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 /* MALL noalloc for sienna_cichlid, reserved for older ASICs */ 80 #define AMDGPU_PTE_NOALLOC (1ULL << 58) 81 82 /* PDE Block Fragment Size for VEGA10 */ 83 #define AMDGPU_PDE_BFS(a) ((uint64_t)a << 59) 84 85 86 /* For GFX9 */ 87 #define AMDGPU_PTE_MTYPE_VG10(a) ((uint64_t)(a) << 57) 88 #define AMDGPU_PTE_MTYPE_VG10_MASK AMDGPU_PTE_MTYPE_VG10(3ULL) 89 90 #define AMDGPU_MTYPE_NC 0 91 #define AMDGPU_MTYPE_CC 2 92 93 #define AMDGPU_PTE_DEFAULT_ATC (AMDGPU_PTE_SYSTEM \ 94 | AMDGPU_PTE_SNOOPED \ 95 | AMDGPU_PTE_EXECUTABLE \ 96 | AMDGPU_PTE_READABLE \ 97 | AMDGPU_PTE_WRITEABLE \ 98 | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_CC)) 99 100 /* gfx10 */ 101 #define AMDGPU_PTE_MTYPE_NV10(a) ((uint64_t)(a) << 48) 102 #define AMDGPU_PTE_MTYPE_NV10_MASK AMDGPU_PTE_MTYPE_NV10(7ULL) 103 104 /* How to program VM fault handling */ 105 #define AMDGPU_VM_FAULT_STOP_NEVER 0 106 #define AMDGPU_VM_FAULT_STOP_FIRST 1 107 #define AMDGPU_VM_FAULT_STOP_ALWAYS 2 108 109 /* Reserve 4MB VRAM for page tables */ 110 #define AMDGPU_VM_RESERVED_VRAM (8ULL << 20) 111 112 /* max number of VMHUB */ 113 #define AMDGPU_MAX_VMHUBS 3 114 #define AMDGPU_GFXHUB_0 0 115 #define AMDGPU_MMHUB_0 1 116 #define AMDGPU_MMHUB_1 2 117 118 /* Reserve 2MB at top/bottom of address space for kernel use */ 119 #define AMDGPU_VA_RESERVED_SIZE (2ULL << 20) 120 121 /* max vmids dedicated for process */ 122 #define AMDGPU_VM_MAX_RESERVED_VMID 1 123 124 #define AMDGPU_VM_CONTEXT_GFX 0 125 #define AMDGPU_VM_CONTEXT_COMPUTE 1 126 127 /* See vm_update_mode */ 128 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0) 129 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1) 130 131 /* VMPT level enumerate, and the hiberachy is: 132 * PDB2->PDB1->PDB0->PTB 133 */ 134 enum amdgpu_vm_level { 135 AMDGPU_VM_PDB2, 136 AMDGPU_VM_PDB1, 137 AMDGPU_VM_PDB0, 138 AMDGPU_VM_PTB 139 }; 140 141 /* base structure for tracking BO usage in a VM */ 142 struct amdgpu_vm_bo_base { 143 /* constant after initialization */ 144 struct amdgpu_vm *vm; 145 struct amdgpu_bo *bo; 146 147 /* protected by bo being reserved */ 148 struct amdgpu_vm_bo_base *next; 149 150 /* protected by spinlock */ 151 struct list_head vm_status; 152 153 /* protected by the BO being reserved */ 154 bool moved; 155 }; 156 157 struct amdgpu_vm_pt { 158 struct amdgpu_vm_bo_base base; 159 160 /* array of page tables, one for each directory entry */ 161 struct amdgpu_vm_pt *entries; 162 }; 163 164 /* provided by hw blocks that can write ptes, e.g., sdma */ 165 struct amdgpu_vm_pte_funcs { 166 /* number of dw to reserve per operation */ 167 unsigned copy_pte_num_dw; 168 169 /* copy pte entries from GART */ 170 void (*copy_pte)(struct amdgpu_ib *ib, 171 uint64_t pe, uint64_t src, 172 unsigned count); 173 174 /* write pte one entry at a time with addr mapping */ 175 void (*write_pte)(struct amdgpu_ib *ib, uint64_t pe, 176 uint64_t value, unsigned count, 177 uint32_t incr); 178 /* for linear pte/pde updates without addr mapping */ 179 void (*set_pte_pde)(struct amdgpu_ib *ib, 180 uint64_t pe, 181 uint64_t addr, unsigned count, 182 uint32_t incr, uint64_t flags); 183 }; 184 185 struct amdgpu_task_info { 186 char process_name[TASK_COMM_LEN]; 187 char task_name[TASK_COMM_LEN]; 188 pid_t pid; 189 pid_t tgid; 190 }; 191 192 /** 193 * struct amdgpu_vm_update_params 194 * 195 * Encapsulate some VM table update parameters to reduce 196 * the number of function parameters 197 * 198 */ 199 struct amdgpu_vm_update_params { 200 201 /** 202 * @adev: amdgpu device we do this update for 203 */ 204 struct amdgpu_device *adev; 205 206 /** 207 * @vm: optional amdgpu_vm we do this update for 208 */ 209 struct amdgpu_vm *vm; 210 211 /** 212 * @immediate: if changes should be made immediately 213 */ 214 bool immediate; 215 216 /** 217 * @unlocked: true if the root BO is not locked 218 */ 219 bool unlocked; 220 221 /** 222 * @pages_addr: 223 * 224 * DMA addresses to use for mapping 225 */ 226 dma_addr_t *pages_addr; 227 228 /** 229 * @job: job to used for hw submission 230 */ 231 struct amdgpu_job *job; 232 233 /** 234 * @num_dw_left: number of dw left for the IB 235 */ 236 unsigned int num_dw_left; 237 }; 238 239 struct amdgpu_vm_update_funcs { 240 int (*map_table)(struct amdgpu_bo *bo); 241 int (*prepare)(struct amdgpu_vm_update_params *p, struct dma_resv *resv, 242 enum amdgpu_sync_mode sync_mode); 243 int (*update)(struct amdgpu_vm_update_params *p, 244 struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, 245 unsigned count, uint32_t incr, uint64_t flags); 246 int (*commit)(struct amdgpu_vm_update_params *p, 247 struct dma_fence **fence); 248 }; 249 250 struct amdgpu_vm { 251 /* tree of virtual addresses mapped */ 252 struct rb_root_cached va; 253 254 /* Lock to prevent eviction while we are updating page tables 255 * use vm_eviction_lock/unlock(vm) 256 */ 257 struct mutex eviction_lock; 258 bool evicting; 259 unsigned int saved_flags; 260 261 /* BOs who needs a validation */ 262 struct list_head evicted; 263 264 /* PT BOs which relocated and their parent need an update */ 265 struct list_head relocated; 266 267 /* per VM BOs moved, but not yet updated in the PT */ 268 struct list_head moved; 269 270 /* All BOs of this VM not currently in the state machine */ 271 struct list_head idle; 272 273 /* regular invalidated BOs, but not yet updated in the PT */ 274 struct list_head invalidated; 275 spinlock_t invalidated_lock; 276 277 /* BO mappings freed, but not yet updated in the PT */ 278 struct list_head freed; 279 280 /* BOs which are invalidated, has been updated in the PTs */ 281 struct list_head done; 282 283 /* contains the page directory */ 284 struct amdgpu_vm_pt root; 285 struct dma_fence *last_update; 286 287 /* Scheduler entities for page table updates */ 288 struct drm_sched_entity immediate; 289 struct drm_sched_entity delayed; 290 291 /* Last unlocked submission to the scheduler entities */ 292 struct dma_fence *last_unlocked; 293 294 unsigned int pasid; 295 /* dedicated to vm */ 296 struct amdgpu_vmid *reserved_vmid[AMDGPU_MAX_VMHUBS]; 297 298 /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */ 299 bool use_cpu_for_update; 300 301 /* Functions to use for VM table updates */ 302 const struct amdgpu_vm_update_funcs *update_funcs; 303 304 /* Flag to indicate ATS support from PTE for GFX9 */ 305 bool pte_support_ats; 306 307 /* Up to 128 pending retry page faults */ 308 DECLARE_KFIFO(faults, u64, 128); 309 310 /* Points to the KFD process VM info */ 311 struct amdkfd_process_info *process_info; 312 313 /* List node in amdkfd_process_info.vm_list_head */ 314 struct list_head vm_list_node; 315 316 /* Valid while the PD is reserved or fenced */ 317 uint64_t pd_phys_addr; 318 319 /* Some basic info about the task */ 320 struct amdgpu_task_info task_info; 321 322 /* Store positions of group of BOs */ 323 struct ttm_lru_bulk_move lru_bulk_move; 324 /* mark whether can do the bulk move */ 325 bool bulk_moveable; 326 /* Flag to indicate if VM is used for compute */ 327 bool is_compute_context; 328 }; 329 330 struct amdgpu_vm_manager { 331 /* Handling of VMIDs */ 332 struct amdgpu_vmid_mgr id_mgr[AMDGPU_MAX_VMHUBS]; 333 unsigned int first_kfd_vmid; 334 bool concurrent_flush; 335 336 /* Handling of VM fences */ 337 u64 fence_context; 338 unsigned seqno[AMDGPU_MAX_RINGS]; 339 340 uint64_t max_pfn; 341 uint32_t num_level; 342 uint32_t block_size; 343 uint32_t fragment_size; 344 enum amdgpu_vm_level root_level; 345 /* vram base address for page table entry */ 346 u64 vram_base_offset; 347 /* vm pte handling */ 348 const struct amdgpu_vm_pte_funcs *vm_pte_funcs; 349 struct drm_gpu_scheduler *vm_pte_scheds[AMDGPU_MAX_RINGS]; 350 unsigned vm_pte_num_scheds; 351 struct amdgpu_ring *page_fault; 352 353 /* partial resident texture handling */ 354 spinlock_t prt_lock; 355 atomic_t num_prt_users; 356 357 /* controls how VM page tables are updated for Graphics and Compute. 358 * BIT0[= 0] Graphics updated by SDMA [= 1] by CPU 359 * BIT1[= 0] Compute updated by SDMA [= 1] by CPU 360 */ 361 int vm_update_mode; 362 363 /* PASID to VM mapping, will be used in interrupt context to 364 * look up VM of a page fault 365 */ 366 struct idr pasid_idr; 367 spinlock_t pasid_lock; 368 }; 369 370 #define amdgpu_vm_copy_pte(adev, ib, pe, src, count) ((adev)->vm_manager.vm_pte_funcs->copy_pte((ib), (pe), (src), (count))) 371 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr))) 372 #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))) 373 374 extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs; 375 extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs; 376 377 void amdgpu_vm_manager_init(struct amdgpu_device *adev); 378 void amdgpu_vm_manager_fini(struct amdgpu_device *adev); 379 380 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout); 381 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, 382 int vm_context, u32 pasid); 383 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid); 384 void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm); 385 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm); 386 void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, 387 struct list_head *validated, 388 struct amdgpu_bo_list_entry *entry); 389 bool amdgpu_vm_ready(struct amdgpu_vm *vm); 390 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, 391 int (*callback)(void *p, struct amdgpu_bo *bo), 392 void *param); 393 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync); 394 int amdgpu_vm_update_pdes(struct amdgpu_device *adev, 395 struct amdgpu_vm *vm, bool immediate); 396 int amdgpu_vm_clear_freed(struct amdgpu_device *adev, 397 struct amdgpu_vm *vm, 398 struct dma_fence **fence); 399 int amdgpu_vm_handle_moved(struct amdgpu_device *adev, 400 struct amdgpu_vm *vm); 401 int amdgpu_vm_bo_update(struct amdgpu_device *adev, 402 struct amdgpu_bo_va *bo_va, 403 bool clear); 404 bool amdgpu_vm_evictable(struct amdgpu_bo *bo); 405 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, 406 struct amdgpu_bo *bo, bool evicted); 407 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr); 408 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, 409 struct amdgpu_bo *bo); 410 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, 411 struct amdgpu_vm *vm, 412 struct amdgpu_bo *bo); 413 int amdgpu_vm_bo_map(struct amdgpu_device *adev, 414 struct amdgpu_bo_va *bo_va, 415 uint64_t addr, uint64_t offset, 416 uint64_t size, uint64_t flags); 417 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev, 418 struct amdgpu_bo_va *bo_va, 419 uint64_t addr, uint64_t offset, 420 uint64_t size, uint64_t flags); 421 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, 422 struct amdgpu_bo_va *bo_va, 423 uint64_t addr); 424 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, 425 struct amdgpu_vm *vm, 426 uint64_t saddr, uint64_t size); 427 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, 428 uint64_t addr); 429 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket); 430 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, 431 struct amdgpu_bo_va *bo_va); 432 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size, 433 uint32_t fragment_size_default, unsigned max_level, 434 unsigned max_bits); 435 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp); 436 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, 437 struct amdgpu_job *job); 438 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev); 439 440 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid, 441 struct amdgpu_task_info *task_info); 442 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, 443 uint64_t addr); 444 445 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm); 446 447 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, 448 struct amdgpu_vm *vm); 449 void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo); 450 451 #if defined(CONFIG_DEBUG_FS) 452 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m); 453 #endif 454 455 #endif 456