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 24 #include <linux/slab.h> 25 #include <linux/mutex.h> 26 #include "kfd_device_queue_manager.h" 27 #include "kfd_kernel_queue.h" 28 #include "kfd_priv.h" 29 #include "kfd_pm4_headers.h" 30 #include "kfd_pm4_opcodes.h" 31 32 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes, 33 unsigned int buffer_size_bytes) 34 { 35 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t); 36 37 BUG_ON((temp * sizeof(uint32_t)) > buffer_size_bytes); 38 *wptr = temp; 39 } 40 41 static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size) 42 { 43 union PM4_MES_TYPE_3_HEADER header; 44 45 header.u32all = 0; 46 header.opcode = opcode; 47 header.count = packet_size/sizeof(uint32_t) - 2; 48 header.type = PM4_TYPE_3; 49 50 return header.u32all; 51 } 52 53 static void pm_calc_rlib_size(struct packet_manager *pm, 54 unsigned int *rlib_size, 55 bool *over_subscription) 56 { 57 unsigned int process_count, queue_count; 58 59 BUG_ON(!pm || !rlib_size || !over_subscription); 60 61 process_count = pm->dqm->processes_count; 62 queue_count = pm->dqm->queue_count; 63 64 /* check if there is over subscription*/ 65 *over_subscription = false; 66 if ((process_count > 1) || 67 queue_count > PIPE_PER_ME_CP_SCHEDULING * QUEUES_PER_PIPE) { 68 *over_subscription = true; 69 pr_debug("kfd: over subscribed runlist\n"); 70 } 71 72 /* calculate run list ib allocation size */ 73 *rlib_size = process_count * sizeof(struct pm4_map_process) + 74 queue_count * sizeof(struct pm4_map_queues); 75 76 /* 77 * Increase the allocation size in case we need a chained run list 78 * when over subscription 79 */ 80 if (*over_subscription) 81 *rlib_size += sizeof(struct pm4_runlist); 82 83 pr_debug("kfd: runlist ib size %d\n", *rlib_size); 84 } 85 86 static int pm_allocate_runlist_ib(struct packet_manager *pm, 87 unsigned int **rl_buffer, 88 uint64_t *rl_gpu_buffer, 89 unsigned int *rl_buffer_size, 90 bool *is_over_subscription) 91 { 92 int retval; 93 94 BUG_ON(!pm); 95 BUG_ON(pm->allocated == true); 96 BUG_ON(is_over_subscription == NULL); 97 98 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); 99 100 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, 101 &pm->ib_buffer_obj); 102 103 if (retval != 0) { 104 pr_err("kfd: failed to allocate runlist IB\n"); 105 return retval; 106 } 107 108 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr; 109 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr; 110 111 memset(*rl_buffer, 0, *rl_buffer_size); 112 pm->allocated = true; 113 return retval; 114 } 115 116 static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer, 117 uint64_t ib, size_t ib_size_in_dwords, bool chain) 118 { 119 struct pm4_runlist *packet; 120 121 BUG_ON(!pm || !buffer || !ib); 122 123 packet = (struct pm4_runlist *)buffer; 124 125 memset(buffer, 0, sizeof(struct pm4_runlist)); 126 packet->header.u32all = build_pm4_header(IT_RUN_LIST, 127 sizeof(struct pm4_runlist)); 128 129 packet->bitfields4.ib_size = ib_size_in_dwords; 130 packet->bitfields4.chain = chain ? 1 : 0; 131 packet->bitfields4.offload_polling = 0; 132 packet->bitfields4.valid = 1; 133 packet->ordinal2 = lower_32_bits(ib); 134 packet->bitfields3.ib_base_hi = upper_32_bits(ib); 135 136 return 0; 137 } 138 139 static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer, 140 struct qcm_process_device *qpd) 141 { 142 struct pm4_map_process *packet; 143 struct queue *cur; 144 uint32_t num_queues; 145 146 BUG_ON(!pm || !buffer || !qpd); 147 148 packet = (struct pm4_map_process *)buffer; 149 150 pr_debug("kfd: In func %s\n", __func__); 151 152 memset(buffer, 0, sizeof(struct pm4_map_process)); 153 154 packet->header.u32all = build_pm4_header(IT_MAP_PROCESS, 155 sizeof(struct pm4_map_process)); 156 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; 157 packet->bitfields2.process_quantum = 1; 158 packet->bitfields2.pasid = qpd->pqm->process->pasid; 159 packet->bitfields3.page_table_base = qpd->page_table_base; 160 packet->bitfields10.gds_size = qpd->gds_size; 161 packet->bitfields10.num_gws = qpd->num_gws; 162 packet->bitfields10.num_oac = qpd->num_oac; 163 num_queues = 0; 164 list_for_each_entry(cur, &qpd->queues_list, list) 165 num_queues++; 166 packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues; 167 168 packet->sh_mem_config = qpd->sh_mem_config; 169 packet->sh_mem_bases = qpd->sh_mem_bases; 170 packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base; 171 packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit; 172 173 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); 174 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); 175 176 return 0; 177 } 178 179 static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer, 180 struct queue *q, bool is_static) 181 { 182 struct pm4_map_queues *packet; 183 bool use_static = is_static; 184 185 BUG_ON(!pm || !buffer || !q); 186 187 pr_debug("kfd: In func %s\n", __func__); 188 189 packet = (struct pm4_map_queues *)buffer; 190 memset(buffer, 0, sizeof(struct pm4_map_queues)); 191 192 packet->header.u32all = build_pm4_header(IT_MAP_QUEUES, 193 sizeof(struct pm4_map_queues)); 194 packet->bitfields2.alloc_format = 195 alloc_format__mes_map_queues__one_per_pipe; 196 packet->bitfields2.num_queues = 1; 197 packet->bitfields2.queue_sel = 198 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots; 199 200 packet->bitfields2.vidmem = (q->properties.is_interop) ? 201 vidmem__mes_map_queues__uses_video_memory : 202 vidmem__mes_map_queues__uses_no_video_memory; 203 204 switch (q->properties.type) { 205 case KFD_QUEUE_TYPE_COMPUTE: 206 case KFD_QUEUE_TYPE_DIQ: 207 packet->bitfields2.engine_sel = 208 engine_sel__mes_map_queues__compute; 209 break; 210 case KFD_QUEUE_TYPE_SDMA: 211 packet->bitfields2.engine_sel = 212 engine_sel__mes_map_queues__sdma0; 213 use_static = false; /* no static queues under SDMA */ 214 break; 215 default: 216 BUG(); 217 break; 218 } 219 220 packet->mes_map_queues_ordinals[0].bitfields3.doorbell_offset = 221 q->properties.doorbell_off; 222 223 packet->mes_map_queues_ordinals[0].bitfields3.is_static = 224 (use_static == true) ? 1 : 0; 225 226 packet->mes_map_queues_ordinals[0].mqd_addr_lo = 227 lower_32_bits(q->gart_mqd_addr); 228 229 packet->mes_map_queues_ordinals[0].mqd_addr_hi = 230 upper_32_bits(q->gart_mqd_addr); 231 232 packet->mes_map_queues_ordinals[0].wptr_addr_lo = 233 lower_32_bits((uint64_t)q->properties.write_ptr); 234 235 packet->mes_map_queues_ordinals[0].wptr_addr_hi = 236 upper_32_bits((uint64_t)q->properties.write_ptr); 237 238 return 0; 239 } 240 241 static int pm_create_runlist_ib(struct packet_manager *pm, 242 struct list_head *queues, 243 uint64_t *rl_gpu_addr, 244 size_t *rl_size_bytes) 245 { 246 unsigned int alloc_size_bytes; 247 unsigned int *rl_buffer, rl_wptr, i; 248 int retval, proccesses_mapped; 249 struct device_process_node *cur; 250 struct qcm_process_device *qpd; 251 struct queue *q; 252 struct kernel_queue *kq; 253 bool is_over_subscription; 254 255 BUG_ON(!pm || !queues || !rl_size_bytes || !rl_gpu_addr); 256 257 rl_wptr = retval = proccesses_mapped = 0; 258 259 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, 260 &alloc_size_bytes, &is_over_subscription); 261 if (retval != 0) 262 return retval; 263 264 *rl_size_bytes = alloc_size_bytes; 265 266 pr_debug("kfd: In func %s\n", __func__); 267 pr_debug("kfd: building runlist ib process count: %d queues count %d\n", 268 pm->dqm->processes_count, pm->dqm->queue_count); 269 270 /* build the run list ib packet */ 271 list_for_each_entry(cur, queues, list) { 272 qpd = cur->qpd; 273 /* build map process packet */ 274 if (proccesses_mapped >= pm->dqm->processes_count) { 275 pr_debug("kfd: not enough space left in runlist IB\n"); 276 pm_release_ib(pm); 277 return -ENOMEM; 278 } 279 280 retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd); 281 if (retval != 0) 282 return retval; 283 284 proccesses_mapped++; 285 inc_wptr(&rl_wptr, sizeof(struct pm4_map_process), 286 alloc_size_bytes); 287 288 list_for_each_entry(kq, &qpd->priv_queue_list, list) { 289 if (kq->queue->properties.is_active != true) 290 continue; 291 292 pr_debug("kfd: static_queue, mapping kernel q %d, is debug status %d\n", 293 kq->queue->queue, qpd->is_debug); 294 295 retval = pm_create_map_queue(pm, &rl_buffer[rl_wptr], 296 kq->queue, qpd->is_debug); 297 if (retval != 0) 298 return retval; 299 300 inc_wptr(&rl_wptr, 301 sizeof(struct pm4_map_queues), 302 alloc_size_bytes); 303 } 304 305 list_for_each_entry(q, &qpd->queues_list, list) { 306 if (q->properties.is_active != true) 307 continue; 308 309 pr_debug("kfd: static_queue, mapping user queue %d, is debug status %d\n", 310 q->queue, qpd->is_debug); 311 312 retval = pm_create_map_queue(pm, &rl_buffer[rl_wptr], 313 q, qpd->is_debug); 314 315 if (retval != 0) 316 return retval; 317 318 inc_wptr(&rl_wptr, 319 sizeof(struct pm4_map_queues), 320 alloc_size_bytes); 321 } 322 } 323 324 pr_debug("kfd: finished map process and queues to runlist\n"); 325 326 if (is_over_subscription) 327 pm_create_runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr, 328 alloc_size_bytes / sizeof(uint32_t), true); 329 330 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) 331 pr_debug("0x%2X ", rl_buffer[i]); 332 pr_debug("\n"); 333 334 return 0; 335 } 336 337 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) 338 { 339 BUG_ON(!dqm); 340 341 pm->dqm = dqm; 342 mutex_init(&pm->lock); 343 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); 344 if (pm->priv_queue == NULL) { 345 mutex_destroy(&pm->lock); 346 return -ENOMEM; 347 } 348 pm->allocated = false; 349 350 return 0; 351 } 352 353 void pm_uninit(struct packet_manager *pm) 354 { 355 BUG_ON(!pm); 356 357 mutex_destroy(&pm->lock); 358 kernel_queue_uninit(pm->priv_queue); 359 } 360 361 int pm_send_set_resources(struct packet_manager *pm, 362 struct scheduling_resources *res) 363 { 364 struct pm4_set_resources *packet; 365 366 BUG_ON(!pm || !res); 367 368 pr_debug("kfd: In func %s\n", __func__); 369 370 mutex_lock(&pm->lock); 371 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 372 sizeof(*packet) / sizeof(uint32_t), 373 (unsigned int **)&packet); 374 if (packet == NULL) { 375 mutex_unlock(&pm->lock); 376 pr_err("kfd: failed to allocate buffer on kernel queue\n"); 377 return -ENOMEM; 378 } 379 380 memset(packet, 0, sizeof(struct pm4_set_resources)); 381 packet->header.u32all = build_pm4_header(IT_SET_RESOURCES, 382 sizeof(struct pm4_set_resources)); 383 384 packet->bitfields2.queue_type = 385 queue_type__mes_set_resources__hsa_interface_queue_hiq; 386 packet->bitfields2.vmid_mask = res->vmid_mask; 387 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY; 388 packet->bitfields7.oac_mask = res->oac_mask; 389 packet->bitfields8.gds_heap_base = res->gds_heap_base; 390 packet->bitfields8.gds_heap_size = res->gds_heap_size; 391 392 packet->gws_mask_lo = lower_32_bits(res->gws_mask); 393 packet->gws_mask_hi = upper_32_bits(res->gws_mask); 394 395 packet->queue_mask_lo = lower_32_bits(res->queue_mask); 396 packet->queue_mask_hi = upper_32_bits(res->queue_mask); 397 398 pm->priv_queue->ops.submit_packet(pm->priv_queue); 399 400 mutex_unlock(&pm->lock); 401 402 return 0; 403 } 404 405 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) 406 { 407 uint64_t rl_gpu_ib_addr; 408 uint32_t *rl_buffer; 409 size_t rl_ib_size, packet_size_dwords; 410 int retval; 411 412 BUG_ON(!pm || !dqm_queues); 413 414 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, 415 &rl_ib_size); 416 if (retval != 0) 417 goto fail_create_runlist_ib; 418 419 pr_debug("kfd: runlist IB address: 0x%llX\n", rl_gpu_ib_addr); 420 421 packet_size_dwords = sizeof(struct pm4_runlist) / sizeof(uint32_t); 422 mutex_lock(&pm->lock); 423 424 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 425 packet_size_dwords, &rl_buffer); 426 if (retval != 0) 427 goto fail_acquire_packet_buffer; 428 429 retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr, 430 rl_ib_size / sizeof(uint32_t), false); 431 if (retval != 0) 432 goto fail_create_runlist; 433 434 pm->priv_queue->ops.submit_packet(pm->priv_queue); 435 436 mutex_unlock(&pm->lock); 437 438 return retval; 439 440 fail_create_runlist: 441 pm->priv_queue->ops.rollback_packet(pm->priv_queue); 442 fail_acquire_packet_buffer: 443 mutex_unlock(&pm->lock); 444 fail_create_runlist_ib: 445 if (pm->allocated == true) 446 pm_release_ib(pm); 447 return retval; 448 } 449 450 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 451 uint32_t fence_value) 452 { 453 int retval; 454 struct pm4_query_status *packet; 455 456 BUG_ON(!pm || !fence_address); 457 458 mutex_lock(&pm->lock); 459 retval = pm->priv_queue->ops.acquire_packet_buffer( 460 pm->priv_queue, 461 sizeof(struct pm4_query_status) / sizeof(uint32_t), 462 (unsigned int **)&packet); 463 if (retval != 0) 464 goto fail_acquire_packet_buffer; 465 466 packet->header.u32all = build_pm4_header(IT_QUERY_STATUS, 467 sizeof(struct pm4_query_status)); 468 469 packet->bitfields2.context_id = 0; 470 packet->bitfields2.interrupt_sel = 471 interrupt_sel__mes_query_status__completion_status; 472 packet->bitfields2.command = 473 command__mes_query_status__fence_only_after_write_ack; 474 475 packet->addr_hi = upper_32_bits((uint64_t)fence_address); 476 packet->addr_lo = lower_32_bits((uint64_t)fence_address); 477 packet->data_hi = upper_32_bits((uint64_t)fence_value); 478 packet->data_lo = lower_32_bits((uint64_t)fence_value); 479 480 pm->priv_queue->ops.submit_packet(pm->priv_queue); 481 mutex_unlock(&pm->lock); 482 483 return 0; 484 485 fail_acquire_packet_buffer: 486 mutex_unlock(&pm->lock); 487 return retval; 488 } 489 490 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 491 enum kfd_preempt_type_filter mode, 492 uint32_t filter_param, bool reset, 493 unsigned int sdma_engine) 494 { 495 int retval; 496 uint32_t *buffer; 497 struct pm4_unmap_queues *packet; 498 499 BUG_ON(!pm); 500 501 mutex_lock(&pm->lock); 502 retval = pm->priv_queue->ops.acquire_packet_buffer( 503 pm->priv_queue, 504 sizeof(struct pm4_unmap_queues) / sizeof(uint32_t), 505 &buffer); 506 if (retval != 0) 507 goto err_acquire_packet_buffer; 508 509 packet = (struct pm4_unmap_queues *)buffer; 510 memset(buffer, 0, sizeof(struct pm4_unmap_queues)); 511 pr_debug("kfd: static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n", 512 mode, reset, type); 513 packet->header.u32all = build_pm4_header(IT_UNMAP_QUEUES, 514 sizeof(struct pm4_unmap_queues)); 515 switch (type) { 516 case KFD_QUEUE_TYPE_COMPUTE: 517 case KFD_QUEUE_TYPE_DIQ: 518 packet->bitfields2.engine_sel = 519 engine_sel__mes_unmap_queues__compute; 520 break; 521 case KFD_QUEUE_TYPE_SDMA: 522 packet->bitfields2.engine_sel = 523 engine_sel__mes_unmap_queues__sdma0 + sdma_engine; 524 break; 525 default: 526 BUG(); 527 break; 528 } 529 530 if (reset) 531 packet->bitfields2.action = 532 action__mes_unmap_queues__reset_queues; 533 else 534 packet->bitfields2.action = 535 action__mes_unmap_queues__preempt_queues; 536 537 switch (mode) { 538 case KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: 539 packet->bitfields2.queue_sel = 540 queue_sel__mes_unmap_queues__perform_request_on_specified_queues; 541 packet->bitfields2.num_queues = 1; 542 packet->bitfields3b.doorbell_offset0 = filter_param; 543 break; 544 case KFD_PREEMPT_TYPE_FILTER_BY_PASID: 545 packet->bitfields2.queue_sel = 546 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues; 547 packet->bitfields3a.pasid = filter_param; 548 break; 549 case KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES: 550 packet->bitfields2.queue_sel = 551 queue_sel__mes_unmap_queues__perform_request_on_all_active_queues; 552 break; 553 case KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES: 554 /* in this case, we do not preempt static queues */ 555 packet->bitfields2.queue_sel = 556 queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only; 557 break; 558 default: 559 BUG(); 560 break; 561 }; 562 563 pm->priv_queue->ops.submit_packet(pm->priv_queue); 564 565 mutex_unlock(&pm->lock); 566 return 0; 567 568 err_acquire_packet_buffer: 569 mutex_unlock(&pm->lock); 570 return retval; 571 } 572 573 void pm_release_ib(struct packet_manager *pm) 574 { 575 BUG_ON(!pm); 576 577 mutex_lock(&pm->lock); 578 if (pm->allocated) { 579 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); 580 pm->allocated = false; 581 } 582 mutex_unlock(&pm->lock); 583 } 584