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 <kgd_kfd_interface.h> 37 38 #include "amd_shared.h" 39 40 #define KFD_SYSFS_FILE_MODE 0444 41 42 #define KFD_MMAP_DOORBELL_MASK 0x8000000000000 43 #define KFD_MMAP_EVENTS_MASK 0x4000000000000 44 45 /* 46 * When working with cp scheduler we should assign the HIQ manually or via 47 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot 48 * definitions for Kaveri. In Kaveri only the first ME queues participates 49 * in the cp scheduling taking that in mind we set the HIQ slot in the 50 * second ME. 51 */ 52 #define KFD_CIK_HIQ_PIPE 4 53 #define KFD_CIK_HIQ_QUEUE 0 54 55 /* GPU ID hash width in bits */ 56 #define KFD_GPU_ID_HASH_WIDTH 16 57 58 /* Macro for allocating structures */ 59 #define kfd_alloc_struct(ptr_to_struct) \ 60 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 61 62 #define KFD_MAX_NUM_OF_PROCESSES 512 63 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 64 65 /* 66 * Kernel module parameter to specify maximum number of supported queues per 67 * device 68 */ 69 extern int max_num_of_queues_per_device; 70 71 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 72 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 73 (KFD_MAX_NUM_OF_PROCESSES * \ 74 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 75 76 #define KFD_KERNEL_QUEUE_SIZE 2048 77 78 /* Kernel module parameter to specify the scheduling policy */ 79 extern int sched_policy; 80 81 /* 82 * Kernel module parameter to specify whether to send sigterm to HSA process on 83 * unhandled exception 84 */ 85 extern int send_sigterm; 86 87 /** 88 * enum kfd_sched_policy 89 * 90 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) 91 * scheduling. In this scheduling mode we're using the firmware code to 92 * schedule the user mode queues and kernel queues such as HIQ and DIQ. 93 * the HIQ queue is used as a special queue that dispatches the configuration 94 * to the cp and the user mode queues list that are currently running. 95 * the DIQ queue is a debugging queue that dispatches debugging commands to the 96 * firmware. 97 * in this scheduling mode user mode queues over subscription feature is 98 * enabled. 99 * 100 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over 101 * subscription feature disabled. 102 * 103 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly 104 * set the command processor registers and sets the queues "manually". This 105 * mode is used *ONLY* for debugging proposes. 106 * 107 */ 108 enum kfd_sched_policy { 109 KFD_SCHED_POLICY_HWS = 0, 110 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, 111 KFD_SCHED_POLICY_NO_HWS 112 }; 113 114 enum cache_policy { 115 cache_policy_coherent, 116 cache_policy_noncoherent 117 }; 118 119 struct kfd_event_interrupt_class { 120 bool (*interrupt_isr)(struct kfd_dev *dev, 121 const uint32_t *ih_ring_entry); 122 void (*interrupt_wq)(struct kfd_dev *dev, 123 const uint32_t *ih_ring_entry); 124 }; 125 126 struct kfd_device_info { 127 enum amd_asic_type asic_family; 128 const struct kfd_event_interrupt_class *event_interrupt_class; 129 unsigned int max_pasid_bits; 130 unsigned int max_no_of_hqd; 131 size_t ih_ring_entry_size; 132 uint8_t num_of_watch_points; 133 uint16_t mqd_size_aligned; 134 }; 135 136 struct kfd_mem_obj { 137 uint32_t range_start; 138 uint32_t range_end; 139 uint64_t gpu_addr; 140 uint32_t *cpu_ptr; 141 }; 142 143 struct kfd_vmid_info { 144 uint32_t first_vmid_kfd; 145 uint32_t last_vmid_kfd; 146 uint32_t vmid_num_kfd; 147 }; 148 149 struct kfd_dev { 150 struct kgd_dev *kgd; 151 152 const struct kfd_device_info *device_info; 153 struct pci_dev *pdev; 154 155 unsigned int id; /* topology stub index */ 156 157 phys_addr_t doorbell_base; /* Start of actual doorbells used by 158 * KFD. It is aligned for mapping 159 * into user mode 160 */ 161 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 162 * to HW doorbell, GFX reserved some 163 * at the start) 164 */ 165 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 166 * page used by kernel queue 167 */ 168 169 struct kgd2kfd_shared_resources shared_resources; 170 struct kfd_vmid_info vm_info; 171 172 const struct kfd2kgd_calls *kfd2kgd; 173 struct mutex doorbell_mutex; 174 DECLARE_BITMAP(doorbell_available_index, 175 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 176 177 void *gtt_mem; 178 uint64_t gtt_start_gpu_addr; 179 void *gtt_start_cpu_ptr; 180 void *gtt_sa_bitmap; 181 struct mutex gtt_sa_lock; 182 unsigned int gtt_sa_chunk_size; 183 unsigned int gtt_sa_num_of_chunks; 184 185 /* Interrupts */ 186 struct kfifo ih_fifo; 187 struct workqueue_struct *ih_wq; 188 struct work_struct interrupt_work; 189 spinlock_t interrupt_lock; 190 191 /* QCM Device instance */ 192 struct device_queue_manager *dqm; 193 194 bool init_complete; 195 /* 196 * Interrupts of interest to KFD are copied 197 * from the HW ring into a SW ring. 198 */ 199 bool interrupts_active; 200 201 /* Debug manager */ 202 struct kfd_dbgmgr *dbgmgr; 203 }; 204 205 /* KGD2KFD callbacks */ 206 void kgd2kfd_exit(void); 207 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 208 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 209 bool kgd2kfd_device_init(struct kfd_dev *kfd, 210 const struct kgd2kfd_shared_resources *gpu_resources); 211 void kgd2kfd_device_exit(struct kfd_dev *kfd); 212 213 enum kfd_mempool { 214 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 215 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 216 KFD_MEMPOOL_FRAMEBUFFER = 3, 217 }; 218 219 /* Character device interface */ 220 int kfd_chardev_init(void); 221 void kfd_chardev_exit(void); 222 struct device *kfd_chardev(void); 223 224 /** 225 * enum kfd_unmap_queues_filter 226 * 227 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 228 * 229 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 230 * running queues list. 231 * 232 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 233 * specific process. 234 * 235 */ 236 enum kfd_unmap_queues_filter { 237 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 238 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 239 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 240 KFD_UNMAP_QUEUES_FILTER_BY_PASID 241 }; 242 243 /** 244 * enum kfd_queue_type 245 * 246 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 247 * 248 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 249 * 250 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 251 * 252 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 253 */ 254 enum kfd_queue_type { 255 KFD_QUEUE_TYPE_COMPUTE, 256 KFD_QUEUE_TYPE_SDMA, 257 KFD_QUEUE_TYPE_HIQ, 258 KFD_QUEUE_TYPE_DIQ 259 }; 260 261 enum kfd_queue_format { 262 KFD_QUEUE_FORMAT_PM4, 263 KFD_QUEUE_FORMAT_AQL 264 }; 265 266 /** 267 * struct queue_properties 268 * 269 * @type: The queue type. 270 * 271 * @queue_id: Queue identifier. 272 * 273 * @queue_address: Queue ring buffer address. 274 * 275 * @queue_size: Queue ring buffer size. 276 * 277 * @priority: Defines the queue priority relative to other queues in the 278 * process. 279 * This is just an indication and HW scheduling may override the priority as 280 * necessary while keeping the relative prioritization. 281 * the priority granularity is from 0 to f which f is the highest priority. 282 * currently all queues are initialized with the highest priority. 283 * 284 * @queue_percent: This field is partially implemented and currently a zero in 285 * this field defines that the queue is non active. 286 * 287 * @read_ptr: User space address which points to the number of dwords the 288 * cp read from the ring buffer. This field updates automatically by the H/W. 289 * 290 * @write_ptr: Defines the number of dwords written to the ring buffer. 291 * 292 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 293 * the queue ring buffer. This field should be similar to write_ptr and the 294 * user should update this field after he updated the write_ptr. 295 * 296 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 297 * 298 * @is_interop: Defines if this is a interop queue. Interop queue means that 299 * the queue can access both graphics and compute resources. 300 * 301 * @is_active: Defines if the queue is active or not. 302 * 303 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 304 * of the queue. 305 * 306 * This structure represents the queue properties for each queue no matter if 307 * it's user mode or kernel mode queue. 308 * 309 */ 310 struct queue_properties { 311 enum kfd_queue_type type; 312 enum kfd_queue_format format; 313 unsigned int queue_id; 314 uint64_t queue_address; 315 uint64_t queue_size; 316 uint32_t priority; 317 uint32_t queue_percent; 318 uint32_t *read_ptr; 319 uint32_t *write_ptr; 320 uint32_t __iomem *doorbell_ptr; 321 uint32_t doorbell_off; 322 bool is_interop; 323 bool is_active; 324 /* Not relevant for user mode queues in cp scheduling */ 325 unsigned int vmid; 326 /* Relevant only for sdma queues*/ 327 uint32_t sdma_engine_id; 328 uint32_t sdma_queue_id; 329 uint32_t sdma_vm_addr; 330 /* Relevant only for VI */ 331 uint64_t eop_ring_buffer_address; 332 uint32_t eop_ring_buffer_size; 333 uint64_t ctx_save_restore_area_address; 334 uint32_t ctx_save_restore_area_size; 335 }; 336 337 /** 338 * struct queue 339 * 340 * @list: Queue linked list. 341 * 342 * @mqd: The queue MQD. 343 * 344 * @mqd_mem_obj: The MQD local gpu memory object. 345 * 346 * @gart_mqd_addr: The MQD gart mc address. 347 * 348 * @properties: The queue properties. 349 * 350 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 351 * that the queue should be execute on. 352 * 353 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 354 * id. 355 * 356 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 357 * 358 * @process: The kfd process that created this queue. 359 * 360 * @device: The kfd device that created this queue. 361 * 362 * This structure represents user mode compute queues. 363 * It contains all the necessary data to handle such queues. 364 * 365 */ 366 367 struct queue { 368 struct list_head list; 369 void *mqd; 370 struct kfd_mem_obj *mqd_mem_obj; 371 uint64_t gart_mqd_addr; 372 struct queue_properties properties; 373 374 uint32_t mec; 375 uint32_t pipe; 376 uint32_t queue; 377 378 unsigned int sdma_id; 379 380 struct kfd_process *process; 381 struct kfd_dev *device; 382 }; 383 384 /* 385 * Please read the kfd_mqd_manager.h description. 386 */ 387 enum KFD_MQD_TYPE { 388 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 389 KFD_MQD_TYPE_HIQ, /* for hiq */ 390 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 391 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 392 KFD_MQD_TYPE_MAX 393 }; 394 395 struct scheduling_resources { 396 unsigned int vmid_mask; 397 enum kfd_queue_type type; 398 uint64_t queue_mask; 399 uint64_t gws_mask; 400 uint32_t oac_mask; 401 uint32_t gds_heap_base; 402 uint32_t gds_heap_size; 403 }; 404 405 struct process_queue_manager { 406 /* data */ 407 struct kfd_process *process; 408 struct list_head queues; 409 unsigned long *queue_slot_bitmap; 410 }; 411 412 struct qcm_process_device { 413 /* The Device Queue Manager that owns this data */ 414 struct device_queue_manager *dqm; 415 struct process_queue_manager *pqm; 416 /* Queues list */ 417 struct list_head queues_list; 418 struct list_head priv_queue_list; 419 420 unsigned int queue_count; 421 unsigned int vmid; 422 bool is_debug; 423 424 /* This flag tells if we should reset all wavefronts on 425 * process termination 426 */ 427 bool reset_wavefronts; 428 429 /* 430 * All the memory management data should be here too 431 */ 432 uint64_t gds_context_area; 433 uint32_t sh_mem_config; 434 uint32_t sh_mem_bases; 435 uint32_t sh_mem_ape1_base; 436 uint32_t sh_mem_ape1_limit; 437 uint32_t page_table_base; 438 uint32_t gds_size; 439 uint32_t num_gws; 440 uint32_t num_oac; 441 uint32_t sh_hidden_private_base; 442 }; 443 444 445 enum kfd_pdd_bound { 446 PDD_UNBOUND = 0, 447 PDD_BOUND, 448 PDD_BOUND_SUSPENDED, 449 }; 450 451 /* Data that is per-process-per device. */ 452 struct kfd_process_device { 453 /* 454 * List of all per-device data for a process. 455 * Starts from kfd_process.per_device_data. 456 */ 457 struct list_head per_device_list; 458 459 /* The device that owns this data. */ 460 struct kfd_dev *dev; 461 462 /* The process that owns this kfd_process_device. */ 463 struct kfd_process *process; 464 465 /* per-process-per device QCM data structure */ 466 struct qcm_process_device qpd; 467 468 /*Apertures*/ 469 uint64_t lds_base; 470 uint64_t lds_limit; 471 uint64_t gpuvm_base; 472 uint64_t gpuvm_limit; 473 uint64_t scratch_base; 474 uint64_t scratch_limit; 475 476 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 477 enum kfd_pdd_bound bound; 478 479 /* Flag used to tell the pdd has dequeued from the dqm. 480 * This is used to prevent dev->dqm->ops.process_termination() from 481 * being called twice when it is already called in IOMMU callback 482 * function. 483 */ 484 bool already_dequeued; 485 }; 486 487 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 488 489 /* Process data */ 490 struct kfd_process { 491 /* 492 * kfd_process are stored in an mm_struct*->kfd_process* 493 * hash table (kfd_processes in kfd_process.c) 494 */ 495 struct hlist_node kfd_processes; 496 497 /* 498 * Opaque pointer to mm_struct. We don't hold a reference to 499 * it so it should never be dereferenced from here. This is 500 * only used for looking up processes by their mm. 501 */ 502 void *mm; 503 504 struct mutex mutex; 505 506 /* 507 * In any process, the thread that started main() is the lead 508 * thread and outlives the rest. 509 * It is here because amd_iommu_bind_pasid wants a task_struct. 510 * It can also be used for safely getting a reference to the 511 * mm_struct of the process. 512 */ 513 struct task_struct *lead_thread; 514 515 /* We want to receive a notification when the mm_struct is destroyed */ 516 struct mmu_notifier mmu_notifier; 517 518 /* Use for delayed freeing of kfd_process structure */ 519 struct rcu_head rcu; 520 521 unsigned int pasid; 522 unsigned int doorbell_index; 523 524 /* 525 * List of kfd_process_device structures, 526 * one for each device the process is using. 527 */ 528 struct list_head per_device_data; 529 530 struct process_queue_manager pqm; 531 532 /*Is the user space process 32 bit?*/ 533 bool is_32bit_user_mode; 534 535 /* Event-related data */ 536 struct mutex event_mutex; 537 /* Event ID allocator and lookup */ 538 struct idr event_idr; 539 /* Event page */ 540 struct kfd_signal_page *signal_page; 541 size_t signal_mapped_size; 542 size_t signal_event_count; 543 bool signal_event_limit_reached; 544 }; 545 546 /** 547 * Ioctl function type. 548 * 549 * \param filep pointer to file structure. 550 * \param p amdkfd process pointer. 551 * \param data pointer to arg that was copied from user. 552 */ 553 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 554 void *data); 555 556 struct amdkfd_ioctl_desc { 557 unsigned int cmd; 558 int flags; 559 amdkfd_ioctl_t *func; 560 unsigned int cmd_drv; 561 const char *name; 562 }; 563 564 void kfd_process_create_wq(void); 565 void kfd_process_destroy_wq(void); 566 struct kfd_process *kfd_create_process(const struct task_struct *); 567 struct kfd_process *kfd_get_process(const struct task_struct *); 568 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 569 570 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 571 struct kfd_process *p); 572 int kfd_bind_processes_to_device(struct kfd_dev *dev); 573 void kfd_unbind_processes_from_device(struct kfd_dev *dev); 574 void kfd_process_iommu_unbind_callback(struct kfd_dev *dev, unsigned int pasid); 575 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 576 struct kfd_process *p); 577 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 578 struct kfd_process *p); 579 580 /* Process device data iterator */ 581 struct kfd_process_device *kfd_get_first_process_device_data( 582 struct kfd_process *p); 583 struct kfd_process_device *kfd_get_next_process_device_data( 584 struct kfd_process *p, 585 struct kfd_process_device *pdd); 586 bool kfd_has_process_device_data(struct kfd_process *p); 587 588 /* PASIDs */ 589 int kfd_pasid_init(void); 590 void kfd_pasid_exit(void); 591 bool kfd_set_pasid_limit(unsigned int new_limit); 592 unsigned int kfd_get_pasid_limit(void); 593 unsigned int kfd_pasid_alloc(void); 594 void kfd_pasid_free(unsigned int pasid); 595 596 /* Doorbells */ 597 int kfd_doorbell_init(struct kfd_dev *kfd); 598 void kfd_doorbell_fini(struct kfd_dev *kfd); 599 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma); 600 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 601 unsigned int *doorbell_off); 602 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 603 u32 read_kernel_doorbell(u32 __iomem *db); 604 void write_kernel_doorbell(u32 __iomem *db, u32 value); 605 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 606 struct kfd_process *process, 607 unsigned int queue_id); 608 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 609 struct kfd_process *process); 610 int kfd_alloc_process_doorbells(struct kfd_process *process); 611 void kfd_free_process_doorbells(struct kfd_process *process); 612 613 /* GTT Sub-Allocator */ 614 615 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 616 struct kfd_mem_obj **mem_obj); 617 618 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 619 620 extern struct device *kfd_device; 621 622 /* Topology */ 623 int kfd_topology_init(void); 624 void kfd_topology_shutdown(void); 625 int kfd_topology_add_device(struct kfd_dev *gpu); 626 int kfd_topology_remove_device(struct kfd_dev *gpu); 627 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 628 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 629 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx); 630 631 /* Interrupts */ 632 int kfd_interrupt_init(struct kfd_dev *dev); 633 void kfd_interrupt_exit(struct kfd_dev *dev); 634 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 635 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 636 bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry); 637 638 /* Power Management */ 639 void kgd2kfd_suspend(struct kfd_dev *kfd); 640 int kgd2kfd_resume(struct kfd_dev *kfd); 641 642 /* amdkfd Apertures */ 643 int kfd_init_apertures(struct kfd_process *process); 644 645 /* Queue Context Management */ 646 struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd); 647 648 int init_queue(struct queue **q, const struct queue_properties *properties); 649 void uninit_queue(struct queue *q); 650 void print_queue_properties(struct queue_properties *q); 651 void print_queue(struct queue *q); 652 653 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 654 struct kfd_dev *dev); 655 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 656 struct kfd_dev *dev); 657 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 658 struct kfd_dev *dev); 659 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 660 void device_queue_manager_uninit(struct device_queue_manager *dqm); 661 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 662 enum kfd_queue_type type); 663 void kernel_queue_uninit(struct kernel_queue *kq); 664 665 /* Process Queue Manager */ 666 struct process_queue_node { 667 struct queue *q; 668 struct kernel_queue *kq; 669 struct list_head process_queue_list; 670 }; 671 672 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 673 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 674 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 675 void pqm_uninit(struct process_queue_manager *pqm); 676 int pqm_create_queue(struct process_queue_manager *pqm, 677 struct kfd_dev *dev, 678 struct file *f, 679 struct queue_properties *properties, 680 unsigned int *qid); 681 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 682 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 683 struct queue_properties *p); 684 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 685 unsigned int qid); 686 687 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 688 unsigned int fence_value, 689 unsigned int timeout_ms); 690 691 /* Packet Manager */ 692 693 #define KFD_FENCE_COMPLETED (100) 694 #define KFD_FENCE_INIT (10) 695 696 struct packet_manager { 697 struct device_queue_manager *dqm; 698 struct kernel_queue *priv_queue; 699 struct mutex lock; 700 bool allocated; 701 struct kfd_mem_obj *ib_buffer_obj; 702 }; 703 704 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 705 void pm_uninit(struct packet_manager *pm); 706 int pm_send_set_resources(struct packet_manager *pm, 707 struct scheduling_resources *res); 708 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 709 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 710 uint32_t fence_value); 711 712 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 713 enum kfd_unmap_queues_filter mode, 714 uint32_t filter_param, bool reset, 715 unsigned int sdma_engine); 716 717 void pm_release_ib(struct packet_manager *pm); 718 719 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 720 721 /* Events */ 722 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 723 extern const struct kfd_device_global_init_class device_global_init_class_cik; 724 725 void kfd_event_init_process(struct kfd_process *p); 726 void kfd_event_free_process(struct kfd_process *p); 727 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 728 int kfd_wait_on_events(struct kfd_process *p, 729 uint32_t num_events, void __user *data, 730 bool all, uint32_t user_timeout_ms, 731 uint32_t *wait_result); 732 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 733 uint32_t valid_id_bits); 734 void kfd_signal_iommu_event(struct kfd_dev *dev, 735 unsigned int pasid, unsigned long address, 736 bool is_write_requested, bool is_execute_requested); 737 void kfd_signal_hw_exception_event(unsigned int pasid); 738 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 739 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 740 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 741 uint32_t event_type, bool auto_reset, uint32_t node_id, 742 uint32_t *event_id, uint32_t *event_trigger_data, 743 uint64_t *event_page_offset, uint32_t *event_slot_index); 744 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 745 746 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 747 748 #endif 749