1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2016-2022 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 25 #include "kfd_kernel_queue.h" 26 #include "kfd_device_queue_manager.h" 27 #include "kfd_pm4_headers_ai.h" 28 #include "kfd_pm4_headers_aldebaran.h" 29 #include "kfd_pm4_opcodes.h" 30 #include "gc/gc_10_1_0_sh_mask.h" 31 32 static int pm_map_process_v9(struct packet_manager *pm, 33 uint32_t *buffer, struct qcm_process_device *qpd) 34 { 35 struct pm4_mes_map_process *packet; 36 uint64_t vm_page_table_base_addr = qpd->page_table_base; 37 struct kfd_node *kfd = pm->dqm->dev; 38 struct kfd_process_device *pdd = 39 container_of(qpd, struct kfd_process_device, qpd); 40 struct amdgpu_device *adev = kfd->adev; 41 42 packet = (struct pm4_mes_map_process *)buffer; 43 memset(buffer, 0, sizeof(struct pm4_mes_map_process)); 44 packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS, 45 sizeof(struct pm4_mes_map_process)); 46 if (adev->enforce_isolation[kfd->node_id] == AMDGPU_ENFORCE_ISOLATION_ENABLE) 47 packet->bitfields2.exec_cleaner_shader = 1; 48 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; 49 packet->bitfields2.process_quantum = 10; 50 packet->bitfields2.pasid = pdd->pasid; 51 packet->bitfields14.gds_size = qpd->gds_size & 0x3F; 52 packet->bitfields14.gds_size_hi = (qpd->gds_size >> 6) & 0xF; 53 packet->bitfields14.num_gws = (qpd->mapped_gws_queue) ? qpd->num_gws : 0; 54 packet->bitfields14.num_oac = qpd->num_oac; 55 packet->bitfields14.sdma_enable = 1; 56 packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; 57 58 if (kfd->dqm->trap_debug_vmid && pdd->process->debug_trap_enabled && 59 pdd->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) { 60 packet->bitfields2.debug_vmid = kfd->dqm->trap_debug_vmid; 61 packet->bitfields2.new_debug = 1; 62 } 63 64 packet->sh_mem_config = qpd->sh_mem_config; 65 packet->sh_mem_bases = qpd->sh_mem_bases; 66 if (qpd->tba_addr) { 67 packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8); 68 /* On GFX9, unlike GFX10, bit TRAP_EN of SQ_SHADER_TBA_HI is 69 * not defined, so setting it won't do any harm. 70 */ 71 packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8) 72 | 1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT; 73 74 packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8); 75 packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8); 76 } 77 78 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); 79 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); 80 81 packet->vm_context_page_table_base_addr_lo32 = 82 lower_32_bits(vm_page_table_base_addr); 83 packet->vm_context_page_table_base_addr_hi32 = 84 upper_32_bits(vm_page_table_base_addr); 85 86 return 0; 87 } 88 89 static int pm_map_process_aldebaran(struct packet_manager *pm, 90 uint32_t *buffer, struct qcm_process_device *qpd) 91 { 92 struct pm4_mes_map_process_aldebaran *packet; 93 uint64_t vm_page_table_base_addr = qpd->page_table_base; 94 struct kfd_dev *kfd = pm->dqm->dev->kfd; 95 struct kfd_node *knode = pm->dqm->dev; 96 struct kfd_process_device *pdd = 97 container_of(qpd, struct kfd_process_device, qpd); 98 int i; 99 struct amdgpu_device *adev = kfd->adev; 100 101 packet = (struct pm4_mes_map_process_aldebaran *)buffer; 102 memset(buffer, 0, sizeof(struct pm4_mes_map_process_aldebaran)); 103 packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS, 104 sizeof(struct pm4_mes_map_process_aldebaran)); 105 if (adev->enforce_isolation[knode->node_id] == 106 AMDGPU_ENFORCE_ISOLATION_ENABLE) 107 packet->bitfields2.exec_cleaner_shader = 1; 108 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; 109 packet->bitfields2.process_quantum = 10; 110 packet->bitfields2.pasid = pdd->pasid; 111 packet->bitfields14.gds_size = qpd->gds_size & 0x3F; 112 packet->bitfields14.gds_size_hi = (qpd->gds_size >> 6) & 0xF; 113 packet->bitfields14.num_gws = (qpd->mapped_gws_queue) ? qpd->num_gws : 0; 114 packet->bitfields14.num_oac = qpd->num_oac; 115 packet->bitfields14.sdma_enable = 1; 116 packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; 117 packet->spi_gdbg_per_vmid_cntl = pdd->spi_dbg_override | 118 pdd->spi_dbg_launch_mode; 119 120 if (pdd->process->debug_trap_enabled) { 121 for (i = 0; i < kfd->device_info.num_of_watch_points; i++) 122 packet->tcp_watch_cntl[i] = pdd->watch_points[i]; 123 124 packet->bitfields2.single_memops = 125 !!(pdd->process->dbg_flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP); 126 } 127 128 packet->sh_mem_config = qpd->sh_mem_config; 129 packet->sh_mem_bases = qpd->sh_mem_bases; 130 if (qpd->tba_addr) { 131 packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8); 132 packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8); 133 packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8); 134 packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8); 135 } 136 137 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); 138 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); 139 140 packet->vm_context_page_table_base_addr_lo32 = 141 lower_32_bits(vm_page_table_base_addr); 142 packet->vm_context_page_table_base_addr_hi32 = 143 upper_32_bits(vm_page_table_base_addr); 144 145 return 0; 146 } 147 148 static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer, 149 uint64_t ib, size_t ib_size_in_dwords, bool chain) 150 { 151 struct pm4_mes_runlist *packet; 152 153 int concurrent_proc_cnt = 0; 154 struct kfd_node *kfd = pm->dqm->dev; 155 struct amdgpu_device *adev = kfd->adev; 156 157 /* Determine the number of processes to map together to HW: 158 * it can not exceed the number of VMIDs available to the 159 * scheduler, and it is determined by the smaller of the number 160 * of processes in the runlist and kfd module parameter 161 * hws_max_conc_proc. 162 * However, if enforce_isolation is set (toggle LDS/VGPRs/SGPRs 163 * cleaner between process switch), enable single-process mode 164 * in HWS. 165 * Note: the arbitration between the number of VMIDs and 166 * hws_max_conc_proc has been done in 167 * kgd2kfd_device_init(). 168 */ 169 concurrent_proc_cnt = (adev->enforce_isolation[kfd->node_id] == 170 AMDGPU_ENFORCE_ISOLATION_ENABLE) ? 171 1 : min(pm->dqm->processes_count, kfd->max_proc_per_quantum); 172 173 packet = (struct pm4_mes_runlist *)buffer; 174 175 memset(buffer, 0, sizeof(struct pm4_mes_runlist)); 176 packet->header.u32All = pm_build_pm4_header(IT_RUN_LIST, 177 sizeof(struct pm4_mes_runlist)); 178 179 packet->bitfields4.ib_size = ib_size_in_dwords; 180 packet->bitfields4.chain = chain ? 1 : 0; 181 packet->bitfields4.offload_polling = 0; 182 packet->bitfields4.chained_runlist_idle_disable = chain ? 1 : 0; 183 packet->bitfields4.valid = 1; 184 packet->bitfields4.process_cnt = concurrent_proc_cnt; 185 packet->ordinal2 = lower_32_bits(ib); 186 packet->ib_base_hi = upper_32_bits(ib); 187 188 return 0; 189 } 190 191 static int pm_set_resources_v9(struct packet_manager *pm, uint32_t *buffer, 192 struct scheduling_resources *res) 193 { 194 struct pm4_mes_set_resources *packet; 195 196 packet = (struct pm4_mes_set_resources *)buffer; 197 memset(buffer, 0, sizeof(struct pm4_mes_set_resources)); 198 199 packet->header.u32All = pm_build_pm4_header(IT_SET_RESOURCES, 200 sizeof(struct pm4_mes_set_resources)); 201 202 packet->bitfields2.queue_type = 203 queue_type__mes_set_resources__hsa_interface_queue_hiq; 204 packet->bitfields2.vmid_mask = res->vmid_mask; 205 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100; 206 if (pm->dqm->dev->adev->gmc.xnack_flags & AMDGPU_GMC_XNACK_FLAG_CHAIN) 207 packet->bitfields2.enb_xnack_retry_disable_check = 1; 208 packet->bitfields7.oac_mask = res->oac_mask; 209 packet->bitfields8.gds_heap_base = res->gds_heap_base; 210 packet->bitfields8.gds_heap_size = res->gds_heap_size; 211 212 packet->gws_mask_lo = lower_32_bits(res->gws_mask); 213 packet->gws_mask_hi = upper_32_bits(res->gws_mask); 214 215 packet->queue_mask_lo = lower_32_bits(res->queue_mask); 216 packet->queue_mask_hi = upper_32_bits(res->queue_mask); 217 218 return 0; 219 } 220 221 static inline bool pm_use_ext_eng(struct kfd_dev *dev) 222 { 223 return amdgpu_ip_version(dev->adev, SDMA0_HWIP, 0) >= 224 IP_VERSION(5, 2, 0); 225 } 226 227 static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, 228 struct queue *q, bool is_static) 229 { 230 struct pm4_mes_map_queues *packet; 231 232 packet = (struct pm4_mes_map_queues *)buffer; 233 memset(buffer, 0, sizeof(struct pm4_mes_map_queues)); 234 235 packet->header.u32All = pm_build_pm4_header(IT_MAP_QUEUES, 236 sizeof(struct pm4_mes_map_queues)); 237 packet->bitfields2.num_queues = 1; 238 packet->bitfields2.queue_sel = 239 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi; 240 241 packet->bitfields2.engine_sel = 242 engine_sel__mes_map_queues__compute_vi; 243 packet->bitfields2.gws_control_queue = q->properties.is_gws ? 1 : 0; 244 packet->bitfields2.extended_engine_sel = 245 extended_engine_sel__mes_map_queues__legacy_engine_sel; 246 packet->bitfields2.queue_type = 247 queue_type__mes_map_queues__normal_compute_vi; 248 249 switch (q->properties.type) { 250 case KFD_QUEUE_TYPE_COMPUTE: 251 if (is_static) 252 packet->bitfields2.queue_type = 253 queue_type__mes_map_queues__normal_latency_static_queue_vi; 254 break; 255 case KFD_QUEUE_TYPE_DIQ: 256 packet->bitfields2.queue_type = 257 queue_type__mes_map_queues__debug_interface_queue_vi; 258 break; 259 case KFD_QUEUE_TYPE_SDMA: 260 case KFD_QUEUE_TYPE_SDMA_XGMI: 261 if (q->properties.sdma_engine_id < 2 && 262 !pm_use_ext_eng(q->device->kfd)) 263 packet->bitfields2.engine_sel = q->properties.sdma_engine_id + 264 engine_sel__mes_map_queues__sdma0_vi; 265 else { 266 /* 267 * For GFX9.4.3, SDMA engine id can be greater than 8. 268 * For such cases, set extended_engine_sel to 2 and 269 * ensure engine_sel lies between 0-7. 270 */ 271 if (q->properties.sdma_engine_id >= 8) 272 packet->bitfields2.extended_engine_sel = 273 extended_engine_sel__mes_map_queues__sdma8_to_15_sel; 274 else 275 packet->bitfields2.extended_engine_sel = 276 extended_engine_sel__mes_map_queues__sdma0_to_7_sel; 277 278 packet->bitfields2.engine_sel = q->properties.sdma_engine_id % 8; 279 } 280 break; 281 default: 282 WARN(1, "queue type %d", q->properties.type); 283 return -EINVAL; 284 } 285 packet->bitfields3.doorbell_offset = 286 q->properties.doorbell_off; 287 288 packet->mqd_addr_lo = 289 lower_32_bits(q->gart_mqd_addr); 290 291 packet->mqd_addr_hi = 292 upper_32_bits(q->gart_mqd_addr); 293 294 packet->wptr_addr_lo = 295 lower_32_bits((uint64_t)q->properties.write_ptr); 296 297 packet->wptr_addr_hi = 298 upper_32_bits((uint64_t)q->properties.write_ptr); 299 300 return 0; 301 } 302 303 static inline void pm_build_dequeue_wait_counts_packet_info(struct packet_manager *pm, 304 uint32_t sch_value, uint32_t que_sleep, uint32_t *reg_offset, 305 uint32_t *reg_data) 306 { 307 pm->dqm->dev->kfd2kgd->build_dequeue_wait_counts_packet_info( 308 pm->dqm->dev->adev, 309 pm->dqm->wait_times, 310 sch_value, 311 que_sleep, 312 reg_offset, 313 reg_data); 314 } 315 316 /* pm_config_dequeue_wait_counts_v9: Builds WRITE_DATA packet with 317 * register/value for configuring dequeue wait counts 318 * 319 * @return: -ve for failure and 0 for success and buffer is 320 * filled in with packet 321 * 322 **/ 323 static int pm_config_dequeue_wait_counts_v9(struct packet_manager *pm, 324 uint32_t *buffer, 325 enum kfd_config_dequeue_wait_counts_cmd cmd, 326 uint32_t value) 327 { 328 struct pm4_mec_write_data_mmio *packet; 329 uint32_t reg_offset = 0; 330 uint32_t reg_data = 0; 331 332 switch (cmd) { 333 case KFD_DEQUEUE_WAIT_INIT: { 334 uint32_t sch_wave = 0, que_sleep = 1; 335 336 /* For all gfx9 ASICs > gfx941, 337 * Reduce CP_IQ_WAIT_TIME2.QUE_SLEEP to 0x1 from default 0x40. 338 * On a 1GHz machine this is roughly 1 microsecond, which is 339 * about how long it takes to load data out of memory during 340 * queue connect 341 * QUE_SLEEP: Wait Count for Dequeue Retry. 342 * 343 * Set CWSR grace period to 1x1000 cycle for GFX9.4.3 APU 344 */ 345 if (KFD_GC_VERSION(pm->dqm->dev) < IP_VERSION(9, 4, 1) || 346 KFD_GC_VERSION(pm->dqm->dev) >= IP_VERSION(10, 0, 0)) 347 return -EPERM; 348 349 if (amdgpu_emu_mode == 0 && pm->dqm->dev->adev->gmc.is_app_apu && 350 (KFD_GC_VERSION(pm->dqm->dev) == IP_VERSION(9, 4, 3))) 351 sch_wave = 1; 352 353 pm_build_dequeue_wait_counts_packet_info(pm, sch_wave, que_sleep, 354 ®_offset, ®_data); 355 356 break; 357 } 358 case KFD_DEQUEUE_WAIT_RESET: 359 /* reg_data would be set to dqm->wait_times */ 360 pm_build_dequeue_wait_counts_packet_info(pm, 0, 0, ®_offset, ®_data); 361 break; 362 363 case KFD_DEQUEUE_WAIT_SET_SCH_WAVE: 364 /* The CP cannot handle value 0 and it will result in 365 * an infinite grace period being set so set to 1 to prevent this. Also 366 * avoid debugger API breakage as it sets 0 and expects a low value. 367 */ 368 if (!value) 369 value = 1; 370 pm_build_dequeue_wait_counts_packet_info(pm, value, 0, ®_offset, ®_data); 371 break; 372 default: 373 pr_err("Invalid dequeue wait cmd\n"); 374 return -EINVAL; 375 } 376 377 packet = (struct pm4_mec_write_data_mmio *)buffer; 378 memset(buffer, 0, sizeof(struct pm4_mec_write_data_mmio)); 379 380 packet->header.u32All = pm_build_pm4_header(IT_WRITE_DATA, 381 sizeof(struct pm4_mec_write_data_mmio)); 382 383 packet->bitfields2.dst_sel = dst_sel___write_data__mem_mapped_register; 384 packet->bitfields2.addr_incr = 385 addr_incr___write_data__do_not_increment_address; 386 387 packet->bitfields3.dst_mmreg_addr = reg_offset; 388 389 packet->data = reg_data; 390 391 return 0; 392 } 393 394 static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, 395 enum kfd_unmap_queues_filter filter, 396 uint32_t filter_param, bool reset) 397 { 398 struct pm4_mes_unmap_queues *packet; 399 400 packet = (struct pm4_mes_unmap_queues *)buffer; 401 memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues)); 402 403 packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES, 404 sizeof(struct pm4_mes_unmap_queues)); 405 406 packet->bitfields2.extended_engine_sel = 407 pm_use_ext_eng(pm->dqm->dev->kfd) ? 408 extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel : 409 extended_engine_sel__mes_unmap_queues__legacy_engine_sel; 410 411 packet->bitfields2.engine_sel = 412 engine_sel__mes_unmap_queues__compute; 413 414 if (reset) 415 packet->bitfields2.action = 416 action__mes_unmap_queues__reset_queues; 417 else 418 packet->bitfields2.action = 419 action__mes_unmap_queues__preempt_queues; 420 421 switch (filter) { 422 case KFD_UNMAP_QUEUES_FILTER_BY_PASID: 423 packet->bitfields2.queue_sel = 424 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues; 425 packet->bitfields3a.pasid = filter_param; 426 break; 427 case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: 428 packet->bitfields2.queue_sel = 429 queue_sel__mes_unmap_queues__unmap_all_queues; 430 break; 431 case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: 432 /* in this case, we do not preempt static queues */ 433 packet->bitfields2.queue_sel = 434 queue_sel__mes_unmap_queues__unmap_all_non_static_queues; 435 break; 436 default: 437 WARN(1, "filter %d", filter); 438 return -EINVAL; 439 } 440 441 return 0; 442 443 } 444 445 static int pm_query_status_v9(struct packet_manager *pm, uint32_t *buffer, 446 uint64_t fence_address, uint64_t fence_value) 447 { 448 struct pm4_mes_query_status *packet; 449 450 packet = (struct pm4_mes_query_status *)buffer; 451 memset(buffer, 0, sizeof(struct pm4_mes_query_status)); 452 453 454 packet->header.u32All = pm_build_pm4_header(IT_QUERY_STATUS, 455 sizeof(struct pm4_mes_query_status)); 456 457 packet->bitfields2.context_id = 0; 458 packet->bitfields2.interrupt_sel = 459 interrupt_sel__mes_query_status__completion_status; 460 packet->bitfields2.command = 461 command__mes_query_status__fence_only_after_write_ack; 462 463 packet->addr_hi = upper_32_bits((uint64_t)fence_address); 464 packet->addr_lo = lower_32_bits((uint64_t)fence_address); 465 packet->data_hi = upper_32_bits((uint64_t)fence_value); 466 packet->data_lo = lower_32_bits((uint64_t)fence_value); 467 468 return 0; 469 } 470 471 const struct packet_manager_funcs kfd_v9_pm_funcs = { 472 .map_process = pm_map_process_v9, 473 .runlist = pm_runlist_v9, 474 .set_resources = pm_set_resources_v9, 475 .map_queues = pm_map_queues_v9, 476 .unmap_queues = pm_unmap_queues_v9, 477 .config_dequeue_wait_counts = pm_config_dequeue_wait_counts_v9, 478 .query_status = pm_query_status_v9, 479 .release_mem = NULL, 480 .map_process_size = sizeof(struct pm4_mes_map_process), 481 .runlist_size = sizeof(struct pm4_mes_runlist), 482 .set_resources_size = sizeof(struct pm4_mes_set_resources), 483 .map_queues_size = sizeof(struct pm4_mes_map_queues), 484 .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), 485 .config_dequeue_wait_counts_size = sizeof(struct pm4_mec_write_data_mmio), 486 .query_status_size = sizeof(struct pm4_mes_query_status), 487 .release_mem_size = 0, 488 }; 489 490 const struct packet_manager_funcs kfd_aldebaran_pm_funcs = { 491 .map_process = pm_map_process_aldebaran, 492 .runlist = pm_runlist_v9, 493 .set_resources = pm_set_resources_v9, 494 .map_queues = pm_map_queues_v9, 495 .unmap_queues = pm_unmap_queues_v9, 496 .config_dequeue_wait_counts = pm_config_dequeue_wait_counts_v9, 497 .query_status = pm_query_status_v9, 498 .release_mem = NULL, 499 .map_process_size = sizeof(struct pm4_mes_map_process_aldebaran), 500 .runlist_size = sizeof(struct pm4_mes_runlist), 501 .set_resources_size = sizeof(struct pm4_mes_set_resources), 502 .map_queues_size = sizeof(struct pm4_mes_map_queues), 503 .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), 504 .config_dequeue_wait_counts_size = sizeof(struct pm4_mec_write_data_mmio), 505 .query_status_size = sizeof(struct pm4_mes_query_status), 506 .release_mem_size = 0, 507 }; 508