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_MAX_RING_ENTRY_SIZE 8 43 44 #define KFD_SYSFS_FILE_MODE 0444 45 46 /* GPU ID hash width in bits */ 47 #define KFD_GPU_ID_HASH_WIDTH 16 48 49 /* Use upper bits of mmap offset to store KFD driver specific information. 50 * BITS[63:62] - Encode MMAP type 51 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 52 * BITS[45:0] - MMAP offset value 53 * 54 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 55 * defines are w.r.t to PAGE_SIZE 56 */ 57 #define KFD_MMAP_TYPE_SHIFT (62 - PAGE_SHIFT) 58 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 59 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 60 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 61 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 62 63 #define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT) 64 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 65 << KFD_MMAP_GPU_ID_SHIFT) 66 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 67 & KFD_MMAP_GPU_ID_MASK) 68 #define KFD_MMAP_GPU_ID_GET(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 69 >> KFD_MMAP_GPU_ID_SHIFT) 70 71 #define KFD_MMAP_OFFSET_VALUE_MASK (0x3FFFFFFFFFFFULL >> PAGE_SHIFT) 72 #define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK) 73 74 /* 75 * When working with cp scheduler we should assign the HIQ manually or via 76 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot 77 * definitions for Kaveri. In Kaveri only the first ME queues participates 78 * in the cp scheduling taking that in mind we set the HIQ slot in the 79 * second ME. 80 */ 81 #define KFD_CIK_HIQ_PIPE 4 82 #define KFD_CIK_HIQ_QUEUE 0 83 84 /* Macro for allocating structures */ 85 #define kfd_alloc_struct(ptr_to_struct) \ 86 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 87 88 #define KFD_MAX_NUM_OF_PROCESSES 512 89 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 90 91 /* 92 * Size of the per-process TBA+TMA buffer: 2 pages 93 * 94 * The first page is the TBA used for the CWSR ISA code. The second 95 * page is used as TMA for daisy changing a user-mode trap handler. 96 */ 97 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 98 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 99 100 /* 101 * Kernel module parameter to specify maximum number of supported queues per 102 * device 103 */ 104 extern int max_num_of_queues_per_device; 105 106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 107 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 108 (KFD_MAX_NUM_OF_PROCESSES * \ 109 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 110 111 #define KFD_KERNEL_QUEUE_SIZE 2048 112 113 /* Kernel module parameter to specify the scheduling policy */ 114 extern int sched_policy; 115 116 /* 117 * Kernel module parameter to specify the maximum process 118 * number per HW scheduler 119 */ 120 extern int hws_max_conc_proc; 121 122 extern int cwsr_enable; 123 124 /* 125 * Kernel module parameter to specify whether to send sigterm to HSA process on 126 * unhandled exception 127 */ 128 extern int send_sigterm; 129 130 /* 131 * This kernel module is used to simulate large bar machine on non-large bar 132 * enabled machines. 133 */ 134 extern int debug_largebar; 135 136 /* 137 * Ignore CRAT table during KFD initialization, can be used to work around 138 * broken CRAT tables on some AMD systems 139 */ 140 extern int ignore_crat; 141 142 /* 143 * Set sh_mem_config.retry_disable on Vega10 144 */ 145 extern int vega10_noretry; 146 147 /** 148 * enum kfd_sched_policy 149 * 150 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) 151 * scheduling. In this scheduling mode we're using the firmware code to 152 * schedule the user mode queues and kernel queues such as HIQ and DIQ. 153 * the HIQ queue is used as a special queue that dispatches the configuration 154 * to the cp and the user mode queues list that are currently running. 155 * the DIQ queue is a debugging queue that dispatches debugging commands to the 156 * firmware. 157 * in this scheduling mode user mode queues over subscription feature is 158 * enabled. 159 * 160 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over 161 * subscription feature disabled. 162 * 163 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly 164 * set the command processor registers and sets the queues "manually". This 165 * mode is used *ONLY* for debugging proposes. 166 * 167 */ 168 enum kfd_sched_policy { 169 KFD_SCHED_POLICY_HWS = 0, 170 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, 171 KFD_SCHED_POLICY_NO_HWS 172 }; 173 174 enum cache_policy { 175 cache_policy_coherent, 176 cache_policy_noncoherent 177 }; 178 179 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) 180 181 struct kfd_event_interrupt_class { 182 bool (*interrupt_isr)(struct kfd_dev *dev, 183 const uint32_t *ih_ring_entry); 184 void (*interrupt_wq)(struct kfd_dev *dev, 185 const uint32_t *ih_ring_entry); 186 }; 187 188 struct kfd_device_info { 189 enum amd_asic_type asic_family; 190 const struct kfd_event_interrupt_class *event_interrupt_class; 191 unsigned int max_pasid_bits; 192 unsigned int max_no_of_hqd; 193 unsigned int doorbell_size; 194 size_t ih_ring_entry_size; 195 uint8_t num_of_watch_points; 196 uint16_t mqd_size_aligned; 197 bool supports_cwsr; 198 bool needs_iommu_device; 199 bool needs_pci_atomics; 200 }; 201 202 struct kfd_mem_obj { 203 uint32_t range_start; 204 uint32_t range_end; 205 uint64_t gpu_addr; 206 uint32_t *cpu_ptr; 207 void *gtt_mem; 208 }; 209 210 struct kfd_vmid_info { 211 uint32_t first_vmid_kfd; 212 uint32_t last_vmid_kfd; 213 uint32_t vmid_num_kfd; 214 }; 215 216 struct kfd_dev { 217 struct kgd_dev *kgd; 218 219 const struct kfd_device_info *device_info; 220 struct pci_dev *pdev; 221 222 unsigned int id; /* topology stub index */ 223 224 phys_addr_t doorbell_base; /* Start of actual doorbells used by 225 * KFD. It is aligned for mapping 226 * into user mode 227 */ 228 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 229 * to HW doorbell, GFX reserved some 230 * at the start) 231 */ 232 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 233 * page used by kernel queue 234 */ 235 236 struct kgd2kfd_shared_resources shared_resources; 237 struct kfd_vmid_info vm_info; 238 239 const struct kfd2kgd_calls *kfd2kgd; 240 struct mutex doorbell_mutex; 241 DECLARE_BITMAP(doorbell_available_index, 242 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 243 244 void *gtt_mem; 245 uint64_t gtt_start_gpu_addr; 246 void *gtt_start_cpu_ptr; 247 void *gtt_sa_bitmap; 248 struct mutex gtt_sa_lock; 249 unsigned int gtt_sa_chunk_size; 250 unsigned int gtt_sa_num_of_chunks; 251 252 /* Interrupts */ 253 struct kfifo ih_fifo; 254 struct workqueue_struct *ih_wq; 255 struct work_struct interrupt_work; 256 spinlock_t interrupt_lock; 257 258 /* QCM Device instance */ 259 struct device_queue_manager *dqm; 260 261 bool init_complete; 262 /* 263 * Interrupts of interest to KFD are copied 264 * from the HW ring into a SW ring. 265 */ 266 bool interrupts_active; 267 268 /* Debug manager */ 269 struct kfd_dbgmgr *dbgmgr; 270 271 /* Maximum process number mapped to HW scheduler */ 272 unsigned int max_proc_per_quantum; 273 274 /* CWSR */ 275 bool cwsr_enabled; 276 const void *cwsr_isa; 277 unsigned int cwsr_isa_size; 278 }; 279 280 /* KGD2KFD callbacks */ 281 void kgd2kfd_exit(void); 282 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 283 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 284 bool kgd2kfd_device_init(struct kfd_dev *kfd, 285 const struct kgd2kfd_shared_resources *gpu_resources); 286 void kgd2kfd_device_exit(struct kfd_dev *kfd); 287 288 enum kfd_mempool { 289 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 290 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 291 KFD_MEMPOOL_FRAMEBUFFER = 3, 292 }; 293 294 /* Character device interface */ 295 int kfd_chardev_init(void); 296 void kfd_chardev_exit(void); 297 struct device *kfd_chardev(void); 298 299 /** 300 * enum kfd_unmap_queues_filter 301 * 302 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 303 * 304 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 305 * running queues list. 306 * 307 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 308 * specific process. 309 * 310 */ 311 enum kfd_unmap_queues_filter { 312 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 313 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 314 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 315 KFD_UNMAP_QUEUES_FILTER_BY_PASID 316 }; 317 318 /** 319 * enum kfd_queue_type 320 * 321 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 322 * 323 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 324 * 325 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 326 * 327 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 328 */ 329 enum kfd_queue_type { 330 KFD_QUEUE_TYPE_COMPUTE, 331 KFD_QUEUE_TYPE_SDMA, 332 KFD_QUEUE_TYPE_HIQ, 333 KFD_QUEUE_TYPE_DIQ 334 }; 335 336 enum kfd_queue_format { 337 KFD_QUEUE_FORMAT_PM4, 338 KFD_QUEUE_FORMAT_AQL 339 }; 340 341 /** 342 * struct queue_properties 343 * 344 * @type: The queue type. 345 * 346 * @queue_id: Queue identifier. 347 * 348 * @queue_address: Queue ring buffer address. 349 * 350 * @queue_size: Queue ring buffer size. 351 * 352 * @priority: Defines the queue priority relative to other queues in the 353 * process. 354 * This is just an indication and HW scheduling may override the priority as 355 * necessary while keeping the relative prioritization. 356 * the priority granularity is from 0 to f which f is the highest priority. 357 * currently all queues are initialized with the highest priority. 358 * 359 * @queue_percent: This field is partially implemented and currently a zero in 360 * this field defines that the queue is non active. 361 * 362 * @read_ptr: User space address which points to the number of dwords the 363 * cp read from the ring buffer. This field updates automatically by the H/W. 364 * 365 * @write_ptr: Defines the number of dwords written to the ring buffer. 366 * 367 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 368 * the queue ring buffer. This field should be similar to write_ptr and the 369 * user should update this field after he updated the write_ptr. 370 * 371 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 372 * 373 * @is_interop: Defines if this is a interop queue. Interop queue means that 374 * the queue can access both graphics and compute resources. 375 * 376 * @is_evicted: Defines if the queue is evicted. Only active queues 377 * are evicted, rendering them inactive. 378 * 379 * @is_active: Defines if the queue is active or not. @is_active and 380 * @is_evicted are protected by the DQM lock. 381 * 382 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 383 * of the queue. 384 * 385 * This structure represents the queue properties for each queue no matter if 386 * it's user mode or kernel mode queue. 387 * 388 */ 389 struct queue_properties { 390 enum kfd_queue_type type; 391 enum kfd_queue_format format; 392 unsigned int queue_id; 393 uint64_t queue_address; 394 uint64_t queue_size; 395 uint32_t priority; 396 uint32_t queue_percent; 397 uint32_t *read_ptr; 398 uint32_t *write_ptr; 399 void __iomem *doorbell_ptr; 400 uint32_t doorbell_off; 401 bool is_interop; 402 bool is_evicted; 403 bool is_active; 404 /* Not relevant for user mode queues in cp scheduling */ 405 unsigned int vmid; 406 /* Relevant only for sdma queues*/ 407 uint32_t sdma_engine_id; 408 uint32_t sdma_queue_id; 409 uint32_t sdma_vm_addr; 410 /* Relevant only for VI */ 411 uint64_t eop_ring_buffer_address; 412 uint32_t eop_ring_buffer_size; 413 uint64_t ctx_save_restore_area_address; 414 uint32_t ctx_save_restore_area_size; 415 uint32_t ctl_stack_size; 416 uint64_t tba_addr; 417 uint64_t tma_addr; 418 }; 419 420 /** 421 * struct queue 422 * 423 * @list: Queue linked list. 424 * 425 * @mqd: The queue MQD. 426 * 427 * @mqd_mem_obj: The MQD local gpu memory object. 428 * 429 * @gart_mqd_addr: The MQD gart mc address. 430 * 431 * @properties: The queue properties. 432 * 433 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 434 * that the queue should be execute on. 435 * 436 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 437 * id. 438 * 439 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 440 * 441 * @process: The kfd process that created this queue. 442 * 443 * @device: The kfd device that created this queue. 444 * 445 * This structure represents user mode compute queues. 446 * It contains all the necessary data to handle such queues. 447 * 448 */ 449 450 struct queue { 451 struct list_head list; 452 void *mqd; 453 struct kfd_mem_obj *mqd_mem_obj; 454 uint64_t gart_mqd_addr; 455 struct queue_properties properties; 456 457 uint32_t mec; 458 uint32_t pipe; 459 uint32_t queue; 460 461 unsigned int sdma_id; 462 unsigned int doorbell_id; 463 464 struct kfd_process *process; 465 struct kfd_dev *device; 466 }; 467 468 /* 469 * Please read the kfd_mqd_manager.h description. 470 */ 471 enum KFD_MQD_TYPE { 472 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 473 KFD_MQD_TYPE_HIQ, /* for hiq */ 474 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 475 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 476 KFD_MQD_TYPE_MAX 477 }; 478 479 struct scheduling_resources { 480 unsigned int vmid_mask; 481 enum kfd_queue_type type; 482 uint64_t queue_mask; 483 uint64_t gws_mask; 484 uint32_t oac_mask; 485 uint32_t gds_heap_base; 486 uint32_t gds_heap_size; 487 }; 488 489 struct process_queue_manager { 490 /* data */ 491 struct kfd_process *process; 492 struct list_head queues; 493 unsigned long *queue_slot_bitmap; 494 }; 495 496 struct qcm_process_device { 497 /* The Device Queue Manager that owns this data */ 498 struct device_queue_manager *dqm; 499 struct process_queue_manager *pqm; 500 /* Queues list */ 501 struct list_head queues_list; 502 struct list_head priv_queue_list; 503 504 unsigned int queue_count; 505 unsigned int vmid; 506 bool is_debug; 507 unsigned int evicted; /* eviction counter, 0=active */ 508 509 /* This flag tells if we should reset all wavefronts on 510 * process termination 511 */ 512 bool reset_wavefronts; 513 514 /* 515 * All the memory management data should be here too 516 */ 517 uint64_t gds_context_area; 518 uint32_t sh_mem_config; 519 uint32_t sh_mem_bases; 520 uint32_t sh_mem_ape1_base; 521 uint32_t sh_mem_ape1_limit; 522 uint32_t page_table_base; 523 uint32_t gds_size; 524 uint32_t num_gws; 525 uint32_t num_oac; 526 uint32_t sh_hidden_private_base; 527 528 /* CWSR memory */ 529 void *cwsr_kaddr; 530 uint64_t cwsr_base; 531 uint64_t tba_addr; 532 uint64_t tma_addr; 533 534 /* IB memory */ 535 uint64_t ib_base; 536 void *ib_kaddr; 537 538 /* doorbell resources per process per device */ 539 unsigned long *doorbell_bitmap; 540 }; 541 542 /* KFD Memory Eviction */ 543 544 /* Approx. wait time before attempting to restore evicted BOs */ 545 #define PROCESS_RESTORE_TIME_MS 100 546 /* Approx. back off time if restore fails due to lack of memory */ 547 #define PROCESS_BACK_OFF_TIME_MS 100 548 /* Approx. time before evicting the process again */ 549 #define PROCESS_ACTIVE_TIME_MS 10 550 551 int kgd2kfd_quiesce_mm(struct mm_struct *mm); 552 int kgd2kfd_resume_mm(struct mm_struct *mm); 553 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 554 struct dma_fence *fence); 555 556 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 557 * idr_handle in the least significant 4 bytes 558 */ 559 #define MAKE_HANDLE(gpu_id, idr_handle) \ 560 (((uint64_t)(gpu_id) << 32) + idr_handle) 561 #define GET_GPU_ID(handle) (handle >> 32) 562 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 563 564 enum kfd_pdd_bound { 565 PDD_UNBOUND = 0, 566 PDD_BOUND, 567 PDD_BOUND_SUSPENDED, 568 }; 569 570 /* Data that is per-process-per device. */ 571 struct kfd_process_device { 572 /* 573 * List of all per-device data for a process. 574 * Starts from kfd_process.per_device_data. 575 */ 576 struct list_head per_device_list; 577 578 /* The device that owns this data. */ 579 struct kfd_dev *dev; 580 581 /* The process that owns this kfd_process_device. */ 582 struct kfd_process *process; 583 584 /* per-process-per device QCM data structure */ 585 struct qcm_process_device qpd; 586 587 /*Apertures*/ 588 uint64_t lds_base; 589 uint64_t lds_limit; 590 uint64_t gpuvm_base; 591 uint64_t gpuvm_limit; 592 uint64_t scratch_base; 593 uint64_t scratch_limit; 594 595 /* VM context for GPUVM allocations */ 596 struct file *drm_file; 597 void *vm; 598 599 /* GPUVM allocations storage */ 600 struct idr alloc_idr; 601 602 /* Flag used to tell the pdd has dequeued from the dqm. 603 * This is used to prevent dev->dqm->ops.process_termination() from 604 * being called twice when it is already called in IOMMU callback 605 * function. 606 */ 607 bool already_dequeued; 608 609 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 610 enum kfd_pdd_bound bound; 611 }; 612 613 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 614 615 /* Process data */ 616 struct kfd_process { 617 /* 618 * kfd_process are stored in an mm_struct*->kfd_process* 619 * hash table (kfd_processes in kfd_process.c) 620 */ 621 struct hlist_node kfd_processes; 622 623 /* 624 * Opaque pointer to mm_struct. We don't hold a reference to 625 * it so it should never be dereferenced from here. This is 626 * only used for looking up processes by their mm. 627 */ 628 void *mm; 629 630 struct kref ref; 631 struct work_struct release_work; 632 633 struct mutex mutex; 634 635 /* 636 * In any process, the thread that started main() is the lead 637 * thread and outlives the rest. 638 * It is here because amd_iommu_bind_pasid wants a task_struct. 639 * It can also be used for safely getting a reference to the 640 * mm_struct of the process. 641 */ 642 struct task_struct *lead_thread; 643 644 /* We want to receive a notification when the mm_struct is destroyed */ 645 struct mmu_notifier mmu_notifier; 646 647 /* Use for delayed freeing of kfd_process structure */ 648 struct rcu_head rcu; 649 650 unsigned int pasid; 651 unsigned int doorbell_index; 652 653 /* 654 * List of kfd_process_device structures, 655 * one for each device the process is using. 656 */ 657 struct list_head per_device_data; 658 659 struct process_queue_manager pqm; 660 661 /*Is the user space process 32 bit?*/ 662 bool is_32bit_user_mode; 663 664 /* Event-related data */ 665 struct mutex event_mutex; 666 /* Event ID allocator and lookup */ 667 struct idr event_idr; 668 /* Event page */ 669 struct kfd_signal_page *signal_page; 670 size_t signal_mapped_size; 671 size_t signal_event_count; 672 bool signal_event_limit_reached; 673 674 /* Information used for memory eviction */ 675 void *kgd_process_info; 676 /* Eviction fence that is attached to all the BOs of this process. The 677 * fence will be triggered during eviction and new one will be created 678 * during restore 679 */ 680 struct dma_fence *ef; 681 682 /* Work items for evicting and restoring BOs */ 683 struct delayed_work eviction_work; 684 struct delayed_work restore_work; 685 /* seqno of the last scheduled eviction */ 686 unsigned int last_eviction_seqno; 687 /* Approx. the last timestamp (in jiffies) when the process was 688 * restored after an eviction 689 */ 690 unsigned long last_restore_timestamp; 691 }; 692 693 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 694 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 695 extern struct srcu_struct kfd_processes_srcu; 696 697 /** 698 * Ioctl function type. 699 * 700 * \param filep pointer to file structure. 701 * \param p amdkfd process pointer. 702 * \param data pointer to arg that was copied from user. 703 */ 704 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 705 void *data); 706 707 struct amdkfd_ioctl_desc { 708 unsigned int cmd; 709 int flags; 710 amdkfd_ioctl_t *func; 711 unsigned int cmd_drv; 712 const char *name; 713 }; 714 715 int kfd_process_create_wq(void); 716 void kfd_process_destroy_wq(void); 717 struct kfd_process *kfd_create_process(struct file *filep); 718 struct kfd_process *kfd_get_process(const struct task_struct *); 719 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 720 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 721 void kfd_unref_process(struct kfd_process *p); 722 int kfd_process_evict_queues(struct kfd_process *p); 723 int kfd_process_restore_queues(struct kfd_process *p); 724 void kfd_suspend_all_processes(void); 725 int kfd_resume_all_processes(void); 726 727 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 728 struct file *drm_file); 729 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 730 struct kfd_process *p); 731 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 732 struct kfd_process *p); 733 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 734 struct kfd_process *p); 735 736 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 737 struct vm_area_struct *vma); 738 739 /* KFD process API for creating and translating handles */ 740 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 741 void *mem); 742 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 743 int handle); 744 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 745 int handle); 746 747 /* Process device data iterator */ 748 struct kfd_process_device *kfd_get_first_process_device_data( 749 struct kfd_process *p); 750 struct kfd_process_device *kfd_get_next_process_device_data( 751 struct kfd_process *p, 752 struct kfd_process_device *pdd); 753 bool kfd_has_process_device_data(struct kfd_process *p); 754 755 /* PASIDs */ 756 int kfd_pasid_init(void); 757 void kfd_pasid_exit(void); 758 bool kfd_set_pasid_limit(unsigned int new_limit); 759 unsigned int kfd_get_pasid_limit(void); 760 unsigned int kfd_pasid_alloc(void); 761 void kfd_pasid_free(unsigned int pasid); 762 763 /* Doorbells */ 764 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 765 int kfd_doorbell_init(struct kfd_dev *kfd); 766 void kfd_doorbell_fini(struct kfd_dev *kfd); 767 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 768 struct vm_area_struct *vma); 769 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 770 unsigned int *doorbell_off); 771 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 772 u32 read_kernel_doorbell(u32 __iomem *db); 773 void write_kernel_doorbell(void __iomem *db, u32 value); 774 void write_kernel_doorbell64(void __iomem *db, u64 value); 775 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd, 776 struct kfd_process *process, 777 unsigned int doorbell_id); 778 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 779 struct kfd_process *process); 780 int kfd_alloc_process_doorbells(struct kfd_process *process); 781 void kfd_free_process_doorbells(struct kfd_process *process); 782 783 /* GTT Sub-Allocator */ 784 785 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 786 struct kfd_mem_obj **mem_obj); 787 788 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 789 790 extern struct device *kfd_device; 791 792 /* Topology */ 793 int kfd_topology_init(void); 794 void kfd_topology_shutdown(void); 795 int kfd_topology_add_device(struct kfd_dev *gpu); 796 int kfd_topology_remove_device(struct kfd_dev *gpu); 797 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 798 uint32_t proximity_domain); 799 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 800 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 801 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 802 int kfd_numa_node_to_apic_id(int numa_node_id); 803 804 /* Interrupts */ 805 int kfd_interrupt_init(struct kfd_dev *dev); 806 void kfd_interrupt_exit(struct kfd_dev *dev); 807 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 808 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 809 bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry); 810 811 /* Power Management */ 812 void kgd2kfd_suspend(struct kfd_dev *kfd); 813 int kgd2kfd_resume(struct kfd_dev *kfd); 814 815 /* amdkfd Apertures */ 816 int kfd_init_apertures(struct kfd_process *process); 817 818 /* Queue Context Management */ 819 int init_queue(struct queue **q, const struct queue_properties *properties); 820 void uninit_queue(struct queue *q); 821 void print_queue_properties(struct queue_properties *q); 822 void print_queue(struct queue *q); 823 824 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 825 struct kfd_dev *dev); 826 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 827 struct kfd_dev *dev); 828 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 829 struct kfd_dev *dev); 830 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 831 struct kfd_dev *dev); 832 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 833 struct kfd_dev *dev); 834 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 835 struct kfd_dev *dev); 836 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 837 void device_queue_manager_uninit(struct device_queue_manager *dqm); 838 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 839 enum kfd_queue_type type); 840 void kernel_queue_uninit(struct kernel_queue *kq); 841 842 /* Process Queue Manager */ 843 struct process_queue_node { 844 struct queue *q; 845 struct kernel_queue *kq; 846 struct list_head process_queue_list; 847 }; 848 849 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 850 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 851 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 852 void pqm_uninit(struct process_queue_manager *pqm); 853 int pqm_create_queue(struct process_queue_manager *pqm, 854 struct kfd_dev *dev, 855 struct file *f, 856 struct queue_properties *properties, 857 unsigned int *qid); 858 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 859 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 860 struct queue_properties *p); 861 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 862 unsigned int qid); 863 864 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 865 unsigned int fence_value, 866 unsigned int timeout_ms); 867 868 /* Packet Manager */ 869 870 #define KFD_FENCE_COMPLETED (100) 871 #define KFD_FENCE_INIT (10) 872 873 struct packet_manager { 874 struct device_queue_manager *dqm; 875 struct kernel_queue *priv_queue; 876 struct mutex lock; 877 bool allocated; 878 struct kfd_mem_obj *ib_buffer_obj; 879 unsigned int ib_size_bytes; 880 881 const struct packet_manager_funcs *pmf; 882 }; 883 884 struct packet_manager_funcs { 885 /* Support ASIC-specific packet formats for PM4 packets */ 886 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 887 struct qcm_process_device *qpd); 888 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 889 uint64_t ib, size_t ib_size_in_dwords, bool chain); 890 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 891 struct scheduling_resources *res); 892 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 893 struct queue *q, bool is_static); 894 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 895 enum kfd_queue_type type, 896 enum kfd_unmap_queues_filter mode, 897 uint32_t filter_param, bool reset, 898 unsigned int sdma_engine); 899 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 900 uint64_t fence_address, uint32_t fence_value); 901 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 902 903 /* Packet sizes */ 904 int map_process_size; 905 int runlist_size; 906 int set_resources_size; 907 int map_queues_size; 908 int unmap_queues_size; 909 int query_status_size; 910 int release_mem_size; 911 }; 912 913 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 914 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 915 916 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 917 void pm_uninit(struct packet_manager *pm); 918 int pm_send_set_resources(struct packet_manager *pm, 919 struct scheduling_resources *res); 920 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 921 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 922 uint32_t fence_value); 923 924 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 925 enum kfd_unmap_queues_filter mode, 926 uint32_t filter_param, bool reset, 927 unsigned int sdma_engine); 928 929 void pm_release_ib(struct packet_manager *pm); 930 931 /* Following PM funcs can be shared among VI and AI */ 932 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 933 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer, 934 struct scheduling_resources *res); 935 936 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 937 938 /* Events */ 939 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 940 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 941 942 extern const struct kfd_device_global_init_class device_global_init_class_cik; 943 944 void kfd_event_init_process(struct kfd_process *p); 945 void kfd_event_free_process(struct kfd_process *p); 946 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 947 int kfd_wait_on_events(struct kfd_process *p, 948 uint32_t num_events, void __user *data, 949 bool all, uint32_t user_timeout_ms, 950 uint32_t *wait_result); 951 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 952 uint32_t valid_id_bits); 953 void kfd_signal_iommu_event(struct kfd_dev *dev, 954 unsigned int pasid, unsigned long address, 955 bool is_write_requested, bool is_execute_requested); 956 void kfd_signal_hw_exception_event(unsigned int pasid); 957 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 958 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 959 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 960 uint64_t size); 961 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 962 uint32_t event_type, bool auto_reset, uint32_t node_id, 963 uint32_t *event_id, uint32_t *event_trigger_data, 964 uint64_t *event_page_offset, uint32_t *event_slot_index); 965 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 966 967 void kfd_flush_tlb(struct kfd_process_device *pdd); 968 969 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 970 971 /* Debugfs */ 972 #if defined(CONFIG_DEBUG_FS) 973 974 void kfd_debugfs_init(void); 975 void kfd_debugfs_fini(void); 976 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 977 int pqm_debugfs_mqds(struct seq_file *m, void *data); 978 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 979 int dqm_debugfs_hqds(struct seq_file *m, void *data); 980 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 981 int pm_debugfs_runlist(struct seq_file *m, void *data); 982 983 #else 984 985 static inline void kfd_debugfs_init(void) {} 986 static inline void kfd_debugfs_fini(void) {} 987 988 #endif 989 990 #endif 991