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 /* RV+ */ 58 #define AMDGPU_PTE_TMZ (1ULL << 3) 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 /* PDE is handled as PTE for VEGA10 */ 72 #define AMDGPU_PDE_PTE (1ULL << 54) 73 74 #define AMDGPU_PTE_LOG (1ULL << 55) 75 76 /* PTE is handled as PDE for VEGA10 (Translate Further) */ 77 #define AMDGPU_PTE_TF (1ULL << 56) 78 79 /* PDE Block Fragment Size for VEGA10 */ 80 #define AMDGPU_PDE_BFS(a) ((uint64_t)a << 59) 81 82 83 /* For GFX9 */ 84 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57) 85 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL) 86 87 #define AMDGPU_MTYPE_NC 0 88 #define AMDGPU_MTYPE_CC 2 89 90 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \ 91 | AMDGPU_PTE_SNOOPED \ 92 | AMDGPU_PTE_EXECUTABLE \ 93 | AMDGPU_PTE_READABLE \ 94 | AMDGPU_PTE_WRITEABLE \ 95 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC)) 96 97 /* gfx10 */ 98 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48) 99 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL) 100 101 /* How to programm VM fault handling */ 102 #define AMDGPU_VM_FAULT_STOP_NEVER 0 103 #define AMDGPU_VM_FAULT_STOP_FIRST 1 104 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2 105 106 /* Reserve 4MB VRAM for page tables */ 107 #define AMDGPU_VM_RESERVED_VRAM (4ULL << 20) 108 109 /* max number of VMHUB */ 110 #define AMDGPU_MAX_VMHUBS 3 111 #define AMDGPU_GFXHUB_0 0 112 #define AMDGPU_MMHUB_0 1 113 #define AMDGPU_MMHUB_1 2 114 115 /* hardcode that limit for now */ 116 #define AMDGPU_VA_RESERVED_SIZE (1ULL << 20) 117 118 /* max vmids dedicated for process */ 119 #define AMDGPU_VM_MAX_RESERVED_VMID 1 120 121 #define AMDGPU_VM_CONTEXT_GFX 0 122 #define AMDGPU_VM_CONTEXT_COMPUTE 1 123 124 /* See vm_update_mode */ 125 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0) 126 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1) 127 128 /* VMPT level enumerate, and the hiberachy is: 129 * PDB2->PDB1->PDB0->PTB 130 */ 131 enum amdgpu_vm_level { 132 AMDGPU_VM_PDB2, 133 AMDGPU_VM_PDB1, 134 AMDGPU_VM_PDB0, 135 AMDGPU_VM_PTB 136 }; 137 138 /* base structure for tracking BO usage in a VM */ 139 struct amdgpu_vm_bo_base { 140 /* constant after initialization */ 141 struct amdgpu_vm *vm; 142 struct amdgpu_bo *bo; 143 144 /* protected by bo being reserved */ 145 struct amdgpu_vm_bo_base *next; 146 147 /* protected by spinlock */ 148 struct list_head vm_status; 149 150 /* protected by the BO being reserved */ 151 bool moved; 152 }; 153 154 struct amdgpu_vm_pt { 155 struct amdgpu_vm_bo_base base; 156 157 /* array of page tables, one for each directory entry */ 158 struct amdgpu_vm_pt *entries; 159 }; 160 161 /* provided by hw blocks that can write ptes, e.g., sdma */ 162 struct amdgpu_vm_pte_funcs { 163 /* number of dw to reserve per operation */ 164 unsigned copy_pte_num_dw; 165 166 /* copy pte entries from GART */ 167 void (*copy_pte)(struct amdgpu_ib *ib, 168 uint64_t pe, uint64_t src, 169 unsigned count); 170 171 /* write pte one entry at a time with addr mapping */ 172 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe, 173 uint64_t value, unsigned count, 174 uint32_t incr); 175 /* for linear pte/pde updates without addr mapping */ 176 void (*set_pte_pde)(struct amdgpu_ib *ib, 177 uint64_t pe, 178 uint64_t addr, unsigned count, 179 uint32_t incr, uint64_t flags); 180 }; 181 182 struct amdgpu_task_info { 183 char process_name[TASK_COMM_LEN]; 184 char task_name[TASK_COMM_LEN]; 185 pid_t pid; 186 pid_t tgid; 187 }; 188 189 /** 190 * struct amdgpu_vm_update_params 191 * 192 * Encapsulate some VM table update parameters to reduce 193 * the number of function parameters 194 * 195 */ 196 struct amdgpu_vm_update_params { 197 198 /** 199 * @adev: amdgpu device we do this update for 200 */ 201 struct amdgpu_device *adev; 202 203 /** 204 * @vm: optional amdgpu_vm we do this update for 205 */ 206 struct amdgpu_vm *vm; 207 208 /** 209 * @immediate: if changes should be made immediately 210 */ 211 bool immediate; 212 213 /** 214 * @unlocked: true if the root BO is not locked 215 */ 216 bool unlocked; 217 218 /** 219 * @pages_addr: 220 * 221 * DMA addresses to use for mapping 222 */ 223 dma_addr_t *pages_addr; 224 225 /** 226 * @job: job to used for hw submission 227 */ 228 struct amdgpu_job *job; 229 230 /** 231 * @num_dw_left: number of dw left for the IB 232 */ 233 unsigned int num_dw_left; 234 }; 235 236 struct amdgpu_vm_update_funcs { 237 int (*map_table)(struct amdgpu_bo *bo); 238 int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv, 239 enum amdgpu_sync_mode sync_mode); 240 int (*update)(struct amdgpu_vm_update_params *p, 241 struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, 242 unsigned count, uint32_t incr, uint64_t flags); 243 int (*commit)(struct amdgpu_vm_update_params *p, 244 struct dma_fence **fence); 245 }; 246 247 struct amdgpu_vm { 248 /* tree of virtual addresses mapped */ 249 struct rb_root_cached va; 250 251 /* Lock to prevent eviction while we are updating page tables 252 * use vm_eviction_lock/unlock(vm) 253 */ 254 struct mutex eviction_lock; 255 bool evicting; 256 unsigned int saved_flags; 257 258 /* BOs who needs a validation */ 259 struct list_head evicted; 260 261 /* PT BOs which relocated and their parent need an update */ 262 struct list_head relocated; 263 264 /* per VM BOs moved, but not yet updated in the PT */ 265 struct list_head moved; 266 267 /* All BOs of this VM not currently in the state machine */ 268 struct list_head idle; 269 270 /* regular invalidated BOs, but not yet updated in the PT */ 271 struct list_head invalidated; 272 spinlock_t invalidated_lock; 273 274 /* BO mappings freed, but not yet updated in the PT */ 275 struct list_head freed; 276 277 /* contains the page directory */ 278 struct amdgpu_vm_pt root; 279 struct dma_fence *last_update; 280 281 /* Scheduler entities for page table updates */ 282 struct drm_sched_entity immediate; 283 struct drm_sched_entity delayed; 284 285 /* Last unlocked submission to the scheduler entities */ 286 struct dma_fence *last_unlocked; 287 288 unsigned int pasid; 289 /* dedicated to vm */ 290 struct amdgpu_vmid *reserved_vmid[AMDGPU_MAX_VMHUBS]; 291 292 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */ 293 bool use_cpu_for_update; 294 295 /* Functions to use for VM table updates */ 296 const struct amdgpu_vm_update_funcs *update_funcs; 297 298 /* Flag to indicate ATS support from PTE for GFX9 */ 299 bool pte_support_ats; 300 301 /* Up to 128 pending retry page faults */ 302 DECLARE_KFIFO(faults, u64, 128); 303 304 /* Points to the KFD process VM info */ 305 struct amdkfd_process_info *process_info; 306 307 /* List node in amdkfd_process_info.vm_list_head */ 308 struct list_head vm_list_node; 309 310 /* Valid while the PD is reserved or fenced */ 311 uint64_t pd_phys_addr; 312 313 /* Some basic info about the task */ 314 struct amdgpu_task_info task_info; 315 316 /* Store positions of group of BOs */ 317 struct ttm_lru_bulk_move lru_bulk_move; 318 /* mark whether can do the bulk move */ 319 bool bulk_moveable; 320 /* Flag to indicate if VM is used for compute */ 321 bool is_compute_context; 322 }; 323 324 struct amdgpu_vm_manager { 325 /* Handling of VMIDs */ 326 struct amdgpu_vmid_mgr id_mgr[AMDGPU_MAX_VMHUBS]; 327 unsigned int first_kfd_vmid; 328 329 /* Handling of VM fences */ 330 u64 fence_context; 331 unsigned seqno[AMDGPU_MAX_RINGS]; 332 333 uint64_t max_pfn; 334 uint32_t num_level; 335 uint32_t block_size; 336 uint32_t fragment_size; 337 enum amdgpu_vm_level root_level; 338 /* vram base address for page table entry */ 339 u64 vram_base_offset; 340 /* vm pte handling */ 341 const struct amdgpu_vm_pte_funcs *vm_pte_funcs; 342 struct drm_gpu_scheduler *vm_pte_scheds[AMDGPU_MAX_RINGS]; 343 unsigned vm_pte_num_scheds; 344 struct amdgpu_ring *page_fault; 345 346 /* partial resident texture handling */ 347 spinlock_t prt_lock; 348 atomic_t num_prt_users; 349 350 /* controls how VM page tables are updated for Graphics and Compute. 351 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU 352 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU 353 */ 354 int vm_update_mode; 355 356 /* PASID to VM mapping, will be used in interrupt context to 357 * look up VM of a page fault 358 */ 359 struct idr pasid_idr; 360 spinlock_t pasid_lock; 361 }; 362 363 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count))) 364 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr))) 365 #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))) 366 367 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs; 368 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs; 369 370 void amdgpu_vm_manager_init(struct amdgpu_device *adev); 371 void amdgpu_vm_manager_fini(struct amdgpu_device *adev); 372 373 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout); 374 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, 375 int vm_context, u32 pasid); 376 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid); 377 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm); 378 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm); 379 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, 380 struct list_head *validated, 381 struct amdgpu_bo_list_entry *entry); 382 bool amdgpu_vm_ready(struct amdgpu_vm *vm); 383 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, 384 int (*callback)(void *p, struct amdgpu_bo *bo), 385 void *param); 386 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync); 387 int amdgpu_vm_update_pdes(struct amdgpu_device *adev, 388 struct amdgpu_vm *vm, bool immediate); 389 int amdgpu_vm_clear_freed(struct amdgpu_device *adev, 390 struct amdgpu_vm *vm, 391 struct dma_fence **fence); 392 int amdgpu_vm_handle_moved(struct amdgpu_device *adev, 393 struct amdgpu_vm *vm); 394 int amdgpu_vm_bo_update(struct amdgpu_device *adev, 395 struct amdgpu_bo_va *bo_va, 396 bool clear); 397 bool amdgpu_vm_evictable(struct amdgpu_bo *bo); 398 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, 399 struct amdgpu_bo *bo, bool evicted); 400 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr); 401 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, 402 struct amdgpu_bo *bo); 403 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, 404 struct amdgpu_vm *vm, 405 struct amdgpu_bo *bo); 406 int amdgpu_vm_bo_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_replace_map(struct amdgpu_device *adev, 411 struct amdgpu_bo_va *bo_va, 412 uint64_t addr, uint64_t offset, 413 uint64_t size, uint64_t flags); 414 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, 415 struct amdgpu_bo_va *bo_va, 416 uint64_t addr); 417 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, 418 struct amdgpu_vm *vm, 419 uint64_t saddr, uint64_t size); 420 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, 421 uint64_t addr); 422 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket); 423 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, 424 struct amdgpu_bo_va *bo_va); 425 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size, 426 uint32_t fragment_size_default, unsigned max_level, 427 unsigned max_bits); 428 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp); 429 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, 430 struct amdgpu_job *job); 431 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev); 432 433 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid, 434 struct amdgpu_task_info *task_info); 435 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, 436 uint64_t addr); 437 438 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm); 439 440 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, 441 struct amdgpu_vm *vm); 442 void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo); 443 444 #endif 445