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 uninit_queue(pqn->q); 157 list_del(&pqn->process_queue_list); 158 kfree(pqn); 159 } 160 161 kfree(pqm->queue_slot_bitmap); 162 pqm->queue_slot_bitmap = NULL; 163 } 164 165 static int init_user_queue(struct process_queue_manager *pqm, 166 struct kfd_dev *dev, struct queue **q, 167 struct queue_properties *q_properties, 168 struct file *f, unsigned int qid) 169 { 170 int retval; 171 172 /* Doorbell initialized in user space*/ 173 q_properties->doorbell_ptr = NULL; 174 175 /* let DQM handle it*/ 176 q_properties->vmid = 0; 177 q_properties->queue_id = qid; 178 179 retval = init_queue(q, q_properties); 180 if (retval != 0) 181 return retval; 182 183 (*q)->device = dev; 184 (*q)->process = pqm->process; 185 186 pr_debug("PQM After init queue"); 187 188 return retval; 189 } 190 191 int pqm_create_queue(struct process_queue_manager *pqm, 192 struct kfd_dev *dev, 193 struct file *f, 194 struct queue_properties *properties, 195 unsigned int *qid, 196 uint32_t *p_doorbell_offset_in_process) 197 { 198 int retval; 199 struct kfd_process_device *pdd; 200 struct queue *q; 201 struct process_queue_node *pqn; 202 struct kernel_queue *kq; 203 enum kfd_queue_type type = properties->type; 204 unsigned int max_queues = 127; /* HWS limit */ 205 206 q = NULL; 207 kq = NULL; 208 209 pdd = kfd_get_process_device_data(dev, pqm->process); 210 if (!pdd) { 211 pr_err("Process device data doesn't exist\n"); 212 return -1; 213 } 214 215 /* 216 * for debug process, verify that it is within the static queues limit 217 * currently limit is set to half of the total avail HQD slots 218 * If we are just about to create DIQ, the is_debug flag is not set yet 219 * Hence we also check the type as well 220 */ 221 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ)) 222 max_queues = dev->device_info->max_no_of_hqd/2; 223 224 if (pdd->qpd.queue_count >= max_queues) 225 return -ENOSPC; 226 227 retval = find_available_queue_slot(pqm, qid); 228 if (retval != 0) 229 return retval; 230 231 if (list_empty(&pdd->qpd.queues_list) && 232 list_empty(&pdd->qpd.priv_queue_list)) 233 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd); 234 235 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL); 236 if (!pqn) { 237 retval = -ENOMEM; 238 goto err_allocate_pqn; 239 } 240 241 switch (type) { 242 case KFD_QUEUE_TYPE_SDMA: 243 case KFD_QUEUE_TYPE_SDMA_XGMI: 244 /* SDMA queues are always allocated statically no matter 245 * which scheduler mode is used. We also do not need to 246 * check whether a SDMA queue can be allocated here, because 247 * allocate_sdma_queue() in create_queue() has the 248 * corresponding check logic. 249 */ 250 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 251 if (retval != 0) 252 goto err_create_queue; 253 pqn->q = q; 254 pqn->kq = NULL; 255 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 256 print_queue(q); 257 break; 258 259 case KFD_QUEUE_TYPE_COMPUTE: 260 /* check if there is over subscription */ 261 if ((dev->dqm->sched_policy == 262 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) && 263 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) || 264 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) { 265 pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n"); 266 retval = -EPERM; 267 goto err_create_queue; 268 } 269 270 retval = init_user_queue(pqm, dev, &q, properties, f, *qid); 271 if (retval != 0) 272 goto err_create_queue; 273 pqn->q = q; 274 pqn->kq = NULL; 275 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 276 print_queue(q); 277 break; 278 case KFD_QUEUE_TYPE_DIQ: 279 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ); 280 if (!kq) { 281 retval = -ENOMEM; 282 goto err_create_queue; 283 } 284 kq->queue->properties.queue_id = *qid; 285 pqn->kq = kq; 286 pqn->q = NULL; 287 retval = dev->dqm->ops.create_kernel_queue(dev->dqm, 288 kq, &pdd->qpd); 289 break; 290 default: 291 WARN(1, "Invalid queue type %d", type); 292 retval = -EINVAL; 293 } 294 295 if (retval != 0) { 296 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n", 297 pqm->process->pasid, type, retval); 298 goto err_create_queue; 299 } 300 301 if (q && p_doorbell_offset_in_process) 302 /* Return the doorbell offset within the doorbell page 303 * to the caller so it can be passed up to user mode 304 * (in bytes). 305 * There are always 1024 doorbells per process, so in case 306 * of 8-byte doorbells, there are two doorbell pages per 307 * process. 308 */ 309 *p_doorbell_offset_in_process = 310 (q->properties.doorbell_off * sizeof(uint32_t)) & 311 (kfd_doorbell_process_slice(dev) - 1); 312 313 pr_debug("PQM After DQM create queue\n"); 314 315 list_add(&pqn->process_queue_list, &pqm->queues); 316 317 if (q) { 318 pr_debug("PQM done creating queue\n"); 319 kfd_procfs_add_queue(q); 320 print_queue_properties(&q->properties); 321 } 322 323 return retval; 324 325 err_create_queue: 326 uninit_queue(q); 327 if (kq) 328 kernel_queue_uninit(kq, false); 329 kfree(pqn); 330 err_allocate_pqn: 331 /* check if queues list is empty unregister process from device */ 332 clear_bit(*qid, pqm->queue_slot_bitmap); 333 if (list_empty(&pdd->qpd.queues_list) && 334 list_empty(&pdd->qpd.priv_queue_list)) 335 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd); 336 return retval; 337 } 338 339 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) 340 { 341 struct process_queue_node *pqn; 342 struct kfd_process_device *pdd; 343 struct device_queue_manager *dqm; 344 struct kfd_dev *dev; 345 int retval; 346 347 dqm = NULL; 348 349 retval = 0; 350 351 pqn = get_queue_by_qid(pqm, qid); 352 if (!pqn) { 353 pr_err("Queue id does not match any known queue\n"); 354 return -EINVAL; 355 } 356 357 dev = NULL; 358 if (pqn->kq) 359 dev = pqn->kq->dev; 360 if (pqn->q) 361 dev = pqn->q->device; 362 if (WARN_ON(!dev)) 363 return -ENODEV; 364 365 pdd = kfd_get_process_device_data(dev, pqm->process); 366 if (!pdd) { 367 pr_err("Process device data doesn't exist\n"); 368 return -1; 369 } 370 371 if (pqn->kq) { 372 /* destroy kernel queue (DIQ) */ 373 dqm = pqn->kq->dev->dqm; 374 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd); 375 kernel_queue_uninit(pqn->kq, false); 376 } 377 378 if (pqn->q) { 379 kfd_procfs_del_queue(pqn->q); 380 dqm = pqn->q->device->dqm; 381 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q); 382 if (retval) { 383 pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n", 384 pqm->process->pasid, 385 pqn->q->properties.queue_id, retval); 386 if (retval != -ETIME) 387 goto err_destroy_queue; 388 } 389 390 if (pqn->q->gws) { 391 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info, 392 pqn->q->gws); 393 pdd->qpd.num_gws = 0; 394 } 395 396 kfree(pqn->q->properties.cu_mask); 397 pqn->q->properties.cu_mask = NULL; 398 uninit_queue(pqn->q); 399 } 400 401 list_del(&pqn->process_queue_list); 402 kfree(pqn); 403 clear_bit(qid, pqm->queue_slot_bitmap); 404 405 if (list_empty(&pdd->qpd.queues_list) && 406 list_empty(&pdd->qpd.priv_queue_list)) 407 dqm->ops.unregister_process(dqm, &pdd->qpd); 408 409 err_destroy_queue: 410 return retval; 411 } 412 413 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 414 struct queue_properties *p) 415 { 416 int retval; 417 struct process_queue_node *pqn; 418 419 pqn = get_queue_by_qid(pqm, qid); 420 if (!pqn) { 421 pr_debug("No queue %d exists for update operation\n", qid); 422 return -EFAULT; 423 } 424 425 pqn->q->properties.queue_address = p->queue_address; 426 pqn->q->properties.queue_size = p->queue_size; 427 pqn->q->properties.queue_percent = p->queue_percent; 428 pqn->q->properties.priority = p->priority; 429 430 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 431 pqn->q); 432 if (retval != 0) 433 return retval; 434 435 return 0; 436 } 437 438 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 439 struct queue_properties *p) 440 { 441 int retval; 442 struct process_queue_node *pqn; 443 444 pqn = get_queue_by_qid(pqm, qid); 445 if (!pqn) { 446 pr_debug("No queue %d exists for update operation\n", qid); 447 return -EFAULT; 448 } 449 450 /* Free the old CU mask memory if it is already allocated, then 451 * allocate memory for the new CU mask. 452 */ 453 kfree(pqn->q->properties.cu_mask); 454 455 pqn->q->properties.cu_mask_count = p->cu_mask_count; 456 pqn->q->properties.cu_mask = p->cu_mask; 457 458 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 459 pqn->q); 460 if (retval != 0) 461 return retval; 462 463 return 0; 464 } 465 466 struct kernel_queue *pqm_get_kernel_queue( 467 struct process_queue_manager *pqm, 468 unsigned int qid) 469 { 470 struct process_queue_node *pqn; 471 472 pqn = get_queue_by_qid(pqm, qid); 473 if (pqn && pqn->kq) 474 return pqn->kq; 475 476 return NULL; 477 } 478 479 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 480 unsigned int qid) 481 { 482 struct process_queue_node *pqn; 483 484 pqn = get_queue_by_qid(pqm, qid); 485 return pqn ? pqn->q : NULL; 486 } 487 488 int pqm_get_wave_state(struct process_queue_manager *pqm, 489 unsigned int qid, 490 void __user *ctl_stack, 491 u32 *ctl_stack_used_size, 492 u32 *save_area_used_size) 493 { 494 struct process_queue_node *pqn; 495 496 pqn = get_queue_by_qid(pqm, qid); 497 if (!pqn) { 498 pr_debug("amdkfd: No queue %d exists for operation\n", 499 qid); 500 return -EFAULT; 501 } 502 503 return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm, 504 pqn->q, 505 ctl_stack, 506 ctl_stack_used_size, 507 save_area_used_size); 508 } 509 510 #if defined(CONFIG_DEBUG_FS) 511 512 int pqm_debugfs_mqds(struct seq_file *m, void *data) 513 { 514 struct process_queue_manager *pqm = data; 515 struct process_queue_node *pqn; 516 struct queue *q; 517 enum KFD_MQD_TYPE mqd_type; 518 struct mqd_manager *mqd_mgr; 519 int r = 0; 520 521 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 522 if (pqn->q) { 523 q = pqn->q; 524 switch (q->properties.type) { 525 case KFD_QUEUE_TYPE_SDMA: 526 case KFD_QUEUE_TYPE_SDMA_XGMI: 527 seq_printf(m, " SDMA queue on device %x\n", 528 q->device->id); 529 mqd_type = KFD_MQD_TYPE_SDMA; 530 break; 531 case KFD_QUEUE_TYPE_COMPUTE: 532 seq_printf(m, " Compute queue on device %x\n", 533 q->device->id); 534 mqd_type = KFD_MQD_TYPE_CP; 535 break; 536 default: 537 seq_printf(m, 538 " Bad user queue type %d on device %x\n", 539 q->properties.type, q->device->id); 540 continue; 541 } 542 mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type]; 543 } else if (pqn->kq) { 544 q = pqn->kq->queue; 545 mqd_mgr = pqn->kq->mqd_mgr; 546 switch (q->properties.type) { 547 case KFD_QUEUE_TYPE_DIQ: 548 seq_printf(m, " DIQ on device %x\n", 549 pqn->kq->dev->id); 550 break; 551 default: 552 seq_printf(m, 553 " Bad kernel queue type %d on device %x\n", 554 q->properties.type, 555 pqn->kq->dev->id); 556 continue; 557 } 558 } else { 559 seq_printf(m, 560 " Weird: Queue node with neither kernel nor user queue\n"); 561 continue; 562 } 563 564 r = mqd_mgr->debugfs_show_mqd(m, q->mqd); 565 if (r != 0) 566 break; 567 } 568 569 return r; 570 } 571 572 #endif 573