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