185 186 /* partial resident texture handling */ 187 spinlock_t prt_lock; 188 atomic_t num_prt_users; 189 }; 190 191 void amdgpu_vm_manager_init(struct amdgpu_device *adev); 192 void amdgpu_vm_manager_fini(struct amdgpu_device *adev); 193 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm); 194 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm); 195 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, 196 struct list_head *validated, 197 struct amdgpu_bo_list_entry *entry); 198 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, 199 int (*callback)(void *p, struct amdgpu_bo *bo), 200 void *param); 201 void amdgpu_vm_move_pt_bos_in_lru(struct amdgpu_device *adev, 202 struct amdgpu_vm *vm); 203 int amdgpu_vm_alloc_pts(struct amdgpu_device *adev, 204 struct amdgpu_vm *vm, 205 uint64_t saddr, uint64_t size); 206 int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring, 207 struct amdgpu_sync *sync, struct dma_fence *fence, 208 struct amdgpu_job *job); 209 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job); 210 void amdgpu_vm_reset_id(struct amdgpu_device *adev, unsigned vmhub, 211 unsigned vmid); 212 void amdgpu_vm_reset_all_ids(struct amdgpu_device *adev); 213 int amdgpu_vm_update_directories(struct amdgpu_device *adev, 214 struct amdgpu_vm *vm); 215 int amdgpu_vm_clear_freed(struct amdgpu_device *adev, 216 struct amdgpu_vm *vm, 217 struct dma_fence **fence); 218 int amdgpu_vm_clear_invalids(struct amdgpu_device *adev, struct amdgpu_vm *vm, 219 struct amdgpu_sync *sync); 220 int amdgpu_vm_bo_update(struct amdgpu_device *adev, 221 struct amdgpu_bo_va *bo_va, 222 bool clear); 223 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, 224 struct amdgpu_bo *bo); 225 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, 226 struct amdgpu_bo *bo); 227 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, 228 struct amdgpu_vm *vm, 229 struct amdgpu_bo *bo); 230 int amdgpu_vm_bo_map(struct amdgpu_device *adev, 231 struct amdgpu_bo_va *bo_va, 232 uint64_t addr, uint64_t offset, 233 uint64_t size, uint64_t flags); 234 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev, 235 struct amdgpu_bo_va *bo_va, 236 uint64_t addr, uint64_t offset, 237 uint64_t size, uint64_t flags); 238 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, 239 struct amdgpu_bo_va *bo_va, 240 uint64_t addr); 241 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, 242 struct amdgpu_vm *vm, 243 uint64_t saddr, uint64_t size); 244 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, 245 struct amdgpu_bo_va *bo_va); 246 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint64_t vm_size); 247 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp); 248 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, 249 struct amdgpu_job *job); 250 251 #endif 252