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

   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_node *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_node *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 (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) && !dev->kfd->shared_resources.enable_mes) {
 127                 if (gws)
 128                         ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
 129                                 gws, &mem);
 130                 else
 131                         ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
 132                                 pqn->q->gws);
 133                 if (unlikely(ret))
 134                         return ret;
 135                 pqn->q->gws = mem;
 136         } else {
 137                 /*
 138                  * Intentionally set GWS to a non-NULL value
 139                  * for devices that do not use GWS for global wave
 140                  * synchronization but require the formality
 141                  * of setting GWS for cooperative groups.
 142                  */
 143                 pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
 144         }
 145
 146         pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0;
 147
 148         return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 149                                                         pqn->q, NULL);
 150 }
 151
 152 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
 153 {
 154         int i;
 155
 156         for (i = 0; i < p->n_pdds; i++)
 157                 kfd_process_dequeue_from_device(p->pdds[i]);
 158 }
 159
 160 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
 161 {
 162         INIT_LIST_HEAD(&pqm->queues);
 163         pqm->queue_slot_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
 164                                                GFP_KERNEL);
 165         if (!pqm->queue_slot_bitmap)
 166                 return -ENOMEM;
 167         pqm->process = p;
 168
 169         return 0;
 170 }
 171
 172 static void pqm_clean_queue_resource(struct process_queue_manager *pqm,
 173                                      struct process_queue_node *pqn)
 174 {
 175         struct kfd_node *dev;
 176         struct kfd_process_device *pdd;
 177
 178         dev = pqn->q->device;
 179
 180         pdd = kfd_get_process_device_data(dev, pqm->process);
 181         if (!pdd) {
 182                 pr_err("Process device data doesn't exist\n");
 183                 return;
 184         }
 185
 186         if (pqn->q->gws) {
 187                 if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
 188                     !dev->kfd->shared_resources.enable_mes)
 189                         amdgpu_amdkfd_remove_gws_from_process(
 190                                 pqm->process->kgd_process_info, pqn->q->gws);
 191                 pdd->qpd.num_gws = 0;
 192         }
 193
 194         if (dev->kfd->shared_resources.enable_mes) {
 195                 amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->gang_ctx_bo);
 196                 if (pqn->q->wptr_bo)
 197                         amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
 198         }
 199 }
 200
 201 void pqm_uninit(struct process_queue_manager *pqm)
 202 {
 203         struct process_queue_node *pqn, *next;
 204
 205         list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
 206                 if (pqn->q)
 207                         pqm_clean_queue_resource(pqm, pqn);
 208
 209                 kfd_procfs_del_queue(pqn->q);
 210                 uninit_queue(pqn->q);
 211                 list_del(&pqn->process_queue_list);
 212                 kfree(pqn);
 213         }
 214
 215         bitmap_free(pqm->queue_slot_bitmap);
 216         pqm->queue_slot_bitmap = NULL;
 217 }
 218
 219 static int init_user_queue(struct process_queue_manager *pqm,
 220                                 struct kfd_node *dev, struct queue **q,
 221                                 struct queue_properties *q_properties,
 222                                 struct file *f, struct amdgpu_bo *wptr_bo,
 223                                 unsigned int qid)
 224 {
 225         int retval;
 226
 227         /* Doorbell initialized in user space*/
 228         q_properties->doorbell_ptr = NULL;
 229         q_properties->exception_status = KFD_EC_MASK(EC_QUEUE_NEW);
 230
 231         /* let DQM handle it*/
 232         q_properties->vmid = 0;
 233         q_properties->queue_id = qid;
 234
 235         retval = init_queue(q, q_properties);
 236         if (retval != 0)
 237                 return retval;
 238
 239         (*q)->device = dev;
 240         (*q)->process = pqm->process;
 241
 242         if (dev->kfd->shared_resources.enable_mes) {
 243                 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
 244                                                 AMDGPU_MES_GANG_CTX_SIZE,
 245                                                 &(*q)->gang_ctx_bo,
 246                                                 &(*q)->gang_ctx_gpu_addr,
 247                                                 &(*q)->gang_ctx_cpu_ptr,
 248                                                 false);
 249                 if (retval) {
 250                         pr_err("failed to allocate gang context bo\n");
 251                         goto cleanup;
 252                 }
 253                 memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
 254                 (*q)->wptr_bo = wptr_bo;
 255         }
 256
 257         pr_debug("PQM After init queue");
 258         return 0;
 259
 260 cleanup:
 261         uninit_queue(*q);
 262         *q = NULL;
 263         return retval;
 264 }
 265
 266 int pqm_create_queue(struct process_queue_manager *pqm,
 267                             struct kfd_node *dev,
 268                             struct file *f,
 269                             struct queue_properties *properties,
 270                             unsigned int *qid,
 271                             struct amdgpu_bo *wptr_bo,
 272                             const struct kfd_criu_queue_priv_data *q_data,
 273                             const void *restore_mqd,
 274                             const void *restore_ctl_stack,
 275                             uint32_t *p_doorbell_offset_in_process)
 276 {
 277         int retval;
 278         struct kfd_process_device *pdd;
 279         struct queue *q;
 280         struct process_queue_node *pqn;
 281         struct kernel_queue *kq;
 282         enum kfd_queue_type type = properties->type;
 283         unsigned int max_queues = 127; /* HWS limit */
 284
 285         /*
 286          * On GFX 9.4.3, increase the number of queues that
 287          * can be created to 255. No HWS limit on GFX 9.4.3.
 288          */
 289         if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3))
 290                 max_queues = 255;
 291
 292         q = NULL;
 293         kq = NULL;
 294
 295         pdd = kfd_get_process_device_data(dev, pqm->process);
 296         if (!pdd) {
 297                 pr_err("Process device data doesn't exist\n");
 298                 return -1;
 299         }
 300
 301         /*
 302          * for debug process, verify that it is within the static queues limit
 303          * currently limit is set to half of the total avail HQD slots
 304          * If we are just about to create DIQ, the is_debug flag is not set yet
 305          * Hence we also check the type as well
 306          */
 307         if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
 308                 max_queues = dev->kfd->device_info.max_no_of_hqd/2;
 309
 310         if (pdd->qpd.queue_count >= max_queues)
 311                 return -ENOSPC;
 312
 313         if (q_data) {
 314                 retval = assign_queue_slot_by_qid(pqm, q_data->q_id);
 315                 *qid = q_data->q_id;
 316         } else
 317                 retval = find_available_queue_slot(pqm, qid);
 318
 319         if (retval != 0)
 320                 return retval;
 321
 322         if (list_empty(&pdd->qpd.queues_list) &&
 323             list_empty(&pdd->qpd.priv_queue_list))
 324                 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
 325
 326         pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
 327         if (!pqn) {
 328                 retval = -ENOMEM;
 329                 goto err_allocate_pqn;
 330         }
 331
 332         switch (type) {
 333         case KFD_QUEUE_TYPE_SDMA:
 334         case KFD_QUEUE_TYPE_SDMA_XGMI:
 335                 /* SDMA queues are always allocated statically no matter
 336                  * which scheduler mode is used. We also do not need to
 337                  * check whether a SDMA queue can be allocated here, because
 338                  * allocate_sdma_queue() in create_queue() has the
 339                  * corresponding check logic.
 340                  */
 341                 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 342                 if (retval != 0)
 343                         goto err_create_queue;
 344                 pqn->q = q;
 345                 pqn->kq = NULL;
 346                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
 347                                                     restore_mqd, restore_ctl_stack);
 348                 print_queue(q);
 349                 break;
 350
 351         case KFD_QUEUE_TYPE_COMPUTE:
 352                 /* check if there is over subscription */
 353                 if ((dev->dqm->sched_policy ==
 354                      KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
 355                 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
 356                 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) {
 357                         pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
 358                         retval = -EPERM;
 359                         goto err_create_queue;
 360                 }
 361
 362                 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 363                 if (retval != 0)
 364                         goto err_create_queue;
 365                 pqn->q = q;
 366                 pqn->kq = NULL;
 367                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
 368                                                     restore_mqd, restore_ctl_stack);
 369                 print_queue(q);
 370                 break;
 371         case KFD_QUEUE_TYPE_DIQ:
 372                 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
 373                 if (!kq) {
 374                         retval = -ENOMEM;
 375                         goto err_create_queue;
 376                 }
 377                 kq->queue->properties.queue_id = *qid;
 378                 pqn->kq = kq;
 379                 pqn->q = NULL;
 380                 retval = kfd_process_drain_interrupts(pdd);
 381                 if (retval)
 382                         break;
 383
 384                 retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
 385                                                         kq, &pdd->qpd);
 386                 break;
 387         default:
 388                 WARN(1, "Invalid queue type %d", type);
 389                 retval = -EINVAL;
 390         }
 391
 392         if (retval != 0) {
 393                 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n",
 394                         pqm->process->pasid, type, retval);
 395                 goto err_create_queue;
 396         }
 397
 398         if (q && p_doorbell_offset_in_process) {
 399                 /* Return the doorbell offset within the doorbell page
 400                  * to the caller so it can be passed up to user mode
 401                  * (in bytes).
 402                  * relative doorbell index = Absolute doorbell index -
 403                  * absolute index of first doorbell in the page.
 404                  */
 405                 uint32_t first_db_index = amdgpu_doorbell_index_on_bar(pdd->dev->adev,
 406                                                                        pdd->qpd.proc_doorbells,
 407                                                                        0,
 408                                                                        pdd->dev->kfd->device_info.doorbell_size);
 409
 410                 *p_doorbell_offset_in_process = (q->properties.doorbell_off
 411                                                 - first_db_index) * sizeof(uint32_t);
 412         }
 413
 414         pr_debug("PQM After DQM create queue\n");
 415
 416         list_add(&pqn->process_queue_list, &pqm->queues);
 417
 418         if (q) {
 419                 pr_debug("PQM done creating queue\n");
 420                 kfd_procfs_add_queue(q);
 421                 print_queue_properties(&q->properties);
 422         }
 423
 424         return retval;
 425
 426 err_create_queue:
 427         uninit_queue(q);
 428         if (kq)
 429                 kernel_queue_uninit(kq, false);
 430         kfree(pqn);
 431 err_allocate_pqn:
 432         /* check if queues list is empty unregister process from device */
 433         clear_bit(*qid, pqm->queue_slot_bitmap);
 434         if (list_empty(&pdd->qpd.queues_list) &&
 435             list_empty(&pdd->qpd.priv_queue_list))
 436                 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
 437         return retval;
 438 }
 439
 440 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 441 {
 442         struct process_queue_node *pqn;
 443         struct kfd_process_device *pdd;
 444         struct device_queue_manager *dqm;
 445         struct kfd_node *dev;
 446         int retval;
 447
 448         dqm = NULL;
 449
 450         retval = 0;
 451
 452         pqn = get_queue_by_qid(pqm, qid);
 453         if (!pqn) {
 454                 pr_err("Queue id does not match any known queue\n");
 455                 return -EINVAL;
 456         }
 457
 458         dev = NULL;
 459         if (pqn->kq)
 460                 dev = pqn->kq->dev;
 461         if (pqn->q)
 462                 dev = pqn->q->device;
 463         if (WARN_ON(!dev))
 464                 return -ENODEV;
 465
 466         pdd = kfd_get_process_device_data(dev, pqm->process);
 467         if (!pdd) {
 468                 pr_err("Process device data doesn't exist\n");
 469                 return -1;
 470         }
 471
 472         if (pqn->kq) {
 473                 /* destroy kernel queue (DIQ) */
 474                 dqm = pqn->kq->dev->dqm;
 475                 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
 476                 kernel_queue_uninit(pqn->kq, false);
 477         }
 478
 479         if (pqn->q) {
 480                 kfd_procfs_del_queue(pqn->q);
 481                 dqm = pqn->q->device->dqm;
 482                 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
 483                 if (retval) {
 484                         pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
 485                                 pqm->process->pasid,
 486                                 pqn->q->properties.queue_id, retval);
 487                         if (retval != -ETIME)
 488                                 goto err_destroy_queue;
 489                 }
 490
 491                 pqm_clean_queue_resource(pqm, pqn);
 492                 uninit_queue(pqn->q);
 493         }
 494
 495         list_del(&pqn->process_queue_list);
 496         kfree(pqn);
 497         clear_bit(qid, pqm->queue_slot_bitmap);
 498
 499         if (list_empty(&pdd->qpd.queues_list) &&
 500             list_empty(&pdd->qpd.priv_queue_list))
 501                 dqm->ops.unregister_process(dqm, &pdd->qpd);
 502
 503 err_destroy_queue:
 504         return retval;
 505 }
 506
 507 int pqm_update_queue_properties(struct process_queue_manager *pqm,
 508                                 unsigned int qid, struct queue_properties *p)
 509 {
 510         int retval;
 511         struct process_queue_node *pqn;
 512
 513         pqn = get_queue_by_qid(pqm, qid);
 514         if (!pqn) {
 515                 pr_debug("No queue %d exists for update operation\n", qid);
 516                 return -EFAULT;
 517         }
 518
 519         pqn->q->properties.queue_address = p->queue_address;
 520         pqn->q->properties.queue_size = p->queue_size;
 521         pqn->q->properties.queue_percent = p->queue_percent;
 522         pqn->q->properties.priority = p->priority;
 523         pqn->q->properties.pm4_target_xcc = p->pm4_target_xcc;
 524
 525         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 526                                                         pqn->q, NULL);
 527         if (retval != 0)
 528                 return retval;
 529
 530         return 0;
 531 }
 532
 533 int pqm_update_mqd(struct process_queue_manager *pqm,
 534                                 unsigned int qid, struct mqd_update_info *minfo)
 535 {
 536         int retval;
 537         struct process_queue_node *pqn;
 538
 539         pqn = get_queue_by_qid(pqm, qid);
 540         if (!pqn) {
 541                 pr_debug("No queue %d exists for update operation\n", qid);
 542                 return -EFAULT;
 543         }
 544
 545         /* CUs are masked for debugger requirements so deny user mask  */
 546         if (pqn->q->properties.is_dbg_wa && minfo && minfo->cu_mask.ptr)
 547                 return -EBUSY;
 548
 549         /* ASICs that have WGPs must enforce pairwise enabled mask checks. */
 550         if (minfo && minfo->cu_mask.ptr &&
 551                         KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) {
 552                 int i;
 553
 554                 for (i = 0; i < minfo->cu_mask.count; i += 2) {
 555                         uint32_t cu_pair = (minfo->cu_mask.ptr[i / 32] >> (i % 32)) & 0x3;
 556
 557                         if (cu_pair && cu_pair != 0x3) {
 558                                 pr_debug("CUs must be adjacent pairwise enabled.\n");
 559                                 return -EINVAL;
 560                         }
 561                 }
 562         }
 563
 564         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 565                                                         pqn->q, minfo);
 566         if (retval != 0)
 567                 return retval;
 568
 569         if (minfo && minfo->cu_mask.ptr)
 570                 pqn->q->properties.is_user_cu_masked = true;
 571
 572         return 0;
 573 }
 574
 575 struct kernel_queue *pqm_get_kernel_queue(
 576                                         struct process_queue_manager *pqm,
 577                                         unsigned int qid)
 578 {
 579         struct process_queue_node *pqn;
 580
 581         pqn = get_queue_by_qid(pqm, qid);
 582         if (pqn && pqn->kq)
 583                 return pqn->kq;
 584
 585         return NULL;
 586 }
 587
 588 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
 589                                         unsigned int qid)
 590 {
 591         struct process_queue_node *pqn;
 592
 593         pqn = get_queue_by_qid(pqm, qid);
 594         return pqn ? pqn->q : NULL;
 595 }
 596
 597 int pqm_get_wave_state(struct process_queue_manager *pqm,
 598                        unsigned int qid,
 599                        void __user *ctl_stack,
 600                        u32 *ctl_stack_used_size,
 601                        u32 *save_area_used_size)
 602 {
 603         struct process_queue_node *pqn;
 604
 605         pqn = get_queue_by_qid(pqm, qid);
 606         if (!pqn) {
 607                 pr_debug("amdkfd: No queue %d exists for operation\n",
 608                          qid);
 609                 return -EFAULT;
 610         }
 611
 612         return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
 613                                                        pqn->q,
 614                                                        ctl_stack,
 615                                                        ctl_stack_used_size,
 616                                                        save_area_used_size);
 617 }
 618
 619 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
 620                            uint64_t exception_clear_mask,
 621                            void __user *buf,
 622                            int *num_qss_entries,
 623                            uint32_t *entry_size)
 624 {
 625         struct process_queue_node *pqn;
 626         struct kfd_queue_snapshot_entry src;
 627         uint32_t tmp_entry_size = *entry_size, tmp_qss_entries = *num_qss_entries;
 628         int r = 0;
 629
 630         *num_qss_entries = 0;
 631         if (!(*entry_size))
 632                 return -EINVAL;
 633
 634         *entry_size = min_t(size_t, *entry_size, sizeof(struct kfd_queue_snapshot_entry));
 635         mutex_lock(&pqm->process->event_mutex);
 636
 637         memset(&src, 0, sizeof(src));
 638
 639         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
 640                 if (!pqn->q)
 641                         continue;
 642
 643                 if (*num_qss_entries < tmp_qss_entries) {
 644                         set_queue_snapshot_entry(pqn->q, exception_clear_mask, &src);
 645
 646                         if (copy_to_user(buf, &src, *entry_size)) {
 647                                 r = -EFAULT;
 648                                 break;
 649                         }
 650                         buf += tmp_entry_size;
 651                 }
 652                 *num_qss_entries += 1;
 653         }
 654
 655         mutex_unlock(&pqm->process->event_mutex);
 656         return r;
 657 }
 658
 659 static int get_queue_data_sizes(struct kfd_process_device *pdd,
 660                                 struct queue *q,
 661                                 uint32_t *mqd_size,
 662                                 uint32_t *ctl_stack_size)
 663 {
 664         int ret;
 665
 666         ret = pqm_get_queue_checkpoint_info(&pdd->process->pqm,
 667                                             q->properties.queue_id,
 668                                             mqd_size,
 669                                             ctl_stack_size);
 670         if (ret)
 671                 pr_err("Failed to get queue dump info (%d)\n", ret);
 672
 673         return ret;
 674 }
 675
 676 int kfd_process_get_queue_info(struct kfd_process *p,
 677                                uint32_t *num_queues,
 678                                uint64_t *priv_data_sizes)
 679 {
 680         uint32_t extra_data_sizes = 0;
 681         struct queue *q;
 682         int i;
 683         int ret;
 684
 685         *num_queues = 0;
 686
 687         /* Run over all PDDs of the process */
 688         for (i = 0; i < p->n_pdds; i++) {
 689                 struct kfd_process_device *pdd = p->pdds[i];
 690
 691                 list_for_each_entry(q, &pdd->qpd.queues_list, list) {
 692                         if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 693                                 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 694                                 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 695                                 uint32_t mqd_size, ctl_stack_size;
 696
 697                                 *num_queues = *num_queues + 1;
 698
 699                                 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
 700                                 if (ret)
 701                                         return ret;
 702
 703                                 extra_data_sizes += mqd_size + ctl_stack_size;
 704                         } else {
 705                                 pr_err("Unsupported queue type (%d)\n", q->properties.type);
 706                                 return -EOPNOTSUPP;
 707                         }
 708                 }
 709         }
 710         *priv_data_sizes = extra_data_sizes +
 711                                 (*num_queues * sizeof(struct kfd_criu_queue_priv_data));
 712
 713         return 0;
 714 }
 715
 716 static int pqm_checkpoint_mqd(struct process_queue_manager *pqm,
 717                               unsigned int qid,
 718                               void *mqd,
 719                               void *ctl_stack)
 720 {
 721         struct process_queue_node *pqn;
 722
 723         pqn = get_queue_by_qid(pqm, qid);
 724         if (!pqn) {
 725                 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
 726                 return -EFAULT;
 727         }
 728
 729         if (!pqn->q->device->dqm->ops.checkpoint_mqd) {
 730                 pr_err("amdkfd: queue dumping not supported on this device\n");
 731                 return -EOPNOTSUPP;
 732         }
 733
 734         return pqn->q->device->dqm->ops.checkpoint_mqd(pqn->q->device->dqm,
 735                                                        pqn->q, mqd, ctl_stack);
 736 }
 737
 738 static int criu_checkpoint_queue(struct kfd_process_device *pdd,
 739                            struct queue *q,
 740                            struct kfd_criu_queue_priv_data *q_data)
 741 {
 742         uint8_t *mqd, *ctl_stack;
 743         int ret;
 744
 745         mqd = (void *)(q_data + 1);
 746         ctl_stack = mqd + q_data->mqd_size;
 747
 748         q_data->gpu_id = pdd->user_gpu_id;
 749         q_data->type = q->properties.type;
 750         q_data->format = q->properties.format;
 751         q_data->q_id =  q->properties.queue_id;
 752         q_data->q_address = q->properties.queue_address;
 753         q_data->q_size = q->properties.queue_size;
 754         q_data->priority = q->properties.priority;
 755         q_data->q_percent = q->properties.queue_percent;
 756         q_data->read_ptr_addr = (uint64_t)q->properties.read_ptr;
 757         q_data->write_ptr_addr = (uint64_t)q->properties.write_ptr;
 758         q_data->doorbell_id = q->doorbell_id;
 759
 760         q_data->sdma_id = q->sdma_id;
 761
 762         q_data->eop_ring_buffer_address =
 763                 q->properties.eop_ring_buffer_address;
 764
 765         q_data->eop_ring_buffer_size = q->properties.eop_ring_buffer_size;
 766
 767         q_data->ctx_save_restore_area_address =
 768                 q->properties.ctx_save_restore_area_address;
 769
 770         q_data->ctx_save_restore_area_size =
 771                 q->properties.ctx_save_restore_area_size;
 772
 773         q_data->gws = !!q->gws;
 774
 775         ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack);
 776         if (ret) {
 777                 pr_err("Failed checkpoint queue_mqd (%d)\n", ret);
 778                 return ret;
 779         }
 780
 781         pr_debug("Dumping Queue: gpu_id:%x queue_id:%u\n", q_data->gpu_id, q_data->q_id);
 782         return ret;
 783 }
 784
 785 static int criu_checkpoint_queues_device(struct kfd_process_device *pdd,
 786                                    uint8_t __user *user_priv,
 787                                    unsigned int *q_index,
 788                                    uint64_t *queues_priv_data_offset)
 789 {
 790         unsigned int q_private_data_size = 0;
 791         uint8_t *q_private_data = NULL; /* Local buffer to store individual queue private data */
 792         struct queue *q;
 793         int ret = 0;
 794
 795         list_for_each_entry(q, &pdd->qpd.queues_list, list) {
 796                 struct kfd_criu_queue_priv_data *q_data;
 797                 uint64_t q_data_size;
 798                 uint32_t mqd_size;
 799                 uint32_t ctl_stack_size;
 800
 801                 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE &&
 802                         q->properties.type != KFD_QUEUE_TYPE_SDMA &&
 803                         q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI) {
 804
 805                         pr_err("Unsupported queue type (%d)\n", q->properties.type);
 806                         ret = -EOPNOTSUPP;
 807                         break;
 808                 }
 809
 810                 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
 811                 if (ret)
 812                         break;
 813
 814                 q_data_size = sizeof(*q_data) + mqd_size + ctl_stack_size;
 815
 816                 /* Increase local buffer space if needed */
 817                 if (q_private_data_size < q_data_size) {
 818                         kfree(q_private_data);
 819
 820                         q_private_data = kzalloc(q_data_size, GFP_KERNEL);
 821                         if (!q_private_data) {
 822                                 ret = -ENOMEM;
 823                                 break;
 824                         }
 825                         q_private_data_size = q_data_size;
 826                 }
 827
 828                 q_data = (struct kfd_criu_queue_priv_data *)q_private_data;
 829
 830                 /* data stored in this order: priv_data, mqd, ctl_stack */
 831                 q_data->mqd_size = mqd_size;
 832                 q_data->ctl_stack_size = ctl_stack_size;
 833
 834                 ret = criu_checkpoint_queue(pdd, q, q_data);
 835                 if (ret)
 836                         break;
 837
 838                 q_data->object_type = KFD_CRIU_OBJECT_TYPE_QUEUE;
 839
 840                 ret = copy_to_user(user_priv + *queues_priv_data_offset,
 841                                 q_data, q_data_size);
 842                 if (ret) {
 843                         ret = -EFAULT;
 844                         break;
 845                 }
 846                 *queues_priv_data_offset += q_data_size;
 847                 *q_index = *q_index + 1;
 848         }
 849
 850         kfree(q_private_data);
 851
 852         return ret;
 853 }
 854
 855 int kfd_criu_checkpoint_queues(struct kfd_process *p,
 856                          uint8_t __user *user_priv_data,
 857                          uint64_t *priv_data_offset)
 858 {
 859         int ret = 0, pdd_index, q_index = 0;
 860
 861         for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
 862                 struct kfd_process_device *pdd = p->pdds[pdd_index];
 863
 864                 /*
 865                  * criu_checkpoint_queues_device will copy data to user and update q_index and
 866                  * queues_priv_data_offset
 867                  */
 868                 ret = criu_checkpoint_queues_device(pdd, user_priv_data, &q_index,
 869                                               priv_data_offset);
 870
 871                 if (ret)
 872                         break;
 873         }
 874
 875         return ret;
 876 }
 877
 878 static void set_queue_properties_from_criu(struct queue_properties *qp,
 879                                           struct kfd_criu_queue_priv_data *q_data)
 880 {
 881         qp->is_interop = false;
 882         qp->queue_percent = q_data->q_percent;
 883         qp->priority = q_data->priority;
 884         qp->queue_address = q_data->q_address;
 885         qp->queue_size = q_data->q_size;
 886         qp->read_ptr = (uint32_t *) q_data->read_ptr_addr;
 887         qp->write_ptr = (uint32_t *) q_data->write_ptr_addr;
 888         qp->eop_ring_buffer_address = q_data->eop_ring_buffer_address;
 889         qp->eop_ring_buffer_size = q_data->eop_ring_buffer_size;
 890         qp->ctx_save_restore_area_address = q_data->ctx_save_restore_area_address;
 891         qp->ctx_save_restore_area_size = q_data->ctx_save_restore_area_size;
 892         qp->ctl_stack_size = q_data->ctl_stack_size;
 893         qp->type = q_data->type;
 894         qp->format = q_data->format;
 895 }
 896
 897 int kfd_criu_restore_queue(struct kfd_process *p,
 898                            uint8_t __user *user_priv_ptr,
 899                            uint64_t *priv_data_offset,
 900                            uint64_t max_priv_data_size)
 901 {
 902         uint8_t *mqd, *ctl_stack, *q_extra_data = NULL;
 903         struct kfd_criu_queue_priv_data *q_data;
 904         struct kfd_process_device *pdd;
 905         uint64_t q_extra_data_size;
 906         struct queue_properties qp;
 907         unsigned int queue_id;
 908         int ret = 0;
 909
 910         if (*priv_data_offset + sizeof(*q_data) > max_priv_data_size)
 911                 return -EINVAL;
 912
 913         q_data = kmalloc(sizeof(*q_data), GFP_KERNEL);
 914         if (!q_data)
 915                 return -ENOMEM;
 916
 917         ret = copy_from_user(q_data, user_priv_ptr + *priv_data_offset, sizeof(*q_data));
 918         if (ret) {
 919                 ret = -EFAULT;
 920                 goto exit;
 921         }
 922
 923         *priv_data_offset += sizeof(*q_data);
 924         q_extra_data_size = (uint64_t)q_data->ctl_stack_size + q_data->mqd_size;
 925
 926         if (*priv_data_offset + q_extra_data_size > max_priv_data_size) {
 927                 ret = -EINVAL;
 928                 goto exit;
 929         }
 930
 931         q_extra_data = kmalloc(q_extra_data_size, GFP_KERNEL);
 932         if (!q_extra_data) {
 933                 ret = -ENOMEM;
 934                 goto exit;
 935         }
 936
 937         ret = copy_from_user(q_extra_data, user_priv_ptr + *priv_data_offset, q_extra_data_size);
 938         if (ret) {
 939                 ret = -EFAULT;
 940                 goto exit;
 941         }
 942
 943         *priv_data_offset += q_extra_data_size;
 944
 945         pdd = kfd_process_device_data_by_id(p, q_data->gpu_id);
 946         if (!pdd) {
 947                 pr_err("Failed to get pdd\n");
 948                 ret = -EINVAL;
 949                 goto exit;
 950         }
 951
 952         /* data stored in this order: mqd, ctl_stack */
 953         mqd = q_extra_data;
 954         ctl_stack = mqd + q_data->mqd_size;
 955
 956         memset(&qp, 0, sizeof(qp));
 957         set_queue_properties_from_criu(&qp, q_data);
 958
 959         print_queue_properties(&qp);
 960
 961         ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
 962                                 NULL);
 963         if (ret) {
 964                 pr_err("Failed to create new queue err:%d\n", ret);
 965                 goto exit;
 966         }
 967
 968         if (q_data->gws)
 969                 ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws);
 970
 971 exit:
 972         if (ret)
 973                 pr_err("Failed to restore queue (%d)\n", ret);
 974         else
 975                 pr_debug("Queue id %d was restored successfully\n", queue_id);
 976
 977         kfree(q_data);
 978
 979         return ret;
 980 }
 981
 982 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
 983                                   unsigned int qid,
 984                                   uint32_t *mqd_size,
 985                                   uint32_t *ctl_stack_size)
 986 {
 987         struct process_queue_node *pqn;
 988
 989         pqn = get_queue_by_qid(pqm, qid);
 990         if (!pqn) {
 991                 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
 992                 return -EFAULT;
 993         }
 994
 995         if (!pqn->q->device->dqm->ops.get_queue_checkpoint_info) {
 996                 pr_err("amdkfd: queue dumping not supported on this device\n");
 997                 return -EOPNOTSUPP;
 998         }
 999
1000         pqn->q->device->dqm->ops.get_queue_checkpoint_info(pqn->q->device->dqm,
1001                                                        pqn->q, mqd_size,
1002                                                        ctl_stack_size);
1003         return 0;
1004 }
1005
1006 #if defined(CONFIG_DEBUG_FS)
1007
1008 int pqm_debugfs_mqds(struct seq_file *m, void *data)
1009 {
1010         struct process_queue_manager *pqm = data;
1011         struct process_queue_node *pqn;
1012         struct queue *q;
1013         enum KFD_MQD_TYPE mqd_type;
1014         struct mqd_manager *mqd_mgr;
1015         int r = 0, xcc, num_xccs = 1;
1016         void *mqd;
1017         uint64_t size = 0;
1018
1019         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1020                 if (pqn->q) {
1021                         q = pqn->q;
1022                         switch (q->properties.type) {
1023                         case KFD_QUEUE_TYPE_SDMA:
1024                         case KFD_QUEUE_TYPE_SDMA_XGMI:
1025                                 seq_printf(m, "  SDMA queue on device %x\n",
1026                                            q->device->id);
1027                                 mqd_type = KFD_MQD_TYPE_SDMA;
1028                                 break;
1029                         case KFD_QUEUE_TYPE_COMPUTE:
1030                                 seq_printf(m, "  Compute queue on device %x\n",
1031                                            q->device->id);
1032                                 mqd_type = KFD_MQD_TYPE_CP;
1033                                 num_xccs = NUM_XCC(q->device->xcc_mask);
1034                                 break;
1035                         default:
1036                                 seq_printf(m,
1037                                 "  Bad user queue type %d on device %x\n",
1038                                            q->properties.type, q->device->id);
1039                                 continue;
1040                         }
1041                         mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
1042                         size = mqd_mgr->mqd_stride(mqd_mgr,
1043                                                         &q->properties);
1044                 } else if (pqn->kq) {
1045                         q = pqn->kq->queue;
1046                         mqd_mgr = pqn->kq->mqd_mgr;
1047                         switch (q->properties.type) {
1048                         case KFD_QUEUE_TYPE_DIQ:
1049                                 seq_printf(m, "  DIQ on device %x\n",
1050                                            pqn->kq->dev->id);
1051                                 break;
1052                         default:
1053                                 seq_printf(m,
1054                                 "  Bad kernel queue type %d on device %x\n",
1055                                            q->properties.type,
1056                                            pqn->kq->dev->id);
1057                                 continue;
1058                         }
1059                 } else {
1060                         seq_printf(m,
1061                 "  Weird: Queue node with neither kernel nor user queue\n");
1062                         continue;
1063                 }
1064
1065                 for (xcc = 0; xcc < num_xccs; xcc++) {
1066                         mqd = q->mqd + size * xcc;
1067                         r = mqd_mgr->debugfs_show_mqd(m, mqd);
1068                         if (r != 0)
1069                                 break;
1070                 }
1071         }
1072
1073         return r;
1074 }
1075
1076 #endif