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 30 static inline struct process_queue_node *get_queue_by_qid( 31 struct process_queue_manager *pqm, unsigned int qid) 32 { 33 struct process_queue_node *pqn; 34 35 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 36 if ((pqn->q && pqn->q->properties.queue_id == qid) || 37 (pqn->kq && pqn->kq->queue->properties.queue_id == qid)) 38 return pqn; 39 } 40 41 return NULL; 42 } 43 44 static int find_available_queue_slot(struct process_queue_manager *pqm, 45 unsigned int *qid) 46 { 47 unsigned long found; 48 49 found = find_first_zero_bit(pqm->queue_slot_bitmap, 50 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 51 52 pr_debug("The new slot id %lu\n", found); 53 54 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { 55 pr_info("Cannot open more queues for process with pasid %d\n", 56 pqm->process->pasid); 57 return -ENOMEM; 58 } 59 60 set_bit(found, pqm->queue_slot_bitmap); 61 *qid = found; 62 63 return 0; 64 } 65 66 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd) 67 { 68 struct kfd_dev *dev = pdd->dev; 69 70 if (pdd->already_dequeued) 71 return; 72 73 dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd); 74 pdd->already_dequeued = true; 75 } 76 77 void kfd_process_dequeue_from_all_devices(struct kfd_process *p) 78 { 79 struct kfd_process_device *pdd; 80 81 list_for_each_entry(pdd, &p->per_device_data, per_device_list) 82 kfd_process_dequeue_from_device(pdd); 83 } 84 85 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p) 86 { 87 INIT_LIST_HEAD(&pqm->queues); 88 pqm->queue_slot_bitmap = 89 kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 90 BITS_PER_BYTE), GFP_KERNEL); 91 if (!pqm->queue_slot_bitmap) 92 return -ENOMEM; 93 pqm->process = p; 94 95 return 0; 96 } 97 98 void pqm_uninit(struct process_queue_manager *pqm) 99 { 100 struct process_queue_node *pqn, *next; 101 102 list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) { 103 uninit_queue(pqn->q); 104 list_del(&pqn->process_queue_list); 105 kfree(pqn); 106 } 107 108 kfree(pqm->queue_slot_bitmap); 109 pqm->queue_slot_bitmap = NULL; 110 } 111 112 static int create_cp_queue(struct process_queue_manager *pqm, 113 struct kfd_dev *dev, struct queue **q, 114 struct queue_properties *q_properties, 115 struct file *f, unsigned int qid) 116 { 117 int retval; 118 119 /* Doorbell initialized in user space*/ 120 q_properties->doorbell_ptr = NULL; 121 122 /* let DQM handle it*/ 123 q_properties->vmid = 0; 124 q_properties->queue_id = qid; 125 126 retval = init_queue(q, q_properties); 127 if (retval != 0) 128 return retval; 129 130 (*q)->device = dev; 131 (*q)->process = pqm->process; 132 133 pr_debug("PQM After init queue"); 134 135 return retval; 136 } 137 138 int pqm_create_queue(struct process_queue_manager *pqm, 139 struct kfd_dev *dev, 140 struct file *f, 141 struct queue_properties *properties, 142 unsigned int *qid) 143 { 144 int retval; 145 struct kfd_process_device *pdd; 146 struct queue *q; 147 struct process_queue_node *pqn; 148 struct kernel_queue *kq; 149 enum kfd_queue_type type = properties->type; 150 unsigned int max_queues = 127; /* HWS limit */ 151 152 q = NULL; 153 kq = NULL; 154 155 pdd = kfd_get_process_device_data(dev, pqm->process); 156 if (!pdd) { 157 pr_err("Process device data doesn't exist\n"); 158 return -1; 159 } 160 161 /* 162 * for debug process, verify that it is within the static queues limit 163 * currently limit is set to half of the total avail HQD slots 164 * If we are just about to create DIQ, the is_debug flag is not set yet 165 * Hence we also check the type as well 166 */ 167 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ)) 168 max_queues = dev->device_info->max_no_of_hqd/2; 169 170 if (pdd->qpd.queue_count >= max_queues) 171 return -ENOSPC; 172 173 retval = find_available_queue_slot(pqm, qid); 174 if (retval != 0) 175 return retval; 176 177 if (list_empty(&pdd->qpd.queues_list) && 178 list_empty(&pdd->qpd.priv_queue_list)) 179 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd); 180 181 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL); 182 if (!pqn) { 183 retval = -ENOMEM; 184 goto err_allocate_pqn; 185 } 186 187 switch (type) { 188 case KFD_QUEUE_TYPE_SDMA: 189 if (dev->dqm->queue_count >= 190 CIK_SDMA_QUEUES_PER_ENGINE * CIK_SDMA_ENGINE_NUM) { 191 pr_err("Over-subscription is not allowed for SDMA.\n"); 192 retval = -EPERM; 193 goto err_create_queue; 194 } 195 196 retval = create_cp_queue(pqm, dev, &q, properties, f, *qid); 197 if (retval != 0) 198 goto err_create_queue; 199 pqn->q = q; 200 pqn->kq = NULL; 201 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 202 pr_debug("DQM returned %d for create_queue\n", retval); 203 print_queue(q); 204 break; 205 206 case KFD_QUEUE_TYPE_COMPUTE: 207 /* check if there is over subscription */ 208 if ((dev->dqm->sched_policy == 209 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) && 210 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) || 211 (dev->dqm->queue_count >= get_queues_num(dev->dqm)))) { 212 pr_err("Over-subscription is not allowed in radeon_kfd.sched_policy == 1\n"); 213 retval = -EPERM; 214 goto err_create_queue; 215 } 216 217 retval = create_cp_queue(pqm, dev, &q, properties, f, *qid); 218 if (retval != 0) 219 goto err_create_queue; 220 pqn->q = q; 221 pqn->kq = NULL; 222 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd); 223 pr_debug("DQM returned %d for create_queue\n", retval); 224 print_queue(q); 225 break; 226 case KFD_QUEUE_TYPE_DIQ: 227 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ); 228 if (!kq) { 229 retval = -ENOMEM; 230 goto err_create_queue; 231 } 232 kq->queue->properties.queue_id = *qid; 233 pqn->kq = kq; 234 pqn->q = NULL; 235 retval = dev->dqm->ops.create_kernel_queue(dev->dqm, 236 kq, &pdd->qpd); 237 break; 238 default: 239 WARN(1, "Invalid queue type %d", type); 240 retval = -EINVAL; 241 } 242 243 if (retval != 0) { 244 pr_err("Pasid %d DQM create queue %d failed. ret %d\n", 245 pqm->process->pasid, type, retval); 246 goto err_create_queue; 247 } 248 249 if (q) 250 /* Return the doorbell offset within the doorbell page 251 * to the caller so it can be passed up to user mode 252 * (in bytes). 253 */ 254 properties->doorbell_off = 255 (q->properties.doorbell_off * sizeof(uint32_t)) & 256 (kfd_doorbell_process_slice(dev) - 1); 257 258 pr_debug("PQM After DQM create queue\n"); 259 260 list_add(&pqn->process_queue_list, &pqm->queues); 261 262 if (q) { 263 pr_debug("PQM done creating queue\n"); 264 print_queue_properties(&q->properties); 265 } 266 267 return retval; 268 269 err_create_queue: 270 kfree(pqn); 271 err_allocate_pqn: 272 /* check if queues list is empty unregister process from device */ 273 clear_bit(*qid, pqm->queue_slot_bitmap); 274 if (list_empty(&pdd->qpd.queues_list) && 275 list_empty(&pdd->qpd.priv_queue_list)) 276 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd); 277 return retval; 278 } 279 280 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) 281 { 282 struct process_queue_node *pqn; 283 struct kfd_process_device *pdd; 284 struct device_queue_manager *dqm; 285 struct kfd_dev *dev; 286 int retval; 287 288 dqm = NULL; 289 290 retval = 0; 291 292 pqn = get_queue_by_qid(pqm, qid); 293 if (!pqn) { 294 pr_err("Queue id does not match any known queue\n"); 295 return -EINVAL; 296 } 297 298 dev = NULL; 299 if (pqn->kq) 300 dev = pqn->kq->dev; 301 if (pqn->q) 302 dev = pqn->q->device; 303 if (WARN_ON(!dev)) 304 return -ENODEV; 305 306 pdd = kfd_get_process_device_data(dev, pqm->process); 307 if (!pdd) { 308 pr_err("Process device data doesn't exist\n"); 309 return -1; 310 } 311 312 if (pqn->kq) { 313 /* destroy kernel queue (DIQ) */ 314 dqm = pqn->kq->dev->dqm; 315 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd); 316 kernel_queue_uninit(pqn->kq); 317 } 318 319 if (pqn->q) { 320 dqm = pqn->q->device->dqm; 321 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q); 322 if (retval) { 323 pr_err("Pasid %d destroy queue %d failed, ret %d\n", 324 pqm->process->pasid, 325 pqn->q->properties.queue_id, retval); 326 if (retval != -ETIME) 327 goto err_destroy_queue; 328 } 329 uninit_queue(pqn->q); 330 } 331 332 list_del(&pqn->process_queue_list); 333 kfree(pqn); 334 clear_bit(qid, pqm->queue_slot_bitmap); 335 336 if (list_empty(&pdd->qpd.queues_list) && 337 list_empty(&pdd->qpd.priv_queue_list)) 338 dqm->ops.unregister_process(dqm, &pdd->qpd); 339 340 err_destroy_queue: 341 return retval; 342 } 343 344 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 345 struct queue_properties *p) 346 { 347 int retval; 348 struct process_queue_node *pqn; 349 350 pqn = get_queue_by_qid(pqm, qid); 351 if (!pqn) { 352 pr_debug("No queue %d exists for update operation\n", qid); 353 return -EFAULT; 354 } 355 356 pqn->q->properties.queue_address = p->queue_address; 357 pqn->q->properties.queue_size = p->queue_size; 358 pqn->q->properties.queue_percent = p->queue_percent; 359 pqn->q->properties.priority = p->priority; 360 361 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, 362 pqn->q); 363 if (retval != 0) 364 return retval; 365 366 return 0; 367 } 368 369 struct kernel_queue *pqm_get_kernel_queue( 370 struct process_queue_manager *pqm, 371 unsigned int qid) 372 { 373 struct process_queue_node *pqn; 374 375 pqn = get_queue_by_qid(pqm, qid); 376 if (pqn && pqn->kq) 377 return pqn->kq; 378 379 return NULL; 380 } 381 382 #if defined(CONFIG_DEBUG_FS) 383 384 int pqm_debugfs_mqds(struct seq_file *m, void *data) 385 { 386 struct process_queue_manager *pqm = data; 387 struct process_queue_node *pqn; 388 struct queue *q; 389 enum KFD_MQD_TYPE mqd_type; 390 struct mqd_manager *mqd_manager; 391 int r = 0; 392 393 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 394 if (pqn->q) { 395 q = pqn->q; 396 switch (q->properties.type) { 397 case KFD_QUEUE_TYPE_SDMA: 398 seq_printf(m, " SDMA queue on device %x\n", 399 q->device->id); 400 mqd_type = KFD_MQD_TYPE_SDMA; 401 break; 402 case KFD_QUEUE_TYPE_COMPUTE: 403 seq_printf(m, " Compute queue on device %x\n", 404 q->device->id); 405 mqd_type = KFD_MQD_TYPE_CP; 406 break; 407 default: 408 seq_printf(m, 409 " Bad user queue type %d on device %x\n", 410 q->properties.type, q->device->id); 411 continue; 412 } 413 mqd_manager = q->device->dqm->ops.get_mqd_manager( 414 q->device->dqm, mqd_type); 415 } else if (pqn->kq) { 416 q = pqn->kq->queue; 417 mqd_manager = pqn->kq->mqd; 418 switch (q->properties.type) { 419 case KFD_QUEUE_TYPE_DIQ: 420 seq_printf(m, " DIQ on device %x\n", 421 pqn->kq->dev->id); 422 mqd_type = KFD_MQD_TYPE_HIQ; 423 break; 424 default: 425 seq_printf(m, 426 " Bad kernel queue type %d on device %x\n", 427 q->properties.type, 428 pqn->kq->dev->id); 429 continue; 430 } 431 } else { 432 seq_printf(m, 433 " Weird: Queue node with neither kernel nor user queue\n"); 434 continue; 435 } 436 437 r = mqd_manager->debugfs_show_mqd(m, q->mqd); 438 if (r != 0) 439 break; 440 } 441 442 return r; 443 } 444 445 #endif 446