| 1 | /* SPDX-License-Identifier: GPL-2.0 OR MIT */ |
| 2 | /* |
| 3 | * Copyright 2014-2022 Advanced Micro Devices, Inc. |
| 4 | * |
| 5 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 6 | * copy of this software and associated documentation files (the "Software"), |
| 7 | * to deal in the Software without restriction, including without limitation |
| 8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 9 | * and/or sell copies of the Software, and to permit persons to whom the |
| 10 | * Software is furnished to do so, subject to the following conditions: |
| 11 | * |
| 12 | * The above copyright notice and this permission notice shall be included in |
| 13 | * all copies or substantial portions of the Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 18 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 19 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 20 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 21 | * OTHER DEALINGS IN THE SOFTWARE. |
| 22 | */ |
| 23 | |
| 24 | #ifndef KFD_PRIV_H_INCLUDED |
| 25 | #define KFD_PRIV_H_INCLUDED |
| 26 | |
| 27 | #include <linux/hashtable.h> |
| 28 | #include <linux/mmu_notifier.h> |
| 29 | #include <linux/memremap.h> |
| 30 | #include <linux/mutex.h> |
| 31 | #include <linux/types.h> |
| 32 | #include <linux/atomic.h> |
| 33 | #include <linux/workqueue.h> |
| 34 | #include <linux/spinlock.h> |
| 35 | #include <uapi/linux/kfd_ioctl.h> |
| 36 | #include <linux/idr.h> |
| 37 | #include <linux/kfifo.h> |
| 38 | #include <linux/seq_file.h> |
| 39 | #include <linux/kref.h> |
| 40 | #include <linux/sysfs.h> |
| 41 | #include <linux/device_cgroup.h> |
| 42 | #include <drm/drm_file.h> |
| 43 | #include <drm/drm_drv.h> |
| 44 | #include <drm/drm_device.h> |
| 45 | #include <drm/drm_ioctl.h> |
| 46 | #include <kgd_kfd_interface.h> |
| 47 | #include <linux/swap.h> |
| 48 | |
| 49 | #include "amd_shared.h" |
| 50 | #include "amdgpu.h" |
| 51 | |
| 52 | #define KFD_MAX_RING_ENTRY_SIZE 8 |
| 53 | |
| 54 | #define KFD_SYSFS_FILE_MODE 0444 |
| 55 | |
| 56 | /* GPU ID hash width in bits */ |
| 57 | #define KFD_GPU_ID_HASH_WIDTH 16 |
| 58 | |
| 59 | /* Use upper bits of mmap offset to store KFD driver specific information. |
| 60 | * BITS[63:62] - Encode MMAP type |
| 61 | * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to |
| 62 | * BITS[45:0] - MMAP offset value |
| 63 | * |
| 64 | * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these |
| 65 | * defines are w.r.t to PAGE_SIZE |
| 66 | */ |
| 67 | #define KFD_MMAP_TYPE_SHIFT 62 |
| 68 | #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) |
| 69 | #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) |
| 70 | #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) |
| 71 | #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) |
| 72 | #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) |
| 73 | |
| 74 | #define KFD_MMAP_GPU_ID_SHIFT 46 |
| 75 | #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ |
| 76 | << KFD_MMAP_GPU_ID_SHIFT) |
| 77 | #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ |
| 78 | & KFD_MMAP_GPU_ID_MASK) |
| 79 | #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ |
| 80 | >> KFD_MMAP_GPU_ID_SHIFT) |
| 81 | |
| 82 | /* |
| 83 | * When working with cp scheduler we should assign the HIQ manually or via |
| 84 | * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot |
| 85 | * definitions for Kaveri. In Kaveri only the first ME queues participates |
| 86 | * in the cp scheduling taking that in mind we set the HIQ slot in the |
| 87 | * second ME. |
| 88 | */ |
| 89 | #define KFD_CIK_HIQ_PIPE 4 |
| 90 | #define KFD_CIK_HIQ_QUEUE 0 |
| 91 | |
| 92 | /* Macro for allocating structures */ |
| 93 | #define kfd_alloc_struct(ptr_to_struct) \ |
| 94 | ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) |
| 95 | |
| 96 | #define KFD_MAX_NUM_OF_PROCESSES 512 |
| 97 | #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 |
| 98 | |
| 99 | /* |
| 100 | * Size of the per-process TBA+TMA buffer: 2 pages |
| 101 | * |
| 102 | * The first chunk is the TBA used for the CWSR ISA code. The second |
| 103 | * chunk is used as TMA for user-mode trap handler setup in daisy-chain mode. |
| 104 | */ |
| 105 | #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) |
| 106 | #define KFD_CWSR_TMA_OFFSET (PAGE_SIZE + 2048) |
| 107 | |
| 108 | #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ |
| 109 | (KFD_MAX_NUM_OF_PROCESSES * \ |
| 110 | KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) |
| 111 | |
| 112 | #define KFD_KERNEL_QUEUE_SIZE 2048 |
| 113 | |
| 114 | /* KFD_UNMAP_LATENCY_MS is the timeout CP waiting for SDMA preemption. One XCC |
| 115 | * can be associated to 2 SDMA engines. queue_preemption_timeout_ms is the time |
| 116 | * driver waiting for CP returning the UNMAP_QUEUE fence. Thus the math is |
| 117 | * queue_preemption_timeout_ms = sdma_preemption_time * 2 + cp workload |
| 118 | * The format here makes CP workload 10% of total timeout |
| 119 | */ |
| 120 | #define KFD_UNMAP_LATENCY_MS \ |
| 121 | ((queue_preemption_timeout_ms - queue_preemption_timeout_ms / 10) >> 1) |
| 122 | |
| 123 | #define KFD_MAX_SDMA_QUEUES 128 |
| 124 | |
| 125 | /* |
| 126 | * 512 = 0x200 |
| 127 | * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the |
| 128 | * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. |
| 129 | * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC |
| 130 | * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in |
| 131 | * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. |
| 132 | */ |
| 133 | #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 |
| 134 | |
| 135 | /** |
| 136 | * enum kfd_ioctl_flags - KFD ioctl flags |
| 137 | * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how |
| 138 | * userspace can use a given ioctl. |
| 139 | */ |
| 140 | enum kfd_ioctl_flags { |
| 141 | /* |
| 142 | * @KFD_IOC_FLAG_CHECKPOINT_RESTORE: |
| 143 | * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially |
| 144 | * perform privileged operations and load arbitrary data into MQDs and |
| 145 | * eventually HQD registers when the queue is mapped by HWS. In order to |
| 146 | * prevent this we should perform additional security checks. |
| 147 | * |
| 148 | * This is equivalent to callers with the CHECKPOINT_RESTORE capability. |
| 149 | * |
| 150 | * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE, |
| 151 | * we also allow ioctls with SYS_ADMIN capability. |
| 152 | */ |
| 153 | KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0), |
| 154 | }; |
| 155 | /* |
| 156 | * Kernel module parameter to specify maximum number of supported queues per |
| 157 | * device |
| 158 | */ |
| 159 | extern int max_num_of_queues_per_device; |
| 160 | |
| 161 | |
| 162 | /* Kernel module parameter to specify the scheduling policy */ |
| 163 | extern int sched_policy; |
| 164 | |
| 165 | /* |
| 166 | * Kernel module parameter to specify the maximum process |
| 167 | * number per HW scheduler |
| 168 | */ |
| 169 | extern int hws_max_conc_proc; |
| 170 | |
| 171 | extern int cwsr_enable; |
| 172 | |
| 173 | /* |
| 174 | * Kernel module parameter to specify whether to send sigterm to HSA process on |
| 175 | * unhandled exception |
| 176 | */ |
| 177 | extern int send_sigterm; |
| 178 | |
| 179 | /* |
| 180 | * This kernel module is used to simulate large bar machine on non-large bar |
| 181 | * enabled machines. |
| 182 | */ |
| 183 | extern int debug_largebar; |
| 184 | |
| 185 | /* Set sh_mem_config.retry_disable on GFX v9 */ |
| 186 | extern int amdgpu_noretry; |
| 187 | |
| 188 | /* Halt if HWS hang is detected */ |
| 189 | extern int halt_if_hws_hang; |
| 190 | |
| 191 | /* Whether MEC FW support GWS barriers */ |
| 192 | extern bool hws_gws_support; |
| 193 | |
| 194 | /* Queue preemption timeout in ms */ |
| 195 | extern int queue_preemption_timeout_ms; |
| 196 | |
| 197 | /* |
| 198 | * Don't evict process queues on vm fault |
| 199 | */ |
| 200 | extern int amdgpu_no_queue_eviction_on_vm_fault; |
| 201 | |
| 202 | /* Enable eviction debug messages */ |
| 203 | extern bool debug_evictions; |
| 204 | |
| 205 | extern struct mutex kfd_processes_mutex; |
| 206 | |
| 207 | enum cache_policy { |
| 208 | cache_policy_coherent, |
| 209 | cache_policy_noncoherent |
| 210 | }; |
| 211 | |
| 212 | #define KFD_GC_VERSION(dev) (amdgpu_ip_version((dev)->adev, GC_HWIP, 0)) |
| 213 | #define KFD_IS_SOC15(dev) ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1))) |
| 214 | #define KFD_SUPPORT_XNACK_PER_PROCESS(dev)\ |
| 215 | ((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) || \ |
| 216 | (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) || \ |
| 217 | (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)) || \ |
| 218 | (KFD_GC_VERSION(dev) == IP_VERSION(9, 5, 0))) |
| 219 | |
| 220 | struct kfd_node; |
| 221 | |
| 222 | struct kfd_event_interrupt_class { |
| 223 | bool (*interrupt_isr)(struct kfd_node *dev, |
| 224 | const uint32_t *ih_ring_entry, uint32_t *patched_ihre, |
| 225 | bool *patched_flag); |
| 226 | void (*interrupt_wq)(struct kfd_node *dev, |
| 227 | const uint32_t *ih_ring_entry); |
| 228 | }; |
| 229 | |
| 230 | struct kfd_device_info { |
| 231 | uint32_t gfx_target_version; |
| 232 | const struct kfd_event_interrupt_class *event_interrupt_class; |
| 233 | unsigned int max_pasid_bits; |
| 234 | unsigned int max_no_of_hqd; |
| 235 | unsigned int doorbell_size; |
| 236 | size_t ih_ring_entry_size; |
| 237 | uint8_t num_of_watch_points; |
| 238 | uint16_t mqd_size_aligned; |
| 239 | bool supports_cwsr; |
| 240 | bool needs_pci_atomics; |
| 241 | uint32_t no_atomic_fw_version; |
| 242 | unsigned int num_sdma_queues_per_engine; |
| 243 | unsigned int num_reserved_sdma_queues_per_engine; |
| 244 | DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES); |
| 245 | }; |
| 246 | |
| 247 | unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev); |
| 248 | unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev); |
| 249 | |
| 250 | struct kfd_mem_obj { |
| 251 | uint32_t range_start; |
| 252 | uint32_t range_end; |
| 253 | uint64_t gpu_addr; |
| 254 | uint32_t *cpu_ptr; |
| 255 | void *gtt_mem; |
| 256 | }; |
| 257 | |
| 258 | struct kfd_vmid_info { |
| 259 | uint32_t first_vmid_kfd; |
| 260 | uint32_t last_vmid_kfd; |
| 261 | uint32_t vmid_num_kfd; |
| 262 | }; |
| 263 | |
| 264 | #define MAX_KFD_NODES 8 |
| 265 | |
| 266 | struct kfd_dev; |
| 267 | |
| 268 | struct kfd_node { |
| 269 | unsigned int node_id; |
| 270 | struct amdgpu_device *adev; /* Duplicated here along with keeping |
| 271 | * a copy in kfd_dev to save a hop |
| 272 | */ |
| 273 | const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with |
| 274 | * keeping a copy in kfd_dev to |
| 275 | * save a hop |
| 276 | */ |
| 277 | struct kfd_vmid_info vm_info; |
| 278 | unsigned int id; /* topology stub index */ |
| 279 | uint32_t xcc_mask; /* Instance mask of XCCs present */ |
| 280 | struct amdgpu_xcp *xcp; |
| 281 | |
| 282 | /* Interrupts */ |
| 283 | struct kfifo ih_fifo; |
| 284 | struct work_struct interrupt_work; |
| 285 | spinlock_t interrupt_lock; |
| 286 | |
| 287 | /* |
| 288 | * Interrupts of interest to KFD are copied |
| 289 | * from the HW ring into a SW ring. |
| 290 | */ |
| 291 | bool interrupts_active; |
| 292 | uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */ |
| 293 | |
| 294 | /* QCM Device instance */ |
| 295 | struct device_queue_manager *dqm; |
| 296 | |
| 297 | /* Global GWS resource shared between processes */ |
| 298 | void *gws; |
| 299 | |
| 300 | /* Clients watching SMI events */ |
| 301 | struct list_head smi_clients; |
| 302 | spinlock_t smi_lock; |
| 303 | uint32_t reset_seq_num; |
| 304 | |
| 305 | /* SRAM ECC flag */ |
| 306 | atomic_t sram_ecc_flag; |
| 307 | |
| 308 | /*spm process id */ |
| 309 | unsigned int spm_pasid; |
| 310 | |
| 311 | /* Maximum process number mapped to HW scheduler */ |
| 312 | unsigned int max_proc_per_quantum; |
| 313 | |
| 314 | unsigned int compute_vmid_bitmap; |
| 315 | |
| 316 | struct kfd_local_mem_info local_mem_info; |
| 317 | |
| 318 | struct kfd_dev *kfd; |
| 319 | |
| 320 | /* Track per device allocated watch points */ |
| 321 | uint32_t alloc_watch_ids; |
| 322 | spinlock_t watch_points_lock; |
| 323 | }; |
| 324 | |
| 325 | struct kfd_dev { |
| 326 | struct amdgpu_device *adev; |
| 327 | |
| 328 | struct kfd_device_info device_info; |
| 329 | |
| 330 | u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells |
| 331 | * page used by kernel queue |
| 332 | */ |
| 333 | |
| 334 | struct kgd2kfd_shared_resources shared_resources; |
| 335 | |
| 336 | const struct kfd2kgd_calls *kfd2kgd; |
| 337 | struct mutex doorbell_mutex; |
| 338 | |
| 339 | void *gtt_mem; |
| 340 | uint64_t gtt_start_gpu_addr; |
| 341 | void *gtt_start_cpu_ptr; |
| 342 | void *gtt_sa_bitmap; |
| 343 | struct mutex gtt_sa_lock; |
| 344 | unsigned int gtt_sa_chunk_size; |
| 345 | unsigned int gtt_sa_num_of_chunks; |
| 346 | |
| 347 | bool init_complete; |
| 348 | |
| 349 | /* Firmware versions */ |
| 350 | uint16_t mec_fw_version; |
| 351 | uint16_t mec2_fw_version; |
| 352 | uint16_t sdma_fw_version; |
| 353 | |
| 354 | /* CWSR */ |
| 355 | bool cwsr_enabled; |
| 356 | const void *cwsr_isa; |
| 357 | unsigned int cwsr_isa_size; |
| 358 | |
| 359 | /* xGMI */ |
| 360 | uint64_t hive_id; |
| 361 | |
| 362 | bool pci_atomic_requested; |
| 363 | |
| 364 | /* Compute Profile ref. count */ |
| 365 | atomic_t compute_profile; |
| 366 | |
| 367 | struct ida doorbell_ida; |
| 368 | unsigned int max_doorbell_slices; |
| 369 | |
| 370 | int noretry; |
| 371 | |
| 372 | struct kfd_node *nodes[MAX_KFD_NODES]; |
| 373 | unsigned int num_nodes; |
| 374 | |
| 375 | struct workqueue_struct *ih_wq; |
| 376 | |
| 377 | /* Kernel doorbells for KFD device */ |
| 378 | struct amdgpu_bo *doorbells; |
| 379 | |
| 380 | /* bitmap for dynamic doorbell allocation from doorbell object */ |
| 381 | unsigned long *doorbell_bitmap; |
| 382 | |
| 383 | /* for dynamic partitioning */ |
| 384 | int kfd_dev_lock; |
| 385 | |
| 386 | atomic_t kfd_processes_count; |
| 387 | }; |
| 388 | |
| 389 | enum kfd_mempool { |
| 390 | KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, |
| 391 | KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, |
| 392 | KFD_MEMPOOL_FRAMEBUFFER = 3, |
| 393 | }; |
| 394 | |
| 395 | /* Character device interface */ |
| 396 | int kfd_chardev_init(void); |
| 397 | void kfd_chardev_exit(void); |
| 398 | |
| 399 | /** |
| 400 | * enum kfd_unmap_queues_filter - Enum for queue filters. |
| 401 | * |
| 402 | * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the |
| 403 | * running queues list. |
| 404 | * |
| 405 | * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues |
| 406 | * in the run list. |
| 407 | * |
| 408 | * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to |
| 409 | * specific process. |
| 410 | * |
| 411 | */ |
| 412 | enum kfd_unmap_queues_filter { |
| 413 | KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1, |
| 414 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2, |
| 415 | KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3 |
| 416 | }; |
| 417 | |
| 418 | /** |
| 419 | * enum kfd_queue_type - Enum for various queue types. |
| 420 | * |
| 421 | * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. |
| 422 | * |
| 423 | * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. |
| 424 | * |
| 425 | * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. |
| 426 | * |
| 427 | * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. |
| 428 | * |
| 429 | * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. |
| 430 | * |
| 431 | * @KFD_QUEUE_TYPE_SDMA_BY_ENG_ID: SDMA user mode queue with target SDMA engine ID. |
| 432 | */ |
| 433 | enum kfd_queue_type { |
| 434 | KFD_QUEUE_TYPE_COMPUTE, |
| 435 | KFD_QUEUE_TYPE_SDMA, |
| 436 | KFD_QUEUE_TYPE_HIQ, |
| 437 | KFD_QUEUE_TYPE_DIQ, |
| 438 | KFD_QUEUE_TYPE_SDMA_XGMI, |
| 439 | KFD_QUEUE_TYPE_SDMA_BY_ENG_ID |
| 440 | }; |
| 441 | |
| 442 | enum kfd_queue_format { |
| 443 | KFD_QUEUE_FORMAT_PM4, |
| 444 | KFD_QUEUE_FORMAT_AQL |
| 445 | }; |
| 446 | |
| 447 | enum KFD_QUEUE_PRIORITY { |
| 448 | KFD_QUEUE_PRIORITY_MINIMUM = 0, |
| 449 | KFD_QUEUE_PRIORITY_MAXIMUM = 15 |
| 450 | }; |
| 451 | |
| 452 | /** |
| 453 | * struct queue_properties |
| 454 | * |
| 455 | * @type: The queue type. |
| 456 | * |
| 457 | * @queue_id: Queue identifier. |
| 458 | * |
| 459 | * @queue_address: Queue ring buffer address. |
| 460 | * |
| 461 | * @queue_size: Queue ring buffer size. |
| 462 | * |
| 463 | * @priority: Defines the queue priority relative to other queues in the |
| 464 | * process. |
| 465 | * This is just an indication and HW scheduling may override the priority as |
| 466 | * necessary while keeping the relative prioritization. |
| 467 | * the priority granularity is from 0 to f which f is the highest priority. |
| 468 | * currently all queues are initialized with the highest priority. |
| 469 | * |
| 470 | * @queue_percent: This field is partially implemented and currently a zero in |
| 471 | * this field defines that the queue is non active. |
| 472 | * |
| 473 | * @read_ptr: User space address which points to the number of dwords the |
| 474 | * cp read from the ring buffer. This field updates automatically by the H/W. |
| 475 | * |
| 476 | * @write_ptr: Defines the number of dwords written to the ring buffer. |
| 477 | * |
| 478 | * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring |
| 479 | * buffer. This field should be similar to write_ptr and the user should |
| 480 | * update this field after updating the write_ptr. |
| 481 | * |
| 482 | * @doorbell_off: The doorbell offset in the doorbell pci-bar. |
| 483 | * |
| 484 | * @is_interop: Defines if this is a interop queue. Interop queue means that |
| 485 | * the queue can access both graphics and compute resources. |
| 486 | * |
| 487 | * @is_evicted: Defines if the queue is evicted. Only active queues |
| 488 | * are evicted, rendering them inactive. |
| 489 | * |
| 490 | * @is_active: Defines if the queue is active or not. @is_active and |
| 491 | * @is_evicted are protected by the DQM lock. |
| 492 | * |
| 493 | * @is_gws: Defines if the queue has been updated to be GWS-capable or not. |
| 494 | * @is_gws should be protected by the DQM lock, since changing it can yield the |
| 495 | * possibility of updating DQM state on number of GWS queues. |
| 496 | * |
| 497 | * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid |
| 498 | * of the queue. |
| 499 | * |
| 500 | * This structure represents the queue properties for each queue no matter if |
| 501 | * it's user mode or kernel mode queue. |
| 502 | * |
| 503 | */ |
| 504 | |
| 505 | struct queue_properties { |
| 506 | enum kfd_queue_type type; |
| 507 | enum kfd_queue_format format; |
| 508 | unsigned int queue_id; |
| 509 | uint64_t queue_address; |
| 510 | uint64_t queue_size; |
| 511 | uint32_t priority; |
| 512 | uint32_t queue_percent; |
| 513 | void __user *read_ptr; |
| 514 | void __user *write_ptr; |
| 515 | void __iomem *doorbell_ptr; |
| 516 | uint32_t doorbell_off; |
| 517 | bool is_interop; |
| 518 | bool is_evicted; |
| 519 | bool is_suspended; |
| 520 | bool is_being_destroyed; |
| 521 | bool is_active; |
| 522 | bool is_gws; |
| 523 | uint32_t pm4_target_xcc; |
| 524 | bool is_dbg_wa; |
| 525 | bool is_user_cu_masked; |
| 526 | /* Not relevant for user mode queues in cp scheduling */ |
| 527 | unsigned int vmid; |
| 528 | /* Relevant only for sdma queues*/ |
| 529 | uint32_t sdma_engine_id; |
| 530 | uint32_t sdma_queue_id; |
| 531 | uint32_t sdma_vm_addr; |
| 532 | /* Relevant only for VI */ |
| 533 | uint64_t eop_ring_buffer_address; |
| 534 | uint32_t eop_ring_buffer_size; |
| 535 | uint64_t ctx_save_restore_area_address; |
| 536 | uint32_t ctx_save_restore_area_size; |
| 537 | uint32_t ctl_stack_size; |
| 538 | uint64_t tba_addr; |
| 539 | uint64_t tma_addr; |
| 540 | uint64_t exception_status; |
| 541 | |
| 542 | struct amdgpu_bo *wptr_bo; |
| 543 | struct amdgpu_bo *rptr_bo; |
| 544 | struct amdgpu_bo *ring_bo; |
| 545 | struct amdgpu_bo *eop_buf_bo; |
| 546 | struct amdgpu_bo *cwsr_bo; |
| 547 | }; |
| 548 | |
| 549 | #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ |
| 550 | (q).queue_address != 0 && \ |
| 551 | (q).queue_percent > 0 && \ |
| 552 | !(q).is_evicted && \ |
| 553 | !(q).is_suspended) |
| 554 | |
| 555 | enum mqd_update_flag { |
| 556 | UPDATE_FLAG_DBG_WA_ENABLE = 1, |
| 557 | UPDATE_FLAG_DBG_WA_DISABLE = 2, |
| 558 | UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */ |
| 559 | }; |
| 560 | |
| 561 | struct mqd_update_info { |
| 562 | union { |
| 563 | struct { |
| 564 | uint32_t count; /* Must be a multiple of 32 */ |
| 565 | uint32_t *ptr; |
| 566 | } cu_mask; |
| 567 | }; |
| 568 | enum mqd_update_flag update_flag; |
| 569 | }; |
| 570 | |
| 571 | /** |
| 572 | * struct queue |
| 573 | * |
| 574 | * @list: Queue linked list. |
| 575 | * |
| 576 | * @mqd: The queue MQD (memory queue descriptor). |
| 577 | * |
| 578 | * @mqd_mem_obj: The MQD local gpu memory object. |
| 579 | * |
| 580 | * @gart_mqd_addr: The MQD gart mc address. |
| 581 | * |
| 582 | * @properties: The queue properties. |
| 583 | * |
| 584 | * @mec: Used only in no cp scheduling mode and identifies to micro engine id |
| 585 | * that the queue should be executed on. |
| 586 | * |
| 587 | * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe |
| 588 | * id. |
| 589 | * |
| 590 | * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. |
| 591 | * |
| 592 | * @process: The kfd process that created this queue. |
| 593 | * |
| 594 | * @device: The kfd device that created this queue. |
| 595 | * |
| 596 | * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL |
| 597 | * otherwise. |
| 598 | * |
| 599 | * This structure represents user mode compute queues. |
| 600 | * It contains all the necessary data to handle such queues. |
| 601 | * |
| 602 | */ |
| 603 | |
| 604 | struct queue { |
| 605 | struct list_head list; |
| 606 | void *mqd; |
| 607 | struct kfd_mem_obj *mqd_mem_obj; |
| 608 | uint64_t gart_mqd_addr; |
| 609 | struct queue_properties properties; |
| 610 | |
| 611 | uint32_t mec; |
| 612 | uint32_t pipe; |
| 613 | uint32_t queue; |
| 614 | |
| 615 | unsigned int sdma_id; |
| 616 | unsigned int doorbell_id; |
| 617 | |
| 618 | struct kfd_process *process; |
| 619 | struct kfd_node *device; |
| 620 | void *gws; |
| 621 | |
| 622 | /* procfs */ |
| 623 | struct kobject kobj; |
| 624 | |
| 625 | void *gang_ctx_bo; |
| 626 | uint64_t gang_ctx_gpu_addr; |
| 627 | void *gang_ctx_cpu_ptr; |
| 628 | |
| 629 | struct amdgpu_bo *wptr_bo_gart; |
| 630 | }; |
| 631 | |
| 632 | enum KFD_MQD_TYPE { |
| 633 | KFD_MQD_TYPE_HIQ = 0, /* for hiq */ |
| 634 | KFD_MQD_TYPE_CP, /* for cp queues and diq */ |
| 635 | KFD_MQD_TYPE_SDMA, /* for sdma queues */ |
| 636 | KFD_MQD_TYPE_DIQ, /* for diq */ |
| 637 | KFD_MQD_TYPE_MAX |
| 638 | }; |
| 639 | |
| 640 | enum KFD_PIPE_PRIORITY { |
| 641 | KFD_PIPE_PRIORITY_CS_LOW = 0, |
| 642 | KFD_PIPE_PRIORITY_CS_MEDIUM, |
| 643 | KFD_PIPE_PRIORITY_CS_HIGH |
| 644 | }; |
| 645 | |
| 646 | struct scheduling_resources { |
| 647 | unsigned int vmid_mask; |
| 648 | enum kfd_queue_type type; |
| 649 | uint64_t queue_mask; |
| 650 | uint64_t gws_mask; |
| 651 | uint32_t oac_mask; |
| 652 | uint32_t gds_heap_base; |
| 653 | uint32_t gds_heap_size; |
| 654 | }; |
| 655 | |
| 656 | struct process_queue_manager { |
| 657 | /* data */ |
| 658 | struct kfd_process *process; |
| 659 | struct list_head queues; |
| 660 | unsigned long *queue_slot_bitmap; |
| 661 | }; |
| 662 | |
| 663 | struct qcm_process_device { |
| 664 | /* The Device Queue Manager that owns this data */ |
| 665 | struct device_queue_manager *dqm; |
| 666 | struct process_queue_manager *pqm; |
| 667 | /* Queues list */ |
| 668 | struct list_head queues_list; |
| 669 | struct list_head priv_queue_list; |
| 670 | |
| 671 | unsigned int queue_count; |
| 672 | unsigned int vmid; |
| 673 | bool is_debug; |
| 674 | unsigned int evicted; /* eviction counter, 0=active */ |
| 675 | |
| 676 | /* This flag tells if we should reset all wavefronts on |
| 677 | * process termination |
| 678 | */ |
| 679 | bool reset_wavefronts; |
| 680 | |
| 681 | /* This flag tells us if this process has a GWS-capable |
| 682 | * queue that will be mapped into the runlist. It's |
| 683 | * possible to request a GWS BO, but not have the queue |
| 684 | * currently mapped, and this changes how the MAP_PROCESS |
| 685 | * PM4 packet is configured. |
| 686 | */ |
| 687 | bool mapped_gws_queue; |
| 688 | |
| 689 | /* All the memory management data should be here too */ |
| 690 | uint64_t gds_context_area; |
| 691 | /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ |
| 692 | uint64_t page_table_base; |
| 693 | uint32_t sh_mem_config; |
| 694 | uint32_t sh_mem_bases; |
| 695 | uint32_t sh_mem_ape1_base; |
| 696 | uint32_t sh_mem_ape1_limit; |
| 697 | uint32_t gds_size; |
| 698 | uint32_t num_gws; |
| 699 | uint32_t num_oac; |
| 700 | uint32_t sh_hidden_private_base; |
| 701 | |
| 702 | /* CWSR memory */ |
| 703 | struct kgd_mem *cwsr_mem; |
| 704 | void *cwsr_kaddr; |
| 705 | uint64_t cwsr_base; |
| 706 | uint64_t tba_addr; |
| 707 | uint64_t tma_addr; |
| 708 | |
| 709 | /* IB memory */ |
| 710 | struct kgd_mem *ib_mem; |
| 711 | uint64_t ib_base; |
| 712 | void *ib_kaddr; |
| 713 | |
| 714 | /* doorbells for kfd process */ |
| 715 | struct amdgpu_bo *proc_doorbells; |
| 716 | |
| 717 | /* bitmap for dynamic doorbell allocation from the bo */ |
| 718 | unsigned long *doorbell_bitmap; |
| 719 | }; |
| 720 | |
| 721 | /* KFD Memory Eviction */ |
| 722 | |
| 723 | /* Approx. wait time before attempting to restore evicted BOs */ |
| 724 | #define PROCESS_RESTORE_TIME_MS 100 |
| 725 | /* Approx. back off time if restore fails due to lack of memory */ |
| 726 | #define PROCESS_BACK_OFF_TIME_MS 100 |
| 727 | /* Approx. time before evicting the process again */ |
| 728 | #define PROCESS_ACTIVE_TIME_MS 10 |
| 729 | |
| 730 | /* 8 byte handle containing GPU ID in the most significant 4 bytes and |
| 731 | * idr_handle in the least significant 4 bytes |
| 732 | */ |
| 733 | #define MAKE_HANDLE(gpu_id, idr_handle) \ |
| 734 | (((uint64_t)(gpu_id) << 32) + idr_handle) |
| 735 | #define GET_GPU_ID(handle) (handle >> 32) |
| 736 | #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) |
| 737 | |
| 738 | enum kfd_pdd_bound { |
| 739 | PDD_UNBOUND = 0, |
| 740 | PDD_BOUND, |
| 741 | PDD_BOUND_SUSPENDED, |
| 742 | }; |
| 743 | |
| 744 | #define MAX_SYSFS_FILENAME_LEN 15 |
| 745 | |
| 746 | /* |
| 747 | * SDMA counter runs at 100MHz frequency. |
| 748 | * We display SDMA activity in microsecond granularity in sysfs. |
| 749 | * As a result, the divisor is 100. |
| 750 | */ |
| 751 | #define SDMA_ACTIVITY_DIVISOR 100 |
| 752 | |
| 753 | /* Data that is per-process-per device. */ |
| 754 | struct kfd_process_device { |
| 755 | /* The device that owns this data. */ |
| 756 | struct kfd_node *dev; |
| 757 | |
| 758 | /* The process that owns this kfd_process_device. */ |
| 759 | struct kfd_process *process; |
| 760 | |
| 761 | /* per-process-per device QCM data structure */ |
| 762 | struct qcm_process_device qpd; |
| 763 | |
| 764 | /*Apertures*/ |
| 765 | uint64_t lds_base; |
| 766 | uint64_t lds_limit; |
| 767 | uint64_t gpuvm_base; |
| 768 | uint64_t gpuvm_limit; |
| 769 | uint64_t scratch_base; |
| 770 | uint64_t scratch_limit; |
| 771 | |
| 772 | /* VM context for GPUVM allocations */ |
| 773 | struct file *drm_file; |
| 774 | void *drm_priv; |
| 775 | |
| 776 | /* GPUVM allocations storage */ |
| 777 | struct idr alloc_idr; |
| 778 | |
| 779 | /* Flag used to tell the pdd has dequeued from the dqm. |
| 780 | * This is used to prevent dev->dqm->ops.process_termination() from |
| 781 | * being called twice when it is already called in IOMMU callback |
| 782 | * function. |
| 783 | */ |
| 784 | bool already_dequeued; |
| 785 | bool runtime_inuse; |
| 786 | |
| 787 | /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ |
| 788 | enum kfd_pdd_bound bound; |
| 789 | |
| 790 | /* VRAM usage */ |
| 791 | atomic64_t vram_usage; |
| 792 | struct attribute attr_vram; |
| 793 | char vram_filename[MAX_SYSFS_FILENAME_LEN]; |
| 794 | |
| 795 | /* SDMA activity tracking */ |
| 796 | uint64_t sdma_past_activity_counter; |
| 797 | struct attribute attr_sdma; |
| 798 | char sdma_filename[MAX_SYSFS_FILENAME_LEN]; |
| 799 | |
| 800 | /* Eviction activity tracking */ |
| 801 | uint64_t last_evict_timestamp; |
| 802 | atomic64_t evict_duration_counter; |
| 803 | struct attribute attr_evict; |
| 804 | |
| 805 | struct kobject *kobj_stats; |
| 806 | |
| 807 | /* |
| 808 | * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process |
| 809 | * that is associated with device encoded by "this" struct instance. The |
| 810 | * value reflects CU usage by all of the waves launched by this process |
| 811 | * on this device. A very important property of occupancy parameter is |
| 812 | * that its value is a snapshot of current use. |
| 813 | * |
| 814 | * Following is to be noted regarding how this parameter is reported: |
| 815 | * |
| 816 | * The number of waves that a CU can launch is limited by couple of |
| 817 | * parameters. These are encoded by struct amdgpu_cu_info instance |
| 818 | * that is part of every device definition. For GFX9 devices this |
| 819 | * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves |
| 820 | * do not use scratch memory and 32 waves (max_scratch_slots_per_cu) |
| 821 | * when they do use scratch memory. This could change for future |
| 822 | * devices and therefore this example should be considered as a guide. |
| 823 | * |
| 824 | * All CU's of a device are available for the process. This may not be true |
| 825 | * under certain conditions - e.g. CU masking. |
| 826 | * |
| 827 | * Finally number of CU's that are occupied by a process is affected by both |
| 828 | * number of CU's a device has along with number of other competing processes |
| 829 | */ |
| 830 | struct attribute attr_cu_occupancy; |
| 831 | |
| 832 | /* sysfs counters for GPU retry fault and page migration tracking */ |
| 833 | struct kobject *kobj_counters; |
| 834 | struct attribute attr_faults; |
| 835 | struct attribute attr_page_in; |
| 836 | struct attribute attr_page_out; |
| 837 | uint64_t faults; |
| 838 | uint64_t page_in; |
| 839 | uint64_t page_out; |
| 840 | |
| 841 | /* Exception code status*/ |
| 842 | uint64_t exception_status; |
| 843 | void *vm_fault_exc_data; |
| 844 | size_t vm_fault_exc_data_size; |
| 845 | |
| 846 | /* Tracks debug per-vmid request settings */ |
| 847 | uint32_t spi_dbg_override; |
| 848 | uint32_t spi_dbg_launch_mode; |
| 849 | uint32_t watch_points[4]; |
| 850 | uint32_t alloc_watch_ids; |
| 851 | |
| 852 | /* |
| 853 | * If this process has been checkpointed before, then the user |
| 854 | * application will use the original gpu_id on the |
| 855 | * checkpointed node to refer to this device. |
| 856 | */ |
| 857 | uint32_t user_gpu_id; |
| 858 | |
| 859 | void *proc_ctx_bo; |
| 860 | uint64_t proc_ctx_gpu_addr; |
| 861 | void *proc_ctx_cpu_ptr; |
| 862 | |
| 863 | /* Tracks queue reset status */ |
| 864 | bool has_reset_queue; |
| 865 | |
| 866 | u32 pasid; |
| 867 | }; |
| 868 | |
| 869 | #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) |
| 870 | |
| 871 | struct svm_range_list { |
| 872 | struct mutex lock; |
| 873 | struct rb_root_cached objects; |
| 874 | struct list_head list; |
| 875 | struct work_struct deferred_list_work; |
| 876 | struct list_head deferred_range_list; |
| 877 | struct list_head criu_svm_metadata_list; |
| 878 | spinlock_t deferred_list_lock; |
| 879 | atomic_t evicted_ranges; |
| 880 | atomic_t drain_pagefaults; |
| 881 | struct delayed_work restore_work; |
| 882 | DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE); |
| 883 | struct task_struct *faulting_task; |
| 884 | /* check point ts decides if page fault recovery need be dropped */ |
| 885 | uint64_t checkpoint_ts[MAX_GPU_INSTANCE]; |
| 886 | |
| 887 | /* Default granularity to use in buffer migration |
| 888 | * and restoration of backing memory while handling |
| 889 | * recoverable page faults |
| 890 | */ |
| 891 | uint8_t default_granularity; |
| 892 | }; |
| 893 | |
| 894 | /* Process data */ |
| 895 | struct kfd_process { |
| 896 | /* |
| 897 | * kfd_process are stored in an mm_struct*->kfd_process* |
| 898 | * hash table (kfd_processes in kfd_process.c) |
| 899 | */ |
| 900 | struct hlist_node kfd_processes; |
| 901 | |
| 902 | /* |
| 903 | * Opaque pointer to mm_struct. We don't hold a reference to |
| 904 | * it so it should never be dereferenced from here. This is |
| 905 | * only used for looking up processes by their mm. |
| 906 | */ |
| 907 | void *mm; |
| 908 | |
| 909 | struct kref ref; |
| 910 | struct work_struct release_work; |
| 911 | |
| 912 | struct mutex mutex; |
| 913 | |
| 914 | /* |
| 915 | * In any process, the thread that started main() is the lead |
| 916 | * thread and outlives the rest. |
| 917 | * It is here because amd_iommu_bind_pasid wants a task_struct. |
| 918 | * It can also be used for safely getting a reference to the |
| 919 | * mm_struct of the process. |
| 920 | */ |
| 921 | struct task_struct *lead_thread; |
| 922 | |
| 923 | /* We want to receive a notification when the mm_struct is destroyed */ |
| 924 | struct mmu_notifier mmu_notifier; |
| 925 | |
| 926 | /* |
| 927 | * Array of kfd_process_device pointers, |
| 928 | * one for each device the process is using. |
| 929 | */ |
| 930 | struct kfd_process_device *pdds[MAX_GPU_INSTANCE]; |
| 931 | uint32_t n_pdds; |
| 932 | |
| 933 | struct process_queue_manager pqm; |
| 934 | |
| 935 | /*Is the user space process 32 bit?*/ |
| 936 | bool is_32bit_user_mode; |
| 937 | |
| 938 | /* Event-related data */ |
| 939 | struct mutex event_mutex; |
| 940 | /* Event ID allocator and lookup */ |
| 941 | struct idr event_idr; |
| 942 | /* Event page */ |
| 943 | u64 signal_handle; |
| 944 | struct kfd_signal_page *signal_page; |
| 945 | size_t signal_mapped_size; |
| 946 | size_t signal_event_count; |
| 947 | bool signal_event_limit_reached; |
| 948 | |
| 949 | /* Information used for memory eviction */ |
| 950 | void *kgd_process_info; |
| 951 | /* Eviction fence that is attached to all the BOs of this process. The |
| 952 | * fence will be triggered during eviction and new one will be created |
| 953 | * during restore |
| 954 | */ |
| 955 | struct dma_fence __rcu *ef; |
| 956 | |
| 957 | /* Work items for evicting and restoring BOs */ |
| 958 | struct delayed_work eviction_work; |
| 959 | struct delayed_work restore_work; |
| 960 | /* seqno of the last scheduled eviction */ |
| 961 | unsigned int last_eviction_seqno; |
| 962 | /* Approx. the last timestamp (in jiffies) when the process was |
| 963 | * restored after an eviction |
| 964 | */ |
| 965 | unsigned long last_restore_timestamp; |
| 966 | |
| 967 | /* Indicates device process is debug attached with reserved vmid. */ |
| 968 | bool debug_trap_enabled; |
| 969 | |
| 970 | /* per-process-per device debug event fd file */ |
| 971 | struct file *dbg_ev_file; |
| 972 | |
| 973 | /* If the process is a kfd debugger, we need to know so we can clean |
| 974 | * up at exit time. If a process enables debugging on itself, it does |
| 975 | * its own clean-up, so we don't set the flag here. We track this by |
| 976 | * counting the number of processes this process is debugging. |
| 977 | */ |
| 978 | atomic_t debugged_process_count; |
| 979 | |
| 980 | /* If the process is a debugged, this is the debugger process */ |
| 981 | struct kfd_process *debugger_process; |
| 982 | |
| 983 | /* Kobj for our procfs */ |
| 984 | struct kobject *kobj; |
| 985 | struct kobject *kobj_queues; |
| 986 | struct attribute attr_pasid; |
| 987 | |
| 988 | /* Keep track cwsr init */ |
| 989 | bool has_cwsr; |
| 990 | |
| 991 | /* Exception code enable mask and status */ |
| 992 | uint64_t exception_enable_mask; |
| 993 | uint64_t exception_status; |
| 994 | |
| 995 | /* Used to drain stale interrupts */ |
| 996 | wait_queue_head_t wait_irq_drain; |
| 997 | bool irq_drain_is_open; |
| 998 | |
| 999 | /* shared virtual memory registered by this process */ |
| 1000 | struct svm_range_list svms; |
| 1001 | |
| 1002 | bool xnack_enabled; |
| 1003 | |
| 1004 | /* Work area for debugger event writer worker. */ |
| 1005 | struct work_struct debug_event_workarea; |
| 1006 | |
| 1007 | /* Tracks debug per-vmid request for debug flags */ |
| 1008 | u32 dbg_flags; |
| 1009 | |
| 1010 | atomic_t poison; |
| 1011 | /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ |
| 1012 | bool queues_paused; |
| 1013 | |
| 1014 | /* Tracks runtime enable status */ |
| 1015 | struct semaphore runtime_enable_sema; |
| 1016 | bool is_runtime_retry; |
| 1017 | struct kfd_runtime_info runtime_info; |
| 1018 | |
| 1019 | /* if gpu page fault sent to KFD */ |
| 1020 | bool gpu_page_fault; |
| 1021 | }; |
| 1022 | |
| 1023 | #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ |
| 1024 | extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); |
| 1025 | extern struct srcu_struct kfd_processes_srcu; |
| 1026 | |
| 1027 | /** |
| 1028 | * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. |
| 1029 | * |
| 1030 | * @filep: pointer to file structure. |
| 1031 | * @p: amdkfd process pointer. |
| 1032 | * @data: pointer to arg that was copied from user. |
| 1033 | * |
| 1034 | * Return: returns ioctl completion code. |
| 1035 | */ |
| 1036 | typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, |
| 1037 | void *data); |
| 1038 | |
| 1039 | struct amdkfd_ioctl_desc { |
| 1040 | unsigned int cmd; |
| 1041 | int flags; |
| 1042 | amdkfd_ioctl_t *func; |
| 1043 | unsigned int cmd_drv; |
| 1044 | const char *name; |
| 1045 | }; |
| 1046 | bool kfd_dev_is_large_bar(struct kfd_node *dev); |
| 1047 | |
| 1048 | int kfd_process_create_wq(void); |
| 1049 | void kfd_process_destroy_wq(void); |
| 1050 | void kfd_cleanup_processes(void); |
| 1051 | struct kfd_process *kfd_create_process(struct task_struct *thread); |
| 1052 | struct kfd_process *kfd_get_process(const struct task_struct *task); |
| 1053 | struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid, |
| 1054 | struct kfd_process_device **pdd); |
| 1055 | struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); |
| 1056 | |
| 1057 | int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id); |
| 1058 | int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node, |
| 1059 | uint32_t *gpuid, uint32_t *gpuidx); |
| 1060 | static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p, |
| 1061 | uint32_t gpuidx, uint32_t *gpuid) { |
| 1062 | return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL; |
| 1063 | } |
| 1064 | static inline struct kfd_process_device *kfd_process_device_from_gpuidx( |
| 1065 | struct kfd_process *p, uint32_t gpuidx) { |
| 1066 | return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL; |
| 1067 | } |
| 1068 | |
| 1069 | void kfd_unref_process(struct kfd_process *p); |
| 1070 | int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger); |
| 1071 | int kfd_process_restore_queues(struct kfd_process *p); |
| 1072 | void kfd_suspend_all_processes(void); |
| 1073 | int kfd_resume_all_processes(void); |
| 1074 | |
| 1075 | struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process, |
| 1076 | uint32_t gpu_id); |
| 1077 | |
| 1078 | int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id); |
| 1079 | |
| 1080 | int kfd_process_device_init_vm(struct kfd_process_device *pdd, |
| 1081 | struct file *drm_file); |
| 1082 | struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev, |
| 1083 | struct kfd_process *p); |
| 1084 | struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev, |
| 1085 | struct kfd_process *p); |
| 1086 | struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, |
| 1087 | struct kfd_process *p); |
| 1088 | |
| 1089 | bool kfd_process_xnack_mode(struct kfd_process *p, bool supported); |
| 1090 | |
| 1091 | int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process, |
| 1092 | struct vm_area_struct *vma); |
| 1093 | |
| 1094 | /* KFD process API for creating and translating handles */ |
| 1095 | int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, |
| 1096 | void *mem); |
| 1097 | void *kfd_process_device_translate_handle(struct kfd_process_device *p, |
| 1098 | int handle); |
| 1099 | void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, |
| 1100 | int handle); |
| 1101 | struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid); |
| 1102 | |
| 1103 | /* PASIDs */ |
| 1104 | int kfd_pasid_init(void); |
| 1105 | void kfd_pasid_exit(void); |
| 1106 | u32 kfd_pasid_alloc(void); |
| 1107 | void kfd_pasid_free(u32 pasid); |
| 1108 | |
| 1109 | /* Doorbells */ |
| 1110 | size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); |
| 1111 | int kfd_doorbell_init(struct kfd_dev *kfd); |
| 1112 | void kfd_doorbell_fini(struct kfd_dev *kfd); |
| 1113 | int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process, |
| 1114 | struct vm_area_struct *vma); |
| 1115 | void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, |
| 1116 | unsigned int *doorbell_off); |
| 1117 | void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); |
| 1118 | u32 read_kernel_doorbell(u32 __iomem *db); |
| 1119 | void write_kernel_doorbell(void __iomem *db, u32 value); |
| 1120 | void write_kernel_doorbell64(void __iomem *db, u64 value); |
| 1121 | unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, |
| 1122 | struct kfd_process_device *pdd, |
| 1123 | unsigned int doorbell_id); |
| 1124 | phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); |
| 1125 | int kfd_alloc_process_doorbells(struct kfd_dev *kfd, |
| 1126 | struct kfd_process_device *pdd); |
| 1127 | void kfd_free_process_doorbells(struct kfd_dev *kfd, |
| 1128 | struct kfd_process_device *pdd); |
| 1129 | /* GTT Sub-Allocator */ |
| 1130 | |
| 1131 | int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size, |
| 1132 | struct kfd_mem_obj **mem_obj); |
| 1133 | |
| 1134 | int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj); |
| 1135 | |
| 1136 | extern struct device *kfd_device; |
| 1137 | |
| 1138 | /* KFD's procfs */ |
| 1139 | void kfd_procfs_init(void); |
| 1140 | void kfd_procfs_shutdown(void); |
| 1141 | int kfd_procfs_add_queue(struct queue *q); |
| 1142 | void kfd_procfs_del_queue(struct queue *q); |
| 1143 | |
| 1144 | /* Topology */ |
| 1145 | int kfd_topology_init(void); |
| 1146 | void kfd_topology_shutdown(void); |
| 1147 | int kfd_topology_add_device(struct kfd_node *gpu); |
| 1148 | int kfd_topology_remove_device(struct kfd_node *gpu); |
| 1149 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain( |
| 1150 | uint32_t proximity_domain); |
| 1151 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( |
| 1152 | uint32_t proximity_domain); |
| 1153 | struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); |
| 1154 | struct kfd_node *kfd_device_by_id(uint32_t gpu_id); |
| 1155 | static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id, |
| 1156 | uint32_t vmid) |
| 1157 | { |
| 1158 | return (node->interrupt_bitmap & (1 << node_id)) != 0 && |
| 1159 | (node->compute_vmid_bitmap & (1 << vmid)) != 0; |
| 1160 | } |
| 1161 | static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev, |
| 1162 | uint32_t node_id, uint32_t vmid) { |
| 1163 | struct kfd_dev *dev = adev->kfd.dev; |
| 1164 | uint32_t i; |
| 1165 | |
| 1166 | if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) && |
| 1167 | KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4) && |
| 1168 | KFD_GC_VERSION(dev) != IP_VERSION(9, 5, 0)) |
| 1169 | return dev->nodes[0]; |
| 1170 | |
| 1171 | for (i = 0; i < dev->num_nodes; i++) |
| 1172 | if (kfd_irq_is_from_node(node: dev->nodes[i], node_id, vmid)) |
| 1173 | return dev->nodes[i]; |
| 1174 | |
| 1175 | return NULL; |
| 1176 | } |
| 1177 | int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev); |
| 1178 | int kfd_numa_node_to_apic_id(int numa_node_id); |
| 1179 | |
| 1180 | /* Interrupts */ |
| 1181 | #define KFD_IRQ_FENCE_CLIENTID 0xff |
| 1182 | #define KFD_IRQ_FENCE_SOURCEID 0xff |
| 1183 | #define KFD_IRQ_IS_FENCE(client, source) \ |
| 1184 | ((client) == KFD_IRQ_FENCE_CLIENTID && \ |
| 1185 | (source) == KFD_IRQ_FENCE_SOURCEID) |
| 1186 | int kfd_interrupt_init(struct kfd_node *dev); |
| 1187 | void kfd_interrupt_exit(struct kfd_node *dev); |
| 1188 | bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry); |
| 1189 | bool interrupt_is_wanted(struct kfd_node *dev, |
| 1190 | const uint32_t *ih_ring_entry, |
| 1191 | uint32_t *patched_ihre, bool *flag); |
| 1192 | int kfd_process_drain_interrupts(struct kfd_process_device *pdd); |
| 1193 | void kfd_process_close_interrupt_drain(unsigned int pasid); |
| 1194 | |
| 1195 | /* amdkfd Apertures */ |
| 1196 | int kfd_init_apertures(struct kfd_process *process); |
| 1197 | |
| 1198 | void kfd_process_set_trap_handler(struct qcm_process_device *qpd, |
| 1199 | uint64_t tba_addr, |
| 1200 | uint64_t tma_addr); |
| 1201 | void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd, |
| 1202 | bool enabled); |
| 1203 | |
| 1204 | /* CWSR initialization */ |
| 1205 | int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep); |
| 1206 | |
| 1207 | /* CRIU */ |
| 1208 | /* |
| 1209 | * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private |
| 1210 | * structures: |
| 1211 | * kfd_criu_process_priv_data |
| 1212 | * kfd_criu_device_priv_data |
| 1213 | * kfd_criu_bo_priv_data |
| 1214 | * kfd_criu_queue_priv_data |
| 1215 | * kfd_criu_event_priv_data |
| 1216 | * kfd_criu_svm_range_priv_data |
| 1217 | */ |
| 1218 | |
| 1219 | #define KFD_CRIU_PRIV_VERSION 1 |
| 1220 | |
| 1221 | struct kfd_criu_process_priv_data { |
| 1222 | uint32_t version; |
| 1223 | uint32_t xnack_mode; |
| 1224 | }; |
| 1225 | |
| 1226 | struct kfd_criu_device_priv_data { |
| 1227 | /* For future use */ |
| 1228 | uint64_t reserved; |
| 1229 | }; |
| 1230 | |
| 1231 | struct kfd_criu_bo_priv_data { |
| 1232 | uint64_t user_addr; |
| 1233 | uint32_t idr_handle; |
| 1234 | uint32_t mapped_gpuids[MAX_GPU_INSTANCE]; |
| 1235 | }; |
| 1236 | |
| 1237 | /* |
| 1238 | * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data, |
| 1239 | * kfd_criu_svm_range_priv_data is the object type |
| 1240 | */ |
| 1241 | enum kfd_criu_object_type { |
| 1242 | KFD_CRIU_OBJECT_TYPE_QUEUE, |
| 1243 | KFD_CRIU_OBJECT_TYPE_EVENT, |
| 1244 | KFD_CRIU_OBJECT_TYPE_SVM_RANGE, |
| 1245 | }; |
| 1246 | |
| 1247 | struct kfd_criu_svm_range_priv_data { |
| 1248 | uint32_t object_type; |
| 1249 | uint64_t start_addr; |
| 1250 | uint64_t size; |
| 1251 | /* Variable length array of attributes */ |
| 1252 | struct kfd_ioctl_svm_attribute attrs[]; |
| 1253 | }; |
| 1254 | |
| 1255 | struct kfd_criu_queue_priv_data { |
| 1256 | uint32_t object_type; |
| 1257 | uint64_t q_address; |
| 1258 | uint64_t q_size; |
| 1259 | uint64_t read_ptr_addr; |
| 1260 | uint64_t write_ptr_addr; |
| 1261 | uint64_t doorbell_off; |
| 1262 | uint64_t eop_ring_buffer_address; |
| 1263 | uint64_t ctx_save_restore_area_address; |
| 1264 | uint32_t gpu_id; |
| 1265 | uint32_t type; |
| 1266 | uint32_t format; |
| 1267 | uint32_t q_id; |
| 1268 | uint32_t priority; |
| 1269 | uint32_t q_percent; |
| 1270 | uint32_t doorbell_id; |
| 1271 | uint32_t gws; |
| 1272 | uint32_t sdma_id; |
| 1273 | uint32_t eop_ring_buffer_size; |
| 1274 | uint32_t ctx_save_restore_area_size; |
| 1275 | uint32_t ctl_stack_size; |
| 1276 | uint32_t mqd_size; |
| 1277 | }; |
| 1278 | |
| 1279 | struct kfd_criu_event_priv_data { |
| 1280 | uint32_t object_type; |
| 1281 | uint64_t user_handle; |
| 1282 | uint32_t event_id; |
| 1283 | uint32_t auto_reset; |
| 1284 | uint32_t type; |
| 1285 | uint32_t signaled; |
| 1286 | |
| 1287 | union { |
| 1288 | struct kfd_hsa_memory_exception_data memory_exception_data; |
| 1289 | struct kfd_hsa_hw_exception_data hw_exception_data; |
| 1290 | }; |
| 1291 | }; |
| 1292 | |
| 1293 | int kfd_process_get_queue_info(struct kfd_process *p, |
| 1294 | uint32_t *num_queues, |
| 1295 | uint64_t *priv_data_sizes); |
| 1296 | |
| 1297 | int kfd_criu_checkpoint_queues(struct kfd_process *p, |
| 1298 | uint8_t __user *user_priv_data, |
| 1299 | uint64_t *priv_data_offset); |
| 1300 | |
| 1301 | int kfd_criu_restore_queue(struct kfd_process *p, |
| 1302 | uint8_t __user *user_priv_data, |
| 1303 | uint64_t *priv_data_offset, |
| 1304 | uint64_t max_priv_data_size); |
| 1305 | |
| 1306 | int kfd_criu_checkpoint_events(struct kfd_process *p, |
| 1307 | uint8_t __user *user_priv_data, |
| 1308 | uint64_t *priv_data_offset); |
| 1309 | |
| 1310 | int kfd_criu_restore_event(struct file *devkfd, |
| 1311 | struct kfd_process *p, |
| 1312 | uint8_t __user *user_priv_data, |
| 1313 | uint64_t *priv_data_offset, |
| 1314 | uint64_t max_priv_data_size); |
| 1315 | /* CRIU - End */ |
| 1316 | |
| 1317 | /* Queue Context Management */ |
| 1318 | int init_queue(struct queue **q, const struct queue_properties *properties); |
| 1319 | void uninit_queue(struct queue *q); |
| 1320 | void print_queue_properties(struct queue_properties *q); |
| 1321 | void print_queue(struct queue *q); |
| 1322 | int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo, |
| 1323 | u64 expected_size); |
| 1324 | void kfd_queue_buffer_put(struct amdgpu_bo **bo); |
| 1325 | int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties); |
| 1326 | int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties); |
| 1327 | void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo); |
| 1328 | int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd, |
| 1329 | struct queue_properties *properties); |
| 1330 | void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev); |
| 1331 | |
| 1332 | struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, |
| 1333 | struct kfd_node *dev); |
| 1334 | struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, |
| 1335 | struct kfd_node *dev); |
| 1336 | struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, |
| 1337 | struct kfd_node *dev); |
| 1338 | struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, |
| 1339 | struct kfd_node *dev); |
| 1340 | struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, |
| 1341 | struct kfd_node *dev); |
| 1342 | struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type, |
| 1343 | struct kfd_node *dev); |
| 1344 | struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev); |
| 1345 | void device_queue_manager_uninit(struct device_queue_manager *dqm); |
| 1346 | struct kernel_queue *kernel_queue_init(struct kfd_node *dev, |
| 1347 | enum kfd_queue_type type); |
| 1348 | void kernel_queue_uninit(struct kernel_queue *kq); |
| 1349 | int kfd_evict_process_device(struct kfd_process_device *pdd); |
| 1350 | int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id); |
| 1351 | |
| 1352 | /* Process Queue Manager */ |
| 1353 | struct process_queue_node { |
| 1354 | struct queue *q; |
| 1355 | struct kernel_queue *kq; |
| 1356 | struct list_head process_queue_list; |
| 1357 | }; |
| 1358 | |
| 1359 | void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); |
| 1360 | void kfd_process_dequeue_from_all_devices(struct kfd_process *p); |
| 1361 | int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); |
| 1362 | void pqm_uninit(struct process_queue_manager *pqm); |
| 1363 | int pqm_create_queue(struct process_queue_manager *pqm, |
| 1364 | struct kfd_node *dev, |
| 1365 | struct queue_properties *properties, |
| 1366 | unsigned int *qid, |
| 1367 | const struct kfd_criu_queue_priv_data *q_data, |
| 1368 | const void *restore_mqd, |
| 1369 | const void *restore_ctl_stack, |
| 1370 | uint32_t *p_doorbell_offset_in_process); |
| 1371 | int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); |
| 1372 | int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid, |
| 1373 | struct queue_properties *p); |
| 1374 | int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid, |
| 1375 | struct mqd_update_info *minfo); |
| 1376 | int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, |
| 1377 | void *gws); |
| 1378 | struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, |
| 1379 | unsigned int qid); |
| 1380 | int pqm_get_wave_state(struct process_queue_manager *pqm, |
| 1381 | unsigned int qid, |
| 1382 | void __user *ctl_stack, |
| 1383 | u32 *ctl_stack_used_size, |
| 1384 | u32 *save_area_used_size); |
| 1385 | int pqm_get_queue_snapshot(struct process_queue_manager *pqm, |
| 1386 | uint64_t exception_clear_mask, |
| 1387 | void __user *buf, |
| 1388 | int *num_qss_entries, |
| 1389 | uint32_t *entry_size); |
| 1390 | |
| 1391 | int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm, |
| 1392 | uint64_t fence_value, |
| 1393 | unsigned int timeout_ms); |
| 1394 | |
| 1395 | int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm, |
| 1396 | unsigned int qid, |
| 1397 | u32 *mqd_size, |
| 1398 | u32 *ctl_stack_size); |
| 1399 | /* Packet Manager */ |
| 1400 | |
| 1401 | #define KFD_FENCE_COMPLETED (100) |
| 1402 | #define KFD_FENCE_INIT (10) |
| 1403 | |
| 1404 | /** |
| 1405 | * enum kfd_config_dequeue_wait_counts_cmd - Command for configuring |
| 1406 | * dequeue wait counts. |
| 1407 | * |
| 1408 | * @KFD_DEQUEUE_WAIT_INIT: Set optimized dequeue wait counts for a |
| 1409 | * certain ASICs. For these ASICs, this is default value used by RESET |
| 1410 | * @KFD_DEQUEUE_WAIT_RESET: Reset dequeue wait counts to the optimized value |
| 1411 | * for certain ASICs. For others set it to default hardware reset value |
| 1412 | * @KFD_DEQUEUE_WAIT_SET_SCH_WAVE: Set context switch latency wait |
| 1413 | * |
| 1414 | */ |
| 1415 | enum kfd_config_dequeue_wait_counts_cmd { |
| 1416 | KFD_DEQUEUE_WAIT_INIT = 1, |
| 1417 | KFD_DEQUEUE_WAIT_RESET = 2, |
| 1418 | KFD_DEQUEUE_WAIT_SET_SCH_WAVE = 3 |
| 1419 | }; |
| 1420 | |
| 1421 | |
| 1422 | struct packet_manager { |
| 1423 | struct device_queue_manager *dqm; |
| 1424 | struct kernel_queue *priv_queue; |
| 1425 | struct mutex lock; |
| 1426 | bool allocated; |
| 1427 | struct kfd_mem_obj *ib_buffer_obj; |
| 1428 | unsigned int ib_size_bytes; |
| 1429 | bool is_over_subscription; |
| 1430 | |
| 1431 | const struct packet_manager_funcs *pmf; |
| 1432 | }; |
| 1433 | |
| 1434 | struct packet_manager_funcs { |
| 1435 | /* Support ASIC-specific packet formats for PM4 packets */ |
| 1436 | int (*map_process)(struct packet_manager *pm, uint32_t *buffer, |
| 1437 | struct qcm_process_device *qpd); |
| 1438 | int (*runlist)(struct packet_manager *pm, uint32_t *buffer, |
| 1439 | uint64_t ib, size_t ib_size_in_dwords, bool chain); |
| 1440 | int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, |
| 1441 | struct scheduling_resources *res); |
| 1442 | int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, |
| 1443 | struct queue *q, bool is_static); |
| 1444 | int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, |
| 1445 | enum kfd_unmap_queues_filter mode, |
| 1446 | uint32_t filter_param, bool reset); |
| 1447 | int (*config_dequeue_wait_counts)(struct packet_manager *pm, uint32_t *buffer, |
| 1448 | enum kfd_config_dequeue_wait_counts_cmd cmd, uint32_t value); |
| 1449 | int (*query_status)(struct packet_manager *pm, uint32_t *buffer, |
| 1450 | uint64_t fence_address, uint64_t fence_value); |
| 1451 | int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); |
| 1452 | |
| 1453 | /* Packet sizes */ |
| 1454 | int map_process_size; |
| 1455 | int runlist_size; |
| 1456 | int set_resources_size; |
| 1457 | int map_queues_size; |
| 1458 | int unmap_queues_size; |
| 1459 | int config_dequeue_wait_counts_size; |
| 1460 | int query_status_size; |
| 1461 | int release_mem_size; |
| 1462 | }; |
| 1463 | |
| 1464 | extern const struct packet_manager_funcs kfd_vi_pm_funcs; |
| 1465 | extern const struct packet_manager_funcs kfd_v9_pm_funcs; |
| 1466 | extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs; |
| 1467 | |
| 1468 | int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); |
| 1469 | void pm_uninit(struct packet_manager *pm); |
| 1470 | int pm_send_set_resources(struct packet_manager *pm, |
| 1471 | struct scheduling_resources *res); |
| 1472 | int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); |
| 1473 | int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, |
| 1474 | uint64_t fence_value); |
| 1475 | |
| 1476 | int pm_send_unmap_queue(struct packet_manager *pm, |
| 1477 | enum kfd_unmap_queues_filter mode, |
| 1478 | uint32_t filter_param, bool reset); |
| 1479 | |
| 1480 | void pm_release_ib(struct packet_manager *pm); |
| 1481 | |
| 1482 | int pm_config_dequeue_wait_counts(struct packet_manager *pm, |
| 1483 | enum kfd_config_dequeue_wait_counts_cmd cmd, |
| 1484 | uint32_t wait_counts_config); |
| 1485 | |
| 1486 | /* Following PM funcs can be shared among VI and AI */ |
| 1487 | unsigned int (unsigned int opcode, size_t packet_size); |
| 1488 | |
| 1489 | uint64_t kfd_get_number_elems(struct kfd_dev *kfd); |
| 1490 | |
| 1491 | /* Events */ |
| 1492 | extern const struct kfd_event_interrupt_class event_interrupt_class_cik; |
| 1493 | extern const struct kfd_event_interrupt_class event_interrupt_class_v9; |
| 1494 | extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3; |
| 1495 | extern const struct kfd_event_interrupt_class event_interrupt_class_v10; |
| 1496 | extern const struct kfd_event_interrupt_class event_interrupt_class_v11; |
| 1497 | |
| 1498 | extern const struct kfd_device_global_init_class device_global_init_class_cik; |
| 1499 | |
| 1500 | int kfd_event_init_process(struct kfd_process *p); |
| 1501 | void kfd_event_free_process(struct kfd_process *p); |
| 1502 | int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); |
| 1503 | int kfd_wait_on_events(struct kfd_process *p, |
| 1504 | uint32_t num_events, void __user *data, |
| 1505 | bool all, uint32_t *user_timeout_ms, |
| 1506 | uint32_t *wait_result); |
| 1507 | void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, |
| 1508 | uint32_t valid_id_bits); |
| 1509 | void kfd_signal_hw_exception_event(u32 pasid); |
| 1510 | int kfd_set_event(struct kfd_process *p, uint32_t event_id); |
| 1511 | int kfd_reset_event(struct kfd_process *p, uint32_t event_id); |
| 1512 | int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset); |
| 1513 | |
| 1514 | int kfd_event_create(struct file *devkfd, struct kfd_process *p, |
| 1515 | uint32_t event_type, bool auto_reset, uint32_t node_id, |
| 1516 | uint32_t *event_id, uint32_t *event_trigger_data, |
| 1517 | uint64_t *event_page_offset, uint32_t *event_slot_index); |
| 1518 | |
| 1519 | int kfd_get_num_events(struct kfd_process *p); |
| 1520 | int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); |
| 1521 | |
| 1522 | void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va); |
| 1523 | |
| 1524 | void kfd_signal_vm_fault_event(struct kfd_process_device *pdd, |
| 1525 | struct kfd_vm_fault_info *info, |
| 1526 | struct kfd_hsa_memory_exception_data *data); |
| 1527 | |
| 1528 | void kfd_signal_reset_event(struct kfd_node *dev); |
| 1529 | |
| 1530 | void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid); |
| 1531 | |
| 1532 | static inline void kfd_flush_tlb(struct kfd_process_device *pdd, |
| 1533 | enum TLB_FLUSH_TYPE type) |
| 1534 | { |
| 1535 | struct amdgpu_device *adev = pdd->dev->adev; |
| 1536 | struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); |
| 1537 | |
| 1538 | amdgpu_vm_flush_compute_tlb(adev, vm, flush_type: type, xcc_mask: pdd->dev->xcc_mask); |
| 1539 | } |
| 1540 | |
| 1541 | static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) |
| 1542 | { |
| 1543 | return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) || |
| 1544 | (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) || |
| 1545 | KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); |
| 1546 | } |
| 1547 | |
| 1548 | int kfd_send_exception_to_runtime(struct kfd_process *p, |
| 1549 | unsigned int queue_id, |
| 1550 | uint64_t error_reason); |
| 1551 | bool kfd_is_locked(struct kfd_dev *kfd); |
| 1552 | |
| 1553 | /* Compute profile */ |
| 1554 | void kfd_inc_compute_active(struct kfd_node *dev); |
| 1555 | void kfd_dec_compute_active(struct kfd_node *dev); |
| 1556 | |
| 1557 | /* Cgroup Support */ |
| 1558 | /* Check with device cgroup if @kfd device is accessible */ |
| 1559 | static inline int kfd_devcgroup_check_permission(struct kfd_node *node) |
| 1560 | { |
| 1561 | #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) |
| 1562 | struct drm_device *ddev; |
| 1563 | |
| 1564 | if (node->xcp) |
| 1565 | ddev = node->xcp->ddev; |
| 1566 | else |
| 1567 | ddev = adev_to_drm(adev: node->adev); |
| 1568 | |
| 1569 | return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, |
| 1570 | minor: ddev->render->index, |
| 1571 | DEVCG_ACC_WRITE | DEVCG_ACC_READ); |
| 1572 | #else |
| 1573 | return 0; |
| 1574 | #endif |
| 1575 | } |
| 1576 | |
| 1577 | static inline bool kfd_is_first_node(struct kfd_node *node) |
| 1578 | { |
| 1579 | return (node == node->kfd->nodes[0]); |
| 1580 | } |
| 1581 | |
| 1582 | /* Debugfs */ |
| 1583 | #if defined(CONFIG_DEBUG_FS) |
| 1584 | |
| 1585 | void kfd_debugfs_init(void); |
| 1586 | void kfd_debugfs_fini(void); |
| 1587 | int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); |
| 1588 | int pqm_debugfs_mqds(struct seq_file *m, void *data); |
| 1589 | int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); |
| 1590 | int dqm_debugfs_hqds(struct seq_file *m, void *data); |
| 1591 | int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); |
| 1592 | int pm_debugfs_runlist(struct seq_file *m, void *data); |
| 1593 | |
| 1594 | int kfd_debugfs_hang_hws(struct kfd_node *dev); |
| 1595 | int pm_debugfs_hang_hws(struct packet_manager *pm); |
| 1596 | int dqm_debugfs_hang_hws(struct device_queue_manager *dqm); |
| 1597 | |
| 1598 | void kfd_debugfs_add_process(struct kfd_process *p); |
| 1599 | void kfd_debugfs_remove_process(struct kfd_process *p); |
| 1600 | |
| 1601 | #else |
| 1602 | |
| 1603 | static inline void kfd_debugfs_init(void) {} |
| 1604 | static inline void kfd_debugfs_fini(void) {} |
| 1605 | static inline void kfd_debugfs_add_process(struct kfd_process *p) {} |
| 1606 | static inline void kfd_debugfs_remove_process(struct kfd_process *p) {} |
| 1607 | |
| 1608 | #endif |
| 1609 | |
| 1610 | #endif |
| 1611 | |