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 <kgd_kfd_interface.h> 35 36 #define KFD_SYSFS_FILE_MODE 0444 37 38 /* 39 * When working with cp scheduler we should assign the HIQ manually or via 40 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot 41 * definitions for Kaveri. In Kaveri only the first ME queues participates 42 * in the cp scheduling taking that in mind we set the HIQ slot in the 43 * second ME. 44 */ 45 #define KFD_CIK_HIQ_PIPE 4 46 #define KFD_CIK_HIQ_QUEUE 0 47 48 /* GPU ID hash width in bits */ 49 #define KFD_GPU_ID_HASH_WIDTH 16 50 51 /* Macro for allocating structures */ 52 #define kfd_alloc_struct(ptr_to_struct) \ 53 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 54 55 #define KFD_MAX_NUM_OF_PROCESSES 512 56 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 57 58 /* 59 * Kernel module parameter to specify maximum number of supported queues per 60 * device 61 */ 62 extern int max_num_of_queues_per_device; 63 64 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096 65 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 66 (KFD_MAX_NUM_OF_PROCESSES * \ 67 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 68 69 #define KFD_KERNEL_QUEUE_SIZE 2048 70 71 /* Kernel module parameter to specify the scheduling policy */ 72 extern int sched_policy; 73 74 /** 75 * enum kfd_sched_policy 76 * 77 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp) 78 * scheduling. In this scheduling mode we're using the firmware code to 79 * schedule the user mode queues and kernel queues such as HIQ and DIQ. 80 * the HIQ queue is used as a special queue that dispatches the configuration 81 * to the cp and the user mode queues list that are currently running. 82 * the DIQ queue is a debugging queue that dispatches debugging commands to the 83 * firmware. 84 * in this scheduling mode user mode queues over subscription feature is 85 * enabled. 86 * 87 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over 88 * subscription feature disabled. 89 * 90 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly 91 * set the command processor registers and sets the queues "manually". This 92 * mode is used *ONLY* for debugging proposes. 93 * 94 */ 95 enum kfd_sched_policy { 96 KFD_SCHED_POLICY_HWS = 0, 97 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION, 98 KFD_SCHED_POLICY_NO_HWS 99 }; 100 101 enum cache_policy { 102 cache_policy_coherent, 103 cache_policy_noncoherent 104 }; 105 106 enum asic_family_type { 107 CHIP_KAVERI = 0, 108 CHIP_CARRIZO 109 }; 110 111 struct kfd_device_info { 112 unsigned int asic_family; 113 unsigned int max_pasid_bits; 114 size_t ih_ring_entry_size; 115 uint8_t num_of_watch_points; 116 uint16_t mqd_size_aligned; 117 }; 118 119 struct kfd_mem_obj { 120 uint32_t range_start; 121 uint32_t range_end; 122 uint64_t gpu_addr; 123 uint32_t *cpu_ptr; 124 }; 125 126 struct kfd_dev { 127 struct kgd_dev *kgd; 128 129 const struct kfd_device_info *device_info; 130 struct pci_dev *pdev; 131 132 unsigned int id; /* topology stub index */ 133 134 phys_addr_t doorbell_base; /* Start of actual doorbells used by 135 * KFD. It is aligned for mapping 136 * into user mode 137 */ 138 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell 139 * to HW doorbell, GFX reserved some 140 * at the start) 141 */ 142 size_t doorbell_process_limit; /* Number of processes we have doorbell 143 * space for. 144 */ 145 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 146 * page used by kernel queue 147 */ 148 149 struct kgd2kfd_shared_resources shared_resources; 150 151 void *gtt_mem; 152 uint64_t gtt_start_gpu_addr; 153 void *gtt_start_cpu_ptr; 154 void *gtt_sa_bitmap; 155 struct mutex gtt_sa_lock; 156 unsigned int gtt_sa_chunk_size; 157 unsigned int gtt_sa_num_of_chunks; 158 159 /* QCM Device instance */ 160 struct device_queue_manager *dqm; 161 162 bool init_complete; 163 }; 164 165 /* KGD2KFD callbacks */ 166 void kgd2kfd_exit(void); 167 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev); 168 bool kgd2kfd_device_init(struct kfd_dev *kfd, 169 const struct kgd2kfd_shared_resources *gpu_resources); 170 void kgd2kfd_device_exit(struct kfd_dev *kfd); 171 172 extern const struct kfd2kgd_calls *kfd2kgd; 173 174 enum kfd_mempool { 175 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 176 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 177 KFD_MEMPOOL_FRAMEBUFFER = 3, 178 }; 179 180 /* Character device interface */ 181 int kfd_chardev_init(void); 182 void kfd_chardev_exit(void); 183 struct device *kfd_chardev(void); 184 185 /** 186 * enum kfd_preempt_type_filter 187 * 188 * @KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: Preempts single queue. 189 * 190 * @KFD_PRERMPT_TYPE_FILTER_ALL_QUEUES: Preempts all queues in the 191 * running queues list. 192 * 193 * @KFD_PRERMPT_TYPE_FILTER_BY_PASID: Preempts queues that belongs to 194 * specific process. 195 * 196 */ 197 enum kfd_preempt_type_filter { 198 KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE, 199 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 200 KFD_PREEMPT_TYPE_FILTER_BY_PASID 201 }; 202 203 enum kfd_preempt_type { 204 KFD_PREEMPT_TYPE_WAVEFRONT, 205 KFD_PREEMPT_TYPE_WAVEFRONT_RESET 206 }; 207 208 /** 209 * enum kfd_queue_type 210 * 211 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 212 * 213 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type. 214 * 215 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 216 * 217 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 218 */ 219 enum kfd_queue_type { 220 KFD_QUEUE_TYPE_COMPUTE, 221 KFD_QUEUE_TYPE_SDMA, 222 KFD_QUEUE_TYPE_HIQ, 223 KFD_QUEUE_TYPE_DIQ 224 }; 225 226 enum kfd_queue_format { 227 KFD_QUEUE_FORMAT_PM4, 228 KFD_QUEUE_FORMAT_AQL 229 }; 230 231 /** 232 * struct queue_properties 233 * 234 * @type: The queue type. 235 * 236 * @queue_id: Queue identifier. 237 * 238 * @queue_address: Queue ring buffer address. 239 * 240 * @queue_size: Queue ring buffer size. 241 * 242 * @priority: Defines the queue priority relative to other queues in the 243 * process. 244 * This is just an indication and HW scheduling may override the priority as 245 * necessary while keeping the relative prioritization. 246 * the priority granularity is from 0 to f which f is the highest priority. 247 * currently all queues are initialized with the highest priority. 248 * 249 * @queue_percent: This field is partially implemented and currently a zero in 250 * this field defines that the queue is non active. 251 * 252 * @read_ptr: User space address which points to the number of dwords the 253 * cp read from the ring buffer. This field updates automatically by the H/W. 254 * 255 * @write_ptr: Defines the number of dwords written to the ring buffer. 256 * 257 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to 258 * the queue ring buffer. This field should be similar to write_ptr and the user 259 * should update this field after he updated the write_ptr. 260 * 261 * @doorbell_off: The doorbell offset in the doorbell pci-bar. 262 * 263 * @is_interop: Defines if this is a interop queue. Interop queue means that the 264 * queue can access both graphics and compute resources. 265 * 266 * @is_active: Defines if the queue is active or not. 267 * 268 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 269 * of the queue. 270 * 271 * This structure represents the queue properties for each queue no matter if 272 * it's user mode or kernel mode queue. 273 * 274 */ 275 struct queue_properties { 276 enum kfd_queue_type type; 277 enum kfd_queue_format format; 278 unsigned int queue_id; 279 uint64_t queue_address; 280 uint64_t queue_size; 281 uint32_t priority; 282 uint32_t queue_percent; 283 uint32_t *read_ptr; 284 uint32_t *write_ptr; 285 uint32_t __iomem *doorbell_ptr; 286 uint32_t doorbell_off; 287 bool is_interop; 288 bool is_active; 289 /* Not relevant for user mode queues in cp scheduling */ 290 unsigned int vmid; 291 /* Relevant only for sdma queues*/ 292 uint32_t sdma_engine_id; 293 uint32_t sdma_queue_id; 294 uint32_t sdma_vm_addr; 295 /* Relevant only for VI */ 296 uint64_t eop_ring_buffer_address; 297 uint32_t eop_ring_buffer_size; 298 uint64_t ctx_save_restore_area_address; 299 uint32_t ctx_save_restore_area_size; 300 }; 301 302 /** 303 * struct queue 304 * 305 * @list: Queue linked list. 306 * 307 * @mqd: The queue MQD. 308 * 309 * @mqd_mem_obj: The MQD local gpu memory object. 310 * 311 * @gart_mqd_addr: The MQD gart mc address. 312 * 313 * @properties: The queue properties. 314 * 315 * @mec: Used only in no cp scheduling mode and identifies to micro engine id 316 * that the queue should be execute on. 317 * 318 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe id. 319 * 320 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 321 * 322 * @process: The kfd process that created this queue. 323 * 324 * @device: The kfd device that created this queue. 325 * 326 * This structure represents user mode compute queues. 327 * It contains all the necessary data to handle such queues. 328 * 329 */ 330 331 struct queue { 332 struct list_head list; 333 void *mqd; 334 struct kfd_mem_obj *mqd_mem_obj; 335 uint64_t gart_mqd_addr; 336 struct queue_properties properties; 337 338 uint32_t mec; 339 uint32_t pipe; 340 uint32_t queue; 341 342 unsigned int sdma_id; 343 344 struct kfd_process *process; 345 struct kfd_dev *device; 346 }; 347 348 /* 349 * Please read the kfd_mqd_manager.h description. 350 */ 351 enum KFD_MQD_TYPE { 352 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ 353 KFD_MQD_TYPE_HIQ, /* for hiq */ 354 KFD_MQD_TYPE_CP, /* for cp queues and diq */ 355 KFD_MQD_TYPE_SDMA, /* for sdma queues */ 356 KFD_MQD_TYPE_MAX 357 }; 358 359 struct scheduling_resources { 360 unsigned int vmid_mask; 361 enum kfd_queue_type type; 362 uint64_t queue_mask; 363 uint64_t gws_mask; 364 uint32_t oac_mask; 365 uint32_t gds_heap_base; 366 uint32_t gds_heap_size; 367 }; 368 369 struct process_queue_manager { 370 /* data */ 371 struct kfd_process *process; 372 unsigned int num_concurrent_processes; 373 struct list_head queues; 374 unsigned long *queue_slot_bitmap; 375 }; 376 377 struct qcm_process_device { 378 /* The Device Queue Manager that owns this data */ 379 struct device_queue_manager *dqm; 380 struct process_queue_manager *pqm; 381 /* Device Queue Manager lock */ 382 struct mutex *lock; 383 /* Queues list */ 384 struct list_head queues_list; 385 struct list_head priv_queue_list; 386 387 unsigned int queue_count; 388 unsigned int vmid; 389 bool is_debug; 390 /* 391 * All the memory management data should be here too 392 */ 393 uint64_t gds_context_area; 394 uint32_t sh_mem_config; 395 uint32_t sh_mem_bases; 396 uint32_t sh_mem_ape1_base; 397 uint32_t sh_mem_ape1_limit; 398 uint32_t page_table_base; 399 uint32_t gds_size; 400 uint32_t num_gws; 401 uint32_t num_oac; 402 }; 403 404 /* Data that is per-process-per device. */ 405 struct kfd_process_device { 406 /* 407 * List of all per-device data for a process. 408 * Starts from kfd_process.per_device_data. 409 */ 410 struct list_head per_device_list; 411 412 /* The device that owns this data. */ 413 struct kfd_dev *dev; 414 415 416 /* per-process-per device QCM data structure */ 417 struct qcm_process_device qpd; 418 419 /*Apertures*/ 420 uint64_t lds_base; 421 uint64_t lds_limit; 422 uint64_t gpuvm_base; 423 uint64_t gpuvm_limit; 424 uint64_t scratch_base; 425 uint64_t scratch_limit; 426 427 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 428 bool bound; 429 }; 430 431 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 432 433 /* Process data */ 434 struct kfd_process { 435 /* 436 * kfd_process are stored in an mm_struct*->kfd_process* 437 * hash table (kfd_processes in kfd_process.c) 438 */ 439 struct hlist_node kfd_processes; 440 441 struct mm_struct *mm; 442 443 struct mutex mutex; 444 445 /* 446 * In any process, the thread that started main() is the lead 447 * thread and outlives the rest. 448 * It is here because amd_iommu_bind_pasid wants a task_struct. 449 */ 450 struct task_struct *lead_thread; 451 452 /* We want to receive a notification when the mm_struct is destroyed */ 453 struct mmu_notifier mmu_notifier; 454 455 /* Use for delayed freeing of kfd_process structure */ 456 struct rcu_head rcu; 457 458 unsigned int pasid; 459 460 /* 461 * List of kfd_process_device structures, 462 * one for each device the process is using. 463 */ 464 struct list_head per_device_data; 465 466 struct process_queue_manager pqm; 467 468 /* The process's queues. */ 469 size_t queue_array_size; 470 471 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */ 472 struct kfd_queue **queues; 473 474 unsigned long allocated_queue_bitmap[DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)]; 475 476 /*Is the user space process 32 bit?*/ 477 bool is_32bit_user_mode; 478 }; 479 480 /** 481 * Ioctl function type. 482 * 483 * \param filep pointer to file structure. 484 * \param p amdkfd process pointer. 485 * \param data pointer to arg that was copied from user. 486 */ 487 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 488 void *data); 489 490 struct amdkfd_ioctl_desc { 491 unsigned int cmd; 492 int flags; 493 amdkfd_ioctl_t *func; 494 unsigned int cmd_drv; 495 const char *name; 496 }; 497 498 void kfd_process_create_wq(void); 499 void kfd_process_destroy_wq(void); 500 struct kfd_process *kfd_create_process(const struct task_struct *); 501 struct kfd_process *kfd_get_process(const struct task_struct *); 502 503 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 504 struct kfd_process *p); 505 void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid); 506 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 507 struct kfd_process *p); 508 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 509 struct kfd_process *p); 510 511 /* Process device data iterator */ 512 struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p); 513 struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p, 514 struct kfd_process_device *pdd); 515 bool kfd_has_process_device_data(struct kfd_process *p); 516 517 /* PASIDs */ 518 int kfd_pasid_init(void); 519 void kfd_pasid_exit(void); 520 bool kfd_set_pasid_limit(unsigned int new_limit); 521 unsigned int kfd_get_pasid_limit(void); 522 unsigned int kfd_pasid_alloc(void); 523 void kfd_pasid_free(unsigned int pasid); 524 525 /* Doorbells */ 526 void kfd_doorbell_init(struct kfd_dev *kfd); 527 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma); 528 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 529 unsigned int *doorbell_off); 530 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 531 u32 read_kernel_doorbell(u32 __iomem *db); 532 void write_kernel_doorbell(u32 __iomem *db, u32 value); 533 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 534 struct kfd_process *process, 535 unsigned int queue_id); 536 537 /* GTT Sub-Allocator */ 538 539 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 540 struct kfd_mem_obj **mem_obj); 541 542 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 543 544 extern struct device *kfd_device; 545 546 /* Topology */ 547 int kfd_topology_init(void); 548 void kfd_topology_shutdown(void); 549 int kfd_topology_add_device(struct kfd_dev *gpu); 550 int kfd_topology_remove_device(struct kfd_dev *gpu); 551 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 552 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 553 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx); 554 555 /* Interrupts */ 556 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 557 558 /* Power Management */ 559 void kgd2kfd_suspend(struct kfd_dev *kfd); 560 int kgd2kfd_resume(struct kfd_dev *kfd); 561 562 /* amdkfd Apertures */ 563 int kfd_init_apertures(struct kfd_process *process); 564 565 /* Queue Context Management */ 566 inline uint32_t lower_32(uint64_t x); 567 inline uint32_t upper_32(uint64_t x); 568 struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd); 569 inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m); 570 571 int init_queue(struct queue **q, struct queue_properties properties); 572 void uninit_queue(struct queue *q); 573 void print_queue_properties(struct queue_properties *q); 574 void print_queue(struct queue *q); 575 576 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type, 577 struct kfd_dev *dev); 578 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 579 struct kfd_dev *dev); 580 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 581 struct kfd_dev *dev); 582 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 583 void device_queue_manager_uninit(struct device_queue_manager *dqm); 584 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 585 enum kfd_queue_type type); 586 void kernel_queue_uninit(struct kernel_queue *kq); 587 588 /* Process Queue Manager */ 589 struct process_queue_node { 590 struct queue *q; 591 struct kernel_queue *kq; 592 struct list_head process_queue_list; 593 }; 594 595 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 596 void pqm_uninit(struct process_queue_manager *pqm); 597 int pqm_create_queue(struct process_queue_manager *pqm, 598 struct kfd_dev *dev, 599 struct file *f, 600 struct queue_properties *properties, 601 unsigned int flags, 602 enum kfd_queue_type type, 603 unsigned int *qid); 604 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 605 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 606 struct queue_properties *p); 607 608 /* Packet Manager */ 609 610 #define KFD_HIQ_TIMEOUT (500) 611 612 #define KFD_FENCE_COMPLETED (100) 613 #define KFD_FENCE_INIT (10) 614 #define KFD_UNMAP_LATENCY (150) 615 616 struct packet_manager { 617 struct device_queue_manager *dqm; 618 struct kernel_queue *priv_queue; 619 struct mutex lock; 620 bool allocated; 621 struct kfd_mem_obj *ib_buffer_obj; 622 }; 623 624 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 625 void pm_uninit(struct packet_manager *pm); 626 int pm_send_set_resources(struct packet_manager *pm, 627 struct scheduling_resources *res); 628 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 629 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 630 uint32_t fence_value); 631 632 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 633 enum kfd_preempt_type_filter mode, 634 uint32_t filter_param, bool reset, 635 unsigned int sdma_engine); 636 637 void pm_release_ib(struct packet_manager *pm); 638 639 uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 640 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 641 struct kfd_process *process); 642 643 #endif 644