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/slab.h> 26 #include <linux/list.h> 27 #include "kfd_device_queue_manager.h" 28 #include "kfd_priv.h" 29 #include "kfd_kernel_queue.h" 30 #include "amdgpu_amdkfd.h" 31 32 static inline struct process_queue_node *get_queue_by_qid( 33 struct process_queue_manager *pqm, unsigned int qid) 34 { 35 struct process_queue_node *pqn; 36 37 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 38 if ((pqn->q && pqn->q->properties.queue_id == qid) || 39 (pqn->kq && pqn->kq->queue->properties.queue_id == qid)) 40 return pqn; 41 } 42 43 return NULL; 44 } 45 46 static int assign_queue_slot_by_qid(struct process_queue_manager *pqm, 47 unsigned int qid) 48 { 49 if (qid >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 50 return -EINVAL; 51 52 if (__test_and_set_bit(qid, pqm->queue_slot_bitmap)) { 53 pr_err("Cannot create new queue because requested qid(%u) is in use\n", qid); 54 return -ENOSPC; 55 } 56 57 return 0; 58 } 59 60 static int find_available_queue_slot(struct process_queue_manager *pqm, 61 unsigned int *qid) 62 { 63 unsigned long found; 64 65 found = find_first_zero_bit(pqm->queue_slot_bitmap, 66 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 67 68 pr_debug("The new slot id %lu\n", found); 69 70 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { 71 pr_info("Cannot open more queues for process with pasid 0x%x\n", 72 pqm->process->pasid); 73 return -ENOMEM; 74 } 75 76 set_bit(found, pqm->queue_slot_bitmap); 77 *qid = found; 78 79 return 0; 80 } 81 82 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd) 83 { 84 struct kfd_dev *dev = pdd->dev; 85 86 if (pdd->already_dequeued) 87 return; 88 89 dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd); 90 pdd->already_dequeued = true; 91 } 92 93 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 94 void *gws) 95 { 96 struct kfd_dev *dev = NULL; 97 struct process_queue_node *pqn; 98 struct kfd_process_device *pdd; 99 struct kgd_mem *mem = NULL; 100 int ret; 101 102 pqn = get_queue_by_qid(pqm, qid); 103 if (!pqn) { 104 pr_err("Queue id does not match any known queue\n"); 105 return -EINVAL; 106 } 107 108 if (pqn->q) 109 dev = pqn->q->device; 110 if (WARN_ON(!dev)) 111 return -ENODEV; 112 113 pdd = kfd_get_process_device_data(dev, pqm->process); 114 if (!pdd) { 115 pr_err("Process device data doesn't exist\n"); 116 return -EINVAL; 117 } 118 119 /* Only allow one queue per process can have GWS assigned */ 120 if (gws && pdd->qpd.num_gws) 121 return -EBUSY; 122 123 if (!gws && pdd->qpd.num_gws == 0) 124 return -EINVAL; 125 126 if (gws) 127 ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info, 128 gws, &mem); 129 else 130 ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info, 131 pqn->q->gws); 132 if (unlikely(ret)) 133 return ret; 134 135 pqn->q->gws = mem; 136 pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0; 137 138 return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 139 pqn->q, NULL); 140 } 141 142 void kfd_process_dequeue_from_all_devices(struct kfd_process *p) 143 { 144 int i; 145 146 for (i = 0; i < p->n_pdds; i++) 147 kfd_process_dequeue_from_device(p->pdds[i]); 148 } 149 150 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p) 151 { 152 INIT_LIST_HEAD(&pqm->queues); 153 pqm->queue_slot_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 154 GFP_KERNEL); 155 if (!pqm->queue_slot_bitmap) 156 return -ENOMEM; 157 pqm->process = p; 158 159 return 0; 160 } 161 162 void pqm_uninit(struct process_queue_manager *pqm) 163 { 164 struct process_queue_node *pqn, *next; 165 166 list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) { 167 if (pqn->q && pqn->q->gws) 168 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info, 169 pqn->q->gws); 170 kfd_procfs_del_queue(pqn->q); 171 uninit_queue(pqn->q); 172 list_del(&pqn->process_queue_list); 173 kfree(pqn); 174 } 175 176 bitmap_free(pqm->queue_slot_bitmap); 177 pqm->queue_slot_bitmap = NULL; 178 } 179 180 static int init_user_queue(struct process_queue_manager *pqm, 181 struct kfd_dev *dev, struct queue **q, 182 struct queue_properties *q_properties, 183 struct file *f, unsigned int qid) 184 { 185 int retval; 186 187 /* Doorbell initialized in user space*/ 188 q_properties->doorbell_ptr = NULL; 189 190 /* let DQM handle it*/ 191 q_properties->vmid = 0; 192 q_properties->queue_id = qid; 193 194 retval = init_queue(q, q_properties); 195 if (retval != 0) 196 return retval; 197 198 (*q)->device = dev; 199 (*q)->process = pqm->process; 200 201 if (dev->shared_resources.enable_mes) { 202 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, 203 AMDGPU_MES_GANG_CTX_SIZE, 204 &(*q)->gang_ctx_bo, 205 &(*q)->gang_ctx_gpu_addr, 206 &(*q)->gang_ctx_cpu_ptr, 207 false); 208 if (retval) { 209 pr_err("failed to allocate gang context bo\n"); 210 goto cleanup; 211 } 212 memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE); 213 } 214 215 pr_debug("PQM After init queue"); 216 return 0; 217 218 cleanup: 219 if (dev->shared_resources.enable_mes) 220 uninit_queue(*q); 221 return retval; 222 } 223 224 int pqm_create_queue(struct process_queue_manager *pqm, 225 struct kfd_dev *dev, 226 struct file *f, 227 struct queue_properties *properties, 228 unsigned int *qid, 229 const struct kfd_criu_queue_priv_data *q_data, 230 const void *restore_mqd, 231 const void *restore_ctl_stack, 232 uint32_t *p_doorbell_offset_in_process) 233 { 234 int retval; 235 struct kfd_process_device *pdd; 236 struct queue *q; 237 struct process_queue_node *pqn; 238 struct kernel_queue *kq; 239 enum kfd_queue_type type = properties->type; 240 unsigned int max_queues = 127; /* HWS limit */ 241 242 q = NULL; 243 kq = NULL; 244 245 pdd = kfd_get_process_device_data(dev, pqm->process); 246 if (!pdd) { 247 pr_err("Process device data doesn't exist\n"); 248 return -1; 249 } 250 251 /* 252 * for debug process, verify that it is within the static queues limit 253 * currently limit is set to half of the total avail HQD slots 254 * If we are just about to create DIQ, the is_debug flag is not set yet 255 * Hence we also check the type as well 256 */ 257 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ)) 258 max_queues = dev->device_info.max_no_of_hqd/2; 259 260 if (pdd->qpd.queue_count >= max_queues) 261 return -ENOSPC; 262 263 if (q_data) { 264 retval = assign_queue_slot_by_qid(pqm, q_data->q_id); 265 *qid = q_data->q_id; 266 } else 267 retval = find_available_queue_slot(pqm, qid); 268 269 if (retval != 0) 270 return retval; 271 272 if (list_empty(&pdd->qpd.queues_list) && 273 list_empty(&pdd->qpd.priv_queue_list)) 274 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd); 275 276 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL); 277 if (!pqn) { 278 retval = -ENOMEM; 279 goto err_allocate_pqn; 280 } 281 282 switch (type) { 283 case KFD_QUEUE_TYPE_SDMA: 284 case KFD_QUEUE_TYPE_SDMA_XGMI: 285 /* SDMA queues are always allocated statically no matter 286 * which scheduler mode is used. We also do not need to 287 * check whether a SDMA queue can be allocated here, because 288 * allocate_sdma_queue() in create_queue() has the 289 * corresponding check logic. 290 */ 291 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 292 if (retval != 0) 293 goto err_create_queue; 294 pqn->q = q; 295 pqn->kq = NULL; 296 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data, 297 restore_mqd, restore_ctl_stack); 298 print_queue(q); 299 break; 300 301 case KFD_QUEUE_TYPE_COMPUTE: 302 /* check if there is over subscription */ 303 if ((dev->dqm->sched_policy == 304 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) && 305 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) || 306 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) { 307 pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n"); 308 retval = -EPERM; 309 goto err_create_queue; 310 } 311 312 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 313 if (retval != 0) 314 goto err_create_queue; 315 pqn->q = q; 316 pqn->kq = NULL; 317 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data, 318 restore_mqd, restore_ctl_stack); 319 print_queue(q); 320 break; 321 case KFD_QUEUE_TYPE_DIQ: 322 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ); 323 if (!kq) { 324 retval = -ENOMEM; 325 goto err_create_queue; 326 } 327 kq->queue->properties.queue_id = *qid; 328 pqn->kq = kq; 329 pqn->q = NULL; 330 retval = dev->dqm->ops.create_kernel_queue(dev->dqm, 331 kq, &pdd->qpd); 332 break; 333 default: 334 WARN(1, "Invalid queue type %d", type); 335 retval = -EINVAL; 336 } 337 338 if (retval != 0) { 339 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n", 340 pqm->process->pasid, type, retval); 341 goto err_create_queue; 342 } 343 344 if (q && p_doorbell_offset_in_process) 345 /* Return the doorbell offset within the doorbell page 346 * to the caller so it can be passed up to user mode 347 * (in bytes). 348 * There are always 1024 doorbells per process, so in case 349 * of 8-byte doorbells, there are two doorbell pages per 350 * process. 351 */ 352 *p_doorbell_offset_in_process = 353 (q->properties.doorbell_off * sizeof(uint32_t)) & 354 (kfd_doorbell_process_slice(dev) - 1); 355 356 pr_debug("PQM After DQM create queue\n"); 357 358 list_add(&pqn->process_queue_list, &pqm->queues); 359 360 if (q) { 361 pr_debug("PQM done creating queue\n"); 362 kfd_procfs_add_queue(q); 363 print_queue_properties(&q->properties); 364 } 365 366 return retval; 367 368 err_create_queue: 369 uninit_queue(q); 370 if (kq) 371 kernel_queue_uninit(kq, false); 372 kfree(pqn); 373 err_allocate_pqn: 374 /* check if queues list is empty unregister process from device */ 375 clear_bit(*qid, pqm->queue_slot_bitmap); 376 if (list_empty(&pdd->qpd.queues_list) && 377 list_empty(&pdd->qpd.priv_queue_list)) 378 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd); 379 return retval; 380 } 381 382 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) 383 { 384 struct process_queue_node *pqn; 385 struct kfd_process_device *pdd; 386 struct device_queue_manager *dqm; 387 struct kfd_dev *dev; 388 int retval; 389 390 dqm = NULL; 391 392 retval = 0; 393 394 pqn = get_queue_by_qid(pqm, qid); 395 if (!pqn) { 396 pr_err("Queue id does not match any known queue\n"); 397 return -EINVAL; 398 } 399 400 dev = NULL; 401 if (pqn->kq) 402 dev = pqn->kq->dev; 403 if (pqn->q) 404 dev = pqn->q->device; 405 if (WARN_ON(!dev)) 406 return -ENODEV; 407 408 pdd = kfd_get_process_device_data(dev, pqm->process); 409 if (!pdd) { 410 pr_err("Process device data doesn't exist\n"); 411 return -1; 412 } 413 414 if (pqn->kq) { 415 /* destroy kernel queue (DIQ) */ 416 dqm = pqn->kq->dev->dqm; 417 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd); 418 kernel_queue_uninit(pqn->kq, false); 419 } 420 421 if (pqn->q) { 422 kfd_procfs_del_queue(pqn->q); 423 dqm = pqn->q->device->dqm; 424 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q); 425 if (retval) { 426 pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n", 427 pqm->process->pasid, 428 pqn->q->properties.queue_id, retval); 429 if (retval != -ETIME) 430 goto err_destroy_queue; 431 } 432 433 if (pqn->q->gws) { 434 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info, 435 pqn->q->gws); 436 pdd->qpd.num_gws = 0; 437 } 438 439 if (dev->shared_resources.enable_mes) 440 amdgpu_amdkfd_free_gtt_mem(dev->adev, 441 pqn->q->gang_ctx_bo); 442 uninit_queue(pqn->q); 443 } 444 445 list_del(&pqn->process_queue_list); 446 kfree(pqn); 447 clear_bit(qid, pqm->queue_slot_bitmap); 448 449 if (list_empty(&pdd->qpd.queues_list) && 450 list_empty(&pdd->qpd.priv_queue_list)) 451 dqm->ops.unregister_process(dqm, &pdd->qpd); 452 453 err_destroy_queue: 454 return retval; 455 } 456 457 int pqm_update_queue_properties(struct process_queue_manager *pqm, 458 unsigned int qid, struct queue_properties *p) 459 { 460 int retval; 461 struct process_queue_node *pqn; 462 463 pqn = get_queue_by_qid(pqm, qid); 464 if (!pqn) { 465 pr_debug("No queue %d exists for update operation\n", qid); 466 return -EFAULT; 467 } 468 469 pqn->q->properties.queue_address = p->queue_address; 470 pqn->q->properties.queue_size = p->queue_size; 471 pqn->q->properties.queue_percent = p->queue_percent; 472 pqn->q->properties.priority = p->priority; 473 474 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 475 pqn->q, NULL); 476 if (retval != 0) 477 return retval; 478 479 return 0; 480 } 481 482 int pqm_update_mqd(struct process_queue_manager *pqm, 483 unsigned int qid, struct mqd_update_info *minfo) 484 { 485 int retval; 486 struct process_queue_node *pqn; 487 488 pqn = get_queue_by_qid(pqm, qid); 489 if (!pqn) { 490 pr_debug("No queue %d exists for update operation\n", qid); 491 return -EFAULT; 492 } 493 494 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 495 pqn->q, minfo); 496 if (retval != 0) 497 return retval; 498 499 return 0; 500 } 501 502 struct kernel_queue *pqm_get_kernel_queue( 503 struct process_queue_manager *pqm, 504 unsigned int qid) 505 { 506 struct process_queue_node *pqn; 507 508 pqn = get_queue_by_qid(pqm, qid); 509 if (pqn && pqn->kq) 510 return pqn->kq; 511 512 return NULL; 513 } 514 515 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 516 unsigned int qid) 517 { 518 struct process_queue_node *pqn; 519 520 pqn = get_queue_by_qid(pqm, qid); 521 return pqn ? pqn->q : NULL; 522 } 523 524 int pqm_get_wave_state(struct process_queue_manager *pqm, 525 unsigned int qid, 526 void __user *ctl_stack, 527 u32 *ctl_stack_used_size, 528 u32 *save_area_used_size) 529 { 530 struct process_queue_node *pqn; 531 532 pqn = get_queue_by_qid(pqm, qid); 533 if (!pqn) { 534 pr_debug("amdkfd: No queue %d exists for operation\n", 535 qid); 536 return -EFAULT; 537 } 538 539 return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm, 540 pqn->q, 541 ctl_stack, 542 ctl_stack_used_size, 543 save_area_used_size); 544 } 545 546 static int get_queue_data_sizes(struct kfd_process_device *pdd, 547 struct queue *q, 548 uint32_t *mqd_size, 549 uint32_t *ctl_stack_size) 550 { 551 int ret; 552 553 ret = pqm_get_queue_checkpoint_info(&pdd->process->pqm, 554 q->properties.queue_id, 555 mqd_size, 556 ctl_stack_size); 557 if (ret) 558 pr_err("Failed to get queue dump info (%d)\n", ret); 559 560 return ret; 561 } 562 563 int kfd_process_get_queue_info(struct kfd_process *p, 564 uint32_t *num_queues, 565 uint64_t *priv_data_sizes) 566 { 567 uint32_t extra_data_sizes = 0; 568 struct queue *q; 569 int i; 570 int ret; 571 572 *num_queues = 0; 573 574 /* Run over all PDDs of the process */ 575 for (i = 0; i < p->n_pdds; i++) { 576 struct kfd_process_device *pdd = p->pdds[i]; 577 578 list_for_each_entry(q, &pdd->qpd.queues_list, list) { 579 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || 580 q->properties.type == KFD_QUEUE_TYPE_SDMA || 581 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { 582 uint32_t mqd_size, ctl_stack_size; 583 584 *num_queues = *num_queues + 1; 585 586 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size); 587 if (ret) 588 return ret; 589 590 extra_data_sizes += mqd_size + ctl_stack_size; 591 } else { 592 pr_err("Unsupported queue type (%d)\n", q->properties.type); 593 return -EOPNOTSUPP; 594 } 595 } 596 } 597 *priv_data_sizes = extra_data_sizes + 598 (*num_queues * sizeof(struct kfd_criu_queue_priv_data)); 599 600 return 0; 601 } 602 603 static int pqm_checkpoint_mqd(struct process_queue_manager *pqm, 604 unsigned int qid, 605 void *mqd, 606 void *ctl_stack) 607 { 608 struct process_queue_node *pqn; 609 610 pqn = get_queue_by_qid(pqm, qid); 611 if (!pqn) { 612 pr_debug("amdkfd: No queue %d exists for operation\n", qid); 613 return -EFAULT; 614 } 615 616 if (!pqn->q->device->dqm->ops.checkpoint_mqd) { 617 pr_err("amdkfd: queue dumping not supported on this device\n"); 618 return -EOPNOTSUPP; 619 } 620 621 return pqn->q->device->dqm->ops.checkpoint_mqd(pqn->q->device->dqm, 622 pqn->q, mqd, ctl_stack); 623 } 624 625 static int criu_checkpoint_queue(struct kfd_process_device *pdd, 626 struct queue *q, 627 struct kfd_criu_queue_priv_data *q_data) 628 { 629 uint8_t *mqd, *ctl_stack; 630 int ret; 631 632 mqd = (void *)(q_data + 1); 633 ctl_stack = mqd + q_data->mqd_size; 634 635 q_data->gpu_id = pdd->user_gpu_id; 636 q_data->type = q->properties.type; 637 q_data->format = q->properties.format; 638 q_data->q_id = q->properties.queue_id; 639 q_data->q_address = q->properties.queue_address; 640 q_data->q_size = q->properties.queue_size; 641 q_data->priority = q->properties.priority; 642 q_data->q_percent = q->properties.queue_percent; 643 q_data->read_ptr_addr = (uint64_t)q->properties.read_ptr; 644 q_data->write_ptr_addr = (uint64_t)q->properties.write_ptr; 645 q_data->doorbell_id = q->doorbell_id; 646 647 q_data->sdma_id = q->sdma_id; 648 649 q_data->eop_ring_buffer_address = 650 q->properties.eop_ring_buffer_address; 651 652 q_data->eop_ring_buffer_size = q->properties.eop_ring_buffer_size; 653 654 q_data->ctx_save_restore_area_address = 655 q->properties.ctx_save_restore_area_address; 656 657 q_data->ctx_save_restore_area_size = 658 q->properties.ctx_save_restore_area_size; 659 660 q_data->gws = !!q->gws; 661 662 ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack); 663 if (ret) { 664 pr_err("Failed checkpoint queue_mqd (%d)\n", ret); 665 return ret; 666 } 667 668 pr_debug("Dumping Queue: gpu_id:%x queue_id:%u\n", q_data->gpu_id, q_data->q_id); 669 return ret; 670 } 671 672 static int criu_checkpoint_queues_device(struct kfd_process_device *pdd, 673 uint8_t __user *user_priv, 674 unsigned int *q_index, 675 uint64_t *queues_priv_data_offset) 676 { 677 unsigned int q_private_data_size = 0; 678 uint8_t *q_private_data = NULL; /* Local buffer to store individual queue private data */ 679 struct queue *q; 680 int ret = 0; 681 682 list_for_each_entry(q, &pdd->qpd.queues_list, list) { 683 struct kfd_criu_queue_priv_data *q_data; 684 uint64_t q_data_size; 685 uint32_t mqd_size; 686 uint32_t ctl_stack_size; 687 688 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE && 689 q->properties.type != KFD_QUEUE_TYPE_SDMA && 690 q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI) { 691 692 pr_err("Unsupported queue type (%d)\n", q->properties.type); 693 ret = -EOPNOTSUPP; 694 break; 695 } 696 697 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size); 698 if (ret) 699 break; 700 701 q_data_size = sizeof(*q_data) + mqd_size + ctl_stack_size; 702 703 /* Increase local buffer space if needed */ 704 if (q_private_data_size < q_data_size) { 705 kfree(q_private_data); 706 707 q_private_data = kzalloc(q_data_size, GFP_KERNEL); 708 if (!q_private_data) { 709 ret = -ENOMEM; 710 break; 711 } 712 q_private_data_size = q_data_size; 713 } 714 715 q_data = (struct kfd_criu_queue_priv_data *)q_private_data; 716 717 /* data stored in this order: priv_data, mqd, ctl_stack */ 718 q_data->mqd_size = mqd_size; 719 q_data->ctl_stack_size = ctl_stack_size; 720 721 ret = criu_checkpoint_queue(pdd, q, q_data); 722 if (ret) 723 break; 724 725 q_data->object_type = KFD_CRIU_OBJECT_TYPE_QUEUE; 726 727 ret = copy_to_user(user_priv + *queues_priv_data_offset, 728 q_data, q_data_size); 729 if (ret) { 730 ret = -EFAULT; 731 break; 732 } 733 *queues_priv_data_offset += q_data_size; 734 *q_index = *q_index + 1; 735 } 736 737 kfree(q_private_data); 738 739 return ret; 740 } 741 742 int kfd_criu_checkpoint_queues(struct kfd_process *p, 743 uint8_t __user *user_priv_data, 744 uint64_t *priv_data_offset) 745 { 746 int ret = 0, pdd_index, q_index = 0; 747 748 for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) { 749 struct kfd_process_device *pdd = p->pdds[pdd_index]; 750 751 /* 752 * criu_checkpoint_queues_device will copy data to user and update q_index and 753 * queues_priv_data_offset 754 */ 755 ret = criu_checkpoint_queues_device(pdd, user_priv_data, &q_index, 756 priv_data_offset); 757 758 if (ret) 759 break; 760 } 761 762 return ret; 763 } 764 765 static void set_queue_properties_from_criu(struct queue_properties *qp, 766 struct kfd_criu_queue_priv_data *q_data) 767 { 768 qp->is_interop = false; 769 qp->queue_percent = q_data->q_percent; 770 qp->priority = q_data->priority; 771 qp->queue_address = q_data->q_address; 772 qp->queue_size = q_data->q_size; 773 qp->read_ptr = (uint32_t *) q_data->read_ptr_addr; 774 qp->write_ptr = (uint32_t *) q_data->write_ptr_addr; 775 qp->eop_ring_buffer_address = q_data->eop_ring_buffer_address; 776 qp->eop_ring_buffer_size = q_data->eop_ring_buffer_size; 777 qp->ctx_save_restore_area_address = q_data->ctx_save_restore_area_address; 778 qp->ctx_save_restore_area_size = q_data->ctx_save_restore_area_size; 779 qp->ctl_stack_size = q_data->ctl_stack_size; 780 qp->type = q_data->type; 781 qp->format = q_data->format; 782 } 783 784 int kfd_criu_restore_queue(struct kfd_process *p, 785 uint8_t __user *user_priv_ptr, 786 uint64_t *priv_data_offset, 787 uint64_t max_priv_data_size) 788 { 789 uint8_t *mqd, *ctl_stack, *q_extra_data = NULL; 790 struct kfd_criu_queue_priv_data *q_data; 791 struct kfd_process_device *pdd; 792 uint64_t q_extra_data_size; 793 struct queue_properties qp; 794 unsigned int queue_id; 795 int ret = 0; 796 797 if (*priv_data_offset + sizeof(*q_data) > max_priv_data_size) 798 return -EINVAL; 799 800 q_data = kmalloc(sizeof(*q_data), GFP_KERNEL); 801 if (!q_data) 802 return -ENOMEM; 803 804 ret = copy_from_user(q_data, user_priv_ptr + *priv_data_offset, sizeof(*q_data)); 805 if (ret) { 806 ret = -EFAULT; 807 goto exit; 808 } 809 810 *priv_data_offset += sizeof(*q_data); 811 q_extra_data_size = (uint64_t)q_data->ctl_stack_size + q_data->mqd_size; 812 813 if (*priv_data_offset + q_extra_data_size > max_priv_data_size) { 814 ret = -EINVAL; 815 goto exit; 816 } 817 818 q_extra_data = kmalloc(q_extra_data_size, GFP_KERNEL); 819 if (!q_extra_data) { 820 ret = -ENOMEM; 821 goto exit; 822 } 823 824 ret = copy_from_user(q_extra_data, user_priv_ptr + *priv_data_offset, q_extra_data_size); 825 if (ret) { 826 ret = -EFAULT; 827 goto exit; 828 } 829 830 *priv_data_offset += q_extra_data_size; 831 832 pdd = kfd_process_device_data_by_id(p, q_data->gpu_id); 833 if (!pdd) { 834 pr_err("Failed to get pdd\n"); 835 ret = -EINVAL; 836 goto exit; 837 } 838 /* data stored in this order: mqd, ctl_stack */ 839 mqd = q_extra_data; 840 ctl_stack = mqd + q_data->mqd_size; 841 842 memset(&qp, 0, sizeof(qp)); 843 set_queue_properties_from_criu(&qp, q_data); 844 845 print_queue_properties(&qp); 846 847 ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, q_data, mqd, ctl_stack, 848 NULL); 849 if (ret) { 850 pr_err("Failed to create new queue err:%d\n", ret); 851 goto exit; 852 } 853 854 if (q_data->gws) 855 ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws); 856 857 exit: 858 if (ret) 859 pr_err("Failed to restore queue (%d)\n", ret); 860 else 861 pr_debug("Queue id %d was restored successfully\n", queue_id); 862 863 kfree(q_data); 864 865 return ret; 866 } 867 868 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm, 869 unsigned int qid, 870 uint32_t *mqd_size, 871 uint32_t *ctl_stack_size) 872 { 873 struct process_queue_node *pqn; 874 875 pqn = get_queue_by_qid(pqm, qid); 876 if (!pqn) { 877 pr_debug("amdkfd: No queue %d exists for operation\n", qid); 878 return -EFAULT; 879 } 880 881 if (!pqn->q->device->dqm->ops.get_queue_checkpoint_info) { 882 pr_err("amdkfd: queue dumping not supported on this device\n"); 883 return -EOPNOTSUPP; 884 } 885 886 pqn->q->device->dqm->ops.get_queue_checkpoint_info(pqn->q->device->dqm, 887 pqn->q, mqd_size, 888 ctl_stack_size); 889 return 0; 890 } 891 892 #if defined(CONFIG_DEBUG_FS) 893 894 int pqm_debugfs_mqds(struct seq_file *m, void *data) 895 { 896 struct process_queue_manager *pqm = data; 897 struct process_queue_node *pqn; 898 struct queue *q; 899 enum KFD_MQD_TYPE mqd_type; 900 struct mqd_manager *mqd_mgr; 901 int r = 0; 902 903 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 904 if (pqn->q) { 905 q = pqn->q; 906 switch (q->properties.type) { 907 case KFD_QUEUE_TYPE_SDMA: 908 case KFD_QUEUE_TYPE_SDMA_XGMI: 909 seq_printf(m, " SDMA queue on device %x\n", 910 q->device->id); 911 mqd_type = KFD_MQD_TYPE_SDMA; 912 break; 913 case KFD_QUEUE_TYPE_COMPUTE: 914 seq_printf(m, " Compute queue on device %x\n", 915 q->device->id); 916 mqd_type = KFD_MQD_TYPE_CP; 917 break; 918 default: 919 seq_printf(m, 920 " Bad user queue type %d on device %x\n", 921 q->properties.type, q->device->id); 922 continue; 923 } 924 mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type]; 925 } else if (pqn->kq) { 926 q = pqn->kq->queue; 927 mqd_mgr = pqn->kq->mqd_mgr; 928 switch (q->properties.type) { 929 case KFD_QUEUE_TYPE_DIQ: 930 seq_printf(m, " DIQ on device %x\n", 931 pqn->kq->dev->id); 932 break; 933 default: 934 seq_printf(m, 935 " Bad kernel queue type %d on device %x\n", 936 q->properties.type, 937 pqn->kq->dev->id); 938 continue; 939 } 940 } else { 941 seq_printf(m, 942 " Weird: Queue node with neither kernel nor user queue\n"); 943 continue; 944 } 945 946 r = mqd_mgr->debugfs_show_mqd(m, q->mqd); 947 if (r != 0) 948 break; 949 } 950 951 return r; 952 } 953 954 #endif 955