1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/slab.h> 25 #include <linux/list.h> 26 #include "kfd_device_queue_manager.h" 27 #include "kfd_priv.h" 28 #include "kfd_kernel_queue.h" 29 #include "amdgpu_amdkfd.h" 30 31 static inline struct process_queue_node *get_queue_by_qid( 32 struct process_queue_manager *pqm, unsigned int qid) 33 { 34 struct process_queue_node *pqn; 35 36 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 37 if ((pqn->q && pqn->q->properties.queue_id == qid) || 38 (pqn->kq && pqn->kq->queue->properties.queue_id == qid)) 39 return pqn; 40 } 41 42 return NULL; 43 } 44 45 static int find_available_queue_slot(struct process_queue_manager *pqm, 46 unsigned int *qid) 47 { 48 unsigned long found; 49 50 found = find_first_zero_bit(pqm->queue_slot_bitmap, 51 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 52 53 pr_debug("The new slot id %lu\n", found); 54 55 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { 56 pr_info("Cannot open more queues for process with pasid 0x%x\n", 57 pqm->process->pasid); 58 return -ENOMEM; 59 } 60 61 set_bit(found, pqm->queue_slot_bitmap); 62 *qid = found; 63 64 return 0; 65 } 66 67 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd) 68 { 69 struct kfd_dev *dev = pdd->dev; 70 71 if (pdd->already_dequeued) 72 return; 73 74 dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd); 75 pdd->already_dequeued = true; 76 } 77 78 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 79 void *gws) 80 { 81 struct kfd_dev *dev = NULL; 82 struct process_queue_node *pqn; 83 struct kfd_process_device *pdd; 84 struct kgd_mem *mem = NULL; 85 int ret; 86 87 pqn = get_queue_by_qid(pqm, qid); 88 if (!pqn) { 89 pr_err("Queue id does not match any known queue\n"); 90 return -EINVAL; 91 } 92 93 if (pqn->q) 94 dev = pqn->q->device; 95 if (WARN_ON(!dev)) 96 return -ENODEV; 97 98 pdd = kfd_get_process_device_data(dev, pqm->process); 99 if (!pdd) { 100 pr_err("Process device data doesn't exist\n"); 101 return -EINVAL; 102 } 103 104 /* Only allow one queue per process can have GWS assigned */ 105 if (gws && pdd->qpd.num_gws) 106 return -EBUSY; 107 108 if (!gws && pdd->qpd.num_gws == 0) 109 return -EINVAL; 110 111 if (gws) 112 ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info, 113 gws, &mem); 114 else 115 ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info, 116 pqn->q->gws); 117 if (unlikely(ret)) 118 return ret; 119 120 pqn->q->gws = mem; 121 pdd->qpd.num_gws = gws ? amdgpu_amdkfd_get_num_gws(dev->kgd) : 0; 122 123 return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 124 pqn->q); 125 } 126 127 void kfd_process_dequeue_from_all_devices(struct kfd_process *p) 128 { 129 int i; 130 131 for (i = 0; i < p->n_pdds; i++) 132 kfd_process_dequeue_from_device(p->pdds[i]); 133 } 134 135 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p) 136 { 137 INIT_LIST_HEAD(&pqm->queues); 138 pqm->queue_slot_bitmap = 139 kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 140 BITS_PER_BYTE), GFP_KERNEL); 141 if (!pqm->queue_slot_bitmap) 142 return -ENOMEM; 143 pqm->process = p; 144 145 return 0; 146 } 147 148 void pqm_uninit(struct process_queue_manager *pqm) 149 { 150 struct process_queue_node *pqn, *next; 151 152 list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) { 153 if (pqn->q && pqn->q->gws) 154 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info, 155 pqn->q->gws); 156 kfd_procfs_del_queue(pqn->q); 157 uninit_queue(pqn->q); 158 list_del(&pqn->process_queue_list); 159 kfree(pqn); 160 } 161 162 kfree(pqm->queue_slot_bitmap); 163 pqm->queue_slot_bitmap = NULL; 164 } 165 166 static int init_user_queue(struct process_queue_manager *pqm, 167 struct kfd_dev *dev, struct queue **q, 168 struct queue_properties *q_properties, 169 struct file *f, unsigned int qid) 170 { 171 int retval; 172 173 /* Doorbell initialized in user space*/ 174 q_properties->doorbell_ptr = NULL; 175 176 /* let DQM handle it*/ 177 q_properties->vmid = 0; 178 q_properties->queue_id = qid; 179 180 retval = init_queue(q, q_properties); 181 if (retval != 0) 182 return retval; 183 184 (*q)->device = dev; 185 (*q)->process = pqm->process; 186 187 pr_debug("PQM After init queue"); 188 189 return retval; 190 } 191 192 int pqm_create_queue(struct process_queue_manager *pqm, 193 struct kfd_dev *dev, 194 struct file *f, 195 struct queue_properties *properties, 196 unsigned int *qid, 197 uint32_t *p_doorbell_offset_in_process) 198 { 199 int retval; 200 struct kfd_process_device *pdd; 201 struct queue *q; 202 struct process_queue_node *pqn; 203 struct kernel_queue *kq; 204 enum kfd_queue_type type = properties->type; 205 unsigned int max_queues = 127; /* HWS limit */ 206 207 q = NULL; 208 kq = NULL; 209 210 pdd = kfd_get_process_device_data(dev, pqm->process); 211 if (!pdd) { 212 pr_err("Process device data doesn't exist\n"); 213 return -1; 214 } 215 216 /* 217 * for debug process, verify that it is within the static queues limit 218 * currently limit is set to half of the total avail HQD slots 219 * If we are just about to create DIQ, the is_debug flag is not set yet 220 * Hence we also check the type as well 221 */ 222 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ)) 223 max_queues = dev->device_info->max_no_of_hqd/2; 224 225 if (pdd->qpd.queue_count >= max_queues) 226 return -ENOSPC; 227 228 retval = find_available_queue_slot(pqm, qid); 229 if (retval != 0) 230 return retval; 231 232 if (list_empty(&pdd->qpd.queues_list) && 233 list_empty(&pdd->qpd.priv_queue_list)) 234 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd); 235 236 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL); 237 if (!pqn) { 238 retval = -ENOMEM; 239 goto err_allocate_pqn; 240 } 241 242 switch (type) { 243 case KFD_QUEUE_TYPE_SDMA: 244 case KFD_QUEUE_TYPE_SDMA_XGMI: 245 /* SDMA queues are always allocated statically no matter 246 * which scheduler mode is used. We also do not need to 247 * check whether a SDMA queue can be allocated here, because 248 * allocate_sdma_queue() in create_queue() has the 249 * corresponding check logic. 250 */ 251 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 252 if (retval != 0) 253 goto err_create_queue; 254 pqn->q = q; 255 pqn->kq = NULL; 256 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 257 print_queue(q); 258 break; 259 260 case KFD_QUEUE_TYPE_COMPUTE: 261 /* check if there is over subscription */ 262 if ((dev->dqm->sched_policy == 263 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) && 264 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) || 265 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) { 266 pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n"); 267 retval = -EPERM; 268 goto err_create_queue; 269 } 270 271 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 272 if (retval != 0) 273 goto err_create_queue; 274 pqn->q = q; 275 pqn->kq = NULL; 276 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 277 print_queue(q); 278 break; 279 case KFD_QUEUE_TYPE_DIQ: 280 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ); 281 if (!kq) { 282 retval = -ENOMEM; 283 goto err_create_queue; 284 } 285 kq->queue->properties.queue_id = *qid; 286 pqn->kq = kq; 287 pqn->q = NULL; 288 retval = dev->dqm->ops.create_kernel_queue(dev->dqm, 289 kq, &pdd->qpd); 290 break; 291 default: 292 WARN(1, "Invalid queue type %d", type); 293 retval = -EINVAL; 294 } 295 296 if (retval != 0) { 297 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n", 298 pqm->process->pasid, type, retval); 299 goto err_create_queue; 300 } 301 302 if (q && p_doorbell_offset_in_process) 303 /* Return the doorbell offset within the doorbell page 304 * to the caller so it can be passed up to user mode 305 * (in bytes). 306 * There are always 1024 doorbells per process, so in case 307 * of 8-byte doorbells, there are two doorbell pages per 308 * process. 309 */ 310 *p_doorbell_offset_in_process = 311 (q->properties.doorbell_off * sizeof(uint32_t)) & 312 (kfd_doorbell_process_slice(dev) - 1); 313 314 pr_debug("PQM After DQM create queue\n"); 315 316 list_add(&pqn->process_queue_list, &pqm->queues); 317 318 if (q) { 319 pr_debug("PQM done creating queue\n"); 320 kfd_procfs_add_queue(q); 321 print_queue_properties(&q->properties); 322 } 323 324 return retval; 325 326 err_create_queue: 327 uninit_queue(q); 328 if (kq) 329 kernel_queue_uninit(kq, false); 330 kfree(pqn); 331 err_allocate_pqn: 332 /* check if queues list is empty unregister process from device */ 333 clear_bit(*qid, pqm->queue_slot_bitmap); 334 if (list_empty(&pdd->qpd.queues_list) && 335 list_empty(&pdd->qpd.priv_queue_list)) 336 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd); 337 return retval; 338 } 339 340 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) 341 { 342 struct process_queue_node *pqn; 343 struct kfd_process_device *pdd; 344 struct device_queue_manager *dqm; 345 struct kfd_dev *dev; 346 int retval; 347 348 dqm = NULL; 349 350 retval = 0; 351 352 pqn = get_queue_by_qid(pqm, qid); 353 if (!pqn) { 354 pr_err("Queue id does not match any known queue\n"); 355 return -EINVAL; 356 } 357 358 dev = NULL; 359 if (pqn->kq) 360 dev = pqn->kq->dev; 361 if (pqn->q) 362 dev = pqn->q->device; 363 if (WARN_ON(!dev)) 364 return -ENODEV; 365 366 pdd = kfd_get_process_device_data(dev, pqm->process); 367 if (!pdd) { 368 pr_err("Process device data doesn't exist\n"); 369 return -1; 370 } 371 372 if (pqn->kq) { 373 /* destroy kernel queue (DIQ) */ 374 dqm = pqn->kq->dev->dqm; 375 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd); 376 kernel_queue_uninit(pqn->kq, false); 377 } 378 379 if (pqn->q) { 380 kfd_procfs_del_queue(pqn->q); 381 dqm = pqn->q->device->dqm; 382 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q); 383 if (retval) { 384 pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n", 385 pqm->process->pasid, 386 pqn->q->properties.queue_id, retval); 387 if (retval != -ETIME) 388 goto err_destroy_queue; 389 } 390 391 if (pqn->q->gws) { 392 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info, 393 pqn->q->gws); 394 pdd->qpd.num_gws = 0; 395 } 396 397 kfree(pqn->q->properties.cu_mask); 398 pqn->q->properties.cu_mask = NULL; 399 uninit_queue(pqn->q); 400 } 401 402 list_del(&pqn->process_queue_list); 403 kfree(pqn); 404 clear_bit(qid, pqm->queue_slot_bitmap); 405 406 if (list_empty(&pdd->qpd.queues_list) && 407 list_empty(&pdd->qpd.priv_queue_list)) 408 dqm->ops.unregister_process(dqm, &pdd->qpd); 409 410 err_destroy_queue: 411 return retval; 412 } 413 414 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 415 struct queue_properties *p) 416 { 417 int retval; 418 struct process_queue_node *pqn; 419 420 pqn = get_queue_by_qid(pqm, qid); 421 if (!pqn) { 422 pr_debug("No queue %d exists for update operation\n", qid); 423 return -EFAULT; 424 } 425 426 pqn->q->properties.queue_address = p->queue_address; 427 pqn->q->properties.queue_size = p->queue_size; 428 pqn->q->properties.queue_percent = p->queue_percent; 429 pqn->q->properties.priority = p->priority; 430 431 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 432 pqn->q); 433 if (retval != 0) 434 return retval; 435 436 return 0; 437 } 438 439 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 440 struct queue_properties *p) 441 { 442 int retval; 443 struct process_queue_node *pqn; 444 445 pqn = get_queue_by_qid(pqm, qid); 446 if (!pqn) { 447 pr_debug("No queue %d exists for update operation\n", qid); 448 return -EFAULT; 449 } 450 451 /* Free the old CU mask memory if it is already allocated, then 452 * allocate memory for the new CU mask. 453 */ 454 kfree(pqn->q->properties.cu_mask); 455 456 pqn->q->properties.cu_mask_count = p->cu_mask_count; 457 pqn->q->properties.cu_mask = p->cu_mask; 458 459 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 460 pqn->q); 461 if (retval != 0) 462 return retval; 463 464 return 0; 465 } 466 467 struct kernel_queue *pqm_get_kernel_queue( 468 struct process_queue_manager *pqm, 469 unsigned int qid) 470 { 471 struct process_queue_node *pqn; 472 473 pqn = get_queue_by_qid(pqm, qid); 474 if (pqn && pqn->kq) 475 return pqn->kq; 476 477 return NULL; 478 } 479 480 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 481 unsigned int qid) 482 { 483 struct process_queue_node *pqn; 484 485 pqn = get_queue_by_qid(pqm, qid); 486 return pqn ? pqn->q : NULL; 487 } 488 489 int pqm_get_wave_state(struct process_queue_manager *pqm, 490 unsigned int qid, 491 void __user *ctl_stack, 492 u32 *ctl_stack_used_size, 493 u32 *save_area_used_size) 494 { 495 struct process_queue_node *pqn; 496 497 pqn = get_queue_by_qid(pqm, qid); 498 if (!pqn) { 499 pr_debug("amdkfd: No queue %d exists for operation\n", 500 qid); 501 return -EFAULT; 502 } 503 504 return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm, 505 pqn->q, 506 ctl_stack, 507 ctl_stack_used_size, 508 save_area_used_size); 509 } 510 511 #if defined(CONFIG_DEBUG_FS) 512 513 int pqm_debugfs_mqds(struct seq_file *m, void *data) 514 { 515 struct process_queue_manager *pqm = data; 516 struct process_queue_node *pqn; 517 struct queue *q; 518 enum KFD_MQD_TYPE mqd_type; 519 struct mqd_manager *mqd_mgr; 520 int r = 0; 521 522 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 523 if (pqn->q) { 524 q = pqn->q; 525 switch (q->properties.type) { 526 case KFD_QUEUE_TYPE_SDMA: 527 case KFD_QUEUE_TYPE_SDMA_XGMI: 528 seq_printf(m, " SDMA queue on device %x\n", 529 q->device->id); 530 mqd_type = KFD_MQD_TYPE_SDMA; 531 break; 532 case KFD_QUEUE_TYPE_COMPUTE: 533 seq_printf(m, " Compute queue on device %x\n", 534 q->device->id); 535 mqd_type = KFD_MQD_TYPE_CP; 536 break; 537 default: 538 seq_printf(m, 539 " Bad user queue type %d on device %x\n", 540 q->properties.type, q->device->id); 541 continue; 542 } 543 mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type]; 544 } else if (pqn->kq) { 545 q = pqn->kq->queue; 546 mqd_mgr = pqn->kq->mqd_mgr; 547 switch (q->properties.type) { 548 case KFD_QUEUE_TYPE_DIQ: 549 seq_printf(m, " DIQ on device %x\n", 550 pqn->kq->dev->id); 551 break; 552 default: 553 seq_printf(m, 554 " Bad kernel queue type %d on device %x\n", 555 q->properties.type, 556 pqn->kq->dev->id); 557 continue; 558 } 559 } else { 560 seq_printf(m, 561 " Weird: Queue node with neither kernel nor user queue\n"); 562 continue; 563 } 564 565 r = mqd_mgr->debugfs_show_mqd(m, q->mqd); 566 if (r != 0) 567 break; 568 } 569 570 return r; 571 } 572 573 #endif 574