drivers/misc/habanalabs/command_submission.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Copyright 2016-2019 HabanaLabs, Ltd.
   5  * All Rights Reserved.
   6  */
   7
   8 #include <uapi/misc/habanalabs.h>
   9 #include "habanalabs.h"
  10
  11 #include <linux/uaccess.h>
  12 #include <linux/slab.h>
  13
  14 static void job_wq_completion(struct work_struct *work);
  15 static long _hl_cs_wait_ioctl(struct hl_device *hdev,
  16                 struct hl_ctx *ctx, u64 timeout_us, u64 seq);
  17 static void cs_do_release(struct kref *ref);
  18
  19 static const char *hl_fence_get_driver_name(struct dma_fence *fence)
  20 {
  21         return "HabanaLabs";
  22 }
  23
  24 static const char *hl_fence_get_timeline_name(struct dma_fence *fence)
  25 {
  26         struct hl_dma_fence *hl_fence =
  27                 container_of(fence, struct hl_dma_fence, base_fence);
  28
  29         return dev_name(hl_fence->hdev->dev);
  30 }
  31
  32 static bool hl_fence_enable_signaling(struct dma_fence *fence)
  33 {
  34         return true;
  35 }
  36
  37 static void hl_fence_release(struct dma_fence *fence)
  38 {
  39         struct hl_dma_fence *hl_fence =
  40                 container_of(fence, struct hl_dma_fence, base_fence);
  41
  42         kfree_rcu(hl_fence, base_fence.rcu);
  43 }
  44
  45 static const struct dma_fence_ops hl_fence_ops = {
  46         .get_driver_name = hl_fence_get_driver_name,
  47         .get_timeline_name = hl_fence_get_timeline_name,
  48         .enable_signaling = hl_fence_enable_signaling,
  49         .wait = dma_fence_default_wait,
  50         .release = hl_fence_release
  51 };
  52
  53 static void cs_get(struct hl_cs *cs)
  54 {
  55         kref_get(&cs->refcount);
  56 }
  57
  58 static int cs_get_unless_zero(struct hl_cs *cs)
  59 {
  60         return kref_get_unless_zero(&cs->refcount);
  61 }
  62
  63 static void cs_put(struct hl_cs *cs)
  64 {
  65         kref_put(&cs->refcount, cs_do_release);
  66 }
  67
  68 static bool is_cb_patched(struct hl_device *hdev, struct hl_cs_job *job)
  69 {
  70         /*
  71          * Patched CB is created for external queues jobs, and for H/W queues
  72          * jobs if the user CB was allocated by driver and MMU is disabled.
  73          */
  74         return (job->queue_type == QUEUE_TYPE_EXT ||
  75                         (job->queue_type == QUEUE_TYPE_HW &&
  76                                         job->is_kernel_allocated_cb &&
  77                                         !hdev->mmu_enable));
  78 }
  79
  80 /*
  81  * cs_parser - parse the user command submission
  82  *
  83  * @hpriv       : pointer to the private data of the fd
  84  * @job        : pointer to the job that holds the command submission info
  85  *
  86  * The function parses the command submission of the user. It calls the
  87  * ASIC specific parser, which returns a list of memory blocks to send
  88  * to the device as different command buffers
  89  *
  90  */
  91 static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
  92 {
  93         struct hl_device *hdev = hpriv->hdev;
  94         struct hl_cs_parser parser;
  95         int rc;
  96
  97         parser.ctx_id = job->cs->ctx->asid;
  98         parser.cs_sequence = job->cs->sequence;
  99         parser.job_id = job->id;
 100
 101         parser.hw_queue_id = job->hw_queue_id;
 102         parser.job_userptr_list = &job->userptr_list;
 103         parser.patched_cb = NULL;
 104         parser.user_cb = job->user_cb;
 105         parser.user_cb_size = job->user_cb_size;
 106         parser.queue_type = job->queue_type;
 107         parser.is_kernel_allocated_cb = job->is_kernel_allocated_cb;
 108         job->patched_cb = NULL;
 109
 110         rc = hdev->asic_funcs->cs_parser(hdev, &parser);
 111
 112         if (is_cb_patched(hdev, job)) {
 113                 if (!rc) {
 114                         job->patched_cb = parser.patched_cb;
 115                         job->job_cb_size = parser.patched_cb_size;
 116
 117                         spin_lock(&job->patched_cb->lock);
 118                         job->patched_cb->cs_cnt++;
 119                         spin_unlock(&job->patched_cb->lock);
 120                 }
 121
 122                 /*
 123                  * Whether the parsing worked or not, we don't need the
 124                  * original CB anymore because it was already parsed and
 125                  * won't be accessed again for this CS
 126                  */
 127                 spin_lock(&job->user_cb->lock);
 128                 job->user_cb->cs_cnt--;
 129                 spin_unlock(&job->user_cb->lock);
 130                 hl_cb_put(job->user_cb);
 131                 job->user_cb = NULL;
 132         } else if (!rc) {
 133                 job->job_cb_size = job->user_cb_size;
 134         }
 135
 136         return rc;
 137 }
 138
 139 static void free_job(struct hl_device *hdev, struct hl_cs_job *job)
 140 {
 141         struct hl_cs *cs = job->cs;
 142
 143         if (is_cb_patched(hdev, job)) {
 144                 hl_userptr_delete_list(hdev, &job->userptr_list);
 145
 146                 /*
 147                  * We might arrive here from rollback and patched CB wasn't
 148                  * created, so we need to check it's not NULL
 149                  */
 150                 if (job->patched_cb) {
 151                         spin_lock(&job->patched_cb->lock);
 152                         job->patched_cb->cs_cnt--;
 153                         spin_unlock(&job->patched_cb->lock);
 154
 155                         hl_cb_put(job->patched_cb);
 156                 }
 157         }
 158
 159         /* For H/W queue jobs, if a user CB was allocated by driver and MMU is
 160          * enabled, the user CB isn't released in cs_parser() and thus should be
 161          * released here.
 162          */
 163         if (job->queue_type == QUEUE_TYPE_HW &&
 164                         job->is_kernel_allocated_cb && hdev->mmu_enable) {
 165                 spin_lock(&job->user_cb->lock);
 166                 job->user_cb->cs_cnt--;
 167                 spin_unlock(&job->user_cb->lock);
 168
 169                 hl_cb_put(job->user_cb);
 170         }
 171
 172         /*
 173          * This is the only place where there can be multiple threads
 174          * modifying the list at the same time
 175          */
 176         spin_lock(&cs->job_lock);
 177         list_del(&job->cs_node);
 178         spin_unlock(&cs->job_lock);
 179
 180         hl_debugfs_remove_job(hdev, job);
 181
 182         if (job->queue_type == QUEUE_TYPE_EXT ||
 183                         job->queue_type == QUEUE_TYPE_HW)
 184                 cs_put(cs);
 185
 186         kfree(job);
 187 }
 188
 189 static void cs_do_release(struct kref *ref)
 190 {
 191         struct hl_cs *cs = container_of(ref, struct hl_cs,
 192                                                 refcount);
 193         struct hl_device *hdev = cs->ctx->hdev;
 194         struct hl_cs_job *job, *tmp;
 195
 196         cs->completed = true;
 197
 198         /*
 199          * Although if we reached here it means that all external jobs have
 200          * finished, because each one of them took refcnt to CS, we still
 201          * need to go over the internal jobs and free them. Otherwise, we
 202          * will have leaked memory and what's worse, the CS object (and
 203          * potentially the CTX object) could be released, while the JOB
 204          * still holds a pointer to them (but no reference).
 205          */
 206         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 207                 free_job(hdev, job);
 208
 209         /* We also need to update CI for internal queues */
 210         if (cs->submitted) {
 211                 hdev->asic_funcs->hw_queues_lock(hdev);
 212
 213                 hdev->cs_active_cnt--;
 214                 if (!hdev->cs_active_cnt) {
 215                         struct hl_device_idle_busy_ts *ts;
 216
 217                         ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx++];
 218                         ts->busy_to_idle_ts = ktime_get();
 219
 220                         if (hdev->idle_busy_ts_idx == HL_IDLE_BUSY_TS_ARR_SIZE)
 221                                 hdev->idle_busy_ts_idx = 0;
 222                 } else if (hdev->cs_active_cnt < 0) {
 223                         dev_crit(hdev->dev, "CS active cnt %d is negative\n",
 224                                 hdev->cs_active_cnt);
 225                 }
 226
 227                 hdev->asic_funcs->hw_queues_unlock(hdev);
 228
 229                 hl_int_hw_queue_update_ci(cs);
 230
 231                 spin_lock(&hdev->hw_queues_mirror_lock);
 232                 /* remove CS from hw_queues mirror list */
 233                 list_del_init(&cs->mirror_node);
 234                 spin_unlock(&hdev->hw_queues_mirror_lock);
 235
 236                 /*
 237                  * Don't cancel TDR in case this CS was timedout because we
 238                  * might be running from the TDR context
 239                  */
 240                 if ((!cs->timedout) &&
 241                         (hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT)) {
 242                         struct hl_cs *next;
 243
 244                         if (cs->tdr_active)
 245                                 cancel_delayed_work_sync(&cs->work_tdr);
 246
 247                         spin_lock(&hdev->hw_queues_mirror_lock);
 248
 249                         /* queue TDR for next CS */
 250                         next = list_first_entry_or_null(
 251                                         &hdev->hw_queues_mirror_list,
 252                                         struct hl_cs, mirror_node);
 253
 254                         if ((next) && (!next->tdr_active)) {
 255                                 next->tdr_active = true;
 256                                 schedule_delayed_work(&next->work_tdr,
 257                                                         hdev->timeout_jiffies);
 258                         }
 259
 260                         spin_unlock(&hdev->hw_queues_mirror_lock);
 261                 }
 262         }
 263
 264         /*
 265          * Must be called before hl_ctx_put because inside we use ctx to get
 266          * the device
 267          */
 268         hl_debugfs_remove_cs(cs);
 269
 270         hl_ctx_put(cs->ctx);
 271
 272         if (cs->timedout)
 273                 dma_fence_set_error(cs->fence, -ETIMEDOUT);
 274         else if (cs->aborted)
 275                 dma_fence_set_error(cs->fence, -EIO);
 276
 277         dma_fence_signal(cs->fence);
 278         dma_fence_put(cs->fence);
 279
 280         kfree(cs);
 281 }
 282
 283 static void cs_timedout(struct work_struct *work)
 284 {
 285         struct hl_device *hdev;
 286         int ctx_asid, rc;
 287         struct hl_cs *cs = container_of(work, struct hl_cs,
 288                                                  work_tdr.work);
 289         rc = cs_get_unless_zero(cs);
 290         if (!rc)
 291                 return;
 292
 293         if ((!cs->submitted) || (cs->completed)) {
 294                 cs_put(cs);
 295                 return;
 296         }
 297
 298         /* Mark the CS is timed out so we won't try to cancel its TDR */
 299         cs->timedout = true;
 300
 301         hdev = cs->ctx->hdev;
 302         ctx_asid = cs->ctx->asid;
 303
 304         /* TODO: add information about last signaled seq and last emitted seq */
 305         dev_err(hdev->dev, "User %d command submission %llu got stuck!\n",
 306                 ctx_asid, cs->sequence);
 307
 308         cs_put(cs);
 309
 310         if (hdev->reset_on_lockup)
 311                 hl_device_reset(hdev, false, false);
 312 }
 313
 314 static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 315                         struct hl_cs **cs_new)
 316 {
 317         struct hl_dma_fence *fence;
 318         struct dma_fence *other = NULL;
 319         struct hl_cs *cs;
 320         int rc;
 321
 322         cs = kzalloc(sizeof(*cs), GFP_ATOMIC);
 323         if (!cs)
 324                 return -ENOMEM;
 325
 326         cs->ctx = ctx;
 327         cs->submitted = false;
 328         cs->completed = false;
 329         INIT_LIST_HEAD(&cs->job_list);
 330         INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
 331         kref_init(&cs->refcount);
 332         spin_lock_init(&cs->job_lock);
 333
 334         fence = kmalloc(sizeof(*fence), GFP_ATOMIC);
 335         if (!fence) {
 336                 rc = -ENOMEM;
 337                 goto free_cs;
 338         }
 339
 340         fence->hdev = hdev;
 341         spin_lock_init(&fence->lock);
 342         cs->fence = &fence->base_fence;
 343
 344         spin_lock(&ctx->cs_lock);
 345
 346         fence->cs_seq = ctx->cs_sequence;
 347         other = ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)];
 348         if ((other) && (!dma_fence_is_signaled(other))) {
 349                 spin_unlock(&ctx->cs_lock);
 350                 dev_dbg(hdev->dev,
 351                         "Rejecting CS because of too many in-flights CS\n");
 352                 rc = -EAGAIN;
 353                 goto free_fence;
 354         }
 355
 356         dma_fence_init(&fence->base_fence, &hl_fence_ops, &fence->lock,
 357                         ctx->asid, ctx->cs_sequence);
 358
 359         cs->sequence = fence->cs_seq;
 360
 361         ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)] =
 362                                                         &fence->base_fence;
 363         ctx->cs_sequence++;
 364
 365         dma_fence_get(&fence->base_fence);
 366
 367         dma_fence_put(other);
 368
 369         spin_unlock(&ctx->cs_lock);
 370
 371         *cs_new = cs;
 372
 373         return 0;
 374
 375 free_fence:
 376         kfree(fence);
 377 free_cs:
 378         kfree(cs);
 379         return rc;
 380 }
 381
 382 static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)
 383 {
 384         struct hl_cs_job *job, *tmp;
 385
 386         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 387                 free_job(hdev, job);
 388 }
 389
 390 void hl_cs_rollback_all(struct hl_device *hdev)
 391 {
 392         struct hl_cs *cs, *tmp;
 393
 394         /* flush all completions */
 395         flush_workqueue(hdev->cq_wq);
 396
 397         /* Make sure we don't have leftovers in the H/W queues mirror list */
 398         list_for_each_entry_safe(cs, tmp, &hdev->hw_queues_mirror_list,
 399                                 mirror_node) {
 400                 cs_get(cs);
 401                 cs->aborted = true;
 402                 dev_warn_ratelimited(hdev->dev, "Killing CS %d.%llu\n",
 403                                         cs->ctx->asid, cs->sequence);
 404                 cs_rollback(hdev, cs);
 405                 cs_put(cs);
 406         }
 407 }
 408
 409 static void job_wq_completion(struct work_struct *work)
 410 {
 411         struct hl_cs_job *job = container_of(work, struct hl_cs_job,
 412                                                 finish_work);
 413         struct hl_cs *cs = job->cs;
 414         struct hl_device *hdev = cs->ctx->hdev;
 415
 416         /* job is no longer needed */
 417         free_job(hdev, job);
 418 }
 419
 420 static int validate_queue_index(struct hl_device *hdev,
 421                                 struct hl_cs_chunk *chunk,
 422                                 enum hl_queue_type *queue_type,
 423                                 bool *is_kernel_allocated_cb)
 424 {
 425         struct asic_fixed_properties *asic = &hdev->asic_prop;
 426         struct hw_queue_properties *hw_queue_prop;
 427
 428         hw_queue_prop = &asic->hw_queues_props[chunk->queue_index];
 429
 430         if ((chunk->queue_index >= HL_MAX_QUEUES) ||
 431                         (hw_queue_prop->type == QUEUE_TYPE_NA)) {
 432                 dev_err(hdev->dev, "Queue index %d is invalid\n",
 433                         chunk->queue_index);
 434                 return -EINVAL;
 435         }
 436
 437         if (hw_queue_prop->driver_only) {
 438                 dev_err(hdev->dev,
 439                         "Queue index %d is restricted for the kernel driver\n",
 440                         chunk->queue_index);
 441                 return -EINVAL;
 442         }
 443
 444         *queue_type = hw_queue_prop->type;
 445         *is_kernel_allocated_cb = !!hw_queue_prop->requires_kernel_cb;
 446
 447         return 0;
 448 }
 449
 450 static struct hl_cb *get_cb_from_cs_chunk(struct hl_device *hdev,
 451                                         struct hl_cb_mgr *cb_mgr,
 452                                         struct hl_cs_chunk *chunk)
 453 {
 454         struct hl_cb *cb;
 455         u32 cb_handle;
 456
 457         cb_handle = (u32) (chunk->cb_handle >> PAGE_SHIFT);
 458
 459         cb = hl_cb_get(hdev, cb_mgr, cb_handle);
 460         if (!cb) {
 461                 dev_err(hdev->dev, "CB handle 0x%x invalid\n", cb_handle);
 462                 return NULL;
 463         }
 464
 465         if ((chunk->cb_size < 8) || (chunk->cb_size > cb->size)) {
 466                 dev_err(hdev->dev, "CB size %u invalid\n", chunk->cb_size);
 467                 goto release_cb;
 468         }
 469
 470         spin_lock(&cb->lock);
 471         cb->cs_cnt++;
 472         spin_unlock(&cb->lock);
 473
 474         return cb;
 475
 476 release_cb:
 477         hl_cb_put(cb);
 478         return NULL;
 479 }
 480
 481 struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
 482                 enum hl_queue_type queue_type, bool is_kernel_allocated_cb)
 483 {
 484         struct hl_cs_job *job;
 485
 486         job = kzalloc(sizeof(*job), GFP_ATOMIC);
 487         if (!job)
 488                 return NULL;
 489
 490         job->queue_type = queue_type;
 491         job->is_kernel_allocated_cb = is_kernel_allocated_cb;
 492
 493         if (is_cb_patched(hdev, job))
 494                 INIT_LIST_HEAD(&job->userptr_list);
 495
 496         if (job->queue_type == QUEUE_TYPE_EXT)
 497                 INIT_WORK(&job->finish_work, job_wq_completion);
 498
 499         return job;
 500 }
 501
 502 static int _hl_cs_ioctl(struct hl_fpriv *hpriv, void __user *chunks,
 503                         u32 num_chunks, u64 *cs_seq)
 504 {
 505         struct hl_device *hdev = hpriv->hdev;
 506         struct hl_cs_chunk *cs_chunk_array;
 507         struct hl_cs_job *job;
 508         struct hl_cs *cs;
 509         struct hl_cb *cb;
 510         bool int_queues_only = true;
 511         u32 size_to_copy;
 512         int rc, i;
 513
 514         *cs_seq = ULLONG_MAX;
 515
 516         if (num_chunks > HL_MAX_JOBS_PER_CS) {
 517                 dev_err(hdev->dev,
 518                         "Number of chunks can NOT be larger than %d\n",
 519                         HL_MAX_JOBS_PER_CS);
 520                 rc = -EINVAL;
 521                 goto out;
 522         }
 523
 524         cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
 525                                         GFP_ATOMIC);
 526         if (!cs_chunk_array) {
 527                 rc = -ENOMEM;
 528                 goto out;
 529         }
 530
 531         size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
 532         if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
 533                 dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
 534                 rc = -EFAULT;
 535                 goto free_cs_chunk_array;
 536         }
 537
 538         /* increment refcnt for context */
 539         hl_ctx_get(hdev, hpriv->ctx);
 540
 541         rc = allocate_cs(hdev, hpriv->ctx, &cs);
 542         if (rc) {
 543                 hl_ctx_put(hpriv->ctx);
 544                 goto free_cs_chunk_array;
 545         }
 546
 547         *cs_seq = cs->sequence;
 548
 549         hl_debugfs_add_cs(cs);
 550
 551         /* Validate ALL the CS chunks before submitting the CS */
 552         for (i = 0 ; i < num_chunks ; i++) {
 553                 struct hl_cs_chunk *chunk = &cs_chunk_array[i];
 554                 enum hl_queue_type queue_type;
 555                 bool is_kernel_allocated_cb;
 556
 557                 rc = validate_queue_index(hdev, chunk, &queue_type,
 558                                                 &is_kernel_allocated_cb);
 559                 if (rc)
 560                         goto free_cs_object;
 561
 562                 if (is_kernel_allocated_cb) {
 563                         cb = get_cb_from_cs_chunk(hdev, &hpriv->cb_mgr, chunk);
 564                         if (!cb) {
 565                                 rc = -EINVAL;
 566                                 goto free_cs_object;
 567                         }
 568                 } else {
 569                         cb = (struct hl_cb *) (uintptr_t) chunk->cb_handle;
 570                 }
 571
 572                 if (queue_type == QUEUE_TYPE_EXT || queue_type == QUEUE_TYPE_HW)
 573                         int_queues_only = false;
 574
 575                 job = hl_cs_allocate_job(hdev, queue_type,
 576                                                 is_kernel_allocated_cb);
 577                 if (!job) {
 578                         dev_err(hdev->dev, "Failed to allocate a new job\n");
 579                         rc = -ENOMEM;
 580                         if (is_kernel_allocated_cb)
 581                                 goto release_cb;
 582                         else
 583                                 goto free_cs_object;
 584                 }
 585
 586                 job->id = i + 1;
 587                 job->cs = cs;
 588                 job->user_cb = cb;
 589                 job->user_cb_size = chunk->cb_size;
 590                 job->hw_queue_id = chunk->queue_index;
 591
 592                 cs->jobs_in_queue_cnt[job->hw_queue_id]++;
 593
 594                 list_add_tail(&job->cs_node, &cs->job_list);
 595
 596                 /*
 597                  * Increment CS reference. When CS reference is 0, CS is
 598                  * done and can be signaled to user and free all its resources
 599                  * Only increment for JOB on external or H/W queues, because
 600                  * only for those JOBs we get completion
 601                  */
 602                 if (job->queue_type == QUEUE_TYPE_EXT ||
 603                                 job->queue_type == QUEUE_TYPE_HW)
 604                         cs_get(cs);
 605
 606                 hl_debugfs_add_job(hdev, job);
 607
 608                 rc = cs_parser(hpriv, job);
 609                 if (rc) {
 610                         dev_err(hdev->dev,
 611                                 "Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
 612                                 cs->ctx->asid, cs->sequence, job->id, rc);
 613                         goto free_cs_object;
 614                 }
 615         }
 616
 617         if (int_queues_only) {
 618                 dev_err(hdev->dev,
 619                         "Reject CS %d.%llu because only internal queues jobs are present\n",
 620                         cs->ctx->asid, cs->sequence);
 621                 rc = -EINVAL;
 622                 goto free_cs_object;
 623         }
 624
 625         rc = hl_hw_queue_schedule_cs(cs);
 626         if (rc) {
 627                 if (rc != -EAGAIN)
 628                         dev_err(hdev->dev,
 629                                 "Failed to submit CS %d.%llu to H/W queues, error %d\n",
 630                                 cs->ctx->asid, cs->sequence, rc);
 631                 goto free_cs_object;
 632         }
 633
 634         rc = HL_CS_STATUS_SUCCESS;
 635         goto put_cs;
 636
 637 release_cb:
 638         spin_lock(&cb->lock);
 639         cb->cs_cnt--;
 640         spin_unlock(&cb->lock);
 641         hl_cb_put(cb);
 642 free_cs_object:
 643         cs_rollback(hdev, cs);
 644         *cs_seq = ULLONG_MAX;
 645         /* The path below is both for good and erroneous exits */
 646 put_cs:
 647         /* We finished with the CS in this function, so put the ref */
 648         cs_put(cs);
 649 free_cs_chunk_array:
 650         kfree(cs_chunk_array);
 651 out:
 652         return rc;
 653 }
 654
 655 int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 656 {
 657         struct hl_device *hdev = hpriv->hdev;
 658         union hl_cs_args *args = data;
 659         struct hl_ctx *ctx = hpriv->ctx;
 660         void __user *chunks_execute, *chunks_restore;
 661         u32 num_chunks_execute, num_chunks_restore;
 662         u64 cs_seq = ULONG_MAX;
 663         int rc, do_ctx_switch;
 664         bool need_soft_reset = false;
 665
 666         if (hl_device_disabled_or_in_reset(hdev)) {
 667                 dev_warn_ratelimited(hdev->dev,
 668                         "Device is %s. Can't submit new CS\n",
 669                         atomic_read(&hdev->in_reset) ? "in_reset" : "disabled");
 670                 rc = -EBUSY;
 671                 goto out;
 672         }
 673
 674         chunks_execute = (void __user *) (uintptr_t) args->in.chunks_execute;
 675         num_chunks_execute = args->in.num_chunks_execute;
 676
 677         if (!num_chunks_execute) {
 678                 dev_err(hdev->dev,
 679                         "Got execute CS with 0 chunks, context %d\n",
 680                         ctx->asid);
 681                 rc = -EINVAL;
 682                 goto out;
 683         }
 684
 685         do_ctx_switch = atomic_cmpxchg(&ctx->thread_ctx_switch_token, 1, 0);
 686
 687         if (do_ctx_switch || (args->in.cs_flags & HL_CS_FLAGS_FORCE_RESTORE)) {
 688                 long ret;
 689
 690                 chunks_restore =
 691                         (void __user *) (uintptr_t) args->in.chunks_restore;
 692                 num_chunks_restore = args->in.num_chunks_restore;
 693
 694                 mutex_lock(&hpriv->restore_phase_mutex);
 695
 696                 if (do_ctx_switch) {
 697                         rc = hdev->asic_funcs->context_switch(hdev, ctx->asid);
 698                         if (rc) {
 699                                 dev_err_ratelimited(hdev->dev,
 700                                         "Failed to switch to context %d, rejecting CS! %d\n",
 701                                         ctx->asid, rc);
 702                                 /*
 703                                  * If we timedout, or if the device is not IDLE
 704                                  * while we want to do context-switch (-EBUSY),
 705                                  * we need to soft-reset because QMAN is
 706                                  * probably stuck. However, we can't call to
 707                                  * reset here directly because of deadlock, so
 708                                  * need to do it at the very end of this
 709                                  * function
 710                                  */
 711                                 if ((rc == -ETIMEDOUT) || (rc == -EBUSY))
 712                                         need_soft_reset = true;
 713                                 mutex_unlock(&hpriv->restore_phase_mutex);
 714                                 goto out;
 715                         }
 716                 }
 717
 718                 hdev->asic_funcs->restore_phase_topology(hdev);
 719
 720                 if (!num_chunks_restore) {
 721                         dev_dbg(hdev->dev,
 722                         "Need to run restore phase but restore CS is empty\n");
 723                         rc = 0;
 724                 } else {
 725                         rc = _hl_cs_ioctl(hpriv, chunks_restore,
 726                                                 num_chunks_restore, &cs_seq);
 727                 }
 728
 729                 mutex_unlock(&hpriv->restore_phase_mutex);
 730
 731                 if (rc) {
 732                         dev_err(hdev->dev,
 733                                 "Failed to submit restore CS for context %d (%d)\n",
 734                                 ctx->asid, rc);
 735                         goto out;
 736                 }
 737
 738                 /* Need to wait for restore completion before execution phase */
 739                 if (num_chunks_restore) {
 740                         ret = _hl_cs_wait_ioctl(hdev, ctx,
 741                                         jiffies_to_usecs(hdev->timeout_jiffies),
 742                                         cs_seq);
 743                         if (ret <= 0) {
 744                                 dev_err(hdev->dev,
 745                                         "Restore CS for context %d failed to complete %ld\n",
 746                                         ctx->asid, ret);
 747                                 rc = -ENOEXEC;
 748                                 goto out;
 749                         }
 750                 }
 751
 752                 ctx->thread_ctx_switch_wait_token = 1;
 753         } else if (!ctx->thread_ctx_switch_wait_token) {
 754                 u32 tmp;
 755
 756                 rc = hl_poll_timeout_memory(hdev,
 757                         &ctx->thread_ctx_switch_wait_token, tmp, (tmp == 1),
 758                         100, jiffies_to_usecs(hdev->timeout_jiffies), false);
 759
 760                 if (rc == -ETIMEDOUT) {
 761                         dev_err(hdev->dev,
 762                                 "context switch phase timeout (%d)\n", tmp);
 763                         goto out;
 764                 }
 765         }
 766
 767         rc = _hl_cs_ioctl(hpriv, chunks_execute, num_chunks_execute, &cs_seq);
 768
 769 out:
 770         if (rc != -EAGAIN) {
 771                 memset(args, 0, sizeof(*args));
 772                 args->out.status = rc;
 773                 args->out.seq = cs_seq;
 774         }
 775
 776         if (((rc == -ETIMEDOUT) || (rc == -EBUSY)) && (need_soft_reset))
 777                 hl_device_reset(hdev, false, false);
 778
 779         return rc;
 780 }
 781
 782 static long _hl_cs_wait_ioctl(struct hl_device *hdev,
 783                 struct hl_ctx *ctx, u64 timeout_us, u64 seq)
 784 {
 785         struct dma_fence *fence;
 786         unsigned long timeout;
 787         long rc;
 788
 789         if (timeout_us == MAX_SCHEDULE_TIMEOUT)
 790                 timeout = timeout_us;
 791         else
 792                 timeout = usecs_to_jiffies(timeout_us);
 793
 794         hl_ctx_get(hdev, ctx);
 795
 796         fence = hl_ctx_get_fence(ctx, seq);
 797         if (IS_ERR(fence)) {
 798                 rc = PTR_ERR(fence);
 799         } else if (fence) {
 800                 rc = dma_fence_wait_timeout(fence, true, timeout);
 801                 if (fence->error == -ETIMEDOUT)
 802                         rc = -ETIMEDOUT;
 803                 else if (fence->error == -EIO)
 804                         rc = -EIO;
 805                 dma_fence_put(fence);
 806         } else
 807                 rc = 1;
 808
 809         hl_ctx_put(ctx);
 810
 811         return rc;
 812 }
 813
 814 int hl_cs_wait_ioctl(struct hl_fpriv *hpriv, void *data)
 815 {
 816         struct hl_device *hdev = hpriv->hdev;
 817         union hl_wait_cs_args *args = data;
 818         u64 seq = args->in.seq;
 819         long rc;
 820
 821         rc = _hl_cs_wait_ioctl(hdev, hpriv->ctx, args->in.timeout_us, seq);
 822
 823         memset(args, 0, sizeof(*args));
 824
 825         if (rc < 0) {
 826                 dev_err_ratelimited(hdev->dev,
 827                                 "Error %ld on waiting for CS handle %llu\n",
 828                                 rc, seq);
 829                 if (rc == -ERESTARTSYS) {
 830                         args->out.status = HL_WAIT_CS_STATUS_INTERRUPTED;
 831                         rc = -EINTR;
 832                 } else if (rc == -ETIMEDOUT) {
 833                         args->out.status = HL_WAIT_CS_STATUS_TIMEDOUT;
 834                 } else if (rc == -EIO) {
 835                         args->out.status = HL_WAIT_CS_STATUS_ABORTED;
 836                 }
 837                 return rc;
 838         }
 839
 840         if (rc == 0)
 841                 args->out.status = HL_WAIT_CS_STATUS_BUSY;
 842         else
 843                 args->out.status = HL_WAIT_CS_STATUS_COMPLETED;
 844
 845         return 0;
 846 }