1 /* 2 * Copyright 2019 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 24 #ifndef __AMDGPU_MES_H__ 25 #define __AMDGPU_MES_H__ 26 27 #include "amdgpu_irq.h" 28 #include "kgd_kfd_interface.h" 29 #include "amdgpu_gfx.h" 30 #include <linux/sched/mm.h> 31 32 #define AMDGPU_MES_MAX_COMPUTE_PIPES 8 33 #define AMDGPU_MES_MAX_GFX_PIPES 2 34 #define AMDGPU_MES_MAX_SDMA_PIPES 2 35 36 #define AMDGPU_MES_API_VERSION_SHIFT 12 37 #define AMDGPU_MES_FEAT_VERSION_SHIFT 24 38 39 #define AMDGPU_MES_VERSION_MASK 0x00000fff 40 #define AMDGPU_MES_API_VERSION_MASK 0x00fff000 41 #define AMDGPU_MES_FEAT_VERSION_MASK 0xff000000 42 43 enum amdgpu_mes_priority_level { 44 AMDGPU_MES_PRIORITY_LEVEL_LOW = 0, 45 AMDGPU_MES_PRIORITY_LEVEL_NORMAL = 1, 46 AMDGPU_MES_PRIORITY_LEVEL_MEDIUM = 2, 47 AMDGPU_MES_PRIORITY_LEVEL_HIGH = 3, 48 AMDGPU_MES_PRIORITY_LEVEL_REALTIME = 4, 49 AMDGPU_MES_PRIORITY_NUM_LEVELS 50 }; 51 52 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */ 53 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */ 54 55 struct amdgpu_mes_funcs; 56 57 enum admgpu_mes_pipe { 58 AMDGPU_MES_SCHED_PIPE = 0, 59 AMDGPU_MES_KIQ_PIPE, 60 AMDGPU_MAX_MES_PIPES = 2, 61 }; 62 63 struct amdgpu_mes { 64 struct amdgpu_device *adev; 65 66 struct mutex mutex_hidden; 67 68 struct idr pasid_idr; 69 struct idr gang_id_idr; 70 struct idr queue_id_idr; 71 struct ida doorbell_ida; 72 73 spinlock_t queue_id_lock; 74 75 uint32_t sched_version; 76 uint32_t kiq_version; 77 78 uint32_t total_max_queue; 79 uint32_t doorbell_id_offset; 80 uint32_t max_doorbell_slices; 81 82 uint64_t default_process_quantum; 83 uint64_t default_gang_quantum; 84 85 struct amdgpu_ring ring; 86 spinlock_t ring_lock; 87 88 const struct firmware *fw[AMDGPU_MAX_MES_PIPES]; 89 90 /* mes ucode */ 91 struct amdgpu_bo *ucode_fw_obj[AMDGPU_MAX_MES_PIPES]; 92 uint64_t ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES]; 93 uint32_t *ucode_fw_ptr[AMDGPU_MAX_MES_PIPES]; 94 uint64_t uc_start_addr[AMDGPU_MAX_MES_PIPES]; 95 96 /* mes ucode data */ 97 struct amdgpu_bo *data_fw_obj[AMDGPU_MAX_MES_PIPES]; 98 uint64_t data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES]; 99 uint32_t *data_fw_ptr[AMDGPU_MAX_MES_PIPES]; 100 uint64_t data_start_addr[AMDGPU_MAX_MES_PIPES]; 101 102 /* eop gpu obj */ 103 struct amdgpu_bo *eop_gpu_obj[AMDGPU_MAX_MES_PIPES]; 104 uint64_t eop_gpu_addr[AMDGPU_MAX_MES_PIPES]; 105 106 void *mqd_backup[AMDGPU_MAX_MES_PIPES]; 107 struct amdgpu_irq_src irq[AMDGPU_MAX_MES_PIPES]; 108 109 uint32_t vmid_mask_gfxhub; 110 uint32_t vmid_mask_mmhub; 111 uint32_t compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES]; 112 uint32_t gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES]; 113 uint32_t sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES]; 114 uint32_t aggregated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS]; 115 uint32_t sch_ctx_offs; 116 uint64_t sch_ctx_gpu_addr; 117 uint64_t *sch_ctx_ptr; 118 uint32_t query_status_fence_offs; 119 uint64_t query_status_fence_gpu_addr; 120 uint64_t *query_status_fence_ptr; 121 uint32_t read_val_offs; 122 uint64_t read_val_gpu_addr; 123 uint32_t *read_val_ptr; 124 125 uint32_t saved_flags; 126 127 /* initialize kiq pipe */ 128 int (*kiq_hw_init)(struct amdgpu_device *adev); 129 int (*kiq_hw_fini)(struct amdgpu_device *adev); 130 131 /* ip specific functions */ 132 const struct amdgpu_mes_funcs *funcs; 133 }; 134 135 struct amdgpu_mes_process { 136 int pasid; 137 struct amdgpu_vm *vm; 138 uint64_t pd_gpu_addr; 139 struct amdgpu_bo *proc_ctx_bo; 140 uint64_t proc_ctx_gpu_addr; 141 void *proc_ctx_cpu_ptr; 142 uint64_t process_quantum; 143 struct list_head gang_list; 144 uint32_t doorbell_index; 145 unsigned long *doorbell_bitmap; 146 struct mutex doorbell_lock; 147 }; 148 149 struct amdgpu_mes_gang { 150 int gang_id; 151 int priority; 152 int inprocess_gang_priority; 153 int global_priority_level; 154 struct list_head list; 155 struct amdgpu_mes_process *process; 156 struct amdgpu_bo *gang_ctx_bo; 157 uint64_t gang_ctx_gpu_addr; 158 void *gang_ctx_cpu_ptr; 159 uint64_t gang_quantum; 160 struct list_head queue_list; 161 }; 162 163 struct amdgpu_mes_queue { 164 struct list_head list; 165 struct amdgpu_mes_gang *gang; 166 int queue_id; 167 uint64_t doorbell_off; 168 struct amdgpu_bo *mqd_obj; 169 void *mqd_cpu_ptr; 170 uint64_t mqd_gpu_addr; 171 uint64_t wptr_gpu_addr; 172 int queue_type; 173 int paging; 174 struct amdgpu_ring *ring; 175 }; 176 177 struct amdgpu_mes_queue_properties { 178 int queue_type; 179 uint64_t hqd_base_gpu_addr; 180 uint64_t rptr_gpu_addr; 181 uint64_t wptr_gpu_addr; 182 uint64_t wptr_mc_addr; 183 uint32_t queue_size; 184 uint64_t eop_gpu_addr; 185 uint32_t hqd_pipe_priority; 186 uint32_t hqd_queue_priority; 187 bool paging; 188 struct amdgpu_ring *ring; 189 /* out */ 190 uint64_t doorbell_off; 191 }; 192 193 struct amdgpu_mes_gang_properties { 194 uint32_t priority; 195 uint32_t gang_quantum; 196 uint32_t inprocess_gang_priority; 197 uint32_t priority_level; 198 int global_priority_level; 199 }; 200 201 struct mes_add_queue_input { 202 uint32_t process_id; 203 uint64_t page_table_base_addr; 204 uint64_t process_va_start; 205 uint64_t process_va_end; 206 uint64_t process_quantum; 207 uint64_t process_context_addr; 208 uint64_t gang_quantum; 209 uint64_t gang_context_addr; 210 uint32_t inprocess_gang_priority; 211 uint32_t gang_global_priority_level; 212 uint32_t doorbell_offset; 213 uint64_t mqd_addr; 214 uint64_t wptr_addr; 215 uint64_t wptr_mc_addr; 216 uint32_t queue_type; 217 uint32_t paging; 218 uint32_t gws_base; 219 uint32_t gws_size; 220 uint64_t tba_addr; 221 uint64_t tma_addr; 222 uint32_t trap_en; 223 uint32_t skip_process_ctx_clear; 224 uint32_t is_kfd_process; 225 uint32_t is_aql_queue; 226 uint32_t queue_size; 227 }; 228 229 struct mes_remove_queue_input { 230 uint32_t doorbell_offset; 231 uint64_t gang_context_addr; 232 }; 233 234 struct mes_unmap_legacy_queue_input { 235 enum amdgpu_unmap_queues_action action; 236 uint32_t queue_type; 237 uint32_t doorbell_offset; 238 uint32_t pipe_id; 239 uint32_t queue_id; 240 uint64_t trail_fence_addr; 241 uint64_t trail_fence_data; 242 }; 243 244 struct mes_suspend_gang_input { 245 bool suspend_all_gangs; 246 uint64_t gang_context_addr; 247 uint64_t suspend_fence_addr; 248 uint32_t suspend_fence_value; 249 }; 250 251 struct mes_resume_gang_input { 252 bool resume_all_gangs; 253 uint64_t gang_context_addr; 254 }; 255 256 enum mes_misc_opcode { 257 MES_MISC_OP_WRITE_REG, 258 MES_MISC_OP_READ_REG, 259 MES_MISC_OP_WRM_REG_WAIT, 260 MES_MISC_OP_WRM_REG_WR_WAIT, 261 MES_MISC_OP_SET_SHADER_DEBUGGER, 262 }; 263 264 struct mes_misc_op_input { 265 enum mes_misc_opcode op; 266 267 union { 268 struct { 269 uint32_t reg_offset; 270 uint64_t buffer_addr; 271 } read_reg; 272 273 struct { 274 uint32_t reg_offset; 275 uint32_t reg_value; 276 } write_reg; 277 278 struct { 279 uint32_t ref; 280 uint32_t mask; 281 uint32_t reg0; 282 uint32_t reg1; 283 } wrm_reg; 284 285 struct { 286 uint64_t process_context_addr; 287 union { 288 struct { 289 uint64_t single_memop : 1; 290 uint64_t single_alu_op : 1; 291 uint64_t reserved: 30; 292 }; 293 uint32_t u32all; 294 } flags; 295 uint32_t spi_gdbg_per_vmid_cntl; 296 uint32_t tcp_watch_cntl[4]; 297 uint32_t trap_en; 298 } set_shader_debugger; 299 }; 300 }; 301 302 struct amdgpu_mes_funcs { 303 int (*add_hw_queue)(struct amdgpu_mes *mes, 304 struct mes_add_queue_input *input); 305 306 int (*remove_hw_queue)(struct amdgpu_mes *mes, 307 struct mes_remove_queue_input *input); 308 309 int (*unmap_legacy_queue)(struct amdgpu_mes *mes, 310 struct mes_unmap_legacy_queue_input *input); 311 312 int (*suspend_gang)(struct amdgpu_mes *mes, 313 struct mes_suspend_gang_input *input); 314 315 int (*resume_gang)(struct amdgpu_mes *mes, 316 struct mes_resume_gang_input *input); 317 318 int (*misc_op)(struct amdgpu_mes *mes, 319 struct mes_misc_op_input *input); 320 }; 321 322 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev)) 323 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev)) 324 325 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs); 326 327 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe); 328 int amdgpu_mes_init(struct amdgpu_device *adev); 329 void amdgpu_mes_fini(struct amdgpu_device *adev); 330 331 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid, 332 struct amdgpu_vm *vm); 333 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid); 334 335 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid, 336 struct amdgpu_mes_gang_properties *gprops, 337 int *gang_id); 338 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id); 339 340 int amdgpu_mes_suspend(struct amdgpu_device *adev); 341 int amdgpu_mes_resume(struct amdgpu_device *adev); 342 343 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id, 344 struct amdgpu_mes_queue_properties *qprops, 345 int *queue_id); 346 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id); 347 348 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev, 349 struct amdgpu_ring *ring, 350 enum amdgpu_unmap_queues_action action, 351 u64 gpu_addr, u64 seq); 352 353 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg); 354 int amdgpu_mes_wreg(struct amdgpu_device *adev, 355 uint32_t reg, uint32_t val); 356 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg, 357 uint32_t val, uint32_t mask); 358 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev, 359 uint32_t reg0, uint32_t reg1, 360 uint32_t ref, uint32_t mask); 361 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev, 362 uint64_t process_context_addr, 363 uint32_t spi_gdbg_per_vmid_cntl, 364 const uint32_t *tcp_watch_cntl, 365 uint32_t flags, 366 bool trap_en); 367 368 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id, 369 int queue_type, int idx, 370 struct amdgpu_mes_ctx_data *ctx_data, 371 struct amdgpu_ring **out); 372 void amdgpu_mes_remove_ring(struct amdgpu_device *adev, 373 struct amdgpu_ring *ring); 374 375 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev, 376 enum amdgpu_mes_priority_level prio); 377 378 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev, 379 struct amdgpu_mes_ctx_data *ctx_data); 380 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data); 381 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev, 382 struct amdgpu_vm *vm, 383 struct amdgpu_mes_ctx_data *ctx_data); 384 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev, 385 struct amdgpu_mes_ctx_data *ctx_data); 386 387 int amdgpu_mes_self_test(struct amdgpu_device *adev); 388 389 int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev, 390 unsigned int *doorbell_index); 391 void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev, 392 unsigned int doorbell_index); 393 unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar( 394 struct amdgpu_device *adev, 395 uint32_t doorbell_index, 396 unsigned int doorbell_id); 397 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev); 398 399 /* 400 * MES lock can be taken in MMU notifiers. 401 * 402 * A bit more detail about why to set no-FS reclaim with MES lock: 403 * 404 * The purpose of the MMU notifier is to stop GPU access to memory so 405 * that the Linux VM subsystem can move pages around safely. This is 406 * done by preempting user mode queues for the affected process. When 407 * MES is used, MES lock needs to be taken to preempt the queues. 408 * 409 * The MMU notifier callback entry point in the driver is 410 * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from 411 * there is: 412 * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm -> 413 * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues 414 * 415 * The last part of the chain is a function pointer where we take the 416 * MES lock. 417 * 418 * The problem with taking locks in the MMU notifier is, that MMU 419 * notifiers can be called in reclaim-FS context. That's where the 420 * kernel frees up pages to make room for new page allocations under 421 * memory pressure. While we are running in reclaim-FS context, we must 422 * not trigger another memory reclaim operation because that would 423 * recursively reenter the reclaim code and cause a deadlock. The 424 * memalloc_nofs_save/restore calls guarantee that. 425 * 426 * In addition we also need to avoid lock dependencies on other locks taken 427 * under the MES lock, for example reservation locks. Here is a possible 428 * scenario of a deadlock: 429 * Thread A: takes and holds reservation lock | triggers reclaim-FS | 430 * MMU notifier | blocks trying to take MES lock 431 * Thread B: takes and holds MES lock | blocks trying to take reservation lock 432 * 433 * In this scenario Thread B gets involved in a deadlock even without 434 * triggering a reclaim-FS operation itself. 435 * To fix this and break the lock dependency chain you'd need to either: 436 * 1. protect reservation locks with memalloc_nofs_save/restore, or 437 * 2. avoid taking reservation locks under the MES lock. 438 * 439 * Reservation locks are taken all over the kernel in different subsystems, we 440 * have no control over them and their lock dependencies.So the only workable 441 * solution is to avoid taking other locks under the MES lock. 442 * As a result, make sure no reclaim-FS happens while holding this lock anywhere 443 * to prevent deadlocks when an MMU notifier runs in reclaim-FS context. 444 */ 445 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes) 446 { 447 mutex_lock(&mes->mutex_hidden); 448 mes->saved_flags = memalloc_noreclaim_save(); 449 } 450 451 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes) 452 { 453 memalloc_noreclaim_restore(mes->saved_flags); 454 mutex_unlock(&mes->mutex_hidden); 455 } 456 #endif /* __AMDGPU_MES_H__ */ 457