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 #ifndef KFD_PRIV_H_INCLUDED 24 #define KFD_PRIV_H_INCLUDED 25 26 #include <linux/hashtable.h> 27 #include <linux/mmu_notifier.h> 28 #include <linux/mutex.h> 29 #include <linux/types.h> 30 #include <linux/atomic.h> 31 #include <linux/workqueue.h> 32 #include <linux/spinlock.h> 33 #include <linux/kfd_ioctl.h> 34 #include <linux/idr.h> 35 #include <linux/kfifo.h> 36 #include <linux/seq_file.h> 37 #include <linux/kref.h> 38 #include <kgd_kfd_interface.h> 39 40 #include "amd_shared.h" 41 42 #define KFD_SYSFS_FILE_MODE 0444 43 44 #define KFD_MMAP_DOORBELL_MASK 0x8000000000000ull 45 #define KFD_MMAP_EVENTS_MASK 0x4000000000000ull 46 #define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000ull 47 48 /* 49 * When working with cp scheduler we should assign the HIQ manually or via 50 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot 51 * definitions for Kaveri. In Kaveri only the first ME queues participates 52 * in the cp scheduling taking that in mind we set the HIQ slot in the 53 * second ME. 54 */ 55 #define KFD_CIK_HIQ_PIPE 4 56 #define KFD_CIK_HIQ_QUEUE 0 57 58 /* GPU ID hash width in bits */ 59 #define KFD_GPU_ID_HASH_WIDTH 16 60 61 /* Macro for allocating structures */ 62 #define kfd_alloc_struct(ptr_to_struct) \ 63 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 64 65 #define KFD_MAX_NUM_OF_PROCESSES 512 66 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 67 68 /* 69 * Size of the per-process TBA+TMA buffer: 2 pages 70 * 71 * The first page is the TBA used for the CWSR ISA code. The second 72 * page is used as TMA for daisy changing a user-mode trap handler. 73 */ 74 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 75 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 76 77 /* 78 * Kernel module parameter to specify maximum number of supported queues per 79 * device 80 */ 81 extern int max_num_of_queues_per_device; 82 83 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 84 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 85 (KFD_MAX_NUM_OF_PROCESSES * \ 86 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 87 88 #define KFD_KERNEL_QUEUE_SIZE 2048 89 90 /* Kernel module parameter to specify the scheduling policy */ 91 extern int sched_policy; 92 93 /* 94 * Kernel module parameter to specify the maximum process 95 * number per HW scheduler 96 */ 97 extern int hws_max_conc_proc; 98 99 extern int cwsr_enable; 100 101 /* 102 * Kernel module parameter to specify whether to send sigterm to HSA process on 103 * unhandled exception 104 */ 105 extern int send_sigterm; 106 107 /* 108 * Ignore CRAT table during KFD initialization, can be used to work around 109 * broken CRAT tables on some AMD systems 110 */ 111 extern int ignore_crat; 112 113 /** 114 * enum kfd_sched_policy 115 * 116 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) 117 * scheduling. In this scheduling mode we're using the firmware code to 118 * schedule the user mode queues and kernel queues such as HIQ and DIQ. 119 * the HIQ queue is used as a special queue that dispatches the configuration 120 * to the cp and the user mode queues list that are currently running. 121 * the DIQ queue is a debugging queue that dispatches debugging commands to the 122 * firmware. 123 * in this scheduling mode user mode queues over subscription feature is 124 * enabled. 125 * 126 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over 127 * subscription feature disabled. 128 * 129 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly 130 * set the command processor registers and sets the queues "manually". This 131 * mode is used *ONLY* for debugging proposes. 132 * 133 */ 134 enum kfd_sched_policy { 135 KFD_SCHED_POLICY_HWS = 0, 136 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, 137 KFD_SCHED_POLICY_NO_HWS 138 }; 139 140 enum cache_policy { 141 cache_policy_coherent, 142 cache_policy_noncoherent 143 }; 144 145 struct kfd_event_interrupt_class { 146 bool (*interrupt_isr)(struct kfd_dev *dev, 147 const uint32_t *ih_ring_entry); 148 void (*interrupt_wq)(struct kfd_dev *dev, 149 const uint32_t *ih_ring_entry); 150 }; 151 152 struct kfd_device_info { 153 enum amd_asic_type asic_family; 154 const struct kfd_event_interrupt_class *event_interrupt_class; 155 unsigned int max_pasid_bits; 156 unsigned int max_no_of_hqd; 157 size_t ih_ring_entry_size; 158 uint8_t num_of_watch_points; 159 uint16_t mqd_size_aligned; 160 bool supports_cwsr; 161 bool needs_iommu_device; 162 bool needs_pci_atomics; 163 }; 164 165 struct kfd_mem_obj { 166 uint32_t range_start; 167 uint32_t range_end; 168 uint64_t gpu_addr; 169 uint32_t *cpu_ptr; 170 }; 171 172 struct kfd_vmid_info { 173 uint32_t first_vmid_kfd; 174 uint32_t last_vmid_kfd; 175 uint32_t vmid_num_kfd; 176 }; 177 178 struct kfd_dev { 179 struct kgd_dev *kgd; 180 181 const struct kfd_device_info *device_info; 182 struct pci_dev *pdev; 183 184 unsigned int id; /* topology stub index */ 185 186 phys_addr_t doorbell_base; /* Start of actual doorbells used by 187 * KFD. It is aligned for mapping 188 * into user mode 189 */ 190 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 191 * to HW doorbell, GFX reserved some 192 * at the start) 193 */ 194 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 195 * page used by kernel queue 196 */ 197 198 struct kgd2kfd_shared_resources shared_resources; 199 struct kfd_vmid_info vm_info; 200 201 const struct kfd2kgd_calls *kfd2kgd; 202 struct mutex doorbell_mutex; 203 DECLARE_BITMAP(doorbell_available_index, 204 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 205 206 void *gtt_mem; 207 uint64_t gtt_start_gpu_addr; 208 void *gtt_start_cpu_ptr; 209 void *gtt_sa_bitmap; 210 struct mutex gtt_sa_lock; 211 unsigned int gtt_sa_chunk_size; 212 unsigned int gtt_sa_num_of_chunks; 213 214 /* Interrupts */ 215 struct kfifo ih_fifo; 216 struct workqueue_struct *ih_wq; 217 struct work_struct interrupt_work; 218 spinlock_t interrupt_lock; 219 220 /* QCM Device instance */ 221 struct device_queue_manager *dqm; 222 223 bool init_complete; 224 /* 225 * Interrupts of interest to KFD are copied 226 * from the HW ring into a SW ring. 227 */ 228 bool interrupts_active; 229 230 /* Debug manager */ 231 struct kfd_dbgmgr *dbgmgr; 232 233 /* Maximum process number mapped to HW scheduler */ 234 unsigned int max_proc_per_quantum; 235 236 /* CWSR */ 237 bool cwsr_enabled; 238 const void *cwsr_isa; 239 unsigned int cwsr_isa_size; 240 }; 241 242 /* KGD2KFD callbacks */ 243 void kgd2kfd_exit(void); 244 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 245 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 246 bool kgd2kfd_device_init(struct kfd_dev *kfd, 247 const struct kgd2kfd_shared_resources *gpu_resources); 248 void kgd2kfd_device_exit(struct kfd_dev *kfd); 249 250 enum kfd_mempool { 251 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 252 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 253 KFD_MEMPOOL_FRAMEBUFFER = 3, 254 }; 255 256 /* Character device interface */ 257 int kfd_chardev_init(void); 258 void kfd_chardev_exit(void); 259 struct device *kfd_chardev(void); 260 261 /** 262 * enum kfd_unmap_queues_filter 263 * 264 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 265 * 266 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 267 * running queues list. 268 * 269 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 270 * specific process. 271 * 272 */ 273 enum kfd_unmap_queues_filter { 274 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 275 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 276 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 277 KFD_UNMAP_QUEUES_FILTER_BY_PASID 278 }; 279 280 /** 281 * enum kfd_queue_type 282 * 283 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 284 * 285 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 286 * 287 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 288 * 289 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 290 */ 291 enum kfd_queue_type { 292 KFD_QUEUE_TYPE_COMPUTE, 293 KFD_QUEUE_TYPE_SDMA, 294 KFD_QUEUE_TYPE_HIQ, 295 KFD_QUEUE_TYPE_DIQ 296 }; 297 298 enum kfd_queue_format { 299 KFD_QUEUE_FORMAT_PM4, 300 KFD_QUEUE_FORMAT_AQL 301 }; 302 303 /** 304 * struct queue_properties 305 * 306 * @type: The queue type. 307 * 308 * @queue_id: Queue identifier. 309 * 310 * @queue_address: Queue ring buffer address. 311 * 312 * @queue_size: Queue ring buffer size. 313 * 314 * @priority: Defines the queue priority relative to other queues in the 315 * process. 316 * This is just an indication and HW scheduling may override the priority as 317 * necessary while keeping the relative prioritization. 318 * the priority granularity is from 0 to f which f is the highest priority. 319 * currently all queues are initialized with the highest priority. 320 * 321 * @queue_percent: This field is partially implemented and currently a zero in 322 * this field defines that the queue is non active. 323 * 324 * @read_ptr: User space address which points to the number of dwords the 325 * cp read from the ring buffer. This field updates automatically by the H/W. 326 * 327 * @write_ptr: Defines the number of dwords written to the ring buffer. 328 * 329 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 330 * the queue ring buffer. This field should be similar to write_ptr and the 331 * user should update this field after he updated the write_ptr. 332 * 333 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 334 * 335 * @is_interop: Defines if this is a interop queue. Interop queue means that 336 * the queue can access both graphics and compute resources. 337 * 338 * @is_evicted: Defines if the queue is evicted. Only active queues 339 * are evicted, rendering them inactive. 340 * 341 * @is_active: Defines if the queue is active or not. @is_active and 342 * @is_evicted are protected by the DQM lock. 343 * 344 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 345 * of the queue. 346 * 347 * This structure represents the queue properties for each queue no matter if 348 * it's user mode or kernel mode queue. 349 * 350 */ 351 struct queue_properties { 352 enum kfd_queue_type type; 353 enum kfd_queue_format format; 354 unsigned int queue_id; 355 uint64_t queue_address; 356 uint64_t queue_size; 357 uint32_t priority; 358 uint32_t queue_percent; 359 uint32_t *read_ptr; 360 uint32_t *write_ptr; 361 uint32_t __iomem *doorbell_ptr; 362 uint32_t doorbell_off; 363 bool is_interop; 364 bool is_evicted; 365 bool is_active; 366 /* Not relevant for user mode queues in cp scheduling */ 367 unsigned int vmid; 368 /* Relevant only for sdma queues*/ 369 uint32_t sdma_engine_id; 370 uint32_t sdma_queue_id; 371 uint32_t sdma_vm_addr; 372 /* Relevant only for VI */ 373 uint64_t eop_ring_buffer_address; 374 uint32_t eop_ring_buffer_size; 375 uint64_t ctx_save_restore_area_address; 376 uint32_t ctx_save_restore_area_size; 377 uint32_t ctl_stack_size; 378 uint64_t tba_addr; 379 uint64_t tma_addr; 380 }; 381 382 /** 383 * struct queue 384 * 385 * @list: Queue linked list. 386 * 387 * @mqd: The queue MQD. 388 * 389 * @mqd_mem_obj: The MQD local gpu memory object. 390 * 391 * @gart_mqd_addr: The MQD gart mc address. 392 * 393 * @properties: The queue properties. 394 * 395 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 396 * that the queue should be execute on. 397 * 398 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 399 * id. 400 * 401 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 402 * 403 * @process: The kfd process that created this queue. 404 * 405 * @device: The kfd device that created this queue. 406 * 407 * This structure represents user mode compute queues. 408 * It contains all the necessary data to handle such queues. 409 * 410 */ 411 412 struct queue { 413 struct list_head list; 414 void *mqd; 415 struct kfd_mem_obj *mqd_mem_obj; 416 uint64_t gart_mqd_addr; 417 struct queue_properties properties; 418 419 uint32_t mec; 420 uint32_t pipe; 421 uint32_t queue; 422 423 unsigned int sdma_id; 424 425 struct kfd_process *process; 426 struct kfd_dev *device; 427 }; 428 429 /* 430 * Please read the kfd_mqd_manager.h description. 431 */ 432 enum KFD_MQD_TYPE { 433 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 434 KFD_MQD_TYPE_HIQ, /* for hiq */ 435 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 436 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 437 KFD_MQD_TYPE_MAX 438 }; 439 440 struct scheduling_resources { 441 unsigned int vmid_mask; 442 enum kfd_queue_type type; 443 uint64_t queue_mask; 444 uint64_t gws_mask; 445 uint32_t oac_mask; 446 uint32_t gds_heap_base; 447 uint32_t gds_heap_size; 448 }; 449 450 struct process_queue_manager { 451 /* data */ 452 struct kfd_process *process; 453 struct list_head queues; 454 unsigned long *queue_slot_bitmap; 455 }; 456 457 struct qcm_process_device { 458 /* The Device Queue Manager that owns this data */ 459 struct device_queue_manager *dqm; 460 struct process_queue_manager *pqm; 461 /* Queues list */ 462 struct list_head queues_list; 463 struct list_head priv_queue_list; 464 465 unsigned int queue_count; 466 unsigned int vmid; 467 bool is_debug; 468 unsigned int evicted; /* eviction counter, 0=active */ 469 470 /* This flag tells if we should reset all wavefronts on 471 * process termination 472 */ 473 bool reset_wavefronts; 474 475 /* 476 * All the memory management data should be here too 477 */ 478 uint64_t gds_context_area; 479 uint32_t sh_mem_config; 480 uint32_t sh_mem_bases; 481 uint32_t sh_mem_ape1_base; 482 uint32_t sh_mem_ape1_limit; 483 uint32_t page_table_base; 484 uint32_t gds_size; 485 uint32_t num_gws; 486 uint32_t num_oac; 487 uint32_t sh_hidden_private_base; 488 489 /* CWSR memory */ 490 void *cwsr_kaddr; 491 uint64_t tba_addr; 492 uint64_t tma_addr; 493 }; 494 495 /* KFD Memory Eviction */ 496 497 /* Approx. wait time before attempting to restore evicted BOs */ 498 #define PROCESS_RESTORE_TIME_MS 100 499 /* Approx. back off time if restore fails due to lack of memory */ 500 #define PROCESS_BACK_OFF_TIME_MS 100 501 /* Approx. time before evicting the process again */ 502 #define PROCESS_ACTIVE_TIME_MS 10 503 504 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 505 struct dma_fence *fence); 506 507 enum kfd_pdd_bound { 508 PDD_UNBOUND = 0, 509 PDD_BOUND, 510 PDD_BOUND_SUSPENDED, 511 }; 512 513 /* Data that is per-process-per device. */ 514 struct kfd_process_device { 515 /* 516 * List of all per-device data for a process. 517 * Starts from kfd_process.per_device_data. 518 */ 519 struct list_head per_device_list; 520 521 /* The device that owns this data. */ 522 struct kfd_dev *dev; 523 524 /* The process that owns this kfd_process_device. */ 525 struct kfd_process *process; 526 527 /* per-process-per device QCM data structure */ 528 struct qcm_process_device qpd; 529 530 /*Apertures*/ 531 uint64_t lds_base; 532 uint64_t lds_limit; 533 uint64_t gpuvm_base; 534 uint64_t gpuvm_limit; 535 uint64_t scratch_base; 536 uint64_t scratch_limit; 537 538 /* VM context for GPUVM allocations */ 539 void *vm; 540 541 /* Flag used to tell the pdd has dequeued from the dqm. 542 * This is used to prevent dev->dqm->ops.process_termination() from 543 * being called twice when it is already called in IOMMU callback 544 * function. 545 */ 546 bool already_dequeued; 547 548 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 549 enum kfd_pdd_bound bound; 550 }; 551 552 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 553 554 /* Process data */ 555 struct kfd_process { 556 /* 557 * kfd_process are stored in an mm_struct*->kfd_process* 558 * hash table (kfd_processes in kfd_process.c) 559 */ 560 struct hlist_node kfd_processes; 561 562 /* 563 * Opaque pointer to mm_struct. We don't hold a reference to 564 * it so it should never be dereferenced from here. This is 565 * only used for looking up processes by their mm. 566 */ 567 void *mm; 568 569 struct kref ref; 570 struct work_struct release_work; 571 572 struct mutex mutex; 573 574 /* 575 * In any process, the thread that started main() is the lead 576 * thread and outlives the rest. 577 * It is here because amd_iommu_bind_pasid wants a task_struct. 578 * It can also be used for safely getting a reference to the 579 * mm_struct of the process. 580 */ 581 struct task_struct *lead_thread; 582 583 /* We want to receive a notification when the mm_struct is destroyed */ 584 struct mmu_notifier mmu_notifier; 585 586 /* Use for delayed freeing of kfd_process structure */ 587 struct rcu_head rcu; 588 589 unsigned int pasid; 590 unsigned int doorbell_index; 591 592 /* 593 * List of kfd_process_device structures, 594 * one for each device the process is using. 595 */ 596 struct list_head per_device_data; 597 598 struct process_queue_manager pqm; 599 600 /*Is the user space process 32 bit?*/ 601 bool is_32bit_user_mode; 602 603 /* Event-related data */ 604 struct mutex event_mutex; 605 /* Event ID allocator and lookup */ 606 struct idr event_idr; 607 /* Event page */ 608 struct kfd_signal_page *signal_page; 609 size_t signal_mapped_size; 610 size_t signal_event_count; 611 bool signal_event_limit_reached; 612 613 /* Information used for memory eviction */ 614 void *kgd_process_info; 615 /* Eviction fence that is attached to all the BOs of this process. The 616 * fence will be triggered during eviction and new one will be created 617 * during restore 618 */ 619 struct dma_fence *ef; 620 621 /* Work items for evicting and restoring BOs */ 622 struct delayed_work eviction_work; 623 struct delayed_work restore_work; 624 /* seqno of the last scheduled eviction */ 625 unsigned int last_eviction_seqno; 626 /* Approx. the last timestamp (in jiffies) when the process was 627 * restored after an eviction 628 */ 629 unsigned long last_restore_timestamp; 630 }; 631 632 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 633 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 634 extern struct srcu_struct kfd_processes_srcu; 635 636 /** 637 * Ioctl function type. 638 * 639 * \param filep pointer to file structure. 640 * \param p amdkfd process pointer. 641 * \param data pointer to arg that was copied from user. 642 */ 643 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 644 void *data); 645 646 struct amdkfd_ioctl_desc { 647 unsigned int cmd; 648 int flags; 649 amdkfd_ioctl_t *func; 650 unsigned int cmd_drv; 651 const char *name; 652 }; 653 654 void kfd_process_create_wq(void); 655 void kfd_process_destroy_wq(void); 656 struct kfd_process *kfd_create_process(struct file *filep); 657 struct kfd_process *kfd_get_process(const struct task_struct *); 658 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 659 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 660 void kfd_unref_process(struct kfd_process *p); 661 void kfd_suspend_all_processes(void); 662 int kfd_resume_all_processes(void); 663 664 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 665 struct kfd_process *p); 666 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 667 struct kfd_process *p); 668 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 669 struct kfd_process *p); 670 671 int kfd_reserved_mem_mmap(struct kfd_process *process, 672 struct vm_area_struct *vma); 673 674 /* Process device data iterator */ 675 struct kfd_process_device *kfd_get_first_process_device_data( 676 struct kfd_process *p); 677 struct kfd_process_device *kfd_get_next_process_device_data( 678 struct kfd_process *p, 679 struct kfd_process_device *pdd); 680 bool kfd_has_process_device_data(struct kfd_process *p); 681 682 /* PASIDs */ 683 int kfd_pasid_init(void); 684 void kfd_pasid_exit(void); 685 bool kfd_set_pasid_limit(unsigned int new_limit); 686 unsigned int kfd_get_pasid_limit(void); 687 unsigned int kfd_pasid_alloc(void); 688 void kfd_pasid_free(unsigned int pasid); 689 690 /* Doorbells */ 691 int kfd_doorbell_init(struct kfd_dev *kfd); 692 void kfd_doorbell_fini(struct kfd_dev *kfd); 693 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma); 694 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 695 unsigned int *doorbell_off); 696 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 697 u32 read_kernel_doorbell(u32 __iomem *db); 698 void write_kernel_doorbell(u32 __iomem *db, u32 value); 699 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 700 struct kfd_process *process, 701 unsigned int queue_id); 702 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 703 struct kfd_process *process); 704 int kfd_alloc_process_doorbells(struct kfd_process *process); 705 void kfd_free_process_doorbells(struct kfd_process *process); 706 707 /* GTT Sub-Allocator */ 708 709 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 710 struct kfd_mem_obj **mem_obj); 711 712 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 713 714 extern struct device *kfd_device; 715 716 /* Topology */ 717 int kfd_topology_init(void); 718 void kfd_topology_shutdown(void); 719 int kfd_topology_add_device(struct kfd_dev *gpu); 720 int kfd_topology_remove_device(struct kfd_dev *gpu); 721 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 722 uint32_t proximity_domain); 723 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 724 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 725 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 726 int kfd_numa_node_to_apic_id(int numa_node_id); 727 728 /* Interrupts */ 729 int kfd_interrupt_init(struct kfd_dev *dev); 730 void kfd_interrupt_exit(struct kfd_dev *dev); 731 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 732 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 733 bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry); 734 735 /* Power Management */ 736 void kgd2kfd_suspend(struct kfd_dev *kfd); 737 int kgd2kfd_resume(struct kfd_dev *kfd); 738 739 /* amdkfd Apertures */ 740 int kfd_init_apertures(struct kfd_process *process); 741 742 /* Queue Context Management */ 743 int init_queue(struct queue **q, const struct queue_properties *properties); 744 void uninit_queue(struct queue *q); 745 void print_queue_properties(struct queue_properties *q); 746 void print_queue(struct queue *q); 747 748 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 749 struct kfd_dev *dev); 750 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 751 struct kfd_dev *dev); 752 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 753 struct kfd_dev *dev); 754 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 755 struct kfd_dev *dev); 756 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 757 struct kfd_dev *dev); 758 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 759 void device_queue_manager_uninit(struct device_queue_manager *dqm); 760 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 761 enum kfd_queue_type type); 762 void kernel_queue_uninit(struct kernel_queue *kq); 763 764 /* Process Queue Manager */ 765 struct process_queue_node { 766 struct queue *q; 767 struct kernel_queue *kq; 768 struct list_head process_queue_list; 769 }; 770 771 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 772 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 773 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 774 void pqm_uninit(struct process_queue_manager *pqm); 775 int pqm_create_queue(struct process_queue_manager *pqm, 776 struct kfd_dev *dev, 777 struct file *f, 778 struct queue_properties *properties, 779 unsigned int *qid); 780 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 781 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 782 struct queue_properties *p); 783 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 784 unsigned int qid); 785 786 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 787 unsigned int fence_value, 788 unsigned int timeout_ms); 789 790 /* Packet Manager */ 791 792 #define KFD_FENCE_COMPLETED (100) 793 #define KFD_FENCE_INIT (10) 794 795 struct packet_manager { 796 struct device_queue_manager *dqm; 797 struct kernel_queue *priv_queue; 798 struct mutex lock; 799 bool allocated; 800 struct kfd_mem_obj *ib_buffer_obj; 801 unsigned int ib_size_bytes; 802 }; 803 804 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 805 void pm_uninit(struct packet_manager *pm); 806 int pm_send_set_resources(struct packet_manager *pm, 807 struct scheduling_resources *res); 808 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 809 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 810 uint32_t fence_value); 811 812 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 813 enum kfd_unmap_queues_filter mode, 814 uint32_t filter_param, bool reset, 815 unsigned int sdma_engine); 816 817 void pm_release_ib(struct packet_manager *pm); 818 819 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 820 821 /* Events */ 822 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 823 extern const struct kfd_device_global_init_class device_global_init_class_cik; 824 825 void kfd_event_init_process(struct kfd_process *p); 826 void kfd_event_free_process(struct kfd_process *p); 827 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 828 int kfd_wait_on_events(struct kfd_process *p, 829 uint32_t num_events, void __user *data, 830 bool all, uint32_t user_timeout_ms, 831 uint32_t *wait_result); 832 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 833 uint32_t valid_id_bits); 834 void kfd_signal_iommu_event(struct kfd_dev *dev, 835 unsigned int pasid, unsigned long address, 836 bool is_write_requested, bool is_execute_requested); 837 void kfd_signal_hw_exception_event(unsigned int pasid); 838 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 839 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 840 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 841 uint32_t event_type, bool auto_reset, uint32_t node_id, 842 uint32_t *event_id, uint32_t *event_trigger_data, 843 uint64_t *event_page_offset, uint32_t *event_slot_index); 844 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 845 846 void kfd_flush_tlb(struct kfd_process_device *pdd); 847 848 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 849 850 /* Debugfs */ 851 #if defined(CONFIG_DEBUG_FS) 852 853 void kfd_debugfs_init(void); 854 void kfd_debugfs_fini(void); 855 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 856 int pqm_debugfs_mqds(struct seq_file *m, void *data); 857 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 858 int dqm_debugfs_hqds(struct seq_file *m, void *data); 859 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 860 int pm_debugfs_runlist(struct seq_file *m, void *data); 861 862 #else 863 864 static inline void kfd_debugfs_init(void) {} 865 static inline void kfd_debugfs_fini(void) {} 866 867 #endif 868 869 #endif 870