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 #define HL_CS_FLAGS_SIG_WAIT    (HL_CS_FLAGS_SIGNAL | HL_CS_FLAGS_WAIT)
  15
  16 static void job_wq_completion(struct work_struct *work);
  17 static long _hl_cs_wait_ioctl(struct hl_device *hdev,
  18                 struct hl_ctx *ctx, u64 timeout_us, u64 seq);
  19 static void cs_do_release(struct kref *ref);
  20
  21 static void hl_sob_reset(struct kref *ref)
  22 {
  23         struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
  24                                                         kref);
  25         struct hl_device *hdev = hw_sob->hdev;
  26
  27         hdev->asic_funcs->reset_sob(hdev, hw_sob);
  28 }
  29
  30 void hl_sob_reset_error(struct kref *ref)
  31 {
  32         struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
  33                                                         kref);
  34         struct hl_device *hdev = hw_sob->hdev;
  35
  36         dev_crit(hdev->dev,
  37                         "SOB release shouldn't be called here, q_idx: %d, sob_id: %d\n",
  38                         hw_sob->q_idx, hw_sob->sob_id);
  39 }
  40
  41 static const char *hl_fence_get_driver_name(struct dma_fence *fence)
  42 {
  43         return "HabanaLabs";
  44 }
  45
  46 static const char *hl_fence_get_timeline_name(struct dma_fence *fence)
  47 {
  48         struct hl_cs_compl *hl_cs_compl =
  49                 container_of(fence, struct hl_cs_compl, base_fence);
  50
  51         return dev_name(hl_cs_compl->hdev->dev);
  52 }
  53
  54 static bool hl_fence_enable_signaling(struct dma_fence *fence)
  55 {
  56         return true;
  57 }
  58
  59 static void hl_fence_release(struct dma_fence *fence)
  60 {
  61         struct hl_cs_compl *hl_cs_cmpl =
  62                 container_of(fence, struct hl_cs_compl, base_fence);
  63         struct hl_device *hdev = hl_cs_cmpl->hdev;
  64
  65         /* EBUSY means the CS was never submitted and hence we don't have
  66          * an attached hw_sob object that we should handle here
  67          */
  68         if (fence->error == -EBUSY)
  69                 goto free;
  70
  71         if ((hl_cs_cmpl->type == CS_TYPE_SIGNAL) ||
  72                         (hl_cs_cmpl->type == CS_TYPE_WAIT)) {
  73
  74                 dev_dbg(hdev->dev,
  75                         "CS 0x%llx type %d finished, sob_id: %d, sob_val: 0x%x\n",
  76                         hl_cs_cmpl->cs_seq,
  77                         hl_cs_cmpl->type,
  78                         hl_cs_cmpl->hw_sob->sob_id,
  79                         hl_cs_cmpl->sob_val);
  80
  81                 /*
  82                  * A signal CS can get completion while the corresponding wait
  83                  * for signal CS is on its way to the PQ. The wait for signal CS
  84                  * will get stuck if the signal CS incremented the SOB to its
  85                  * max value and there are no pending (submitted) waits on this
  86                  * SOB.
  87                  * We do the following to void this situation:
  88                  * 1. The wait for signal CS must get a ref for the signal CS as
  89                  *    soon as possible in cs_ioctl_signal_wait() and put it
  90                  *    before being submitted to the PQ but after it incremented
  91                  *    the SOB refcnt in init_signal_wait_cs().
  92                  * 2. Signal/Wait for signal CS will decrement the SOB refcnt
  93                  *    here.
  94                  * These two measures guarantee that the wait for signal CS will
  95                  * reset the SOB upon completion rather than the signal CS and
  96                  * hence the above scenario is avoided.
  97                  */
  98                 kref_put(&hl_cs_cmpl->hw_sob->kref, hl_sob_reset);
  99         }
 100
 101 free:
 102         kfree_rcu(hl_cs_cmpl, base_fence.rcu);
 103 }
 104
 105 static const struct dma_fence_ops hl_fence_ops = {
 106         .get_driver_name = hl_fence_get_driver_name,
 107         .get_timeline_name = hl_fence_get_timeline_name,
 108         .enable_signaling = hl_fence_enable_signaling,
 109         .release = hl_fence_release
 110 };
 111
 112 static void cs_get(struct hl_cs *cs)
 113 {
 114         kref_get(&cs->refcount);
 115 }
 116
 117 static int cs_get_unless_zero(struct hl_cs *cs)
 118 {
 119         return kref_get_unless_zero(&cs->refcount);
 120 }
 121
 122 static void cs_put(struct hl_cs *cs)
 123 {
 124         kref_put(&cs->refcount, cs_do_release);
 125 }
 126
 127 static bool is_cb_patched(struct hl_device *hdev, struct hl_cs_job *job)
 128 {
 129         /*
 130          * Patched CB is created for external queues jobs, and for H/W queues
 131          * jobs if the user CB was allocated by driver and MMU is disabled.
 132          */
 133         return (job->queue_type == QUEUE_TYPE_EXT ||
 134                         (job->queue_type == QUEUE_TYPE_HW &&
 135                                         job->is_kernel_allocated_cb &&
 136                                         !hdev->mmu_enable));
 137 }
 138
 139 /*
 140  * cs_parser - parse the user command submission
 141  *
 142  * @hpriv       : pointer to the private data of the fd
 143  * @job        : pointer to the job that holds the command submission info
 144  *
 145  * The function parses the command submission of the user. It calls the
 146  * ASIC specific parser, which returns a list of memory blocks to send
 147  * to the device as different command buffers
 148  *
 149  */
 150 static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
 151 {
 152         struct hl_device *hdev = hpriv->hdev;
 153         struct hl_cs_parser parser;
 154         int rc;
 155
 156         parser.ctx_id = job->cs->ctx->asid;
 157         parser.cs_sequence = job->cs->sequence;
 158         parser.job_id = job->id;
 159
 160         parser.hw_queue_id = job->hw_queue_id;
 161         parser.job_userptr_list = &job->userptr_list;
 162         parser.patched_cb = NULL;
 163         parser.user_cb = job->user_cb;
 164         parser.user_cb_size = job->user_cb_size;
 165         parser.queue_type = job->queue_type;
 166         parser.is_kernel_allocated_cb = job->is_kernel_allocated_cb;
 167         job->patched_cb = NULL;
 168
 169         rc = hdev->asic_funcs->cs_parser(hdev, &parser);
 170
 171         if (is_cb_patched(hdev, job)) {
 172                 if (!rc) {
 173                         job->patched_cb = parser.patched_cb;
 174                         job->job_cb_size = parser.patched_cb_size;
 175                         job->contains_dma_pkt = parser.contains_dma_pkt;
 176
 177                         spin_lock(&job->patched_cb->lock);
 178                         job->patched_cb->cs_cnt++;
 179                         spin_unlock(&job->patched_cb->lock);
 180                 }
 181
 182                 /*
 183                  * Whether the parsing worked or not, we don't need the
 184                  * original CB anymore because it was already parsed and
 185                  * won't be accessed again for this CS
 186                  */
 187                 spin_lock(&job->user_cb->lock);
 188                 job->user_cb->cs_cnt--;
 189                 spin_unlock(&job->user_cb->lock);
 190                 hl_cb_put(job->user_cb);
 191                 job->user_cb = NULL;
 192         } else if (!rc) {
 193                 job->job_cb_size = job->user_cb_size;
 194         }
 195
 196         return rc;
 197 }
 198
 199 static void free_job(struct hl_device *hdev, struct hl_cs_job *job)
 200 {
 201         struct hl_cs *cs = job->cs;
 202
 203         if (is_cb_patched(hdev, job)) {
 204                 hl_userptr_delete_list(hdev, &job->userptr_list);
 205
 206                 /*
 207                  * We might arrive here from rollback and patched CB wasn't
 208                  * created, so we need to check it's not NULL
 209                  */
 210                 if (job->patched_cb) {
 211                         spin_lock(&job->patched_cb->lock);
 212                         job->patched_cb->cs_cnt--;
 213                         spin_unlock(&job->patched_cb->lock);
 214
 215                         hl_cb_put(job->patched_cb);
 216                 }
 217         }
 218
 219         /* For H/W queue jobs, if a user CB was allocated by driver and MMU is
 220          * enabled, the user CB isn't released in cs_parser() and thus should be
 221          * released here.
 222          */
 223         if (job->queue_type == QUEUE_TYPE_HW &&
 224                         job->is_kernel_allocated_cb && hdev->mmu_enable) {
 225                 spin_lock(&job->user_cb->lock);
 226                 job->user_cb->cs_cnt--;
 227                 spin_unlock(&job->user_cb->lock);
 228
 229                 hl_cb_put(job->user_cb);
 230         }
 231
 232         /*
 233          * This is the only place where there can be multiple threads
 234          * modifying the list at the same time
 235          */
 236         spin_lock(&cs->job_lock);
 237         list_del(&job->cs_node);
 238         spin_unlock(&cs->job_lock);
 239
 240         hl_debugfs_remove_job(hdev, job);
 241
 242         if (job->queue_type == QUEUE_TYPE_EXT ||
 243                         job->queue_type == QUEUE_TYPE_HW)
 244                 cs_put(cs);
 245
 246         kfree(job);
 247 }
 248
 249 static void cs_counters_aggregate(struct hl_device *hdev, struct hl_ctx *ctx)
 250 {
 251         hdev->aggregated_cs_counters.device_in_reset_drop_cnt +=
 252                         ctx->cs_counters.device_in_reset_drop_cnt;
 253         hdev->aggregated_cs_counters.out_of_mem_drop_cnt +=
 254                         ctx->cs_counters.out_of_mem_drop_cnt;
 255         hdev->aggregated_cs_counters.parsing_drop_cnt +=
 256                         ctx->cs_counters.parsing_drop_cnt;
 257         hdev->aggregated_cs_counters.queue_full_drop_cnt +=
 258                         ctx->cs_counters.queue_full_drop_cnt;
 259 }
 260
 261 static void cs_do_release(struct kref *ref)
 262 {
 263         struct hl_cs *cs = container_of(ref, struct hl_cs,
 264                                                 refcount);
 265         struct hl_device *hdev = cs->ctx->hdev;
 266         struct hl_cs_job *job, *tmp;
 267
 268         cs->completed = true;
 269
 270         /*
 271          * Although if we reached here it means that all external jobs have
 272          * finished, because each one of them took refcnt to CS, we still
 273          * need to go over the internal jobs and free them. Otherwise, we
 274          * will have leaked memory and what's worse, the CS object (and
 275          * potentially the CTX object) could be released, while the JOB
 276          * still holds a pointer to them (but no reference).
 277          */
 278         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 279                 free_job(hdev, job);
 280
 281         /* We also need to update CI for internal queues */
 282         if (cs->submitted) {
 283                 hdev->asic_funcs->hw_queues_lock(hdev);
 284
 285                 hdev->cs_active_cnt--;
 286                 if (!hdev->cs_active_cnt) {
 287                         struct hl_device_idle_busy_ts *ts;
 288
 289                         ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx++];
 290                         ts->busy_to_idle_ts = ktime_get();
 291
 292                         if (hdev->idle_busy_ts_idx == HL_IDLE_BUSY_TS_ARR_SIZE)
 293                                 hdev->idle_busy_ts_idx = 0;
 294                 } else if (hdev->cs_active_cnt < 0) {
 295                         dev_crit(hdev->dev, "CS active cnt %d is negative\n",
 296                                 hdev->cs_active_cnt);
 297                 }
 298
 299                 hdev->asic_funcs->hw_queues_unlock(hdev);
 300
 301                 hl_int_hw_queue_update_ci(cs);
 302
 303                 spin_lock(&hdev->hw_queues_mirror_lock);
 304                 /* remove CS from hw_queues mirror list */
 305                 list_del_init(&cs->mirror_node);
 306                 spin_unlock(&hdev->hw_queues_mirror_lock);
 307
 308                 /*
 309                  * Don't cancel TDR in case this CS was timedout because we
 310                  * might be running from the TDR context
 311                  */
 312                 if ((!cs->timedout) &&
 313                         (hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT)) {
 314                         struct hl_cs *next;
 315
 316                         if (cs->tdr_active)
 317                                 cancel_delayed_work_sync(&cs->work_tdr);
 318
 319                         spin_lock(&hdev->hw_queues_mirror_lock);
 320
 321                         /* queue TDR for next CS */
 322                         next = list_first_entry_or_null(
 323                                         &hdev->hw_queues_mirror_list,
 324                                         struct hl_cs, mirror_node);
 325
 326                         if ((next) && (!next->tdr_active)) {
 327                                 next->tdr_active = true;
 328                                 schedule_delayed_work(&next->work_tdr,
 329                                                         hdev->timeout_jiffies);
 330                         }
 331
 332                         spin_unlock(&hdev->hw_queues_mirror_lock);
 333                 }
 334         } else if (cs->type == CS_TYPE_WAIT) {
 335                 /*
 336                  * In case the wait for signal CS was submitted, the put occurs
 337                  * in init_signal_wait_cs() right before hanging on the PQ.
 338                  */
 339                 dma_fence_put(cs->signal_fence);
 340         }
 341
 342         /*
 343          * Must be called before hl_ctx_put because inside we use ctx to get
 344          * the device
 345          */
 346         hl_debugfs_remove_cs(cs);
 347
 348         hl_ctx_put(cs->ctx);
 349
 350         /* We need to mark an error for not submitted because in that case
 351          * the dma fence release flow is different. Mainly, we don't need
 352          * to handle hw_sob for signal/wait
 353          */
 354         if (cs->timedout)
 355                 dma_fence_set_error(cs->fence, -ETIMEDOUT);
 356         else if (cs->aborted)
 357                 dma_fence_set_error(cs->fence, -EIO);
 358         else if (!cs->submitted)
 359                 dma_fence_set_error(cs->fence, -EBUSY);
 360
 361         dma_fence_signal(cs->fence);
 362         dma_fence_put(cs->fence);
 363
 364         cs_counters_aggregate(hdev, cs->ctx);
 365
 366         kfree(cs->jobs_in_queue_cnt);
 367         kfree(cs);
 368 }
 369
 370 static void cs_timedout(struct work_struct *work)
 371 {
 372         struct hl_device *hdev;
 373         int ctx_asid, rc;
 374         struct hl_cs *cs = container_of(work, struct hl_cs,
 375                                                  work_tdr.work);
 376         rc = cs_get_unless_zero(cs);
 377         if (!rc)
 378                 return;
 379
 380         if ((!cs->submitted) || (cs->completed)) {
 381                 cs_put(cs);
 382                 return;
 383         }
 384
 385         /* Mark the CS is timed out so we won't try to cancel its TDR */
 386         cs->timedout = true;
 387
 388         hdev = cs->ctx->hdev;
 389         ctx_asid = cs->ctx->asid;
 390
 391         dev_err(hdev->dev,
 392                 "Command submission %llu has not finished in time!\n",
 393                 cs->sequence);
 394
 395         cs_put(cs);
 396
 397         if (hdev->reset_on_lockup)
 398                 hl_device_reset(hdev, false, false);
 399 }
 400
 401 static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 402                         enum hl_cs_type cs_type, struct hl_cs **cs_new)
 403 {
 404         struct hl_cs_compl *cs_cmpl;
 405         struct dma_fence *other = NULL;
 406         struct hl_cs *cs;
 407         int rc;
 408
 409         cs = kzalloc(sizeof(*cs), GFP_ATOMIC);
 410         if (!cs)
 411                 return -ENOMEM;
 412
 413         cs->ctx = ctx;
 414         cs->submitted = false;
 415         cs->completed = false;
 416         cs->type = cs_type;
 417         INIT_LIST_HEAD(&cs->job_list);
 418         INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
 419         kref_init(&cs->refcount);
 420         spin_lock_init(&cs->job_lock);
 421
 422         cs_cmpl = kmalloc(sizeof(*cs_cmpl), GFP_ATOMIC);
 423         if (!cs_cmpl) {
 424                 rc = -ENOMEM;
 425                 goto free_cs;
 426         }
 427
 428         cs_cmpl->hdev = hdev;
 429         cs_cmpl->type = cs->type;
 430         spin_lock_init(&cs_cmpl->lock);
 431         cs->fence = &cs_cmpl->base_fence;
 432
 433         spin_lock(&ctx->cs_lock);
 434
 435         cs_cmpl->cs_seq = ctx->cs_sequence;
 436         other = ctx->cs_pending[cs_cmpl->cs_seq &
 437                                 (hdev->asic_prop.max_pending_cs - 1)];
 438         if ((other) && (!dma_fence_is_signaled(other))) {
 439                 dev_dbg(hdev->dev,
 440                         "Rejecting CS because of too many in-flights CS\n");
 441                 rc = -EAGAIN;
 442                 goto free_fence;
 443         }
 444
 445         cs->jobs_in_queue_cnt = kcalloc(hdev->asic_prop.max_queues,
 446                         sizeof(*cs->jobs_in_queue_cnt), GFP_ATOMIC);
 447         if (!cs->jobs_in_queue_cnt) {
 448                 rc = -ENOMEM;
 449                 goto free_fence;
 450         }
 451
 452         dma_fence_init(&cs_cmpl->base_fence, &hl_fence_ops, &cs_cmpl->lock,
 453                         ctx->asid, ctx->cs_sequence);
 454
 455         cs->sequence = cs_cmpl->cs_seq;
 456
 457         ctx->cs_pending[cs_cmpl->cs_seq &
 458                         (hdev->asic_prop.max_pending_cs - 1)] =
 459                                                         &cs_cmpl->base_fence;
 460         ctx->cs_sequence++;
 461
 462         dma_fence_get(&cs_cmpl->base_fence);
 463
 464         dma_fence_put(other);
 465
 466         spin_unlock(&ctx->cs_lock);
 467
 468         *cs_new = cs;
 469
 470         return 0;
 471
 472 free_fence:
 473         spin_unlock(&ctx->cs_lock);
 474         kfree(cs_cmpl);
 475 free_cs:
 476         kfree(cs);
 477         return rc;
 478 }
 479
 480 static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)
 481 {
 482         struct hl_cs_job *job, *tmp;
 483
 484         list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 485                 free_job(hdev, job);
 486 }
 487
 488 void hl_cs_rollback_all(struct hl_device *hdev)
 489 {
 490         int i;
 491         struct hl_cs *cs, *tmp;
 492
 493         /* flush all completions */
 494         for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
 495                 flush_workqueue(hdev->cq_wq[i]);
 496
 497         /* Make sure we don't have leftovers in the H/W queues mirror list */
 498         list_for_each_entry_safe(cs, tmp, &hdev->hw_queues_mirror_list,
 499                                 mirror_node) {
 500                 cs_get(cs);
 501                 cs->aborted = true;
 502                 dev_warn_ratelimited(hdev->dev, "Killing CS %d.%llu\n",
 503                                         cs->ctx->asid, cs->sequence);
 504                 cs_rollback(hdev, cs);
 505                 cs_put(cs);
 506         }
 507 }
 508
 509 static void job_wq_completion(struct work_struct *work)
 510 {
 511         struct hl_cs_job *job = container_of(work, struct hl_cs_job,
 512                                                 finish_work);
 513         struct hl_cs *cs = job->cs;
 514         struct hl_device *hdev = cs->ctx->hdev;
 515
 516         /* job is no longer needed */
 517         free_job(hdev, job);
 518 }
 519
 520 static int validate_queue_index(struct hl_device *hdev,
 521                                 struct hl_cs_chunk *chunk,
 522                                 enum hl_queue_type *queue_type,
 523                                 bool *is_kernel_allocated_cb)
 524 {
 525         struct asic_fixed_properties *asic = &hdev->asic_prop;
 526         struct hw_queue_properties *hw_queue_prop;
 527
 528         /* This must be checked here to prevent out-of-bounds access to
 529          * hw_queues_props array
 530          */
 531         if (chunk->queue_index >= asic->max_queues) {
 532                 dev_err(hdev->dev, "Queue index %d is invalid\n",
 533                         chunk->queue_index);
 534                 return -EINVAL;
 535         }
 536
 537         hw_queue_prop = &asic->hw_queues_props[chunk->queue_index];
 538
 539         if (hw_queue_prop->type == QUEUE_TYPE_NA) {
 540                 dev_err(hdev->dev, "Queue index %d is invalid\n",
 541                         chunk->queue_index);
 542                 return -EINVAL;
 543         }
 544
 545         if (hw_queue_prop->driver_only) {
 546                 dev_err(hdev->dev,
 547                         "Queue index %d is restricted for the kernel driver\n",
 548                         chunk->queue_index);
 549                 return -EINVAL;
 550         }
 551
 552         *queue_type = hw_queue_prop->type;
 553         *is_kernel_allocated_cb = !!hw_queue_prop->requires_kernel_cb;
 554
 555         return 0;
 556 }
 557
 558 static struct hl_cb *get_cb_from_cs_chunk(struct hl_device *hdev,
 559                                         struct hl_cb_mgr *cb_mgr,
 560                                         struct hl_cs_chunk *chunk)
 561 {
 562         struct hl_cb *cb;
 563         u32 cb_handle;
 564
 565         cb_handle = (u32) (chunk->cb_handle >> PAGE_SHIFT);
 566
 567         cb = hl_cb_get(hdev, cb_mgr, cb_handle);
 568         if (!cb) {
 569                 dev_err(hdev->dev, "CB handle 0x%x invalid\n", cb_handle);
 570                 return NULL;
 571         }
 572
 573         if ((chunk->cb_size < 8) || (chunk->cb_size > cb->size)) {
 574                 dev_err(hdev->dev, "CB size %u invalid\n", chunk->cb_size);
 575                 goto release_cb;
 576         }
 577
 578         spin_lock(&cb->lock);
 579         cb->cs_cnt++;
 580         spin_unlock(&cb->lock);
 581
 582         return cb;
 583
 584 release_cb:
 585         hl_cb_put(cb);
 586         return NULL;
 587 }
 588
 589 struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
 590                 enum hl_queue_type queue_type, bool is_kernel_allocated_cb)
 591 {
 592         struct hl_cs_job *job;
 593
 594         job = kzalloc(sizeof(*job), GFP_ATOMIC);
 595         if (!job)
 596                 return NULL;
 597
 598         job->queue_type = queue_type;
 599         job->is_kernel_allocated_cb = is_kernel_allocated_cb;
 600
 601         if (is_cb_patched(hdev, job))
 602                 INIT_LIST_HEAD(&job->userptr_list);
 603
 604         if (job->queue_type == QUEUE_TYPE_EXT)
 605                 INIT_WORK(&job->finish_work, job_wq_completion);
 606
 607         return job;
 608 }
 609
 610 static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
 611                                 u32 num_chunks, u64 *cs_seq)
 612 {
 613         struct hl_device *hdev = hpriv->hdev;
 614         struct hl_cs_chunk *cs_chunk_array;
 615         struct hl_cs_job *job;
 616         struct hl_cs *cs;
 617         struct hl_cb *cb;
 618         bool int_queues_only = true;
 619         u32 size_to_copy;
 620         int rc, i;
 621
 622         *cs_seq = ULLONG_MAX;
 623
 624         if (num_chunks > HL_MAX_JOBS_PER_CS) {
 625                 dev_err(hdev->dev,
 626                         "Number of chunks can NOT be larger than %d\n",
 627                         HL_MAX_JOBS_PER_CS);
 628                 rc = -EINVAL;
 629                 goto out;
 630         }
 631
 632         cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
 633                                         GFP_ATOMIC);
 634         if (!cs_chunk_array) {
 635                 rc = -ENOMEM;
 636                 goto out;
 637         }
 638
 639         size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
 640         if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
 641                 dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
 642                 rc = -EFAULT;
 643                 goto free_cs_chunk_array;
 644         }
 645
 646         /* increment refcnt for context */
 647         hl_ctx_get(hdev, hpriv->ctx);
 648
 649         rc = allocate_cs(hdev, hpriv->ctx, CS_TYPE_DEFAULT, &cs);
 650         if (rc) {
 651                 hl_ctx_put(hpriv->ctx);
 652                 goto free_cs_chunk_array;
 653         }
 654
 655         *cs_seq = cs->sequence;
 656
 657         hl_debugfs_add_cs(cs);
 658
 659         /* Validate ALL the CS chunks before submitting the CS */
 660         for (i = 0 ; i < num_chunks ; i++) {
 661                 struct hl_cs_chunk *chunk = &cs_chunk_array[i];
 662                 enum hl_queue_type queue_type;
 663                 bool is_kernel_allocated_cb;
 664
 665                 rc = validate_queue_index(hdev, chunk, &queue_type,
 666                                                 &is_kernel_allocated_cb);
 667                 if (rc) {
 668                         hpriv->ctx->cs_counters.parsing_drop_cnt++;
 669                         goto free_cs_object;
 670                 }
 671
 672                 if (is_kernel_allocated_cb) {
 673                         cb = get_cb_from_cs_chunk(hdev, &hpriv->cb_mgr, chunk);
 674                         if (!cb) {
 675                                 hpriv->ctx->cs_counters.parsing_drop_cnt++;
 676                                 rc = -EINVAL;
 677                                 goto free_cs_object;
 678                         }
 679                 } else {
 680                         cb = (struct hl_cb *) (uintptr_t) chunk->cb_handle;
 681                 }
 682
 683                 if (queue_type == QUEUE_TYPE_EXT || queue_type == QUEUE_TYPE_HW)
 684                         int_queues_only = false;
 685
 686                 job = hl_cs_allocate_job(hdev, queue_type,
 687                                                 is_kernel_allocated_cb);
 688                 if (!job) {
 689                         hpriv->ctx->cs_counters.out_of_mem_drop_cnt++;
 690                         dev_err(hdev->dev, "Failed to allocate a new job\n");
 691                         rc = -ENOMEM;
 692                         if (is_kernel_allocated_cb)
 693                                 goto release_cb;
 694                         else
 695                                 goto free_cs_object;
 696                 }
 697
 698                 job->id = i + 1;
 699                 job->cs = cs;
 700                 job->user_cb = cb;
 701                 job->user_cb_size = chunk->cb_size;
 702                 job->hw_queue_id = chunk->queue_index;
 703
 704                 cs->jobs_in_queue_cnt[job->hw_queue_id]++;
 705
 706                 list_add_tail(&job->cs_node, &cs->job_list);
 707
 708                 /*
 709                  * Increment CS reference. When CS reference is 0, CS is
 710                  * done and can be signaled to user and free all its resources
 711                  * Only increment for JOB on external or H/W queues, because
 712                  * only for those JOBs we get completion
 713                  */
 714                 if (job->queue_type == QUEUE_TYPE_EXT ||
 715                                 job->queue_type == QUEUE_TYPE_HW)
 716                         cs_get(cs);
 717
 718                 hl_debugfs_add_job(hdev, job);
 719
 720                 rc = cs_parser(hpriv, job);
 721                 if (rc) {
 722                         hpriv->ctx->cs_counters.parsing_drop_cnt++;
 723                         dev_err(hdev->dev,
 724                                 "Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
 725                                 cs->ctx->asid, cs->sequence, job->id, rc);
 726                         goto free_cs_object;
 727                 }
 728         }
 729
 730         if (int_queues_only) {
 731                 hpriv->ctx->cs_counters.parsing_drop_cnt++;
 732                 dev_err(hdev->dev,
 733                         "Reject CS %d.%llu because only internal queues jobs are present\n",
 734                         cs->ctx->asid, cs->sequence);
 735                 rc = -EINVAL;
 736                 goto free_cs_object;
 737         }
 738
 739         rc = hl_hw_queue_schedule_cs(cs);
 740         if (rc) {
 741                 if (rc != -EAGAIN)
 742                         dev_err(hdev->dev,
 743                                 "Failed to submit CS %d.%llu to H/W queues, error %d\n",
 744                                 cs->ctx->asid, cs->sequence, rc);
 745                 goto free_cs_object;
 746         }
 747
 748         rc = HL_CS_STATUS_SUCCESS;
 749         goto put_cs;
 750
 751 release_cb:
 752         spin_lock(&cb->lock);
 753         cb->cs_cnt--;
 754         spin_unlock(&cb->lock);
 755         hl_cb_put(cb);
 756 free_cs_object:
 757         cs_rollback(hdev, cs);
 758         *cs_seq = ULLONG_MAX;
 759         /* The path below is both for good and erroneous exits */
 760 put_cs:
 761         /* We finished with the CS in this function, so put the ref */
 762         cs_put(cs);
 763 free_cs_chunk_array:
 764         kfree(cs_chunk_array);
 765 out:
 766         return rc;
 767 }
 768
 769 static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
 770                                 void __user *chunks, u32 num_chunks,
 771                                 u64 *cs_seq)
 772 {
 773         struct hl_device *hdev = hpriv->hdev;
 774         struct hl_ctx *ctx = hpriv->ctx;
 775         struct hl_cs_chunk *cs_chunk_array, *chunk;
 776         struct hw_queue_properties *hw_queue_prop;
 777         struct dma_fence *sig_fence = NULL;
 778         struct hl_cs_job *job;
 779         struct hl_cs *cs;
 780         struct hl_cb *cb;
 781         enum hl_queue_type q_type;
 782         u64 *signal_seq_arr = NULL, signal_seq;
 783         u32 size_to_copy, q_idx, signal_seq_arr_len, cb_size;
 784         int rc;
 785
 786         *cs_seq = ULLONG_MAX;
 787
 788         if (num_chunks > HL_MAX_JOBS_PER_CS) {
 789                 dev_err(hdev->dev,
 790                         "Number of chunks can NOT be larger than %d\n",
 791                         HL_MAX_JOBS_PER_CS);
 792                 rc = -EINVAL;
 793                 goto out;
 794         }
 795
 796         cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
 797                                         GFP_ATOMIC);
 798         if (!cs_chunk_array) {
 799                 rc = -ENOMEM;
 800                 goto out;
 801         }
 802
 803         size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
 804         if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
 805                 dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
 806                 rc = -EFAULT;
 807                 goto free_cs_chunk_array;
 808         }
 809
 810         /* currently it is guaranteed to have only one chunk */
 811         chunk = &cs_chunk_array[0];
 812         q_idx = chunk->queue_index;
 813         hw_queue_prop = &hdev->asic_prop.hw_queues_props[q_idx];
 814         q_type = hw_queue_prop->type;
 815
 816         if ((q_idx >= hdev->asic_prop.max_queues) ||
 817                         (!hw_queue_prop->supports_sync_stream)) {
 818                 dev_err(hdev->dev, "Queue index %d is invalid\n", q_idx);
 819                 rc = -EINVAL;
 820                 goto free_cs_chunk_array;
 821         }
 822
 823         if (cs_type == CS_TYPE_WAIT) {
 824                 struct hl_cs_compl *sig_waitcs_cmpl;
 825
 826                 signal_seq_arr_len = chunk->num_signal_seq_arr;
 827
 828                 /* currently only one signal seq is supported */
 829                 if (signal_seq_arr_len != 1) {
 830                         dev_err(hdev->dev,
 831                                 "Wait for signal CS supports only one signal CS seq\n");
 832                         rc = -EINVAL;
 833                         goto free_cs_chunk_array;
 834                 }
 835
 836                 signal_seq_arr = kmalloc_array(signal_seq_arr_len,
 837                                                 sizeof(*signal_seq_arr),
 838                                                 GFP_ATOMIC);
 839                 if (!signal_seq_arr) {
 840                         rc = -ENOMEM;
 841                         goto free_cs_chunk_array;
 842                 }
 843
 844                 size_to_copy = chunk->num_signal_seq_arr *
 845                                 sizeof(*signal_seq_arr);
 846                 if (copy_from_user(signal_seq_arr,
 847                                         u64_to_user_ptr(chunk->signal_seq_arr),
 848                                         size_to_copy)) {
 849                         dev_err(hdev->dev,
 850                                 "Failed to copy signal seq array from user\n");
 851                         rc = -EFAULT;
 852                         goto free_signal_seq_array;
 853                 }
 854
 855                 /* currently it is guaranteed to have only one signal seq */
 856                 signal_seq = signal_seq_arr[0];
 857                 sig_fence = hl_ctx_get_fence(ctx, signal_seq);
 858                 if (IS_ERR(sig_fence)) {
 859                         dev_err(hdev->dev,
 860                                 "Failed to get signal CS with seq 0x%llx\n",
 861                                 signal_seq);
 862                         rc = PTR_ERR(sig_fence);
 863                         goto free_signal_seq_array;
 864                 }
 865
 866                 if (!sig_fence) {
 867                         /* signal CS already finished */
 868                         rc = 0;
 869                         goto free_signal_seq_array;
 870                 }
 871
 872                 sig_waitcs_cmpl =
 873                         container_of(sig_fence, struct hl_cs_compl, base_fence);
 874
 875                 if (sig_waitcs_cmpl->type != CS_TYPE_SIGNAL) {
 876                         dev_err(hdev->dev,
 877                                 "CS seq 0x%llx is not of a signal CS\n",
 878                                 signal_seq);
 879                         dma_fence_put(sig_fence);
 880                         rc = -EINVAL;
 881                         goto free_signal_seq_array;
 882                 }
 883
 884                 if (dma_fence_is_signaled(sig_fence)) {
 885                         /* signal CS already finished */
 886                         dma_fence_put(sig_fence);
 887                         rc = 0;
 888                         goto free_signal_seq_array;
 889                 }
 890         }
 891
 892         /* increment refcnt for context */
 893         hl_ctx_get(hdev, ctx);
 894
 895         rc = allocate_cs(hdev, ctx, cs_type, &cs);
 896         if (rc) {
 897                 if (cs_type == CS_TYPE_WAIT)
 898                         dma_fence_put(sig_fence);
 899                 hl_ctx_put(ctx);
 900                 goto free_signal_seq_array;
 901         }
 902
 903         /*
 904          * Save the signal CS fence for later initialization right before
 905          * hanging the wait CS on the queue.
 906          */
 907         if (cs->type == CS_TYPE_WAIT)
 908                 cs->signal_fence = sig_fence;
 909
 910         hl_debugfs_add_cs(cs);
 911
 912         *cs_seq = cs->sequence;
 913
 914         job = hl_cs_allocate_job(hdev, q_type, true);
 915         if (!job) {
 916                 ctx->cs_counters.out_of_mem_drop_cnt++;
 917                 dev_err(hdev->dev, "Failed to allocate a new job\n");
 918                 rc = -ENOMEM;
 919                 goto put_cs;
 920         }
 921
 922         cb = hl_cb_kernel_create(hdev, PAGE_SIZE);
 923         if (!cb) {
 924                 ctx->cs_counters.out_of_mem_drop_cnt++;
 925                 kfree(job);
 926                 rc = -EFAULT;
 927                 goto put_cs;
 928         }
 929
 930         if (cs->type == CS_TYPE_WAIT)
 931                 cb_size = hdev->asic_funcs->get_wait_cb_size(hdev);
 932         else
 933                 cb_size = hdev->asic_funcs->get_signal_cb_size(hdev);
 934
 935         job->id = 0;
 936         job->cs = cs;
 937         job->user_cb = cb;
 938         job->user_cb->cs_cnt++;
 939         job->user_cb_size = cb_size;
 940         job->hw_queue_id = q_idx;
 941
 942         /*
 943          * No need in parsing, user CB is the patched CB.
 944          * We call hl_cb_destroy() out of two reasons - we don't need the CB in
 945          * the CB idr anymore and to decrement its refcount as it was
 946          * incremented inside hl_cb_kernel_create().
 947          */
 948         job->patched_cb = job->user_cb;
 949         job->job_cb_size = job->user_cb_size;
 950         hl_cb_destroy(hdev, &hdev->kernel_cb_mgr, cb->id << PAGE_SHIFT);
 951
 952         cs->jobs_in_queue_cnt[job->hw_queue_id]++;
 953
 954         list_add_tail(&job->cs_node, &cs->job_list);
 955
 956         /* increment refcount as for external queues we get completion */
 957         cs_get(cs);
 958
 959         hl_debugfs_add_job(hdev, job);
 960
 961         rc = hl_hw_queue_schedule_cs(cs);
 962         if (rc) {
 963                 if (rc != -EAGAIN)
 964                         dev_err(hdev->dev,
 965                                 "Failed to submit CS %d.%llu to H/W queues, error %d\n",
 966                                 ctx->asid, cs->sequence, rc);
 967                 goto free_cs_object;
 968         }
 969
 970         rc = HL_CS_STATUS_SUCCESS;
 971         goto put_cs;
 972
 973 free_cs_object:
 974         cs_rollback(hdev, cs);
 975         *cs_seq = ULLONG_MAX;
 976         /* The path below is both for good and erroneous exits */
 977 put_cs:
 978         /* We finished with the CS in this function, so put the ref */
 979         cs_put(cs);
 980 free_signal_seq_array:
 981         if (cs_type == CS_TYPE_WAIT)
 982                 kfree(signal_seq_arr);
 983 free_cs_chunk_array:
 984         kfree(cs_chunk_array);
 985 out:
 986         return rc;
 987 }
 988
 989 int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 990 {
 991         struct hl_device *hdev = hpriv->hdev;
 992         union hl_cs_args *args = data;
 993         struct hl_ctx *ctx = hpriv->ctx;
 994         void __user *chunks_execute, *chunks_restore;
 995         enum hl_cs_type cs_type;
 996         u32 num_chunks_execute, num_chunks_restore, sig_wait_flags;
 997         u64 cs_seq = ULONG_MAX;
 998         int rc, do_ctx_switch;
 999         bool need_soft_reset = false;
1000
1001         if (hl_device_disabled_or_in_reset(hdev)) {
1002                 dev_warn_ratelimited(hdev->dev,
1003                         "Device is %s. Can't submit new CS\n",
1004                         atomic_read(&hdev->in_reset) ? "in_reset" : "disabled");
1005                 rc = -EBUSY;
1006                 goto out;
1007         }
1008
1009         sig_wait_flags = args->in.cs_flags & HL_CS_FLAGS_SIG_WAIT;
1010
1011         if (unlikely(sig_wait_flags == HL_CS_FLAGS_SIG_WAIT)) {
1012                 dev_err(hdev->dev,
1013                         "Signal and wait CS flags are mutually exclusive, context %d\n",
1014                 ctx->asid);
1015                 rc = -EINVAL;
1016                 goto out;
1017         }
1018
1019         if (unlikely((sig_wait_flags & HL_CS_FLAGS_SIG_WAIT) &&
1020                         (!hdev->supports_sync_stream))) {
1021                 dev_err(hdev->dev, "Sync stream CS is not supported\n");
1022                 rc = -EINVAL;
1023                 goto out;
1024         }
1025
1026         if (args->in.cs_flags & HL_CS_FLAGS_SIGNAL)
1027                 cs_type = CS_TYPE_SIGNAL;
1028         else if (args->in.cs_flags & HL_CS_FLAGS_WAIT)
1029                 cs_type = CS_TYPE_WAIT;
1030         else
1031                 cs_type = CS_TYPE_DEFAULT;
1032
1033         chunks_execute = (void __user *) (uintptr_t) args->in.chunks_execute;
1034         num_chunks_execute = args->in.num_chunks_execute;
1035
1036         if (cs_type == CS_TYPE_DEFAULT) {
1037                 if (!num_chunks_execute) {
1038                         dev_err(hdev->dev,
1039                                 "Got execute CS with 0 chunks, context %d\n",
1040                                 ctx->asid);
1041                         rc = -EINVAL;
1042                         goto out;
1043                 }
1044         } else if (num_chunks_execute != 1) {
1045                 dev_err(hdev->dev,
1046                         "Sync stream CS mandates one chunk only, context %d\n",
1047                         ctx->asid);
1048                 rc = -EINVAL;
1049                 goto out;
1050         }
1051
1052         do_ctx_switch = atomic_cmpxchg(&ctx->thread_ctx_switch_token, 1, 0);
1053
1054         if (do_ctx_switch || (args->in.cs_flags & HL_CS_FLAGS_FORCE_RESTORE)) {
1055                 long ret;
1056
1057                 chunks_restore =
1058                         (void __user *) (uintptr_t) args->in.chunks_restore;
1059                 num_chunks_restore = args->in.num_chunks_restore;
1060
1061                 mutex_lock(&hpriv->restore_phase_mutex);
1062
1063                 if (do_ctx_switch) {
1064                         rc = hdev->asic_funcs->context_switch(hdev, ctx->asid);
1065                         if (rc) {
1066                                 dev_err_ratelimited(hdev->dev,
1067                                         "Failed to switch to context %d, rejecting CS! %d\n",
1068                                         ctx->asid, rc);
1069                                 /*
1070                                  * If we timedout, or if the device is not IDLE
1071                                  * while we want to do context-switch (-EBUSY),
1072                                  * we need to soft-reset because QMAN is
1073                                  * probably stuck. However, we can't call to
1074                                  * reset here directly because of deadlock, so
1075                                  * need to do it at the very end of this
1076                                  * function
1077                                  */
1078                                 if ((rc == -ETIMEDOUT) || (rc == -EBUSY))
1079                                         need_soft_reset = true;
1080                                 mutex_unlock(&hpriv->restore_phase_mutex);
1081                                 goto out;
1082                         }
1083                 }
1084
1085                 hdev->asic_funcs->restore_phase_topology(hdev);
1086
1087                 if (!num_chunks_restore) {
1088                         dev_dbg(hdev->dev,
1089                         "Need to run restore phase but restore CS is empty\n");
1090                         rc = 0;
1091                 } else {
1092                         rc = cs_ioctl_default(hpriv, chunks_restore,
1093                                                 num_chunks_restore, &cs_seq);
1094                 }
1095
1096                 mutex_unlock(&hpriv->restore_phase_mutex);
1097
1098                 if (rc) {
1099                         dev_err(hdev->dev,
1100                                 "Failed to submit restore CS for context %d (%d)\n",
1101                                 ctx->asid, rc);
1102                         goto out;
1103                 }
1104
1105                 /* Need to wait for restore completion before execution phase */
1106                 if (num_chunks_restore) {
1107                         ret = _hl_cs_wait_ioctl(hdev, ctx,
1108                                         jiffies_to_usecs(hdev->timeout_jiffies),
1109                                         cs_seq);
1110                         if (ret <= 0) {
1111                                 dev_err(hdev->dev,
1112                                         "Restore CS for context %d failed to complete %ld\n",
1113                                         ctx->asid, ret);
1114                                 rc = -ENOEXEC;
1115                                 goto out;
1116                         }
1117                 }
1118
1119                 ctx->thread_ctx_switch_wait_token = 1;
1120         } else if (!ctx->thread_ctx_switch_wait_token) {
1121                 u32 tmp;
1122
1123                 rc = hl_poll_timeout_memory(hdev,
1124                         &ctx->thread_ctx_switch_wait_token, tmp, (tmp == 1),
1125                         100, jiffies_to_usecs(hdev->timeout_jiffies), false);
1126
1127                 if (rc == -ETIMEDOUT) {
1128                         dev_err(hdev->dev,
1129                                 "context switch phase timeout (%d)\n", tmp);
1130                         goto out;
1131                 }
1132         }
1133
1134         if (cs_type == CS_TYPE_DEFAULT)
1135                 rc = cs_ioctl_default(hpriv, chunks_execute, num_chunks_execute,
1136                                         &cs_seq);
1137         else
1138                 rc = cs_ioctl_signal_wait(hpriv, cs_type, chunks_execute,
1139                                                 num_chunks_execute, &cs_seq);
1140
1141 out:
1142         if (rc != -EAGAIN) {
1143                 memset(args, 0, sizeof(*args));
1144                 args->out.status = rc;
1145                 args->out.seq = cs_seq;
1146         }
1147
1148         if (((rc == -ETIMEDOUT) || (rc == -EBUSY)) && (need_soft_reset))
1149                 hl_device_reset(hdev, false, false);
1150
1151         return rc;
1152 }
1153
1154 static long _hl_cs_wait_ioctl(struct hl_device *hdev,
1155                 struct hl_ctx *ctx, u64 timeout_us, u64 seq)
1156 {
1157         struct dma_fence *fence;
1158         unsigned long timeout;
1159         long rc;
1160
1161         if (timeout_us == MAX_SCHEDULE_TIMEOUT)
1162                 timeout = timeout_us;
1163         else
1164                 timeout = usecs_to_jiffies(timeout_us);
1165
1166         hl_ctx_get(hdev, ctx);
1167
1168         fence = hl_ctx_get_fence(ctx, seq);
1169         if (IS_ERR(fence)) {
1170                 rc = PTR_ERR(fence);
1171                 if (rc == -EINVAL)
1172                         dev_notice_ratelimited(hdev->dev,
1173                                 "Can't wait on CS %llu because current CS is at seq %llu\n",
1174                                 seq, ctx->cs_sequence);
1175         } else if (fence) {
1176                 rc = dma_fence_wait_timeout(fence, true, timeout);
1177                 if (fence->error == -ETIMEDOUT)
1178                         rc = -ETIMEDOUT;
1179                 else if (fence->error == -EIO)
1180                         rc = -EIO;
1181                 dma_fence_put(fence);
1182         } else {
1183                 dev_dbg(hdev->dev,
1184                         "Can't wait on seq %llu because current CS is at seq %llu (Fence is gone)\n",
1185                         seq, ctx->cs_sequence);
1186                 rc = 1;
1187         }
1188
1189         hl_ctx_put(ctx);
1190
1191         return rc;
1192 }
1193
1194 int hl_cs_wait_ioctl(struct hl_fpriv *hpriv, void *data)
1195 {
1196         struct hl_device *hdev = hpriv->hdev;
1197         union hl_wait_cs_args *args = data;
1198         u64 seq = args->in.seq;
1199         long rc;
1200
1201         rc = _hl_cs_wait_ioctl(hdev, hpriv->ctx, args->in.timeout_us, seq);
1202
1203         memset(args, 0, sizeof(*args));
1204
1205         if (rc < 0) {
1206                 if (rc == -ERESTARTSYS) {
1207                         dev_err_ratelimited(hdev->dev,
1208                                 "user process got signal while waiting for CS handle %llu\n",
1209                                 seq);
1210                         args->out.status = HL_WAIT_CS_STATUS_INTERRUPTED;
1211                         rc = -EINTR;
1212                 } else if (rc == -ETIMEDOUT) {
1213                         dev_err_ratelimited(hdev->dev,
1214                                 "CS %llu has timed-out while user process is waiting for it\n",
1215                                 seq);
1216                         args->out.status = HL_WAIT_CS_STATUS_TIMEDOUT;
1217                 } else if (rc == -EIO) {
1218                         dev_err_ratelimited(hdev->dev,
1219                                 "CS %llu has been aborted while user process is waiting for it\n",
1220                                 seq);
1221                         args->out.status = HL_WAIT_CS_STATUS_ABORTED;
1222                 }
1223                 return rc;
1224         }
1225
1226         if (rc == 0)
1227                 args->out.status = HL_WAIT_CS_STATUS_BUSY;
1228         else
1229                 args->out.status = HL_WAIT_CS_STATUS_COMPLETED;
1230
1231         return 0;
1232 }