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 amdgpu 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 \ 107 (KFD_MAX_NUM_OF_PROCESSES * \ 108 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 109 110 #define KFD_KERNEL_QUEUE_SIZE 2048 111 112 /* Kernel module parameter to specify the scheduling policy */ 113 extern int sched_policy; 114 115 /* 116 * Kernel module parameter to specify the maximum process 117 * number per HW scheduler 118 */ 119 extern int hws_max_conc_proc; 120 121 extern int cwsr_enable; 122 123 /* 124 * Kernel module parameter to specify whether to send sigterm to HSA process on 125 * unhandled exception 126 */ 127 extern int send_sigterm; 128 129 /* 130 * This kernel module is used to simulate large bar machine on non-large bar 131 * enabled machines. 132 */ 133 extern int debug_largebar; 134 135 /* 136 * Ignore CRAT table during KFD initialization, can be used to work around 137 * broken CRAT tables on some AMD systems 138 */ 139 extern int ignore_crat; 140 141 /* 142 * Set sh_mem_config.retry_disable on Vega10 143 */ 144 extern int noretry; 145 146 /* 147 * Halt if HWS hang is detected 148 */ 149 extern int halt_if_hws_hang; 150 151 enum cache_policy { 152 cache_policy_coherent, 153 cache_policy_noncoherent 154 }; 155 156 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) 157 158 struct kfd_event_interrupt_class { 159 bool (*interrupt_isr)(struct kfd_dev *dev, 160 const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 161 bool *patched_flag); 162 void (*interrupt_wq)(struct kfd_dev *dev, 163 const uint32_t *ih_ring_entry); 164 }; 165 166 struct kfd_device_info { 167 enum amd_asic_type asic_family; 168 const struct kfd_event_interrupt_class *event_interrupt_class; 169 unsigned int max_pasid_bits; 170 unsigned int max_no_of_hqd; 171 unsigned int doorbell_size; 172 size_t ih_ring_entry_size; 173 uint8_t num_of_watch_points; 174 uint16_t mqd_size_aligned; 175 bool supports_cwsr; 176 bool needs_iommu_device; 177 bool needs_pci_atomics; 178 unsigned int num_sdma_engines; 179 }; 180 181 struct kfd_mem_obj { 182 uint32_t range_start; 183 uint32_t range_end; 184 uint64_t gpu_addr; 185 uint32_t *cpu_ptr; 186 void *gtt_mem; 187 }; 188 189 struct kfd_vmid_info { 190 uint32_t first_vmid_kfd; 191 uint32_t last_vmid_kfd; 192 uint32_t vmid_num_kfd; 193 }; 194 195 struct kfd_dev { 196 struct kgd_dev *kgd; 197 198 const struct kfd_device_info *device_info; 199 struct pci_dev *pdev; 200 201 unsigned int id; /* topology stub index */ 202 203 phys_addr_t doorbell_base; /* Start of actual doorbells used by 204 * KFD. It is aligned for mapping 205 * into user mode 206 */ 207 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 208 * to HW doorbell, GFX reserved some 209 * at the start) 210 */ 211 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 212 * page used by kernel queue 213 */ 214 215 struct kgd2kfd_shared_resources shared_resources; 216 struct kfd_vmid_info vm_info; 217 218 const struct kfd2kgd_calls *kfd2kgd; 219 struct mutex doorbell_mutex; 220 DECLARE_BITMAP(doorbell_available_index, 221 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 222 223 void *gtt_mem; 224 uint64_t gtt_start_gpu_addr; 225 void *gtt_start_cpu_ptr; 226 void *gtt_sa_bitmap; 227 struct mutex gtt_sa_lock; 228 unsigned int gtt_sa_chunk_size; 229 unsigned int gtt_sa_num_of_chunks; 230 231 /* Interrupts */ 232 struct kfifo ih_fifo; 233 struct workqueue_struct *ih_wq; 234 struct work_struct interrupt_work; 235 spinlock_t interrupt_lock; 236 237 /* QCM Device instance */ 238 struct device_queue_manager *dqm; 239 240 bool init_complete; 241 /* 242 * Interrupts of interest to KFD are copied 243 * from the HW ring into a SW ring. 244 */ 245 bool interrupts_active; 246 247 /* Debug manager */ 248 struct kfd_dbgmgr *dbgmgr; 249 250 /* Maximum process number mapped to HW scheduler */ 251 unsigned int max_proc_per_quantum; 252 253 /* CWSR */ 254 bool cwsr_enabled; 255 const void *cwsr_isa; 256 unsigned int cwsr_isa_size; 257 258 /* xGMI */ 259 uint64_t hive_id; 260 }; 261 262 /* KGD2KFD callbacks */ 263 void kgd2kfd_exit(void); 264 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 265 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 266 bool kgd2kfd_device_init(struct kfd_dev *kfd, 267 const struct kgd2kfd_shared_resources *gpu_resources); 268 void kgd2kfd_device_exit(struct kfd_dev *kfd); 269 270 enum kfd_mempool { 271 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 272 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 273 KFD_MEMPOOL_FRAMEBUFFER = 3, 274 }; 275 276 /* Character device interface */ 277 int kfd_chardev_init(void); 278 void kfd_chardev_exit(void); 279 struct device *kfd_chardev(void); 280 281 /** 282 * enum kfd_unmap_queues_filter 283 * 284 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 285 * 286 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 287 * running queues list. 288 * 289 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 290 * specific process. 291 * 292 */ 293 enum kfd_unmap_queues_filter { 294 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 295 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 296 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 297 KFD_UNMAP_QUEUES_FILTER_BY_PASID 298 }; 299 300 /** 301 * enum kfd_queue_type 302 * 303 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 304 * 305 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 306 * 307 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 308 * 309 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 310 */ 311 enum kfd_queue_type { 312 KFD_QUEUE_TYPE_COMPUTE, 313 KFD_QUEUE_TYPE_SDMA, 314 KFD_QUEUE_TYPE_HIQ, 315 KFD_QUEUE_TYPE_DIQ 316 }; 317 318 enum kfd_queue_format { 319 KFD_QUEUE_FORMAT_PM4, 320 KFD_QUEUE_FORMAT_AQL 321 }; 322 323 /** 324 * struct queue_properties 325 * 326 * @type: The queue type. 327 * 328 * @queue_id: Queue identifier. 329 * 330 * @queue_address: Queue ring buffer address. 331 * 332 * @queue_size: Queue ring buffer size. 333 * 334 * @priority: Defines the queue priority relative to other queues in the 335 * process. 336 * This is just an indication and HW scheduling may override the priority as 337 * necessary while keeping the relative prioritization. 338 * the priority granularity is from 0 to f which f is the highest priority. 339 * currently all queues are initialized with the highest priority. 340 * 341 * @queue_percent: This field is partially implemented and currently a zero in 342 * this field defines that the queue is non active. 343 * 344 * @read_ptr: User space address which points to the number of dwords the 345 * cp read from the ring buffer. This field updates automatically by the H/W. 346 * 347 * @write_ptr: Defines the number of dwords written to the ring buffer. 348 * 349 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 350 * the queue ring buffer. This field should be similar to write_ptr and the 351 * user should update this field after he updated the write_ptr. 352 * 353 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 354 * 355 * @is_interop: Defines if this is a interop queue. Interop queue means that 356 * the queue can access both graphics and compute resources. 357 * 358 * @is_evicted: Defines if the queue is evicted. Only active queues 359 * are evicted, rendering them inactive. 360 * 361 * @is_active: Defines if the queue is active or not. @is_active and 362 * @is_evicted are protected by the DQM lock. 363 * 364 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 365 * of the queue. 366 * 367 * This structure represents the queue properties for each queue no matter if 368 * it's user mode or kernel mode queue. 369 * 370 */ 371 struct queue_properties { 372 enum kfd_queue_type type; 373 enum kfd_queue_format format; 374 unsigned int queue_id; 375 uint64_t queue_address; 376 uint64_t queue_size; 377 uint32_t priority; 378 uint32_t queue_percent; 379 uint32_t *read_ptr; 380 uint32_t *write_ptr; 381 void __iomem *doorbell_ptr; 382 uint32_t doorbell_off; 383 bool is_interop; 384 bool is_evicted; 385 bool is_active; 386 /* Not relevant for user mode queues in cp scheduling */ 387 unsigned int vmid; 388 /* Relevant only for sdma queues*/ 389 uint32_t sdma_engine_id; 390 uint32_t sdma_queue_id; 391 uint32_t sdma_vm_addr; 392 /* Relevant only for VI */ 393 uint64_t eop_ring_buffer_address; 394 uint32_t eop_ring_buffer_size; 395 uint64_t ctx_save_restore_area_address; 396 uint32_t ctx_save_restore_area_size; 397 uint32_t ctl_stack_size; 398 uint64_t tba_addr; 399 uint64_t tma_addr; 400 /* Relevant for CU */ 401 uint32_t cu_mask_count; /* Must be a multiple of 32 */ 402 uint32_t *cu_mask; 403 }; 404 405 /** 406 * struct queue 407 * 408 * @list: Queue linked list. 409 * 410 * @mqd: The queue MQD. 411 * 412 * @mqd_mem_obj: The MQD local gpu memory object. 413 * 414 * @gart_mqd_addr: The MQD gart mc address. 415 * 416 * @properties: The queue properties. 417 * 418 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 419 * that the queue should be execute on. 420 * 421 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 422 * id. 423 * 424 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 425 * 426 * @process: The kfd process that created this queue. 427 * 428 * @device: The kfd device that created this queue. 429 * 430 * This structure represents user mode compute queues. 431 * It contains all the necessary data to handle such queues. 432 * 433 */ 434 435 struct queue { 436 struct list_head list; 437 void *mqd; 438 struct kfd_mem_obj *mqd_mem_obj; 439 uint64_t gart_mqd_addr; 440 struct queue_properties properties; 441 442 uint32_t mec; 443 uint32_t pipe; 444 uint32_t queue; 445 446 unsigned int sdma_id; 447 unsigned int doorbell_id; 448 449 struct kfd_process *process; 450 struct kfd_dev *device; 451 }; 452 453 /* 454 * Please read the kfd_mqd_manager.h description. 455 */ 456 enum KFD_MQD_TYPE { 457 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 458 KFD_MQD_TYPE_HIQ, /* for hiq */ 459 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 460 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 461 KFD_MQD_TYPE_MAX 462 }; 463 464 struct scheduling_resources { 465 unsigned int vmid_mask; 466 enum kfd_queue_type type; 467 uint64_t queue_mask; 468 uint64_t gws_mask; 469 uint32_t oac_mask; 470 uint32_t gds_heap_base; 471 uint32_t gds_heap_size; 472 }; 473 474 struct process_queue_manager { 475 /* data */ 476 struct kfd_process *process; 477 struct list_head queues; 478 unsigned long *queue_slot_bitmap; 479 }; 480 481 struct qcm_process_device { 482 /* The Device Queue Manager that owns this data */ 483 struct device_queue_manager *dqm; 484 struct process_queue_manager *pqm; 485 /* Queues list */ 486 struct list_head queues_list; 487 struct list_head priv_queue_list; 488 489 unsigned int queue_count; 490 unsigned int vmid; 491 bool is_debug; 492 unsigned int evicted; /* eviction counter, 0=active */ 493 494 /* This flag tells if we should reset all wavefronts on 495 * process termination 496 */ 497 bool reset_wavefronts; 498 499 /* 500 * All the memory management data should be here too 501 */ 502 uint64_t gds_context_area; 503 uint32_t sh_mem_config; 504 uint32_t sh_mem_bases; 505 uint32_t sh_mem_ape1_base; 506 uint32_t sh_mem_ape1_limit; 507 uint32_t page_table_base; 508 uint32_t gds_size; 509 uint32_t num_gws; 510 uint32_t num_oac; 511 uint32_t sh_hidden_private_base; 512 513 /* CWSR memory */ 514 void *cwsr_kaddr; 515 uint64_t cwsr_base; 516 uint64_t tba_addr; 517 uint64_t tma_addr; 518 519 /* IB memory */ 520 uint64_t ib_base; 521 void *ib_kaddr; 522 523 /* doorbell resources per process per device */ 524 unsigned long *doorbell_bitmap; 525 }; 526 527 /* KFD Memory Eviction */ 528 529 /* Approx. wait time before attempting to restore evicted BOs */ 530 #define PROCESS_RESTORE_TIME_MS 100 531 /* Approx. back off time if restore fails due to lack of memory */ 532 #define PROCESS_BACK_OFF_TIME_MS 100 533 /* Approx. time before evicting the process again */ 534 #define PROCESS_ACTIVE_TIME_MS 10 535 536 int kgd2kfd_quiesce_mm(struct mm_struct *mm); 537 int kgd2kfd_resume_mm(struct mm_struct *mm); 538 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 539 struct dma_fence *fence); 540 541 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 542 * idr_handle in the least significant 4 bytes 543 */ 544 #define MAKE_HANDLE(gpu_id, idr_handle) \ 545 (((uint64_t)(gpu_id) << 32) + idr_handle) 546 #define GET_GPU_ID(handle) (handle >> 32) 547 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 548 549 enum kfd_pdd_bound { 550 PDD_UNBOUND = 0, 551 PDD_BOUND, 552 PDD_BOUND_SUSPENDED, 553 }; 554 555 /* Data that is per-process-per device. */ 556 struct kfd_process_device { 557 /* 558 * List of all per-device data for a process. 559 * Starts from kfd_process.per_device_data. 560 */ 561 struct list_head per_device_list; 562 563 /* The device that owns this data. */ 564 struct kfd_dev *dev; 565 566 /* The process that owns this kfd_process_device. */ 567 struct kfd_process *process; 568 569 /* per-process-per device QCM data structure */ 570 struct qcm_process_device qpd; 571 572 /*Apertures*/ 573 uint64_t lds_base; 574 uint64_t lds_limit; 575 uint64_t gpuvm_base; 576 uint64_t gpuvm_limit; 577 uint64_t scratch_base; 578 uint64_t scratch_limit; 579 580 /* VM context for GPUVM allocations */ 581 struct file *drm_file; 582 void *vm; 583 584 /* GPUVM allocations storage */ 585 struct idr alloc_idr; 586 587 /* Flag used to tell the pdd has dequeued from the dqm. 588 * This is used to prevent dev->dqm->ops.process_termination() from 589 * being called twice when it is already called in IOMMU callback 590 * function. 591 */ 592 bool already_dequeued; 593 594 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 595 enum kfd_pdd_bound bound; 596 }; 597 598 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 599 600 /* Process data */ 601 struct kfd_process { 602 /* 603 * kfd_process are stored in an mm_struct*->kfd_process* 604 * hash table (kfd_processes in kfd_process.c) 605 */ 606 struct hlist_node kfd_processes; 607 608 /* 609 * Opaque pointer to mm_struct. We don't hold a reference to 610 * it so it should never be dereferenced from here. This is 611 * only used for looking up processes by their mm. 612 */ 613 void *mm; 614 615 struct kref ref; 616 struct work_struct release_work; 617 618 struct mutex mutex; 619 620 /* 621 * In any process, the thread that started main() is the lead 622 * thread and outlives the rest. 623 * It is here because amd_iommu_bind_pasid wants a task_struct. 624 * It can also be used for safely getting a reference to the 625 * mm_struct of the process. 626 */ 627 struct task_struct *lead_thread; 628 629 /* We want to receive a notification when the mm_struct is destroyed */ 630 struct mmu_notifier mmu_notifier; 631 632 /* Use for delayed freeing of kfd_process structure */ 633 struct rcu_head rcu; 634 635 unsigned int pasid; 636 unsigned int doorbell_index; 637 638 /* 639 * List of kfd_process_device structures, 640 * one for each device the process is using. 641 */ 642 struct list_head per_device_data; 643 644 struct process_queue_manager pqm; 645 646 /*Is the user space process 32 bit?*/ 647 bool is_32bit_user_mode; 648 649 /* Event-related data */ 650 struct mutex event_mutex; 651 /* Event ID allocator and lookup */ 652 struct idr event_idr; 653 /* Event page */ 654 struct kfd_signal_page *signal_page; 655 size_t signal_mapped_size; 656 size_t signal_event_count; 657 bool signal_event_limit_reached; 658 659 /* Information used for memory eviction */ 660 void *kgd_process_info; 661 /* Eviction fence that is attached to all the BOs of this process. The 662 * fence will be triggered during eviction and new one will be created 663 * during restore 664 */ 665 struct dma_fence *ef; 666 667 /* Work items for evicting and restoring BOs */ 668 struct delayed_work eviction_work; 669 struct delayed_work restore_work; 670 /* seqno of the last scheduled eviction */ 671 unsigned int last_eviction_seqno; 672 /* Approx. the last timestamp (in jiffies) when the process was 673 * restored after an eviction 674 */ 675 unsigned long last_restore_timestamp; 676 }; 677 678 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 679 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 680 extern struct srcu_struct kfd_processes_srcu; 681 682 /** 683 * Ioctl function type. 684 * 685 * \param filep pointer to file structure. 686 * \param p amdkfd process pointer. 687 * \param data pointer to arg that was copied from user. 688 */ 689 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 690 void *data); 691 692 struct amdkfd_ioctl_desc { 693 unsigned int cmd; 694 int flags; 695 amdkfd_ioctl_t *func; 696 unsigned int cmd_drv; 697 const char *name; 698 }; 699 bool kfd_dev_is_large_bar(struct kfd_dev *dev); 700 701 int kfd_process_create_wq(void); 702 void kfd_process_destroy_wq(void); 703 struct kfd_process *kfd_create_process(struct file *filep); 704 struct kfd_process *kfd_get_process(const struct task_struct *); 705 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 706 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 707 void kfd_unref_process(struct kfd_process *p); 708 int kfd_process_evict_queues(struct kfd_process *p); 709 int kfd_process_restore_queues(struct kfd_process *p); 710 void kfd_suspend_all_processes(void); 711 int kfd_resume_all_processes(void); 712 713 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 714 struct file *drm_file); 715 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 716 struct kfd_process *p); 717 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 718 struct kfd_process *p); 719 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 720 struct kfd_process *p); 721 722 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 723 struct vm_area_struct *vma); 724 725 /* KFD process API for creating and translating handles */ 726 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 727 void *mem); 728 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 729 int handle); 730 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 731 int handle); 732 733 /* Process device data iterator */ 734 struct kfd_process_device *kfd_get_first_process_device_data( 735 struct kfd_process *p); 736 struct kfd_process_device *kfd_get_next_process_device_data( 737 struct kfd_process *p, 738 struct kfd_process_device *pdd); 739 bool kfd_has_process_device_data(struct kfd_process *p); 740 741 /* PASIDs */ 742 int kfd_pasid_init(void); 743 void kfd_pasid_exit(void); 744 bool kfd_set_pasid_limit(unsigned int new_limit); 745 unsigned int kfd_get_pasid_limit(void); 746 unsigned int kfd_pasid_alloc(void); 747 void kfd_pasid_free(unsigned int pasid); 748 749 /* Doorbells */ 750 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 751 int kfd_doorbell_init(struct kfd_dev *kfd); 752 void kfd_doorbell_fini(struct kfd_dev *kfd); 753 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 754 struct vm_area_struct *vma); 755 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 756 unsigned int *doorbell_off); 757 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 758 u32 read_kernel_doorbell(u32 __iomem *db); 759 void write_kernel_doorbell(void __iomem *db, u32 value); 760 void write_kernel_doorbell64(void __iomem *db, u64 value); 761 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd, 762 struct kfd_process *process, 763 unsigned int doorbell_id); 764 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 765 struct kfd_process *process); 766 int kfd_alloc_process_doorbells(struct kfd_process *process); 767 void kfd_free_process_doorbells(struct kfd_process *process); 768 769 /* GTT Sub-Allocator */ 770 771 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 772 struct kfd_mem_obj **mem_obj); 773 774 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 775 776 extern struct device *kfd_device; 777 778 /* Topology */ 779 int kfd_topology_init(void); 780 void kfd_topology_shutdown(void); 781 int kfd_topology_add_device(struct kfd_dev *gpu); 782 int kfd_topology_remove_device(struct kfd_dev *gpu); 783 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 784 uint32_t proximity_domain); 785 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 786 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 787 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 788 int kfd_numa_node_to_apic_id(int numa_node_id); 789 790 /* Interrupts */ 791 int kfd_interrupt_init(struct kfd_dev *dev); 792 void kfd_interrupt_exit(struct kfd_dev *dev); 793 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 794 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 795 bool interrupt_is_wanted(struct kfd_dev *dev, 796 const uint32_t *ih_ring_entry, 797 uint32_t *patched_ihre, bool *flag); 798 799 /* Power Management */ 800 void kgd2kfd_suspend(struct kfd_dev *kfd); 801 int kgd2kfd_resume(struct kfd_dev *kfd); 802 803 /* GPU reset */ 804 int kgd2kfd_pre_reset(struct kfd_dev *kfd); 805 int kgd2kfd_post_reset(struct kfd_dev *kfd); 806 807 /* amdkfd Apertures */ 808 int kfd_init_apertures(struct kfd_process *process); 809 810 /* Queue Context Management */ 811 int init_queue(struct queue **q, const struct queue_properties *properties); 812 void uninit_queue(struct queue *q); 813 void print_queue_properties(struct queue_properties *q); 814 void print_queue(struct queue *q); 815 816 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 817 struct kfd_dev *dev); 818 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 819 struct kfd_dev *dev); 820 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 821 struct kfd_dev *dev); 822 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 823 struct kfd_dev *dev); 824 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 825 struct kfd_dev *dev); 826 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 827 struct kfd_dev *dev); 828 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 829 void device_queue_manager_uninit(struct device_queue_manager *dqm); 830 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 831 enum kfd_queue_type type); 832 void kernel_queue_uninit(struct kernel_queue *kq); 833 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); 834 835 /* Process Queue Manager */ 836 struct process_queue_node { 837 struct queue *q; 838 struct kernel_queue *kq; 839 struct list_head process_queue_list; 840 }; 841 842 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 843 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 844 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 845 void pqm_uninit(struct process_queue_manager *pqm); 846 int pqm_create_queue(struct process_queue_manager *pqm, 847 struct kfd_dev *dev, 848 struct file *f, 849 struct queue_properties *properties, 850 unsigned int *qid); 851 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 852 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 853 struct queue_properties *p); 854 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 855 struct queue_properties *p); 856 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 857 unsigned int qid); 858 859 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 860 unsigned int fence_value, 861 unsigned int timeout_ms); 862 863 /* Packet Manager */ 864 865 #define KFD_FENCE_COMPLETED (100) 866 #define KFD_FENCE_INIT (10) 867 868 struct packet_manager { 869 struct device_queue_manager *dqm; 870 struct kernel_queue *priv_queue; 871 struct mutex lock; 872 bool allocated; 873 struct kfd_mem_obj *ib_buffer_obj; 874 unsigned int ib_size_bytes; 875 876 const struct packet_manager_funcs *pmf; 877 }; 878 879 struct packet_manager_funcs { 880 /* Support ASIC-specific packet formats for PM4 packets */ 881 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 882 struct qcm_process_device *qpd); 883 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 884 uint64_t ib, size_t ib_size_in_dwords, bool chain); 885 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 886 struct scheduling_resources *res); 887 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 888 struct queue *q, bool is_static); 889 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 890 enum kfd_queue_type type, 891 enum kfd_unmap_queues_filter mode, 892 uint32_t filter_param, bool reset, 893 unsigned int sdma_engine); 894 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 895 uint64_t fence_address, uint32_t fence_value); 896 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 897 898 /* Packet sizes */ 899 int map_process_size; 900 int runlist_size; 901 int set_resources_size; 902 int map_queues_size; 903 int unmap_queues_size; 904 int query_status_size; 905 int release_mem_size; 906 }; 907 908 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 909 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 910 911 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 912 void pm_uninit(struct packet_manager *pm); 913 int pm_send_set_resources(struct packet_manager *pm, 914 struct scheduling_resources *res); 915 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 916 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 917 uint32_t fence_value); 918 919 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 920 enum kfd_unmap_queues_filter mode, 921 uint32_t filter_param, bool reset, 922 unsigned int sdma_engine); 923 924 void pm_release_ib(struct packet_manager *pm); 925 926 /* Following PM funcs can be shared among VI and AI */ 927 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 928 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer, 929 struct scheduling_resources *res); 930 931 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 932 933 /* Events */ 934 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 935 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 936 937 extern const struct kfd_device_global_init_class device_global_init_class_cik; 938 939 void kfd_event_init_process(struct kfd_process *p); 940 void kfd_event_free_process(struct kfd_process *p); 941 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 942 int kfd_wait_on_events(struct kfd_process *p, 943 uint32_t num_events, void __user *data, 944 bool all, uint32_t user_timeout_ms, 945 uint32_t *wait_result); 946 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 947 uint32_t valid_id_bits); 948 void kfd_signal_iommu_event(struct kfd_dev *dev, 949 unsigned int pasid, unsigned long address, 950 bool is_write_requested, bool is_execute_requested); 951 void kfd_signal_hw_exception_event(unsigned int pasid); 952 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 953 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 954 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 955 uint64_t size); 956 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 957 uint32_t event_type, bool auto_reset, uint32_t node_id, 958 uint32_t *event_id, uint32_t *event_trigger_data, 959 uint64_t *event_page_offset, uint32_t *event_slot_index); 960 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 961 962 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid, 963 struct kfd_vm_fault_info *info); 964 965 void kfd_signal_reset_event(struct kfd_dev *dev); 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 bool kfd_is_locked(void); 972 973 /* Debugfs */ 974 #if defined(CONFIG_DEBUG_FS) 975 976 void kfd_debugfs_init(void); 977 void kfd_debugfs_fini(void); 978 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 979 int pqm_debugfs_mqds(struct seq_file *m, void *data); 980 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 981 int dqm_debugfs_hqds(struct seq_file *m, void *data); 982 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 983 int pm_debugfs_runlist(struct seq_file *m, void *data); 984 985 int kfd_debugfs_hang_hws(struct kfd_dev *dev); 986 int pm_debugfs_hang_hws(struct packet_manager *pm); 987 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 988 989 #else 990 991 static inline void kfd_debugfs_init(void) {} 992 static inline void kfd_debugfs_fini(void) {} 993 994 #endif 995 996 #endif 997