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_vi.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 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes, 38 "Runlist IB overflow"); 39 *wptr = temp; 40 } 41 42 static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size) 43 { 44 union PM4_MES_TYPE_3_HEADER header; 45 46 header.u32All = 0; 47 header.opcode = opcode; 48 header.count = packet_size / 4 - 2; 49 header.type = PM4_TYPE_3; 50 51 return header.u32All; 52 } 53 54 static void pm_calc_rlib_size(struct packet_manager *pm, 55 unsigned int *rlib_size, 56 bool *over_subscription) 57 { 58 unsigned int process_count, queue_count, compute_queue_count; 59 unsigned int map_queue_size; 60 unsigned int max_proc_per_quantum = 1; 61 struct kfd_dev *dev = pm->dqm->dev; 62 63 process_count = pm->dqm->processes_count; 64 queue_count = pm->dqm->queue_count; 65 compute_queue_count = queue_count - pm->dqm->sdma_queue_count; 66 67 /* check if there is over subscription 68 * Note: the arbitration between the number of VMIDs and 69 * hws_max_conc_proc has been done in 70 * kgd2kfd_device_init(). 71 */ 72 *over_subscription = false; 73 74 if (dev->max_proc_per_quantum > 1) 75 max_proc_per_quantum = dev->max_proc_per_quantum; 76 77 if ((process_count > max_proc_per_quantum) || 78 compute_queue_count > get_queues_num(pm->dqm)) { 79 *over_subscription = true; 80 pr_debug("Over subscribed runlist\n"); 81 } 82 83 map_queue_size = sizeof(struct pm4_mes_map_queues); 84 /* calculate run list ib allocation size */ 85 *rlib_size = process_count * sizeof(struct pm4_mes_map_process) + 86 queue_count * map_queue_size; 87 88 /* 89 * Increase the allocation size in case we need a chained run list 90 * when over subscription 91 */ 92 if (*over_subscription) 93 *rlib_size += sizeof(struct pm4_mes_runlist); 94 95 pr_debug("runlist ib size %d\n", *rlib_size); 96 } 97 98 static int pm_allocate_runlist_ib(struct packet_manager *pm, 99 unsigned int **rl_buffer, 100 uint64_t *rl_gpu_buffer, 101 unsigned int *rl_buffer_size, 102 bool *is_over_subscription) 103 { 104 int retval; 105 106 if (WARN_ON(pm->allocated)) 107 return -EINVAL; 108 109 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); 110 111 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, 112 &pm->ib_buffer_obj); 113 114 if (retval) { 115 pr_err("Failed to allocate runlist IB\n"); 116 return retval; 117 } 118 119 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr; 120 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr; 121 122 memset(*rl_buffer, 0, *rl_buffer_size); 123 pm->allocated = true; 124 return retval; 125 } 126 127 static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer, 128 uint64_t ib, size_t ib_size_in_dwords, bool chain) 129 { 130 struct pm4_mes_runlist *packet; 131 int concurrent_proc_cnt = 0; 132 struct kfd_dev *kfd = pm->dqm->dev; 133 134 if (WARN_ON(!ib)) 135 return -EFAULT; 136 137 /* Determine the number of processes to map together to HW: 138 * it can not exceed the number of VMIDs available to the 139 * scheduler, and it is determined by the smaller of the number 140 * of processes in the runlist and kfd module parameter 141 * hws_max_conc_proc. 142 * Note: the arbitration between the number of VMIDs and 143 * hws_max_conc_proc has been done in 144 * kgd2kfd_device_init(). 145 */ 146 concurrent_proc_cnt = min(pm->dqm->processes_count, 147 kfd->max_proc_per_quantum); 148 149 packet = (struct pm4_mes_runlist *)buffer; 150 151 memset(buffer, 0, sizeof(struct pm4_mes_runlist)); 152 packet->header.u32All = build_pm4_header(IT_RUN_LIST, 153 sizeof(struct pm4_mes_runlist)); 154 155 packet->bitfields4.ib_size = ib_size_in_dwords; 156 packet->bitfields4.chain = chain ? 1 : 0; 157 packet->bitfields4.offload_polling = 0; 158 packet->bitfields4.valid = 1; 159 packet->bitfields4.process_cnt = concurrent_proc_cnt; 160 packet->ordinal2 = lower_32_bits(ib); 161 packet->bitfields3.ib_base_hi = upper_32_bits(ib); 162 163 return 0; 164 } 165 166 static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer, 167 struct qcm_process_device *qpd) 168 { 169 struct pm4_mes_map_process *packet; 170 171 packet = (struct pm4_mes_map_process *)buffer; 172 173 memset(buffer, 0, sizeof(struct pm4_mes_map_process)); 174 175 packet->header.u32All = build_pm4_header(IT_MAP_PROCESS, 176 sizeof(struct pm4_mes_map_process)); 177 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; 178 packet->bitfields2.process_quantum = 1; 179 packet->bitfields2.pasid = qpd->pqm->process->pasid; 180 packet->bitfields3.page_table_base = qpd->page_table_base; 181 packet->bitfields10.gds_size = qpd->gds_size; 182 packet->bitfields10.num_gws = qpd->num_gws; 183 packet->bitfields10.num_oac = qpd->num_oac; 184 packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; 185 186 packet->sh_mem_config = qpd->sh_mem_config; 187 packet->sh_mem_bases = qpd->sh_mem_bases; 188 packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base; 189 packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit; 190 191 /* TODO: scratch support */ 192 packet->sh_hidden_private_base_vmid = 0; 193 194 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); 195 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); 196 197 return 0; 198 } 199 200 static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer, 201 struct queue *q, bool is_static) 202 { 203 struct pm4_mes_map_queues *packet; 204 bool use_static = is_static; 205 206 packet = (struct pm4_mes_map_queues *)buffer; 207 memset(buffer, 0, sizeof(struct pm4_mes_map_queues)); 208 209 packet->header.u32All = build_pm4_header(IT_MAP_QUEUES, 210 sizeof(struct pm4_mes_map_queues)); 211 packet->bitfields2.alloc_format = 212 alloc_format__mes_map_queues__one_per_pipe_vi; 213 packet->bitfields2.num_queues = 1; 214 packet->bitfields2.queue_sel = 215 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi; 216 217 packet->bitfields2.engine_sel = 218 engine_sel__mes_map_queues__compute_vi; 219 packet->bitfields2.queue_type = 220 queue_type__mes_map_queues__normal_compute_vi; 221 222 switch (q->properties.type) { 223 case KFD_QUEUE_TYPE_COMPUTE: 224 if (use_static) 225 packet->bitfields2.queue_type = 226 queue_type__mes_map_queues__normal_latency_static_queue_vi; 227 break; 228 case KFD_QUEUE_TYPE_DIQ: 229 packet->bitfields2.queue_type = 230 queue_type__mes_map_queues__debug_interface_queue_vi; 231 break; 232 case KFD_QUEUE_TYPE_SDMA: 233 packet->bitfields2.engine_sel = q->properties.sdma_engine_id + 234 engine_sel__mes_map_queues__sdma0_vi; 235 use_static = false; /* no static queues under SDMA */ 236 break; 237 default: 238 WARN(1, "queue type %d", q->properties.type); 239 return -EINVAL; 240 } 241 packet->bitfields3.doorbell_offset = 242 q->properties.doorbell_off; 243 244 packet->mqd_addr_lo = 245 lower_32_bits(q->gart_mqd_addr); 246 247 packet->mqd_addr_hi = 248 upper_32_bits(q->gart_mqd_addr); 249 250 packet->wptr_addr_lo = 251 lower_32_bits((uint64_t)q->properties.write_ptr); 252 253 packet->wptr_addr_hi = 254 upper_32_bits((uint64_t)q->properties.write_ptr); 255 256 return 0; 257 } 258 259 static int pm_create_runlist_ib(struct packet_manager *pm, 260 struct list_head *queues, 261 uint64_t *rl_gpu_addr, 262 size_t *rl_size_bytes) 263 { 264 unsigned int alloc_size_bytes; 265 unsigned int *rl_buffer, rl_wptr, i; 266 int retval, proccesses_mapped; 267 struct device_process_node *cur; 268 struct qcm_process_device *qpd; 269 struct queue *q; 270 struct kernel_queue *kq; 271 bool is_over_subscription; 272 273 rl_wptr = retval = proccesses_mapped = 0; 274 275 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, 276 &alloc_size_bytes, &is_over_subscription); 277 if (retval) 278 return retval; 279 280 *rl_size_bytes = alloc_size_bytes; 281 pm->ib_size_bytes = alloc_size_bytes; 282 283 pr_debug("Building runlist ib process count: %d queues count %d\n", 284 pm->dqm->processes_count, pm->dqm->queue_count); 285 286 /* build the run list ib packet */ 287 list_for_each_entry(cur, queues, list) { 288 qpd = cur->qpd; 289 /* build map process packet */ 290 if (proccesses_mapped >= pm->dqm->processes_count) { 291 pr_debug("Not enough space left in runlist IB\n"); 292 pm_release_ib(pm); 293 return -ENOMEM; 294 } 295 296 retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd); 297 if (retval) 298 return retval; 299 300 proccesses_mapped++; 301 inc_wptr(&rl_wptr, sizeof(struct pm4_mes_map_process), 302 alloc_size_bytes); 303 304 list_for_each_entry(kq, &qpd->priv_queue_list, list) { 305 if (!kq->queue->properties.is_active) 306 continue; 307 308 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n", 309 kq->queue->queue, qpd->is_debug); 310 311 retval = pm_create_map_queue(pm, 312 &rl_buffer[rl_wptr], 313 kq->queue, 314 qpd->is_debug); 315 if (retval) 316 return retval; 317 318 inc_wptr(&rl_wptr, 319 sizeof(struct pm4_mes_map_queues), 320 alloc_size_bytes); 321 } 322 323 list_for_each_entry(q, &qpd->queues_list, list) { 324 if (!q->properties.is_active) 325 continue; 326 327 pr_debug("static_queue, mapping user queue %d, is debug status %d\n", 328 q->queue, qpd->is_debug); 329 330 retval = pm_create_map_queue(pm, 331 &rl_buffer[rl_wptr], 332 q, 333 qpd->is_debug); 334 335 if (retval) 336 return retval; 337 338 inc_wptr(&rl_wptr, 339 sizeof(struct pm4_mes_map_queues), 340 alloc_size_bytes); 341 } 342 } 343 344 pr_debug("Finished map process and queues to runlist\n"); 345 346 if (is_over_subscription) 347 retval = pm_create_runlist(pm, &rl_buffer[rl_wptr], 348 *rl_gpu_addr, 349 alloc_size_bytes / sizeof(uint32_t), 350 true); 351 352 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) 353 pr_debug("0x%2X ", rl_buffer[i]); 354 pr_debug("\n"); 355 356 return retval; 357 } 358 359 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) 360 { 361 pm->dqm = dqm; 362 mutex_init(&pm->lock); 363 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); 364 if (!pm->priv_queue) { 365 mutex_destroy(&pm->lock); 366 return -ENOMEM; 367 } 368 pm->allocated = false; 369 370 return 0; 371 } 372 373 void pm_uninit(struct packet_manager *pm) 374 { 375 mutex_destroy(&pm->lock); 376 kernel_queue_uninit(pm->priv_queue); 377 } 378 379 int pm_send_set_resources(struct packet_manager *pm, 380 struct scheduling_resources *res) 381 { 382 struct pm4_mes_set_resources *packet; 383 int retval = 0; 384 385 mutex_lock(&pm->lock); 386 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 387 sizeof(*packet) / sizeof(uint32_t), 388 (unsigned int **)&packet); 389 if (!packet) { 390 pr_err("Failed to allocate buffer on kernel queue\n"); 391 retval = -ENOMEM; 392 goto out; 393 } 394 395 memset(packet, 0, sizeof(struct pm4_mes_set_resources)); 396 packet->header.u32All = build_pm4_header(IT_SET_RESOURCES, 397 sizeof(struct pm4_mes_set_resources)); 398 399 packet->bitfields2.queue_type = 400 queue_type__mes_set_resources__hsa_interface_queue_hiq; 401 packet->bitfields2.vmid_mask = res->vmid_mask; 402 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100; 403 packet->bitfields7.oac_mask = res->oac_mask; 404 packet->bitfields8.gds_heap_base = res->gds_heap_base; 405 packet->bitfields8.gds_heap_size = res->gds_heap_size; 406 407 packet->gws_mask_lo = lower_32_bits(res->gws_mask); 408 packet->gws_mask_hi = upper_32_bits(res->gws_mask); 409 410 packet->queue_mask_lo = lower_32_bits(res->queue_mask); 411 packet->queue_mask_hi = upper_32_bits(res->queue_mask); 412 413 pm->priv_queue->ops.submit_packet(pm->priv_queue); 414 415 out: 416 mutex_unlock(&pm->lock); 417 418 return retval; 419 } 420 421 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) 422 { 423 uint64_t rl_gpu_ib_addr; 424 uint32_t *rl_buffer; 425 size_t rl_ib_size, packet_size_dwords; 426 int retval; 427 428 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, 429 &rl_ib_size); 430 if (retval) 431 goto fail_create_runlist_ib; 432 433 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr); 434 435 packet_size_dwords = sizeof(struct pm4_mes_runlist) / sizeof(uint32_t); 436 mutex_lock(&pm->lock); 437 438 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 439 packet_size_dwords, &rl_buffer); 440 if (retval) 441 goto fail_acquire_packet_buffer; 442 443 retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr, 444 rl_ib_size / sizeof(uint32_t), false); 445 if (retval) 446 goto fail_create_runlist; 447 448 pm->priv_queue->ops.submit_packet(pm->priv_queue); 449 450 mutex_unlock(&pm->lock); 451 452 return retval; 453 454 fail_create_runlist: 455 pm->priv_queue->ops.rollback_packet(pm->priv_queue); 456 fail_acquire_packet_buffer: 457 mutex_unlock(&pm->lock); 458 fail_create_runlist_ib: 459 pm_release_ib(pm); 460 return retval; 461 } 462 463 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 464 uint32_t fence_value) 465 { 466 int retval; 467 struct pm4_mes_query_status *packet; 468 469 if (WARN_ON(!fence_address)) 470 return -EFAULT; 471 472 mutex_lock(&pm->lock); 473 retval = pm->priv_queue->ops.acquire_packet_buffer( 474 pm->priv_queue, 475 sizeof(struct pm4_mes_query_status) / sizeof(uint32_t), 476 (unsigned int **)&packet); 477 if (retval) 478 goto fail_acquire_packet_buffer; 479 480 packet->header.u32All = build_pm4_header(IT_QUERY_STATUS, 481 sizeof(struct pm4_mes_query_status)); 482 483 packet->bitfields2.context_id = 0; 484 packet->bitfields2.interrupt_sel = 485 interrupt_sel__mes_query_status__completion_status; 486 packet->bitfields2.command = 487 command__mes_query_status__fence_only_after_write_ack; 488 489 packet->addr_hi = upper_32_bits((uint64_t)fence_address); 490 packet->addr_lo = lower_32_bits((uint64_t)fence_address); 491 packet->data_hi = upper_32_bits((uint64_t)fence_value); 492 packet->data_lo = lower_32_bits((uint64_t)fence_value); 493 494 pm->priv_queue->ops.submit_packet(pm->priv_queue); 495 496 fail_acquire_packet_buffer: 497 mutex_unlock(&pm->lock); 498 return retval; 499 } 500 501 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 502 enum kfd_unmap_queues_filter filter, 503 uint32_t filter_param, bool reset, 504 unsigned int sdma_engine) 505 { 506 int retval; 507 uint32_t *buffer; 508 struct pm4_mes_unmap_queues *packet; 509 510 mutex_lock(&pm->lock); 511 retval = pm->priv_queue->ops.acquire_packet_buffer( 512 pm->priv_queue, 513 sizeof(struct pm4_mes_unmap_queues) / sizeof(uint32_t), 514 &buffer); 515 if (retval) 516 goto err_acquire_packet_buffer; 517 518 packet = (struct pm4_mes_unmap_queues *)buffer; 519 memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues)); 520 pr_debug("static_queue: unmapping queues: filter is %d , reset is %d , type is %d\n", 521 filter, reset, type); 522 packet->header.u32All = build_pm4_header(IT_UNMAP_QUEUES, 523 sizeof(struct pm4_mes_unmap_queues)); 524 switch (type) { 525 case KFD_QUEUE_TYPE_COMPUTE: 526 case KFD_QUEUE_TYPE_DIQ: 527 packet->bitfields2.engine_sel = 528 engine_sel__mes_unmap_queues__compute; 529 break; 530 case KFD_QUEUE_TYPE_SDMA: 531 packet->bitfields2.engine_sel = 532 engine_sel__mes_unmap_queues__sdma0 + sdma_engine; 533 break; 534 default: 535 WARN(1, "queue type %d", type); 536 retval = -EINVAL; 537 goto err_invalid; 538 } 539 540 if (reset) 541 packet->bitfields2.action = 542 action__mes_unmap_queues__reset_queues; 543 else 544 packet->bitfields2.action = 545 action__mes_unmap_queues__preempt_queues; 546 547 switch (filter) { 548 case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: 549 packet->bitfields2.queue_sel = 550 queue_sel__mes_unmap_queues__perform_request_on_specified_queues; 551 packet->bitfields2.num_queues = 1; 552 packet->bitfields3b.doorbell_offset0 = filter_param; 553 break; 554 case KFD_UNMAP_QUEUES_FILTER_BY_PASID: 555 packet->bitfields2.queue_sel = 556 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues; 557 packet->bitfields3a.pasid = filter_param; 558 break; 559 case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: 560 packet->bitfields2.queue_sel = 561 queue_sel__mes_unmap_queues__unmap_all_queues; 562 break; 563 case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: 564 /* in this case, we do not preempt static queues */ 565 packet->bitfields2.queue_sel = 566 queue_sel__mes_unmap_queues__unmap_all_non_static_queues; 567 break; 568 default: 569 WARN(1, "filter %d", filter); 570 retval = -EINVAL; 571 goto err_invalid; 572 } 573 574 pm->priv_queue->ops.submit_packet(pm->priv_queue); 575 576 mutex_unlock(&pm->lock); 577 return 0; 578 579 err_invalid: 580 pm->priv_queue->ops.rollback_packet(pm->priv_queue); 581 err_acquire_packet_buffer: 582 mutex_unlock(&pm->lock); 583 return retval; 584 } 585 586 void pm_release_ib(struct packet_manager *pm) 587 { 588 mutex_lock(&pm->lock); 589 if (pm->allocated) { 590 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); 591 pm->allocated = false; 592 } 593 mutex_unlock(&pm->lock); 594 } 595 596 #if defined(CONFIG_DEBUG_FS) 597 598 int pm_debugfs_runlist(struct seq_file *m, void *data) 599 { 600 struct packet_manager *pm = data; 601 602 mutex_lock(&pm->lock); 603 604 if (!pm->allocated) { 605 seq_puts(m, " No active runlist\n"); 606 goto out; 607 } 608 609 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 610 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false); 611 612 out: 613 mutex_unlock(&pm->lock); 614 return 0; 615 } 616 617 #endif 618