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/idr.h> 28 #include <linux/kfifo.h> 29 #include <linux/rbtree.h> 30 #include <drm/gpu_scheduler.h> 31 #include <drm/drm_file.h> 32 #include <drm/ttm/ttm_bo_driver.h> 33 #include <linux/sched/mm.h> 34 35 #include "amdgpu_sync.h" 36 #include "amdgpu_ring.h" 37 #include "amdgpu_ids.h" 38 39 struct amdgpu_bo_va; 40 struct amdgpu_job; 41 struct amdgpu_bo_list_entry; 42 43 /* 44 * GPUVM handling 45 */ 46 47 /* Maximum number of PTEs the hardware can write with one command */ 48 #define AMDGPU_VM_MAX_UPDATE_SIZE 0x3FFFF 49 50 /* number of entries in page table */ 51 #define AMDGPU_VM_PTE_COUNT(adev) (1 << (adev)->vm_manager.block_size) 52 53 #define AMDGPU_PTE_VALID (1ULL << 0) 54 #define AMDGPU_PTE_SYSTEM (1ULL << 1) 55 #define AMDGPU_PTE_SNOOPED (1ULL << 2) 56 57 /* VI only */ 58 #define AMDGPU_PTE_EXECUTABLE (1ULL << 4) 59 60 #define AMDGPU_PTE_READABLE (1ULL << 5) 61 #define AMDGPU_PTE_WRITEABLE (1ULL << 6) 62 63 #define AMDGPU_PTE_FRAG(x) ((x & 0x1fULL) << 7) 64 65 /* TILED for VEGA10, reserved for older ASICs */ 66 #define AMDGPU_PTE_PRT (1ULL << 51) 67 68 /* PDE is handled as PTE for VEGA10 */ 69 #define AMDGPU_PDE_PTE (1ULL << 54) 70 71 #define AMDGPU_PTE_LOG (1ULL << 55) 72 73 /* PTE is handled as PDE for VEGA10 (Translate Further) */ 74 #define AMDGPU_PTE_TF (1ULL << 56) 75 76 /* PDE Block Fragment Size for VEGA10 */ 77 #define AMDGPU_PDE_BFS(a) ((uint64_t)a << 59) 78 79 80 /* For GFX9 */ 81 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57) 82 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL) 83 84 #define AMDGPU_MTYPE_NC 0 85 #define AMDGPU_MTYPE_CC 2 86 87 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \ 88 | AMDGPU_PTE_SNOOPED \ 89 | AMDGPU_PTE_EXECUTABLE \ 90 | AMDGPU_PTE_READABLE \ 91 | AMDGPU_PTE_WRITEABLE \ 92 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC)) 93 94 /* gfx10 */ 95 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48) 96 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL) 97 98 /* How to programm VM fault handling */ 99 #define AMDGPU_VM_FAULT_STOP_NEVER 0 100 #define AMDGPU_VM_FAULT_STOP_FIRST 1 101 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2 102 103 /* Reserve 4MB VRAM for page tables */ 104 #define AMDGPU_VM_RESERVED_VRAM (4ULL << 20) 105 106 /* max number of VMHUB */ 107 #define AMDGPU_MAX_VMHUBS 3 108 #define AMDGPU_GFXHUB_0 0 109 #define AMDGPU_MMHUB_0 1 110 #define AMDGPU_MMHUB_1 2 111 112 /* hardcode that limit for now */ 113 #define AMDGPU_VA_RESERVED_SIZE (1ULL << 20) 114 115 /* max vmids dedicated for process */ 116 #define AMDGPU_VM_MAX_RESERVED_VMID 1 117 118 #define AMDGPU_VM_CONTEXT_GFX 0 119 #define AMDGPU_VM_CONTEXT_COMPUTE 1 120 121 /* See vm_update_mode */ 122 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0) 123 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1) 124 125 /* VMPT level enumerate, and the hiberachy is: 126 * PDB2->PDB1->PDB0->PTB 127 */ 128 enum amdgpu_vm_level { 129 AMDGPU_VM_PDB2, 130 AMDGPU_VM_PDB1, 131 AMDGPU_VM_PDB0, 132 AMDGPU_VM_PTB 133 }; 134 135 /* base structure for tracking BO usage in a VM */ 136 struct amdgpu_vm_bo_base { 137 /* constant after initialization */ 138 struct amdgpu_vm *vm; 139 struct amdgpu_bo *bo; 140 141 /* protected by bo being reserved */ 142 struct amdgpu_vm_bo_base *next; 143 144 /* protected by spinlock */ 145 struct list_head vm_status; 146 147 /* protected by the BO being reserved */ 148 bool moved; 149 }; 150 151 struct amdgpu_vm_pt { 152 struct amdgpu_vm_bo_base base; 153 154 /* array of page tables, one for each directory entry */ 155 struct amdgpu_vm_pt *entries; 156 }; 157 158 /* provided by hw blocks that can write ptes, e.g., sdma */ 159 struct amdgpu_vm_pte_funcs { 160 /* number of dw to reserve per operation */ 161 unsigned copy_pte_num_dw; 162 163 /* copy pte entries from GART */ 164 void (*copy_pte)(struct amdgpu_ib *ib, 165 uint64_t pe, uint64_t src, 166 unsigned count); 167 168 /* write pte one entry at a time with addr mapping */ 169 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe, 170 uint64_t value, unsigned count, 171 uint32_t incr); 172 /* for linear pte/pde updates without addr mapping */ 173 void (*set_pte_pde)(struct amdgpu_ib *ib, 174 uint64_t pe, 175 uint64_t addr, unsigned count, 176 uint32_t incr, uint64_t flags); 177 }; 178 179 struct amdgpu_task_info { 180 char process_name[TASK_COMM_LEN]; 181 char task_name[TASK_COMM_LEN]; 182 pid_t pid; 183 pid_t tgid; 184 }; 185 186 /** 187 * struct amdgpu_vm_update_params 188 * 189 * Encapsulate some VM table update parameters to reduce 190 * the number of function parameters 191 * 192 */ 193 struct amdgpu_vm_update_params { 194 195 /** 196 * @adev: amdgpu device we do this update for 197 */ 198 struct amdgpu_device *adev; 199 200 /** 201 * @vm: optional amdgpu_vm we do this update for 202 */ 203 struct amdgpu_vm *vm; 204 205 /** 206 * @direct: if changes should be made directly 207 */ 208 bool direct; 209 210 /** 211 * @pages_addr: 212 * 213 * DMA addresses to use for mapping 214 */ 215 dma_addr_t *pages_addr; 216 217 /** 218 * @job: job to used for hw submission 219 */ 220 struct amdgpu_job *job; 221 222 /** 223 * @num_dw_left: number of dw left for the IB 224 */ 225 unsigned int num_dw_left; 226 }; 227 228 struct amdgpu_vm_update_funcs { 229 int (*map_table)(struct amdgpu_bo *bo); 230 int (*prepare)(struct amdgpu_vm_update_params *p, void * owner, 231 struct dma_fence *exclusive); 232 int (*update)(struct amdgpu_vm_update_params *p, 233 struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, 234 unsigned count, uint32_t incr, uint64_t flags); 235 int (*commit)(struct amdgpu_vm_update_params *p, 236 struct dma_fence **fence); 237 }; 238 239 struct amdgpu_vm { 240 /* tree of virtual addresses mapped */ 241 struct rb_root_cached va; 242 243 /* Lock to prevent eviction while we are updating page tables 244 * use vm_eviction_lock/unlock(vm) 245 */ 246 struct mutex eviction_lock; 247 bool evicting; 248 unsigned int saved_flags; 249 250 /* BOs who needs a validation */ 251 struct list_head evicted; 252 253 /* PT BOs which relocated and their parent need an update */ 254 struct list_head relocated; 255 256 /* per VM BOs moved, but not yet updated in the PT */ 257 struct list_head moved; 258 259 /* All BOs of this VM not currently in the state machine */ 260 struct list_head idle; 261 262 /* regular invalidated BOs, but not yet updated in the PT */ 263 struct list_head invalidated; 264 spinlock_t invalidated_lock; 265 266 /* BO mappings freed, but not yet updated in the PT */ 267 struct list_head freed; 268 269 /* contains the page directory */ 270 struct amdgpu_vm_pt root; 271 struct dma_fence *last_update; 272 273 /* Scheduler entities for page table updates */ 274 struct drm_sched_entity direct; 275 struct drm_sched_entity delayed; 276 277 /* Last submission to the scheduler entities */ 278 struct dma_fence *last_direct; 279 struct dma_fence *last_delayed; 280 281 unsigned int pasid; 282 /* dedicated to vm */ 283 struct amdgpu_vmid *reserved_vmid[AMDGPU_MAX_VMHUBS]; 284 285 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */ 286 bool use_cpu_for_update; 287 288 /* Functions to use for VM table updates */ 289 const struct amdgpu_vm_update_funcs *update_funcs; 290 291 /* Flag to indicate ATS support from PTE for GFX9 */ 292 bool pte_support_ats; 293 294 /* Up to 128 pending retry page faults */ 295 DECLARE_KFIFO(faults, u64, 128); 296 297 /* Points to the KFD process VM info */ 298 struct amdkfd_process_info *process_info; 299 300 /* List node in amdkfd_process_info.vm_list_head */ 301 struct list_head vm_list_node; 302 303 /* Valid while the PD is reserved or fenced */ 304 uint64_t pd_phys_addr; 305 306 /* Some basic info about the task */ 307 struct amdgpu_task_info task_info; 308 309 /* Store positions of group of BOs */ 310 struct ttm_lru_bulk_move lru_bulk_move; 311 /* mark whether can do the bulk move */ 312 bool bulk_moveable; 313 /* Flag to indicate if VM is used for compute */ 314 bool is_compute_context; 315 }; 316 317 struct amdgpu_vm_manager { 318 /* Handling of VMIDs */ 319 struct amdgpu_vmid_mgr id_mgr[AMDGPU_MAX_VMHUBS]; 320 321 /* Handling of VM fences */ 322 u64 fence_context; 323 unsigned seqno[AMDGPU_MAX_RINGS]; 324 325 uint64_t max_pfn; 326 uint32_t num_level; 327 uint32_t block_size; 328 uint32_t fragment_size; 329 enum amdgpu_vm_level root_level; 330 /* vram base address for page table entry */ 331 u64 vram_base_offset; 332 /* vm pte handling */ 333 const struct amdgpu_vm_pte_funcs *vm_pte_funcs; 334 struct drm_gpu_scheduler *vm_pte_scheds[AMDGPU_MAX_RINGS]; 335 unsigned vm_pte_num_scheds; 336 struct amdgpu_ring *page_fault; 337 338 /* partial resident texture handling */ 339 spinlock_t prt_lock; 340 atomic_t num_prt_users; 341 342 /* controls how VM page tables are updated for Graphics and Compute. 343 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU 344 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU 345 */ 346 int vm_update_mode; 347 348 /* PASID to VM mapping, will be used in interrupt context to 349 * look up VM of a page fault 350 */ 351 struct idr pasid_idr; 352 spinlock_t pasid_lock; 353 354 /* counter of mapped memory through xgmi */ 355 uint32_t xgmi_map_counter; 356 struct mutex lock_pstate; 357 }; 358 359 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count))) 360 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr))) 361 #define amdgpu_vm_set_pte_pde(adev, ib, pe, addr, count, incr, flags) ((adev)->vm_manager.vm_pte_funcs->set_pte_pde((ib), (pe), (addr), (count), (incr), (flags))) 362 363 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs; 364 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs; 365 366 void amdgpu_vm_manager_init(struct amdgpu_device *adev); 367 void amdgpu_vm_manager_fini(struct amdgpu_device *adev); 368 369 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout); 370 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, 371 int vm_context, unsigned int pasid); 372 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid); 373 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm); 374 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm); 375 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, 376 struct list_head *validated, 377 struct amdgpu_bo_list_entry *entry); 378 bool amdgpu_vm_ready(struct amdgpu_vm *vm); 379 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, 380 int (*callback)(void *p, struct amdgpu_bo *bo), 381 void *param); 382 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync); 383 int amdgpu_vm_update_pdes(struct amdgpu_device *adev, 384 struct amdgpu_vm *vm, bool direct); 385 int amdgpu_vm_clear_freed(struct amdgpu_device *adev, 386 struct amdgpu_vm *vm, 387 struct dma_fence **fence); 388 int amdgpu_vm_handle_moved(struct amdgpu_device *adev, 389 struct amdgpu_vm *vm); 390 int amdgpu_vm_bo_update(struct amdgpu_device *adev, 391 struct amdgpu_bo_va *bo_va, 392 bool clear); 393 bool amdgpu_vm_evictable(struct amdgpu_bo *bo); 394 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, 395 struct amdgpu_bo *bo, bool evicted); 396 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr); 397 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, 398 struct amdgpu_bo *bo); 399 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, 400 struct amdgpu_vm *vm, 401 struct amdgpu_bo *bo); 402 int amdgpu_vm_bo_map(struct amdgpu_device *adev, 403 struct amdgpu_bo_va *bo_va, 404 uint64_t addr, uint64_t offset, 405 uint64_t size, uint64_t flags); 406 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev, 407 struct amdgpu_bo_va *bo_va, 408 uint64_t addr, uint64_t offset, 409 uint64_t size, uint64_t flags); 410 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, 411 struct amdgpu_bo_va *bo_va, 412 uint64_t addr); 413 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, 414 struct amdgpu_vm *vm, 415 uint64_t saddr, uint64_t size); 416 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, 417 uint64_t addr); 418 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket); 419 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, 420 struct amdgpu_bo_va *bo_va); 421 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size, 422 uint32_t fragment_size_default, unsigned max_level, 423 unsigned max_bits); 424 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp); 425 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, 426 struct amdgpu_job *job); 427 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev); 428 429 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid, 430 struct amdgpu_task_info *task_info); 431 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, unsigned int pasid, 432 uint64_t addr); 433 434 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm); 435 436 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, 437 struct amdgpu_vm *vm); 438 void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo); 439 440 #endif 441