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 <kgd_kfd_interface.h> 35 36#define KFD_SYSFS_FILE_MODE 0444 37 38/* 39 * When working with cp scheduler we should assign the HIQ manually or via 40 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot 41 * definitions for Kaveri. In Kaveri only the first ME queues participates 42 * in the cp scheduling taking that in mind we set the HIQ slot in the 43 * second ME. 44 */ 45#define KFD_CIK_HIQ_PIPE 4 46#define KFD_CIK_HIQ_QUEUE 0 47 48/* GPU ID hash width in bits */ 49#define KFD_GPU_ID_HASH_WIDTH 16 50 51/* Macro for allocating structures */ 52#define kfd_alloc_struct(ptr_to_struct) \ 53 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 54 55#define KFD_MAX_NUM_OF_PROCESSES 512 56#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 57 58/* 59 * Kernel module parameter to specify maximum number of supported queues per 60 * device 61 */ 62extern int max_num_of_queues_per_device; 63 64#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 65#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 66 (KFD_MAX_NUM_OF_PROCESSES * \ 67 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 68 69#define KFD_KERNEL_QUEUE_SIZE 2048 70 71/* Kernel module parameter to specify the scheduling policy */ 72extern int sched_policy; 73 74/** 75 * enum kfd_sched_policy 76 * 77 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) 78 * scheduling. In this scheduling mode we're using the firmware code to 79 * schedule the user mode queues and kernel queues such as HIQ and DIQ. 80 * the HIQ queue is used as a special queue that dispatches the configuration 81 * to the cp and the user mode queues list that are currently running. 82 * the DIQ queue is a debugging queue that dispatches debugging commands to the 83 * firmware. 84 * in this scheduling mode user mode queues over subscription feature is 85 * enabled. 86 * 87 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over 88 * subscription feature disabled. 89 * 90 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly 91 * set the command processor registers and sets the queues "manually". This 92 * mode is used *ONLY* for debugging proposes. 93 * 94 */ 95enum kfd_sched_policy { 96 KFD_SCHED_POLICY_HWS = 0, 97 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, 98 KFD_SCHED_POLICY_NO_HWS 99}; 100 101enum cache_policy { 102 cache_policy_coherent, 103 cache_policy_noncoherent 104}; 105 106enum asic_family_type { 107 CHIP_KAVERI = 0, 108 CHIP_CARRIZO 109}; 110 111struct kfd_device_info { 112 unsigned int asic_family; 113 unsigned int max_pasid_bits; 114 size_t ih_ring_entry_size; 115 uint8_t num_of_watch_points; 116 uint16_t mqd_size_aligned; 117}; 118 119struct kfd_mem_obj { 120 uint32_t range_start; 121 uint32_t range_end; 122 uint64_t gpu_addr; 123 uint32_t *cpu_ptr; 124}; 125 126struct kfd_dev { 127 struct kgd_dev *kgd; 128 129 const struct kfd_device_info *device_info; 130 struct pci_dev *pdev; 131 132 unsigned int id; /* topology stub index */ 133 134 phys_addr_t doorbell_base; /* Start of actual doorbells used by 135 * KFD. It is aligned for mapping 136 * into user mode 137 */ 138 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 139 * to HW doorbell, GFX reserved some 140 * at the start) 141 */ 142 size_t doorbell_process_limit; /* Number of processes we have doorbell 143 * space for. 144 */ 145 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 146 * page used by kernel queue 147 */ 148 149 struct kgd2kfd_shared_resources shared_resources; 150 151 const struct kfd2kgd_calls *kfd2kgd; 152 struct mutex doorbell_mutex; 153 unsigned long doorbell_available_index[DIV_ROUND_UP( 154 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)]; 155 156 void *gtt_mem; 157 uint64_t gtt_start_gpu_addr; 158 void *gtt_start_cpu_ptr; 159 void *gtt_sa_bitmap; 160 struct mutex gtt_sa_lock; 161 unsigned int gtt_sa_chunk_size; 162 unsigned int gtt_sa_num_of_chunks; 163 164 /* QCM Device instance */ 165 struct device_queue_manager *dqm; 166 167 bool init_complete; 168}; 169 170/* KGD2KFD callbacks */ 171void kgd2kfd_exit(void); 172struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 173 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 174bool kgd2kfd_device_init(struct kfd_dev *kfd, 175 const struct kgd2kfd_shared_resources *gpu_resources); 176void kgd2kfd_device_exit(struct kfd_dev *kfd); 177 178enum kfd_mempool { 179 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 180 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 181 KFD_MEMPOOL_FRAMEBUFFER = 3, 182}; 183 184/* Character device interface */ 185int kfd_chardev_init(void); 186void kfd_chardev_exit(void); 187struct device *kfd_chardev(void); 188 189/** 190 * enum kfd_preempt_type_filter 191 * 192 * @KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: Preempts single queue. 193 * 194 * @KFD_PRERMPT_TYPE_FILTER_ALL_QUEUES: Preempts all queues in the 195 * running queues list. 196 * 197 * @KFD_PRERMPT_TYPE_FILTER_BY_PASID: Preempts queues that belongs to 198 * specific process. 199 * 200 */ 201enum kfd_preempt_type_filter { 202 KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE, 203 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 204 KFD_PREEMPT_TYPE_FILTER_BY_PASID 205}; 206 207enum kfd_preempt_type { 208 KFD_PREEMPT_TYPE_WAVEFRONT, 209 KFD_PREEMPT_TYPE_WAVEFRONT_RESET 210}; 211 212/** 213 * enum kfd_queue_type 214 * 215 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 216 * 217 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 218 * 219 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 220 * 221 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 222 */ 223enum kfd_queue_type { 224 KFD_QUEUE_TYPE_COMPUTE, 225 KFD_QUEUE_TYPE_SDMA, 226 KFD_QUEUE_TYPE_HIQ, 227 KFD_QUEUE_TYPE_DIQ 228}; 229 230enum kfd_queue_format { 231 KFD_QUEUE_FORMAT_PM4, 232 KFD_QUEUE_FORMAT_AQL 233}; 234 235/** 236 * struct queue_properties 237 * 238 * @type: The queue type. 239 * 240 * @queue_id: Queue identifier. 241 * 242 * @queue_address: Queue ring buffer address. 243 * 244 * @queue_size: Queue ring buffer size. 245 * 246 * @priority: Defines the queue priority relative to other queues in the 247 * process. 248 * This is just an indication and HW scheduling may override the priority as 249 * necessary while keeping the relative prioritization. 250 * the priority granularity is from 0 to f which f is the highest priority. 251 * currently all queues are initialized with the highest priority. 252 * 253 * @queue_percent: This field is partially implemented and currently a zero in 254 * this field defines that the queue is non active. 255 * 256 * @read_ptr: User space address which points to the number of dwords the 257 * cp read from the ring buffer. This field updates automatically by the H/W. 258 * 259 * @write_ptr: Defines the number of dwords written to the ring buffer. 260 * 261 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 262 * the queue ring buffer. This field should be similar to write_ptr and the user 263 * should update this field after he updated the write_ptr. 264 * 265 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 266 * 267 * @is_interop: Defines if this is a interop queue. Interop queue means that the 268 * queue can access both graphics and compute resources. 269 * 270 * @is_active: Defines if the queue is active or not. 271 * 272 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 273 * of the queue. 274 * 275 * This structure represents the queue properties for each queue no matter if 276 * it's user mode or kernel mode queue. 277 * 278 */ 279struct queue_properties { 280 enum kfd_queue_type type; 281 enum kfd_queue_format format; 282 unsigned int queue_id; 283 uint64_t queue_address; 284 uint64_t queue_size; 285 uint32_t priority; 286 uint32_t queue_percent; 287 uint32_t *read_ptr; 288 uint32_t *write_ptr; 289 uint32_t __iomem *doorbell_ptr; 290 uint32_t doorbell_off; 291 bool is_interop; 292 bool is_active; 293 /* Not relevant for user mode queues in cp scheduling */ 294 unsigned int vmid; 295 /* Relevant only for sdma queues*/ 296 uint32_t sdma_engine_id; 297 uint32_t sdma_queue_id; 298 uint32_t sdma_vm_addr; 299 /* Relevant only for VI */ 300 uint64_t eop_ring_buffer_address; 301 uint32_t eop_ring_buffer_size; 302 uint64_t ctx_save_restore_area_address; 303 uint32_t ctx_save_restore_area_size; 304}; 305 306/** 307 * struct queue 308 * 309 * @list: Queue linked list. 310 * 311 * @mqd: The queue MQD. 312 * 313 * @mqd_mem_obj: The MQD local gpu memory object. 314 * 315 * @gart_mqd_addr: The MQD gart mc address. 316 * 317 * @properties: The queue properties. 318 * 319 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 320 * that the queue should be execute on. 321 * 322 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe id. 323 * 324 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 325 * 326 * @process: The kfd process that created this queue. 327 * 328 * @device: The kfd device that created this queue. 329 * 330 * This structure represents user mode compute queues. 331 * It contains all the necessary data to handle such queues. 332 * 333 */ 334 335struct queue { 336 struct list_head list; 337 void *mqd; 338 struct kfd_mem_obj *mqd_mem_obj; 339 uint64_t gart_mqd_addr; 340 struct queue_properties properties; 341 342 uint32_t mec; 343 uint32_t pipe; 344 uint32_t queue; 345 346 unsigned int sdma_id; 347 348 struct kfd_process *process; 349 struct kfd_dev *device; 350}; 351 352/* 353 * Please read the kfd_mqd_manager.h description. 354 */ 355enum KFD_MQD_TYPE { 356 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 357 KFD_MQD_TYPE_HIQ, /* for hiq */ 358 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 359 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 360 KFD_MQD_TYPE_MAX 361}; 362 363struct scheduling_resources { 364 unsigned int vmid_mask; 365 enum kfd_queue_type type; 366 uint64_t queue_mask; 367 uint64_t gws_mask; 368 uint32_t oac_mask; 369 uint32_t gds_heap_base; 370 uint32_t gds_heap_size; 371}; 372 373struct process_queue_manager { 374 /* data */ 375 struct kfd_process *process; 376 unsigned int num_concurrent_processes; 377 struct list_head queues; 378 unsigned long *queue_slot_bitmap; 379}; 380 381struct qcm_process_device { 382 /* The Device Queue Manager that owns this data */ 383 struct device_queue_manager *dqm; 384 struct process_queue_manager *pqm; 385 /* Queues list */ 386 struct list_head queues_list; 387 struct list_head priv_queue_list; 388 389 unsigned int queue_count; 390 unsigned int vmid; 391 bool is_debug; 392 /* 393 * All the memory management data should be here too 394 */ 395 uint64_t gds_context_area; 396 uint32_t sh_mem_config; 397 uint32_t sh_mem_bases; 398 uint32_t sh_mem_ape1_base; 399 uint32_t sh_mem_ape1_limit; 400 uint32_t page_table_base; 401 uint32_t gds_size; 402 uint32_t num_gws; 403 uint32_t num_oac; 404}; 405 406/* Data that is per-process-per device. */ 407struct kfd_process_device { 408 /* 409 * List of all per-device data for a process. 410 * Starts from kfd_process.per_device_data. 411 */ 412 struct list_head per_device_list; 413 414 /* The device that owns this data. */ 415 struct kfd_dev *dev; 416 417 418 /* per-process-per device QCM data structure */ 419 struct qcm_process_device qpd; 420 421 /*Apertures*/ 422 uint64_t lds_base; 423 uint64_t lds_limit; 424 uint64_t gpuvm_base; 425 uint64_t gpuvm_limit; 426 uint64_t scratch_base; 427 uint64_t scratch_limit; 428 429 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 430 bool bound; 431}; 432 433#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 434 435/* Process data */ 436struct kfd_process { 437 /* 438 * kfd_process are stored in an mm_struct*->kfd_process* 439 * hash table (kfd_processes in kfd_process.c) 440 */ 441 struct hlist_node kfd_processes; 442 443 struct mm_struct *mm; 444 445 struct mutex mutex; 446 447 /* 448 * In any process, the thread that started main() is the lead 449 * thread and outlives the rest. 450 * It is here because amd_iommu_bind_pasid wants a task_struct. 451 */ 452 struct task_struct *lead_thread; 453 454 /* We want to receive a notification when the mm_struct is destroyed */ 455 struct mmu_notifier mmu_notifier; 456 457 /* Use for delayed freeing of kfd_process structure */ 458 struct rcu_head rcu; 459 460 unsigned int pasid; 461 462 /* 463 * List of kfd_process_device structures, 464 * one for each device the process is using. 465 */ 466 struct list_head per_device_data; 467 468 struct process_queue_manager pqm; 469 470 /* The process's queues. */ 471 size_t queue_array_size; 472 473 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */ 474 struct kfd_queue **queues; 475 476 unsigned long allocated_queue_bitmap[DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)]; 477 478 /*Is the user space process 32 bit?*/ 479 bool is_32bit_user_mode; 480}; 481 482/** 483 * Ioctl function type. 484 * 485 * \param filep pointer to file structure. 486 * \param p amdkfd process pointer. 487 * \param data pointer to arg that was copied from user. 488 */ 489typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 490 void *data); 491 492struct amdkfd_ioctl_desc { 493 unsigned int cmd; 494 int flags; 495 amdkfd_ioctl_t *func; 496 unsigned int cmd_drv; 497 const char *name; 498}; 499 500void kfd_process_create_wq(void); 501void kfd_process_destroy_wq(void); 502struct kfd_process *kfd_create_process(const struct task_struct *); 503struct kfd_process *kfd_get_process(const struct task_struct *); 504 505struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 506 struct kfd_process *p); 507void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid); 508struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 509 struct kfd_process *p); 510struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 511 struct kfd_process *p); 512 513/* Process device data iterator */ 514struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p); 515struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p, 516 struct kfd_process_device *pdd); 517bool kfd_has_process_device_data(struct kfd_process *p); 518 519/* PASIDs */ 520int kfd_pasid_init(void); 521void kfd_pasid_exit(void); 522bool kfd_set_pasid_limit(unsigned int new_limit); 523unsigned int kfd_get_pasid_limit(void); 524unsigned int kfd_pasid_alloc(void); 525void kfd_pasid_free(unsigned int pasid); 526 527/* Doorbells */ 528void kfd_doorbell_init(struct kfd_dev *kfd); 529int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma); 530u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 531 unsigned int *doorbell_off); 532void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 533u32 read_kernel_doorbell(u32 __iomem *db); 534void write_kernel_doorbell(u32 __iomem *db, u32 value); 535unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 536 struct kfd_process *process, 537 unsigned int queue_id); 538 539/* GTT Sub-Allocator */ 540 541int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 542 struct kfd_mem_obj **mem_obj); 543 544int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 545 546extern struct device *kfd_device; 547 548/* Topology */ 549int kfd_topology_init(void); 550void kfd_topology_shutdown(void); 551int kfd_topology_add_device(struct kfd_dev *gpu); 552int kfd_topology_remove_device(struct kfd_dev *gpu); 553struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 554struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 555struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx); 556 557/* Interrupts */ 558void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 559 560/* Power Management */ 561void kgd2kfd_suspend(struct kfd_dev *kfd); 562int kgd2kfd_resume(struct kfd_dev *kfd); 563 564/* amdkfd Apertures */ 565int kfd_init_apertures(struct kfd_process *process); 566 567/* Queue Context Management */ 568inline uint32_t lower_32(uint64_t x); 569inline uint32_t upper_32(uint64_t x); 570struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd); 571inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m); 572 573int init_queue(struct queue **q, struct queue_properties properties); 574void uninit_queue(struct queue *q); 575void print_queue_properties(struct queue_properties *q); 576void print_queue(struct queue *q); 577 578struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 579 struct kfd_dev *dev); 580struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 581 struct kfd_dev *dev); 582struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 583 struct kfd_dev *dev); 584struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 585void device_queue_manager_uninit(struct device_queue_manager *dqm); 586struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 587 enum kfd_queue_type type); 588void kernel_queue_uninit(struct kernel_queue *kq); 589 590/* Process Queue Manager */ 591struct process_queue_node { 592 struct queue *q; 593 struct kernel_queue *kq; 594 struct list_head process_queue_list; 595}; 596 597int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 598void pqm_uninit(struct process_queue_manager *pqm); 599int pqm_create_queue(struct process_queue_manager *pqm, 600 struct kfd_dev *dev, 601 struct file *f, 602 struct queue_properties *properties, 603 unsigned int flags, 604 enum kfd_queue_type type, 605 unsigned int *qid); 606int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 607int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 608 struct queue_properties *p); 609 610/* Packet Manager */ 611 612#define KFD_HIQ_TIMEOUT (500) 613 614#define KFD_FENCE_COMPLETED (100) 615#define KFD_FENCE_INIT (10) 616#define KFD_UNMAP_LATENCY (150) 617 618struct packet_manager { 619 struct device_queue_manager *dqm; 620 struct kernel_queue *priv_queue; 621 struct mutex lock; 622 bool allocated; 623 struct kfd_mem_obj *ib_buffer_obj; 624}; 625 626int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 627void pm_uninit(struct packet_manager *pm); 628int pm_send_set_resources(struct packet_manager *pm, 629 struct scheduling_resources *res); 630int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 631int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 632 uint32_t fence_value); 633 634int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 635 enum kfd_preempt_type_filter mode, 636 uint32_t filter_param, bool reset, 637 unsigned int sdma_engine); 638 639void pm_release_ib(struct packet_manager *pm); 640 641uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 642phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 643 struct kfd_process *process); 644 645#endif 646