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 #include "amdgpu_reset.h"
  32
  33 static inline struct process_queue_node *get_queue_by_qid(
  34                         struct process_queue_manager *pqm, unsigned int qid)
  35 {
  36         struct process_queue_node *pqn;
  37
  38         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
  39                 if ((pqn->q && pqn->q->properties.queue_id == qid) ||
  40                     (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
  41                         return pqn;
  42         }
  43
  44         return NULL;
  45 }
  46
  47 static int assign_queue_slot_by_qid(struct process_queue_manager *pqm,
  48                                     unsigned int qid)
  49 {
  50         if (qid >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
  51                 return -EINVAL;
  52
  53         if (__test_and_set_bit(qid, pqm->queue_slot_bitmap)) {
  54                 pr_err("Cannot create new queue because requested qid(%u) is in use\n", qid);
  55                 return -ENOSPC;
  56         }
  57
  58         return 0;
  59 }
  60
  61 static int find_available_queue_slot(struct process_queue_manager *pqm,
  62                                         unsigned int *qid)
  63 {
  64         unsigned long found;
  65
  66         found = find_first_zero_bit(pqm->queue_slot_bitmap,
  67                         KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
  68
  69         pr_debug("The new slot id %lu\n", found);
  70
  71         if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
  72                 pr_info("Cannot open more queues for process with pasid 0x%x\n",
  73                                 pqm->process->pasid);
  74                 return -ENOMEM;
  75         }
  76
  77         set_bit(found, pqm->queue_slot_bitmap);
  78         *qid = found;
  79
  80         return 0;
  81 }
  82
  83 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
  84 {
  85         struct kfd_node *dev = pdd->dev;
  86
  87         if (pdd->already_dequeued)
  88                 return;
  89
  90         dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
  91         if (dev->kfd->shared_resources.enable_mes &&
  92             down_read_trylock(&dev->adev->reset_domain->sem)) {
  93                 amdgpu_mes_flush_shader_debugger(dev->adev,
  94                                                  pdd->proc_ctx_gpu_addr);
  95                 up_read(&dev->adev->reset_domain->sem);
  96         }
  97         pdd->already_dequeued = true;
  98 }
  99
 100 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
 101                         void *gws)
 102 {
 103         struct mqd_update_info minfo = {0};
 104         struct kfd_node *dev = NULL;
 105         struct process_queue_node *pqn;
 106         struct kfd_process_device *pdd;
 107         struct kgd_mem *mem = NULL;
 108         int ret;
 109
 110         pqn = get_queue_by_qid(pqm, qid);
 111         if (!pqn) {
 112                 pr_err("Queue id does not match any known queue\n");
 113                 return -EINVAL;
 114         }
 115
 116         if (pqn->q)
 117                 dev = pqn->q->device;
 118         if (WARN_ON(!dev))
 119                 return -ENODEV;
 120
 121         pdd = kfd_get_process_device_data(dev, pqm->process);
 122         if (!pdd) {
 123                 pr_err("Process device data doesn't exist\n");
 124                 return -EINVAL;
 125         }
 126
 127         /* Only allow one queue per process can have GWS assigned */
 128         if (gws && pdd->qpd.num_gws)
 129                 return -EBUSY;
 130
 131         if (!gws && pdd->qpd.num_gws == 0)
 132                 return -EINVAL;
 133
 134         if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) &&
 135             KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4) &&
 136             !dev->kfd->shared_resources.enable_mes) {
 137                 if (gws)
 138                         ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
 139                                 gws, &mem);
 140                 else
 141                         ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
 142                                 pqn->q->gws);
 143                 if (unlikely(ret))
 144                         return ret;
 145                 pqn->q->gws = mem;
 146         } else {
 147                 /*
 148                  * Intentionally set GWS to a non-NULL value
 149                  * for devices that do not use GWS for global wave
 150                  * synchronization but require the formality
 151                  * of setting GWS for cooperative groups.
 152                  */
 153                 pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
 154         }
 155
 156         pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0;
 157         minfo.update_flag = gws ? UPDATE_FLAG_IS_GWS : 0;
 158
 159         return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 160                                                         pqn->q, &minfo);
 161 }
 162
 163 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
 164 {
 165         int i;
 166
 167         for (i = 0; i < p->n_pdds; i++)
 168                 kfd_process_dequeue_from_device(p->pdds[i]);
 169 }
 170
 171 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
 172 {
 173         INIT_LIST_HEAD(&pqm->queues);
 174         pqm->queue_slot_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
 175                                                GFP_KERNEL);
 176         if (!pqm->queue_slot_bitmap)
 177                 return -ENOMEM;
 178         pqm->process = p;
 179
 180         return 0;
 181 }
 182
 183 static void pqm_clean_queue_resource(struct process_queue_manager *pqm,
 184                                      struct process_queue_node *pqn)
 185 {
 186         struct kfd_node *dev;
 187         struct kfd_process_device *pdd;
 188
 189         dev = pqn->q->device;
 190
 191         pdd = kfd_get_process_device_data(dev, pqm->process);
 192         if (!pdd) {
 193                 pr_err("Process device data doesn't exist\n");
 194                 return;
 195         }
 196
 197         if (pqn->q->gws) {
 198                 if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
 199                     KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 4) &&
 200                     !dev->kfd->shared_resources.enable_mes)
 201                         amdgpu_amdkfd_remove_gws_from_process(
 202                                 pqm->process->kgd_process_info, pqn->q->gws);
 203                 pdd->qpd.num_gws = 0;
 204         }
 205
 206         if (dev->kfd->shared_resources.enable_mes) {
 207                 amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->gang_ctx_bo);
 208                 if (pqn->q->wptr_bo)
 209                         amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
 210         }
 211 }
 212
 213 void pqm_uninit(struct process_queue_manager *pqm)
 214 {
 215         struct process_queue_node *pqn, *next;
 216
 217         list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
 218                 if (pqn->q)
 219                         pqm_clean_queue_resource(pqm, pqn);
 220
 221                 kfd_procfs_del_queue(pqn->q);
 222                 uninit_queue(pqn->q);
 223                 list_del(&pqn->process_queue_list);
 224                 kfree(pqn);
 225         }
 226
 227         bitmap_free(pqm->queue_slot_bitmap);
 228         pqm->queue_slot_bitmap = NULL;
 229 }
 230
 231 static int init_user_queue(struct process_queue_manager *pqm,
 232                                 struct kfd_node *dev, struct queue **q,
 233                                 struct queue_properties *q_properties,
 234                                 struct file *f, struct amdgpu_bo *wptr_bo,
 235                                 unsigned int qid)
 236 {
 237         int retval;
 238
 239         /* Doorbell initialized in user space*/
 240         q_properties->doorbell_ptr = NULL;
 241         q_properties->exception_status = KFD_EC_MASK(EC_QUEUE_NEW);
 242
 243         /* let DQM handle it*/
 244         q_properties->vmid = 0;
 245         q_properties->queue_id = qid;
 246
 247         retval = init_queue(q, q_properties);
 248         if (retval != 0)
 249                 return retval;
 250
 251         (*q)->device = dev;
 252         (*q)->process = pqm->process;
 253
 254         if (dev->kfd->shared_resources.enable_mes) {
 255                 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
 256                                                 AMDGPU_MES_GANG_CTX_SIZE,
 257                                                 &(*q)->gang_ctx_bo,
 258                                                 &(*q)->gang_ctx_gpu_addr,
 259                                                 &(*q)->gang_ctx_cpu_ptr,
 260                                                 false);
 261                 if (retval) {
 262                         pr_err("failed to allocate gang context bo\n");
 263                         goto cleanup;
 264                 }
 265                 memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
 266                 (*q)->wptr_bo = wptr_bo;
 267         }
 268
 269         pr_debug("PQM After init queue");
 270         return 0;
 271
 272 cleanup:
 273         uninit_queue(*q);
 274         *q = NULL;
 275         return retval;
 276 }
 277
 278 int pqm_create_queue(struct process_queue_manager *pqm,
 279                             struct kfd_node *dev,
 280                             struct file *f,
 281                             struct queue_properties *properties,
 282                             unsigned int *qid,
 283                             struct amdgpu_bo *wptr_bo,
 284                             const struct kfd_criu_queue_priv_data *q_data,
 285                             const void *restore_mqd,
 286                             const void *restore_ctl_stack,
 287                             uint32_t *p_doorbell_offset_in_process)
 288 {
 289         int retval;
 290         struct kfd_process_device *pdd;
 291         struct queue *q;
 292         struct process_queue_node *pqn;
 293         struct kernel_queue *kq;
 294         enum kfd_queue_type type = properties->type;
 295         unsigned int max_queues = 127; /* HWS limit */
 296
 297         /*
 298          * On GFX 9.4.3, increase the number of queues that
 299          * can be created to 255. No HWS limit on GFX 9.4.3.
 300          */
 301         if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3) ||
 302             KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4))
 303                 max_queues = 255;
 304
 305         q = NULL;
 306         kq = NULL;
 307
 308         pdd = kfd_get_process_device_data(dev, pqm->process);
 309         if (!pdd) {
 310                 pr_err("Process device data doesn't exist\n");
 311                 return -1;
 312         }
 313
 314         /*
 315          * for debug process, verify that it is within the static queues limit
 316          * currently limit is set to half of the total avail HQD slots
 317          * If we are just about to create DIQ, the is_debug flag is not set yet
 318          * Hence we also check the type as well
 319          */
 320         if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
 321                 max_queues = dev->kfd->device_info.max_no_of_hqd/2;
 322
 323         if (pdd->qpd.queue_count >= max_queues)
 324                 return -ENOSPC;
 325
 326         if (q_data) {
 327                 retval = assign_queue_slot_by_qid(pqm, q_data->q_id);
 328                 *qid = q_data->q_id;
 329         } else
 330                 retval = find_available_queue_slot(pqm, qid);
 331
 332         if (retval != 0)
 333                 return retval;
 334
 335         if (list_empty(&pdd->qpd.queues_list) &&
 336             list_empty(&pdd->qpd.priv_queue_list))
 337                 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
 338
 339         pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
 340         if (!pqn) {
 341                 retval = -ENOMEM;
 342                 goto err_allocate_pqn;
 343         }
 344
 345         switch (type) {
 346         case KFD_QUEUE_TYPE_SDMA:
 347         case KFD_QUEUE_TYPE_SDMA_XGMI:
 348                 /* SDMA queues are always allocated statically no matter
 349                  * which scheduler mode is used. We also do not need to
 350                  * check whether a SDMA queue can be allocated here, because
 351                  * allocate_sdma_queue() in create_queue() has the
 352                  * corresponding check logic.
 353                  */
 354                 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 355                 if (retval != 0)
 356                         goto err_create_queue;
 357                 pqn->q = q;
 358                 pqn->kq = NULL;
 359                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
 360                                                     restore_mqd, restore_ctl_stack);
 361                 print_queue(q);
 362                 break;
 363
 364         case KFD_QUEUE_TYPE_COMPUTE:
 365                 /* check if there is over subscription */
 366                 if ((dev->dqm->sched_policy ==
 367                      KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
 368                 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
 369                 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) {
 370                         pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
 371                         retval = -EPERM;
 372                         goto err_create_queue;
 373                 }
 374
 375                 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
 376                 if (retval != 0)
 377                         goto err_create_queue;
 378                 pqn->q = q;
 379                 pqn->kq = NULL;
 380                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
 381                                                     restore_mqd, restore_ctl_stack);
 382                 print_queue(q);
 383                 break;
 384         case KFD_QUEUE_TYPE_DIQ:
 385                 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
 386                 if (!kq) {
 387                         retval = -ENOMEM;
 388                         goto err_create_queue;
 389                 }
 390                 kq->queue->properties.queue_id = *qid;
 391                 pqn->kq = kq;
 392                 pqn->q = NULL;
 393                 retval = kfd_process_drain_interrupts(pdd);
 394                 if (retval)
 395                         break;
 396
 397                 retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
 398                                                         kq, &pdd->qpd);
 399                 break;
 400         default:
 401                 WARN(1, "Invalid queue type %d", type);
 402                 retval = -EINVAL;
 403         }
 404
 405         if (retval != 0) {
 406                 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n",
 407                         pqm->process->pasid, type, retval);
 408                 goto err_create_queue;
 409         }
 410
 411         if (q && p_doorbell_offset_in_process) {
 412                 /* Return the doorbell offset within the doorbell page
 413                  * to the caller so it can be passed up to user mode
 414                  * (in bytes).
 415                  * relative doorbell index = Absolute doorbell index -
 416                  * absolute index of first doorbell in the page.
 417                  */
 418                 uint32_t first_db_index = amdgpu_doorbell_index_on_bar(pdd->dev->adev,
 419                                                                        pdd->qpd.proc_doorbells,
 420                                                                        0,
 421                                                                        pdd->dev->kfd->device_info.doorbell_size);
 422
 423                 *p_doorbell_offset_in_process = (q->properties.doorbell_off
 424                                                 - first_db_index) * sizeof(uint32_t);
 425         }
 426
 427         pr_debug("PQM After DQM create queue\n");
 428
 429         list_add(&pqn->process_queue_list, &pqm->queues);
 430
 431         if (q) {
 432                 pr_debug("PQM done creating queue\n");
 433                 kfd_procfs_add_queue(q);
 434                 print_queue_properties(&q->properties);
 435         }
 436
 437         return retval;
 438
 439 err_create_queue:
 440         uninit_queue(q);
 441         if (kq)
 442                 kernel_queue_uninit(kq);
 443         kfree(pqn);
 444 err_allocate_pqn:
 445         /* check if queues list is empty unregister process from device */
 446         clear_bit(*qid, pqm->queue_slot_bitmap);
 447         if (list_empty(&pdd->qpd.queues_list) &&
 448             list_empty(&pdd->qpd.priv_queue_list))
 449                 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
 450         return retval;
 451 }
 452
 453 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 454 {
 455         struct process_queue_node *pqn;
 456         struct kfd_process_device *pdd;
 457         struct device_queue_manager *dqm;
 458         struct kfd_node *dev;
 459         int retval;
 460
 461         dqm = NULL;
 462
 463         retval = 0;
 464
 465         pqn = get_queue_by_qid(pqm, qid);
 466         if (!pqn) {
 467                 pr_err("Queue id does not match any known queue\n");
 468                 return -EINVAL;
 469         }
 470
 471         dev = NULL;
 472         if (pqn->kq)
 473                 dev = pqn->kq->dev;
 474         if (pqn->q)
 475                 dev = pqn->q->device;
 476         if (WARN_ON(!dev))
 477                 return -ENODEV;
 478
 479         pdd = kfd_get_process_device_data(dev, pqm->process);
 480         if (!pdd) {
 481                 pr_err("Process device data doesn't exist\n");
 482                 return -1;
 483         }
 484
 485         if (pqn->kq) {
 486                 /* destroy kernel queue (DIQ) */
 487                 dqm = pqn->kq->dev->dqm;
 488                 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
 489                 kernel_queue_uninit(pqn->kq);
 490         }
 491
 492         if (pqn->q) {
 493                 kfd_procfs_del_queue(pqn->q);
 494                 dqm = pqn->q->device->dqm;
 495                 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
 496                 if (retval) {
 497                         pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
 498                                 pqm->process->pasid,
 499                                 pqn->q->properties.queue_id, retval);
 500                         if (retval != -ETIME)
 501                                 goto err_destroy_queue;
 502                 }
 503
 504                 pqm_clean_queue_resource(pqm, pqn);
 505                 uninit_queue(pqn->q);
 506         }
 507
 508         list_del(&pqn->process_queue_list);
 509         kfree(pqn);
 510         clear_bit(qid, pqm->queue_slot_bitmap);
 511
 512         if (list_empty(&pdd->qpd.queues_list) &&
 513             list_empty(&pdd->qpd.priv_queue_list))
 514                 dqm->ops.unregister_process(dqm, &pdd->qpd);
 515
 516 err_destroy_queue:
 517         return retval;
 518 }
 519
 520 int pqm_update_queue_properties(struct process_queue_manager *pqm,
 521                                 unsigned int qid, struct queue_properties *p)
 522 {
 523         int retval;
 524         struct process_queue_node *pqn;
 525
 526         pqn = get_queue_by_qid(pqm, qid);
 527         if (!pqn) {
 528                 pr_debug("No queue %d exists for update operation\n", qid);
 529                 return -EFAULT;
 530         }
 531
 532         pqn->q->properties.queue_address = p->queue_address;
 533         pqn->q->properties.queue_size = p->queue_size;
 534         pqn->q->properties.queue_percent = p->queue_percent;
 535         pqn->q->properties.priority = p->priority;
 536         pqn->q->properties.pm4_target_xcc = p->pm4_target_xcc;
 537
 538         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 539                                                         pqn->q, NULL);
 540         if (retval != 0)
 541                 return retval;
 542
 543         return 0;
 544 }
 545
 546 int pqm_update_mqd(struct process_queue_manager *pqm,
 547                                 unsigned int qid, struct mqd_update_info *minfo)
 548 {
 549         int retval;
 550         struct process_queue_node *pqn;
 551
 552         pqn = get_queue_by_qid(pqm, qid);
 553         if (!pqn) {
 554                 pr_debug("No queue %d exists for update operation\n", qid);
 555                 return -EFAULT;
 556         }
 557
 558         /* CUs are masked for debugger requirements so deny user mask  */
 559         if (pqn->q->properties.is_dbg_wa && minfo && minfo->cu_mask.ptr)
 560                 return -EBUSY;
 561
 562         /* ASICs that have WGPs must enforce pairwise enabled mask checks. */
 563         if (minfo && minfo->cu_mask.ptr &&
 564                         KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) {
 565                 int i;
 566
 567                 for (i = 0; i < minfo->cu_mask.count; i += 2) {
 568                         uint32_t cu_pair = (minfo->cu_mask.ptr[i / 32] >> (i % 32)) & 0x3;
 569
 570                         if (cu_pair && cu_pair != 0x3) {
 571                                 pr_debug("CUs must be adjacent pairwise enabled.\n");
 572                                 return -EINVAL;
 573                         }
 574                 }
 575         }
 576
 577         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 578                                                         pqn->q, minfo);
 579         if (retval != 0)
 580                 return retval;
 581
 582         if (minfo && minfo->cu_mask.ptr)
 583                 pqn->q->properties.is_user_cu_masked = true;
 584
 585         return 0;
 586 }
 587
 588 struct kernel_queue *pqm_get_kernel_queue(
 589                                         struct process_queue_manager *pqm,
 590                                         unsigned int qid)
 591 {
 592         struct process_queue_node *pqn;
 593
 594         pqn = get_queue_by_qid(pqm, qid);
 595         if (pqn && pqn->kq)
 596                 return pqn->kq;
 597
 598         return NULL;
 599 }
 600
 601 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
 602                                         unsigned int qid)
 603 {
 604         struct process_queue_node *pqn;
 605
 606         pqn = get_queue_by_qid(pqm, qid);
 607         return pqn ? pqn->q : NULL;
 608 }
 609
 610 int pqm_get_wave_state(struct process_queue_manager *pqm,
 611                        unsigned int qid,
 612                        void __user *ctl_stack,
 613                        u32 *ctl_stack_used_size,
 614                        u32 *save_area_used_size)
 615 {
 616         struct process_queue_node *pqn;
 617
 618         pqn = get_queue_by_qid(pqm, qid);
 619         if (!pqn) {
 620                 pr_debug("amdkfd: No queue %d exists for operation\n",
 621                          qid);
 622                 return -EFAULT;
 623         }
 624
 625         return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
 626                                                        pqn->q,
 627                                                        ctl_stack,
 628                                                        ctl_stack_used_size,
 629                                                        save_area_used_size);
 630 }
 631
 632 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
 633                            uint64_t exception_clear_mask,
 634                            void __user *buf,
 635                            int *num_qss_entries,
 636                            uint32_t *entry_size)
 637 {
 638         struct process_queue_node *pqn;
 639         struct kfd_queue_snapshot_entry src;
 640         uint32_t tmp_entry_size = *entry_size, tmp_qss_entries = *num_qss_entries;
 641         int r = 0;
 642
 643         *num_qss_entries = 0;
 644         if (!(*entry_size))
 645                 return -EINVAL;
 646
 647         *entry_size = min_t(size_t, *entry_size, sizeof(struct kfd_queue_snapshot_entry));
 648         mutex_lock(&pqm->process->event_mutex);
 649
 650         memset(&src, 0, sizeof(src));
 651
 652         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
 653                 if (!pqn->q)
 654                         continue;
 655
 656                 if (*num_qss_entries < tmp_qss_entries) {
 657                         set_queue_snapshot_entry(pqn->q, exception_clear_mask, &src);
 658
 659                         if (copy_to_user(buf, &src, *entry_size)) {
 660                                 r = -EFAULT;
 661                                 break;
 662                         }
 663                         buf += tmp_entry_size;
 664                 }
 665                 *num_qss_entries += 1;
 666         }
 667
 668         mutex_unlock(&pqm->process->event_mutex);
 669         return r;
 670 }
 671
 672 static int get_queue_data_sizes(struct kfd_process_device *pdd,
 673                                 struct queue *q,
 674                                 uint32_t *mqd_size,
 675                                 uint32_t *ctl_stack_size)
 676 {
 677         int ret;
 678
 679         ret = pqm_get_queue_checkpoint_info(&pdd->process->pqm,
 680                                             q->properties.queue_id,
 681                                             mqd_size,
 682                                             ctl_stack_size);
 683         if (ret)
 684                 pr_err("Failed to get queue dump info (%d)\n", ret);
 685
 686         return ret;
 687 }
 688
 689 int kfd_process_get_queue_info(struct kfd_process *p,
 690                                uint32_t *num_queues,
 691                                uint64_t *priv_data_sizes)
 692 {
 693         uint32_t extra_data_sizes = 0;
 694         struct queue *q;
 695         int i;
 696         int ret;
 697
 698         *num_queues = 0;
 699
 700         /* Run over all PDDs of the process */
 701         for (i = 0; i < p->n_pdds; i++) {
 702                 struct kfd_process_device *pdd = p->pdds[i];
 703
 704                 list_for_each_entry(q, &pdd->qpd.queues_list, list) {
 705                         if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 706                                 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 707                                 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 708                                 uint32_t mqd_size, ctl_stack_size;
 709
 710                                 *num_queues = *num_queues + 1;
 711
 712                                 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
 713                                 if (ret)
 714                                         return ret;
 715
 716                                 extra_data_sizes += mqd_size + ctl_stack_size;
 717                         } else {
 718                                 pr_err("Unsupported queue type (%d)\n", q->properties.type);
 719                                 return -EOPNOTSUPP;
 720                         }
 721                 }
 722         }
 723         *priv_data_sizes = extra_data_sizes +
 724                                 (*num_queues * sizeof(struct kfd_criu_queue_priv_data));
 725
 726         return 0;
 727 }
 728
 729 static int pqm_checkpoint_mqd(struct process_queue_manager *pqm,
 730                               unsigned int qid,
 731                               void *mqd,
 732                               void *ctl_stack)
 733 {
 734         struct process_queue_node *pqn;
 735
 736         pqn = get_queue_by_qid(pqm, qid);
 737         if (!pqn) {
 738                 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
 739                 return -EFAULT;
 740         }
 741
 742         if (!pqn->q->device->dqm->ops.checkpoint_mqd) {
 743                 pr_err("amdkfd: queue dumping not supported on this device\n");
 744                 return -EOPNOTSUPP;
 745         }
 746
 747         return pqn->q->device->dqm->ops.checkpoint_mqd(pqn->q->device->dqm,
 748                                                        pqn->q, mqd, ctl_stack);
 749 }
 750
 751 static int criu_checkpoint_queue(struct kfd_process_device *pdd,
 752                            struct queue *q,
 753                            struct kfd_criu_queue_priv_data *q_data)
 754 {
 755         uint8_t *mqd, *ctl_stack;
 756         int ret;
 757
 758         mqd = (void *)(q_data + 1);
 759         ctl_stack = mqd + q_data->mqd_size;
 760
 761         q_data->gpu_id = pdd->user_gpu_id;
 762         q_data->type = q->properties.type;
 763         q_data->format = q->properties.format;
 764         q_data->q_id =  q->properties.queue_id;
 765         q_data->q_address = q->properties.queue_address;
 766         q_data->q_size = q->properties.queue_size;
 767         q_data->priority = q->properties.priority;
 768         q_data->q_percent = q->properties.queue_percent;
 769         q_data->read_ptr_addr = (uint64_t)q->properties.read_ptr;
 770         q_data->write_ptr_addr = (uint64_t)q->properties.write_ptr;
 771         q_data->doorbell_id = q->doorbell_id;
 772
 773         q_data->sdma_id = q->sdma_id;
 774
 775         q_data->eop_ring_buffer_address =
 776                 q->properties.eop_ring_buffer_address;
 777
 778         q_data->eop_ring_buffer_size = q->properties.eop_ring_buffer_size;
 779
 780         q_data->ctx_save_restore_area_address =
 781                 q->properties.ctx_save_restore_area_address;
 782
 783         q_data->ctx_save_restore_area_size =
 784                 q->properties.ctx_save_restore_area_size;
 785
 786         q_data->gws = !!q->gws;
 787
 788         ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack);
 789         if (ret) {
 790                 pr_err("Failed checkpoint queue_mqd (%d)\n", ret);
 791                 return ret;
 792         }
 793
 794         pr_debug("Dumping Queue: gpu_id:%x queue_id:%u\n", q_data->gpu_id, q_data->q_id);
 795         return ret;
 796 }
 797
 798 static int criu_checkpoint_queues_device(struct kfd_process_device *pdd,
 799                                    uint8_t __user *user_priv,
 800                                    unsigned int *q_index,
 801                                    uint64_t *queues_priv_data_offset)
 802 {
 803         unsigned int q_private_data_size = 0;
 804         uint8_t *q_private_data = NULL; /* Local buffer to store individual queue private data */
 805         struct queue *q;
 806         int ret = 0;
 807
 808         list_for_each_entry(q, &pdd->qpd.queues_list, list) {
 809                 struct kfd_criu_queue_priv_data *q_data;
 810                 uint64_t q_data_size;
 811                 uint32_t mqd_size;
 812                 uint32_t ctl_stack_size;
 813
 814                 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE &&
 815                         q->properties.type != KFD_QUEUE_TYPE_SDMA &&
 816                         q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI) {
 817
 818                         pr_err("Unsupported queue type (%d)\n", q->properties.type);
 819                         ret = -EOPNOTSUPP;
 820                         break;
 821                 }
 822
 823                 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
 824                 if (ret)
 825                         break;
 826
 827                 q_data_size = sizeof(*q_data) + mqd_size + ctl_stack_size;
 828
 829                 /* Increase local buffer space if needed */
 830                 if (q_private_data_size < q_data_size) {
 831                         kfree(q_private_data);
 832
 833                         q_private_data = kzalloc(q_data_size, GFP_KERNEL);
 834                         if (!q_private_data) {
 835                                 ret = -ENOMEM;
 836                                 break;
 837                         }
 838                         q_private_data_size = q_data_size;
 839                 }
 840
 841                 q_data = (struct kfd_criu_queue_priv_data *)q_private_data;
 842
 843                 /* data stored in this order: priv_data, mqd, ctl_stack */
 844                 q_data->mqd_size = mqd_size;
 845                 q_data->ctl_stack_size = ctl_stack_size;
 846
 847                 ret = criu_checkpoint_queue(pdd, q, q_data);
 848                 if (ret)
 849                         break;
 850
 851                 q_data->object_type = KFD_CRIU_OBJECT_TYPE_QUEUE;
 852
 853                 ret = copy_to_user(user_priv + *queues_priv_data_offset,
 854                                 q_data, q_data_size);
 855                 if (ret) {
 856                         ret = -EFAULT;
 857                         break;
 858                 }
 859                 *queues_priv_data_offset += q_data_size;
 860                 *q_index = *q_index + 1;
 861         }
 862
 863         kfree(q_private_data);
 864
 865         return ret;
 866 }
 867
 868 int kfd_criu_checkpoint_queues(struct kfd_process *p,
 869                          uint8_t __user *user_priv_data,
 870                          uint64_t *priv_data_offset)
 871 {
 872         int ret = 0, pdd_index, q_index = 0;
 873
 874         for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
 875                 struct kfd_process_device *pdd = p->pdds[pdd_index];
 876
 877                 /*
 878                  * criu_checkpoint_queues_device will copy data to user and update q_index and
 879                  * queues_priv_data_offset
 880                  */
 881                 ret = criu_checkpoint_queues_device(pdd, user_priv_data, &q_index,
 882                                               priv_data_offset);
 883
 884                 if (ret)
 885                         break;
 886         }
 887
 888         return ret;
 889 }
 890
 891 static void set_queue_properties_from_criu(struct queue_properties *qp,
 892                                           struct kfd_criu_queue_priv_data *q_data)
 893 {
 894         qp->is_interop = false;
 895         qp->queue_percent = q_data->q_percent;
 896         qp->priority = q_data->priority;
 897         qp->queue_address = q_data->q_address;
 898         qp->queue_size = q_data->q_size;
 899         qp->read_ptr = (uint32_t *) q_data->read_ptr_addr;
 900         qp->write_ptr = (uint32_t *) q_data->write_ptr_addr;
 901         qp->eop_ring_buffer_address = q_data->eop_ring_buffer_address;
 902         qp->eop_ring_buffer_size = q_data->eop_ring_buffer_size;
 903         qp->ctx_save_restore_area_address = q_data->ctx_save_restore_area_address;
 904         qp->ctx_save_restore_area_size = q_data->ctx_save_restore_area_size;
 905         qp->ctl_stack_size = q_data->ctl_stack_size;
 906         qp->type = q_data->type;
 907         qp->format = q_data->format;
 908 }
 909
 910 int kfd_criu_restore_queue(struct kfd_process *p,
 911                            uint8_t __user *user_priv_ptr,
 912                            uint64_t *priv_data_offset,
 913                            uint64_t max_priv_data_size)
 914 {
 915         uint8_t *mqd, *ctl_stack, *q_extra_data = NULL;
 916         struct kfd_criu_queue_priv_data *q_data;
 917         struct kfd_process_device *pdd;
 918         uint64_t q_extra_data_size;
 919         struct queue_properties qp;
 920         unsigned int queue_id;
 921         int ret = 0;
 922
 923         if (*priv_data_offset + sizeof(*q_data) > max_priv_data_size)
 924                 return -EINVAL;
 925
 926         q_data = kmalloc(sizeof(*q_data), GFP_KERNEL);
 927         if (!q_data)
 928                 return -ENOMEM;
 929
 930         ret = copy_from_user(q_data, user_priv_ptr + *priv_data_offset, sizeof(*q_data));
 931         if (ret) {
 932                 ret = -EFAULT;
 933                 goto exit;
 934         }
 935
 936         *priv_data_offset += sizeof(*q_data);
 937         q_extra_data_size = (uint64_t)q_data->ctl_stack_size + q_data->mqd_size;
 938
 939         if (*priv_data_offset + q_extra_data_size > max_priv_data_size) {
 940                 ret = -EINVAL;
 941                 goto exit;
 942         }
 943
 944         q_extra_data = kmalloc(q_extra_data_size, GFP_KERNEL);
 945         if (!q_extra_data) {
 946                 ret = -ENOMEM;
 947                 goto exit;
 948         }
 949
 950         ret = copy_from_user(q_extra_data, user_priv_ptr + *priv_data_offset, q_extra_data_size);
 951         if (ret) {
 952                 ret = -EFAULT;
 953                 goto exit;
 954         }
 955
 956         *priv_data_offset += q_extra_data_size;
 957
 958         pdd = kfd_process_device_data_by_id(p, q_data->gpu_id);
 959         if (!pdd) {
 960                 pr_err("Failed to get pdd\n");
 961                 ret = -EINVAL;
 962                 goto exit;
 963         }
 964
 965         /* data stored in this order: mqd, ctl_stack */
 966         mqd = q_extra_data;
 967         ctl_stack = mqd + q_data->mqd_size;
 968
 969         memset(&qp, 0, sizeof(qp));
 970         set_queue_properties_from_criu(&qp, q_data);
 971
 972         print_queue_properties(&qp);
 973
 974         ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
 975                                 NULL);
 976         if (ret) {
 977                 pr_err("Failed to create new queue err:%d\n", ret);
 978                 goto exit;
 979         }
 980
 981         if (q_data->gws)
 982                 ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws);
 983
 984 exit:
 985         if (ret)
 986                 pr_err("Failed to restore queue (%d)\n", ret);
 987         else
 988                 pr_debug("Queue id %d was restored successfully\n", queue_id);
 989
 990         kfree(q_data);
 991
 992         return ret;
 993 }
 994
 995 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
 996                                   unsigned int qid,
 997                                   uint32_t *mqd_size,
 998                                   uint32_t *ctl_stack_size)
 999 {
1000         struct process_queue_node *pqn;
1001
1002         pqn = get_queue_by_qid(pqm, qid);
1003         if (!pqn) {
1004                 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
1005                 return -EFAULT;
1006         }
1007
1008         if (!pqn->q->device->dqm->ops.get_queue_checkpoint_info) {
1009                 pr_err("amdkfd: queue dumping not supported on this device\n");
1010                 return -EOPNOTSUPP;
1011         }
1012
1013         pqn->q->device->dqm->ops.get_queue_checkpoint_info(pqn->q->device->dqm,
1014                                                        pqn->q, mqd_size,
1015                                                        ctl_stack_size);
1016         return 0;
1017 }
1018
1019 #if defined(CONFIG_DEBUG_FS)
1020
1021 int pqm_debugfs_mqds(struct seq_file *m, void *data)
1022 {
1023         struct process_queue_manager *pqm = data;
1024         struct process_queue_node *pqn;
1025         struct queue *q;
1026         enum KFD_MQD_TYPE mqd_type;
1027         struct mqd_manager *mqd_mgr;
1028         int r = 0, xcc, num_xccs = 1;
1029         void *mqd;
1030         uint64_t size = 0;
1031
1032         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1033                 if (pqn->q) {
1034                         q = pqn->q;
1035                         switch (q->properties.type) {
1036                         case KFD_QUEUE_TYPE_SDMA:
1037                         case KFD_QUEUE_TYPE_SDMA_XGMI:
1038                                 seq_printf(m, "  SDMA queue on device %x\n",
1039                                            q->device->id);
1040                                 mqd_type = KFD_MQD_TYPE_SDMA;
1041                                 break;
1042                         case KFD_QUEUE_TYPE_COMPUTE:
1043                                 seq_printf(m, "  Compute queue on device %x\n",
1044                                            q->device->id);
1045                                 mqd_type = KFD_MQD_TYPE_CP;
1046                                 num_xccs = NUM_XCC(q->device->xcc_mask);
1047                                 break;
1048                         default:
1049                                 seq_printf(m,
1050                                 "  Bad user queue type %d on device %x\n",
1051                                            q->properties.type, q->device->id);
1052                                 continue;
1053                         }
1054                         mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
1055                         size = mqd_mgr->mqd_stride(mqd_mgr,
1056                                                         &q->properties);
1057                 } else if (pqn->kq) {
1058                         q = pqn->kq->queue;
1059                         mqd_mgr = pqn->kq->mqd_mgr;
1060                         switch (q->properties.type) {
1061                         case KFD_QUEUE_TYPE_DIQ:
1062                                 seq_printf(m, "  DIQ on device %x\n",
1063                                            pqn->kq->dev->id);
1064                                 break;
1065                         default:
1066                                 seq_printf(m,
1067                                 "  Bad kernel queue type %d on device %x\n",
1068                                            q->properties.type,
1069                                            pqn->kq->dev->id);
1070                                 continue;
1071                         }
1072                 } else {
1073                         seq_printf(m,
1074                 "  Weird: Queue node with neither kernel nor user queue\n");
1075                         continue;
1076                 }
1077
1078                 for (xcc = 0; xcc < num_xccs; xcc++) {
1079                         mqd = q->mqd + size * xcc;
1080                         r = mqd_mgr->debugfs_show_mqd(m, mqd);
1081                         if (r != 0)
1082                                 break;
1083                 }
1084         }
1085
1086         return r;
1087 }
1088
1089 #endif