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 25 #include <linux/ratelimit.h> 26 #include <linux/printk.h> 27 #include <linux/slab.h> 28 #include <linux/list.h> 29 #include <linux/types.h> 30 #include <linux/bitops.h> 31 #include <linux/sched.h> 32 #include "kfd_priv.h" 33 #include "kfd_device_queue_manager.h" 34 #include "kfd_mqd_manager.h" 35 #include "cik_regs.h" 36 #include "kfd_kernel_queue.h" 37 #include "amdgpu_amdkfd.h" 38 #include "amdgpu_reset.h" 39 #include "amdgpu_sdma.h" 40 #include "mes_v11_api_def.h" 41 #include "kfd_debug.h" 42 43 /* Size of the per-pipe EOP queue */ 44 #define CIK_HPD_EOP_BYTES_LOG2 11 45 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2) 46 /* See unmap_queues_cpsch() */ 47 #define USE_DEFAULT_GRACE_PERIOD 0xffffffff 48 49 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, 50 u32 pasid, unsigned int vmid); 51 52 static int execute_queues_cpsch(struct device_queue_manager *dqm, 53 enum kfd_unmap_queues_filter filter, 54 uint32_t filter_param, 55 uint32_t grace_period); 56 static int unmap_queues_cpsch(struct device_queue_manager *dqm, 57 enum kfd_unmap_queues_filter filter, 58 uint32_t filter_param, 59 uint32_t grace_period, 60 bool reset); 61 62 static int map_queues_cpsch(struct device_queue_manager *dqm); 63 64 static void deallocate_sdma_queue(struct device_queue_manager *dqm, 65 struct queue *q); 66 67 static inline void deallocate_hqd(struct device_queue_manager *dqm, 68 struct queue *q); 69 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q); 70 static int allocate_sdma_queue(struct device_queue_manager *dqm, 71 struct queue *q, const uint32_t *restore_sdma_id); 72 73 static int reset_queues_on_hws_hang(struct device_queue_manager *dqm, bool is_sdma); 74 75 static inline 76 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type) 77 { 78 if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI) 79 return KFD_MQD_TYPE_SDMA; 80 return KFD_MQD_TYPE_CP; 81 } 82 83 static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe) 84 { 85 int i; 86 int pipe_offset = (mec * dqm->dev->kfd->shared_resources.num_pipe_per_mec 87 + pipe) * dqm->dev->kfd->shared_resources.num_queue_per_pipe; 88 89 /* queue is available for KFD usage if bit is 1 */ 90 for (i = 0; i < dqm->dev->kfd->shared_resources.num_queue_per_pipe; ++i) 91 if (test_bit(pipe_offset + i, 92 dqm->dev->kfd->shared_resources.cp_queue_bitmap)) 93 return true; 94 return false; 95 } 96 97 unsigned int get_cp_queues_num(struct device_queue_manager *dqm) 98 { 99 return bitmap_weight(dqm->dev->kfd->shared_resources.cp_queue_bitmap, 100 AMDGPU_MAX_QUEUES); 101 } 102 103 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm) 104 { 105 return dqm->dev->kfd->shared_resources.num_queue_per_pipe; 106 } 107 108 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm) 109 { 110 return dqm->dev->kfd->shared_resources.num_pipe_per_mec; 111 } 112 113 static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm) 114 { 115 return kfd_get_num_sdma_engines(dqm->dev) + 116 kfd_get_num_xgmi_sdma_engines(dqm->dev); 117 } 118 119 unsigned int get_num_sdma_queues(struct device_queue_manager *dqm) 120 { 121 return kfd_get_num_sdma_engines(dqm->dev) * 122 dqm->dev->kfd->device_info.num_sdma_queues_per_engine; 123 } 124 125 unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) 126 { 127 return kfd_get_num_xgmi_sdma_engines(dqm->dev) * 128 dqm->dev->kfd->device_info.num_sdma_queues_per_engine; 129 } 130 131 static void init_sdma_bitmaps(struct device_queue_manager *dqm) 132 { 133 bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES); 134 bitmap_set(dqm->sdma_bitmap, 0, get_num_sdma_queues(dqm)); 135 136 bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES); 137 bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm)); 138 139 /* Mask out the reserved queues */ 140 bitmap_andnot(dqm->sdma_bitmap, dqm->sdma_bitmap, 141 dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap, 142 KFD_MAX_SDMA_QUEUES); 143 } 144 145 void program_sh_mem_settings(struct device_queue_manager *dqm, 146 struct qcm_process_device *qpd) 147 { 148 uint32_t xcc_mask = dqm->dev->xcc_mask; 149 int xcc_id; 150 151 for_each_inst(xcc_id, xcc_mask) 152 dqm->dev->kfd2kgd->program_sh_mem_settings( 153 dqm->dev->adev, qpd->vmid, qpd->sh_mem_config, 154 qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit, 155 qpd->sh_mem_bases, xcc_id); 156 } 157 158 static void kfd_hws_hang(struct device_queue_manager *dqm) 159 { 160 struct device_process_node *cur; 161 struct qcm_process_device *qpd; 162 struct queue *q; 163 164 /* Mark all device queues as reset. */ 165 list_for_each_entry(cur, &dqm->queues, list) { 166 qpd = cur->qpd; 167 list_for_each_entry(q, &qpd->queues_list, list) { 168 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 169 170 pdd->has_reset_queue = true; 171 } 172 } 173 174 /* 175 * Issue a GPU reset if HWS is unresponsive 176 */ 177 amdgpu_amdkfd_gpu_reset(dqm->dev->adev); 178 } 179 180 static int convert_to_mes_queue_type(int queue_type) 181 { 182 int mes_queue_type; 183 184 switch (queue_type) { 185 case KFD_QUEUE_TYPE_COMPUTE: 186 mes_queue_type = MES_QUEUE_TYPE_COMPUTE; 187 break; 188 case KFD_QUEUE_TYPE_SDMA: 189 mes_queue_type = MES_QUEUE_TYPE_SDMA; 190 break; 191 default: 192 WARN(1, "Invalid queue type %d", queue_type); 193 mes_queue_type = -EINVAL; 194 break; 195 } 196 197 return mes_queue_type; 198 } 199 200 static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, 201 struct qcm_process_device *qpd) 202 { 203 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; 204 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 205 struct mes_add_queue_input queue_input; 206 int r, queue_type; 207 uint64_t wptr_addr_off; 208 209 if (!dqm->sched_running || dqm->sched_halt) 210 return 0; 211 if (!down_read_trylock(&adev->reset_domain->sem)) 212 return -EIO; 213 214 memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input)); 215 queue_input.process_id = pdd->pasid; 216 queue_input.page_table_base_addr = qpd->page_table_base; 217 queue_input.process_va_start = 0; 218 queue_input.process_va_end = adev->vm_manager.max_pfn - 1; 219 /* MES unit for quantum is 100ns */ 220 queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM; /* Equivalent to 10ms. */ 221 queue_input.process_context_addr = pdd->proc_ctx_gpu_addr; 222 queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */ 223 queue_input.gang_context_addr = q->gang_ctx_gpu_addr; 224 queue_input.inprocess_gang_priority = q->properties.priority; 225 queue_input.gang_global_priority_level = 226 AMDGPU_MES_PRIORITY_LEVEL_NORMAL; 227 queue_input.doorbell_offset = q->properties.doorbell_off; 228 queue_input.mqd_addr = q->gart_mqd_addr; 229 queue_input.wptr_addr = (uint64_t)q->properties.write_ptr; 230 231 wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1); 232 queue_input.wptr_mc_addr = amdgpu_bo_gpu_offset(q->properties.wptr_bo) + wptr_addr_off; 233 234 queue_input.is_kfd_process = 1; 235 queue_input.is_aql_queue = (q->properties.format == KFD_QUEUE_FORMAT_AQL); 236 queue_input.queue_size = q->properties.queue_size >> 2; 237 238 queue_input.paging = false; 239 queue_input.tba_addr = qpd->tba_addr; 240 queue_input.tma_addr = qpd->tma_addr; 241 queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device); 242 queue_input.skip_process_ctx_clear = 243 qpd->pqm->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED && 244 (qpd->pqm->process->debug_trap_enabled || 245 kfd_dbg_has_ttmps_always_setup(q->device)); 246 247 queue_type = convert_to_mes_queue_type(q->properties.type); 248 if (queue_type < 0) { 249 dev_err(adev->dev, "Queue type not supported with MES, queue:%d\n", 250 q->properties.type); 251 up_read(&adev->reset_domain->sem); 252 return -EINVAL; 253 } 254 queue_input.queue_type = (uint32_t)queue_type; 255 256 queue_input.exclusively_scheduled = q->properties.is_gws; 257 258 amdgpu_mes_lock(&adev->mes); 259 r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input); 260 amdgpu_mes_unlock(&adev->mes); 261 up_read(&adev->reset_domain->sem); 262 if (r) { 263 dev_err(adev->dev, "failed to add hardware queue to MES, doorbell=0x%x\n", 264 q->properties.doorbell_off); 265 dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n"); 266 kfd_hws_hang(dqm); 267 } 268 269 return r; 270 } 271 272 static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q, 273 struct qcm_process_device *qpd) 274 { 275 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; 276 int r; 277 struct mes_remove_queue_input queue_input; 278 279 if (!dqm->sched_running || dqm->sched_halt) 280 return 0; 281 if (!down_read_trylock(&adev->reset_domain->sem)) 282 return -EIO; 283 284 memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input)); 285 queue_input.doorbell_offset = q->properties.doorbell_off; 286 queue_input.gang_context_addr = q->gang_ctx_gpu_addr; 287 288 amdgpu_mes_lock(&adev->mes); 289 r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input); 290 amdgpu_mes_unlock(&adev->mes); 291 up_read(&adev->reset_domain->sem); 292 293 if (r) { 294 dev_err(adev->dev, "failed to remove hardware queue from MES, doorbell=0x%x\n", 295 q->properties.doorbell_off); 296 dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n"); 297 kfd_hws_hang(dqm); 298 } 299 300 return r; 301 } 302 303 static int remove_all_kfd_queues_mes(struct device_queue_manager *dqm) 304 { 305 struct device_process_node *cur; 306 struct device *dev = dqm->dev->adev->dev; 307 struct qcm_process_device *qpd; 308 struct queue *q; 309 int retval = 0; 310 311 list_for_each_entry(cur, &dqm->queues, list) { 312 qpd = cur->qpd; 313 list_for_each_entry(q, &qpd->queues_list, list) { 314 if (q->properties.is_active) { 315 retval = remove_queue_mes(dqm, q, qpd); 316 if (retval) { 317 dev_err(dev, "%s: Failed to remove queue %d for dev %d", 318 __func__, 319 q->properties.queue_id, 320 dqm->dev->id); 321 return retval; 322 } 323 } 324 } 325 } 326 327 return retval; 328 } 329 330 static int add_all_kfd_queues_mes(struct device_queue_manager *dqm) 331 { 332 struct device_process_node *cur; 333 struct device *dev = dqm->dev->adev->dev; 334 struct qcm_process_device *qpd; 335 struct queue *q; 336 int retval = 0; 337 338 list_for_each_entry(cur, &dqm->queues, list) { 339 qpd = cur->qpd; 340 list_for_each_entry(q, &qpd->queues_list, list) { 341 if (!q->properties.is_active) 342 continue; 343 retval = add_queue_mes(dqm, q, qpd); 344 if (retval) { 345 dev_err(dev, "%s: Failed to add queue %d for dev %d", 346 __func__, 347 q->properties.queue_id, 348 dqm->dev->id); 349 return retval; 350 } 351 } 352 } 353 354 return retval; 355 } 356 357 static int suspend_all_queues_mes(struct device_queue_manager *dqm) 358 { 359 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; 360 int r = 0; 361 362 if (!down_read_trylock(&adev->reset_domain->sem)) 363 return -EIO; 364 365 r = amdgpu_mes_suspend(adev); 366 up_read(&adev->reset_domain->sem); 367 368 if (r) { 369 dev_err(adev->dev, "failed to suspend gangs from MES\n"); 370 dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n"); 371 kfd_hws_hang(dqm); 372 } 373 374 return r; 375 } 376 377 static int resume_all_queues_mes(struct device_queue_manager *dqm) 378 { 379 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; 380 int r = 0; 381 382 if (!down_read_trylock(&adev->reset_domain->sem)) 383 return -EIO; 384 385 r = amdgpu_mes_resume(adev); 386 up_read(&adev->reset_domain->sem); 387 388 if (r) { 389 dev_err(adev->dev, "failed to resume gangs from MES\n"); 390 dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n"); 391 kfd_hws_hang(dqm); 392 } 393 394 return r; 395 } 396 397 static void increment_queue_count(struct device_queue_manager *dqm, 398 struct qcm_process_device *qpd, 399 struct queue *q) 400 { 401 dqm->active_queue_count++; 402 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || 403 q->properties.type == KFD_QUEUE_TYPE_DIQ) 404 dqm->active_cp_queue_count++; 405 406 if (q->properties.is_gws) { 407 dqm->gws_queue_count++; 408 qpd->mapped_gws_queue = true; 409 } 410 } 411 412 static void decrement_queue_count(struct device_queue_manager *dqm, 413 struct qcm_process_device *qpd, 414 struct queue *q) 415 { 416 dqm->active_queue_count--; 417 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || 418 q->properties.type == KFD_QUEUE_TYPE_DIQ) 419 dqm->active_cp_queue_count--; 420 421 if (q->properties.is_gws) { 422 dqm->gws_queue_count--; 423 qpd->mapped_gws_queue = false; 424 } 425 } 426 427 /* 428 * Allocate a doorbell ID to this queue. 429 * If doorbell_id is passed in, make sure requested ID is valid then allocate it. 430 */ 431 static int allocate_doorbell(struct qcm_process_device *qpd, 432 struct queue *q, 433 uint32_t const *restore_id) 434 { 435 struct kfd_node *dev = qpd->dqm->dev; 436 437 if (!KFD_IS_SOC15(dev)) { 438 /* On pre-SOC15 chips we need to use the queue ID to 439 * preserve the user mode ABI. 440 */ 441 442 if (restore_id && *restore_id != q->properties.queue_id) 443 return -EINVAL; 444 445 q->doorbell_id = q->properties.queue_id; 446 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 447 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 448 /* For SDMA queues on SOC15 with 8-byte doorbell, use static 449 * doorbell assignments based on the engine and queue id. 450 * The doobell index distance between RLC (2*i) and (2*i+1) 451 * for a SDMA engine is 512. 452 */ 453 454 uint32_t *idx_offset = dev->kfd->shared_resources.sdma_doorbell_idx; 455 456 /* 457 * q->properties.sdma_engine_id corresponds to the virtual 458 * sdma engine number. However, for doorbell allocation, 459 * we need the physical sdma engine id in order to get the 460 * correct doorbell offset. 461 */ 462 uint32_t valid_id = idx_offset[qpd->dqm->dev->node_id * 463 get_num_all_sdma_engines(qpd->dqm) + 464 q->properties.sdma_engine_id] 465 + (q->properties.sdma_queue_id & 1) 466 * KFD_QUEUE_DOORBELL_MIRROR_OFFSET 467 + (q->properties.sdma_queue_id >> 1); 468 469 if (restore_id && *restore_id != valid_id) 470 return -EINVAL; 471 q->doorbell_id = valid_id; 472 } else { 473 /* For CP queues on SOC15 */ 474 if (restore_id) { 475 /* make sure that ID is free */ 476 if (__test_and_set_bit(*restore_id, qpd->doorbell_bitmap)) 477 return -EINVAL; 478 479 q->doorbell_id = *restore_id; 480 } else { 481 /* or reserve a free doorbell ID */ 482 unsigned int found; 483 484 found = find_first_zero_bit(qpd->doorbell_bitmap, 485 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 486 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { 487 pr_debug("No doorbells available"); 488 return -EBUSY; 489 } 490 set_bit(found, qpd->doorbell_bitmap); 491 q->doorbell_id = found; 492 } 493 } 494 495 q->properties.doorbell_off = amdgpu_doorbell_index_on_bar(dev->adev, 496 qpd->proc_doorbells, 497 q->doorbell_id, 498 dev->kfd->device_info.doorbell_size); 499 return 0; 500 } 501 502 static void deallocate_doorbell(struct qcm_process_device *qpd, 503 struct queue *q) 504 { 505 unsigned int old; 506 struct kfd_node *dev = qpd->dqm->dev; 507 508 if (!KFD_IS_SOC15(dev) || 509 q->properties.type == KFD_QUEUE_TYPE_SDMA || 510 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) 511 return; 512 513 old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap); 514 WARN_ON(!old); 515 } 516 517 static void program_trap_handler_settings(struct device_queue_manager *dqm, 518 struct qcm_process_device *qpd) 519 { 520 uint32_t xcc_mask = dqm->dev->xcc_mask; 521 int xcc_id; 522 523 if (dqm->dev->kfd2kgd->program_trap_handler_settings) 524 for_each_inst(xcc_id, xcc_mask) 525 dqm->dev->kfd2kgd->program_trap_handler_settings( 526 dqm->dev->adev, qpd->vmid, qpd->tba_addr, 527 qpd->tma_addr, xcc_id); 528 } 529 530 static int allocate_vmid(struct device_queue_manager *dqm, 531 struct qcm_process_device *qpd, 532 struct queue *q) 533 { 534 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 535 struct device *dev = dqm->dev->adev->dev; 536 int allocated_vmid = -1, i; 537 538 for (i = dqm->dev->vm_info.first_vmid_kfd; 539 i <= dqm->dev->vm_info.last_vmid_kfd; i++) { 540 if (!dqm->vmid_pasid[i]) { 541 allocated_vmid = i; 542 break; 543 } 544 } 545 546 if (allocated_vmid < 0) { 547 dev_err(dev, "no more vmid to allocate\n"); 548 return -ENOSPC; 549 } 550 551 pr_debug("vmid allocated: %d\n", allocated_vmid); 552 553 dqm->vmid_pasid[allocated_vmid] = pdd->pasid; 554 555 set_pasid_vmid_mapping(dqm, pdd->pasid, allocated_vmid); 556 557 qpd->vmid = allocated_vmid; 558 q->properties.vmid = allocated_vmid; 559 560 program_sh_mem_settings(dqm, qpd); 561 562 if (KFD_IS_SOC15(dqm->dev) && dqm->dev->kfd->cwsr_enabled) 563 program_trap_handler_settings(dqm, qpd); 564 565 /* qpd->page_table_base is set earlier when register_process() 566 * is called, i.e. when the first queue is created. 567 */ 568 dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->adev, 569 qpd->vmid, 570 qpd->page_table_base); 571 /* invalidate the VM context after pasid and vmid mapping is set up */ 572 kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY); 573 574 if (dqm->dev->kfd2kgd->set_scratch_backing_va) 575 dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->adev, 576 qpd->sh_hidden_private_base, qpd->vmid); 577 578 return 0; 579 } 580 581 static int flush_texture_cache_nocpsch(struct kfd_node *kdev, 582 struct qcm_process_device *qpd) 583 { 584 const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf; 585 int ret; 586 587 if (!qpd->ib_kaddr) 588 return -ENOMEM; 589 590 ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr); 591 if (ret) 592 return ret; 593 594 return amdgpu_amdkfd_submit_ib(kdev->adev, KGD_ENGINE_MEC1, qpd->vmid, 595 qpd->ib_base, (uint32_t *)qpd->ib_kaddr, 596 pmf->release_mem_size / sizeof(uint32_t)); 597 } 598 599 static void deallocate_vmid(struct device_queue_manager *dqm, 600 struct qcm_process_device *qpd, 601 struct queue *q) 602 { 603 struct device *dev = dqm->dev->adev->dev; 604 605 /* On GFX v7, CP doesn't flush TC at dequeue */ 606 if (q->device->adev->asic_type == CHIP_HAWAII) 607 if (flush_texture_cache_nocpsch(q->device, qpd)) 608 dev_err(dev, "Failed to flush TC\n"); 609 610 kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY); 611 612 /* Release the vmid mapping */ 613 set_pasid_vmid_mapping(dqm, 0, qpd->vmid); 614 dqm->vmid_pasid[qpd->vmid] = 0; 615 616 qpd->vmid = 0; 617 q->properties.vmid = 0; 618 } 619 620 static int create_queue_nocpsch(struct device_queue_manager *dqm, 621 struct queue *q, 622 struct qcm_process_device *qpd, 623 const struct kfd_criu_queue_priv_data *qd, 624 const void *restore_mqd, const void *restore_ctl_stack) 625 { 626 struct mqd_manager *mqd_mgr; 627 int retval; 628 629 dqm_lock(dqm); 630 631 if (dqm->total_queue_count >= max_num_of_queues_per_device) { 632 pr_warn("Can't create new usermode queue because %d queues were already created\n", 633 dqm->total_queue_count); 634 retval = -EPERM; 635 goto out_unlock; 636 } 637 638 if (list_empty(&qpd->queues_list)) { 639 retval = allocate_vmid(dqm, qpd, q); 640 if (retval) 641 goto out_unlock; 642 } 643 q->properties.vmid = qpd->vmid; 644 /* 645 * Eviction state logic: mark all queues as evicted, even ones 646 * not currently active. Restoring inactive queues later only 647 * updates the is_evicted flag but is a no-op otherwise. 648 */ 649 q->properties.is_evicted = !!qpd->evicted; 650 651 q->properties.tba_addr = qpd->tba_addr; 652 q->properties.tma_addr = qpd->tma_addr; 653 654 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 655 q->properties.type)]; 656 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { 657 retval = allocate_hqd(dqm, q); 658 if (retval) 659 goto deallocate_vmid; 660 pr_debug("Loading mqd to hqd on pipe %d, queue %d\n", 661 q->pipe, q->queue); 662 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 663 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 664 retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL); 665 if (retval) 666 goto deallocate_vmid; 667 dqm->asic_ops.init_sdma_vm(dqm, q, qpd); 668 } 669 670 retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL); 671 if (retval) 672 goto out_deallocate_hqd; 673 674 /* Temporarily release dqm lock to avoid a circular lock dependency */ 675 dqm_unlock(dqm); 676 q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); 677 dqm_lock(dqm); 678 679 if (!q->mqd_mem_obj) { 680 retval = -ENOMEM; 681 goto out_deallocate_doorbell; 682 } 683 684 if (qd) 685 mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, 686 &q->properties, restore_mqd, restore_ctl_stack, 687 qd->ctl_stack_size); 688 else 689 mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, 690 &q->gart_mqd_addr, &q->properties); 691 692 if (q->properties.is_active) { 693 if (!dqm->sched_running) { 694 WARN_ONCE(1, "Load non-HWS mqd while stopped\n"); 695 goto add_queue_to_list; 696 } 697 698 if (WARN(q->process->mm != current->mm, 699 "should only run in user thread")) 700 retval = -EFAULT; 701 else 702 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, 703 q->queue, &q->properties, current->mm); 704 if (retval) 705 goto out_free_mqd; 706 } 707 708 add_queue_to_list: 709 list_add(&q->list, &qpd->queues_list); 710 qpd->queue_count++; 711 if (q->properties.is_active) 712 increment_queue_count(dqm, qpd, q); 713 714 /* 715 * Unconditionally increment this counter, regardless of the queue's 716 * type or whether the queue is active. 717 */ 718 dqm->total_queue_count++; 719 pr_debug("Total of %d queues are accountable so far\n", 720 dqm->total_queue_count); 721 goto out_unlock; 722 723 out_free_mqd: 724 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 725 out_deallocate_doorbell: 726 deallocate_doorbell(qpd, q); 727 out_deallocate_hqd: 728 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) 729 deallocate_hqd(dqm, q); 730 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 731 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) 732 deallocate_sdma_queue(dqm, q); 733 deallocate_vmid: 734 if (list_empty(&qpd->queues_list)) 735 deallocate_vmid(dqm, qpd, q); 736 out_unlock: 737 dqm_unlock(dqm); 738 return retval; 739 } 740 741 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q) 742 { 743 bool set; 744 int pipe, bit, i; 745 746 set = false; 747 748 for (pipe = dqm->next_pipe_to_allocate, i = 0; 749 i < get_pipes_per_mec(dqm); 750 pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) { 751 752 if (!is_pipe_enabled(dqm, 0, pipe)) 753 continue; 754 755 if (dqm->allocated_queues[pipe] != 0) { 756 bit = ffs(dqm->allocated_queues[pipe]) - 1; 757 dqm->allocated_queues[pipe] &= ~(1 << bit); 758 q->pipe = pipe; 759 q->queue = bit; 760 set = true; 761 break; 762 } 763 } 764 765 if (!set) 766 return -EBUSY; 767 768 pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue); 769 /* horizontal hqd allocation */ 770 dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm); 771 772 return 0; 773 } 774 775 static inline void deallocate_hqd(struct device_queue_manager *dqm, 776 struct queue *q) 777 { 778 dqm->allocated_queues[q->pipe] |= (1 << q->queue); 779 } 780 781 #define SQ_IND_CMD_CMD_KILL 0x00000003 782 #define SQ_IND_CMD_MODE_BROADCAST 0x00000001 783 784 static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process *p) 785 { 786 int status = 0; 787 unsigned int vmid; 788 uint16_t queried_pasid; 789 union SQ_CMD_BITS reg_sq_cmd; 790 union GRBM_GFX_INDEX_BITS reg_gfx_index; 791 struct kfd_process_device *pdd; 792 int first_vmid_to_scan = dev->vm_info.first_vmid_kfd; 793 int last_vmid_to_scan = dev->vm_info.last_vmid_kfd; 794 uint32_t xcc_mask = dev->xcc_mask; 795 int xcc_id; 796 797 reg_sq_cmd.u32All = 0; 798 reg_gfx_index.u32All = 0; 799 800 pr_debug("Killing all process wavefronts\n"); 801 802 if (!dev->kfd2kgd->get_atc_vmid_pasid_mapping_info) { 803 dev_err(dev->adev->dev, "no vmid pasid mapping supported\n"); 804 return -EOPNOTSUPP; 805 } 806 807 /* taking the VMID for that process on the safe way using PDD */ 808 pdd = kfd_get_process_device_data(dev, p); 809 if (!pdd) 810 return -EFAULT; 811 812 /* Scan all registers in the range ATC_VMID8_PASID_MAPPING .. 813 * ATC_VMID15_PASID_MAPPING 814 * to check which VMID the current process is mapped to. 815 */ 816 817 for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) { 818 status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info 819 (dev->adev, vmid, &queried_pasid); 820 821 if (status && queried_pasid == pdd->pasid) { 822 pr_debug("Killing wave fronts of vmid %d and process pid %d\n", 823 vmid, p->lead_thread->pid); 824 break; 825 } 826 } 827 828 if (vmid > last_vmid_to_scan) { 829 dev_err(dev->adev->dev, "Didn't find vmid for process pid %d\n", 830 p->lead_thread->pid); 831 return -EFAULT; 832 } 833 834 reg_gfx_index.bits.sh_broadcast_writes = 1; 835 reg_gfx_index.bits.se_broadcast_writes = 1; 836 reg_gfx_index.bits.instance_broadcast_writes = 1; 837 reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST; 838 reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL; 839 reg_sq_cmd.bits.vm_id = vmid; 840 841 for_each_inst(xcc_id, xcc_mask) 842 dev->kfd2kgd->wave_control_execute( 843 dev->adev, reg_gfx_index.u32All, 844 reg_sq_cmd.u32All, xcc_id); 845 846 return 0; 847 } 848 849 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked 850 * to avoid asynchronized access 851 */ 852 static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm, 853 struct qcm_process_device *qpd, 854 struct queue *q) 855 { 856 int retval; 857 struct mqd_manager *mqd_mgr; 858 859 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 860 q->properties.type)]; 861 862 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) 863 deallocate_hqd(dqm, q); 864 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) 865 deallocate_sdma_queue(dqm, q); 866 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) 867 deallocate_sdma_queue(dqm, q); 868 else { 869 pr_debug("q->properties.type %d is invalid\n", 870 q->properties.type); 871 return -EINVAL; 872 } 873 dqm->total_queue_count--; 874 875 deallocate_doorbell(qpd, q); 876 877 if (!dqm->sched_running) { 878 WARN_ONCE(1, "Destroy non-HWS queue while stopped\n"); 879 return 0; 880 } 881 882 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, 883 KFD_PREEMPT_TYPE_WAVEFRONT_RESET, 884 KFD_UNMAP_LATENCY_MS, 885 q->pipe, q->queue); 886 if (retval == -ETIME) 887 qpd->reset_wavefronts = true; 888 889 list_del(&q->list); 890 if (list_empty(&qpd->queues_list)) { 891 if (qpd->reset_wavefronts) { 892 pr_warn("Resetting wave fronts (nocpsch) on dev %p\n", 893 dqm->dev); 894 /* dbgdev_wave_reset_wavefronts has to be called before 895 * deallocate_vmid(), i.e. when vmid is still in use. 896 */ 897 dbgdev_wave_reset_wavefronts(dqm->dev, 898 qpd->pqm->process); 899 qpd->reset_wavefronts = false; 900 } 901 902 deallocate_vmid(dqm, qpd, q); 903 } 904 qpd->queue_count--; 905 if (q->properties.is_active) 906 decrement_queue_count(dqm, qpd, q); 907 908 return retval; 909 } 910 911 static int destroy_queue_nocpsch(struct device_queue_manager *dqm, 912 struct qcm_process_device *qpd, 913 struct queue *q) 914 { 915 int retval; 916 uint64_t sdma_val = 0; 917 struct device *dev = dqm->dev->adev->dev; 918 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 919 struct mqd_manager *mqd_mgr = 920 dqm->mqd_mgrs[get_mqd_type_from_queue_type(q->properties.type)]; 921 922 /* Get the SDMA queue stats */ 923 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || 924 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { 925 retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr, 926 &sdma_val); 927 if (retval) 928 dev_err(dev, "Failed to read SDMA queue counter for queue: %d\n", 929 q->properties.queue_id); 930 } 931 932 dqm_lock(dqm); 933 retval = destroy_queue_nocpsch_locked(dqm, qpd, q); 934 if (!retval) 935 pdd->sdma_past_activity_counter += sdma_val; 936 dqm_unlock(dqm); 937 938 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 939 940 return retval; 941 } 942 943 static int update_queue(struct device_queue_manager *dqm, struct queue *q, 944 struct mqd_update_info *minfo) 945 { 946 int retval = 0; 947 struct device *dev = dqm->dev->adev->dev; 948 struct mqd_manager *mqd_mgr; 949 struct kfd_process_device *pdd; 950 bool prev_active = false; 951 952 dqm_lock(dqm); 953 pdd = kfd_get_process_device_data(q->device, q->process); 954 if (!pdd) { 955 retval = -ENODEV; 956 goto out_unlock; 957 } 958 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 959 q->properties.type)]; 960 961 /* Save previous activity state for counters */ 962 prev_active = q->properties.is_active; 963 964 /* Make sure the queue is unmapped before updating the MQD */ 965 if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { 966 if (!dqm->dev->kfd->shared_resources.enable_mes) 967 retval = unmap_queues_cpsch(dqm, 968 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); 969 else if (prev_active) 970 retval = remove_queue_mes(dqm, q, &pdd->qpd); 971 972 /* queue is reset so inaccessable */ 973 if (pdd->has_reset_queue) { 974 retval = -EACCES; 975 goto out_unlock; 976 } 977 978 if (retval) { 979 dev_err(dev, "unmap queue failed\n"); 980 goto out_unlock; 981 } 982 } else if (prev_active && 983 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || 984 q->properties.type == KFD_QUEUE_TYPE_SDMA || 985 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { 986 987 if (!dqm->sched_running) { 988 WARN_ONCE(1, "Update non-HWS queue while stopped\n"); 989 goto out_unlock; 990 } 991 992 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, 993 (dqm->dev->kfd->cwsr_enabled ? 994 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : 995 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), 996 KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); 997 if (retval) { 998 dev_err(dev, "destroy mqd failed\n"); 999 goto out_unlock; 1000 } 1001 } 1002 1003 mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties, minfo); 1004 1005 /* 1006 * check active state vs. the previous state and modify 1007 * counter accordingly. map_queues_cpsch uses the 1008 * dqm->active_queue_count to determine whether a new runlist must be 1009 * uploaded. 1010 */ 1011 if (q->properties.is_active && !prev_active) { 1012 increment_queue_count(dqm, &pdd->qpd, q); 1013 } else if (!q->properties.is_active && prev_active) { 1014 decrement_queue_count(dqm, &pdd->qpd, q); 1015 } else if (q->gws && !q->properties.is_gws) { 1016 if (q->properties.is_active) { 1017 dqm->gws_queue_count++; 1018 pdd->qpd.mapped_gws_queue = true; 1019 } 1020 q->properties.is_gws = true; 1021 } else if (!q->gws && q->properties.is_gws) { 1022 if (q->properties.is_active) { 1023 dqm->gws_queue_count--; 1024 pdd->qpd.mapped_gws_queue = false; 1025 } 1026 q->properties.is_gws = false; 1027 } 1028 1029 if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { 1030 if (!dqm->dev->kfd->shared_resources.enable_mes) 1031 retval = map_queues_cpsch(dqm); 1032 else if (q->properties.is_active) 1033 retval = add_queue_mes(dqm, q, &pdd->qpd); 1034 } else if (q->properties.is_active && 1035 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || 1036 q->properties.type == KFD_QUEUE_TYPE_SDMA || 1037 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { 1038 if (WARN(q->process->mm != current->mm, 1039 "should only run in user thread")) 1040 retval = -EFAULT; 1041 else 1042 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, 1043 q->pipe, q->queue, 1044 &q->properties, current->mm); 1045 } 1046 1047 out_unlock: 1048 dqm_unlock(dqm); 1049 return retval; 1050 } 1051 1052 /* suspend_single_queue does not lock the dqm like the 1053 * evict_process_queues_cpsch or evict_process_queues_nocpsch. You should 1054 * lock the dqm before calling, and unlock after calling. 1055 * 1056 * The reason we don't lock the dqm is because this function may be 1057 * called on multiple queues in a loop, so rather than locking/unlocking 1058 * multiple times, we will just keep the dqm locked for all of the calls. 1059 */ 1060 static int suspend_single_queue(struct device_queue_manager *dqm, 1061 struct kfd_process_device *pdd, 1062 struct queue *q) 1063 { 1064 bool is_new; 1065 1066 if (q->properties.is_suspended) 1067 return 0; 1068 1069 pr_debug("Suspending process pid %d queue [%i]\n", 1070 pdd->process->lead_thread->pid, 1071 q->properties.queue_id); 1072 1073 is_new = q->properties.exception_status & KFD_EC_MASK(EC_QUEUE_NEW); 1074 1075 if (is_new || q->properties.is_being_destroyed) { 1076 pr_debug("Suspend: skip %s queue id %i\n", 1077 is_new ? "new" : "destroyed", 1078 q->properties.queue_id); 1079 return -EBUSY; 1080 } 1081 1082 q->properties.is_suspended = true; 1083 if (q->properties.is_active) { 1084 if (dqm->dev->kfd->shared_resources.enable_mes) { 1085 int r = remove_queue_mes(dqm, q, &pdd->qpd); 1086 1087 if (r) 1088 return r; 1089 } 1090 1091 decrement_queue_count(dqm, &pdd->qpd, q); 1092 q->properties.is_active = false; 1093 } 1094 1095 return 0; 1096 } 1097 1098 /* resume_single_queue does not lock the dqm like the functions 1099 * restore_process_queues_cpsch or restore_process_queues_nocpsch. You should 1100 * lock the dqm before calling, and unlock after calling. 1101 * 1102 * The reason we don't lock the dqm is because this function may be 1103 * called on multiple queues in a loop, so rather than locking/unlocking 1104 * multiple times, we will just keep the dqm locked for all of the calls. 1105 */ 1106 static int resume_single_queue(struct device_queue_manager *dqm, 1107 struct qcm_process_device *qpd, 1108 struct queue *q) 1109 { 1110 struct kfd_process_device *pdd; 1111 1112 if (!q->properties.is_suspended) 1113 return 0; 1114 1115 pdd = qpd_to_pdd(qpd); 1116 1117 pr_debug("Restoring from suspend process pid %d queue [%i]\n", 1118 pdd->process->lead_thread->pid, 1119 q->properties.queue_id); 1120 1121 q->properties.is_suspended = false; 1122 1123 if (QUEUE_IS_ACTIVE(q->properties)) { 1124 if (dqm->dev->kfd->shared_resources.enable_mes) { 1125 int r = add_queue_mes(dqm, q, &pdd->qpd); 1126 1127 if (r) 1128 return r; 1129 } 1130 1131 q->properties.is_active = true; 1132 increment_queue_count(dqm, qpd, q); 1133 } 1134 1135 return 0; 1136 } 1137 1138 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, 1139 struct qcm_process_device *qpd) 1140 { 1141 struct queue *q; 1142 struct mqd_manager *mqd_mgr; 1143 struct kfd_process_device *pdd; 1144 int retval, ret = 0; 1145 1146 dqm_lock(dqm); 1147 if (qpd->evicted++ > 0) /* already evicted, do nothing */ 1148 goto out; 1149 1150 pdd = qpd_to_pdd(qpd); 1151 pr_debug_ratelimited("Evicting process pid %d queues\n", 1152 pdd->process->lead_thread->pid); 1153 1154 pdd->last_evict_timestamp = get_jiffies_64(); 1155 /* Mark all queues as evicted. Deactivate all active queues on 1156 * the qpd. 1157 */ 1158 list_for_each_entry(q, &qpd->queues_list, list) { 1159 q->properties.is_evicted = true; 1160 if (!q->properties.is_active) 1161 continue; 1162 1163 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 1164 q->properties.type)]; 1165 q->properties.is_active = false; 1166 decrement_queue_count(dqm, qpd, q); 1167 1168 if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n")) 1169 continue; 1170 1171 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, 1172 (dqm->dev->kfd->cwsr_enabled ? 1173 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : 1174 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), 1175 KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); 1176 if (retval && !ret) 1177 /* Return the first error, but keep going to 1178 * maintain a consistent eviction state 1179 */ 1180 ret = retval; 1181 } 1182 1183 out: 1184 dqm_unlock(dqm); 1185 return ret; 1186 } 1187 1188 static int evict_process_queues_cpsch(struct device_queue_manager *dqm, 1189 struct qcm_process_device *qpd) 1190 { 1191 struct queue *q; 1192 struct device *dev = dqm->dev->adev->dev; 1193 struct kfd_process_device *pdd; 1194 int retval = 0; 1195 1196 dqm_lock(dqm); 1197 if (qpd->evicted++ > 0) /* already evicted, do nothing */ 1198 goto out; 1199 1200 pdd = qpd_to_pdd(qpd); 1201 1202 /* The debugger creates processes that temporarily have not acquired 1203 * all VMs for all devices and has no VMs itself. 1204 * Skip queue eviction on process eviction. 1205 */ 1206 if (!pdd->drm_priv) 1207 goto out; 1208 1209 pr_debug_ratelimited("Evicting process pid %d queues\n", 1210 pdd->process->lead_thread->pid); 1211 1212 if (dqm->dev->kfd->shared_resources.enable_mes) { 1213 pdd->last_evict_timestamp = get_jiffies_64(); 1214 retval = suspend_all_queues_mes(dqm); 1215 if (retval) { 1216 dev_err(dev, "Suspending all queues failed"); 1217 goto out; 1218 } 1219 } 1220 1221 /* Mark all queues as evicted. Deactivate all active queues on 1222 * the qpd. 1223 */ 1224 list_for_each_entry(q, &qpd->queues_list, list) { 1225 q->properties.is_evicted = true; 1226 if (!q->properties.is_active) 1227 continue; 1228 1229 q->properties.is_active = false; 1230 decrement_queue_count(dqm, qpd, q); 1231 1232 if (dqm->dev->kfd->shared_resources.enable_mes) { 1233 retval = remove_queue_mes(dqm, q, qpd); 1234 if (retval) { 1235 dev_err(dev, "Failed to evict queue %d\n", 1236 q->properties.queue_id); 1237 goto out; 1238 } 1239 } 1240 } 1241 1242 if (!dqm->dev->kfd->shared_resources.enable_mes) { 1243 pdd->last_evict_timestamp = get_jiffies_64(); 1244 retval = execute_queues_cpsch(dqm, 1245 qpd->is_debug ? 1246 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : 1247 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, 1248 USE_DEFAULT_GRACE_PERIOD); 1249 } else { 1250 retval = resume_all_queues_mes(dqm); 1251 if (retval) 1252 dev_err(dev, "Resuming all queues failed"); 1253 } 1254 1255 out: 1256 dqm_unlock(dqm); 1257 return retval; 1258 } 1259 1260 static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, 1261 struct qcm_process_device *qpd) 1262 { 1263 struct mm_struct *mm = NULL; 1264 struct queue *q; 1265 struct mqd_manager *mqd_mgr; 1266 struct kfd_process_device *pdd; 1267 uint64_t pd_base; 1268 uint64_t eviction_duration; 1269 int retval, ret = 0; 1270 1271 pdd = qpd_to_pdd(qpd); 1272 /* Retrieve PD base */ 1273 pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); 1274 1275 dqm_lock(dqm); 1276 if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ 1277 goto out; 1278 if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ 1279 qpd->evicted--; 1280 goto out; 1281 } 1282 1283 pr_debug_ratelimited("Restoring process pid %d queues\n", 1284 pdd->process->lead_thread->pid); 1285 1286 /* Update PD Base in QPD */ 1287 qpd->page_table_base = pd_base; 1288 pr_debug("Updated PD address to 0x%llx\n", pd_base); 1289 1290 if (!list_empty(&qpd->queues_list)) { 1291 dqm->dev->kfd2kgd->set_vm_context_page_table_base( 1292 dqm->dev->adev, 1293 qpd->vmid, 1294 qpd->page_table_base); 1295 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); 1296 } 1297 1298 /* Take a safe reference to the mm_struct, which may otherwise 1299 * disappear even while the kfd_process is still referenced. 1300 */ 1301 mm = get_task_mm(pdd->process->lead_thread); 1302 if (!mm) { 1303 ret = -EFAULT; 1304 goto out; 1305 } 1306 1307 /* Remove the eviction flags. Activate queues that are not 1308 * inactive for other reasons. 1309 */ 1310 list_for_each_entry(q, &qpd->queues_list, list) { 1311 q->properties.is_evicted = false; 1312 if (!QUEUE_IS_ACTIVE(q->properties)) 1313 continue; 1314 1315 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 1316 q->properties.type)]; 1317 q->properties.is_active = true; 1318 increment_queue_count(dqm, qpd, q); 1319 1320 if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n")) 1321 continue; 1322 1323 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, 1324 q->queue, &q->properties, mm); 1325 if (retval && !ret) 1326 /* Return the first error, but keep going to 1327 * maintain a consistent eviction state 1328 */ 1329 ret = retval; 1330 } 1331 qpd->evicted = 0; 1332 eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; 1333 atomic64_add(eviction_duration, &pdd->evict_duration_counter); 1334 out: 1335 if (mm) 1336 mmput(mm); 1337 dqm_unlock(dqm); 1338 return ret; 1339 } 1340 1341 static int restore_process_queues_cpsch(struct device_queue_manager *dqm, 1342 struct qcm_process_device *qpd) 1343 { 1344 struct queue *q; 1345 struct device *dev = dqm->dev->adev->dev; 1346 struct kfd_process_device *pdd; 1347 uint64_t eviction_duration; 1348 int retval = 0; 1349 1350 pdd = qpd_to_pdd(qpd); 1351 1352 dqm_lock(dqm); 1353 if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ 1354 goto out; 1355 if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ 1356 qpd->evicted--; 1357 goto out; 1358 } 1359 1360 /* The debugger creates processes that temporarily have not acquired 1361 * all VMs for all devices and has no VMs itself. 1362 * Skip queue restore on process restore. 1363 */ 1364 if (!pdd->drm_priv) 1365 goto vm_not_acquired; 1366 1367 pr_debug_ratelimited("Restoring process pid %d queues\n", 1368 pdd->process->lead_thread->pid); 1369 1370 /* Update PD Base in QPD */ 1371 qpd->page_table_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); 1372 pr_debug("Updated PD address to 0x%llx\n", qpd->page_table_base); 1373 1374 /* activate all active queues on the qpd */ 1375 list_for_each_entry(q, &qpd->queues_list, list) { 1376 q->properties.is_evicted = false; 1377 if (!QUEUE_IS_ACTIVE(q->properties)) 1378 continue; 1379 1380 q->properties.is_active = true; 1381 increment_queue_count(dqm, &pdd->qpd, q); 1382 1383 if (dqm->dev->kfd->shared_resources.enable_mes) { 1384 retval = add_queue_mes(dqm, q, qpd); 1385 if (retval) { 1386 dev_err(dev, "Failed to restore queue %d\n", 1387 q->properties.queue_id); 1388 goto out; 1389 } 1390 } 1391 } 1392 if (!dqm->dev->kfd->shared_resources.enable_mes) 1393 retval = execute_queues_cpsch(dqm, 1394 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); 1395 eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; 1396 atomic64_add(eviction_duration, &pdd->evict_duration_counter); 1397 vm_not_acquired: 1398 qpd->evicted = 0; 1399 out: 1400 dqm_unlock(dqm); 1401 return retval; 1402 } 1403 1404 static int register_process(struct device_queue_manager *dqm, 1405 struct qcm_process_device *qpd) 1406 { 1407 struct device_process_node *n; 1408 struct kfd_process_device *pdd; 1409 uint64_t pd_base; 1410 int retval; 1411 1412 n = kzalloc(sizeof(*n), GFP_KERNEL); 1413 if (!n) 1414 return -ENOMEM; 1415 1416 n->qpd = qpd; 1417 1418 pdd = qpd_to_pdd(qpd); 1419 /* Retrieve PD base */ 1420 pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); 1421 1422 dqm_lock(dqm); 1423 list_add(&n->list, &dqm->queues); 1424 1425 /* Update PD Base in QPD */ 1426 qpd->page_table_base = pd_base; 1427 pr_debug("Updated PD address to 0x%llx\n", pd_base); 1428 1429 retval = dqm->asic_ops.update_qpd(dqm, qpd); 1430 1431 dqm->processes_count++; 1432 1433 dqm_unlock(dqm); 1434 1435 /* Outside the DQM lock because under the DQM lock we can't do 1436 * reclaim or take other locks that others hold while reclaiming. 1437 */ 1438 kfd_inc_compute_active(dqm->dev); 1439 1440 return retval; 1441 } 1442 1443 static int unregister_process(struct device_queue_manager *dqm, 1444 struct qcm_process_device *qpd) 1445 { 1446 int retval; 1447 struct device_process_node *cur, *next; 1448 1449 pr_debug("qpd->queues_list is %s\n", 1450 list_empty(&qpd->queues_list) ? "empty" : "not empty"); 1451 1452 retval = 0; 1453 dqm_lock(dqm); 1454 1455 list_for_each_entry_safe(cur, next, &dqm->queues, list) { 1456 if (qpd == cur->qpd) { 1457 list_del(&cur->list); 1458 kfree(cur); 1459 dqm->processes_count--; 1460 goto out; 1461 } 1462 } 1463 /* qpd not found in dqm list */ 1464 retval = 1; 1465 out: 1466 dqm_unlock(dqm); 1467 1468 /* Outside the DQM lock because under the DQM lock we can't do 1469 * reclaim or take other locks that others hold while reclaiming. 1470 */ 1471 if (!retval) 1472 kfd_dec_compute_active(dqm->dev); 1473 1474 return retval; 1475 } 1476 1477 static int 1478 set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid, 1479 unsigned int vmid) 1480 { 1481 uint32_t xcc_mask = dqm->dev->xcc_mask; 1482 int xcc_id, ret; 1483 1484 for_each_inst(xcc_id, xcc_mask) { 1485 ret = dqm->dev->kfd2kgd->set_pasid_vmid_mapping( 1486 dqm->dev->adev, pasid, vmid, xcc_id); 1487 if (ret) 1488 break; 1489 } 1490 1491 return ret; 1492 } 1493 1494 static void init_interrupts(struct device_queue_manager *dqm) 1495 { 1496 uint32_t xcc_mask = dqm->dev->xcc_mask; 1497 unsigned int i, xcc_id; 1498 1499 for_each_inst(xcc_id, xcc_mask) { 1500 for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { 1501 if (is_pipe_enabled(dqm, 0, i)) { 1502 dqm->dev->kfd2kgd->init_interrupts( 1503 dqm->dev->adev, i, xcc_id); 1504 } 1505 } 1506 } 1507 } 1508 1509 static int initialize_nocpsch(struct device_queue_manager *dqm) 1510 { 1511 int pipe, queue; 1512 1513 pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); 1514 1515 dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm), 1516 sizeof(unsigned int), GFP_KERNEL); 1517 if (!dqm->allocated_queues) 1518 return -ENOMEM; 1519 1520 mutex_init(&dqm->lock_hidden); 1521 INIT_LIST_HEAD(&dqm->queues); 1522 dqm->active_queue_count = dqm->next_pipe_to_allocate = 0; 1523 dqm->active_cp_queue_count = 0; 1524 dqm->gws_queue_count = 0; 1525 1526 for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { 1527 int pipe_offset = pipe * get_queues_per_pipe(dqm); 1528 1529 for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) 1530 if (test_bit(pipe_offset + queue, 1531 dqm->dev->kfd->shared_resources.cp_queue_bitmap)) 1532 dqm->allocated_queues[pipe] |= 1 << queue; 1533 } 1534 1535 memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid)); 1536 1537 init_sdma_bitmaps(dqm); 1538 1539 return 0; 1540 } 1541 1542 static void uninitialize(struct device_queue_manager *dqm) 1543 { 1544 int i; 1545 1546 WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0); 1547 1548 kfree(dqm->allocated_queues); 1549 for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) 1550 kfree(dqm->mqd_mgrs[i]); 1551 mutex_destroy(&dqm->lock_hidden); 1552 } 1553 1554 static int start_nocpsch(struct device_queue_manager *dqm) 1555 { 1556 int r = 0; 1557 1558 pr_info("SW scheduler is used"); 1559 init_interrupts(dqm); 1560 1561 if (dqm->dev->adev->asic_type == CHIP_HAWAII) 1562 r = pm_init(&dqm->packet_mgr, dqm); 1563 if (!r) 1564 dqm->sched_running = true; 1565 1566 return r; 1567 } 1568 1569 static int stop_nocpsch(struct device_queue_manager *dqm) 1570 { 1571 dqm_lock(dqm); 1572 if (!dqm->sched_running) { 1573 dqm_unlock(dqm); 1574 return 0; 1575 } 1576 1577 if (dqm->dev->adev->asic_type == CHIP_HAWAII) 1578 pm_uninit(&dqm->packet_mgr); 1579 dqm->sched_running = false; 1580 dqm_unlock(dqm); 1581 1582 return 0; 1583 } 1584 1585 static int allocate_sdma_queue(struct device_queue_manager *dqm, 1586 struct queue *q, const uint32_t *restore_sdma_id) 1587 { 1588 struct device *dev = dqm->dev->adev->dev; 1589 int bit; 1590 1591 if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { 1592 if (bitmap_empty(dqm->sdma_bitmap, get_num_sdma_queues(dqm))) { 1593 dev_warn(dev, "No more SDMA queue to allocate (%d total queues)\n", 1594 get_num_sdma_queues(dqm)); 1595 return -ENOMEM; 1596 } 1597 1598 if (restore_sdma_id) { 1599 /* Re-use existing sdma_id */ 1600 if (!test_bit(*restore_sdma_id, dqm->sdma_bitmap)) { 1601 dev_err(dev, "SDMA queue already in use\n"); 1602 return -EBUSY; 1603 } 1604 clear_bit(*restore_sdma_id, dqm->sdma_bitmap); 1605 q->sdma_id = *restore_sdma_id; 1606 } else { 1607 /* Find first available sdma_id */ 1608 bit = find_first_bit(dqm->sdma_bitmap, 1609 get_num_sdma_queues(dqm)); 1610 clear_bit(bit, dqm->sdma_bitmap); 1611 q->sdma_id = bit; 1612 } 1613 1614 q->properties.sdma_engine_id = 1615 q->sdma_id % kfd_get_num_sdma_engines(dqm->dev); 1616 q->properties.sdma_queue_id = q->sdma_id / 1617 kfd_get_num_sdma_engines(dqm->dev); 1618 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 1619 if (bitmap_empty(dqm->xgmi_sdma_bitmap, get_num_xgmi_sdma_queues(dqm))) { 1620 dev_warn(dev, "No more XGMI SDMA queue to allocate (%d total queues)\n", 1621 get_num_xgmi_sdma_queues(dqm)); 1622 return -ENOMEM; 1623 } 1624 if (restore_sdma_id) { 1625 /* Re-use existing sdma_id */ 1626 if (!test_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap)) { 1627 dev_err(dev, "SDMA queue already in use\n"); 1628 return -EBUSY; 1629 } 1630 clear_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap); 1631 q->sdma_id = *restore_sdma_id; 1632 } else { 1633 bit = find_first_bit(dqm->xgmi_sdma_bitmap, 1634 get_num_xgmi_sdma_queues(dqm)); 1635 clear_bit(bit, dqm->xgmi_sdma_bitmap); 1636 q->sdma_id = bit; 1637 } 1638 /* sdma_engine_id is sdma id including 1639 * both PCIe-optimized SDMAs and XGMI- 1640 * optimized SDMAs. The calculation below 1641 * assumes the first N engines are always 1642 * PCIe-optimized ones 1643 */ 1644 q->properties.sdma_engine_id = 1645 kfd_get_num_sdma_engines(dqm->dev) + 1646 q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev); 1647 q->properties.sdma_queue_id = q->sdma_id / 1648 kfd_get_num_xgmi_sdma_engines(dqm->dev); 1649 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) { 1650 int i, num_queues, num_engines, eng_offset = 0, start_engine; 1651 bool free_bit_found = false, is_xgmi = false; 1652 1653 if (q->properties.sdma_engine_id < kfd_get_num_sdma_engines(dqm->dev)) { 1654 num_queues = get_num_sdma_queues(dqm); 1655 num_engines = kfd_get_num_sdma_engines(dqm->dev); 1656 q->properties.type = KFD_QUEUE_TYPE_SDMA; 1657 } else { 1658 num_queues = get_num_xgmi_sdma_queues(dqm); 1659 num_engines = kfd_get_num_xgmi_sdma_engines(dqm->dev); 1660 eng_offset = kfd_get_num_sdma_engines(dqm->dev); 1661 q->properties.type = KFD_QUEUE_TYPE_SDMA_XGMI; 1662 is_xgmi = true; 1663 } 1664 1665 /* Scan available bit based on target engine ID. */ 1666 start_engine = q->properties.sdma_engine_id - eng_offset; 1667 for (i = start_engine; i < num_queues; i += num_engines) { 1668 1669 if (!test_bit(i, is_xgmi ? dqm->xgmi_sdma_bitmap : dqm->sdma_bitmap)) 1670 continue; 1671 1672 clear_bit(i, is_xgmi ? dqm->xgmi_sdma_bitmap : dqm->sdma_bitmap); 1673 q->sdma_id = i; 1674 q->properties.sdma_queue_id = q->sdma_id / num_engines; 1675 free_bit_found = true; 1676 break; 1677 } 1678 1679 if (!free_bit_found) { 1680 dev_warn(dev, "No more SDMA queue to allocate for target ID %i (%d total queues)\n", 1681 q->properties.sdma_engine_id, num_queues); 1682 return -ENOMEM; 1683 } 1684 } 1685 1686 pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id); 1687 pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id); 1688 1689 return 0; 1690 } 1691 1692 static void deallocate_sdma_queue(struct device_queue_manager *dqm, 1693 struct queue *q) 1694 { 1695 if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { 1696 if (q->sdma_id >= get_num_sdma_queues(dqm)) 1697 return; 1698 set_bit(q->sdma_id, dqm->sdma_bitmap); 1699 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 1700 if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm)) 1701 return; 1702 set_bit(q->sdma_id, dqm->xgmi_sdma_bitmap); 1703 } 1704 } 1705 1706 /* 1707 * Device Queue Manager implementation for cp scheduler 1708 */ 1709 1710 static int set_sched_resources(struct device_queue_manager *dqm) 1711 { 1712 int i, mec; 1713 struct scheduling_resources res; 1714 struct device *dev = dqm->dev->adev->dev; 1715 1716 res.vmid_mask = dqm->dev->compute_vmid_bitmap; 1717 1718 res.queue_mask = 0; 1719 for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) { 1720 mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe) 1721 / dqm->dev->kfd->shared_resources.num_pipe_per_mec; 1722 1723 if (!test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap)) 1724 continue; 1725 1726 /* only acquire queues from the first MEC */ 1727 if (mec > 0) 1728 continue; 1729 1730 /* This situation may be hit in the future if a new HW 1731 * generation exposes more than 64 queues. If so, the 1732 * definition of res.queue_mask needs updating 1733 */ 1734 if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) { 1735 dev_err(dev, "Invalid queue enabled by amdgpu: %d\n", i); 1736 break; 1737 } 1738 1739 res.queue_mask |= 1ull 1740 << amdgpu_queue_mask_bit_to_set_resource_bit( 1741 dqm->dev->adev, i); 1742 } 1743 res.gws_mask = ~0ull; 1744 res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0; 1745 1746 pr_debug("Scheduling resources:\n" 1747 "vmid mask: 0x%8X\n" 1748 "queue mask: 0x%8llX\n", 1749 res.vmid_mask, res.queue_mask); 1750 1751 return pm_send_set_resources(&dqm->packet_mgr, &res); 1752 } 1753 1754 static int initialize_cpsch(struct device_queue_manager *dqm) 1755 { 1756 pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); 1757 1758 mutex_init(&dqm->lock_hidden); 1759 INIT_LIST_HEAD(&dqm->queues); 1760 dqm->active_queue_count = dqm->processes_count = 0; 1761 dqm->active_cp_queue_count = 0; 1762 dqm->gws_queue_count = 0; 1763 dqm->active_runlist = false; 1764 dqm->trap_debug_vmid = 0; 1765 1766 init_sdma_bitmaps(dqm); 1767 1768 update_dqm_wait_times(dqm); 1769 return 0; 1770 } 1771 1772 /* halt_cpsch: 1773 * Unmap queues so the schedule doesn't continue remaining jobs in the queue. 1774 * Then set dqm->sched_halt so queues don't map to runlist until unhalt_cpsch 1775 * is called. 1776 */ 1777 static int halt_cpsch(struct device_queue_manager *dqm) 1778 { 1779 int ret = 0; 1780 1781 dqm_lock(dqm); 1782 if (!dqm->sched_running) { 1783 dqm_unlock(dqm); 1784 return 0; 1785 } 1786 1787 WARN_ONCE(dqm->sched_halt, "Scheduling is already on halt\n"); 1788 1789 if (!dqm->is_hws_hang) { 1790 if (!dqm->dev->kfd->shared_resources.enable_mes) 1791 ret = unmap_queues_cpsch(dqm, 1792 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 1793 USE_DEFAULT_GRACE_PERIOD, false); 1794 else 1795 ret = remove_all_kfd_queues_mes(dqm); 1796 } 1797 dqm->sched_halt = true; 1798 dqm_unlock(dqm); 1799 1800 return ret; 1801 } 1802 1803 /* unhalt_cpsch 1804 * Unset dqm->sched_halt and map queues back to runlist 1805 */ 1806 static int unhalt_cpsch(struct device_queue_manager *dqm) 1807 { 1808 int ret = 0; 1809 1810 dqm_lock(dqm); 1811 if (!dqm->sched_running || !dqm->sched_halt) { 1812 WARN_ONCE(!dqm->sched_halt, "Scheduling is not on halt.\n"); 1813 dqm_unlock(dqm); 1814 return 0; 1815 } 1816 dqm->sched_halt = false; 1817 if (!dqm->dev->kfd->shared_resources.enable_mes) 1818 ret = execute_queues_cpsch(dqm, 1819 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 1820 0, USE_DEFAULT_GRACE_PERIOD); 1821 else 1822 ret = add_all_kfd_queues_mes(dqm); 1823 1824 dqm_unlock(dqm); 1825 1826 return ret; 1827 } 1828 1829 static int start_cpsch(struct device_queue_manager *dqm) 1830 { 1831 struct device *dev = dqm->dev->adev->dev; 1832 int retval, num_hw_queue_slots; 1833 1834 retval = 0; 1835 1836 dqm_lock(dqm); 1837 1838 if (!dqm->dev->kfd->shared_resources.enable_mes) { 1839 retval = pm_init(&dqm->packet_mgr, dqm); 1840 if (retval) 1841 goto fail_packet_manager_init; 1842 1843 retval = set_sched_resources(dqm); 1844 if (retval) 1845 goto fail_set_sched_resources; 1846 } 1847 pr_debug("Allocating fence memory\n"); 1848 1849 /* allocate fence memory on the gart */ 1850 retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr), 1851 &dqm->fence_mem); 1852 1853 if (retval) 1854 goto fail_allocate_vidmem; 1855 1856 dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr; 1857 dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr; 1858 1859 init_interrupts(dqm); 1860 1861 /* clear hang status when driver try to start the hw scheduler */ 1862 dqm->sched_running = true; 1863 1864 if (!dqm->dev->kfd->shared_resources.enable_mes) { 1865 if (pm_config_dequeue_wait_counts(&dqm->packet_mgr, 1866 KFD_DEQUEUE_WAIT_INIT, 0 /* unused */)) 1867 dev_err(dev, "Setting optimized dequeue wait failed. Using default values\n"); 1868 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); 1869 } 1870 1871 /* setup per-queue reset detection buffer */ 1872 num_hw_queue_slots = dqm->dev->kfd->shared_resources.num_queue_per_pipe * 1873 dqm->dev->kfd->shared_resources.num_pipe_per_mec * 1874 NUM_XCC(dqm->dev->xcc_mask); 1875 1876 dqm->detect_hang_info_size = num_hw_queue_slots * sizeof(struct dqm_detect_hang_info); 1877 dqm->detect_hang_info = kzalloc(dqm->detect_hang_info_size, GFP_KERNEL); 1878 1879 if (!dqm->detect_hang_info) { 1880 retval = -ENOMEM; 1881 goto fail_detect_hang_buffer; 1882 } 1883 1884 dqm_unlock(dqm); 1885 1886 return 0; 1887 fail_detect_hang_buffer: 1888 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); 1889 fail_allocate_vidmem: 1890 fail_set_sched_resources: 1891 if (!dqm->dev->kfd->shared_resources.enable_mes) 1892 pm_uninit(&dqm->packet_mgr); 1893 fail_packet_manager_init: 1894 dqm_unlock(dqm); 1895 return retval; 1896 } 1897 1898 static int stop_cpsch(struct device_queue_manager *dqm) 1899 { 1900 int ret = 0; 1901 1902 dqm_lock(dqm); 1903 if (!dqm->sched_running) { 1904 dqm_unlock(dqm); 1905 return 0; 1906 } 1907 1908 if (!dqm->dev->kfd->shared_resources.enable_mes) 1909 ret = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 1910 0, USE_DEFAULT_GRACE_PERIOD, false); 1911 else 1912 ret = remove_all_kfd_queues_mes(dqm); 1913 1914 dqm->sched_running = false; 1915 1916 if (!dqm->dev->kfd->shared_resources.enable_mes) 1917 pm_release_ib(&dqm->packet_mgr); 1918 1919 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); 1920 if (!dqm->dev->kfd->shared_resources.enable_mes) 1921 pm_uninit(&dqm->packet_mgr); 1922 kfree(dqm->detect_hang_info); 1923 dqm->detect_hang_info = NULL; 1924 dqm_unlock(dqm); 1925 1926 return ret; 1927 } 1928 1929 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, 1930 struct kernel_queue *kq, 1931 struct qcm_process_device *qpd) 1932 { 1933 dqm_lock(dqm); 1934 if (dqm->total_queue_count >= max_num_of_queues_per_device) { 1935 pr_warn("Can't create new kernel queue because %d queues were already created\n", 1936 dqm->total_queue_count); 1937 dqm_unlock(dqm); 1938 return -EPERM; 1939 } 1940 1941 /* 1942 * Unconditionally increment this counter, regardless of the queue's 1943 * type or whether the queue is active. 1944 */ 1945 dqm->total_queue_count++; 1946 pr_debug("Total of %d queues are accountable so far\n", 1947 dqm->total_queue_count); 1948 1949 list_add(&kq->list, &qpd->priv_queue_list); 1950 increment_queue_count(dqm, qpd, kq->queue); 1951 qpd->is_debug = true; 1952 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, 1953 USE_DEFAULT_GRACE_PERIOD); 1954 dqm_unlock(dqm); 1955 1956 return 0; 1957 } 1958 1959 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, 1960 struct kernel_queue *kq, 1961 struct qcm_process_device *qpd) 1962 { 1963 dqm_lock(dqm); 1964 list_del(&kq->list); 1965 decrement_queue_count(dqm, qpd, kq->queue); 1966 qpd->is_debug = false; 1967 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 1968 USE_DEFAULT_GRACE_PERIOD); 1969 /* 1970 * Unconditionally decrement this counter, regardless of the queue's 1971 * type. 1972 */ 1973 dqm->total_queue_count--; 1974 pr_debug("Total of %d queues are accountable so far\n", 1975 dqm->total_queue_count); 1976 dqm_unlock(dqm); 1977 } 1978 1979 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, 1980 struct qcm_process_device *qpd, 1981 const struct kfd_criu_queue_priv_data *qd, 1982 const void *restore_mqd, const void *restore_ctl_stack) 1983 { 1984 int retval; 1985 struct mqd_manager *mqd_mgr; 1986 1987 if (dqm->total_queue_count >= max_num_of_queues_per_device) { 1988 pr_warn("Can't create new usermode queue because %d queues were already created\n", 1989 dqm->total_queue_count); 1990 retval = -EPERM; 1991 goto out; 1992 } 1993 1994 if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 1995 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI || 1996 q->properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) { 1997 dqm_lock(dqm); 1998 retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL); 1999 dqm_unlock(dqm); 2000 if (retval) 2001 goto out; 2002 } 2003 2004 retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL); 2005 if (retval) 2006 goto out_deallocate_sdma_queue; 2007 2008 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 2009 q->properties.type)]; 2010 2011 if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 2012 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) 2013 dqm->asic_ops.init_sdma_vm(dqm, q, qpd); 2014 q->properties.tba_addr = qpd->tba_addr; 2015 q->properties.tma_addr = qpd->tma_addr; 2016 q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); 2017 if (!q->mqd_mem_obj) { 2018 retval = -ENOMEM; 2019 goto out_deallocate_doorbell; 2020 } 2021 2022 dqm_lock(dqm); 2023 /* 2024 * Eviction state logic: mark all queues as evicted, even ones 2025 * not currently active. Restoring inactive queues later only 2026 * updates the is_evicted flag but is a no-op otherwise. 2027 */ 2028 q->properties.is_evicted = !!qpd->evicted; 2029 q->properties.is_dbg_wa = qpd->pqm->process->debug_trap_enabled && 2030 kfd_dbg_has_cwsr_workaround(q->device); 2031 2032 if (qd) 2033 mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, 2034 &q->properties, restore_mqd, restore_ctl_stack, 2035 qd->ctl_stack_size); 2036 else 2037 mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, 2038 &q->gart_mqd_addr, &q->properties); 2039 2040 list_add(&q->list, &qpd->queues_list); 2041 qpd->queue_count++; 2042 2043 if (q->properties.is_active) { 2044 increment_queue_count(dqm, qpd, q); 2045 2046 if (!dqm->dev->kfd->shared_resources.enable_mes) 2047 retval = execute_queues_cpsch(dqm, 2048 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); 2049 else 2050 retval = add_queue_mes(dqm, q, qpd); 2051 if (retval) 2052 goto cleanup_queue; 2053 } 2054 2055 /* 2056 * Unconditionally increment this counter, regardless of the queue's 2057 * type or whether the queue is active. 2058 */ 2059 dqm->total_queue_count++; 2060 2061 pr_debug("Total of %d queues are accountable so far\n", 2062 dqm->total_queue_count); 2063 2064 dqm_unlock(dqm); 2065 return retval; 2066 2067 cleanup_queue: 2068 qpd->queue_count--; 2069 list_del(&q->list); 2070 if (q->properties.is_active) 2071 decrement_queue_count(dqm, qpd, q); 2072 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 2073 dqm_unlock(dqm); 2074 out_deallocate_doorbell: 2075 deallocate_doorbell(qpd, q); 2076 out_deallocate_sdma_queue: 2077 if (q->properties.type == KFD_QUEUE_TYPE_SDMA || 2078 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 2079 dqm_lock(dqm); 2080 deallocate_sdma_queue(dqm, q); 2081 dqm_unlock(dqm); 2082 } 2083 out: 2084 return retval; 2085 } 2086 2087 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm, 2088 uint64_t fence_value, 2089 unsigned int timeout_ms) 2090 { 2091 unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies; 2092 struct device *dev = dqm->dev->adev->dev; 2093 uint64_t *fence_addr = dqm->fence_addr; 2094 2095 while (*fence_addr != fence_value) { 2096 /* Fatal err detected, this response won't come */ 2097 if (amdgpu_amdkfd_is_fed(dqm->dev->adev)) 2098 return -EIO; 2099 2100 if (time_after(jiffies, end_jiffies)) { 2101 dev_err(dev, "qcm fence wait loop timeout expired\n"); 2102 /* In HWS case, this is used to halt the driver thread 2103 * in order not to mess up CP states before doing 2104 * scandumps for FW debugging. 2105 */ 2106 while (halt_if_hws_hang) 2107 schedule(); 2108 2109 return -ETIME; 2110 } 2111 schedule(); 2112 } 2113 2114 return 0; 2115 } 2116 2117 /* dqm->lock mutex has to be locked before calling this function */ 2118 static int map_queues_cpsch(struct device_queue_manager *dqm) 2119 { 2120 struct device *dev = dqm->dev->adev->dev; 2121 int retval; 2122 2123 if (!dqm->sched_running || dqm->sched_halt) 2124 return 0; 2125 if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0) 2126 return 0; 2127 if (dqm->active_runlist) 2128 return 0; 2129 2130 retval = pm_send_runlist(&dqm->packet_mgr, &dqm->queues); 2131 pr_debug("%s sent runlist\n", __func__); 2132 if (retval) { 2133 dev_err(dev, "failed to execute runlist\n"); 2134 return retval; 2135 } 2136 dqm->active_runlist = true; 2137 2138 return retval; 2139 } 2140 2141 static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q, 2142 struct qcm_process_device *qpd) 2143 { 2144 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 2145 2146 dev_err(dqm->dev->adev->dev, "queue id 0x%0x at pasid %d is reset\n", 2147 q->properties.queue_id, pdd->process->lead_thread->pid); 2148 2149 pdd->has_reset_queue = true; 2150 if (q->properties.is_active) { 2151 q->properties.is_active = false; 2152 decrement_queue_count(dqm, qpd, q); 2153 } 2154 } 2155 2156 static int detect_queue_hang(struct device_queue_manager *dqm) 2157 { 2158 int i; 2159 2160 /* detect should be used only in dqm locked queue reset */ 2161 if (WARN_ON(dqm->detect_hang_count > 0)) 2162 return 0; 2163 2164 memset(dqm->detect_hang_info, 0, dqm->detect_hang_info_size); 2165 2166 for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) { 2167 uint32_t mec, pipe, queue; 2168 int xcc_id; 2169 2170 mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe) 2171 / dqm->dev->kfd->shared_resources.num_pipe_per_mec; 2172 2173 if (mec || !test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap)) 2174 continue; 2175 2176 amdgpu_queue_mask_bit_to_mec_queue(dqm->dev->adev, i, &mec, &pipe, &queue); 2177 2178 for_each_inst(xcc_id, dqm->dev->xcc_mask) { 2179 uint64_t queue_addr = dqm->dev->kfd2kgd->hqd_get_pq_addr( 2180 dqm->dev->adev, pipe, queue, xcc_id); 2181 struct dqm_detect_hang_info hang_info; 2182 2183 if (!queue_addr) 2184 continue; 2185 2186 hang_info.pipe_id = pipe; 2187 hang_info.queue_id = queue; 2188 hang_info.xcc_id = xcc_id; 2189 hang_info.queue_address = queue_addr; 2190 2191 dqm->detect_hang_info[dqm->detect_hang_count] = hang_info; 2192 dqm->detect_hang_count++; 2193 } 2194 } 2195 2196 return dqm->detect_hang_count; 2197 } 2198 2199 static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uint64_t queue_address) 2200 { 2201 struct device_process_node *cur; 2202 struct qcm_process_device *qpd; 2203 struct queue *q; 2204 2205 list_for_each_entry(cur, &dqm->queues, list) { 2206 qpd = cur->qpd; 2207 list_for_each_entry(q, &qpd->queues_list, list) { 2208 if (queue_address == q->properties.queue_address) 2209 return q; 2210 } 2211 } 2212 2213 return NULL; 2214 } 2215 2216 static int reset_hung_queues(struct device_queue_manager *dqm) 2217 { 2218 int r = 0, reset_count = 0, i; 2219 2220 if (!dqm->detect_hang_info || dqm->is_hws_hang) 2221 return -EIO; 2222 2223 /* assume dqm locked. */ 2224 if (!detect_queue_hang(dqm)) 2225 return -ENOTRECOVERABLE; 2226 2227 for (i = 0; i < dqm->detect_hang_count; i++) { 2228 struct dqm_detect_hang_info hang_info = dqm->detect_hang_info[i]; 2229 struct queue *q = find_queue_by_address(dqm, hang_info.queue_address); 2230 struct kfd_process_device *pdd; 2231 uint64_t queue_addr = 0; 2232 2233 if (!q) { 2234 r = -ENOTRECOVERABLE; 2235 goto reset_fail; 2236 } 2237 2238 pdd = kfd_get_process_device_data(dqm->dev, q->process); 2239 if (!pdd) { 2240 r = -ENOTRECOVERABLE; 2241 goto reset_fail; 2242 } 2243 2244 queue_addr = dqm->dev->kfd2kgd->hqd_reset(dqm->dev->adev, 2245 hang_info.pipe_id, hang_info.queue_id, hang_info.xcc_id, 2246 KFD_UNMAP_LATENCY_MS); 2247 2248 /* either reset failed or we reset an unexpected queue. */ 2249 if (queue_addr != q->properties.queue_address) { 2250 r = -ENOTRECOVERABLE; 2251 goto reset_fail; 2252 } 2253 2254 set_queue_as_reset(dqm, q, &pdd->qpd); 2255 reset_count++; 2256 } 2257 2258 if (reset_count == dqm->detect_hang_count) 2259 kfd_signal_reset_event(dqm->dev); 2260 else 2261 r = -ENOTRECOVERABLE; 2262 2263 reset_fail: 2264 dqm->detect_hang_count = 0; 2265 2266 return r; 2267 } 2268 2269 static bool sdma_has_hang(struct device_queue_manager *dqm) 2270 { 2271 int engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm); 2272 int engine_end = engine_start + get_num_all_sdma_engines(dqm); 2273 int num_queues_per_eng = dqm->dev->kfd->device_info.num_sdma_queues_per_engine; 2274 int i, j; 2275 2276 for (i = engine_start; i < engine_end; i++) { 2277 for (j = 0; j < num_queues_per_eng; j++) { 2278 if (!dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j)) 2279 continue; 2280 2281 return true; 2282 } 2283 } 2284 2285 return false; 2286 } 2287 2288 static bool set_sdma_queue_as_reset(struct device_queue_manager *dqm, 2289 uint32_t doorbell_off) 2290 { 2291 struct device_process_node *cur; 2292 struct qcm_process_device *qpd; 2293 struct queue *q; 2294 2295 list_for_each_entry(cur, &dqm->queues, list) { 2296 qpd = cur->qpd; 2297 list_for_each_entry(q, &qpd->queues_list, list) { 2298 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA || 2299 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) && 2300 q->properties.doorbell_off == doorbell_off) { 2301 set_queue_as_reset(dqm, q, qpd); 2302 return true; 2303 } 2304 } 2305 } 2306 2307 return false; 2308 } 2309 2310 static int reset_hung_queues_sdma(struct device_queue_manager *dqm) 2311 { 2312 int engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm); 2313 int engine_end = engine_start + get_num_all_sdma_engines(dqm); 2314 int num_queues_per_eng = dqm->dev->kfd->device_info.num_sdma_queues_per_engine; 2315 int r = 0, i, j; 2316 2317 if (dqm->is_hws_hang) 2318 return -EIO; 2319 2320 /* Scan for hung HW queues and reset engine. */ 2321 dqm->detect_hang_count = 0; 2322 for (i = engine_start; i < engine_end; i++) { 2323 for (j = 0; j < num_queues_per_eng; j++) { 2324 uint32_t doorbell_off = 2325 dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j); 2326 2327 if (!doorbell_off) 2328 continue; 2329 2330 /* Reset engine and check. */ 2331 if (amdgpu_sdma_reset_engine(dqm->dev->adev, i, false) || 2332 dqm->dev->kfd2kgd->hqd_sdma_get_doorbell(dqm->dev->adev, i, j) || 2333 !set_sdma_queue_as_reset(dqm, doorbell_off)) { 2334 r = -ENOTRECOVERABLE; 2335 goto reset_fail; 2336 } 2337 2338 /* Should only expect one queue active per engine */ 2339 dqm->detect_hang_count++; 2340 break; 2341 } 2342 } 2343 2344 /* Signal process reset */ 2345 if (dqm->detect_hang_count) 2346 kfd_signal_reset_event(dqm->dev); 2347 else 2348 r = -ENOTRECOVERABLE; 2349 2350 reset_fail: 2351 dqm->detect_hang_count = 0; 2352 2353 return r; 2354 } 2355 2356 static int reset_queues_on_hws_hang(struct device_queue_manager *dqm, bool is_sdma) 2357 { 2358 struct amdgpu_device *adev = dqm->dev->adev; 2359 2360 while (halt_if_hws_hang) 2361 schedule(); 2362 2363 if (adev->debug_disable_gpu_ring_reset) { 2364 dev_info_once(adev->dev, 2365 "%s queue hung, but ring reset disabled", 2366 is_sdma ? "sdma" : "compute"); 2367 2368 return -EPERM; 2369 } 2370 if (!amdgpu_gpu_recovery) 2371 return -ENOTRECOVERABLE; 2372 2373 return is_sdma ? reset_hung_queues_sdma(dqm) : reset_hung_queues(dqm); 2374 } 2375 2376 /* dqm->lock mutex has to be locked before calling this function 2377 * 2378 * @grace_period: If USE_DEFAULT_GRACE_PERIOD then default wait time 2379 * for context switch latency. Lower values are used by debugger 2380 * since context switching are triggered at high frequency. 2381 * This is configured by setting CP_IQ_WAIT_TIME2.SCH_WAVE 2382 * 2383 */ 2384 static int unmap_queues_cpsch(struct device_queue_manager *dqm, 2385 enum kfd_unmap_queues_filter filter, 2386 uint32_t filter_param, 2387 uint32_t grace_period, 2388 bool reset) 2389 { 2390 struct device *dev = dqm->dev->adev->dev; 2391 struct mqd_manager *mqd_mgr; 2392 int retval; 2393 2394 if (!dqm->sched_running) 2395 return 0; 2396 if (!dqm->active_runlist) 2397 return 0; 2398 if (!down_read_trylock(&dqm->dev->adev->reset_domain->sem)) 2399 return -EIO; 2400 2401 if (grace_period != USE_DEFAULT_GRACE_PERIOD) { 2402 retval = pm_config_dequeue_wait_counts(&dqm->packet_mgr, 2403 KFD_DEQUEUE_WAIT_SET_SCH_WAVE, grace_period); 2404 if (retval) 2405 goto out; 2406 } 2407 2408 retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset); 2409 if (retval) 2410 goto out; 2411 2412 *dqm->fence_addr = KFD_FENCE_INIT; 2413 mb(); 2414 pm_send_query_status(&dqm->packet_mgr, dqm->fence_gpu_addr, 2415 KFD_FENCE_COMPLETED); 2416 /* should be timed out */ 2417 retval = amdkfd_fence_wait_timeout(dqm, KFD_FENCE_COMPLETED, 2418 queue_preemption_timeout_ms); 2419 if (retval) { 2420 dev_err(dev, "The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); 2421 kfd_hws_hang(dqm); 2422 goto out; 2423 } 2424 2425 /* In the current MEC firmware implementation, if compute queue 2426 * doesn't response to the preemption request in time, HIQ will 2427 * abandon the unmap request without returning any timeout error 2428 * to driver. Instead, MEC firmware will log the doorbell of the 2429 * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields. 2430 * To make sure the queue unmap was successful, driver need to 2431 * check those fields 2432 */ 2433 mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]; 2434 if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd) && 2435 reset_queues_on_hws_hang(dqm, false)) 2436 goto reset_fail; 2437 2438 /* Check for SDMA hang and attempt SDMA reset */ 2439 if (sdma_has_hang(dqm) && reset_queues_on_hws_hang(dqm, true)) 2440 goto reset_fail; 2441 2442 /* We need to reset the grace period value for this device */ 2443 if (grace_period != USE_DEFAULT_GRACE_PERIOD) { 2444 if (pm_config_dequeue_wait_counts(&dqm->packet_mgr, 2445 KFD_DEQUEUE_WAIT_RESET, 0 /* unused */)) 2446 dev_err(dev, "Failed to reset grace period\n"); 2447 } 2448 2449 pm_release_ib(&dqm->packet_mgr); 2450 dqm->active_runlist = false; 2451 out: 2452 up_read(&dqm->dev->adev->reset_domain->sem); 2453 return retval; 2454 2455 reset_fail: 2456 dqm->is_hws_hang = true; 2457 kfd_hws_hang(dqm); 2458 up_read(&dqm->dev->adev->reset_domain->sem); 2459 return -ETIME; 2460 } 2461 2462 /* only for compute queue */ 2463 static int reset_queues_cpsch(struct device_queue_manager *dqm, uint16_t pasid) 2464 { 2465 int retval; 2466 2467 dqm_lock(dqm); 2468 2469 retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID, 2470 pasid, USE_DEFAULT_GRACE_PERIOD, true); 2471 2472 dqm_unlock(dqm); 2473 return retval; 2474 } 2475 2476 /* dqm->lock mutex has to be locked before calling this function */ 2477 static int execute_queues_cpsch(struct device_queue_manager *dqm, 2478 enum kfd_unmap_queues_filter filter, 2479 uint32_t filter_param, 2480 uint32_t grace_period) 2481 { 2482 int retval; 2483 2484 if (!down_read_trylock(&dqm->dev->adev->reset_domain->sem)) 2485 return -EIO; 2486 retval = unmap_queues_cpsch(dqm, filter, filter_param, grace_period, false); 2487 if (!retval) 2488 retval = map_queues_cpsch(dqm); 2489 up_read(&dqm->dev->adev->reset_domain->sem); 2490 return retval; 2491 } 2492 2493 static int wait_on_destroy_queue(struct device_queue_manager *dqm, 2494 struct queue *q) 2495 { 2496 struct kfd_process_device *pdd = kfd_get_process_device_data(q->device, 2497 q->process); 2498 int ret = 0; 2499 2500 if (WARN_ON(!pdd)) 2501 return ret; 2502 2503 if (pdd->qpd.is_debug) 2504 return ret; 2505 2506 q->properties.is_being_destroyed = true; 2507 2508 if (pdd->process->debug_trap_enabled && q->properties.is_suspended) { 2509 dqm_unlock(dqm); 2510 mutex_unlock(&q->process->mutex); 2511 ret = wait_event_interruptible(dqm->destroy_wait, 2512 !q->properties.is_suspended); 2513 2514 mutex_lock(&q->process->mutex); 2515 dqm_lock(dqm); 2516 } 2517 2518 return ret; 2519 } 2520 2521 static int destroy_queue_cpsch(struct device_queue_manager *dqm, 2522 struct qcm_process_device *qpd, 2523 struct queue *q) 2524 { 2525 int retval; 2526 struct mqd_manager *mqd_mgr; 2527 uint64_t sdma_val = 0; 2528 struct kfd_process_device *pdd = qpd_to_pdd(qpd); 2529 struct device *dev = dqm->dev->adev->dev; 2530 2531 /* Get the SDMA queue stats */ 2532 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || 2533 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { 2534 retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr, 2535 &sdma_val); 2536 if (retval) 2537 dev_err(dev, "Failed to read SDMA queue counter for queue: %d\n", 2538 q->properties.queue_id); 2539 } 2540 2541 /* remove queue from list to prevent rescheduling after preemption */ 2542 dqm_lock(dqm); 2543 2544 retval = wait_on_destroy_queue(dqm, q); 2545 2546 if (retval) { 2547 dqm_unlock(dqm); 2548 return retval; 2549 } 2550 2551 if (qpd->is_debug) { 2552 /* 2553 * error, currently we do not allow to destroy a queue 2554 * of a currently debugged process 2555 */ 2556 retval = -EBUSY; 2557 goto failed_try_destroy_debugged_queue; 2558 2559 } 2560 2561 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 2562 q->properties.type)]; 2563 2564 deallocate_doorbell(qpd, q); 2565 2566 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || 2567 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { 2568 deallocate_sdma_queue(dqm, q); 2569 pdd->sdma_past_activity_counter += sdma_val; 2570 } 2571 2572 if (q->properties.is_active) { 2573 decrement_queue_count(dqm, qpd, q); 2574 q->properties.is_active = false; 2575 if (!dqm->dev->kfd->shared_resources.enable_mes) { 2576 retval = execute_queues_cpsch(dqm, 2577 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, 2578 USE_DEFAULT_GRACE_PERIOD); 2579 if (retval == -ETIME) 2580 qpd->reset_wavefronts = true; 2581 } else { 2582 retval = remove_queue_mes(dqm, q, qpd); 2583 } 2584 } 2585 list_del(&q->list); 2586 qpd->queue_count--; 2587 2588 /* 2589 * Unconditionally decrement this counter, regardless of the queue's 2590 * type 2591 */ 2592 dqm->total_queue_count--; 2593 pr_debug("Total of %d queues are accountable so far\n", 2594 dqm->total_queue_count); 2595 2596 dqm_unlock(dqm); 2597 2598 /* 2599 * Do free_mqd and raise delete event after dqm_unlock(dqm) to avoid 2600 * circular locking 2601 */ 2602 kfd_dbg_ev_raise(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE), 2603 qpd->pqm->process, q->device, 2604 -1, false, NULL, 0); 2605 2606 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 2607 2608 return retval; 2609 2610 failed_try_destroy_debugged_queue: 2611 2612 dqm_unlock(dqm); 2613 return retval; 2614 } 2615 2616 static bool set_cache_memory_policy(struct device_queue_manager *dqm, 2617 struct qcm_process_device *qpd, 2618 enum cache_policy default_policy, 2619 enum cache_policy alternate_policy, 2620 void __user *alternate_aperture_base, 2621 uint64_t alternate_aperture_size, 2622 u32 misc_process_properties) 2623 { 2624 bool retval = true; 2625 2626 if (!dqm->asic_ops.set_cache_memory_policy) 2627 return retval; 2628 2629 dqm_lock(dqm); 2630 2631 retval = dqm->asic_ops.set_cache_memory_policy( 2632 dqm, 2633 qpd, 2634 default_policy, 2635 alternate_policy, 2636 alternate_aperture_base, 2637 alternate_aperture_size, 2638 misc_process_properties); 2639 2640 if (retval) 2641 goto out; 2642 2643 if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0)) 2644 program_sh_mem_settings(dqm, qpd); 2645 2646 pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n", 2647 qpd->sh_mem_config, qpd->sh_mem_ape1_base, 2648 qpd->sh_mem_ape1_limit); 2649 2650 out: 2651 dqm_unlock(dqm); 2652 return retval; 2653 } 2654 2655 static int process_termination_nocpsch(struct device_queue_manager *dqm, 2656 struct qcm_process_device *qpd) 2657 { 2658 struct queue *q; 2659 struct device_process_node *cur, *next_dpn; 2660 int retval = 0; 2661 bool found = false; 2662 2663 dqm_lock(dqm); 2664 2665 /* Clear all user mode queues */ 2666 while (!list_empty(&qpd->queues_list)) { 2667 struct mqd_manager *mqd_mgr; 2668 int ret; 2669 2670 q = list_first_entry(&qpd->queues_list, struct queue, list); 2671 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 2672 q->properties.type)]; 2673 ret = destroy_queue_nocpsch_locked(dqm, qpd, q); 2674 if (ret) 2675 retval = ret; 2676 dqm_unlock(dqm); 2677 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 2678 dqm_lock(dqm); 2679 } 2680 2681 /* Unregister process */ 2682 list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { 2683 if (qpd == cur->qpd) { 2684 list_del(&cur->list); 2685 kfree(cur); 2686 dqm->processes_count--; 2687 found = true; 2688 break; 2689 } 2690 } 2691 2692 dqm_unlock(dqm); 2693 2694 /* Outside the DQM lock because under the DQM lock we can't do 2695 * reclaim or take other locks that others hold while reclaiming. 2696 */ 2697 if (found) 2698 kfd_dec_compute_active(dqm->dev); 2699 2700 return retval; 2701 } 2702 2703 static int get_wave_state(struct device_queue_manager *dqm, 2704 struct queue *q, 2705 void __user *ctl_stack, 2706 u32 *ctl_stack_used_size, 2707 u32 *save_area_used_size) 2708 { 2709 struct mqd_manager *mqd_mgr; 2710 2711 dqm_lock(dqm); 2712 2713 mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; 2714 2715 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE || 2716 q->properties.is_active || !q->device->kfd->cwsr_enabled || 2717 !mqd_mgr->get_wave_state) { 2718 dqm_unlock(dqm); 2719 return -EINVAL; 2720 } 2721 2722 dqm_unlock(dqm); 2723 2724 /* 2725 * get_wave_state is outside the dqm lock to prevent circular locking 2726 * and the queue should be protected against destruction by the process 2727 * lock. 2728 */ 2729 return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, &q->properties, 2730 ctl_stack, ctl_stack_used_size, save_area_used_size); 2731 } 2732 2733 static void get_queue_checkpoint_info(struct device_queue_manager *dqm, 2734 const struct queue *q, 2735 u32 *mqd_size, 2736 u32 *ctl_stack_size) 2737 { 2738 struct mqd_manager *mqd_mgr; 2739 enum KFD_MQD_TYPE mqd_type = 2740 get_mqd_type_from_queue_type(q->properties.type); 2741 2742 dqm_lock(dqm); 2743 mqd_mgr = dqm->mqd_mgrs[mqd_type]; 2744 *mqd_size = mqd_mgr->mqd_size * NUM_XCC(mqd_mgr->dev->xcc_mask); 2745 *ctl_stack_size = 0; 2746 2747 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE && mqd_mgr->get_checkpoint_info) 2748 mqd_mgr->get_checkpoint_info(mqd_mgr, q->mqd, ctl_stack_size); 2749 2750 dqm_unlock(dqm); 2751 } 2752 2753 static int checkpoint_mqd(struct device_queue_manager *dqm, 2754 const struct queue *q, 2755 void *mqd, 2756 void *ctl_stack) 2757 { 2758 struct mqd_manager *mqd_mgr; 2759 int r = 0; 2760 enum KFD_MQD_TYPE mqd_type = 2761 get_mqd_type_from_queue_type(q->properties.type); 2762 2763 dqm_lock(dqm); 2764 2765 if (q->properties.is_active || !q->device->kfd->cwsr_enabled) { 2766 r = -EINVAL; 2767 goto dqm_unlock; 2768 } 2769 2770 mqd_mgr = dqm->mqd_mgrs[mqd_type]; 2771 if (!mqd_mgr->checkpoint_mqd) { 2772 r = -EOPNOTSUPP; 2773 goto dqm_unlock; 2774 } 2775 2776 mqd_mgr->checkpoint_mqd(mqd_mgr, q->mqd, mqd, ctl_stack); 2777 2778 dqm_unlock: 2779 dqm_unlock(dqm); 2780 return r; 2781 } 2782 2783 static int process_termination_cpsch(struct device_queue_manager *dqm, 2784 struct qcm_process_device *qpd) 2785 { 2786 int retval; 2787 struct queue *q; 2788 struct device *dev = dqm->dev->adev->dev; 2789 struct kernel_queue *kq, *kq_next; 2790 struct mqd_manager *mqd_mgr; 2791 struct device_process_node *cur, *next_dpn; 2792 enum kfd_unmap_queues_filter filter = 2793 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES; 2794 bool found = false; 2795 2796 retval = 0; 2797 2798 dqm_lock(dqm); 2799 2800 /* Clean all kernel queues */ 2801 list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) { 2802 list_del(&kq->list); 2803 decrement_queue_count(dqm, qpd, kq->queue); 2804 qpd->is_debug = false; 2805 dqm->total_queue_count--; 2806 filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES; 2807 } 2808 2809 /* Clear all user mode queues */ 2810 list_for_each_entry(q, &qpd->queues_list, list) { 2811 if (q->properties.type == KFD_QUEUE_TYPE_SDMA) 2812 deallocate_sdma_queue(dqm, q); 2813 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) 2814 deallocate_sdma_queue(dqm, q); 2815 2816 if (q->properties.is_active) { 2817 decrement_queue_count(dqm, qpd, q); 2818 2819 if (dqm->dev->kfd->shared_resources.enable_mes) { 2820 retval = remove_queue_mes(dqm, q, qpd); 2821 if (retval) 2822 dev_err(dev, "Failed to remove queue %d\n", 2823 q->properties.queue_id); 2824 } 2825 } 2826 2827 dqm->total_queue_count--; 2828 } 2829 2830 /* Unregister process */ 2831 list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { 2832 if (qpd == cur->qpd) { 2833 list_del(&cur->list); 2834 kfree(cur); 2835 dqm->processes_count--; 2836 found = true; 2837 break; 2838 } 2839 } 2840 2841 if (!dqm->dev->kfd->shared_resources.enable_mes) 2842 retval = execute_queues_cpsch(dqm, filter, 0, USE_DEFAULT_GRACE_PERIOD); 2843 2844 if ((retval || qpd->reset_wavefronts) && 2845 down_read_trylock(&dqm->dev->adev->reset_domain->sem)) { 2846 pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); 2847 dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process); 2848 qpd->reset_wavefronts = false; 2849 up_read(&dqm->dev->adev->reset_domain->sem); 2850 } 2851 2852 /* Lastly, free mqd resources. 2853 * Do free_mqd() after dqm_unlock to avoid circular locking. 2854 */ 2855 while (!list_empty(&qpd->queues_list)) { 2856 q = list_first_entry(&qpd->queues_list, struct queue, list); 2857 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( 2858 q->properties.type)]; 2859 list_del(&q->list); 2860 qpd->queue_count--; 2861 dqm_unlock(dqm); 2862 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); 2863 dqm_lock(dqm); 2864 } 2865 dqm_unlock(dqm); 2866 2867 /* Outside the DQM lock because under the DQM lock we can't do 2868 * reclaim or take other locks that others hold while reclaiming. 2869 */ 2870 if (found) 2871 kfd_dec_compute_active(dqm->dev); 2872 2873 return retval; 2874 } 2875 2876 static int init_mqd_managers(struct device_queue_manager *dqm) 2877 { 2878 int i, j; 2879 struct device *dev = dqm->dev->adev->dev; 2880 struct mqd_manager *mqd_mgr; 2881 2882 for (i = 0; i < KFD_MQD_TYPE_MAX; i++) { 2883 mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev); 2884 if (!mqd_mgr) { 2885 dev_err(dev, "mqd manager [%d] initialization failed\n", i); 2886 goto out_free; 2887 } 2888 dqm->mqd_mgrs[i] = mqd_mgr; 2889 } 2890 2891 return 0; 2892 2893 out_free: 2894 for (j = 0; j < i; j++) { 2895 kfree(dqm->mqd_mgrs[j]); 2896 dqm->mqd_mgrs[j] = NULL; 2897 } 2898 2899 return -ENOMEM; 2900 } 2901 2902 /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/ 2903 static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) 2904 { 2905 int retval; 2906 struct kfd_node *dev = dqm->dev; 2907 struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; 2908 uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * 2909 get_num_all_sdma_engines(dqm) * 2910 dev->kfd->device_info.num_sdma_queues_per_engine + 2911 (dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * 2912 NUM_XCC(dqm->dev->xcc_mask)); 2913 2914 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size, 2915 &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), 2916 (void *)&(mem_obj->cpu_ptr), false); 2917 2918 return retval; 2919 } 2920 2921 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev) 2922 { 2923 struct device_queue_manager *dqm; 2924 2925 pr_debug("Loading device queue manager\n"); 2926 2927 dqm = kzalloc(sizeof(*dqm), GFP_KERNEL); 2928 if (!dqm) 2929 return NULL; 2930 2931 switch (dev->adev->asic_type) { 2932 /* HWS is not available on Hawaii. */ 2933 case CHIP_HAWAII: 2934 /* HWS depends on CWSR for timely dequeue. CWSR is not 2935 * available on Tonga. 2936 * 2937 * FIXME: This argument also applies to Kaveri. 2938 */ 2939 case CHIP_TONGA: 2940 dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS; 2941 break; 2942 default: 2943 dqm->sched_policy = sched_policy; 2944 break; 2945 } 2946 2947 dqm->dev = dev; 2948 switch (dqm->sched_policy) { 2949 case KFD_SCHED_POLICY_HWS: 2950 case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: 2951 /* initialize dqm for cp scheduling */ 2952 dqm->ops.create_queue = create_queue_cpsch; 2953 dqm->ops.initialize = initialize_cpsch; 2954 dqm->ops.start = start_cpsch; 2955 dqm->ops.stop = stop_cpsch; 2956 dqm->ops.halt = halt_cpsch; 2957 dqm->ops.unhalt = unhalt_cpsch; 2958 dqm->ops.destroy_queue = destroy_queue_cpsch; 2959 dqm->ops.update_queue = update_queue; 2960 dqm->ops.register_process = register_process; 2961 dqm->ops.unregister_process = unregister_process; 2962 dqm->ops.uninitialize = uninitialize; 2963 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch; 2964 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch; 2965 dqm->ops.set_cache_memory_policy = set_cache_memory_policy; 2966 dqm->ops.process_termination = process_termination_cpsch; 2967 dqm->ops.evict_process_queues = evict_process_queues_cpsch; 2968 dqm->ops.restore_process_queues = restore_process_queues_cpsch; 2969 dqm->ops.get_wave_state = get_wave_state; 2970 dqm->ops.reset_queues = reset_queues_cpsch; 2971 dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info; 2972 dqm->ops.checkpoint_mqd = checkpoint_mqd; 2973 break; 2974 case KFD_SCHED_POLICY_NO_HWS: 2975 /* initialize dqm for no cp scheduling */ 2976 dqm->ops.start = start_nocpsch; 2977 dqm->ops.stop = stop_nocpsch; 2978 dqm->ops.create_queue = create_queue_nocpsch; 2979 dqm->ops.destroy_queue = destroy_queue_nocpsch; 2980 dqm->ops.update_queue = update_queue; 2981 dqm->ops.register_process = register_process; 2982 dqm->ops.unregister_process = unregister_process; 2983 dqm->ops.initialize = initialize_nocpsch; 2984 dqm->ops.uninitialize = uninitialize; 2985 dqm->ops.set_cache_memory_policy = set_cache_memory_policy; 2986 dqm->ops.process_termination = process_termination_nocpsch; 2987 dqm->ops.evict_process_queues = evict_process_queues_nocpsch; 2988 dqm->ops.restore_process_queues = 2989 restore_process_queues_nocpsch; 2990 dqm->ops.get_wave_state = get_wave_state; 2991 dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info; 2992 dqm->ops.checkpoint_mqd = checkpoint_mqd; 2993 break; 2994 default: 2995 dev_err(dev->adev->dev, "Invalid scheduling policy %d\n", dqm->sched_policy); 2996 goto out_free; 2997 } 2998 2999 switch (dev->adev->asic_type) { 3000 case CHIP_KAVERI: 3001 case CHIP_HAWAII: 3002 device_queue_manager_init_cik(&dqm->asic_ops); 3003 break; 3004 3005 case CHIP_CARRIZO: 3006 case CHIP_TONGA: 3007 case CHIP_FIJI: 3008 case CHIP_POLARIS10: 3009 case CHIP_POLARIS11: 3010 case CHIP_POLARIS12: 3011 case CHIP_VEGAM: 3012 device_queue_manager_init_vi(&dqm->asic_ops); 3013 break; 3014 3015 default: 3016 if (KFD_GC_VERSION(dev) >= IP_VERSION(12, 0, 0)) 3017 device_queue_manager_init_v12(&dqm->asic_ops); 3018 else if (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0)) 3019 device_queue_manager_init_v11(&dqm->asic_ops); 3020 else if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) 3021 device_queue_manager_init_v10(&dqm->asic_ops); 3022 else if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1)) 3023 device_queue_manager_init_v9(&dqm->asic_ops); 3024 else { 3025 WARN(1, "Unexpected ASIC family %u", 3026 dev->adev->asic_type); 3027 goto out_free; 3028 } 3029 } 3030 3031 if (init_mqd_managers(dqm)) 3032 goto out_free; 3033 3034 if (!dev->kfd->shared_resources.enable_mes && allocate_hiq_sdma_mqd(dqm)) { 3035 dev_err(dev->adev->dev, "Failed to allocate hiq sdma mqd trunk buffer\n"); 3036 goto out_free; 3037 } 3038 3039 if (!dqm->ops.initialize(dqm)) { 3040 init_waitqueue_head(&dqm->destroy_wait); 3041 return dqm; 3042 } 3043 3044 out_free: 3045 kfree(dqm); 3046 return NULL; 3047 } 3048 3049 static void deallocate_hiq_sdma_mqd(struct kfd_node *dev, 3050 struct kfd_mem_obj *mqd) 3051 { 3052 WARN(!mqd, "No hiq sdma mqd trunk to free"); 3053 3054 amdgpu_amdkfd_free_gtt_mem(dev->adev, &mqd->gtt_mem); 3055 } 3056 3057 void device_queue_manager_uninit(struct device_queue_manager *dqm) 3058 { 3059 dqm->ops.stop(dqm); 3060 dqm->ops.uninitialize(dqm); 3061 if (!dqm->dev->kfd->shared_resources.enable_mes) 3062 deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); 3063 kfree(dqm); 3064 } 3065 3066 int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id) 3067 { 3068 struct kfd_process_device *pdd = NULL; 3069 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, &pdd); 3070 struct device_queue_manager *dqm = knode->dqm; 3071 struct device *dev = dqm->dev->adev->dev; 3072 struct qcm_process_device *qpd; 3073 struct queue *q = NULL; 3074 int ret = 0; 3075 3076 if (!pdd) 3077 return -EINVAL; 3078 3079 dqm_lock(dqm); 3080 3081 if (pdd) { 3082 qpd = &pdd->qpd; 3083 3084 list_for_each_entry(q, &qpd->queues_list, list) { 3085 if (q->doorbell_id == doorbell_id && q->properties.is_active) { 3086 ret = suspend_all_queues_mes(dqm); 3087 if (ret) { 3088 dev_err(dev, "Suspending all queues failed"); 3089 goto out; 3090 } 3091 3092 q->properties.is_evicted = true; 3093 q->properties.is_active = false; 3094 decrement_queue_count(dqm, qpd, q); 3095 3096 ret = remove_queue_mes(dqm, q, qpd); 3097 if (ret) { 3098 dev_err(dev, "Removing bad queue failed"); 3099 goto out; 3100 } 3101 3102 ret = resume_all_queues_mes(dqm); 3103 if (ret) 3104 dev_err(dev, "Resuming all queues failed"); 3105 3106 break; 3107 } 3108 } 3109 } 3110 3111 out: 3112 dqm_unlock(dqm); 3113 kfd_unref_process(p); 3114 return ret; 3115 } 3116 3117 int kfd_evict_process_device(struct kfd_process_device *pdd) 3118 { 3119 struct device_queue_manager *dqm; 3120 struct kfd_process *p; 3121 3122 p = pdd->process; 3123 dqm = pdd->dev->dqm; 3124 3125 WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid); 3126 3127 return dqm->ops.evict_process_queues(dqm, &pdd->qpd); 3128 } 3129 3130 int reserve_debug_trap_vmid(struct device_queue_manager *dqm, 3131 struct qcm_process_device *qpd) 3132 { 3133 int r; 3134 struct device *dev = dqm->dev->adev->dev; 3135 int updated_vmid_mask; 3136 3137 if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 3138 dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy); 3139 return -EINVAL; 3140 } 3141 3142 dqm_lock(dqm); 3143 3144 if (dqm->trap_debug_vmid != 0) { 3145 dev_err(dev, "Trap debug id already reserved\n"); 3146 r = -EBUSY; 3147 goto out_unlock; 3148 } 3149 3150 r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 3151 USE_DEFAULT_GRACE_PERIOD, false); 3152 if (r) 3153 goto out_unlock; 3154 3155 updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; 3156 updated_vmid_mask &= ~(1 << dqm->dev->vm_info.last_vmid_kfd); 3157 3158 dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; 3159 dqm->trap_debug_vmid = dqm->dev->vm_info.last_vmid_kfd; 3160 r = set_sched_resources(dqm); 3161 if (r) 3162 goto out_unlock; 3163 3164 r = map_queues_cpsch(dqm); 3165 if (r) 3166 goto out_unlock; 3167 3168 pr_debug("Reserved VMID for trap debug: %i\n", dqm->trap_debug_vmid); 3169 3170 out_unlock: 3171 dqm_unlock(dqm); 3172 return r; 3173 } 3174 3175 /* 3176 * Releases vmid for the trap debugger 3177 */ 3178 int release_debug_trap_vmid(struct device_queue_manager *dqm, 3179 struct qcm_process_device *qpd) 3180 { 3181 struct device *dev = dqm->dev->adev->dev; 3182 int r; 3183 int updated_vmid_mask; 3184 uint32_t trap_debug_vmid; 3185 3186 if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 3187 dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy); 3188 return -EINVAL; 3189 } 3190 3191 dqm_lock(dqm); 3192 trap_debug_vmid = dqm->trap_debug_vmid; 3193 if (dqm->trap_debug_vmid == 0) { 3194 dev_err(dev, "Trap debug id is not reserved\n"); 3195 r = -EINVAL; 3196 goto out_unlock; 3197 } 3198 3199 r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 3200 USE_DEFAULT_GRACE_PERIOD, false); 3201 if (r) 3202 goto out_unlock; 3203 3204 updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; 3205 updated_vmid_mask |= (1 << dqm->dev->vm_info.last_vmid_kfd); 3206 3207 dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; 3208 dqm->trap_debug_vmid = 0; 3209 r = set_sched_resources(dqm); 3210 if (r) 3211 goto out_unlock; 3212 3213 r = map_queues_cpsch(dqm); 3214 if (r) 3215 goto out_unlock; 3216 3217 pr_debug("Released VMID for trap debug: %i\n", trap_debug_vmid); 3218 3219 out_unlock: 3220 dqm_unlock(dqm); 3221 return r; 3222 } 3223 3224 #define QUEUE_NOT_FOUND -1 3225 /* invalidate queue operation in array */ 3226 static void q_array_invalidate(uint32_t num_queues, uint32_t *queue_ids) 3227 { 3228 int i; 3229 3230 for (i = 0; i < num_queues; i++) 3231 queue_ids[i] |= KFD_DBG_QUEUE_INVALID_MASK; 3232 } 3233 3234 /* find queue index in array */ 3235 static int q_array_get_index(unsigned int queue_id, 3236 uint32_t num_queues, 3237 uint32_t *queue_ids) 3238 { 3239 int i; 3240 3241 for (i = 0; i < num_queues; i++) 3242 if (queue_id == (queue_ids[i] & ~KFD_DBG_QUEUE_INVALID_MASK)) 3243 return i; 3244 3245 return QUEUE_NOT_FOUND; 3246 } 3247 3248 struct copy_context_work_handler_workarea { 3249 struct work_struct copy_context_work; 3250 struct kfd_process *p; 3251 }; 3252 3253 static void copy_context_work_handler(struct work_struct *work) 3254 { 3255 struct copy_context_work_handler_workarea *workarea; 3256 struct mqd_manager *mqd_mgr; 3257 struct queue *q; 3258 struct mm_struct *mm; 3259 struct kfd_process *p; 3260 uint32_t tmp_ctl_stack_used_size, tmp_save_area_used_size; 3261 int i; 3262 3263 workarea = container_of(work, 3264 struct copy_context_work_handler_workarea, 3265 copy_context_work); 3266 3267 p = workarea->p; 3268 mm = get_task_mm(p->lead_thread); 3269 3270 if (!mm) 3271 return; 3272 3273 kthread_use_mm(mm); 3274 for (i = 0; i < p->n_pdds; i++) { 3275 struct kfd_process_device *pdd = p->pdds[i]; 3276 struct device_queue_manager *dqm = pdd->dev->dqm; 3277 struct qcm_process_device *qpd = &pdd->qpd; 3278 3279 list_for_each_entry(q, &qpd->queues_list, list) { 3280 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE) 3281 continue; 3282 3283 mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; 3284 3285 /* We ignore the return value from get_wave_state 3286 * because 3287 * i) right now, it always returns 0, and 3288 * ii) if we hit an error, we would continue to the 3289 * next queue anyway. 3290 */ 3291 mqd_mgr->get_wave_state(mqd_mgr, 3292 q->mqd, 3293 &q->properties, 3294 (void __user *) q->properties.ctx_save_restore_area_address, 3295 &tmp_ctl_stack_used_size, 3296 &tmp_save_area_used_size); 3297 } 3298 } 3299 kthread_unuse_mm(mm); 3300 mmput(mm); 3301 } 3302 3303 static uint32_t *get_queue_ids(uint32_t num_queues, uint32_t *usr_queue_id_array) 3304 { 3305 size_t array_size = num_queues * sizeof(uint32_t); 3306 3307 if (!usr_queue_id_array) 3308 return NULL; 3309 3310 return memdup_user(usr_queue_id_array, array_size); 3311 } 3312 3313 int resume_queues(struct kfd_process *p, 3314 uint32_t num_queues, 3315 uint32_t *usr_queue_id_array) 3316 { 3317 uint32_t *queue_ids = NULL; 3318 int total_resumed = 0; 3319 int i; 3320 3321 if (usr_queue_id_array) { 3322 queue_ids = get_queue_ids(num_queues, usr_queue_id_array); 3323 3324 if (IS_ERR(queue_ids)) 3325 return PTR_ERR(queue_ids); 3326 3327 /* mask all queues as invalid. unmask per successful request */ 3328 q_array_invalidate(num_queues, queue_ids); 3329 } 3330 3331 for (i = 0; i < p->n_pdds; i++) { 3332 struct kfd_process_device *pdd = p->pdds[i]; 3333 struct device_queue_manager *dqm = pdd->dev->dqm; 3334 struct device *dev = dqm->dev->adev->dev; 3335 struct qcm_process_device *qpd = &pdd->qpd; 3336 struct queue *q; 3337 int r, per_device_resumed = 0; 3338 3339 dqm_lock(dqm); 3340 3341 /* unmask queues that resume or already resumed as valid */ 3342 list_for_each_entry(q, &qpd->queues_list, list) { 3343 int q_idx = QUEUE_NOT_FOUND; 3344 3345 if (queue_ids) 3346 q_idx = q_array_get_index( 3347 q->properties.queue_id, 3348 num_queues, 3349 queue_ids); 3350 3351 if (!queue_ids || q_idx != QUEUE_NOT_FOUND) { 3352 int err = resume_single_queue(dqm, &pdd->qpd, q); 3353 3354 if (queue_ids) { 3355 if (!err) { 3356 queue_ids[q_idx] &= 3357 ~KFD_DBG_QUEUE_INVALID_MASK; 3358 } else { 3359 queue_ids[q_idx] |= 3360 KFD_DBG_QUEUE_ERROR_MASK; 3361 break; 3362 } 3363 } 3364 3365 if (dqm->dev->kfd->shared_resources.enable_mes) { 3366 wake_up_all(&dqm->destroy_wait); 3367 if (!err) 3368 total_resumed++; 3369 } else { 3370 per_device_resumed++; 3371 } 3372 } 3373 } 3374 3375 if (!per_device_resumed) { 3376 dqm_unlock(dqm); 3377 continue; 3378 } 3379 3380 r = execute_queues_cpsch(dqm, 3381 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 3382 0, 3383 USE_DEFAULT_GRACE_PERIOD); 3384 if (r) { 3385 dev_err(dev, "Failed to resume process queues\n"); 3386 if (queue_ids) { 3387 list_for_each_entry(q, &qpd->queues_list, list) { 3388 int q_idx = q_array_get_index( 3389 q->properties.queue_id, 3390 num_queues, 3391 queue_ids); 3392 3393 /* mask queue as error on resume fail */ 3394 if (q_idx != QUEUE_NOT_FOUND) 3395 queue_ids[q_idx] |= 3396 KFD_DBG_QUEUE_ERROR_MASK; 3397 } 3398 } 3399 } else { 3400 wake_up_all(&dqm->destroy_wait); 3401 total_resumed += per_device_resumed; 3402 } 3403 3404 dqm_unlock(dqm); 3405 } 3406 3407 if (queue_ids) { 3408 if (copy_to_user((void __user *)usr_queue_id_array, queue_ids, 3409 num_queues * sizeof(uint32_t))) 3410 pr_err("copy_to_user failed on queue resume\n"); 3411 3412 kfree(queue_ids); 3413 } 3414 3415 return total_resumed; 3416 } 3417 3418 int suspend_queues(struct kfd_process *p, 3419 uint32_t num_queues, 3420 uint32_t grace_period, 3421 uint64_t exception_clear_mask, 3422 uint32_t *usr_queue_id_array) 3423 { 3424 uint32_t *queue_ids = get_queue_ids(num_queues, usr_queue_id_array); 3425 int total_suspended = 0; 3426 int i; 3427 3428 if (IS_ERR(queue_ids)) 3429 return PTR_ERR(queue_ids); 3430 3431 /* mask all queues as invalid. umask on successful request */ 3432 q_array_invalidate(num_queues, queue_ids); 3433 3434 for (i = 0; i < p->n_pdds; i++) { 3435 struct kfd_process_device *pdd = p->pdds[i]; 3436 struct device_queue_manager *dqm = pdd->dev->dqm; 3437 struct device *dev = dqm->dev->adev->dev; 3438 struct qcm_process_device *qpd = &pdd->qpd; 3439 struct queue *q; 3440 int r, per_device_suspended = 0; 3441 3442 mutex_lock(&p->event_mutex); 3443 dqm_lock(dqm); 3444 3445 /* unmask queues that suspend or already suspended */ 3446 list_for_each_entry(q, &qpd->queues_list, list) { 3447 int q_idx = q_array_get_index(q->properties.queue_id, 3448 num_queues, 3449 queue_ids); 3450 3451 if (q_idx != QUEUE_NOT_FOUND) { 3452 int err = suspend_single_queue(dqm, pdd, q); 3453 bool is_mes = dqm->dev->kfd->shared_resources.enable_mes; 3454 3455 if (!err) { 3456 queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK; 3457 if (exception_clear_mask && is_mes) 3458 q->properties.exception_status &= 3459 ~exception_clear_mask; 3460 3461 if (is_mes) 3462 total_suspended++; 3463 else 3464 per_device_suspended++; 3465 } else if (err != -EBUSY) { 3466 r = err; 3467 queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK; 3468 break; 3469 } 3470 } 3471 } 3472 3473 if (!per_device_suspended) { 3474 dqm_unlock(dqm); 3475 mutex_unlock(&p->event_mutex); 3476 if (total_suspended) 3477 amdgpu_amdkfd_debug_mem_fence(dqm->dev->adev); 3478 continue; 3479 } 3480 3481 r = execute_queues_cpsch(dqm, 3482 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, 3483 grace_period); 3484 3485 if (r) 3486 dev_err(dev, "Failed to suspend process queues.\n"); 3487 else 3488 total_suspended += per_device_suspended; 3489 3490 list_for_each_entry(q, &qpd->queues_list, list) { 3491 int q_idx = q_array_get_index(q->properties.queue_id, 3492 num_queues, queue_ids); 3493 3494 if (q_idx == QUEUE_NOT_FOUND) 3495 continue; 3496 3497 /* mask queue as error on suspend fail */ 3498 if (r) 3499 queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK; 3500 else if (exception_clear_mask) 3501 q->properties.exception_status &= 3502 ~exception_clear_mask; 3503 } 3504 3505 dqm_unlock(dqm); 3506 mutex_unlock(&p->event_mutex); 3507 amdgpu_device_flush_hdp(dqm->dev->adev, NULL); 3508 } 3509 3510 if (total_suspended) { 3511 struct copy_context_work_handler_workarea copy_context_worker; 3512 3513 INIT_WORK_ONSTACK( 3514 ©_context_worker.copy_context_work, 3515 copy_context_work_handler); 3516 3517 copy_context_worker.p = p; 3518 3519 schedule_work(©_context_worker.copy_context_work); 3520 3521 3522 flush_work(©_context_worker.copy_context_work); 3523 destroy_work_on_stack(©_context_worker.copy_context_work); 3524 } 3525 3526 if (copy_to_user((void __user *)usr_queue_id_array, queue_ids, 3527 num_queues * sizeof(uint32_t))) 3528 pr_err("copy_to_user failed on queue suspend\n"); 3529 3530 kfree(queue_ids); 3531 3532 return total_suspended; 3533 } 3534 3535 static uint32_t set_queue_type_for_user(struct queue_properties *q_props) 3536 { 3537 switch (q_props->type) { 3538 case KFD_QUEUE_TYPE_COMPUTE: 3539 return q_props->format == KFD_QUEUE_FORMAT_PM4 3540 ? KFD_IOC_QUEUE_TYPE_COMPUTE 3541 : KFD_IOC_QUEUE_TYPE_COMPUTE_AQL; 3542 case KFD_QUEUE_TYPE_SDMA: 3543 return KFD_IOC_QUEUE_TYPE_SDMA; 3544 case KFD_QUEUE_TYPE_SDMA_XGMI: 3545 return KFD_IOC_QUEUE_TYPE_SDMA_XGMI; 3546 default: 3547 WARN_ONCE(true, "queue type not recognized!"); 3548 return 0xffffffff; 3549 }; 3550 } 3551 3552 void set_queue_snapshot_entry(struct queue *q, 3553 uint64_t exception_clear_mask, 3554 struct kfd_queue_snapshot_entry *qss_entry) 3555 { 3556 qss_entry->ring_base_address = q->properties.queue_address; 3557 qss_entry->write_pointer_address = (uint64_t)q->properties.write_ptr; 3558 qss_entry->read_pointer_address = (uint64_t)q->properties.read_ptr; 3559 qss_entry->ctx_save_restore_address = 3560 q->properties.ctx_save_restore_area_address; 3561 qss_entry->ctx_save_restore_area_size = 3562 q->properties.ctx_save_restore_area_size; 3563 qss_entry->exception_status = q->properties.exception_status; 3564 qss_entry->queue_id = q->properties.queue_id; 3565 qss_entry->gpu_id = q->device->id; 3566 qss_entry->ring_size = (uint32_t)q->properties.queue_size; 3567 qss_entry->queue_type = set_queue_type_for_user(&q->properties); 3568 q->properties.exception_status &= ~exception_clear_mask; 3569 } 3570 3571 int debug_lock_and_unmap(struct device_queue_manager *dqm) 3572 { 3573 struct device *dev = dqm->dev->adev->dev; 3574 int r; 3575 3576 if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 3577 dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy); 3578 return -EINVAL; 3579 } 3580 3581 if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) 3582 return 0; 3583 3584 dqm_lock(dqm); 3585 3586 r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 0, false); 3587 if (r) 3588 dqm_unlock(dqm); 3589 3590 return r; 3591 } 3592 3593 int debug_map_and_unlock(struct device_queue_manager *dqm) 3594 { 3595 struct device *dev = dqm->dev->adev->dev; 3596 int r; 3597 3598 if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 3599 dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy); 3600 return -EINVAL; 3601 } 3602 3603 if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) 3604 return 0; 3605 3606 r = map_queues_cpsch(dqm); 3607 3608 dqm_unlock(dqm); 3609 3610 return r; 3611 } 3612 3613 int debug_refresh_runlist(struct device_queue_manager *dqm) 3614 { 3615 int r = debug_lock_and_unmap(dqm); 3616 3617 if (r) 3618 return r; 3619 3620 return debug_map_and_unlock(dqm); 3621 } 3622 3623 bool kfd_dqm_is_queue_in_process(struct device_queue_manager *dqm, 3624 struct qcm_process_device *qpd, 3625 int doorbell_off, u32 *queue_format) 3626 { 3627 struct queue *q; 3628 bool r = false; 3629 3630 if (!queue_format) 3631 return r; 3632 3633 dqm_lock(dqm); 3634 3635 list_for_each_entry(q, &qpd->queues_list, list) { 3636 if (q->properties.doorbell_off == doorbell_off) { 3637 *queue_format = q->properties.format; 3638 r = true; 3639 goto out; 3640 } 3641 } 3642 3643 out: 3644 dqm_unlock(dqm); 3645 return r; 3646 } 3647 #if defined(CONFIG_DEBUG_FS) 3648 3649 static void seq_reg_dump(struct seq_file *m, 3650 uint32_t (*dump)[2], uint32_t n_regs) 3651 { 3652 uint32_t i, count; 3653 3654 for (i = 0, count = 0; i < n_regs; i++) { 3655 if (count == 0 || 3656 dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) { 3657 seq_printf(m, "%s %08x: %08x", 3658 i ? "\n" : "", 3659 dump[i][0], dump[i][1]); 3660 count = 7; 3661 } else { 3662 seq_printf(m, " %08x", dump[i][1]); 3663 count--; 3664 } 3665 } 3666 3667 seq_puts(m, "\n"); 3668 } 3669 3670 int dqm_debugfs_hqds(struct seq_file *m, void *data) 3671 { 3672 struct device_queue_manager *dqm = data; 3673 uint32_t xcc_mask = dqm->dev->xcc_mask; 3674 uint32_t (*dump)[2], n_regs; 3675 int pipe, queue; 3676 int r = 0, xcc_id; 3677 uint32_t sdma_engine_start; 3678 3679 if (!dqm->sched_running) { 3680 seq_puts(m, " Device is stopped\n"); 3681 return 0; 3682 } 3683 3684 for_each_inst(xcc_id, xcc_mask) { 3685 r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, 3686 KFD_CIK_HIQ_PIPE, 3687 KFD_CIK_HIQ_QUEUE, &dump, 3688 &n_regs, xcc_id); 3689 if (!r) { 3690 seq_printf( 3691 m, 3692 " Inst %d, HIQ on MEC %d Pipe %d Queue %d\n", 3693 xcc_id, 3694 KFD_CIK_HIQ_PIPE / get_pipes_per_mec(dqm) + 1, 3695 KFD_CIK_HIQ_PIPE % get_pipes_per_mec(dqm), 3696 KFD_CIK_HIQ_QUEUE); 3697 seq_reg_dump(m, dump, n_regs); 3698 3699 kfree(dump); 3700 } 3701 3702 for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { 3703 int pipe_offset = pipe * get_queues_per_pipe(dqm); 3704 3705 for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { 3706 if (!test_bit(pipe_offset + queue, 3707 dqm->dev->kfd->shared_resources.cp_queue_bitmap)) 3708 continue; 3709 3710 r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, 3711 pipe, queue, 3712 &dump, &n_regs, 3713 xcc_id); 3714 if (r) 3715 break; 3716 3717 seq_printf(m, 3718 " Inst %d, CP Pipe %d, Queue %d\n", 3719 xcc_id, pipe, queue); 3720 seq_reg_dump(m, dump, n_regs); 3721 3722 kfree(dump); 3723 } 3724 } 3725 } 3726 3727 sdma_engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm); 3728 for (pipe = sdma_engine_start; 3729 pipe < (sdma_engine_start + get_num_all_sdma_engines(dqm)); 3730 pipe++) { 3731 for (queue = 0; 3732 queue < dqm->dev->kfd->device_info.num_sdma_queues_per_engine; 3733 queue++) { 3734 r = dqm->dev->kfd2kgd->hqd_sdma_dump( 3735 dqm->dev->adev, pipe, queue, &dump, &n_regs); 3736 if (r) 3737 break; 3738 3739 seq_printf(m, " SDMA Engine %d, RLC %d\n", 3740 pipe, queue); 3741 seq_reg_dump(m, dump, n_regs); 3742 3743 kfree(dump); 3744 } 3745 } 3746 3747 return r; 3748 } 3749 3750 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm) 3751 { 3752 int r = 0; 3753 3754 dqm_lock(dqm); 3755 r = pm_debugfs_hang_hws(&dqm->packet_mgr); 3756 if (r) { 3757 dqm_unlock(dqm); 3758 return r; 3759 } 3760 dqm->active_runlist = true; 3761 r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 3762 0, USE_DEFAULT_GRACE_PERIOD); 3763 dqm_unlock(dqm); 3764 3765 return r; 3766 } 3767 3768 #endif 3769