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 bool pci_atomic_requested; 296 297 /* Use IOMMU v2 flag */ 298 bool use_iommu_v2; 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 uint32_t reset_seq_num; 314 315 struct ida doorbell_ida; 316 unsigned int max_doorbell_slices; 317 318 int noretry; 319 }; 320 321 enum kfd_mempool { 322 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 323 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 324 KFD_MEMPOOL_FRAMEBUFFER = 3, 325 }; 326 327 /* Character device interface */ 328 int kfd_chardev_init(void); 329 void kfd_chardev_exit(void); 330 struct device *kfd_chardev(void); 331 332 /** 333 * enum kfd_unmap_queues_filter - Enum for queue filters. 334 * 335 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 336 * 337 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 338 * running queues list. 339 * 340 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 341 * specific process. 342 * 343 */ 344 enum kfd_unmap_queues_filter { 345 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 346 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 347 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 348 KFD_UNMAP_QUEUES_FILTER_BY_PASID 349 }; 350 351 /** 352 * enum kfd_queue_type - Enum for various queue types. 353 * 354 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 355 * 356 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 357 * 358 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 359 * 360 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 361 * 362 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 363 */ 364 enum kfd_queue_type { 365 KFD_QUEUE_TYPE_COMPUTE, 366 KFD_QUEUE_TYPE_SDMA, 367 KFD_QUEUE_TYPE_HIQ, 368 KFD_QUEUE_TYPE_DIQ, 369 KFD_QUEUE_TYPE_SDMA_XGMI 370 }; 371 372 enum kfd_queue_format { 373 KFD_QUEUE_FORMAT_PM4, 374 KFD_QUEUE_FORMAT_AQL 375 }; 376 377 enum KFD_QUEUE_PRIORITY { 378 KFD_QUEUE_PRIORITY_MINIMUM = 0, 379 KFD_QUEUE_PRIORITY_MAXIMUM = 15 380 }; 381 382 /** 383 * struct queue_properties 384 * 385 * @type: The queue type. 386 * 387 * @queue_id: Queue identifier. 388 * 389 * @queue_address: Queue ring buffer address. 390 * 391 * @queue_size: Queue ring buffer size. 392 * 393 * @priority: Defines the queue priority relative to other queues in the 394 * process. 395 * This is just an indication and HW scheduling may override the priority as 396 * necessary while keeping the relative prioritization. 397 * the priority granularity is from 0 to f which f is the highest priority. 398 * currently all queues are initialized with the highest priority. 399 * 400 * @queue_percent: This field is partially implemented and currently a zero in 401 * this field defines that the queue is non active. 402 * 403 * @read_ptr: User space address which points to the number of dwords the 404 * cp read from the ring buffer. This field updates automatically by the H/W. 405 * 406 * @write_ptr: Defines the number of dwords written to the ring buffer. 407 * 408 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 409 * buffer. This field should be similar to write_ptr and the user should 410 * update this field after updating the write_ptr. 411 * 412 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 413 * 414 * @is_interop: Defines if this is a interop queue. Interop queue means that 415 * the queue can access both graphics and compute resources. 416 * 417 * @is_evicted: Defines if the queue is evicted. Only active queues 418 * are evicted, rendering them inactive. 419 * 420 * @is_active: Defines if the queue is active or not. @is_active and 421 * @is_evicted are protected by the DQM lock. 422 * 423 * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 424 * @is_gws should be protected by the DQM lock, since changing it can yield the 425 * possibility of updating DQM state on number of GWS queues. 426 * 427 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 428 * of the queue. 429 * 430 * This structure represents the queue properties for each queue no matter if 431 * it's user mode or kernel mode queue. 432 * 433 */ 434 struct queue_properties { 435 enum kfd_queue_type type; 436 enum kfd_queue_format format; 437 unsigned int queue_id; 438 uint64_t queue_address; 439 uint64_t queue_size; 440 uint32_t priority; 441 uint32_t queue_percent; 442 uint32_t *read_ptr; 443 uint32_t *write_ptr; 444 void __iomem *doorbell_ptr; 445 uint32_t doorbell_off; 446 bool is_interop; 447 bool is_evicted; 448 bool is_active; 449 bool is_gws; 450 /* Not relevant for user mode queues in cp scheduling */ 451 unsigned int vmid; 452 /* Relevant only for sdma queues*/ 453 uint32_t sdma_engine_id; 454 uint32_t sdma_queue_id; 455 uint32_t sdma_vm_addr; 456 /* Relevant only for VI */ 457 uint64_t eop_ring_buffer_address; 458 uint32_t eop_ring_buffer_size; 459 uint64_t ctx_save_restore_area_address; 460 uint32_t ctx_save_restore_area_size; 461 uint32_t ctl_stack_size; 462 uint64_t tba_addr; 463 uint64_t tma_addr; 464 /* Relevant for CU */ 465 uint32_t cu_mask_count; /* Must be a multiple of 32 */ 466 uint32_t *cu_mask; 467 }; 468 469 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 470 (q).queue_address != 0 && \ 471 (q).queue_percent > 0 && \ 472 !(q).is_evicted) 473 474 /** 475 * struct queue 476 * 477 * @list: Queue linked list. 478 * 479 * @mqd: The queue MQD (memory queue descriptor). 480 * 481 * @mqd_mem_obj: The MQD local gpu memory object. 482 * 483 * @gart_mqd_addr: The MQD gart mc address. 484 * 485 * @properties: The queue properties. 486 * 487 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 488 * that the queue should be executed on. 489 * 490 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 491 * id. 492 * 493 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 494 * 495 * @process: The kfd process that created this queue. 496 * 497 * @device: The kfd device that created this queue. 498 * 499 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 500 * otherwise. 501 * 502 * This structure represents user mode compute queues. 503 * It contains all the necessary data to handle such queues. 504 * 505 */ 506 507 struct queue { 508 struct list_head list; 509 void *mqd; 510 struct kfd_mem_obj *mqd_mem_obj; 511 uint64_t gart_mqd_addr; 512 struct queue_properties properties; 513 514 uint32_t mec; 515 uint32_t pipe; 516 uint32_t queue; 517 518 unsigned int sdma_id; 519 unsigned int doorbell_id; 520 521 struct kfd_process *process; 522 struct kfd_dev *device; 523 void *gws; 524 525 /* procfs */ 526 struct kobject kobj; 527 }; 528 529 enum KFD_MQD_TYPE { 530 KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 531 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 532 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 533 KFD_MQD_TYPE_DIQ, /* for diq */ 534 KFD_MQD_TYPE_MAX 535 }; 536 537 enum KFD_PIPE_PRIORITY { 538 KFD_PIPE_PRIORITY_CS_LOW = 0, 539 KFD_PIPE_PRIORITY_CS_MEDIUM, 540 KFD_PIPE_PRIORITY_CS_HIGH 541 }; 542 543 struct scheduling_resources { 544 unsigned int vmid_mask; 545 enum kfd_queue_type type; 546 uint64_t queue_mask; 547 uint64_t gws_mask; 548 uint32_t oac_mask; 549 uint32_t gds_heap_base; 550 uint32_t gds_heap_size; 551 }; 552 553 struct process_queue_manager { 554 /* data */ 555 struct kfd_process *process; 556 struct list_head queues; 557 unsigned long *queue_slot_bitmap; 558 }; 559 560 struct qcm_process_device { 561 /* The Device Queue Manager that owns this data */ 562 struct device_queue_manager *dqm; 563 struct process_queue_manager *pqm; 564 /* Queues list */ 565 struct list_head queues_list; 566 struct list_head priv_queue_list; 567 568 unsigned int queue_count; 569 unsigned int vmid; 570 bool is_debug; 571 unsigned int evicted; /* eviction counter, 0=active */ 572 573 /* This flag tells if we should reset all wavefronts on 574 * process termination 575 */ 576 bool reset_wavefronts; 577 578 /* This flag tells us if this process has a GWS-capable 579 * queue that will be mapped into the runlist. It's 580 * possible to request a GWS BO, but not have the queue 581 * currently mapped, and this changes how the MAP_PROCESS 582 * PM4 packet is configured. 583 */ 584 bool mapped_gws_queue; 585 586 /* All the memory management data should be here too */ 587 uint64_t gds_context_area; 588 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 589 uint64_t page_table_base; 590 uint32_t sh_mem_config; 591 uint32_t sh_mem_bases; 592 uint32_t sh_mem_ape1_base; 593 uint32_t sh_mem_ape1_limit; 594 uint32_t gds_size; 595 uint32_t num_gws; 596 uint32_t num_oac; 597 uint32_t sh_hidden_private_base; 598 599 /* CWSR memory */ 600 void *cwsr_kaddr; 601 uint64_t cwsr_base; 602 uint64_t tba_addr; 603 uint64_t tma_addr; 604 605 /* IB memory */ 606 uint64_t ib_base; 607 void *ib_kaddr; 608 609 /* doorbell resources per process per device */ 610 unsigned long *doorbell_bitmap; 611 }; 612 613 /* KFD Memory Eviction */ 614 615 /* Approx. wait time before attempting to restore evicted BOs */ 616 #define PROCESS_RESTORE_TIME_MS 100 617 /* Approx. back off time if restore fails due to lack of memory */ 618 #define PROCESS_BACK_OFF_TIME_MS 100 619 /* Approx. time before evicting the process again */ 620 #define PROCESS_ACTIVE_TIME_MS 10 621 622 /* 8 byte handle containing GPU ID in the most significant 4 bytes and 623 * idr_handle in the least significant 4 bytes 624 */ 625 #define MAKE_HANDLE(gpu_id, idr_handle) \ 626 (((uint64_t)(gpu_id) << 32) + idr_handle) 627 #define GET_GPU_ID(handle) (handle >> 32) 628 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 629 630 enum kfd_pdd_bound { 631 PDD_UNBOUND = 0, 632 PDD_BOUND, 633 PDD_BOUND_SUSPENDED, 634 }; 635 636 #define MAX_SYSFS_FILENAME_LEN 15 637 638 /* 639 * SDMA counter runs at 100MHz frequency. 640 * We display SDMA activity in microsecond granularity in sysfs. 641 * As a result, the divisor is 100. 642 */ 643 #define SDMA_ACTIVITY_DIVISOR 100 644 645 /* Data that is per-process-per device. */ 646 struct kfd_process_device { 647 /* 648 * List of all per-device data for a process. 649 * Starts from kfd_process.per_device_data. 650 */ 651 struct list_head per_device_list; 652 653 /* The device that owns this data. */ 654 struct kfd_dev *dev; 655 656 /* The process that owns this kfd_process_device. */ 657 struct kfd_process *process; 658 659 /* per-process-per device QCM data structure */ 660 struct qcm_process_device qpd; 661 662 /*Apertures*/ 663 uint64_t lds_base; 664 uint64_t lds_limit; 665 uint64_t gpuvm_base; 666 uint64_t gpuvm_limit; 667 uint64_t scratch_base; 668 uint64_t scratch_limit; 669 670 /* VM context for GPUVM allocations */ 671 struct file *drm_file; 672 void *vm; 673 674 /* GPUVM allocations storage */ 675 struct idr alloc_idr; 676 677 /* Flag used to tell the pdd has dequeued from the dqm. 678 * This is used to prevent dev->dqm->ops.process_termination() from 679 * being called twice when it is already called in IOMMU callback 680 * function. 681 */ 682 bool already_dequeued; 683 bool runtime_inuse; 684 685 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 686 enum kfd_pdd_bound bound; 687 688 /* VRAM usage */ 689 uint64_t vram_usage; 690 struct attribute attr_vram; 691 char vram_filename[MAX_SYSFS_FILENAME_LEN]; 692 693 /* SDMA activity tracking */ 694 uint64_t sdma_past_activity_counter; 695 struct attribute attr_sdma; 696 char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 697 698 /* Eviction activity tracking */ 699 uint64_t last_evict_timestamp; 700 atomic64_t evict_duration_counter; 701 struct attribute attr_evict; 702 703 struct kobject *kobj_stats; 704 unsigned int doorbell_index; 705 706 /* 707 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process 708 * that is associated with device encoded by "this" struct instance. The 709 * value reflects CU usage by all of the waves launched by this process 710 * on this device. A very important property of occupancy parameter is 711 * that its value is a snapshot of current use. 712 * 713 * Following is to be noted regarding how this parameter is reported: 714 * 715 * The number of waves that a CU can launch is limited by couple of 716 * parameters. These are encoded by struct amdgpu_cu_info instance 717 * that is part of every device definition. For GFX9 devices this 718 * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves 719 * do not use scratch memory and 32 waves (max_scratch_slots_per_cu) 720 * when they do use scratch memory. This could change for future 721 * devices and therefore this example should be considered as a guide. 722 * 723 * All CU's of a device are available for the process. This may not be true 724 * under certain conditions - e.g. CU masking. 725 * 726 * Finally number of CU's that are occupied by a process is affected by both 727 * number of CU's a device has along with number of other competing processes 728 */ 729 struct attribute attr_cu_occupancy; 730 }; 731 732 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 733 734 /* Process data */ 735 struct kfd_process { 736 /* 737 * kfd_process are stored in an mm_struct*->kfd_process* 738 * hash table (kfd_processes in kfd_process.c) 739 */ 740 struct hlist_node kfd_processes; 741 742 /* 743 * Opaque pointer to mm_struct. We don't hold a reference to 744 * it so it should never be dereferenced from here. This is 745 * only used for looking up processes by their mm. 746 */ 747 void *mm; 748 749 struct kref ref; 750 struct work_struct release_work; 751 752 struct mutex mutex; 753 754 /* 755 * In any process, the thread that started main() is the lead 756 * thread and outlives the rest. 757 * It is here because amd_iommu_bind_pasid wants a task_struct. 758 * It can also be used for safely getting a reference to the 759 * mm_struct of the process. 760 */ 761 struct task_struct *lead_thread; 762 763 /* We want to receive a notification when the mm_struct is destroyed */ 764 struct mmu_notifier mmu_notifier; 765 766 u32 pasid; 767 768 /* 769 * List of kfd_process_device structures, 770 * one for each device the process is using. 771 */ 772 struct list_head per_device_data; 773 774 struct process_queue_manager pqm; 775 776 /*Is the user space process 32 bit?*/ 777 bool is_32bit_user_mode; 778 779 /* Event-related data */ 780 struct mutex event_mutex; 781 /* Event ID allocator and lookup */ 782 struct idr event_idr; 783 /* Event page */ 784 struct kfd_signal_page *signal_page; 785 size_t signal_mapped_size; 786 size_t signal_event_count; 787 bool signal_event_limit_reached; 788 789 /* Information used for memory eviction */ 790 void *kgd_process_info; 791 /* Eviction fence that is attached to all the BOs of this process. The 792 * fence will be triggered during eviction and new one will be created 793 * during restore 794 */ 795 struct dma_fence *ef; 796 797 /* Work items for evicting and restoring BOs */ 798 struct delayed_work eviction_work; 799 struct delayed_work restore_work; 800 /* seqno of the last scheduled eviction */ 801 unsigned int last_eviction_seqno; 802 /* Approx. the last timestamp (in jiffies) when the process was 803 * restored after an eviction 804 */ 805 unsigned long last_restore_timestamp; 806 807 /* Kobj for our procfs */ 808 struct kobject *kobj; 809 struct kobject *kobj_queues; 810 struct attribute attr_pasid; 811 }; 812 813 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 814 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 815 extern struct srcu_struct kfd_processes_srcu; 816 817 /** 818 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 819 * 820 * @filep: pointer to file structure. 821 * @p: amdkfd process pointer. 822 * @data: pointer to arg that was copied from user. 823 * 824 * Return: returns ioctl completion code. 825 */ 826 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 827 void *data); 828 829 struct amdkfd_ioctl_desc { 830 unsigned int cmd; 831 int flags; 832 amdkfd_ioctl_t *func; 833 unsigned int cmd_drv; 834 const char *name; 835 }; 836 bool kfd_dev_is_large_bar(struct kfd_dev *dev); 837 838 int kfd_process_create_wq(void); 839 void kfd_process_destroy_wq(void); 840 struct kfd_process *kfd_create_process(struct file *filep); 841 struct kfd_process *kfd_get_process(const struct task_struct *); 842 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); 843 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 844 void kfd_unref_process(struct kfd_process *p); 845 int kfd_process_evict_queues(struct kfd_process *p); 846 int kfd_process_restore_queues(struct kfd_process *p); 847 void kfd_suspend_all_processes(void); 848 int kfd_resume_all_processes(void); 849 850 int kfd_process_device_init_vm(struct kfd_process_device *pdd, 851 struct file *drm_file); 852 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 853 struct kfd_process *p); 854 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 855 struct kfd_process *p); 856 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 857 struct kfd_process *p); 858 859 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 860 struct vm_area_struct *vma); 861 862 /* KFD process API for creating and translating handles */ 863 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 864 void *mem); 865 void *kfd_process_device_translate_handle(struct kfd_process_device *p, 866 int handle); 867 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 868 int handle); 869 870 /* Process device data iterator */ 871 struct kfd_process_device *kfd_get_first_process_device_data( 872 struct kfd_process *p); 873 struct kfd_process_device *kfd_get_next_process_device_data( 874 struct kfd_process *p, 875 struct kfd_process_device *pdd); 876 bool kfd_has_process_device_data(struct kfd_process *p); 877 878 /* PASIDs */ 879 int kfd_pasid_init(void); 880 void kfd_pasid_exit(void); 881 bool kfd_set_pasid_limit(unsigned int new_limit); 882 unsigned int kfd_get_pasid_limit(void); 883 u32 kfd_pasid_alloc(void); 884 void kfd_pasid_free(u32 pasid); 885 886 /* Doorbells */ 887 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 888 int kfd_doorbell_init(struct kfd_dev *kfd); 889 void kfd_doorbell_fini(struct kfd_dev *kfd); 890 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 891 struct vm_area_struct *vma); 892 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 893 unsigned int *doorbell_off); 894 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 895 u32 read_kernel_doorbell(u32 __iomem *db); 896 void write_kernel_doorbell(void __iomem *db, u32 value); 897 void write_kernel_doorbell64(void __iomem *db, u64 value); 898 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 899 struct kfd_process_device *pdd, 900 unsigned int doorbell_id); 901 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); 902 int kfd_alloc_process_doorbells(struct kfd_dev *kfd, 903 unsigned int *doorbell_index); 904 void kfd_free_process_doorbells(struct kfd_dev *kfd, 905 unsigned int doorbell_index); 906 /* GTT Sub-Allocator */ 907 908 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 909 struct kfd_mem_obj **mem_obj); 910 911 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 912 913 extern struct device *kfd_device; 914 915 /* KFD's procfs */ 916 void kfd_procfs_init(void); 917 void kfd_procfs_shutdown(void); 918 int kfd_procfs_add_queue(struct queue *q); 919 void kfd_procfs_del_queue(struct queue *q); 920 921 /* Topology */ 922 int kfd_topology_init(void); 923 void kfd_topology_shutdown(void); 924 int kfd_topology_add_device(struct kfd_dev *gpu); 925 int kfd_topology_remove_device(struct kfd_dev *gpu); 926 struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 927 uint32_t proximity_domain); 928 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 929 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 930 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 931 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd); 932 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 933 int kfd_numa_node_to_apic_id(int numa_node_id); 934 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); 935 936 /* Interrupts */ 937 int kfd_interrupt_init(struct kfd_dev *dev); 938 void kfd_interrupt_exit(struct kfd_dev *dev); 939 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 940 bool interrupt_is_wanted(struct kfd_dev *dev, 941 const uint32_t *ih_ring_entry, 942 uint32_t *patched_ihre, bool *flag); 943 944 /* amdkfd Apertures */ 945 int kfd_init_apertures(struct kfd_process *process); 946 947 /* Queue Context Management */ 948 int init_queue(struct queue **q, const struct queue_properties *properties); 949 void uninit_queue(struct queue *q); 950 void print_queue_properties(struct queue_properties *q); 951 void print_queue(struct queue *q); 952 953 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 954 struct kfd_dev *dev); 955 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 956 struct kfd_dev *dev); 957 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 958 struct kfd_dev *dev); 959 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 960 struct kfd_dev *dev); 961 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 962 struct kfd_dev *dev); 963 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 964 struct kfd_dev *dev); 965 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 966 void device_queue_manager_uninit(struct device_queue_manager *dqm); 967 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 968 enum kfd_queue_type type); 969 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 970 int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid); 971 972 /* Process Queue Manager */ 973 struct process_queue_node { 974 struct queue *q; 975 struct kernel_queue *kq; 976 struct list_head process_queue_list; 977 }; 978 979 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 980 void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 981 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 982 void pqm_uninit(struct process_queue_manager *pqm); 983 int pqm_create_queue(struct process_queue_manager *pqm, 984 struct kfd_dev *dev, 985 struct file *f, 986 struct queue_properties *properties, 987 unsigned int *qid, 988 uint32_t *p_doorbell_offset_in_process); 989 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 990 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 991 struct queue_properties *p); 992 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 993 struct queue_properties *p); 994 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 995 void *gws); 996 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 997 unsigned int qid); 998 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 999 unsigned int qid); 1000 int pqm_get_wave_state(struct process_queue_manager *pqm, 1001 unsigned int qid, 1002 void __user *ctl_stack, 1003 u32 *ctl_stack_used_size, 1004 u32 *save_area_used_size); 1005 1006 int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 1007 unsigned int fence_value, 1008 unsigned int timeout_ms); 1009 1010 /* Packet Manager */ 1011 1012 #define KFD_FENCE_COMPLETED (100) 1013 #define KFD_FENCE_INIT (10) 1014 1015 struct packet_manager { 1016 struct device_queue_manager *dqm; 1017 struct kernel_queue *priv_queue; 1018 struct mutex lock; 1019 bool allocated; 1020 struct kfd_mem_obj *ib_buffer_obj; 1021 unsigned int ib_size_bytes; 1022 bool is_over_subscription; 1023 1024 const struct packet_manager_funcs *pmf; 1025 }; 1026 1027 struct packet_manager_funcs { 1028 /* Support ASIC-specific packet formats for PM4 packets */ 1029 int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 1030 struct qcm_process_device *qpd); 1031 int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 1032 uint64_t ib, size_t ib_size_in_dwords, bool chain); 1033 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 1034 struct scheduling_resources *res); 1035 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 1036 struct queue *q, bool is_static); 1037 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 1038 enum kfd_queue_type type, 1039 enum kfd_unmap_queues_filter mode, 1040 uint32_t filter_param, bool reset, 1041 unsigned int sdma_engine); 1042 int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1043 uint64_t fence_address, uint32_t fence_value); 1044 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1045 1046 /* Packet sizes */ 1047 int map_process_size; 1048 int runlist_size; 1049 int set_resources_size; 1050 int map_queues_size; 1051 int unmap_queues_size; 1052 int query_status_size; 1053 int release_mem_size; 1054 }; 1055 1056 extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1057 extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1058 1059 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1060 void pm_uninit(struct packet_manager *pm, bool hanging); 1061 int pm_send_set_resources(struct packet_manager *pm, 1062 struct scheduling_resources *res); 1063 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 1064 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 1065 uint32_t fence_value); 1066 1067 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 1068 enum kfd_unmap_queues_filter mode, 1069 uint32_t filter_param, bool reset, 1070 unsigned int sdma_engine); 1071 1072 void pm_release_ib(struct packet_manager *pm); 1073 1074 /* Following PM funcs can be shared among VI and AI */ 1075 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 1076 1077 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 1078 1079 /* Events */ 1080 extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1081 extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1082 1083 extern const struct kfd_device_global_init_class device_global_init_class_cik; 1084 1085 void kfd_event_init_process(struct kfd_process *p); 1086 void kfd_event_free_process(struct kfd_process *p); 1087 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1088 int kfd_wait_on_events(struct kfd_process *p, 1089 uint32_t num_events, void __user *data, 1090 bool all, uint32_t user_timeout_ms, 1091 uint32_t *wait_result); 1092 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 1093 uint32_t valid_id_bits); 1094 void kfd_signal_iommu_event(struct kfd_dev *dev, 1095 u32 pasid, unsigned long address, 1096 bool is_write_requested, bool is_execute_requested); 1097 void kfd_signal_hw_exception_event(u32 pasid); 1098 int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1099 int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 1100 int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 1101 uint64_t size); 1102 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1103 uint32_t event_type, bool auto_reset, uint32_t node_id, 1104 uint32_t *event_id, uint32_t *event_trigger_data, 1105 uint64_t *event_page_offset, uint32_t *event_slot_index); 1106 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1107 1108 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, 1109 struct kfd_vm_fault_info *info); 1110 1111 void kfd_signal_reset_event(struct kfd_dev *dev); 1112 1113 void kfd_flush_tlb(struct kfd_process_device *pdd); 1114 1115 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1116 1117 bool kfd_is_locked(void); 1118 1119 /* Compute profile */ 1120 void kfd_inc_compute_active(struct kfd_dev *dev); 1121 void kfd_dec_compute_active(struct kfd_dev *dev); 1122 1123 /* Cgroup Support */ 1124 /* Check with device cgroup if @kfd device is accessible */ 1125 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 1126 { 1127 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 1128 struct drm_device *ddev = kfd->ddev; 1129 1130 return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 1131 ddev->render->index, 1132 DEVCG_ACC_WRITE | DEVCG_ACC_READ); 1133 #else 1134 return 0; 1135 #endif 1136 } 1137 1138 /* Debugfs */ 1139 #if defined(CONFIG_DEBUG_FS) 1140 1141 void kfd_debugfs_init(void); 1142 void kfd_debugfs_fini(void); 1143 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1144 int pqm_debugfs_mqds(struct seq_file *m, void *data); 1145 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1146 int dqm_debugfs_hqds(struct seq_file *m, void *data); 1147 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1148 int pm_debugfs_runlist(struct seq_file *m, void *data); 1149 1150 int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1151 int pm_debugfs_hang_hws(struct packet_manager *pm); 1152 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1153 1154 #else 1155 1156 static inline void kfd_debugfs_init(void) {} 1157 static inline void kfd_debugfs_fini(void) {} 1158 1159 #endif 1160 1161 #endif 1162