summaryrefslogtreecommitdiffstats
path: root/drivers/misc/habanalabs/command_submission.c
blob: 6ad83d5ef4b006759351962d54795b724e0dc590 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
// SPDX-License-Identifier: GPL-2.0

/*
 * Copyright 2016-2019 HabanaLabs, Ltd.
 * All Rights Reserved.
 */

#include <uapi/misc/habanalabs.h>
#include "habanalabs.h"

#include <linux/uaccess.h>
#include <linux/slab.h>

static void job_wq_completion(struct work_struct *work);
static long _hl_cs_wait_ioctl(struct hl_device *hdev,
		struct hl_ctx *ctx, u64 timeout_us, u64 seq);
static void cs_do_release(struct kref *ref);

static const char *hl_fence_get_driver_name(struct dma_fence *fence)
{
	return "HabanaLabs";
}

static const char *hl_fence_get_timeline_name(struct dma_fence *fence)
{
	struct hl_dma_fence *hl_fence =
		container_of(fence, struct hl_dma_fence, base_fence);

	return dev_name(hl_fence->hdev->dev);
}

static bool hl_fence_enable_signaling(struct dma_fence *fence)
{
	return true;
}

static void hl_fence_release(struct dma_fence *fence)
{
	struct hl_dma_fence *hl_fence =
		container_of(fence, struct hl_dma_fence, base_fence);

	kfree_rcu(hl_fence, base_fence.rcu);
}

static const struct dma_fence_ops hl_fence_ops = {
	.get_driver_name = hl_fence_get_driver_name,
	.get_timeline_name = hl_fence_get_timeline_name,
	.enable_signaling = hl_fence_enable_signaling,
	.wait = dma_fence_default_wait,
	.release = hl_fence_release
};

static void cs_get(struct hl_cs *cs)
{
	kref_get(&cs->refcount);
}

static int cs_get_unless_zero(struct hl_cs *cs)
{
	return kref_get_unless_zero(&cs->refcount);
}

static void cs_put(struct hl_cs *cs)
{
	kref_put(&cs->refcount, cs_do_release);
}

/*
 * cs_parser - parse the user command submission
 *
 * @hpriv	: pointer to the private data of the fd
 * @job        : pointer to the job that holds the command submission info
 *
 * The function parses the command submission of the user. It calls the
 * ASIC specific parser, which returns a list of memory blocks to send
 * to the device as different command buffers
 *
 */
static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
{
	struct hl_device *hdev = hpriv->hdev;
	struct hl_cs_parser parser;
	int rc;

	parser.ctx_id = job->cs->ctx->asid;
	parser.cs_sequence = job->cs->sequence;
	parser.job_id = job->id;

	parser.hw_queue_id = job->hw_queue_id;
	parser.job_userptr_list = &job->userptr_list;
	parser.patched_cb = NULL;
	parser.user_cb = job->user_cb;
	parser.user_cb_size = job->user_cb_size;
	parser.ext_queue = job->ext_queue;
	job->patched_cb = NULL;

	rc = hdev->asic_funcs->cs_parser(hdev, &parser);
	if (job->ext_queue) {
		if (!rc) {
			job->patched_cb = parser.patched_cb;
			job->job_cb_size = parser.patched_cb_size;

			spin_lock(&job->patched_cb->lock);
			job->patched_cb->cs_cnt++;
			spin_unlock(&job->patched_cb->lock);
		}

		/*
		 * Whether the parsing worked or not, we don't need the
		 * original CB anymore because it was already parsed and
		 * won't be accessed again for this CS
		 */
		spin_lock(&job->user_cb->lock);
		job->user_cb->cs_cnt--;
		spin_unlock(&job->user_cb->lock);
		hl_cb_put(job->user_cb);
		job->user_cb = NULL;
	}

	return rc;
}

static void free_job(struct hl_device *hdev, struct hl_cs_job *job)
{
	struct hl_cs *cs = job->cs;

	if (job->ext_queue) {
		hl_userptr_delete_list(hdev, &job->userptr_list);

		/*
		 * We might arrive here from rollback and patched CB wasn't
		 * created, so we need to check it's not NULL
		 */
		if (job->patched_cb) {
			spin_lock(&job->patched_cb->lock);
			job->patched_cb->cs_cnt--;
			spin_unlock(&job->patched_cb->lock);

			hl_cb_put(job->patched_cb);
		}
	}

	/*
	 * This is the only place where there can be multiple threads
	 * modifying the list at the same time
	 */
	spin_lock(&cs->job_lock);
	list_del(&job->cs_node);
	spin_unlock(&cs->job_lock);

	hl_debugfs_remove_job(hdev, job);

	if (job->ext_queue)
		cs_put(cs);

	kfree(job);
}

static void cs_do_release(struct kref *ref)
{
	struct hl_cs *cs = container_of(ref, struct hl_cs,
						refcount);
	struct hl_device *hdev = cs->ctx->hdev;
	struct hl_cs_job *job, *tmp;

	cs->completed = true;

	/*
	 * Although if we reached here it means that all external jobs have
	 * finished, because each one of them took refcnt to CS, we still
	 * need to go over the internal jobs and free them. Otherwise, we
	 * will have leaked memory and what's worse, the CS object (and
	 * potentially the CTX object) could be released, while the JOB
	 * still holds a pointer to them (but no reference).
	 */
	list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
		free_job(hdev, job);

	/* We also need to update CI for internal queues */
	if (cs->submitted) {
		int cs_cnt = atomic_dec_return(&hdev->cs_active_cnt);

		WARN_ONCE((cs_cnt < 0),
			"hl%d: error in CS active cnt %d\n",
			hdev->id, cs_cnt);

		hl_int_hw_queue_update_ci(cs);

		spin_lock(&hdev->hw_queues_mirror_lock);
		/* remove CS from hw_queues mirror list */
		list_del_init(&cs->mirror_node);
		spin_unlock(&hdev->hw_queues_mirror_lock);

		/*
		 * Don't cancel TDR in case this CS was timedout because we
		 * might be running from the TDR context
		 */
		if ((!cs->timedout) &&
			(hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT)) {
			struct hl_cs *next;

			if (cs->tdr_active)
				cancel_delayed_work_sync(&cs->work_tdr);

			spin_lock(&hdev->hw_queues_mirror_lock);

			/* queue TDR for next CS */
			next = list_first_entry_or_null(
					&hdev->hw_queues_mirror_list,
					struct hl_cs, mirror_node);

			if ((next) && (!next->tdr_active)) {
				next->tdr_active = true;
				schedule_delayed_work(&next->work_tdr,
							hdev->timeout_jiffies);
			}

			spin_unlock(&hdev->hw_queues_mirror_lock);
		}
	}

	/*
	 * Must be called before hl_ctx_put because inside we use ctx to get
	 * the device
	 */
	hl_debugfs_remove_cs(cs);

	hl_ctx_put(cs->ctx);

	if (cs->timedout)
		dma_fence_set_error(cs->fence, -ETIMEDOUT);
	else if (cs->aborted)
		dma_fence_set_error(cs->fence, -EIO);

	dma_fence_signal(cs->fence);
	dma_fence_put(cs->fence);

	kfree(cs);
}

static void cs_timedout(struct work_struct *work)
{
	struct hl_device *hdev;
	int ctx_asid, rc;
	struct hl_cs *cs = container_of(work, struct hl_cs,
						 work_tdr.work);
	rc = cs_get_unless_zero(cs);
	if (!rc)
		return;

	if ((!cs->submitted) || (cs->completed)) {
		cs_put(cs);
		return;
	}

	/* Mark the CS is timed out so we won't try to cancel its TDR */
	cs->timedout = true;

	hdev = cs->ctx->hdev;
	ctx_asid = cs->ctx->asid;

	/* TODO: add information about last signaled seq and last emitted seq */
	dev_err(hdev->dev, "User %d command submission %llu got stuck!\n",
		ctx_asid, cs->sequence);

	cs_put(cs);

	if (hdev->reset_on_lockup)
		hl_device_reset(hdev, false, false);
}

static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
			struct hl_cs **cs_new)
{
	struct hl_dma_fence *fence;
	struct dma_fence *other = NULL;
	struct hl_cs *cs;
	int rc;

	cs = kzalloc(sizeof(*cs), GFP_ATOMIC);
	if (!cs)
		return -ENOMEM;

	cs->ctx = ctx;
	cs->submitted = false;
	cs->completed = false;
	INIT_LIST_HEAD(&cs->job_list);
	INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
	kref_init(&cs->refcount);
	spin_lock_init(&cs->job_lock);

	fence = kmalloc(sizeof(*fence), GFP_ATOMIC);
	if (!fence) {
		rc = -ENOMEM;
		goto free_cs;
	}

	fence->hdev = hdev;
	spin_lock_init(&fence->lock);
	cs->fence = &fence->base_fence;

	spin_lock(&ctx->cs_lock);

	fence->cs_seq = ctx->cs_sequence;
	other = ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)];
	if ((other) && (!dma_fence_is_signaled(other))) {
		spin_unlock(&ctx->cs_lock);
		rc = -EAGAIN;
		goto free_fence;
	}

	dma_fence_init(&fence->base_fence, &hl_fence_ops, &fence->lock,
			ctx->asid, ctx->cs_sequence);

	cs->sequence = fence->cs_seq;

	ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)] =
							&fence->base_fence;
	ctx->cs_sequence++;

	dma_fence_get(&fence->base_fence);

	dma_fence_put(other);

	spin_unlock(&ctx->cs_lock);

	*cs_new = cs;

	return 0;

free_fence:
	kfree(fence);
free_cs:
	kfree(cs);
	return rc;
}

static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)
{
	struct hl_cs_job *job, *tmp;

	list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
		free_job(hdev, job);
}

void hl_cs_rollback_all(struct hl_device *hdev)
{
	struct hl_cs *cs, *tmp;

	/* flush all completions */
	flush_workqueue(hdev->cq_wq);

	/* Make sure we don't have leftovers in the H/W queues mirror list */
	list_for_each_entry_safe(cs, tmp, &hdev->hw_queues_mirror_list,
				mirror_node) {
		cs_get(cs);
		cs->aborted = true;
		dev_warn_ratelimited(hdev->dev, "Killing CS %d.%llu\n",
					cs->ctx->asid, cs->sequence);
		cs_rollback(hdev, cs);
		cs_put(cs);
	}
}

static void job_wq_completion(struct work_struct *work)
{
	struct hl_cs_job *job = container_of(work, struct hl_cs_job,
						finish_work);
	struct hl_cs *cs = job->cs;
	struct hl_device *hdev = cs->ctx->hdev;

	/* job is no longer needed */
	free_job(hdev, job);
}

static struct hl_cb *validate_queue_index(struct hl_device *hdev,
					struct hl_cb_mgr *cb_mgr,
					struct hl_cs_chunk *chunk,
					bool *ext_queue)
{
	struct asic_fixed_properties *asic = &hdev->asic_prop;
	struct hw_queue_properties *hw_queue_prop;
	u32 cb_handle;
	struct hl_cb *cb;

	/* Assume external queue */
	*ext_queue = true;

	hw_queue_prop = &asic->hw_queues_props[chunk->queue_index];

	if ((chunk->queue_index >= HL_MAX_QUEUES) ||
			(hw_queue_prop->type == QUEUE_TYPE_NA)) {
		dev_err(hdev->dev, "Queue index %d is invalid\n",
			chunk->queue_index);
		return NULL;
	}

	if (hw_queue_prop->kmd_only) {
		dev_err(hdev->dev, "Queue index %d is restricted for KMD\n",
			chunk->queue_index);
		return NULL;
	} else if (hw_queue_prop->type == QUEUE_TYPE_INT) {
		*ext_queue = false;
		return (struct hl_cb *) (uintptr_t) chunk->cb_handle;
	}

	/* Retrieve CB object */
	cb_handle = (u32) (chunk->cb_handle >> PAGE_SHIFT);

	cb = hl_cb_get(hdev, cb_mgr, cb_handle);
	if (!cb) {
		dev_err(hdev->dev, "CB handle 0x%x invalid\n", cb_handle);
		return NULL;
	}

	if ((chunk->cb_size < 8) || (chunk->cb_size > cb->size)) {
		dev_err(hdev->dev, "CB size %u invalid\n", chunk->cb_size);
		goto release_cb;
	}

	spin_lock(&cb->lock);
	cb->cs_cnt++;
	spin_unlock(&cb->lock);

	return cb;

release_cb:
	hl_cb_put(cb);
	return NULL;
}

struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev, bool ext_queue)
{
	struct hl_cs_job *job;

	job = kzalloc(sizeof(*job), GFP_ATOMIC);
	if (!job)
		return NULL;

	job->ext_queue = ext_queue;

	if (job->ext_queue) {
		INIT_LIST_HEAD(&job->userptr_list);
		INIT_WORK(&job->finish_work, job_wq_completion);
	}

	return job;
}

static int _hl_cs_ioctl(struct hl_fpriv *hpriv, void __user *chunks,
			u32 num_chunks, u64 *cs_seq)
{
	struct hl_device *hdev = hpriv->hdev;
	struct hl_cs_chunk *cs_chunk_array;
	struct hl_cs_job *job;
	struct hl_cs *cs;
	struct hl_cb *cb;
	bool ext_queue_present = false;
	u32 size_to_copy;
	int rc, i, parse_cnt;

	*cs_seq = ULLONG_MAX;

	if (num_chunks > HL_MAX_JOBS_PER_CS) {
		dev_err(hdev->dev,
			"Number of chunks can NOT be larger than %d\n",
			HL_MAX_JOBS_PER_CS);
		rc = -EINVAL;
		goto out;
	}

	cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
					GFP_ATOMIC);
	if (!cs_chunk_array) {
		rc = -ENOMEM;
		goto out;
	}

	size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
	if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
		dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
		rc = -EFAULT;
		goto free_cs_chunk_array;
	}

	/* increment refcnt for context */
	hl_ctx_get(hdev, hpriv->ctx);

	rc = allocate_cs(hdev, hpriv->ctx, &cs);
	if (rc) {
		hl_ctx_put(hpriv->ctx);
		goto free_cs_chunk_array;
	}

	*cs_seq = cs->sequence;

	hl_debugfs_add_cs(cs);

	/* Validate ALL the CS chunks before submitting the CS */
	for (i = 0, parse_cnt = 0 ; i < num_chunks ; i++, parse_cnt++) {
		struct hl_cs_chunk *chunk = &cs_chunk_array[i];
		bool ext_queue;

		cb = validate_queue_index(hdev, &hpriv->cb_mgr, chunk,
					&ext_queue);
		if (ext_queue) {
			ext_queue_present = true;
			if (!cb) {
				rc = -EINVAL;
				goto free_cs_object;
			}
		}

		job = hl_cs_allocate_job(hdev, ext_queue);
		if (!job) {
			dev_err(hdev->dev, "Failed to allocate a new job\n");
			rc = -ENOMEM;
			if (ext_queue)
				goto release_cb;
			else
				goto free_cs_object;
		}

		job->id = i + 1;
		job->cs = cs;
		job->user_cb = cb;
		job->user_cb_size = chunk->cb_size;
		if (job->ext_queue)
			job->job_cb_size = cb->size;
		else
			job->job_cb_size = chunk->cb_size;
		job->hw_queue_id = chunk->queue_index;

		cs->jobs_in_queue_cnt[job->hw_queue_id]++;

		list_add_tail(&job->cs_node, &cs->job_list);

		/*
		 * Increment CS reference. When CS reference is 0, CS is
		 * done and can be signaled to user and free all its resources
		 * Only increment for JOB on external queues, because only
		 * for those JOBs we get completion
		 */
		if (job->ext_queue)
			cs_get(cs);

		hl_debugfs_add_job(hdev, job);

		rc = cs_parser(hpriv, job);
		if (rc) {
			dev_err(hdev->dev,
				"Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
				cs->ctx->asid, cs->sequence, job->id, rc);
			goto free_cs_object;
		}
	}

	if (!ext_queue_present) {
		dev_err(hdev->dev,
			"Reject CS %d.%llu because no external queues jobs\n",
			cs->ctx->asid, cs->sequence);
		rc = -EINVAL;
		goto free_cs_object;
	}

	rc = hl_hw_queue_schedule_cs(cs);
	if (rc) {
		dev_err(hdev->dev,
			"Failed to submit CS %d.%llu to H/W queues, error %d\n",
			cs->ctx->asid, cs->sequence, rc);
		goto free_cs_object;
	}

	rc = HL_CS_STATUS_SUCCESS;
	goto put_cs;

release_cb:
	spin_lock(&cb->lock);
	cb->cs_cnt--;
	spin_unlock(&cb->lock);
	hl_cb_put(cb);
free_cs_object:
	cs_rollback(hdev, cs);
	*cs_seq = ULLONG_MAX;
	/* The path below is both for good and erroneous exits */
put_cs:
	/* We finished with the CS in this function, so put the ref */
	cs_put(cs);
free_cs_chunk_array:
	kfree(cs_chunk_array);
out:
	return rc;
}

int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
{
	struct hl_device *hdev = hpriv->hdev;
	union hl_cs_args *args = data;
	struct hl_ctx *ctx = hpriv->ctx;
	void __user *chunks;
	u32 num_chunks;
	u64 cs_seq = ULONG_MAX;
	int rc, do_ctx_switch;
	bool need_soft_reset = false;

	if (hl_device_disabled_or_in_reset(hdev)) {
		dev_warn_ratelimited(hdev->dev,
			"Device is %s. Can't submit new CS\n",
			atomic_read(&hdev->in_reset) ? "in_reset" : "disabled");
		rc = -EBUSY;
		goto out;
	}

	do_ctx_switch = atomic_cmpxchg(&ctx->thread_ctx_switch_token, 1, 0);

	if (do_ctx_switch || (args->in.cs_flags & HL_CS_FLAGS_FORCE_RESTORE)) {
		long ret;

		chunks = (void __user *)(uintptr_t)args->in.chunks_restore;
		num_chunks = args->in.num_chunks_restore;

		mutex_lock(&hpriv->restore_phase_mutex);

		if (do_ctx_switch) {
			rc = hdev->asic_funcs->context_switch(hdev, ctx->asid);
			if (rc) {
				dev_err_ratelimited(hdev->dev,
					"Failed to switch to context %d, rejecting CS! %d\n",
					ctx->asid, rc);
				/*
				 * If we timedout, or if the device is not IDLE
				 * while we want to do context-switch (-EBUSY),
				 * we need to soft-reset because QMAN is
				 * probably stuck. However, we can't call to
				 * reset here directly because of deadlock, so
				 * need to do it at the very end of this
				 * function
				 */
				if ((rc == -ETIMEDOUT) || (rc == -EBUSY))
					need_soft_reset = true;
				mutex_unlock(&hpriv->restore_phase_mutex);
				goto out;
			}
		}

		hdev->asic_funcs->restore_phase_topology(hdev);

		if (num_chunks == 0) {
			dev_dbg(hdev->dev,
			"Need to run restore phase but restore CS is empty\n");
			rc = 0;
		} else {
			rc = _hl_cs_ioctl(hpriv, chunks, num_chunks,
						&cs_seq);
		}

		mutex_unlock(&hpriv->restore_phase_mutex);

		if (rc) {
			dev_err(hdev->dev,
				"Failed to submit restore CS for context %d (%d)\n",
				ctx->asid, rc);
			goto out;
		}

		/* Need to wait for restore completion before execution phase */
		if (num_chunks > 0) {
			ret = _hl_cs_wait_ioctl(hdev, ctx,
					jiffies_to_usecs(hdev->timeout_jiffies),
					cs_seq);
			if (ret <= 0) {
				dev_err(hdev->dev,
					"Restore CS for context %d failed to complete %ld\n",
					ctx->asid, ret);
				rc = -ENOEXEC;
				goto out;
			}
		}

		ctx->thread_ctx_switch_wait_token = 1;
	} else if (!ctx->thread_ctx_switch_wait_token) {
		u32 tmp;

		rc = hl_poll_timeout_memory(hdev,
			&ctx->thread_ctx_switch_wait_token, tmp, (tmp == 1),
			100, jiffies_to_usecs(hdev->timeout_jiffies));

		if (rc == -ETIMEDOUT) {
			dev_err(hdev->dev,
				"context switch phase timeout (%d)\n", tmp);
			goto out;
		}
	}

	chunks = (void __user *)(uintptr_t)args->in.chunks_execute;
	num_chunks = args->in.num_chunks_execute;

	if (num_chunks == 0) {
		dev_err(hdev->dev,
			"Got execute CS with 0 chunks, context %d\n",
			ctx->asid);
		rc = -EINVAL;
		goto out;
	}

	rc = _hl_cs_ioctl(hpriv, chunks, num_chunks, &cs_seq);

out:
	if (rc != -EAGAIN) {
		memset(args, 0, sizeof(*args));
		args->out.status = rc;
		args->out.seq = cs_seq;
	}

	if (((rc == -ETIMEDOUT) || (rc == -EBUSY)) && (need_soft_reset))
		hl_device_reset(hdev, false, false);

	return rc;
}

static long _hl_cs_wait_ioctl(struct hl_device *hdev,
		struct hl_ctx *ctx, u64 timeout_us, u64 seq)
{
	struct dma_fence *fence;
	unsigned long timeout;
	long rc;

	if (timeout_us == MAX_SCHEDULE_TIMEOUT)
		timeout = timeout_us;
	else
		timeout = usecs_to_jiffies(timeout_us);

	hl_ctx_get(hdev, ctx);

	fence = hl_ctx_get_fence(ctx, seq);
	if (IS_ERR(fence)) {
		rc = PTR_ERR(fence);
	} else if (fence) {
		rc = dma_fence_wait_timeout(fence, true, timeout);
		if (fence->error == -ETIMEDOUT)
			rc = -ETIMEDOUT;
		else if (fence->error == -EIO)
			rc = -EIO;
		dma_fence_put(fence);
	} else
		rc = 1;

	hl_ctx_put(ctx);

	return rc;
}

int hl_cs_wait_ioctl(struct hl_fpriv *hpriv, void *data)
{
	struct hl_device *hdev = hpriv->hdev;
	union hl_wait_cs_args *args = data;
	u64 seq = args->in.seq;
	long rc;

	rc = _hl_cs_wait_ioctl(hdev, hpriv->ctx, args->in.timeout_us, seq);

	memset(args, 0, sizeof(*args));

	if (rc < 0) {
		dev_err(hdev->dev, "Error %ld on waiting for CS handle %llu\n",
			rc, seq);
		if (rc == -ERESTARTSYS) {
			args->out.status = HL_WAIT_CS_STATUS_INTERRUPTED;
			rc = -EINTR;
		} else if (rc == -ETIMEDOUT) {
			args->out.status = HL_WAIT_CS_STATUS_TIMEDOUT;
		} else if (rc == -EIO) {
			args->out.status = HL_WAIT_CS_STATUS_ABORTED;
		}
		return rc;
	}

	if (rc == 0)
		args->out.status = HL_WAIT_CS_STATUS_BUSY;
	else
		args->out.status = HL_WAIT_CS_STATUS_COMPLETED;

	return 0;
}