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 <drm/drm_ioctl.h> 44 #include <kgd_kfd_interface.h> 45 #include <linux/swap.h> 46 47 #include "amd_shared.h" 48 49 #define KFD_MAX_RING_ENTRY_SIZE 8 50 51 #define KFD_SYSFS_FILE_MODE 0444 52 53 /* GPU ID hash width in bits */ 54 #define KFD_GPU_ID_HASH_WIDTH 16 55 56 /* Use upper bits of mmap offset to store KFD driver specific information. 57 * BITS[63:62] - Encode MMAP type 58 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 59 * BITS[45:0] - MMAP offset value 60 * 61 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 62 * defines are w.r.t to PAGE_SIZE 63 */ 64 #define KFD_MMAP_TYPE_SHIFT 62 65 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 66 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 67 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 68 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 69 #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) 70 71 #define KFD_MMAP_GPU_ID_SHIFT 46 72 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 73 << KFD_MMAP_GPU_ID_SHIFT) 74 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 75 & KFD_MMAP_GPU_ID_MASK) 76 #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 77 >> KFD_MMAP_GPU_ID_SHIFT) 78 79 /* 80 * When working with cp scheduler we should assign the HIQ manually or via 81 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot 82 * definitions for Kaveri. In Kaveri only the first ME queues participates 83 * in the cp scheduling taking that in mind we set the HIQ slot in the 84 * second ME. 85 */ 86 #define KFD_CIK_HIQ_PIPE 4 87 #define KFD_CIK_HIQ_QUEUE 0 88 89 /* Macro for allocating structures */ 90 #define kfd_alloc_struct(ptr_to_struct) \ 91 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 92 93 #define KFD_MAX_NUM_OF_PROCESSES 512 94 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 95 96 /* 97 * Size of the per-process TBA+TMA buffer: 2 pages 98 * 99 * The first page is the TBA used for the CWSR ISA code. The second 100 * page is used as TMA for user-mode trap handler setup in daisy-chain mode. 101 */ 102 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 103 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 104 105 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 106 (KFD_MAX_NUM_OF_PROCESSES * \ 107 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 108 109 #define KFD_KERNEL_QUEUE_SIZE 2048 110 111 #define KFD_UNMAP_LATENCY_MS (4000) 112 113 /* 114 * 512 = 0x200 115 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the 116 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. 117 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC 118 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in 119 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. 120 */ 121 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 122 123 124 /* 125 * Kernel module parameter to specify maximum number of supported queues per 126 * device 127 */ 128 extern int max_num_of_queues_per_device; 129 130 131 /* Kernel module parameter to specify the scheduling policy */ 132 extern int sched_policy; 133 134 /* 135 * Kernel module parameter to specify the maximum process 136 * number per HW scheduler 137 */ 138 extern int hws_max_conc_proc; 139 140 extern int cwsr_enable; 141 142 /* 143 * Kernel module parameter to specify whether to send sigterm to HSA process on 144 * unhandled exception 145 */ 146 extern int send_sigterm; 147 148 /* 149 * This kernel module is used to simulate large bar machine on non-large bar 150 * enabled machines. 151 */ 152 extern int debug_largebar; 153 154 /* 155 * Ignore CRAT table during KFD initialization, can be used to work around 156 * broken CRAT tables on some AMD systems 157 */ 158 extern int ignore_crat; 159 160 /* Set sh_mem_config.retry_disable on GFX v9 */ 161 extern int amdgpu_noretry; 162 163 /* Halt if HWS hang is detected */ 164 extern int halt_if_hws_hang; 165 166 /* Whether MEC FW support GWS barriers */ 167 extern bool hws_gws_support; 168 169 /* Queue preemption timeout in ms */ 170 extern int queue_preemption_timeout_ms; 171 172 /* Enable eviction debug messages */ 173 extern bool debug_evictions; 174 175 enum cache_policy { 176 cache_policy_coherent, 177 cache_policy_noncoherent 178 }; 179 180 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) 181 182 struct kfd_event_interrupt_class { 183 bool (*interrupt_isr)(struct kfd_dev *dev, 184 const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 185 bool *patched_flag); 186 void (*interrupt_wq)(struct kfd_dev *dev, 187 const uint32_t *ih_ring_entry); 188 }; 189 190 struct kfd_device_info { 191 enum amd_asic_type asic_family; 192 const char *asic_name; 193 const struct kfd_event_interrupt_class *event_interrupt_class; 194 unsigned int max_pasid_bits; 195 unsigned int max_no_of_hqd; 196 unsigned int doorbell_size; 197 size_t ih_ring_entry_size; 198 uint8_t num_of_watch_points; 199 uint16_t mqd_size_aligned; 200 bool supports_cwsr; 201 bool needs_iommu_device; 202 bool needs_pci_atomics; 203 unsigned int num_sdma_engines; 204 unsigned int num_xgmi_sdma_engines; 205 unsigned int num_sdma_queues_per_engine; 206 }; 207 208 struct kfd_mem_obj { 209 uint32_t range_start; 210 uint32_t range_end; 211 uint64_t gpu_addr; 212 uint32_t *cpu_ptr; 213 void *gtt_mem; 214 }; 215 216 struct kfd_vmid_info { 217 uint32_t first_vmid_kfd; 218 uint32_t last_vmid_kfd; 219 uint32_t vmid_num_kfd; 220 }; 221 222 struct kfd_dev { 223 struct kgd_dev *kgd; 224 225 const struct kfd_device_info *device_info; 226 struct pci_dev *pdev; 227 struct drm_device *ddev; 228 229 unsigned int id; /* topology stub index */ 230 231 phys_addr_t doorbell_base; /* Start of actual doorbells used by 232 * KFD. It is aligned for mapping 233 * into user mode 234 */ 235 size_t doorbell_base_dw_offset; /* Offset from the start of the PCI 236 * doorbell BAR to the first KFD 237 * doorbell in dwords. GFX reserves 238 * the segment before this offset. 239 */ 240 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 241 * page used by kernel queue 242 */ 243 244 struct kgd2kfd_shared_resources shared_resources; 245 struct kfd_vmid_info vm_info; 246 247 const struct kfd2kgd_calls *kfd2kgd; 248 struct mutex doorbell_mutex; 249 DECLARE_BITMAP(doorbell_available_index, 250 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 251 252 void *gtt_mem; 253 uint64_t gtt_start_gpu_addr; 254 void *gtt_start_cpu_ptr; 255 void *gtt_sa_bitmap; 256 struct mutex gtt_sa_lock; 257 unsigned int gtt_sa_chunk_size; 258 unsigned int gtt_sa_num_of_chunks; 259 260 /* Interrupts */ 261 struct kfifo ih_fifo; 262 struct workqueue_struct *ih_wq; 263 struct work_struct interrupt_work; 264 spinlock_t interrupt_lock; 265 266 /* QCM Device instance */ 267 struct device_queue_manager *dqm; 268 269 bool init_complete; 270 /* 271 * Interrupts of interest to KFD are copied 272 * from the HW ring into a SW ring. 273 */ 274 bool interrupts_active; 275 276 /* Debug manager */ 277 struct kfd_dbgmgr *dbgmgr; 278 279 /* Firmware versions */ 280 uint16_t mec_fw_version; 281 uint16_t mec2_fw_version; 282 uint16_t sdma_fw_version; 283 284 /* Maximum process number mapped to HW scheduler */ 285 unsigned int max_proc_per_quantum; 286 287 /* CWSR */ 288 bool cwsr_enabled; 289 const void *cwsr_isa; 290 unsigned int cwsr_isa_size; 291 292 /* xGMI */ 293 uint64_t hive_id; 294 295 /* UUID */ 296 uint64_t unique_id; 297 298 bool pci_atomic_requested; 299 300 /* SRAM ECC flag */ 301 atomic_t sram_ecc_flag; 302 303 /* Compute Profile ref. count */ 304 atomic_t compute_profile; 305 306 /* Global GWS resource shared between processes */ 307 void *gws; 308 309 /* Clients watching SMI events */ 310 struct list_head smi_clients; 311 spinlock_t smi_lock; 312 }; 313 314 enum kfd_mempool { 315 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 316 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 317 KFD_MEMPOOL_FRAMEBUFFER = 3, 318 }; 319 320 /* Character device interface */ 321 int kfd_chardev_init(void); 322 void kfd_chardev_exit(void); 323 struct device *kfd_chardev(void); 324 325 /** 326 * enum kfd_unmap_queues_filter - Enum for queue filters. 327 * 328 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 329 * 330 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 331 * running queues list. 332 * 333 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 334 * specific process. 335 * 336 */ 337 enum kfd_unmap_queues_filter { 338 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 339 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 340 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 341 KFD_UNMAP_QUEUES_FILTER_BY_PASID 342 }; 343 344 /** 345 * enum kfd_queue_type - Enum for various queue types. 346 * 347 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 348 * 349 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 350 * 351 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 352 * 353 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 354 * 355 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 356 */ 357 enum kfd_queue_type { 358 KFD_QUEUE_TYPE_COMPUTE, 359 KFD_QUEUE_TYPE_SDMA, 360 KFD_QUEUE_TYPE_HIQ, 361 KFD_QUEUE_TYPE_DIQ, 362 KFD_QUEUE_TYPE_SDMA_XGMI 363 }; 364 365 enum kfd_queue_format { 366 KFD_QUEUE_FORMAT_PM4, 367 KFD_QUEUE_FORMAT_AQL 368 }; 369 370 enum KFD_QUEUE_PRIORITY { 371 KFD_QUEUE_PRIORITY_MINIMUM = 0, 372 KFD_QUEUE_PRIORITY_MAXIMUM = 15 373 }; 374 375 /** 376 * struct queue_properties 377 * 378 * @type: The queue type. 379 * 380 * @queue_id: Queue identifier. 381 * 382 * @queue_address: Queue ring buffer address. 383 * 384 * @queue_size: Queue ring buffer size. 385 * 386 * @priority: Defines the queue priority relative to other queues in the 387 * process. 388 * This is just an indication and HW scheduling may override the priority as 389 * necessary while keeping the relative prioritization. 390 * the priority granularity is from 0 to f which f is the highest priority. 391 * currently all queues are initialized with the highest priority. 392 * 393 * @queue_percent: This field is partially implemented and currently a zero in 394 * this field defines that the queue is non active. 395 * 396 * @read_ptr: User space address which points to the number of dwords the 397 * cp read from the ring buffer. This field updates automatically by the H/W. 398 * 399 * @write_ptr: Defines the number of dwords written to the ring buffer. 400 * 401 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 402 * buffer. This field should be similar to write_ptr and the user should 403 * update this field after updating the write_ptr. 404 * 405 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 406 * 407 * @is_interop: Defines if this is a interop queue. Interop queue means that 408 * the queue can access both graphics and compute resources. 409 * 410 * @is_evicted: Defines if the queue is evicted. Only active queues 411 * are evicted, rendering them inactive. 412 * 413 * @is_active: Defines if the queue is active or not. @is_active and 414 * @is_evicted are protected by the DQM lock. 415 * 416 * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 417 * @is_gws should be protected by the DQM lock, since changing it can yield the 418 * possibility of updating DQM state on number of GWS queues. 419 * 420 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 421 * of the queue. 422 * 423 * This structure represents the queue properties for each queue no matter if 424 * it's user mode or kernel mode queue. 425 * 426 */ 427 struct queue_properties { 428 enum kfd_queue_type type; 429 enum kfd_queue_format format; 430 unsigned int queue_id; 431 uint64_t queue_address; 432 uint64_t queue_size; 433 uint32_t priority; 434 uint32_t queue_percent; 435 uint32_t *read_ptr; 436 uint32_t *write_ptr; 437 void __iomem *doorbell_ptr; 438 uint32_t doorbell_off; 439 bool is_interop; 440 bool is_evicted; 441 bool is_active; 442 bool is_gws; 443 /* Not relevant for user mode queues in cp scheduling */ 444 unsigned int vmid; 445 /* Relevant only for sdma queues*/ 446 uint32_t sdma_engine_id; 447 uint32_t sdma_queue_id; 448 uint32_t sdma_vm_addr; 449 /* Relevant only for VI */ 450 uint64_t eop_ring_buffer_address; 451 uint32_t eop_ring_buffer_size; 452 uint64_t ctx_save_restore_area_address; 453 uint32_t ctx_save_restore_area_size; 454 uint32_t ctl_stack_size; 455 uint64_t tba_addr; 456 uint64_t tma_addr; 457 /* Relevant for CU */ 458 uint32_t cu_mask_count; /* Must be a multiple of 32 */ 459 uint32_t *cu_mask; 460 }; 461 462 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 463 (q).queue_address != 0 && \ 464 (q).queue_percent > 0 && \ 465 !(q).is_evicted) 466 467 /** 468 * struct queue 469 * 470 * @list: Queue linked list. 471 * 472 * @mqd: The queue MQD (memory queue descriptor). 473 * 474 * @mqd_mem_obj: The MQD local gpu memory object. 475 * 476 * @gart_mqd_addr: The MQD gart mc address. 477 * 478 * @properties: The queue properties. 479 * 480 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 481 * that the queue should be executed on. 482 * 483 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 484 * id. 485 * 486 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 487 * 488 * @process: The kfd process that created this queue. 489 * 490 * @device: The kfd device that created this queue. 491 * 492 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 493 * otherwise. 494 * 495 * This structure represents user mode compute queues. 496 * It contains all the necessary data to handle such queues. 497 * 498 */ 499 500 struct queue { 501 struct list_head list; 502 void *mqd; 503 struct kfd_mem_obj *mqd_mem_obj; 504 uint64_t gart_mqd_addr; 505 struct queue_properties properties; 506 507 uint32_t mec; 508 uint32_t pipe; 509 uint32_t queue; 510 511 unsigned int sdma_id; 512 unsigned int doorbell_id; 513 514 struct kfd_process *process; 515 struct kfd_dev *device; 516 void *gws; 517 518 /* procfs */ 519 struct kobject kobj; 520 }; 521 522 enum KFD_MQD_TYPE { 523 KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 524 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 525 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 526 KFD_MQD_TYPE_DIQ, /* for diq */ 527 KFD_MQD_TYPE_MAX 528 }; 529 530 enum KFD_PIPE_PRIORITY { 531 KFD_PIPE_PRIORITY_CS_LOW = 0, 532 KFD_PIPE_PRIORITY_CS_MEDIUM, 533 KFD_PIPE_PRIORITY_CS_HIGH 534 }; 535 536 struct scheduling_resources { 537 unsigned int vmid_mask; 538 enum kfd_queue_type type; 539 uint64_t queue_mask; 540 uint64_t gws_mask; 541 uint32_t oac_mask; 542 uint32_t gds_heap_base; 543 uint32_t gds_heap_size; 544 }; 545 546 struct process_queue_manager { 547 /* data */ 548 struct kfd_process *process; 549 struct list_head queues; 550 unsigned long *queue_slot_bitmap; 551 }; 552 553 struct qcm_process_device { 554 /* The Device Queue Manager that owns this data */ 555 struct device_queue_manager *dqm; 556 struct process_queue_manager *pqm; 557 /* Queues list */ 558 struct list_head queues_list; 559 struct list_head priv_queue_list; 560 561 unsigned int queue_count; 562 unsigned int vmid; 563 bool is_debug; 564 unsigned int evicted; /* eviction counter, 0=active */ 565 566 /* This flag tells if we should reset all wavefronts on 567 * process termination 568 */ 569 bool reset_wavefronts; 570 571 /* This flag tells us if this process has a GWS-capable 572 * queue that will be mapped into the runlist. It's 573 * possible to request a GWS BO, but not have the queue 574 * currently mapped, and this changes how the MAP_PROCESS 575 * PM4 packet is configured. 576 */ 577 bool mapped_gws_queue; 578 579 /* All the memory management data should be here too */ 580 uint64_t gds_context_area; 581 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 582 uint64_t page_table_base; 583 uint32_t sh_mem_config; 584 uint32_t sh_mem_bases; 585 uint32_t sh_mem_ape1_base; 586 uint32_t sh_mem_ape1_limit; 587 uint32_t gds_size; 588 uint32_t num_gws; 589 uint32_t num_oac; 590 uint32_t sh_hidden_private_base; 591 592 /* CWSR memory */ 593 void *cwsr_kaddr; 594 uint64_t cwsr_base; 595 uint64_t tba_addr; 596 uint64_t tma_addr; 597 598 /* IB memory */ 599 uint64_t ib_base; 600 void *ib_kaddr; 601 602 /* doorbell resources per process per device */ 603 unsigned long *doorbell_bitmap; 604 }; 605 606 /* KFD Memory Eviction */ 607 608 /* Approx. wait time before attempting to restore evicted BOs */ 609 #define PROCESS_RESTORE_TIME_MS 100 610 /* Approx. back off time if restore fails due to lack of memory */ 611 #define PROCESS_BACK_OFF_TIME_MS 100 612 /* Approx. time before evicting the process again */ 613 #define PROCESS_ACTIVE_TIME_MS 10 614 615 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 616 * idr_handle in the least significant 4 bytes 617 */ 618 #define MAKE_HANDLE(gpu_id, idr_handle) \ 619 (((uint64_t)(gpu_id) << 32) + idr_handle) 620 #define GET_GPU_ID(handle) (handle >> 32) 621 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 622 623 enum kfd_pdd_bound { 624 PDD_UNBOUND = 0, 625 PDD_BOUND, 626 PDD_BOUND_SUSPENDED, 627 }; 628 629 #define MAX_SYSFS_FILENAME_LEN 11 630 631 /* 632 * SDMA counter runs at 100MHz frequency. 633 * We display SDMA activity in microsecond granularity in sysfs. 634 * As a result, the divisor is 100. 635 */ 636 #define SDMA_ACTIVITY_DIVISOR 100 637 638 /* Data that is per-process-per device. */ 639 struct kfd_process_device { 640 /* 641 * List of all per-device data for a process. 642 * Starts from kfd_process.per_device_data. 643 */ 644 struct list_head per_device_list; 645 646 /* The device that owns this data. */ 647 struct kfd_dev *dev; 648 649 /* The process that owns this kfd_process_device. */ 650 struct kfd_process *process; 651 652 /* per-process-per device QCM data structure */ 653 struct qcm_process_device qpd; 654 655 /*Apertures*/ 656 uint64_t lds_base; 657 uint64_t lds_limit; 658 uint64_t gpuvm_base; 659 uint64_t gpuvm_limit; 660 uint64_t scratch_base; 661 uint64_t scratch_limit; 662 663 /* VM context for GPUVM allocations */ 664 struct file *drm_file; 665 void *vm; 666 667 /* GPUVM allocations storage */ 668 struct idr alloc_idr; 669 670 /* Flag used to tell the pdd has dequeued from the dqm. 671 * This is used to prevent dev->dqm->ops.process_termination() from 672 * being called twice when it is already called in IOMMU callback 673 * function. 674 */ 675 bool already_dequeued; 676 bool runtime_inuse; 677 678 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 679 enum kfd_pdd_bound bound; 680 681 /* VRAM usage */ 682 uint64_t vram_usage; 683 struct attribute attr_vram; 684 char vram_filename[MAX_SYSFS_FILENAME_LEN]; 685 686 /* SDMA activity tracking */ 687 uint64_t sdma_past_activity_counter; 688 struct attribute attr_sdma; 689 char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 690 }; 691 692 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 693 694 /* Process data */ 695 struct kfd_process { 696 /* 697 * kfd_process are stored in an mm_struct*->kfd_process* 698 * hash table (kfd_processes in kfd_process.c) 699 */ 700 struct hlist_node kfd_processes; 701 702 /* 703 * Opaque pointer to mm_struct. We don't hold a reference to 704 * it so it should never be dereferenced from here. This is 705 * only used for looking up processes by their mm. 706 */ 707 void *mm; 708 709 struct kref ref; 710 struct work_struct release_work; 711 712 struct mutex mutex; 713 714 /* 715 * In any process, the thread that started main() is the lead 716 * thread and outlives the rest. 717 * It is here because amd_iommu_bind_pasid wants a task_struct. 718 * It can also be used for safely getting a reference to the 719 * mm_struct of the process. 720 */ 721 struct task_struct *lead_thread; 722 723 /* We want to receive a notification when the mm_struct is destroyed */ 724 struct mmu_notifier mmu_notifier; 725 726 uint16_t pasid; 727 unsigned int doorbell_index; 728 729 /* 730 * List of kfd_process_device structures, 731 * one for each device the process is using. 732 */ 733 struct list_head per_device_data; 734 735 struct process_queue_manager pqm; 736 737 /*Is the user space process 32 bit?*/ 738 bool is_32bit_user_mode; 739 740 /* Event-related data */ 741 struct mutex event_mutex; 742 /* Event ID allocator and lookup */ 743 struct idr event_idr; 744 /* Event page */ 745 struct kfd_signal_page *signal_page; 746 size_t signal_mapped_size; 747 size_t signal_event_count; 748 bool signal_event_limit_reached; 749 750 /* Information used for memory eviction */ 751 void *kgd_process_info; 752 /* Eviction fence that is attached to all the BOs of this process. The 753 * fence will be triggered during eviction and new one will be created 754 * during restore 755 */ 756 struct dma_fence *ef; 757 758 /* Work items for evicting and restoring BOs */ 759 struct delayed_work eviction_work; 760 struct delayed_work restore_work; 761 /* seqno of the last scheduled eviction */ 762 unsigned int last_eviction_seqno; 763 /* Approx. the last timestamp (in jiffies) when the process was 764 * restored after an eviction 765 */ 766 unsigned long last_restore_timestamp; 767 768 /* Kobj for our procfs */ 769 struct kobject *kobj; 770 struct kobject *kobj_queues; 771 struct attribute attr_pasid; 772 }; 773 774 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 775 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 776 extern struct srcu_struct kfd_processes_srcu; 777 778 /** 779 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 780 * 781 * @filep: pointer to file structure. 782 * @p: amdkfd process pointer. 783 * @data: pointer to arg that was copied from user. 784 * 785 * Return: returns ioctl completion code. 786 */ 787 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 788 void *data); 789 790 struct amdkfd_ioctl_desc { 791 unsigned int cmd; 792 int flags; 793 amdkfd_ioctl_t *func; 794 unsigned int cmd_drv; 795 const char *name; 796 }; 797 bool kfd_dev_is_large_bar(struct kfd_dev *dev); 798 799 int kfd_process_create_wq(void); 800 void kfd_process_destroy_wq(void); 801 struct kfd_process *kfd_create_process(struct file *filep); 802 struct kfd_process *kfd_get_process(const struct task_struct *); 803 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 804 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 805 void kfd_unref_process(struct kfd_process *p); 806 int kfd_process_evict_queues(struct kfd_process *p); 807 int kfd_process_restore_queues(struct kfd_process *p); 808 void kfd_suspend_all_processes(void); 809 int kfd_resume_all_processes(void); 810 811 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 812 struct file *drm_file); 813 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 814 struct kfd_process *p); 815 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 816 struct kfd_process *p); 817 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 818 struct kfd_process *p); 819 820 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 821 struct vm_area_struct *vma); 822 823 /* KFD process API for creating and translating handles */ 824 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 825 void *mem); 826 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 827 int handle); 828 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 829 int handle); 830 831 /* Process device data iterator */ 832 struct kfd_process_device *kfd_get_first_process_device_data( 833 struct kfd_process *p); 834 struct kfd_process_device *kfd_get_next_process_device_data( 835 struct kfd_process *p, 836 struct kfd_process_device *pdd); 837 bool kfd_has_process_device_data(struct kfd_process *p); 838 839 /* PASIDs */ 840 int kfd_pasid_init(void); 841 void kfd_pasid_exit(void); 842 bool kfd_set_pasid_limit(unsigned int new_limit); 843 unsigned int kfd_get_pasid_limit(void); 844 unsigned int kfd_pasid_alloc(void); 845 void kfd_pasid_free(unsigned int pasid); 846 847 /* Doorbells */ 848 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 849 int kfd_doorbell_init(struct kfd_dev *kfd); 850 void kfd_doorbell_fini(struct kfd_dev *kfd); 851 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 852 struct vm_area_struct *vma); 853 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 854 unsigned int *doorbell_off); 855 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 856 u32 read_kernel_doorbell(u32 __iomem *db); 857 void write_kernel_doorbell(void __iomem *db, u32 value); 858 void write_kernel_doorbell64(void __iomem *db, u64 value); 859 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 860 struct kfd_process *process, 861 unsigned int doorbell_id); 862 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 863 struct kfd_process *process); 864 int kfd_alloc_process_doorbells(struct kfd_process *process); 865 void kfd_free_process_doorbells(struct kfd_process *process); 866 867 /* GTT Sub-Allocator */ 868 869 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 870 struct kfd_mem_obj **mem_obj); 871 872 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 873 874 extern struct device *kfd_device; 875 876 /* KFD's procfs */ 877 void kfd_procfs_init(void); 878 void kfd_procfs_shutdown(void); 879 int kfd_procfs_add_queue(struct queue *q); 880 void kfd_procfs_del_queue(struct queue *q); 881 882 /* Topology */ 883 int kfd_topology_init(void); 884 void kfd_topology_shutdown(void); 885 int kfd_topology_add_device(struct kfd_dev *gpu); 886 int kfd_topology_remove_device(struct kfd_dev *gpu); 887 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 888 uint32_t proximity_domain); 889 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 890 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 891 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 892 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd); 893 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 894 int kfd_numa_node_to_apic_id(int numa_node_id); 895 896 /* Interrupts */ 897 int kfd_interrupt_init(struct kfd_dev *dev); 898 void kfd_interrupt_exit(struct kfd_dev *dev); 899 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 900 bool interrupt_is_wanted(struct kfd_dev *dev, 901 const uint32_t *ih_ring_entry, 902 uint32_t *patched_ihre, bool *flag); 903 904 /* amdkfd Apertures */ 905 int kfd_init_apertures(struct kfd_process *process); 906 907 /* Queue Context Management */ 908 int init_queue(struct queue **q, const struct queue_properties *properties); 909 void uninit_queue(struct queue *q); 910 void print_queue_properties(struct queue_properties *q); 911 void print_queue(struct queue *q); 912 913 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 914 struct kfd_dev *dev); 915 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 916 struct kfd_dev *dev); 917 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 918 struct kfd_dev *dev); 919 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 920 struct kfd_dev *dev); 921 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 922 struct kfd_dev *dev); 923 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 924 struct kfd_dev *dev); 925 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 926 void device_queue_manager_uninit(struct device_queue_manager *dqm); 927 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 928 enum kfd_queue_type type); 929 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 930 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); 931 932 /* Process Queue Manager */ 933 struct process_queue_node { 934 struct queue *q; 935 struct kernel_queue *kq; 936 struct list_head process_queue_list; 937 }; 938 939 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 940 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 941 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 942 void pqm_uninit(struct process_queue_manager *pqm); 943 int pqm_create_queue(struct process_queue_manager *pqm, 944 struct kfd_dev *dev, 945 struct file *f, 946 struct queue_properties *properties, 947 unsigned int *qid, 948 uint32_t *p_doorbell_offset_in_process); 949 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 950 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 951 struct queue_properties *p); 952 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 953 struct queue_properties *p); 954 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 955 void *gws); 956 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 957 unsigned int qid); 958 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 959 unsigned int qid); 960 int pqm_get_wave_state(struct process_queue_manager *pqm, 961 unsigned int qid, 962 void __user *ctl_stack, 963 u32 *ctl_stack_used_size, 964 u32 *save_area_used_size); 965 966 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 967 unsigned int fence_value, 968 unsigned int timeout_ms); 969 970 /* Packet Manager */ 971 972 #define KFD_FENCE_COMPLETED (100) 973 #define KFD_FENCE_INIT (10) 974 975 struct packet_manager { 976 struct device_queue_manager *dqm; 977 struct kernel_queue *priv_queue; 978 struct mutex lock; 979 bool allocated; 980 struct kfd_mem_obj *ib_buffer_obj; 981 unsigned int ib_size_bytes; 982 bool is_over_subscription; 983 984 const struct packet_manager_funcs *pmf; 985 }; 986 987 struct packet_manager_funcs { 988 /* Support ASIC-specific packet formats for PM4 packets */ 989 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 990 struct qcm_process_device *qpd); 991 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 992 uint64_t ib, size_t ib_size_in_dwords, bool chain); 993 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 994 struct scheduling_resources *res); 995 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 996 struct queue *q, bool is_static); 997 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 998 enum kfd_queue_type type, 999 enum kfd_unmap_queues_filter mode, 1000 uint32_t filter_param, bool reset, 1001 unsigned int sdma_engine); 1002 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1003 uint64_t fence_address, uint32_t fence_value); 1004 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1005 1006 /* Packet sizes */ 1007 int map_process_size; 1008 int runlist_size; 1009 int set_resources_size; 1010 int map_queues_size; 1011 int unmap_queues_size; 1012 int query_status_size; 1013 int release_mem_size; 1014 }; 1015 1016 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1017 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1018 1019 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1020 void pm_uninit(struct packet_manager *pm, bool hanging); 1021 int pm_send_set_resources(struct packet_manager *pm, 1022 struct scheduling_resources *res); 1023 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 1024 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 1025 uint32_t fence_value); 1026 1027 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 1028 enum kfd_unmap_queues_filter mode, 1029 uint32_t filter_param, bool reset, 1030 unsigned int sdma_engine); 1031 1032 void pm_release_ib(struct packet_manager *pm); 1033 1034 /* Following PM funcs can be shared among VI and AI */ 1035 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 1036 1037 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 1038 1039 /* Events */ 1040 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1041 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1042 1043 extern const struct kfd_device_global_init_class device_global_init_class_cik; 1044 1045 void kfd_event_init_process(struct kfd_process *p); 1046 void kfd_event_free_process(struct kfd_process *p); 1047 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1048 int kfd_wait_on_events(struct kfd_process *p, 1049 uint32_t num_events, void __user *data, 1050 bool all, uint32_t user_timeout_ms, 1051 uint32_t *wait_result); 1052 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 1053 uint32_t valid_id_bits); 1054 void kfd_signal_iommu_event(struct kfd_dev *dev, 1055 unsigned int pasid, unsigned long address, 1056 bool is_write_requested, bool is_execute_requested); 1057 void kfd_signal_hw_exception_event(unsigned int pasid); 1058 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1059 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 1060 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 1061 uint64_t size); 1062 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1063 uint32_t event_type, bool auto_reset, uint32_t node_id, 1064 uint32_t *event_id, uint32_t *event_trigger_data, 1065 uint64_t *event_page_offset, uint32_t *event_slot_index); 1066 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1067 1068 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid, 1069 struct kfd_vm_fault_info *info); 1070 1071 void kfd_signal_reset_event(struct kfd_dev *dev); 1072 1073 void kfd_flush_tlb(struct kfd_process_device *pdd); 1074 1075 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1076 1077 bool kfd_is_locked(void); 1078 1079 /* Compute profile */ 1080 void kfd_inc_compute_active(struct kfd_dev *dev); 1081 void kfd_dec_compute_active(struct kfd_dev *dev); 1082 1083 /* Cgroup Support */ 1084 /* Check with device cgroup if @kfd device is accessible */ 1085 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 1086 { 1087 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 1088 struct drm_device *ddev = kfd->ddev; 1089 1090 return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 1091 ddev->render->index, 1092 DEVCG_ACC_WRITE | DEVCG_ACC_READ); 1093 #else 1094 return 0; 1095 #endif 1096 } 1097 1098 /* Debugfs */ 1099 #if defined(CONFIG_DEBUG_FS) 1100 1101 void kfd_debugfs_init(void); 1102 void kfd_debugfs_fini(void); 1103 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1104 int pqm_debugfs_mqds(struct seq_file *m, void *data); 1105 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1106 int dqm_debugfs_hqds(struct seq_file *m, void *data); 1107 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1108 int pm_debugfs_runlist(struct seq_file *m, void *data); 1109 1110 int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1111 int pm_debugfs_hang_hws(struct packet_manager *pm); 1112 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1113 1114 #else 1115 1116 static inline void kfd_debugfs_init(void) {} 1117 static inline void kfd_debugfs_fini(void) {} 1118 1119 #endif 1120 1121 #endif 1122