drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

   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 <linux/types.h>
  27 #include <linux/printk.h>
  28 #include <linux/bitops.h>
  29 #include <linux/sched.h>
  30 #include "kfd_priv.h"
  31 #include "kfd_device_queue_manager.h"
  32 #include "kfd_mqd_manager.h"
  33 #include "cik_regs.h"
  34 #include "kfd_kernel_queue.h"
  35
  36 /* Size of the per-pipe EOP queue */
  37 #define CIK_HPD_EOP_BYTES_LOG2 11
  38 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
  39
  40 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
  41                                         unsigned int pasid, unsigned int vmid);
  42
  43 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
  44                                         struct queue *q,
  45                                         struct qcm_process_device *qpd);
  46
  47 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock);
  48 static int destroy_queues_cpsch(struct device_queue_manager *dqm, bool lock);
  49
  50 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
  51                                         struct queue *q,
  52                                         struct qcm_process_device *qpd);
  53
  54 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
  55                                 unsigned int sdma_queue_id);
  56
  57 static inline
  58 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
  59 {
  60         if (type == KFD_QUEUE_TYPE_SDMA)
  61                 return KFD_MQD_TYPE_SDMA;
  62         return KFD_MQD_TYPE_CP;
  63 }
  64
  65 unsigned int get_first_pipe(struct device_queue_manager *dqm)
  66 {
  67         BUG_ON(!dqm || !dqm->dev);
  68         return dqm->dev->shared_resources.first_compute_pipe;
  69 }
  70
  71 unsigned int get_pipes_num(struct device_queue_manager *dqm)
  72 {
  73         BUG_ON(!dqm || !dqm->dev);
  74         return dqm->dev->shared_resources.compute_pipe_count;
  75 }
  76
  77 static inline unsigned int get_pipes_num_cpsch(void)
  78 {
  79         return PIPE_PER_ME_CP_SCHEDULING;
  80 }
  81
  82 void program_sh_mem_settings(struct device_queue_manager *dqm,
  83                                         struct qcm_process_device *qpd)
  84 {
  85         return kfd2kgd->program_sh_mem_settings(dqm->dev->kgd, qpd->vmid,
  86                                                 qpd->sh_mem_config,
  87                                                 qpd->sh_mem_ape1_base,
  88                                                 qpd->sh_mem_ape1_limit,
  89                                                 qpd->sh_mem_bases);
  90 }
  91
  92 static int allocate_vmid(struct device_queue_manager *dqm,
  93                         struct qcm_process_device *qpd,
  94                         struct queue *q)
  95 {
  96         int bit, allocated_vmid;
  97
  98         if (dqm->vmid_bitmap == 0)
  99                 return -ENOMEM;
 100
 101         bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap, CIK_VMID_NUM);
 102         clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
 103
 104         /* Kaveri kfd vmid's starts from vmid 8 */
 105         allocated_vmid = bit + KFD_VMID_START_OFFSET;
 106         pr_debug("kfd: vmid allocation %d\n", allocated_vmid);
 107         qpd->vmid = allocated_vmid;
 108         q->properties.vmid = allocated_vmid;
 109
 110         set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
 111         program_sh_mem_settings(dqm, qpd);
 112
 113         return 0;
 114 }
 115
 116 static void deallocate_vmid(struct device_queue_manager *dqm,
 117                                 struct qcm_process_device *qpd,
 118                                 struct queue *q)
 119 {
 120         int bit = qpd->vmid - KFD_VMID_START_OFFSET;
 121
 122         /* Release the vmid mapping */
 123         set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
 124
 125         set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
 126         qpd->vmid = 0;
 127         q->properties.vmid = 0;
 128 }
 129
 130 static int create_queue_nocpsch(struct device_queue_manager *dqm,
 131                                 struct queue *q,
 132                                 struct qcm_process_device *qpd,
 133                                 int *allocated_vmid)
 134 {
 135         int retval;
 136
 137         BUG_ON(!dqm || !q || !qpd || !allocated_vmid);
 138
 139         pr_debug("kfd: In func %s\n", __func__);
 140         print_queue(q);
 141
 142         mutex_lock(&dqm->lock);
 143
 144         if (dqm->total_queue_count >= max_num_of_queues_per_device) {
 145                 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
 146                                 dqm->total_queue_count);
 147                 mutex_unlock(&dqm->lock);
 148                 return -EPERM;
 149         }
 150
 151         if (list_empty(&qpd->queues_list)) {
 152                 retval = allocate_vmid(dqm, qpd, q);
 153                 if (retval != 0) {
 154                         mutex_unlock(&dqm->lock);
 155                         return retval;
 156                 }
 157         }
 158         *allocated_vmid = qpd->vmid;
 159         q->properties.vmid = qpd->vmid;
 160
 161         if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
 162                 retval = create_compute_queue_nocpsch(dqm, q, qpd);
 163         if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 164                 retval = create_sdma_queue_nocpsch(dqm, q, qpd);
 165
 166         if (retval != 0) {
 167                 if (list_empty(&qpd->queues_list)) {
 168                         deallocate_vmid(dqm, qpd, q);
 169                         *allocated_vmid = 0;
 170                 }
 171                 mutex_unlock(&dqm->lock);
 172                 return retval;
 173         }
 174
 175         list_add(&q->list, &qpd->queues_list);
 176         if (q->properties.is_active)
 177                 dqm->queue_count++;
 178
 179         if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 180                 dqm->sdma_queue_count++;
 181
 182         /*
 183          * Unconditionally increment this counter, regardless of the queue's
 184          * type or whether the queue is active.
 185          */
 186         dqm->total_queue_count++;
 187         pr_debug("Total of %d queues are accountable so far\n",
 188                         dqm->total_queue_count);
 189
 190         mutex_unlock(&dqm->lock);
 191         return 0;
 192 }
 193
 194 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 195 {
 196         bool set;
 197         int pipe, bit, i;
 198
 199         set = false;
 200
 201         for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_num(dqm);
 202                         pipe = ((pipe + 1) % get_pipes_num(dqm)), ++i) {
 203                 if (dqm->allocated_queues[pipe] != 0) {
 204                         bit = find_first_bit(
 205                                 (unsigned long *)&dqm->allocated_queues[pipe],
 206                                 QUEUES_PER_PIPE);
 207
 208                         clear_bit(bit,
 209                                 (unsigned long *)&dqm->allocated_queues[pipe]);
 210                         q->pipe = pipe;
 211                         q->queue = bit;
 212                         set = true;
 213                         break;
 214                 }
 215         }
 216
 217         if (set == false)
 218                 return -EBUSY;
 219
 220         pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
 221                                 __func__, q->pipe, q->queue);
 222         /* horizontal hqd allocation */
 223         dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_num(dqm);
 224
 225         return 0;
 226 }
 227
 228 static inline void deallocate_hqd(struct device_queue_manager *dqm,
 229                                 struct queue *q)
 230 {
 231         set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
 232 }
 233
 234 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
 235                                         struct queue *q,
 236                                         struct qcm_process_device *qpd)
 237 {
 238         int retval;
 239         struct mqd_manager *mqd;
 240
 241         BUG_ON(!dqm || !q || !qpd);
 242
 243         mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
 244         if (mqd == NULL)
 245                 return -ENOMEM;
 246
 247         retval = allocate_hqd(dqm, q);
 248         if (retval != 0)
 249                 return retval;
 250
 251         retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 252                                 &q->gart_mqd_addr, &q->properties);
 253         if (retval != 0) {
 254                 deallocate_hqd(dqm, q);
 255                 return retval;
 256         }
 257
 258         pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
 259                         q->pipe,
 260                         q->queue);
 261
 262         retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
 263                         q->queue, (uint32_t __user *) q->properties.write_ptr);
 264         if (retval != 0) {
 265                 deallocate_hqd(dqm, q);
 266                 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
 267                 return retval;
 268         }
 269
 270         return 0;
 271 }
 272
 273 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
 274                                 struct qcm_process_device *qpd,
 275                                 struct queue *q)
 276 {
 277         int retval;
 278         struct mqd_manager *mqd;
 279
 280         BUG_ON(!dqm || !q || !q->mqd || !qpd);
 281
 282         retval = 0;
 283
 284         pr_debug("kfd: In Func %s\n", __func__);
 285
 286         mutex_lock(&dqm->lock);
 287
 288         if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
 289                 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
 290                 if (mqd == NULL) {
 291                         retval = -ENOMEM;
 292                         goto out;
 293                 }
 294                 deallocate_hqd(dqm, q);
 295         } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 296                 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
 297                 if (mqd == NULL) {
 298                         retval = -ENOMEM;
 299                         goto out;
 300                 }
 301                 dqm->sdma_queue_count--;
 302                 deallocate_sdma_queue(dqm, q->sdma_id);
 303         } else {
 304                 pr_debug("q->properties.type is invalid (%d)\n",
 305                                 q->properties.type);
 306                 retval = -EINVAL;
 307                 goto out;
 308         }
 309
 310         retval = mqd->destroy_mqd(mqd, q->mqd,
 311                                 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 312                                 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
 313                                 q->pipe, q->queue);
 314
 315         if (retval != 0)
 316                 goto out;
 317
 318         mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
 319
 320         list_del(&q->list);
 321         if (list_empty(&qpd->queues_list))
 322                 deallocate_vmid(dqm, qpd, q);
 323         if (q->properties.is_active)
 324                 dqm->queue_count--;
 325
 326         /*
 327          * Unconditionally decrement this counter, regardless of the queue's
 328          * type
 329          */
 330         dqm->total_queue_count--;
 331         pr_debug("Total of %d queues are accountable so far\n",
 332                         dqm->total_queue_count);
 333
 334 out:
 335         mutex_unlock(&dqm->lock);
 336         return retval;
 337 }
 338
 339 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
 340 {
 341         int retval;
 342         struct mqd_manager *mqd;
 343         bool prev_active = false;
 344
 345         BUG_ON(!dqm || !q || !q->mqd);
 346
 347         mutex_lock(&dqm->lock);
 348         mqd = dqm->ops.get_mqd_manager(dqm,
 349                         get_mqd_type_from_queue_type(q->properties.type));
 350         if (mqd == NULL) {
 351                 mutex_unlock(&dqm->lock);
 352                 return -ENOMEM;
 353         }
 354
 355         if (q->properties.is_active == true)
 356                 prev_active = true;
 357
 358         /*
 359          *
 360          * check active state vs. the previous state
 361          * and modify counter accordingly
 362          */
 363         retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
 364         if ((q->properties.is_active == true) && (prev_active == false))
 365                 dqm->queue_count++;
 366         else if ((q->properties.is_active == false) && (prev_active == true))
 367                 dqm->queue_count--;
 368
 369         if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
 370                 retval = execute_queues_cpsch(dqm, false);
 371
 372         mutex_unlock(&dqm->lock);
 373         return retval;
 374 }
 375
 376 static struct mqd_manager *get_mqd_manager_nocpsch(
 377                 struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
 378 {
 379         struct mqd_manager *mqd;
 380
 381         BUG_ON(!dqm || type >= KFD_MQD_TYPE_MAX);
 382
 383         pr_debug("kfd: In func %s mqd type %d\n", __func__, type);
 384
 385         mqd = dqm->mqds[type];
 386         if (!mqd) {
 387                 mqd = mqd_manager_init(type, dqm->dev);
 388                 if (mqd == NULL)
 389                         pr_err("kfd: mqd manager is NULL");
 390                 dqm->mqds[type] = mqd;
 391         }
 392
 393         return mqd;
 394 }
 395
 396 static int register_process_nocpsch(struct device_queue_manager *dqm,
 397                                         struct qcm_process_device *qpd)
 398 {
 399         struct device_process_node *n;
 400         int retval;
 401
 402         BUG_ON(!dqm || !qpd);
 403
 404         pr_debug("kfd: In func %s\n", __func__);
 405
 406         n = kzalloc(sizeof(struct device_process_node), GFP_KERNEL);
 407         if (!n)
 408                 return -ENOMEM;
 409
 410         n->qpd = qpd;
 411
 412         mutex_lock(&dqm->lock);
 413         list_add(&n->list, &dqm->queues);
 414
 415         retval = dqm->ops_asic_specific.register_process(dqm, qpd);
 416
 417         dqm->processes_count++;
 418
 419         mutex_unlock(&dqm->lock);
 420
 421         return retval;
 422 }
 423
 424 static int unregister_process_nocpsch(struct device_queue_manager *dqm,
 425                                         struct qcm_process_device *qpd)
 426 {
 427         int retval;
 428         struct device_process_node *cur, *next;
 429
 430         BUG_ON(!dqm || !qpd);
 431
 432         BUG_ON(!list_empty(&qpd->queues_list));
 433
 434         pr_debug("kfd: In func %s\n", __func__);
 435
 436         retval = 0;
 437         mutex_lock(&dqm->lock);
 438
 439         list_for_each_entry_safe(cur, next, &dqm->queues, list) {
 440                 if (qpd == cur->qpd) {
 441                         list_del(&cur->list);
 442                         kfree(cur);
 443                         dqm->processes_count--;
 444                         goto out;
 445                 }
 446         }
 447         /* qpd not found in dqm list */
 448         retval = 1;
 449 out:
 450         mutex_unlock(&dqm->lock);
 451         return retval;
 452 }
 453
 454 static int
 455 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
 456                         unsigned int vmid)
 457 {
 458         uint32_t pasid_mapping;
 459
 460         pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
 461                                                 ATC_VMID_PASID_MAPPING_VALID;
 462         return kfd2kgd->set_pasid_vmid_mapping(dqm->dev->kgd, pasid_mapping,
 463                                                 vmid);
 464 }
 465
 466 int init_pipelines(struct device_queue_manager *dqm,
 467                         unsigned int pipes_num, unsigned int first_pipe)
 468 {
 469         void *hpdptr;
 470         struct mqd_manager *mqd;
 471         unsigned int i, err, inx;
 472         uint64_t pipe_hpd_addr;
 473
 474         BUG_ON(!dqm || !dqm->dev);
 475
 476         pr_debug("kfd: In func %s\n", __func__);
 477
 478         /*
 479          * Allocate memory for the HPDs. This is hardware-owned per-pipe data.
 480          * The driver never accesses this memory after zeroing it.
 481          * It doesn't even have to be saved/restored on suspend/resume
 482          * because it contains no data when there are no active queues.
 483          */
 484
 485         err = kfd_gtt_sa_allocate(dqm->dev, CIK_HPD_EOP_BYTES * pipes_num,
 486                                         &dqm->pipeline_mem);
 487
 488         if (err) {
 489                 pr_err("kfd: error allocate vidmem num pipes: %d\n",
 490                         pipes_num);
 491                 return -ENOMEM;
 492         }
 493
 494         hpdptr = dqm->pipeline_mem->cpu_ptr;
 495         dqm->pipelines_addr = dqm->pipeline_mem->gpu_addr;
 496
 497         memset(hpdptr, 0, CIK_HPD_EOP_BYTES * pipes_num);
 498
 499         mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
 500         if (mqd == NULL) {
 501                 kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
 502                 return -ENOMEM;
 503         }
 504
 505         for (i = 0; i < pipes_num; i++) {
 506                 inx = i + first_pipe;
 507                 /*
 508                  * HPD buffer on GTT is allocated by amdkfd, no need to waste
 509                  * space in GTT for pipelines we don't initialize
 510                  */
 511                 pipe_hpd_addr = dqm->pipelines_addr + i * CIK_HPD_EOP_BYTES;
 512                 pr_debug("kfd: pipeline address %llX\n", pipe_hpd_addr);
 513                 /* = log2(bytes/4)-1 */
 514                 kfd2kgd->init_pipeline(dqm->dev->kgd, inx,
 515                                 CIK_HPD_EOP_BYTES_LOG2 - 3, pipe_hpd_addr);
 516         }
 517
 518         return 0;
 519 }
 520
 521 static int init_scheduler(struct device_queue_manager *dqm)
 522 {
 523         int retval;
 524
 525         BUG_ON(!dqm);
 526
 527         pr_debug("kfd: In %s\n", __func__);
 528
 529         retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
 530         return retval;
 531 }
 532
 533 static int initialize_nocpsch(struct device_queue_manager *dqm)
 534 {
 535         int i;
 536
 537         BUG_ON(!dqm);
 538
 539         pr_debug("kfd: In func %s num of pipes: %d\n",
 540                         __func__, get_pipes_num(dqm));
 541
 542         mutex_init(&dqm->lock);
 543         INIT_LIST_HEAD(&dqm->queues);
 544         dqm->queue_count = dqm->next_pipe_to_allocate = 0;
 545         dqm->sdma_queue_count = 0;
 546         dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
 547                                         sizeof(unsigned int), GFP_KERNEL);
 548         if (!dqm->allocated_queues) {
 549                 mutex_destroy(&dqm->lock);
 550                 return -ENOMEM;
 551         }
 552
 553         for (i = 0; i < get_pipes_num(dqm); i++)
 554                 dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
 555
 556         dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
 557         dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
 558
 559         init_scheduler(dqm);
 560         return 0;
 561 }
 562
 563 static void uninitialize_nocpsch(struct device_queue_manager *dqm)
 564 {
 565         int i;
 566
 567         BUG_ON(!dqm);
 568
 569         BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
 570
 571         kfree(dqm->allocated_queues);
 572         for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
 573                 kfree(dqm->mqds[i]);
 574         mutex_destroy(&dqm->lock);
 575         kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
 576 }
 577
 578 static int start_nocpsch(struct device_queue_manager *dqm)
 579 {
 580         return 0;
 581 }
 582
 583 static int stop_nocpsch(struct device_queue_manager *dqm)
 584 {
 585         return 0;
 586 }
 587
 588 static int allocate_sdma_queue(struct device_queue_manager *dqm,
 589                                 unsigned int *sdma_queue_id)
 590 {
 591         int bit;
 592
 593         if (dqm->sdma_bitmap == 0)
 594                 return -ENOMEM;
 595
 596         bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
 597                                 CIK_SDMA_QUEUES);
 598
 599         clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
 600         *sdma_queue_id = bit;
 601
 602         return 0;
 603 }
 604
 605 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
 606                                 unsigned int sdma_queue_id)
 607 {
 608         if (sdma_queue_id >= CIK_SDMA_QUEUES)
 609                 return;
 610         set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
 611 }
 612
 613 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
 614                                 struct qcm_process_device *qpd)
 615 {
 616         uint32_t value = SDMA_ATC;
 617
 618         if (q->process->is_32bit_user_mode)
 619                 value |= SDMA_VA_PTR32 | get_sh_mem_bases_32(qpd_to_pdd(qpd));
 620         else
 621                 value |= SDMA_VA_SHARED_BASE(get_sh_mem_bases_nybble_64(
 622                                                         qpd_to_pdd(qpd)));
 623         q->properties.sdma_vm_addr = value;
 624 }
 625
 626 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 627                                         struct queue *q,
 628                                         struct qcm_process_device *qpd)
 629 {
 630         struct mqd_manager *mqd;
 631         int retval;
 632
 633         mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
 634         if (!mqd)
 635                 return -ENOMEM;
 636
 637         retval = allocate_sdma_queue(dqm, &q->sdma_id);
 638         if (retval != 0)
 639                 return retval;
 640
 641         q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
 642         q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
 643
 644         pr_debug("kfd: sdma id is:    %d\n", q->sdma_id);
 645         pr_debug("     sdma queue id: %d\n", q->properties.sdma_queue_id);
 646         pr_debug("     sdma engine id: %d\n", q->properties.sdma_engine_id);
 647
 648         retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 649                                 &q->gart_mqd_addr, &q->properties);
 650         if (retval != 0) {
 651                 deallocate_sdma_queue(dqm, q->sdma_id);
 652                 return retval;
 653         }
 654
 655         init_sdma_vm(dqm, q, qpd);
 656         return 0;
 657 }
 658
 659 /*
 660  * Device Queue Manager implementation for cp scheduler
 661  */
 662
 663 static int set_sched_resources(struct device_queue_manager *dqm)
 664 {
 665         struct scheduling_resources res;
 666         unsigned int queue_num, queue_mask;
 667
 668         BUG_ON(!dqm);
 669
 670         pr_debug("kfd: In func %s\n", __func__);
 671
 672         queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
 673         queue_mask = (1 << queue_num) - 1;
 674         res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
 675         res.vmid_mask <<= KFD_VMID_START_OFFSET;
 676         res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
 677         res.gws_mask = res.oac_mask = res.gds_heap_base =
 678                                                 res.gds_heap_size = 0;
 679
 680         pr_debug("kfd: scheduling resources:\n"
 681                         "      vmid mask: 0x%8X\n"
 682                         "      queue mask: 0x%8llX\n",
 683                         res.vmid_mask, res.queue_mask);
 684
 685         return pm_send_set_resources(&dqm->packets, &res);
 686 }
 687
 688 static int initialize_cpsch(struct device_queue_manager *dqm)
 689 {
 690         int retval;
 691
 692         BUG_ON(!dqm);
 693
 694         pr_debug("kfd: In func %s num of pipes: %d\n",
 695                         __func__, get_pipes_num_cpsch());
 696
 697         mutex_init(&dqm->lock);
 698         INIT_LIST_HEAD(&dqm->queues);
 699         dqm->queue_count = dqm->processes_count = 0;
 700         dqm->sdma_queue_count = 0;
 701         dqm->active_runlist = false;
 702         retval = dqm->ops_asic_specific.initialize(dqm);
 703         if (retval != 0)
 704                 goto fail_init_pipelines;
 705
 706         return 0;
 707
 708 fail_init_pipelines:
 709         mutex_destroy(&dqm->lock);
 710         return retval;
 711 }
 712
 713 static int start_cpsch(struct device_queue_manager *dqm)
 714 {
 715         struct device_process_node *node;
 716         int retval;
 717
 718         BUG_ON(!dqm);
 719
 720         retval = 0;
 721
 722         retval = pm_init(&dqm->packets, dqm);
 723         if (retval != 0)
 724                 goto fail_packet_manager_init;
 725
 726         retval = set_sched_resources(dqm);
 727         if (retval != 0)
 728                 goto fail_set_sched_resources;
 729
 730         pr_debug("kfd: allocating fence memory\n");
 731
 732         /* allocate fence memory on the gart */
 733         retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
 734                                         &dqm->fence_mem);
 735
 736         if (retval != 0)
 737                 goto fail_allocate_vidmem;
 738
 739         dqm->fence_addr = dqm->fence_mem->cpu_ptr;
 740         dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
 741         list_for_each_entry(node, &dqm->queues, list)
 742                 if (node->qpd->pqm->process && dqm->dev)
 743                         kfd_bind_process_to_device(dqm->dev,
 744                                                 node->qpd->pqm->process);
 745
 746         execute_queues_cpsch(dqm, true);
 747
 748         return 0;
 749 fail_allocate_vidmem:
 750 fail_set_sched_resources:
 751         pm_uninit(&dqm->packets);
 752 fail_packet_manager_init:
 753         return retval;
 754 }
 755
 756 static int stop_cpsch(struct device_queue_manager *dqm)
 757 {
 758         struct device_process_node *node;
 759         struct kfd_process_device *pdd;
 760
 761         BUG_ON(!dqm);
 762
 763         destroy_queues_cpsch(dqm, true);
 764
 765         list_for_each_entry(node, &dqm->queues, list) {
 766                 pdd = qpd_to_pdd(node->qpd);
 767                 pdd->bound = false;
 768         }
 769         kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
 770         pm_uninit(&dqm->packets);
 771
 772         return 0;
 773 }
 774
 775 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
 776                                         struct kernel_queue *kq,
 777                                         struct qcm_process_device *qpd)
 778 {
 779         BUG_ON(!dqm || !kq || !qpd);
 780
 781         pr_debug("kfd: In func %s\n", __func__);
 782
 783         mutex_lock(&dqm->lock);
 784         if (dqm->total_queue_count >= max_num_of_queues_per_device) {
 785                 pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
 786                                 dqm->total_queue_count);
 787                 mutex_unlock(&dqm->lock);
 788                 return -EPERM;
 789         }
 790
 791         /*
 792          * Unconditionally increment this counter, regardless of the queue's
 793          * type or whether the queue is active.
 794          */
 795         dqm->total_queue_count++;
 796         pr_debug("Total of %d queues are accountable so far\n",
 797                         dqm->total_queue_count);
 798
 799         list_add(&kq->list, &qpd->priv_queue_list);
 800         dqm->queue_count++;
 801         qpd->is_debug = true;
 802         execute_queues_cpsch(dqm, false);
 803         mutex_unlock(&dqm->lock);
 804
 805         return 0;
 806 }
 807
 808 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
 809                                         struct kernel_queue *kq,
 810                                         struct qcm_process_device *qpd)
 811 {
 812         BUG_ON(!dqm || !kq);
 813
 814         pr_debug("kfd: In %s\n", __func__);
 815
 816         mutex_lock(&dqm->lock);
 817         destroy_queues_cpsch(dqm, false);
 818         list_del(&kq->list);
 819         dqm->queue_count--;
 820         qpd->is_debug = false;
 821         execute_queues_cpsch(dqm, false);
 822         /*
 823          * Unconditionally decrement this counter, regardless of the queue's
 824          * type.
 825          */
 826         dqm->total_queue_count--;
 827         pr_debug("Total of %d queues are accountable so far\n",
 828                         dqm->total_queue_count);
 829         mutex_unlock(&dqm->lock);
 830 }
 831
 832 static void select_sdma_engine_id(struct queue *q)
 833 {
 834         static int sdma_id;
 835
 836         q->sdma_id = sdma_id;
 837         sdma_id = (sdma_id + 1) % 2;
 838 }
 839
 840 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 841                         struct qcm_process_device *qpd, int *allocate_vmid)
 842 {
 843         int retval;
 844         struct mqd_manager *mqd;
 845
 846         BUG_ON(!dqm || !q || !qpd);
 847
 848         retval = 0;
 849
 850         if (allocate_vmid)
 851                 *allocate_vmid = 0;
 852
 853         mutex_lock(&dqm->lock);
 854
 855         if (dqm->total_queue_count >= max_num_of_queues_per_device) {
 856                 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
 857                                 dqm->total_queue_count);
 858                 retval = -EPERM;
 859                 goto out;
 860         }
 861
 862         if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 863                 select_sdma_engine_id(q);
 864
 865         mqd = dqm->ops.get_mqd_manager(dqm,
 866                         get_mqd_type_from_queue_type(q->properties.type));
 867
 868         if (mqd == NULL) {
 869                 mutex_unlock(&dqm->lock);
 870                 return -ENOMEM;
 871         }
 872
 873         retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 874                                 &q->gart_mqd_addr, &q->properties);
 875         if (retval != 0)
 876                 goto out;
 877
 878         list_add(&q->list, &qpd->queues_list);
 879         if (q->properties.is_active) {
 880                 dqm->queue_count++;
 881                 retval = execute_queues_cpsch(dqm, false);
 882         }
 883
 884         if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 885                         dqm->sdma_queue_count++;
 886         /*
 887          * Unconditionally increment this counter, regardless of the queue's
 888          * type or whether the queue is active.
 889          */
 890         dqm->total_queue_count++;
 891
 892         pr_debug("Total of %d queues are accountable so far\n",
 893                         dqm->total_queue_count);
 894
 895 out:
 896         mutex_unlock(&dqm->lock);
 897         return retval;
 898 }
 899
 900 static int fence_wait_timeout(unsigned int *fence_addr,
 901                                 unsigned int fence_value,
 902                                 unsigned long timeout)
 903 {
 904         BUG_ON(!fence_addr);
 905         timeout += jiffies;
 906
 907         while (*fence_addr != fence_value) {
 908                 if (time_after(jiffies, timeout)) {
 909                         pr_err("kfd: qcm fence wait loop timeout expired\n");
 910                         return -ETIME;
 911                 }
 912                 schedule();
 913         }
 914
 915         return 0;
 916 }
 917
 918 static int destroy_sdma_queues(struct device_queue_manager *dqm,
 919                                 unsigned int sdma_engine)
 920 {
 921         return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
 922                         KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false,
 923                         sdma_engine);
 924 }
 925
 926 static int destroy_queues_cpsch(struct device_queue_manager *dqm, bool lock)
 927 {
 928         int retval;
 929
 930         BUG_ON(!dqm);
 931
 932         retval = 0;
 933
 934         if (lock)
 935                 mutex_lock(&dqm->lock);
 936         if (dqm->active_runlist == false)
 937                 goto out;
 938
 939         pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
 940                 dqm->sdma_queue_count);
 941
 942         if (dqm->sdma_queue_count > 0) {
 943                 destroy_sdma_queues(dqm, 0);
 944                 destroy_sdma_queues(dqm, 1);
 945         }
 946
 947         retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
 948                         KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false, 0);
 949         if (retval != 0)
 950                 goto out;
 951
 952         *dqm->fence_addr = KFD_FENCE_INIT;
 953         pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
 954                                 KFD_FENCE_COMPLETED);
 955         /* should be timed out */
 956         fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
 957                                 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
 958         pm_release_ib(&dqm->packets);
 959         dqm->active_runlist = false;
 960
 961 out:
 962         if (lock)
 963                 mutex_unlock(&dqm->lock);
 964         return retval;
 965 }
 966
 967 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
 968 {
 969         int retval;
 970
 971         BUG_ON(!dqm);
 972
 973         if (lock)
 974                 mutex_lock(&dqm->lock);
 975
 976         retval = destroy_queues_cpsch(dqm, false);
 977         if (retval != 0) {
 978                 pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
 979                 goto out;
 980         }
 981
 982         if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
 983                 retval = 0;
 984                 goto out;
 985         }
 986
 987         if (dqm->active_runlist) {
 988                 retval = 0;
 989                 goto out;
 990         }
 991
 992         retval = pm_send_runlist(&dqm->packets, &dqm->queues);
 993         if (retval != 0) {
 994                 pr_err("kfd: failed to execute runlist");
 995                 goto out;
 996         }
 997         dqm->active_runlist = true;
 998
 999 out:
1000         if (lock)
1001                 mutex_unlock(&dqm->lock);
1002         return retval;
1003 }
1004
1005 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1006                                 struct qcm_process_device *qpd,
1007                                 struct queue *q)
1008 {
1009         int retval;
1010         struct mqd_manager *mqd;
1011
1012         BUG_ON(!dqm || !qpd || !q);
1013
1014         retval = 0;
1015
1016         /* remove queue from list to prevent rescheduling after preemption */
1017         mutex_lock(&dqm->lock);
1018         mqd = dqm->ops.get_mqd_manager(dqm,
1019                         get_mqd_type_from_queue_type(q->properties.type));
1020         if (!mqd) {
1021                 retval = -ENOMEM;
1022                 goto failed;
1023         }
1024
1025         if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1026                 dqm->sdma_queue_count--;
1027
1028         list_del(&q->list);
1029         if (q->properties.is_active)
1030                 dqm->queue_count--;
1031
1032         execute_queues_cpsch(dqm, false);
1033
1034         mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1035
1036         /*
1037          * Unconditionally decrement this counter, regardless of the queue's
1038          * type
1039          */
1040         dqm->total_queue_count--;
1041         pr_debug("Total of %d queues are accountable so far\n",
1042                         dqm->total_queue_count);
1043
1044         mutex_unlock(&dqm->lock);
1045
1046         return 0;
1047
1048 failed:
1049         mutex_unlock(&dqm->lock);
1050         return retval;
1051 }
1052
1053 /*
1054  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1055  * stay in user mode.
1056  */
1057 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1058 /* APE1 limit is inclusive and 64K aligned. */
1059 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1060
1061 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1062                                    struct qcm_process_device *qpd,
1063                                    enum cache_policy default_policy,
1064                                    enum cache_policy alternate_policy,
1065                                    void __user *alternate_aperture_base,
1066                                    uint64_t alternate_aperture_size)
1067 {
1068         bool retval;
1069
1070         pr_debug("kfd: In func %s\n", __func__);
1071
1072         mutex_lock(&dqm->lock);
1073
1074         if (alternate_aperture_size == 0) {
1075                 /* base > limit disables APE1 */
1076                 qpd->sh_mem_ape1_base = 1;
1077                 qpd->sh_mem_ape1_limit = 0;
1078         } else {
1079                 /*
1080                  * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1081                  *                      SH_MEM_APE1_BASE[31:0], 0x0000 }
1082                  * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1083                  *                      SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1084                  * Verify that the base and size parameters can be
1085                  * represented in this format and convert them.
1086                  * Additionally restrict APE1 to user-mode addresses.
1087                  */
1088
1089                 uint64_t base = (uintptr_t)alternate_aperture_base;
1090                 uint64_t limit = base + alternate_aperture_size - 1;
1091
1092                 if (limit <= base)
1093                         goto out;
1094
1095                 if ((base & APE1_FIXED_BITS_MASK) != 0)
1096                         goto out;
1097
1098                 if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1099                         goto out;
1100
1101                 qpd->sh_mem_ape1_base = base >> 16;
1102                 qpd->sh_mem_ape1_limit = limit >> 16;
1103         }
1104
1105         retval = dqm->ops_asic_specific.set_cache_memory_policy(
1106                         dqm,
1107                         qpd,
1108                         default_policy,
1109                         alternate_policy,
1110                         alternate_aperture_base,
1111                         alternate_aperture_size);
1112
1113         if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1114                 program_sh_mem_settings(dqm, qpd);
1115
1116         pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1117                 qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1118                 qpd->sh_mem_ape1_limit);
1119
1120         mutex_unlock(&dqm->lock);
1121         return retval;
1122
1123 out:
1124         mutex_unlock(&dqm->lock);
1125         return false;
1126 }
1127
1128 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1129 {
1130         struct device_queue_manager *dqm;
1131
1132         BUG_ON(!dev);
1133
1134         pr_debug("kfd: loading device queue manager\n");
1135
1136         dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1137         if (!dqm)
1138                 return NULL;
1139
1140         dqm->dev = dev;
1141         switch (sched_policy) {
1142         case KFD_SCHED_POLICY_HWS:
1143         case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1144                 /* initialize dqm for cp scheduling */
1145                 dqm->ops.create_queue = create_queue_cpsch;
1146                 dqm->ops.initialize = initialize_cpsch;
1147                 dqm->ops.start = start_cpsch;
1148                 dqm->ops.stop = stop_cpsch;
1149                 dqm->ops.destroy_queue = destroy_queue_cpsch;
1150                 dqm->ops.update_queue = update_queue;
1151                 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1152                 dqm->ops.register_process = register_process_nocpsch;
1153                 dqm->ops.unregister_process = unregister_process_nocpsch;
1154                 dqm->ops.uninitialize = uninitialize_nocpsch;
1155                 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1156                 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1157                 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1158                 break;
1159         case KFD_SCHED_POLICY_NO_HWS:
1160                 /* initialize dqm for no cp scheduling */
1161                 dqm->ops.start = start_nocpsch;
1162                 dqm->ops.stop = stop_nocpsch;
1163                 dqm->ops.create_queue = create_queue_nocpsch;
1164                 dqm->ops.destroy_queue = destroy_queue_nocpsch;
1165                 dqm->ops.update_queue = update_queue;
1166                 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1167                 dqm->ops.register_process = register_process_nocpsch;
1168                 dqm->ops.unregister_process = unregister_process_nocpsch;
1169                 dqm->ops.initialize = initialize_nocpsch;
1170                 dqm->ops.uninitialize = uninitialize_nocpsch;
1171                 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1172                 break;
1173         default:
1174                 BUG();
1175                 break;
1176         }
1177
1178         switch (dev->device_info->asic_family) {
1179         case CHIP_CARRIZO:
1180                 device_queue_manager_init_vi(&dqm->ops_asic_specific);
1181                 break;
1182
1183         case CHIP_KAVERI:
1184                 device_queue_manager_init_cik(&dqm->ops_asic_specific);
1185                 break;
1186         }
1187
1188         if (dqm->ops.initialize(dqm) != 0) {
1189                 kfree(dqm);
1190                 return NULL;
1191         }
1192
1193         return dqm;
1194 }
1195
1196 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1197 {
1198         BUG_ON(!dqm);
1199
1200         dqm->ops.uninitialize(dqm);
1201         kfree(dqm);
1202 }