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/sizeof(uint32_t) - 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; 59 unsigned int map_queue_size; 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) || queue_count > get_queues_num(pm->dqm)) { 67 *over_subscription = true; 68 pr_debug("Over subscribed runlist\n"); 69 } 70 71 map_queue_size = sizeof(struct pm4_mes_map_queues); 72 /* calculate run list ib allocation size */ 73 *rlib_size = process_count * sizeof(struct pm4_mes_map_process) + 74 queue_count * map_queue_size; 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_mes_runlist); 82 83 pr_debug("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 if (WARN_ON(pm->allocated)) 95 return -EINVAL; 96 97 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); 98 99 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, 100 &pm->ib_buffer_obj); 101 102 if (retval) { 103 pr_err("Failed to allocate runlist IB\n"); 104 return retval; 105 } 106 107 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr; 108 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr; 109 110 memset(*rl_buffer, 0, *rl_buffer_size); 111 pm->allocated = true; 112 return retval; 113 } 114 115 static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer, 116 uint64_t ib, size_t ib_size_in_dwords, bool chain) 117 { 118 struct pm4_mes_runlist *packet; 119 120 if (WARN_ON(!ib)) 121 return -EFAULT; 122 123 packet = (struct pm4_mes_runlist *)buffer; 124 125 memset(buffer, 0, sizeof(struct pm4_mes_runlist)); 126 packet->header.u32All = build_pm4_header(IT_RUN_LIST, 127 sizeof(struct pm4_mes_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_mes_map_process *packet; 143 144 packet = (struct pm4_mes_map_process *)buffer; 145 146 memset(buffer, 0, sizeof(struct pm4_mes_map_process)); 147 148 packet->header.u32All = build_pm4_header(IT_MAP_PROCESS, 149 sizeof(struct pm4_mes_map_process)); 150 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; 151 packet->bitfields2.process_quantum = 1; 152 packet->bitfields2.pasid = qpd->pqm->process->pasid; 153 packet->bitfields3.page_table_base = qpd->page_table_base; 154 packet->bitfields10.gds_size = qpd->gds_size; 155 packet->bitfields10.num_gws = qpd->num_gws; 156 packet->bitfields10.num_oac = qpd->num_oac; 157 packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; 158 159 packet->sh_mem_config = qpd->sh_mem_config; 160 packet->sh_mem_bases = qpd->sh_mem_bases; 161 packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base; 162 packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit; 163 164 /* TODO: scratch support */ 165 packet->sh_hidden_private_base_vmid = 0; 166 167 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); 168 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); 169 170 return 0; 171 } 172 173 static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer, 174 struct queue *q, bool is_static) 175 { 176 struct pm4_mes_map_queues *packet; 177 bool use_static = is_static; 178 179 packet = (struct pm4_mes_map_queues *)buffer; 180 memset(buffer, 0, sizeof(struct pm4_mes_map_queues)); 181 182 packet->header.u32All = build_pm4_header(IT_MAP_QUEUES, 183 sizeof(struct pm4_mes_map_queues)); 184 packet->bitfields2.alloc_format = 185 alloc_format__mes_map_queues__one_per_pipe_vi; 186 packet->bitfields2.num_queues = 1; 187 packet->bitfields2.queue_sel = 188 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi; 189 190 packet->bitfields2.engine_sel = 191 engine_sel__mes_map_queues__compute_vi; 192 packet->bitfields2.queue_type = 193 queue_type__mes_map_queues__normal_compute_vi; 194 195 switch (q->properties.type) { 196 case KFD_QUEUE_TYPE_COMPUTE: 197 if (use_static) 198 packet->bitfields2.queue_type = 199 queue_type__mes_map_queues__normal_latency_static_queue_vi; 200 break; 201 case KFD_QUEUE_TYPE_DIQ: 202 packet->bitfields2.queue_type = 203 queue_type__mes_map_queues__debug_interface_queue_vi; 204 break; 205 case KFD_QUEUE_TYPE_SDMA: 206 packet->bitfields2.engine_sel = q->properties.sdma_engine_id + 207 engine_sel__mes_map_queues__sdma0_vi; 208 use_static = false; /* no static queues under SDMA */ 209 break; 210 default: 211 WARN(1, "queue type %d", q->properties.type); 212 return -EINVAL; 213 } 214 packet->bitfields3.doorbell_offset = 215 q->properties.doorbell_off; 216 217 packet->mqd_addr_lo = 218 lower_32_bits(q->gart_mqd_addr); 219 220 packet->mqd_addr_hi = 221 upper_32_bits(q->gart_mqd_addr); 222 223 packet->wptr_addr_lo = 224 lower_32_bits((uint64_t)q->properties.write_ptr); 225 226 packet->wptr_addr_hi = 227 upper_32_bits((uint64_t)q->properties.write_ptr); 228 229 return 0; 230 } 231 232 static int pm_create_runlist_ib(struct packet_manager *pm, 233 struct list_head *queues, 234 uint64_t *rl_gpu_addr, 235 size_t *rl_size_bytes) 236 { 237 unsigned int alloc_size_bytes; 238 unsigned int *rl_buffer, rl_wptr, i; 239 int retval, proccesses_mapped; 240 struct device_process_node *cur; 241 struct qcm_process_device *qpd; 242 struct queue *q; 243 struct kernel_queue *kq; 244 bool is_over_subscription; 245 246 rl_wptr = retval = proccesses_mapped = 0; 247 248 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, 249 &alloc_size_bytes, &is_over_subscription); 250 if (retval) 251 return retval; 252 253 *rl_size_bytes = alloc_size_bytes; 254 255 pr_debug("Building runlist ib process count: %d queues count %d\n", 256 pm->dqm->processes_count, pm->dqm->queue_count); 257 258 /* build the run list ib packet */ 259 list_for_each_entry(cur, queues, list) { 260 qpd = cur->qpd; 261 /* build map process packet */ 262 if (proccesses_mapped >= pm->dqm->processes_count) { 263 pr_debug("Not enough space left in runlist IB\n"); 264 pm_release_ib(pm); 265 return -ENOMEM; 266 } 267 268 retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd); 269 if (retval) 270 return retval; 271 272 proccesses_mapped++; 273 inc_wptr(&rl_wptr, sizeof(struct pm4_mes_map_process), 274 alloc_size_bytes); 275 276 list_for_each_entry(kq, &qpd->priv_queue_list, list) { 277 if (!kq->queue->properties.is_active) 278 continue; 279 280 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n", 281 kq->queue->queue, qpd->is_debug); 282 283 retval = pm_create_map_queue(pm, 284 &rl_buffer[rl_wptr], 285 kq->queue, 286 qpd->is_debug); 287 if (retval) 288 return retval; 289 290 inc_wptr(&rl_wptr, 291 sizeof(struct pm4_mes_map_queues), 292 alloc_size_bytes); 293 } 294 295 list_for_each_entry(q, &qpd->queues_list, list) { 296 if (!q->properties.is_active) 297 continue; 298 299 pr_debug("static_queue, mapping user queue %d, is debug status %d\n", 300 q->queue, qpd->is_debug); 301 302 retval = pm_create_map_queue(pm, 303 &rl_buffer[rl_wptr], 304 q, 305 qpd->is_debug); 306 307 if (retval) 308 return retval; 309 310 inc_wptr(&rl_wptr, 311 sizeof(struct pm4_mes_map_queues), 312 alloc_size_bytes); 313 } 314 } 315 316 pr_debug("Finished map process and queues to runlist\n"); 317 318 if (is_over_subscription) 319 retval = pm_create_runlist(pm, &rl_buffer[rl_wptr], 320 *rl_gpu_addr, 321 alloc_size_bytes / sizeof(uint32_t), 322 true); 323 324 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) 325 pr_debug("0x%2X ", rl_buffer[i]); 326 pr_debug("\n"); 327 328 return retval; 329 } 330 331 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) 332 { 333 pm->dqm = dqm; 334 mutex_init(&pm->lock); 335 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); 336 if (!pm->priv_queue) { 337 mutex_destroy(&pm->lock); 338 return -ENOMEM; 339 } 340 pm->allocated = false; 341 342 return 0; 343 } 344 345 void pm_uninit(struct packet_manager *pm) 346 { 347 mutex_destroy(&pm->lock); 348 kernel_queue_uninit(pm->priv_queue); 349 } 350 351 int pm_send_set_resources(struct packet_manager *pm, 352 struct scheduling_resources *res) 353 { 354 struct pm4_mes_set_resources *packet; 355 int retval = 0; 356 357 mutex_lock(&pm->lock); 358 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 359 sizeof(*packet) / sizeof(uint32_t), 360 (unsigned int **)&packet); 361 if (!packet) { 362 pr_err("Failed to allocate buffer on kernel queue\n"); 363 retval = -ENOMEM; 364 goto out; 365 } 366 367 memset(packet, 0, sizeof(struct pm4_mes_set_resources)); 368 packet->header.u32All = build_pm4_header(IT_SET_RESOURCES, 369 sizeof(struct pm4_mes_set_resources)); 370 371 packet->bitfields2.queue_type = 372 queue_type__mes_set_resources__hsa_interface_queue_hiq; 373 packet->bitfields2.vmid_mask = res->vmid_mask; 374 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100; 375 packet->bitfields7.oac_mask = res->oac_mask; 376 packet->bitfields8.gds_heap_base = res->gds_heap_base; 377 packet->bitfields8.gds_heap_size = res->gds_heap_size; 378 379 packet->gws_mask_lo = lower_32_bits(res->gws_mask); 380 packet->gws_mask_hi = upper_32_bits(res->gws_mask); 381 382 packet->queue_mask_lo = lower_32_bits(res->queue_mask); 383 packet->queue_mask_hi = upper_32_bits(res->queue_mask); 384 385 pm->priv_queue->ops.submit_packet(pm->priv_queue); 386 387 out: 388 mutex_unlock(&pm->lock); 389 390 return retval; 391 } 392 393 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) 394 { 395 uint64_t rl_gpu_ib_addr; 396 uint32_t *rl_buffer; 397 size_t rl_ib_size, packet_size_dwords; 398 int retval; 399 400 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, 401 &rl_ib_size); 402 if (retval) 403 goto fail_create_runlist_ib; 404 405 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr); 406 407 packet_size_dwords = sizeof(struct pm4_mes_runlist) / sizeof(uint32_t); 408 mutex_lock(&pm->lock); 409 410 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, 411 packet_size_dwords, &rl_buffer); 412 if (retval) 413 goto fail_acquire_packet_buffer; 414 415 retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr, 416 rl_ib_size / sizeof(uint32_t), false); 417 if (retval) 418 goto fail_create_runlist; 419 420 pm->priv_queue->ops.submit_packet(pm->priv_queue); 421 422 mutex_unlock(&pm->lock); 423 424 return retval; 425 426 fail_create_runlist: 427 pm->priv_queue->ops.rollback_packet(pm->priv_queue); 428 fail_acquire_packet_buffer: 429 mutex_unlock(&pm->lock); 430 fail_create_runlist_ib: 431 pm_release_ib(pm); 432 return retval; 433 } 434 435 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 436 uint32_t fence_value) 437 { 438 int retval; 439 struct pm4_mes_query_status *packet; 440 441 if (WARN_ON(!fence_address)) 442 return -EFAULT; 443 444 mutex_lock(&pm->lock); 445 retval = pm->priv_queue->ops.acquire_packet_buffer( 446 pm->priv_queue, 447 sizeof(struct pm4_mes_query_status) / sizeof(uint32_t), 448 (unsigned int **)&packet); 449 if (retval) 450 goto fail_acquire_packet_buffer; 451 452 packet->header.u32All = build_pm4_header(IT_QUERY_STATUS, 453 sizeof(struct pm4_mes_query_status)); 454 455 packet->bitfields2.context_id = 0; 456 packet->bitfields2.interrupt_sel = 457 interrupt_sel__mes_query_status__completion_status; 458 packet->bitfields2.command = 459 command__mes_query_status__fence_only_after_write_ack; 460 461 packet->addr_hi = upper_32_bits((uint64_t)fence_address); 462 packet->addr_lo = lower_32_bits((uint64_t)fence_address); 463 packet->data_hi = upper_32_bits((uint64_t)fence_value); 464 packet->data_lo = lower_32_bits((uint64_t)fence_value); 465 466 pm->priv_queue->ops.submit_packet(pm->priv_queue); 467 468 fail_acquire_packet_buffer: 469 mutex_unlock(&pm->lock); 470 return retval; 471 } 472 473 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 474 enum kfd_unmap_queues_filter filter, 475 uint32_t filter_param, bool reset, 476 unsigned int sdma_engine) 477 { 478 int retval; 479 uint32_t *buffer; 480 struct pm4_mes_unmap_queues *packet; 481 482 mutex_lock(&pm->lock); 483 retval = pm->priv_queue->ops.acquire_packet_buffer( 484 pm->priv_queue, 485 sizeof(struct pm4_mes_unmap_queues) / sizeof(uint32_t), 486 &buffer); 487 if (retval) 488 goto err_acquire_packet_buffer; 489 490 packet = (struct pm4_mes_unmap_queues *)buffer; 491 memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues)); 492 pr_debug("static_queue: unmapping queues: filter is %d , reset is %d , type is %d\n", 493 filter, reset, type); 494 packet->header.u32All = build_pm4_header(IT_UNMAP_QUEUES, 495 sizeof(struct pm4_mes_unmap_queues)); 496 switch (type) { 497 case KFD_QUEUE_TYPE_COMPUTE: 498 case KFD_QUEUE_TYPE_DIQ: 499 packet->bitfields2.engine_sel = 500 engine_sel__mes_unmap_queues__compute; 501 break; 502 case KFD_QUEUE_TYPE_SDMA: 503 packet->bitfields2.engine_sel = 504 engine_sel__mes_unmap_queues__sdma0 + sdma_engine; 505 break; 506 default: 507 WARN(1, "queue type %d", type); 508 retval = -EINVAL; 509 goto err_invalid; 510 } 511 512 if (reset) 513 packet->bitfields2.action = 514 action__mes_unmap_queues__reset_queues; 515 else 516 packet->bitfields2.action = 517 action__mes_unmap_queues__preempt_queues; 518 519 switch (filter) { 520 case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: 521 packet->bitfields2.queue_sel = 522 queue_sel__mes_unmap_queues__perform_request_on_specified_queues; 523 packet->bitfields2.num_queues = 1; 524 packet->bitfields3b.doorbell_offset0 = filter_param; 525 break; 526 case KFD_UNMAP_QUEUES_FILTER_BY_PASID: 527 packet->bitfields2.queue_sel = 528 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues; 529 packet->bitfields3a.pasid = filter_param; 530 break; 531 case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: 532 packet->bitfields2.queue_sel = 533 queue_sel__mes_unmap_queues__unmap_all_queues; 534 break; 535 case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: 536 /* in this case, we do not preempt static queues */ 537 packet->bitfields2.queue_sel = 538 queue_sel__mes_unmap_queues__unmap_all_non_static_queues; 539 break; 540 default: 541 WARN(1, "filter %d", filter); 542 retval = -EINVAL; 543 goto err_invalid; 544 } 545 546 pm->priv_queue->ops.submit_packet(pm->priv_queue); 547 548 mutex_unlock(&pm->lock); 549 return 0; 550 551 err_invalid: 552 pm->priv_queue->ops.rollback_packet(pm->priv_queue); 553 err_acquire_packet_buffer: 554 mutex_unlock(&pm->lock); 555 return retval; 556 } 557 558 void pm_release_ib(struct packet_manager *pm) 559 { 560 mutex_lock(&pm->lock); 561 if (pm->allocated) { 562 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); 563 pm->allocated = false; 564 } 565 mutex_unlock(&pm->lock); 566 } 567