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 unsigned int num_sdma_queues_per_engine; 180 }; 181 182 struct kfd_mem_obj { 183 uint32_t range_start; 184 uint32_t range_end; 185 uint64_t gpu_addr; 186 uint32_t *cpu_ptr; 187 void *gtt_mem; 188 }; 189 190 struct kfd_vmid_info { 191 uint32_t first_vmid_kfd; 192 uint32_t last_vmid_kfd; 193 uint32_t vmid_num_kfd; 194 }; 195 196 struct kfd_dev { 197 struct kgd_dev *kgd; 198 199 const struct kfd_device_info *device_info; 200 struct pci_dev *pdev; 201 202 unsigned int id; /* topology stub index */ 203 204 phys_addr_t doorbell_base; /* Start of actual doorbells used by 205 * KFD. It is aligned for mapping 206 * into user mode 207 */ 208 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 209 * to HW doorbell, GFX reserved some 210 * at the start) 211 */ 212 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 213 * page used by kernel queue 214 */ 215 216 struct kgd2kfd_shared_resources shared_resources; 217 struct kfd_vmid_info vm_info; 218 219 const struct kfd2kgd_calls *kfd2kgd; 220 struct mutex doorbell_mutex; 221 DECLARE_BITMAP(doorbell_available_index, 222 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 223 224 void *gtt_mem; 225 uint64_t gtt_start_gpu_addr; 226 void *gtt_start_cpu_ptr; 227 void *gtt_sa_bitmap; 228 struct mutex gtt_sa_lock; 229 unsigned int gtt_sa_chunk_size; 230 unsigned int gtt_sa_num_of_chunks; 231 232 /* Interrupts */ 233 struct kfifo ih_fifo; 234 struct workqueue_struct *ih_wq; 235 struct work_struct interrupt_work; 236 spinlock_t interrupt_lock; 237 238 /* QCM Device instance */ 239 struct device_queue_manager *dqm; 240 241 bool init_complete; 242 /* 243 * Interrupts of interest to KFD are copied 244 * from the HW ring into a SW ring. 245 */ 246 bool interrupts_active; 247 248 /* Debug manager */ 249 struct kfd_dbgmgr *dbgmgr; 250 251 /* Firmware versions */ 252 uint16_t mec_fw_version; 253 uint16_t sdma_fw_version; 254 255 /* Maximum process number mapped to HW scheduler */ 256 unsigned int max_proc_per_quantum; 257 258 /* CWSR */ 259 bool cwsr_enabled; 260 const void *cwsr_isa; 261 unsigned int cwsr_isa_size; 262 263 /* xGMI */ 264 uint64_t hive_id; 265 266 bool pci_atomic_requested; 267 }; 268 269 /* KGD2KFD callbacks */ 270 void kgd2kfd_exit(void); 271 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 272 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g); 273 bool kgd2kfd_device_init(struct kfd_dev *kfd, 274 const struct kgd2kfd_shared_resources *gpu_resources); 275 void kgd2kfd_device_exit(struct kfd_dev *kfd); 276 277 enum kfd_mempool { 278 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 279 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 280 KFD_MEMPOOL_FRAMEBUFFER = 3, 281 }; 282 283 /* Character device interface */ 284 int kfd_chardev_init(void); 285 void kfd_chardev_exit(void); 286 struct device *kfd_chardev(void); 287 288 /** 289 * enum kfd_unmap_queues_filter 290 * 291 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 292 * 293 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 294 * running queues list. 295 * 296 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 297 * specific process. 298 * 299 */ 300 enum kfd_unmap_queues_filter { 301 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 302 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 303 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 304 KFD_UNMAP_QUEUES_FILTER_BY_PASID 305 }; 306 307 /** 308 * enum kfd_queue_type 309 * 310 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 311 * 312 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 313 * 314 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 315 * 316 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 317 */ 318 enum kfd_queue_type { 319 KFD_QUEUE_TYPE_COMPUTE, 320 KFD_QUEUE_TYPE_SDMA, 321 KFD_QUEUE_TYPE_HIQ, 322 KFD_QUEUE_TYPE_DIQ 323 }; 324 325 enum kfd_queue_format { 326 KFD_QUEUE_FORMAT_PM4, 327 KFD_QUEUE_FORMAT_AQL 328 }; 329 330 /** 331 * struct queue_properties 332 * 333 * @type: The queue type. 334 * 335 * @queue_id: Queue identifier. 336 * 337 * @queue_address: Queue ring buffer address. 338 * 339 * @queue_size: Queue ring buffer size. 340 * 341 * @priority: Defines the queue priority relative to other queues in the 342 * process. 343 * This is just an indication and HW scheduling may override the priority as 344 * necessary while keeping the relative prioritization. 345 * the priority granularity is from 0 to f which f is the highest priority. 346 * currently all queues are initialized with the highest priority. 347 * 348 * @queue_percent: This field is partially implemented and currently a zero in 349 * this field defines that the queue is non active. 350 * 351 * @read_ptr: User space address which points to the number of dwords the 352 * cp read from the ring buffer. This field updates automatically by the H/W. 353 * 354 * @write_ptr: Defines the number of dwords written to the ring buffer. 355 * 356 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 357 * the queue ring buffer. This field should be similar to write_ptr and the 358 * user should update this field after he updated the write_ptr. 359 * 360 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 361 * 362 * @is_interop: Defines if this is a interop queue. Interop queue means that 363 * the queue can access both graphics and compute resources. 364 * 365 * @is_evicted: Defines if the queue is evicted. Only active queues 366 * are evicted, rendering them inactive. 367 * 368 * @is_active: Defines if the queue is active or not. @is_active and 369 * @is_evicted are protected by the DQM lock. 370 * 371 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 372 * of the queue. 373 * 374 * This structure represents the queue properties for each queue no matter if 375 * it's user mode or kernel mode queue. 376 * 377 */ 378 struct queue_properties { 379 enum kfd_queue_type type; 380 enum kfd_queue_format format; 381 unsigned int queue_id; 382 uint64_t queue_address; 383 uint64_t queue_size; 384 uint32_t priority; 385 uint32_t queue_percent; 386 uint32_t *read_ptr; 387 uint32_t *write_ptr; 388 void __iomem *doorbell_ptr; 389 uint32_t doorbell_off; 390 bool is_interop; 391 bool is_evicted; 392 bool is_active; 393 /* Not relevant for user mode queues in cp scheduling */ 394 unsigned int vmid; 395 /* Relevant only for sdma queues*/ 396 uint32_t sdma_engine_id; 397 uint32_t sdma_queue_id; 398 uint32_t sdma_vm_addr; 399 /* Relevant only for VI */ 400 uint64_t eop_ring_buffer_address; 401 uint32_t eop_ring_buffer_size; 402 uint64_t ctx_save_restore_area_address; 403 uint32_t ctx_save_restore_area_size; 404 uint32_t ctl_stack_size; 405 uint64_t tba_addr; 406 uint64_t tma_addr; 407 /* Relevant for CU */ 408 uint32_t cu_mask_count; /* Must be a multiple of 32 */ 409 uint32_t *cu_mask; 410 }; 411 412 /** 413 * struct queue 414 * 415 * @list: Queue linked list. 416 * 417 * @mqd: The queue MQD. 418 * 419 * @mqd_mem_obj: The MQD local gpu memory object. 420 * 421 * @gart_mqd_addr: The MQD gart mc address. 422 * 423 * @properties: The queue properties. 424 * 425 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 426 * that the queue should be execute on. 427 * 428 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 429 * id. 430 * 431 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 432 * 433 * @process: The kfd process that created this queue. 434 * 435 * @device: The kfd device that created this queue. 436 * 437 * This structure represents user mode compute queues. 438 * It contains all the necessary data to handle such queues. 439 * 440 */ 441 442 struct queue { 443 struct list_head list; 444 void *mqd; 445 struct kfd_mem_obj *mqd_mem_obj; 446 uint64_t gart_mqd_addr; 447 struct queue_properties properties; 448 449 uint32_t mec; 450 uint32_t pipe; 451 uint32_t queue; 452 453 unsigned int sdma_id; 454 unsigned int doorbell_id; 455 456 struct kfd_process *process; 457 struct kfd_dev *device; 458 }; 459 460 /* 461 * Please read the kfd_mqd_manager.h description. 462 */ 463 enum KFD_MQD_TYPE { 464 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 465 KFD_MQD_TYPE_HIQ, /* for hiq */ 466 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 467 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 468 KFD_MQD_TYPE_MAX 469 }; 470 471 struct scheduling_resources { 472 unsigned int vmid_mask; 473 enum kfd_queue_type type; 474 uint64_t queue_mask; 475 uint64_t gws_mask; 476 uint32_t oac_mask; 477 uint32_t gds_heap_base; 478 uint32_t gds_heap_size; 479 }; 480 481 struct process_queue_manager { 482 /* data */ 483 struct kfd_process *process; 484 struct list_head queues; 485 unsigned long *queue_slot_bitmap; 486 }; 487 488 struct qcm_process_device { 489 /* The Device Queue Manager that owns this data */ 490 struct device_queue_manager *dqm; 491 struct process_queue_manager *pqm; 492 /* Queues list */ 493 struct list_head queues_list; 494 struct list_head priv_queue_list; 495 496 unsigned int queue_count; 497 unsigned int vmid; 498 bool is_debug; 499 unsigned int evicted; /* eviction counter, 0=active */ 500 501 /* This flag tells if we should reset all wavefronts on 502 * process termination 503 */ 504 bool reset_wavefronts; 505 506 /* 507 * All the memory management data should be here too 508 */ 509 uint64_t gds_context_area; 510 uint64_t page_table_base; 511 uint32_t sh_mem_config; 512 uint32_t sh_mem_bases; 513 uint32_t sh_mem_ape1_base; 514 uint32_t sh_mem_ape1_limit; 515 uint32_t gds_size; 516 uint32_t num_gws; 517 uint32_t num_oac; 518 uint32_t sh_hidden_private_base; 519 520 /* CWSR memory */ 521 void *cwsr_kaddr; 522 uint64_t cwsr_base; 523 uint64_t tba_addr; 524 uint64_t tma_addr; 525 526 /* IB memory */ 527 uint64_t ib_base; 528 void *ib_kaddr; 529 530 /* doorbell resources per process per device */ 531 unsigned long *doorbell_bitmap; 532 }; 533 534 /* KFD Memory Eviction */ 535 536 /* Approx. wait time before attempting to restore evicted BOs */ 537 #define PROCESS_RESTORE_TIME_MS 100 538 /* Approx. back off time if restore fails due to lack of memory */ 539 #define PROCESS_BACK_OFF_TIME_MS 100 540 /* Approx. time before evicting the process again */ 541 #define PROCESS_ACTIVE_TIME_MS 10 542 543 int kgd2kfd_quiesce_mm(struct mm_struct *mm); 544 int kgd2kfd_resume_mm(struct mm_struct *mm); 545 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 546 struct dma_fence *fence); 547 548 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 549 * idr_handle in the least significant 4 bytes 550 */ 551 #define MAKE_HANDLE(gpu_id, idr_handle) \ 552 (((uint64_t)(gpu_id) << 32) + idr_handle) 553 #define GET_GPU_ID(handle) (handle >> 32) 554 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 555 556 enum kfd_pdd_bound { 557 PDD_UNBOUND = 0, 558 PDD_BOUND, 559 PDD_BOUND_SUSPENDED, 560 }; 561 562 /* Data that is per-process-per device. */ 563 struct kfd_process_device { 564 /* 565 * List of all per-device data for a process. 566 * Starts from kfd_process.per_device_data. 567 */ 568 struct list_head per_device_list; 569 570 /* The device that owns this data. */ 571 struct kfd_dev *dev; 572 573 /* The process that owns this kfd_process_device. */ 574 struct kfd_process *process; 575 576 /* per-process-per device QCM data structure */ 577 struct qcm_process_device qpd; 578 579 /*Apertures*/ 580 uint64_t lds_base; 581 uint64_t lds_limit; 582 uint64_t gpuvm_base; 583 uint64_t gpuvm_limit; 584 uint64_t scratch_base; 585 uint64_t scratch_limit; 586 587 /* VM context for GPUVM allocations */ 588 struct file *drm_file; 589 void *vm; 590 591 /* GPUVM allocations storage */ 592 struct idr alloc_idr; 593 594 /* Flag used to tell the pdd has dequeued from the dqm. 595 * This is used to prevent dev->dqm->ops.process_termination() from 596 * being called twice when it is already called in IOMMU callback 597 * function. 598 */ 599 bool already_dequeued; 600 601 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 602 enum kfd_pdd_bound bound; 603 }; 604 605 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 606 607 /* Process data */ 608 struct kfd_process { 609 /* 610 * kfd_process are stored in an mm_struct*->kfd_process* 611 * hash table (kfd_processes in kfd_process.c) 612 */ 613 struct hlist_node kfd_processes; 614 615 /* 616 * Opaque pointer to mm_struct. We don't hold a reference to 617 * it so it should never be dereferenced from here. This is 618 * only used for looking up processes by their mm. 619 */ 620 void *mm; 621 622 struct kref ref; 623 struct work_struct release_work; 624 625 struct mutex mutex; 626 627 /* 628 * In any process, the thread that started main() is the lead 629 * thread and outlives the rest. 630 * It is here because amd_iommu_bind_pasid wants a task_struct. 631 * It can also be used for safely getting a reference to the 632 * mm_struct of the process. 633 */ 634 struct task_struct *lead_thread; 635 636 /* We want to receive a notification when the mm_struct is destroyed */ 637 struct mmu_notifier mmu_notifier; 638 639 /* Use for delayed freeing of kfd_process structure */ 640 struct rcu_head rcu; 641 642 unsigned int pasid; 643 unsigned int doorbell_index; 644 645 /* 646 * List of kfd_process_device structures, 647 * one for each device the process is using. 648 */ 649 struct list_head per_device_data; 650 651 struct process_queue_manager pqm; 652 653 /*Is the user space process 32 bit?*/ 654 bool is_32bit_user_mode; 655 656 /* Event-related data */ 657 struct mutex event_mutex; 658 /* Event ID allocator and lookup */ 659 struct idr event_idr; 660 /* Event page */ 661 struct kfd_signal_page *signal_page; 662 size_t signal_mapped_size; 663 size_t signal_event_count; 664 bool signal_event_limit_reached; 665 666 /* Information used for memory eviction */ 667 void *kgd_process_info; 668 /* Eviction fence that is attached to all the BOs of this process. The 669 * fence will be triggered during eviction and new one will be created 670 * during restore 671 */ 672 struct dma_fence *ef; 673 674 /* Work items for evicting and restoring BOs */ 675 struct delayed_work eviction_work; 676 struct delayed_work restore_work; 677 /* seqno of the last scheduled eviction */ 678 unsigned int last_eviction_seqno; 679 /* Approx. the last timestamp (in jiffies) when the process was 680 * restored after an eviction 681 */ 682 unsigned long last_restore_timestamp; 683 }; 684 685 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 686 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 687 extern struct srcu_struct kfd_processes_srcu; 688 689 /** 690 * Ioctl function type. 691 * 692 * \param filep pointer to file structure. 693 * \param p amdkfd process pointer. 694 * \param data pointer to arg that was copied from user. 695 */ 696 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 697 void *data); 698 699 struct amdkfd_ioctl_desc { 700 unsigned int cmd; 701 int flags; 702 amdkfd_ioctl_t *func; 703 unsigned int cmd_drv; 704 const char *name; 705 }; 706 bool kfd_dev_is_large_bar(struct kfd_dev *dev); 707 708 int kfd_process_create_wq(void); 709 void kfd_process_destroy_wq(void); 710 struct kfd_process *kfd_create_process(struct file *filep); 711 struct kfd_process *kfd_get_process(const struct task_struct *); 712 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 713 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 714 void kfd_unref_process(struct kfd_process *p); 715 int kfd_process_evict_queues(struct kfd_process *p); 716 int kfd_process_restore_queues(struct kfd_process *p); 717 void kfd_suspend_all_processes(void); 718 int kfd_resume_all_processes(void); 719 720 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 721 struct file *drm_file); 722 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 723 struct kfd_process *p); 724 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 725 struct kfd_process *p); 726 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 727 struct kfd_process *p); 728 729 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 730 struct vm_area_struct *vma); 731 732 /* KFD process API for creating and translating handles */ 733 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 734 void *mem); 735 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 736 int handle); 737 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 738 int handle); 739 740 /* Process device data iterator */ 741 struct kfd_process_device *kfd_get_first_process_device_data( 742 struct kfd_process *p); 743 struct kfd_process_device *kfd_get_next_process_device_data( 744 struct kfd_process *p, 745 struct kfd_process_device *pdd); 746 bool kfd_has_process_device_data(struct kfd_process *p); 747 748 /* PASIDs */ 749 int kfd_pasid_init(void); 750 void kfd_pasid_exit(void); 751 bool kfd_set_pasid_limit(unsigned int new_limit); 752 unsigned int kfd_get_pasid_limit(void); 753 unsigned int kfd_pasid_alloc(void); 754 void kfd_pasid_free(unsigned int pasid); 755 756 /* Doorbells */ 757 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 758 int kfd_doorbell_init(struct kfd_dev *kfd); 759 void kfd_doorbell_fini(struct kfd_dev *kfd); 760 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 761 struct vm_area_struct *vma); 762 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 763 unsigned int *doorbell_off); 764 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 765 u32 read_kernel_doorbell(u32 __iomem *db); 766 void write_kernel_doorbell(void __iomem *db, u32 value); 767 void write_kernel_doorbell64(void __iomem *db, u64 value); 768 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd, 769 struct kfd_process *process, 770 unsigned int doorbell_id); 771 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 772 struct kfd_process *process); 773 int kfd_alloc_process_doorbells(struct kfd_process *process); 774 void kfd_free_process_doorbells(struct kfd_process *process); 775 776 /* GTT Sub-Allocator */ 777 778 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 779 struct kfd_mem_obj **mem_obj); 780 781 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 782 783 extern struct device *kfd_device; 784 785 /* Topology */ 786 int kfd_topology_init(void); 787 void kfd_topology_shutdown(void); 788 int kfd_topology_add_device(struct kfd_dev *gpu); 789 int kfd_topology_remove_device(struct kfd_dev *gpu); 790 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 791 uint32_t proximity_domain); 792 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 793 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 794 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 795 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 796 int kfd_numa_node_to_apic_id(int numa_node_id); 797 798 /* Interrupts */ 799 int kfd_interrupt_init(struct kfd_dev *dev); 800 void kfd_interrupt_exit(struct kfd_dev *dev); 801 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 802 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 803 bool interrupt_is_wanted(struct kfd_dev *dev, 804 const uint32_t *ih_ring_entry, 805 uint32_t *patched_ihre, bool *flag); 806 807 /* Power Management */ 808 void kgd2kfd_suspend(struct kfd_dev *kfd); 809 int kgd2kfd_resume(struct kfd_dev *kfd); 810 811 /* GPU reset */ 812 int kgd2kfd_pre_reset(struct kfd_dev *kfd); 813 int kgd2kfd_post_reset(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 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); 842 843 /* Process Queue Manager */ 844 struct process_queue_node { 845 struct queue *q; 846 struct kernel_queue *kq; 847 struct list_head process_queue_list; 848 }; 849 850 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 851 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 852 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 853 void pqm_uninit(struct process_queue_manager *pqm); 854 int pqm_create_queue(struct process_queue_manager *pqm, 855 struct kfd_dev *dev, 856 struct file *f, 857 struct queue_properties *properties, 858 unsigned int *qid); 859 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 860 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 861 struct queue_properties *p); 862 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 863 struct queue_properties *p); 864 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 865 unsigned int qid); 866 int pqm_get_wave_state(struct process_queue_manager *pqm, 867 unsigned int qid, 868 void __user *ctl_stack, 869 u32 *ctl_stack_used_size, 870 u32 *save_area_used_size); 871 872 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 873 unsigned int fence_value, 874 unsigned int timeout_ms); 875 876 /* Packet Manager */ 877 878 #define KFD_FENCE_COMPLETED (100) 879 #define KFD_FENCE_INIT (10) 880 881 struct packet_manager { 882 struct device_queue_manager *dqm; 883 struct kernel_queue *priv_queue; 884 struct mutex lock; 885 bool allocated; 886 struct kfd_mem_obj *ib_buffer_obj; 887 unsigned int ib_size_bytes; 888 889 const struct packet_manager_funcs *pmf; 890 }; 891 892 struct packet_manager_funcs { 893 /* Support ASIC-specific packet formats for PM4 packets */ 894 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 895 struct qcm_process_device *qpd); 896 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 897 uint64_t ib, size_t ib_size_in_dwords, bool chain); 898 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 899 struct scheduling_resources *res); 900 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 901 struct queue *q, bool is_static); 902 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 903 enum kfd_queue_type type, 904 enum kfd_unmap_queues_filter mode, 905 uint32_t filter_param, bool reset, 906 unsigned int sdma_engine); 907 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 908 uint64_t fence_address, uint32_t fence_value); 909 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 910 911 /* Packet sizes */ 912 int map_process_size; 913 int runlist_size; 914 int set_resources_size; 915 int map_queues_size; 916 int unmap_queues_size; 917 int query_status_size; 918 int release_mem_size; 919 }; 920 921 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 922 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 923 924 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 925 void pm_uninit(struct packet_manager *pm); 926 int pm_send_set_resources(struct packet_manager *pm, 927 struct scheduling_resources *res); 928 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 929 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 930 uint32_t fence_value); 931 932 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 933 enum kfd_unmap_queues_filter mode, 934 uint32_t filter_param, bool reset, 935 unsigned int sdma_engine); 936 937 void pm_release_ib(struct packet_manager *pm); 938 939 /* Following PM funcs can be shared among VI and AI */ 940 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 941 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer, 942 struct scheduling_resources *res); 943 944 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 945 946 /* Events */ 947 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 948 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 949 950 extern const struct kfd_device_global_init_class device_global_init_class_cik; 951 952 void kfd_event_init_process(struct kfd_process *p); 953 void kfd_event_free_process(struct kfd_process *p); 954 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 955 int kfd_wait_on_events(struct kfd_process *p, 956 uint32_t num_events, void __user *data, 957 bool all, uint32_t user_timeout_ms, 958 uint32_t *wait_result); 959 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 960 uint32_t valid_id_bits); 961 void kfd_signal_iommu_event(struct kfd_dev *dev, 962 unsigned int pasid, unsigned long address, 963 bool is_write_requested, bool is_execute_requested); 964 void kfd_signal_hw_exception_event(unsigned int pasid); 965 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 966 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 967 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 968 uint64_t size); 969 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 970 uint32_t event_type, bool auto_reset, uint32_t node_id, 971 uint32_t *event_id, uint32_t *event_trigger_data, 972 uint64_t *event_page_offset, uint32_t *event_slot_index); 973 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 974 975 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid, 976 struct kfd_vm_fault_info *info); 977 978 void kfd_signal_reset_event(struct kfd_dev *dev); 979 980 void kfd_flush_tlb(struct kfd_process_device *pdd); 981 982 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 983 984 bool kfd_is_locked(void); 985 986 /* Debugfs */ 987 #if defined(CONFIG_DEBUG_FS) 988 989 void kfd_debugfs_init(void); 990 void kfd_debugfs_fini(void); 991 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 992 int pqm_debugfs_mqds(struct seq_file *m, void *data); 993 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 994 int dqm_debugfs_hqds(struct seq_file *m, void *data); 995 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 996 int pm_debugfs_runlist(struct seq_file *m, void *data); 997 998 int kfd_debugfs_hang_hws(struct kfd_dev *dev); 999 int pm_debugfs_hang_hws(struct packet_manager *pm); 1000 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1001 1002 #else 1003 1004 static inline void kfd_debugfs_init(void) {} 1005 static inline void kfd_debugfs_fini(void) {} 1006 1007 #endif 1008 1009 #endif 1010