1 /* 2 * Copyright 2014 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 23 /* amdgpu_amdkfd.h defines the private interface between amdgpu and amdkfd. */ 24 25 #ifndef AMDGPU_AMDKFD_H_INCLUDED 26 #define AMDGPU_AMDKFD_H_INCLUDED 27 28 #include <linux/list.h> 29 #include <linux/types.h> 30 #include <linux/mm.h> 31 #include <linux/kthread.h> 32 #include <linux/workqueue.h> 33 #include <linux/mmu_notifier.h> 34 #include <linux/memremap.h> 35 #include <kgd_kfd_interface.h> 36 #include <drm/drm_client.h> 37 #include "amdgpu_sync.h" 38 #include "amdgpu_vm.h" 39 #include "amdgpu_xcp.h" 40 41 extern uint64_t amdgpu_amdkfd_total_mem_size; 42 43 enum TLB_FLUSH_TYPE { 44 TLB_FLUSH_LEGACY = 0, 45 TLB_FLUSH_LIGHTWEIGHT, 46 TLB_FLUSH_HEAVYWEIGHT 47 }; 48 49 struct amdgpu_device; 50 struct amdgpu_reset_context; 51 52 enum kfd_mem_attachment_type { 53 KFD_MEM_ATT_SHARED, /* Share kgd_mem->bo or another attachment's */ 54 KFD_MEM_ATT_USERPTR, /* SG bo to DMA map pages from a userptr bo */ 55 KFD_MEM_ATT_DMABUF, /* DMAbuf to DMA map TTM BOs */ 56 KFD_MEM_ATT_SG /* Tag to DMA map SG BOs */ 57 }; 58 59 struct kfd_mem_attachment { 60 struct list_head list; 61 enum kfd_mem_attachment_type type; 62 bool is_mapped; 63 struct amdgpu_bo_va *bo_va; 64 struct amdgpu_device *adev; 65 uint64_t va; 66 uint64_t pte_flags; 67 }; 68 69 struct kgd_mem { 70 struct mutex lock; 71 struct amdgpu_bo *bo; 72 struct dma_buf *dmabuf; 73 struct hmm_range *range; 74 struct list_head attachments; 75 /* protected by amdkfd_process_info.lock */ 76 struct list_head validate_list; 77 uint32_t domain; 78 unsigned int mapped_to_gpu_memory; 79 uint64_t va; 80 81 uint32_t alloc_flags; 82 83 uint32_t invalid; 84 struct amdkfd_process_info *process_info; 85 86 struct amdgpu_sync sync; 87 88 uint32_t gem_handle; 89 bool aql_queue; 90 bool is_imported; 91 }; 92 93 /* KFD Memory Eviction */ 94 struct amdgpu_amdkfd_fence { 95 struct dma_fence base; 96 struct mm_struct *mm; 97 spinlock_t lock; 98 char timeline_name[TASK_COMM_LEN]; 99 struct svm_range_bo *svm_bo; 100 }; 101 102 struct amdgpu_kfd_dev { 103 struct kfd_dev *dev; 104 int64_t vram_used[MAX_XCP]; 105 uint64_t vram_used_aligned[MAX_XCP]; 106 bool init_complete; 107 struct work_struct reset_work; 108 109 /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ 110 struct dev_pagemap pgmap; 111 112 /* Client for KFD BO GEM handle allocations */ 113 struct drm_client_dev client; 114 }; 115 116 enum kgd_engine_type { 117 KGD_ENGINE_PFP = 1, 118 KGD_ENGINE_ME, 119 KGD_ENGINE_CE, 120 KGD_ENGINE_MEC1, 121 KGD_ENGINE_MEC2, 122 KGD_ENGINE_RLC, 123 KGD_ENGINE_SDMA1, 124 KGD_ENGINE_SDMA2, 125 KGD_ENGINE_MAX 126 }; 127 128 129 struct amdkfd_process_info { 130 /* List head of all VMs that belong to a KFD process */ 131 struct list_head vm_list_head; 132 /* List head for all KFD BOs that belong to a KFD process. */ 133 struct list_head kfd_bo_list; 134 /* List of userptr BOs that are valid or invalid */ 135 struct list_head userptr_valid_list; 136 struct list_head userptr_inval_list; 137 /* Lock to protect kfd_bo_list */ 138 struct mutex lock; 139 140 /* Number of VMs */ 141 unsigned int n_vms; 142 /* Eviction Fence */ 143 struct amdgpu_amdkfd_fence *eviction_fence; 144 145 /* MMU-notifier related fields */ 146 struct mutex notifier_lock; 147 uint32_t evicted_bos; 148 struct delayed_work restore_userptr_work; 149 struct pid *pid; 150 bool block_mmu_notifications; 151 }; 152 153 int amdgpu_amdkfd_init(void); 154 void amdgpu_amdkfd_fini(void); 155 156 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm); 157 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm); 158 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, 159 const void *ih_ring_entry); 160 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev); 161 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev); 162 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev); 163 int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev); 164 void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev); 165 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, 166 enum kgd_engine_type engine, 167 uint32_t vmid, uint64_t gpu_addr, 168 uint32_t *ib_cmd, uint32_t ib_len); 169 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle); 170 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev); 171 172 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid); 173 174 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev, 175 struct amdgpu_reset_context *reset_context); 176 177 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev); 178 179 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev); 180 181 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev, 182 int queue_bit); 183 184 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context, 185 struct mm_struct *mm, 186 struct svm_range_bo *svm_bo); 187 188 int amdgpu_amdkfd_drm_client_create(struct amdgpu_device *adev); 189 #if defined(CONFIG_DEBUG_FS) 190 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data); 191 #endif 192 #if IS_ENABLED(CONFIG_HSA_AMD) 193 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm); 194 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f); 195 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo); 196 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, 197 unsigned long cur_seq, struct kgd_mem *mem); 198 int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo, 199 uint32_t domain, 200 struct dma_fence *fence); 201 #else 202 static inline 203 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) 204 { 205 return false; 206 } 207 208 static inline 209 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f) 210 { 211 return NULL; 212 } 213 214 static inline 215 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo) 216 { 217 return 0; 218 } 219 220 static inline 221 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, 222 unsigned long cur_seq, struct kgd_mem *mem) 223 { 224 return 0; 225 } 226 static inline 227 int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo, 228 uint32_t domain, 229 struct dma_fence *fence) 230 { 231 return 0; 232 } 233 #endif 234 /* Shared API */ 235 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size, 236 void **mem_obj, uint64_t *gpu_addr, 237 void **cpu_ptr, bool mqd_gfx9); 238 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void **mem_obj); 239 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size, 240 void **mem_obj); 241 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj); 242 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem); 243 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem); 244 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, 245 enum kgd_engine_type type); 246 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, 247 struct kfd_local_mem_info *mem_info, 248 struct amdgpu_xcp *xcp); 249 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev); 250 251 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev); 252 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, 253 struct amdgpu_device **dmabuf_adev, 254 uint64_t *bo_size, void *metadata_buffer, 255 size_t buffer_size, uint32_t *metadata_size, 256 uint32_t *flags, int8_t *xcp_id); 257 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst, 258 struct amdgpu_device *src); 259 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst, 260 struct amdgpu_device *src, 261 bool is_min); 262 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min); 263 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, 264 uint32_t *payload); 265 int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off, 266 u32 inst); 267 int amdgpu_amdkfd_start_sched(struct amdgpu_device *adev, uint32_t node_id); 268 int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id); 269 int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id, 270 bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable); 271 bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id); 272 273 274 /* Read user wptr from a specified user address space with page fault 275 * disabled. The memory must be pinned and mapped to the hardware when 276 * this is called in hqd_load functions, so it should never fault in 277 * the first place. This resolves a circular lock dependency involving 278 * four locks, including the DQM lock and mmap_lock. 279 */ 280 #define read_user_wptr(mmptr, wptr, dst) \ 281 ({ \ 282 bool valid = false; \ 283 if ((mmptr) && (wptr)) { \ 284 pagefault_disable(); \ 285 if ((mmptr) == current->mm) { \ 286 valid = !get_user((dst), (wptr)); \ 287 } else if (current->flags & PF_KTHREAD) { \ 288 kthread_use_mm(mmptr); \ 289 valid = !get_user((dst), (wptr)); \ 290 kthread_unuse_mm(mmptr); \ 291 } \ 292 pagefault_enable(); \ 293 } \ 294 valid; \ 295 }) 296 297 /* GPUVM API */ 298 #define drm_priv_to_vm(drm_priv) \ 299 (&((struct amdgpu_fpriv *) \ 300 ((struct drm_file *)(drm_priv))->driver_priv)->vm) 301 302 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev, 303 struct amdgpu_vm *avm, u32 pasid); 304 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev, 305 struct amdgpu_vm *avm, 306 void **process_info, 307 struct dma_fence **ef); 308 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev, 309 void *drm_priv); 310 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv); 311 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, 312 uint8_t xcp_id); 313 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 314 struct amdgpu_device *adev, uint64_t va, uint64_t size, 315 void *drm_priv, struct kgd_mem **mem, 316 uint64_t *offset, uint32_t flags, bool criu_resume); 317 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( 318 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv, 319 uint64_t *size); 320 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev, 321 struct kgd_mem *mem, void *drm_priv); 322 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 323 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv); 324 int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv); 325 int amdgpu_amdkfd_gpuvm_sync_memory( 326 struct amdgpu_device *adev, struct kgd_mem *mem, bool intr); 327 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, 328 void **kptr, uint64_t *size); 329 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem); 330 331 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo, struct amdgpu_bo **bo_gart); 332 333 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info, 334 struct dma_fence __rcu **ef); 335 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev, 336 struct kfd_vm_fault_info *info); 337 int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd, 338 uint64_t va, void *drm_priv, 339 struct kgd_mem **mem, uint64_t *size, 340 uint64_t *mmap_offset); 341 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, 342 struct dma_buf **dmabuf); 343 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev); 344 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev, 345 struct tile_config *config); 346 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, 347 enum amdgpu_ras_block block, uint32_t reset); 348 349 void amdgpu_amdkfd_ras_pasid_poison_consumption_handler(struct amdgpu_device *adev, 350 enum amdgpu_ras_block block, uint16_t pasid, 351 pasid_notify pasid_fn, void *data, uint32_t reset); 352 353 bool amdgpu_amdkfd_is_fed(struct amdgpu_device *adev); 354 bool amdgpu_amdkfd_bo_mapped_to_dev(void *drm_priv, struct kgd_mem *mem); 355 void amdgpu_amdkfd_block_mmu_notifications(void *p); 356 int amdgpu_amdkfd_criu_resume(void *p); 357 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, 358 uint64_t size, u32 alloc_flag, int8_t xcp_id); 359 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, 360 uint64_t size, u32 alloc_flag, int8_t xcp_id); 361 362 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id); 363 364 #define KFD_XCP_MEM_ID(adev, xcp_id) \ 365 ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ 366 (adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1) 367 368 #define KFD_XCP_MEMORY_SIZE(adev, xcp_id) amdgpu_amdkfd_xcp_memory_size((adev), (xcp_id)) 369 370 371 #if IS_ENABLED(CONFIG_HSA_AMD) 372 void amdgpu_amdkfd_gpuvm_init_mem_limits(void); 373 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, 374 struct amdgpu_vm *vm); 375 376 /** 377 * @amdgpu_amdkfd_release_notify() - Notify KFD when GEM object is released 378 * 379 * Allows KFD to release its resources associated with the GEM object. 380 */ 381 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo); 382 void amdgpu_amdkfd_reserve_system_mem(uint64_t size); 383 #else 384 static inline 385 void amdgpu_amdkfd_gpuvm_init_mem_limits(void) 386 { 387 } 388 389 static inline 390 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, 391 struct amdgpu_vm *vm) 392 { 393 } 394 395 static inline 396 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) 397 { 398 } 399 #endif 400 401 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 402 int kgd2kfd_init_zone_device(struct amdgpu_device *adev); 403 #else 404 static inline 405 int kgd2kfd_init_zone_device(struct amdgpu_device *adev) 406 { 407 return 0; 408 } 409 #endif 410 411 /* KGD2KFD callbacks */ 412 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger); 413 int kgd2kfd_resume_mm(struct mm_struct *mm); 414 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 415 struct dma_fence *fence); 416 #if IS_ENABLED(CONFIG_HSA_AMD) 417 int kgd2kfd_init(void); 418 void kgd2kfd_exit(void); 419 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf); 420 bool kgd2kfd_device_init(struct kfd_dev *kfd, 421 const struct kgd2kfd_shared_resources *gpu_resources); 422 void kgd2kfd_device_exit(struct kfd_dev *kfd); 423 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm); 424 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm); 425 int kgd2kfd_pre_reset(struct kfd_dev *kfd, 426 struct amdgpu_reset_context *reset_context); 427 int kgd2kfd_post_reset(struct kfd_dev *kfd); 428 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 429 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd); 430 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask); 431 int kgd2kfd_check_and_lock_kfd(void); 432 void kgd2kfd_unlock_kfd(void); 433 int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t node_id); 434 int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id); 435 bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id); 436 #else 437 static inline int kgd2kfd_init(void) 438 { 439 return -ENOENT; 440 } 441 442 static inline void kgd2kfd_exit(void) 443 { 444 } 445 446 static inline 447 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) 448 { 449 return NULL; 450 } 451 452 static inline 453 bool kgd2kfd_device_init(struct kfd_dev *kfd, 454 const struct kgd2kfd_shared_resources *gpu_resources) 455 { 456 return false; 457 } 458 459 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd) 460 { 461 } 462 463 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) 464 { 465 } 466 467 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm) 468 { 469 return 0; 470 } 471 472 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd, 473 struct amdgpu_reset_context *reset_context) 474 { 475 return 0; 476 } 477 478 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd) 479 { 480 return 0; 481 } 482 483 static inline 484 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) 485 { 486 } 487 488 static inline 489 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) 490 { 491 } 492 493 static inline 494 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) 495 { 496 } 497 498 static inline int kgd2kfd_check_and_lock_kfd(void) 499 { 500 return 0; 501 } 502 503 static inline void kgd2kfd_unlock_kfd(void) 504 { 505 } 506 507 static inline int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t node_id) 508 { 509 return 0; 510 } 511 512 static inline int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id) 513 { 514 return 0; 515 } 516 517 static inline bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id) 518 { 519 return false; 520 } 521 #endif 522 #endif /* AMDGPU_AMDKFD_H_INCLUDED */ 523