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 "amdgpu_sync.h" 37 #include "amdgpu_vm.h" 38 #include "amdgpu_xcp.h" 39 40 extern uint64_t amdgpu_amdkfd_total_mem_size; 41 42 enum TLB_FLUSH_TYPE { 43 TLB_FLUSH_LEGACY = 0, 44 TLB_FLUSH_LIGHTWEIGHT, 45 TLB_FLUSH_HEAVYWEIGHT 46 }; 47 48 struct amdgpu_device; 49 50 enum kfd_mem_attachment_type { 51 KFD_MEM_ATT_SHARED, /* Share kgd_mem->bo or another attachment's */ 52 KFD_MEM_ATT_USERPTR, /* SG bo to DMA map pages from a userptr bo */ 53 KFD_MEM_ATT_DMABUF, /* DMAbuf to DMA map TTM BOs */ 54 KFD_MEM_ATT_SG /* Tag to DMA map SG BOs */ 55 }; 56 57 struct kfd_mem_attachment { 58 struct list_head list; 59 enum kfd_mem_attachment_type type; 60 bool is_mapped; 61 struct amdgpu_bo_va *bo_va; 62 struct amdgpu_device *adev; 63 uint64_t va; 64 uint64_t pte_flags; 65 }; 66 67 struct kgd_mem { 68 struct mutex lock; 69 struct amdgpu_bo *bo; 70 struct dma_buf *dmabuf; 71 struct hmm_range *range; 72 struct list_head attachments; 73 /* protected by amdkfd_process_info.lock */ 74 struct list_head validate_list; 75 uint32_t domain; 76 unsigned int mapped_to_gpu_memory; 77 uint64_t va; 78 79 uint32_t alloc_flags; 80 81 uint32_t invalid; 82 struct amdkfd_process_info *process_info; 83 84 struct amdgpu_sync sync; 85 86 bool aql_queue; 87 bool is_imported; 88 }; 89 90 /* KFD Memory Eviction */ 91 struct amdgpu_amdkfd_fence { 92 struct dma_fence base; 93 struct mm_struct *mm; 94 spinlock_t lock; 95 char timeline_name[TASK_COMM_LEN]; 96 struct svm_range_bo *svm_bo; 97 }; 98 99 struct amdgpu_kfd_dev { 100 struct kfd_dev *dev; 101 int64_t vram_used[MAX_XCP]; 102 uint64_t vram_used_aligned[MAX_XCP]; 103 bool init_complete; 104 struct work_struct reset_work; 105 106 /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ 107 struct dev_pagemap pgmap; 108 }; 109 110 enum kgd_engine_type { 111 KGD_ENGINE_PFP = 1, 112 KGD_ENGINE_ME, 113 KGD_ENGINE_CE, 114 KGD_ENGINE_MEC1, 115 KGD_ENGINE_MEC2, 116 KGD_ENGINE_RLC, 117 KGD_ENGINE_SDMA1, 118 KGD_ENGINE_SDMA2, 119 KGD_ENGINE_MAX 120 }; 121 122 123 struct amdkfd_process_info { 124 /* List head of all VMs that belong to a KFD process */ 125 struct list_head vm_list_head; 126 /* List head for all KFD BOs that belong to a KFD process. */ 127 struct list_head kfd_bo_list; 128 /* List of userptr BOs that are valid or invalid */ 129 struct list_head userptr_valid_list; 130 struct list_head userptr_inval_list; 131 /* Lock to protect kfd_bo_list */ 132 struct mutex lock; 133 134 /* Number of VMs */ 135 unsigned int n_vms; 136 /* Eviction Fence */ 137 struct amdgpu_amdkfd_fence *eviction_fence; 138 139 /* MMU-notifier related fields */ 140 struct mutex notifier_lock; 141 uint32_t evicted_bos; 142 struct delayed_work restore_userptr_work; 143 struct pid *pid; 144 bool block_mmu_notifications; 145 }; 146 147 int amdgpu_amdkfd_init(void); 148 void amdgpu_amdkfd_fini(void); 149 150 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm); 151 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm); 152 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, 153 const void *ih_ring_entry); 154 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev); 155 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev); 156 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev); 157 int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev); 158 void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev); 159 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, 160 enum kgd_engine_type engine, 161 uint32_t vmid, uint64_t gpu_addr, 162 uint32_t *ib_cmd, uint32_t ib_len); 163 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle); 164 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev); 165 166 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid); 167 168 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev); 169 170 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev); 171 172 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev); 173 174 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev, 175 int queue_bit); 176 177 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context, 178 struct mm_struct *mm, 179 struct svm_range_bo *svm_bo); 180 #if defined(CONFIG_DEBUG_FS) 181 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data); 182 #endif 183 #if IS_ENABLED(CONFIG_HSA_AMD) 184 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm); 185 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f); 186 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo); 187 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, 188 unsigned long cur_seq, struct kgd_mem *mem); 189 #else 190 static inline 191 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) 192 { 193 return false; 194 } 195 196 static inline 197 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f) 198 { 199 return NULL; 200 } 201 202 static inline 203 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo) 204 { 205 return 0; 206 } 207 208 static inline 209 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, 210 unsigned long cur_seq, struct kgd_mem *mem) 211 { 212 return 0; 213 } 214 #endif 215 /* Shared API */ 216 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size, 217 void **mem_obj, uint64_t *gpu_addr, 218 void **cpu_ptr, bool mqd_gfx9); 219 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj); 220 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size, 221 void **mem_obj); 222 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj); 223 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem); 224 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem); 225 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, 226 enum kgd_engine_type type); 227 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, 228 struct kfd_local_mem_info *mem_info, 229 struct amdgpu_xcp *xcp); 230 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev); 231 232 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev); 233 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, 234 struct amdgpu_device **dmabuf_adev, 235 uint64_t *bo_size, void *metadata_buffer, 236 size_t buffer_size, uint32_t *metadata_size, 237 uint32_t *flags, int8_t *xcp_id); 238 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst, 239 struct amdgpu_device *src); 240 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst, 241 struct amdgpu_device *src, 242 bool is_min); 243 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min); 244 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, 245 uint32_t *payload); 246 int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off, 247 u32 inst); 248 249 /* Read user wptr from a specified user address space with page fault 250 * disabled. The memory must be pinned and mapped to the hardware when 251 * this is called in hqd_load functions, so it should never fault in 252 * the first place. This resolves a circular lock dependency involving 253 * four locks, including the DQM lock and mmap_lock. 254 */ 255 #define read_user_wptr(mmptr, wptr, dst) \ 256 ({ \ 257 bool valid = false; \ 258 if ((mmptr) && (wptr)) { \ 259 pagefault_disable(); \ 260 if ((mmptr) == current->mm) { \ 261 valid = !get_user((dst), (wptr)); \ 262 } else if (current->flags & PF_KTHREAD) { \ 263 kthread_use_mm(mmptr); \ 264 valid = !get_user((dst), (wptr)); \ 265 kthread_unuse_mm(mmptr); \ 266 } \ 267 pagefault_enable(); \ 268 } \ 269 valid; \ 270 }) 271 272 /* GPUVM API */ 273 #define drm_priv_to_vm(drm_priv) \ 274 (&((struct amdgpu_fpriv *) \ 275 ((struct drm_file *)(drm_priv))->driver_priv)->vm) 276 277 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev, 278 struct amdgpu_vm *avm, u32 pasid); 279 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev, 280 struct amdgpu_vm *avm, 281 void **process_info, 282 struct dma_fence **ef); 283 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev, 284 void *drm_priv); 285 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv); 286 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, 287 uint8_t xcp_id); 288 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 289 struct amdgpu_device *adev, uint64_t va, uint64_t size, 290 void *drm_priv, struct kgd_mem **mem, 291 uint64_t *offset, uint32_t flags, bool criu_resume); 292 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( 293 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv, 294 uint64_t *size); 295 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev, 296 struct kgd_mem *mem, void *drm_priv); 297 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 298 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv); 299 void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv); 300 int amdgpu_amdkfd_gpuvm_sync_memory( 301 struct amdgpu_device *adev, struct kgd_mem *mem, bool intr); 302 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, 303 void **kptr, uint64_t *size); 304 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem); 305 306 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo); 307 308 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info, 309 struct dma_fence **ef); 310 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev, 311 struct kfd_vm_fault_info *info); 312 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, 313 struct dma_buf *dmabuf, 314 uint64_t va, void *drm_priv, 315 struct kgd_mem **mem, uint64_t *size, 316 uint64_t *mmap_offset); 317 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, 318 struct dma_buf **dmabuf); 319 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev); 320 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev, 321 struct tile_config *config); 322 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, 323 bool reset); 324 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem); 325 void amdgpu_amdkfd_block_mmu_notifications(void *p); 326 int amdgpu_amdkfd_criu_resume(void *p); 327 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev); 328 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, 329 uint64_t size, u32 alloc_flag, int8_t xcp_id); 330 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, 331 uint64_t size, u32 alloc_flag, int8_t xcp_id); 332 333 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id); 334 335 #define KFD_XCP_MEM_ID(adev, xcp_id) \ 336 ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ 337 (adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1) 338 339 #define KFD_XCP_MEMORY_SIZE(adev, xcp_id) amdgpu_amdkfd_xcp_memory_size((adev), (xcp_id)) 340 341 342 #if IS_ENABLED(CONFIG_HSA_AMD) 343 void amdgpu_amdkfd_gpuvm_init_mem_limits(void); 344 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, 345 struct amdgpu_vm *vm); 346 347 /** 348 * @amdgpu_amdkfd_release_notify() - Notify KFD when GEM object is released 349 * 350 * Allows KFD to release its resources associated with the GEM object. 351 */ 352 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo); 353 void amdgpu_amdkfd_reserve_system_mem(uint64_t size); 354 #else 355 static inline 356 void amdgpu_amdkfd_gpuvm_init_mem_limits(void) 357 { 358 } 359 360 static inline 361 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, 362 struct amdgpu_vm *vm) 363 { 364 } 365 366 static inline 367 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) 368 { 369 } 370 #endif 371 372 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 373 int kgd2kfd_init_zone_device(struct amdgpu_device *adev); 374 #else 375 static inline 376 int kgd2kfd_init_zone_device(struct amdgpu_device *adev) 377 { 378 return 0; 379 } 380 #endif 381 382 /* KGD2KFD callbacks */ 383 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger); 384 int kgd2kfd_resume_mm(struct mm_struct *mm); 385 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 386 struct dma_fence *fence); 387 #if IS_ENABLED(CONFIG_HSA_AMD) 388 int kgd2kfd_init(void); 389 void kgd2kfd_exit(void); 390 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf); 391 bool kgd2kfd_device_init(struct kfd_dev *kfd, 392 const struct kgd2kfd_shared_resources *gpu_resources); 393 void kgd2kfd_device_exit(struct kfd_dev *kfd); 394 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm); 395 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm); 396 int kgd2kfd_pre_reset(struct kfd_dev *kfd); 397 int kgd2kfd_post_reset(struct kfd_dev *kfd); 398 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 399 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd); 400 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask); 401 int kgd2kfd_check_and_lock_kfd(void); 402 void kgd2kfd_unlock_kfd(void); 403 #else 404 static inline int kgd2kfd_init(void) 405 { 406 return -ENOENT; 407 } 408 409 static inline void kgd2kfd_exit(void) 410 { 411 } 412 413 static inline 414 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) 415 { 416 return NULL; 417 } 418 419 static inline 420 bool kgd2kfd_device_init(struct kfd_dev *kfd, 421 const struct kgd2kfd_shared_resources *gpu_resources) 422 { 423 return false; 424 } 425 426 static inline void kgd2kfd_device_exit(struct kfd_dev *kfd) 427 { 428 } 429 430 static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) 431 { 432 } 433 434 static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm) 435 { 436 return 0; 437 } 438 439 static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd) 440 { 441 return 0; 442 } 443 444 static inline int kgd2kfd_post_reset(struct kfd_dev *kfd) 445 { 446 return 0; 447 } 448 449 static inline 450 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) 451 { 452 } 453 454 static inline 455 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) 456 { 457 } 458 459 static inline 460 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) 461 { 462 } 463 464 static inline int kgd2kfd_check_and_lock_kfd(void) 465 { 466 return 0; 467 } 468 469 static inline void kgd2kfd_unlock_kfd(void) 470 { 471 } 472 #endif 473 #endif /* AMDGPU_AMDKFD_H_INCLUDED */ 474