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 <linux/sysfs.h> 39 #include <linux/device_cgroup.h> 40 #include <drm/drm_file.h> 41 #include <drm/drm_drv.h> 42 #include <drm/drm_device.h> 43 #include <kgd_kfd_interface.h> 44 #include <linux/swap.h> 45 46 #include "amd_shared.h" 47 48 #define KFD_MAX_RING_ENTRY_SIZE 8 49 50 #define KFD_SYSFS_FILE_MODE 0444 51 52 /* GPU ID hash width in bits */ 53 #define KFD_GPU_ID_HASH_WIDTH 16 54 55 /* Use upper bits of mmap offset to store KFD driver specific information. 56 * BITS[63:62] - Encode MMAP type 57 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 58 * BITS[45:0] - MMAP offset value 59 * 60 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 61 * defines are w.r.t to PAGE_SIZE 62 */ 63 #define KFD_MMAP_TYPE_SHIFT 62 64 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 65 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 66 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 67 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 68 #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) 69 70 #define KFD_MMAP_GPU_ID_SHIFT 46 71 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 72 << KFD_MMAP_GPU_ID_SHIFT) 73 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 74 & KFD_MMAP_GPU_ID_MASK) 75 #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 76 >> KFD_MMAP_GPU_ID_SHIFT) 77 78 /* 79 * When working with cp scheduler we should assign the HIQ manually or via 80 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot 81 * definitions for Kaveri. In Kaveri only the first ME queues participates 82 * in the cp scheduling taking that in mind we set the HIQ slot in the 83 * second ME. 84 */ 85 #define KFD_CIK_HIQ_PIPE 4 86 #define KFD_CIK_HIQ_QUEUE 0 87 88 /* Macro for allocating structures */ 89 #define kfd_alloc_struct(ptr_to_struct) \ 90 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 91 92 #define KFD_MAX_NUM_OF_PROCESSES 512 93 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 94 95 /* 96 * Size of the per-process TBA+TMA buffer: 2 pages 97 * 98 * The first page is the TBA used for the CWSR ISA code. The second 99 * page is used as TMA for daisy changing a user-mode trap handler. 100 */ 101 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 102 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 103 104 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 105 (KFD_MAX_NUM_OF_PROCESSES * \ 106 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 107 108 #define KFD_KERNEL_QUEUE_SIZE 2048 109 110 #define KFD_UNMAP_LATENCY_MS (4000) 111 112 /* 113 * 512 = 0x200 114 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the 115 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. 116 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC 117 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in 118 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. 119 */ 120 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 121 122 123 /* 124 * Kernel module parameter to specify maximum number of supported queues per 125 * device 126 */ 127 extern int max_num_of_queues_per_device; 128 129 130 /* Kernel module parameter to specify the scheduling policy */ 131 extern int sched_policy; 132 133 /* 134 * Kernel module parameter to specify the maximum process 135 * number per HW scheduler 136 */ 137 extern int hws_max_conc_proc; 138 139 extern int cwsr_enable; 140 141 /* 142 * Kernel module parameter to specify whether to send sigterm to HSA process on 143 * unhandled exception 144 */ 145 extern int send_sigterm; 146 147 /* 148 * This kernel module is used to simulate large bar machine on non-large bar 149 * enabled machines. 150 */ 151 extern int debug_largebar; 152 153 /* 154 * Ignore CRAT table during KFD initialization, can be used to work around 155 * broken CRAT tables on some AMD systems 156 */ 157 extern int ignore_crat; 158 159 /* 160 * Set sh_mem_config.retry_disable on Vega10 161 */ 162 extern int amdgpu_noretry; 163 164 /* 165 * Halt if HWS hang is detected 166 */ 167 extern int halt_if_hws_hang; 168 169 /* 170 * Whether MEC FW support GWS barriers 171 */ 172 extern bool hws_gws_support; 173 174 /* 175 * Queue preemption timeout in ms 176 */ 177 extern int queue_preemption_timeout_ms; 178 179 enum cache_policy { 180 cache_policy_coherent, 181 cache_policy_noncoherent 182 }; 183 184 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) 185 186 struct kfd_event_interrupt_class { 187 bool (*interrupt_isr)(struct kfd_dev *dev, 188 const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 189 bool *patched_flag); 190 void (*interrupt_wq)(struct kfd_dev *dev, 191 const uint32_t *ih_ring_entry); 192 }; 193 194 struct kfd_device_info { 195 enum amd_asic_type asic_family; 196 const char *asic_name; 197 const struct kfd_event_interrupt_class *event_interrupt_class; 198 unsigned int max_pasid_bits; 199 unsigned int max_no_of_hqd; 200 unsigned int doorbell_size; 201 size_t ih_ring_entry_size; 202 uint8_t num_of_watch_points; 203 uint16_t mqd_size_aligned; 204 bool supports_cwsr; 205 bool needs_iommu_device; 206 bool needs_pci_atomics; 207 unsigned int num_sdma_engines; 208 unsigned int num_xgmi_sdma_engines; 209 unsigned int num_sdma_queues_per_engine; 210 }; 211 212 struct kfd_mem_obj { 213 uint32_t range_start; 214 uint32_t range_end; 215 uint64_t gpu_addr; 216 uint32_t *cpu_ptr; 217 void *gtt_mem; 218 }; 219 220 struct kfd_vmid_info { 221 uint32_t first_vmid_kfd; 222 uint32_t last_vmid_kfd; 223 uint32_t vmid_num_kfd; 224 }; 225 226 struct kfd_dev { 227 struct kgd_dev *kgd; 228 229 const struct kfd_device_info *device_info; 230 struct pci_dev *pdev; 231 struct drm_device *ddev; 232 233 unsigned int id; /* topology stub index */ 234 235 phys_addr_t doorbell_base; /* Start of actual doorbells used by 236 * KFD. It is aligned for mapping 237 * into user mode 238 */ 239 size_t doorbell_base_dw_offset; /* Offset from the start of the PCI 240 * doorbell BAR to the first KFD 241 * doorbell in dwords. GFX reserves 242 * the segment before this offset. 243 */ 244 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 245 * page used by kernel queue 246 */ 247 248 struct kgd2kfd_shared_resources shared_resources; 249 struct kfd_vmid_info vm_info; 250 251 const struct kfd2kgd_calls *kfd2kgd; 252 struct mutex doorbell_mutex; 253 DECLARE_BITMAP(doorbell_available_index, 254 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 255 256 void *gtt_mem; 257 uint64_t gtt_start_gpu_addr; 258 void *gtt_start_cpu_ptr; 259 void *gtt_sa_bitmap; 260 struct mutex gtt_sa_lock; 261 unsigned int gtt_sa_chunk_size; 262 unsigned int gtt_sa_num_of_chunks; 263 264 /* Interrupts */ 265 struct kfifo ih_fifo; 266 struct workqueue_struct *ih_wq; 267 struct work_struct interrupt_work; 268 spinlock_t interrupt_lock; 269 270 /* QCM Device instance */ 271 struct device_queue_manager *dqm; 272 273 bool init_complete; 274 /* 275 * Interrupts of interest to KFD are copied 276 * from the HW ring into a SW ring. 277 */ 278 bool interrupts_active; 279 280 /* Debug manager */ 281 struct kfd_dbgmgr *dbgmgr; 282 283 /* Firmware versions */ 284 uint16_t mec_fw_version; 285 uint16_t sdma_fw_version; 286 287 /* Maximum process number mapped to HW scheduler */ 288 unsigned int max_proc_per_quantum; 289 290 /* CWSR */ 291 bool cwsr_enabled; 292 const void *cwsr_isa; 293 unsigned int cwsr_isa_size; 294 295 /* xGMI */ 296 uint64_t hive_id; 297 298 /* UUID */ 299 uint64_t unique_id; 300 301 bool pci_atomic_requested; 302 303 /* SRAM ECC flag */ 304 atomic_t sram_ecc_flag; 305 306 /* Compute Profile ref. count */ 307 atomic_t compute_profile; 308 309 /* Global GWS resource shared b/t processes*/ 310 void *gws; 311 }; 312 313 enum kfd_mempool { 314 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 315 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 316 KFD_MEMPOOL_FRAMEBUFFER = 3, 317 }; 318 319 /* Character device interface */ 320 int kfd_chardev_init(void); 321 void kfd_chardev_exit(void); 322 struct device *kfd_chardev(void); 323 324 /** 325 * enum kfd_unmap_queues_filter 326 * 327 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 328 * 329 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 330 * running queues list. 331 * 332 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 333 * specific process. 334 * 335 */ 336 enum kfd_unmap_queues_filter { 337 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 338 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 339 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 340 KFD_UNMAP_QUEUES_FILTER_BY_PASID 341 }; 342 343 /** 344 * enum kfd_queue_type 345 * 346 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 347 * 348 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 349 * 350 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 351 * 352 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 353 */ 354 enum kfd_queue_type { 355 KFD_QUEUE_TYPE_COMPUTE, 356 KFD_QUEUE_TYPE_SDMA, 357 KFD_QUEUE_TYPE_HIQ, 358 KFD_QUEUE_TYPE_DIQ, 359 KFD_QUEUE_TYPE_SDMA_XGMI 360 }; 361 362 enum kfd_queue_format { 363 KFD_QUEUE_FORMAT_PM4, 364 KFD_QUEUE_FORMAT_AQL 365 }; 366 367 enum KFD_QUEUE_PRIORITY { 368 KFD_QUEUE_PRIORITY_MINIMUM = 0, 369 KFD_QUEUE_PRIORITY_MAXIMUM = 15 370 }; 371 372 /** 373 * struct queue_properties 374 * 375 * @type: The queue type. 376 * 377 * @queue_id: Queue identifier. 378 * 379 * @queue_address: Queue ring buffer address. 380 * 381 * @queue_size: Queue ring buffer size. 382 * 383 * @priority: Defines the queue priority relative to other queues in the 384 * process. 385 * This is just an indication and HW scheduling may override the priority as 386 * necessary while keeping the relative prioritization. 387 * the priority granularity is from 0 to f which f is the highest priority. 388 * currently all queues are initialized with the highest priority. 389 * 390 * @queue_percent: This field is partially implemented and currently a zero in 391 * this field defines that the queue is non active. 392 * 393 * @read_ptr: User space address which points to the number of dwords the 394 * cp read from the ring buffer. This field updates automatically by the H/W. 395 * 396 * @write_ptr: Defines the number of dwords written to the ring buffer. 397 * 398 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 399 * the queue ring buffer. This field should be similar to write_ptr and the 400 * user should update this field after he updated the write_ptr. 401 * 402 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 403 * 404 * @is_interop: Defines if this is a interop queue. Interop queue means that 405 * the queue can access both graphics and compute resources. 406 * 407 * @is_evicted: Defines if the queue is evicted. Only active queues 408 * are evicted, rendering them inactive. 409 * 410 * @is_active: Defines if the queue is active or not. @is_active and 411 * @is_evicted are protected by the DQM lock. 412 * 413 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 414 * of the queue. 415 * 416 * This structure represents the queue properties for each queue no matter if 417 * it's user mode or kernel mode queue. 418 * 419 */ 420 struct queue_properties { 421 enum kfd_queue_type type; 422 enum kfd_queue_format format; 423 unsigned int queue_id; 424 uint64_t queue_address; 425 uint64_t queue_size; 426 uint32_t priority; 427 uint32_t queue_percent; 428 uint32_t *read_ptr; 429 uint32_t *write_ptr; 430 void __iomem *doorbell_ptr; 431 uint32_t doorbell_off; 432 bool is_interop; 433 bool is_evicted; 434 bool is_active; 435 /* Not relevant for user mode queues in cp scheduling */ 436 unsigned int vmid; 437 /* Relevant only for sdma queues*/ 438 uint32_t sdma_engine_id; 439 uint32_t sdma_queue_id; 440 uint32_t sdma_vm_addr; 441 /* Relevant only for VI */ 442 uint64_t eop_ring_buffer_address; 443 uint32_t eop_ring_buffer_size; 444 uint64_t ctx_save_restore_area_address; 445 uint32_t ctx_save_restore_area_size; 446 uint32_t ctl_stack_size; 447 uint64_t tba_addr; 448 uint64_t tma_addr; 449 /* Relevant for CU */ 450 uint32_t cu_mask_count; /* Must be a multiple of 32 */ 451 uint32_t *cu_mask; 452 }; 453 454 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 455 (q).queue_address != 0 && \ 456 (q).queue_percent > 0 && \ 457 !(q).is_evicted) 458 459 /** 460 * struct queue 461 * 462 * @list: Queue linked list. 463 * 464 * @mqd: The queue MQD. 465 * 466 * @mqd_mem_obj: The MQD local gpu memory object. 467 * 468 * @gart_mqd_addr: The MQD gart mc address. 469 * 470 * @properties: The queue properties. 471 * 472 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 473 * that the queue should be execute on. 474 * 475 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 476 * id. 477 * 478 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 479 * 480 * @process: The kfd process that created this queue. 481 * 482 * @device: The kfd device that created this queue. 483 * 484 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 485 * otherwise. 486 * 487 * This structure represents user mode compute queues. 488 * It contains all the necessary data to handle such queues. 489 * 490 */ 491 492 struct queue { 493 struct list_head list; 494 void *mqd; 495 struct kfd_mem_obj *mqd_mem_obj; 496 uint64_t gart_mqd_addr; 497 struct queue_properties properties; 498 499 uint32_t mec; 500 uint32_t pipe; 501 uint32_t queue; 502 503 unsigned int sdma_id; 504 unsigned int doorbell_id; 505 506 struct kfd_process *process; 507 struct kfd_dev *device; 508 void *gws; 509 510 /* procfs */ 511 struct kobject kobj; 512 }; 513 514 /* 515 * Please read the kfd_mqd_manager.h description. 516 */ 517 enum KFD_MQD_TYPE { 518 KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 519 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 520 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 521 KFD_MQD_TYPE_DIQ, /* for diq */ 522 KFD_MQD_TYPE_MAX 523 }; 524 525 enum KFD_PIPE_PRIORITY { 526 KFD_PIPE_PRIORITY_CS_LOW = 0, 527 KFD_PIPE_PRIORITY_CS_MEDIUM, 528 KFD_PIPE_PRIORITY_CS_HIGH 529 }; 530 531 struct scheduling_resources { 532 unsigned int vmid_mask; 533 enum kfd_queue_type type; 534 uint64_t queue_mask; 535 uint64_t gws_mask; 536 uint32_t oac_mask; 537 uint32_t gds_heap_base; 538 uint32_t gds_heap_size; 539 }; 540 541 struct process_queue_manager { 542 /* data */ 543 struct kfd_process *process; 544 struct list_head queues; 545 unsigned long *queue_slot_bitmap; 546 }; 547 548 struct qcm_process_device { 549 /* The Device Queue Manager that owns this data */ 550 struct device_queue_manager *dqm; 551 struct process_queue_manager *pqm; 552 /* Queues list */ 553 struct list_head queues_list; 554 struct list_head priv_queue_list; 555 556 unsigned int queue_count; 557 unsigned int vmid; 558 bool is_debug; 559 unsigned int evicted; /* eviction counter, 0=active */ 560 561 /* This flag tells if we should reset all wavefronts on 562 * process termination 563 */ 564 bool reset_wavefronts; 565 566 /* 567 * All the memory management data should be here too 568 */ 569 uint64_t gds_context_area; 570 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 571 uint64_t page_table_base; 572 uint32_t sh_mem_config; 573 uint32_t sh_mem_bases; 574 uint32_t sh_mem_ape1_base; 575 uint32_t sh_mem_ape1_limit; 576 uint32_t gds_size; 577 uint32_t num_gws; 578 uint32_t num_oac; 579 uint32_t sh_hidden_private_base; 580 581 /* CWSR memory */ 582 void *cwsr_kaddr; 583 uint64_t cwsr_base; 584 uint64_t tba_addr; 585 uint64_t tma_addr; 586 587 /* IB memory */ 588 uint64_t ib_base; 589 void *ib_kaddr; 590 591 /* doorbell resources per process per device */ 592 unsigned long *doorbell_bitmap; 593 }; 594 595 /* KFD Memory Eviction */ 596 597 /* Approx. wait time before attempting to restore evicted BOs */ 598 #define PROCESS_RESTORE_TIME_MS 100 599 /* Approx. back off time if restore fails due to lack of memory */ 600 #define PROCESS_BACK_OFF_TIME_MS 100 601 /* Approx. time before evicting the process again */ 602 #define PROCESS_ACTIVE_TIME_MS 10 603 604 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 605 * idr_handle in the least significant 4 bytes 606 */ 607 #define MAKE_HANDLE(gpu_id, idr_handle) \ 608 (((uint64_t)(gpu_id) << 32) + idr_handle) 609 #define GET_GPU_ID(handle) (handle >> 32) 610 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 611 612 enum kfd_pdd_bound { 613 PDD_UNBOUND = 0, 614 PDD_BOUND, 615 PDD_BOUND_SUSPENDED, 616 }; 617 618 /* Data that is per-process-per device. */ 619 struct kfd_process_device { 620 /* 621 * List of all per-device data for a process. 622 * Starts from kfd_process.per_device_data. 623 */ 624 struct list_head per_device_list; 625 626 /* The device that owns this data. */ 627 struct kfd_dev *dev; 628 629 /* The process that owns this kfd_process_device. */ 630 struct kfd_process *process; 631 632 /* per-process-per device QCM data structure */ 633 struct qcm_process_device qpd; 634 635 /*Apertures*/ 636 uint64_t lds_base; 637 uint64_t lds_limit; 638 uint64_t gpuvm_base; 639 uint64_t gpuvm_limit; 640 uint64_t scratch_base; 641 uint64_t scratch_limit; 642 643 /* VM context for GPUVM allocations */ 644 struct file *drm_file; 645 void *vm; 646 647 /* GPUVM allocations storage */ 648 struct idr alloc_idr; 649 650 /* Flag used to tell the pdd has dequeued from the dqm. 651 * This is used to prevent dev->dqm->ops.process_termination() from 652 * being called twice when it is already called in IOMMU callback 653 * function. 654 */ 655 bool already_dequeued; 656 bool runtime_inuse; 657 658 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 659 enum kfd_pdd_bound bound; 660 }; 661 662 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 663 664 /* Process data */ 665 struct kfd_process { 666 /* 667 * kfd_process are stored in an mm_struct*->kfd_process* 668 * hash table (kfd_processes in kfd_process.c) 669 */ 670 struct hlist_node kfd_processes; 671 672 /* 673 * Opaque pointer to mm_struct. We don't hold a reference to 674 * it so it should never be dereferenced from here. This is 675 * only used for looking up processes by their mm. 676 */ 677 void *mm; 678 679 struct kref ref; 680 struct work_struct release_work; 681 682 struct mutex mutex; 683 684 /* 685 * In any process, the thread that started main() is the lead 686 * thread and outlives the rest. 687 * It is here because amd_iommu_bind_pasid wants a task_struct. 688 * It can also be used for safely getting a reference to the 689 * mm_struct of the process. 690 */ 691 struct task_struct *lead_thread; 692 693 /* We want to receive a notification when the mm_struct is destroyed */ 694 struct mmu_notifier mmu_notifier; 695 696 uint16_t pasid; 697 unsigned int doorbell_index; 698 699 /* 700 * List of kfd_process_device structures, 701 * one for each device the process is using. 702 */ 703 struct list_head per_device_data; 704 705 struct process_queue_manager pqm; 706 707 /*Is the user space process 32 bit?*/ 708 bool is_32bit_user_mode; 709 710 /* Event-related data */ 711 struct mutex event_mutex; 712 /* Event ID allocator and lookup */ 713 struct idr event_idr; 714 /* Event page */ 715 struct kfd_signal_page *signal_page; 716 size_t signal_mapped_size; 717 size_t signal_event_count; 718 bool signal_event_limit_reached; 719 720 /* Information used for memory eviction */ 721 void *kgd_process_info; 722 /* Eviction fence that is attached to all the BOs of this process. The 723 * fence will be triggered during eviction and new one will be created 724 * during restore 725 */ 726 struct dma_fence *ef; 727 728 /* Work items for evicting and restoring BOs */ 729 struct delayed_work eviction_work; 730 struct delayed_work restore_work; 731 /* seqno of the last scheduled eviction */ 732 unsigned int last_eviction_seqno; 733 /* Approx. the last timestamp (in jiffies) when the process was 734 * restored after an eviction 735 */ 736 unsigned long last_restore_timestamp; 737 738 /* Kobj for our procfs */ 739 struct kobject *kobj; 740 struct kobject *kobj_queues; 741 struct attribute attr_pasid; 742 }; 743 744 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 745 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 746 extern struct srcu_struct kfd_processes_srcu; 747 748 /** 749 * Ioctl function type. 750 * 751 * \param filep pointer to file structure. 752 * \param p amdkfd process pointer. 753 * \param data pointer to arg that was copied from user. 754 */ 755 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 756 void *data); 757 758 struct amdkfd_ioctl_desc { 759 unsigned int cmd; 760 int flags; 761 amdkfd_ioctl_t *func; 762 unsigned int cmd_drv; 763 const char *name; 764 }; 765 bool kfd_dev_is_large_bar(struct kfd_dev *dev); 766 767 int kfd_process_create_wq(void); 768 void kfd_process_destroy_wq(void); 769 struct kfd_process *kfd_create_process(struct file *filep); 770 struct kfd_process *kfd_get_process(const struct task_struct *); 771 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 772 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 773 void kfd_unref_process(struct kfd_process *p); 774 int kfd_process_evict_queues(struct kfd_process *p); 775 int kfd_process_restore_queues(struct kfd_process *p); 776 void kfd_suspend_all_processes(void); 777 int kfd_resume_all_processes(void); 778 779 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 780 struct file *drm_file); 781 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 782 struct kfd_process *p); 783 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 784 struct kfd_process *p); 785 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 786 struct kfd_process *p); 787 788 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 789 struct vm_area_struct *vma); 790 791 /* KFD process API for creating and translating handles */ 792 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 793 void *mem); 794 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 795 int handle); 796 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 797 int handle); 798 799 /* Process device data iterator */ 800 struct kfd_process_device *kfd_get_first_process_device_data( 801 struct kfd_process *p); 802 struct kfd_process_device *kfd_get_next_process_device_data( 803 struct kfd_process *p, 804 struct kfd_process_device *pdd); 805 bool kfd_has_process_device_data(struct kfd_process *p); 806 807 /* PASIDs */ 808 int kfd_pasid_init(void); 809 void kfd_pasid_exit(void); 810 bool kfd_set_pasid_limit(unsigned int new_limit); 811 unsigned int kfd_get_pasid_limit(void); 812 unsigned int kfd_pasid_alloc(void); 813 void kfd_pasid_free(unsigned int pasid); 814 815 /* Doorbells */ 816 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 817 int kfd_doorbell_init(struct kfd_dev *kfd); 818 void kfd_doorbell_fini(struct kfd_dev *kfd); 819 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 820 struct vm_area_struct *vma); 821 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 822 unsigned int *doorbell_off); 823 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 824 u32 read_kernel_doorbell(u32 __iomem *db); 825 void write_kernel_doorbell(void __iomem *db, u32 value); 826 void write_kernel_doorbell64(void __iomem *db, u64 value); 827 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 828 struct kfd_process *process, 829 unsigned int doorbell_id); 830 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 831 struct kfd_process *process); 832 int kfd_alloc_process_doorbells(struct kfd_process *process); 833 void kfd_free_process_doorbells(struct kfd_process *process); 834 835 /* GTT Sub-Allocator */ 836 837 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 838 struct kfd_mem_obj **mem_obj); 839 840 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 841 842 extern struct device *kfd_device; 843 844 /* KFD's procfs */ 845 void kfd_procfs_init(void); 846 void kfd_procfs_shutdown(void); 847 int kfd_procfs_add_queue(struct queue *q); 848 void kfd_procfs_del_queue(struct queue *q); 849 850 /* Topology */ 851 int kfd_topology_init(void); 852 void kfd_topology_shutdown(void); 853 int kfd_topology_add_device(struct kfd_dev *gpu); 854 int kfd_topology_remove_device(struct kfd_dev *gpu); 855 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 856 uint32_t proximity_domain); 857 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 858 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 859 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 860 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd); 861 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 862 int kfd_numa_node_to_apic_id(int numa_node_id); 863 864 /* Interrupts */ 865 int kfd_interrupt_init(struct kfd_dev *dev); 866 void kfd_interrupt_exit(struct kfd_dev *dev); 867 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 868 bool interrupt_is_wanted(struct kfd_dev *dev, 869 const uint32_t *ih_ring_entry, 870 uint32_t *patched_ihre, bool *flag); 871 872 /* amdkfd Apertures */ 873 int kfd_init_apertures(struct kfd_process *process); 874 875 /* Queue Context Management */ 876 int init_queue(struct queue **q, const struct queue_properties *properties); 877 void uninit_queue(struct queue *q); 878 void print_queue_properties(struct queue_properties *q); 879 void print_queue(struct queue *q); 880 881 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 882 struct kfd_dev *dev); 883 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 884 struct kfd_dev *dev); 885 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 886 struct kfd_dev *dev); 887 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 888 struct kfd_dev *dev); 889 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 890 struct kfd_dev *dev); 891 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 892 struct kfd_dev *dev); 893 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 894 void device_queue_manager_uninit(struct device_queue_manager *dqm); 895 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 896 enum kfd_queue_type type); 897 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 898 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); 899 900 /* Process Queue Manager */ 901 struct process_queue_node { 902 struct queue *q; 903 struct kernel_queue *kq; 904 struct list_head process_queue_list; 905 }; 906 907 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 908 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 909 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 910 void pqm_uninit(struct process_queue_manager *pqm); 911 int pqm_create_queue(struct process_queue_manager *pqm, 912 struct kfd_dev *dev, 913 struct file *f, 914 struct queue_properties *properties, 915 unsigned int *qid, 916 uint32_t *p_doorbell_offset_in_process); 917 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 918 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 919 struct queue_properties *p); 920 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 921 struct queue_properties *p); 922 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 923 void *gws); 924 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 925 unsigned int qid); 926 int pqm_get_wave_state(struct process_queue_manager *pqm, 927 unsigned int qid, 928 void __user *ctl_stack, 929 u32 *ctl_stack_used_size, 930 u32 *save_area_used_size); 931 932 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 933 unsigned int fence_value, 934 unsigned int timeout_ms); 935 936 /* Packet Manager */ 937 938 #define KFD_FENCE_COMPLETED (100) 939 #define KFD_FENCE_INIT (10) 940 941 struct packet_manager { 942 struct device_queue_manager *dqm; 943 struct kernel_queue *priv_queue; 944 struct mutex lock; 945 bool allocated; 946 struct kfd_mem_obj *ib_buffer_obj; 947 unsigned int ib_size_bytes; 948 bool is_over_subscription; 949 950 const struct packet_manager_funcs *pmf; 951 }; 952 953 struct packet_manager_funcs { 954 /* Support ASIC-specific packet formats for PM4 packets */ 955 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 956 struct qcm_process_device *qpd); 957 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 958 uint64_t ib, size_t ib_size_in_dwords, bool chain); 959 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 960 struct scheduling_resources *res); 961 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 962 struct queue *q, bool is_static); 963 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 964 enum kfd_queue_type type, 965 enum kfd_unmap_queues_filter mode, 966 uint32_t filter_param, bool reset, 967 unsigned int sdma_engine); 968 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 969 uint64_t fence_address, uint32_t fence_value); 970 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 971 972 /* Packet sizes */ 973 int map_process_size; 974 int runlist_size; 975 int set_resources_size; 976 int map_queues_size; 977 int unmap_queues_size; 978 int query_status_size; 979 int release_mem_size; 980 }; 981 982 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 983 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 984 985 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 986 void pm_uninit(struct packet_manager *pm, bool hanging); 987 int pm_send_set_resources(struct packet_manager *pm, 988 struct scheduling_resources *res); 989 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 990 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 991 uint32_t fence_value); 992 993 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 994 enum kfd_unmap_queues_filter mode, 995 uint32_t filter_param, bool reset, 996 unsigned int sdma_engine); 997 998 void pm_release_ib(struct packet_manager *pm); 999 1000 /* Following PM funcs can be shared among VI and AI */ 1001 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 1002 1003 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 1004 1005 /* Events */ 1006 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1007 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1008 1009 extern const struct kfd_device_global_init_class device_global_init_class_cik; 1010 1011 void kfd_event_init_process(struct kfd_process *p); 1012 void kfd_event_free_process(struct kfd_process *p); 1013 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1014 int kfd_wait_on_events(struct kfd_process *p, 1015 uint32_t num_events, void __user *data, 1016 bool all, uint32_t user_timeout_ms, 1017 uint32_t *wait_result); 1018 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 1019 uint32_t valid_id_bits); 1020 void kfd_signal_iommu_event(struct kfd_dev *dev, 1021 unsigned int pasid, unsigned long address, 1022 bool is_write_requested, bool is_execute_requested); 1023 void kfd_signal_hw_exception_event(unsigned int pasid); 1024 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1025 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 1026 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 1027 uint64_t size); 1028 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1029 uint32_t event_type, bool auto_reset, uint32_t node_id, 1030 uint32_t *event_id, uint32_t *event_trigger_data, 1031 uint64_t *event_page_offset, uint32_t *event_slot_index); 1032 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1033 1034 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid, 1035 struct kfd_vm_fault_info *info); 1036 1037 void kfd_signal_reset_event(struct kfd_dev *dev); 1038 1039 void kfd_flush_tlb(struct kfd_process_device *pdd); 1040 1041 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1042 1043 bool kfd_is_locked(void); 1044 1045 /* Compute profile */ 1046 void kfd_inc_compute_active(struct kfd_dev *dev); 1047 void kfd_dec_compute_active(struct kfd_dev *dev); 1048 1049 /* Cgroup Support */ 1050 /* Check with device cgroup if @kfd device is accessible */ 1051 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 1052 { 1053 #if defined(CONFIG_CGROUP_DEVICE) 1054 struct drm_device *ddev = kfd->ddev; 1055 1056 return devcgroup_check_permission(DEVCG_DEV_CHAR, ddev->driver->major, 1057 ddev->render->index, 1058 DEVCG_ACC_WRITE | DEVCG_ACC_READ); 1059 #else 1060 return 0; 1061 #endif 1062 } 1063 1064 /* Debugfs */ 1065 #if defined(CONFIG_DEBUG_FS) 1066 1067 void kfd_debugfs_init(void); 1068 void kfd_debugfs_fini(void); 1069 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1070 int pqm_debugfs_mqds(struct seq_file *m, void *data); 1071 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1072 int dqm_debugfs_hqds(struct seq_file *m, void *data); 1073 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1074 int pm_debugfs_runlist(struct seq_file *m, void *data); 1075 1076 int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1077 int pm_debugfs_hang_hws(struct packet_manager *pm); 1078 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1079 1080 #else 1081 1082 static inline void kfd_debugfs_init(void) {} 1083 static inline void kfd_debugfs_fini(void) {} 1084 1085 #endif 1086 1087 #endif 1088