summaryrefslogtreecommitdiffstats
path: root/drivers/accel/habanalabs/common/hw_queue.c
blob: d0087c0ec48c9f5758a2eb8614ae31efb27ce9fa (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
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
// SPDX-License-Identifier: GPL-2.0

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

#include "habanalabs.h"

#include <linux/slab.h>

/*
 * hl_queue_add_ptr - add to pi or ci and checks if it wraps around
 *
 * @ptr: the current pi/ci value
 * @val: the amount to add
 *
 * Add val to ptr. It can go until twice the queue length.
 */
inline u32 hl_hw_queue_add_ptr(u32 ptr, u16 val)
{
	ptr += val;
	ptr &= ((HL_QUEUE_LENGTH << 1) - 1);
	return ptr;
}
static inline int queue_ci_get(atomic_t *ci, u32 queue_len)
{
	return atomic_read(ci) & ((queue_len << 1) - 1);
}

static inline int queue_free_slots(struct hl_hw_queue *q, u32 queue_len)
{
	int delta = (q->pi - queue_ci_get(&q->ci, queue_len));

	if (delta >= 0)
		return (queue_len - delta);
	else
		return (abs(delta) - queue_len);
}

void hl_hw_queue_update_ci(struct hl_cs *cs)
{
	struct hl_device *hdev = cs->ctx->hdev;
	struct hl_hw_queue *q;
	int i;

	if (hdev->disabled)
		return;

	q = &hdev->kernel_queues[0];

	/* There are no internal queues if H/W queues are being used */
	if (!hdev->asic_prop.max_queues || q->queue_type == QUEUE_TYPE_HW)
		return;

	/* We must increment CI for every queue that will never get a
	 * completion, there are 2 scenarios this can happen:
	 * 1. All queues of a non completion CS will never get a completion.
	 * 2. Internal queues never gets completion.
	 */
	for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
		if (!cs_needs_completion(cs) || q->queue_type == QUEUE_TYPE_INT)
			atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
	}
}

/*
 * hl_hw_queue_submit_bd() - Submit a buffer descriptor to an external or a
 *                                H/W queue.
 * @hdev: pointer to habanalabs device structure
 * @q: pointer to habanalabs queue structure
 * @ctl: BD's control word
 * @len: BD's length
 * @ptr: BD's pointer
 *
 * This function assumes there is enough space on the queue to submit a new
 * BD to it. It initializes the next BD and calls the device specific
 * function to set the pi (and doorbell)
 *
 * This function must be called when the scheduler mutex is taken
 *
 */
void hl_hw_queue_submit_bd(struct hl_device *hdev, struct hl_hw_queue *q,
		u32 ctl, u32 len, u64 ptr)
{
	struct hl_bd *bd;

	bd = q->kernel_address;
	bd += hl_pi_2_offset(q->pi);
	bd->ctl = cpu_to_le32(ctl);
	bd->len = cpu_to_le32(len);
	bd->ptr = cpu_to_le64(ptr);

	q->pi = hl_queue_inc_ptr(q->pi);
	hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
}

/*
 * ext_queue_sanity_checks - perform some sanity checks on external queue
 *
 * @hdev              : pointer to hl_device structure
 * @q                 :	pointer to hl_hw_queue structure
 * @num_of_entries    : how many entries to check for space
 * @reserve_cq_entry  :	whether to reserve an entry in the cq
 *
 * H/W queues spinlock should be taken before calling this function
 *
 * Perform the following:
 * - Make sure we have enough space in the h/w queue
 * - Make sure we have enough space in the completion queue
 * - Reserve space in the completion queue (needs to be reversed if there
 *   is a failure down the road before the actual submission of work). Only
 *   do this action if reserve_cq_entry is true
 *
 */
static int ext_queue_sanity_checks(struct hl_device *hdev,
				struct hl_hw_queue *q, int num_of_entries,
				bool reserve_cq_entry)
{
	atomic_t *free_slots =
			&hdev->completion_queue[q->cq_id].free_slots_cnt;
	int free_slots_cnt;

	/* Check we have enough space in the queue */
	free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);

	if (free_slots_cnt < num_of_entries) {
		dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
			q->hw_queue_id, num_of_entries);
		return -EAGAIN;
	}

	if (reserve_cq_entry) {
		/*
		 * Check we have enough space in the completion queue
		 * Add -1 to counter (decrement) unless counter was already 0
		 * In that case, CQ is full so we can't submit a new CB because
		 * we won't get ack on its completion
		 * atomic_add_unless will return 0 if counter was already 0
		 */
		if (atomic_add_negative(num_of_entries * -1, free_slots)) {
			dev_dbg(hdev->dev, "No space for %d on CQ %d\n",
				num_of_entries, q->hw_queue_id);
			atomic_add(num_of_entries, free_slots);
			return -EAGAIN;
		}
	}

	return 0;
}

/*
 * int_queue_sanity_checks - perform some sanity checks on internal queue
 *
 * @hdev              : pointer to hl_device structure
 * @q                 :	pointer to hl_hw_queue structure
 * @num_of_entries    : how many entries to check for space
 *
 * H/W queues spinlock should be taken before calling this function
 *
 * Perform the following:
 * - Make sure we have enough space in the h/w queue
 *
 */
static int int_queue_sanity_checks(struct hl_device *hdev,
					struct hl_hw_queue *q,
					int num_of_entries)
{
	int free_slots_cnt;

	if (num_of_entries > q->int_queue_len) {
		dev_err(hdev->dev,
			"Cannot populate queue %u with %u jobs\n",
			q->hw_queue_id, num_of_entries);
		return -ENOMEM;
	}

	/* Check we have enough space in the queue */
	free_slots_cnt = queue_free_slots(q, q->int_queue_len);

	if (free_slots_cnt < num_of_entries) {
		dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
			q->hw_queue_id, num_of_entries);
		return -EAGAIN;
	}

	return 0;
}

/*
 * hw_queue_sanity_checks() - Make sure we have enough space in the h/w queue
 * @hdev: Pointer to hl_device structure.
 * @q: Pointer to hl_hw_queue structure.
 * @num_of_entries: How many entries to check for space.
 *
 * Notice: We do not reserve queue entries so this function mustn't be called
 *         more than once per CS for the same queue
 *
 */
static int hw_queue_sanity_checks(struct hl_device *hdev, struct hl_hw_queue *q,
					int num_of_entries)
{
	int free_slots_cnt;

	/* Check we have enough space in the queue */
	free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);

	if (free_slots_cnt < num_of_entries) {
		dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
			q->hw_queue_id, num_of_entries);
		return -EAGAIN;
	}

	return 0;
}

/*
 * hl_hw_queue_send_cb_no_cmpl - send a single CB (not a JOB) without completion
 *
 * @hdev: pointer to hl_device structure
 * @hw_queue_id: Queue's type
 * @cb_size: size of CB
 * @cb_ptr: pointer to CB location
 *
 * This function sends a single CB, that must NOT generate a completion entry.
 * Sending CPU messages can be done instead via 'hl_hw_queue_submit_bd()'
 */
int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
				u32 cb_size, u64 cb_ptr)
{
	struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
	int rc = 0;

	hdev->asic_funcs->hw_queues_lock(hdev);

	if (hdev->disabled) {
		rc = -EPERM;
		goto out;
	}

	/*
	 * hl_hw_queue_send_cb_no_cmpl() is called for queues of a H/W queue
	 * type only on init phase, when the queues are empty and being tested,
	 * so there is no need for sanity checks.
	 */
	if (q->queue_type != QUEUE_TYPE_HW) {
		rc = ext_queue_sanity_checks(hdev, q, 1, false);
		if (rc)
			goto out;
	}

	hl_hw_queue_submit_bd(hdev, q, 0, cb_size, cb_ptr);

out:
	hdev->asic_funcs->hw_queues_unlock(hdev);

	return rc;
}

/*
 * ext_queue_schedule_job - submit a JOB to an external queue
 *
 * @job: pointer to the job that needs to be submitted to the queue
 *
 * This function must be called when the scheduler mutex is taken
 *
 */
static void ext_queue_schedule_job(struct hl_cs_job *job)
{
	struct hl_device *hdev = job->cs->ctx->hdev;
	struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
	struct hl_cq_entry cq_pkt;
	struct hl_cq *cq;
	u64 cq_addr;
	struct hl_cb *cb;
	u32 ctl;
	u32 len;
	u64 ptr;

	/*
	 * Update the JOB ID inside the BD CTL so the device would know what
	 * to write in the completion queue
	 */
	ctl = ((q->pi << BD_CTL_SHADOW_INDEX_SHIFT) & BD_CTL_SHADOW_INDEX_MASK);

	cb = job->patched_cb;
	len = job->job_cb_size;
	ptr = cb->bus_address;

	/* Skip completion flow in case this is a non completion CS */
	if (!cs_needs_completion(job->cs))
		goto submit_bd;

	cq_pkt.data = cpu_to_le32(
			((q->pi << CQ_ENTRY_SHADOW_INDEX_SHIFT)
				& CQ_ENTRY_SHADOW_INDEX_MASK) |
			FIELD_PREP(CQ_ENTRY_SHADOW_INDEX_VALID_MASK, 1) |
			FIELD_PREP(CQ_ENTRY_READY_MASK, 1));

	/*
	 * No need to protect pi_offset because scheduling to the
	 * H/W queues is done under the scheduler mutex
	 *
	 * No need to check if CQ is full because it was already
	 * checked in ext_queue_sanity_checks
	 */
	cq = &hdev->completion_queue[q->cq_id];
	cq_addr = cq->bus_address + cq->pi * sizeof(struct hl_cq_entry);

	hdev->asic_funcs->add_end_of_cb_packets(hdev, cb->kernel_address, len,
						job->user_cb_size,
						cq_addr,
						le32_to_cpu(cq_pkt.data),
						q->msi_vec,
						job->contains_dma_pkt);

	q->shadow_queue[hl_pi_2_offset(q->pi)] = job;

	cq->pi = hl_cq_inc_ptr(cq->pi);

submit_bd:
	hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
}

/*
 * int_queue_schedule_job - submit a JOB to an internal queue
 *
 * @job: pointer to the job that needs to be submitted to the queue
 *
 * This function must be called when the scheduler mutex is taken
 *
 */
static void int_queue_schedule_job(struct hl_cs_job *job)
{
	struct hl_device *hdev = job->cs->ctx->hdev;
	struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
	struct hl_bd bd;
	__le64 *pi;

	bd.ctl = 0;
	bd.len = cpu_to_le32(job->job_cb_size);

	if (job->is_kernel_allocated_cb)
		/* bus_address is actually a mmu mapped address
		 * allocated from an internal pool
		 */
		bd.ptr = cpu_to_le64(job->user_cb->bus_address);
	else
		bd.ptr = cpu_to_le64((u64) (uintptr_t) job->user_cb);

	pi = q->kernel_address + (q->pi & (q->int_queue_len - 1)) * sizeof(bd);

	q->pi++;
	q->pi &= ((q->int_queue_len << 1) - 1);

	hdev->asic_funcs->pqe_write(hdev, pi, &bd);

	hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
}

/*
 * hw_queue_schedule_job - submit a JOB to a H/W queue
 *
 * @job: pointer to the job that needs to be submitted to the queue
 *
 * This function must be called when the scheduler mutex is taken
 *
 */
static void hw_queue_schedule_job(struct hl_cs_job *job)
{
	struct hl_device *hdev = job->cs->ctx->hdev;
	struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
	u64 ptr;
	u32 offset, ctl, len;

	/*
	 * Upon PQE completion, COMP_DATA is used as the write data to the
	 * completion queue (QMAN HBW message), and COMP_OFFSET is used as the
	 * write address offset in the SM block (QMAN LBW message).
	 * The write address offset is calculated as "COMP_OFFSET << 2".
	 */
	offset = job->cs->sequence & (hdev->asic_prop.max_pending_cs - 1);
	ctl = ((offset << BD_CTL_COMP_OFFSET_SHIFT) & BD_CTL_COMP_OFFSET_MASK) |
		((q->pi << BD_CTL_COMP_DATA_SHIFT) & BD_CTL_COMP_DATA_MASK);

	len = job->job_cb_size;

	/*
	 * A patched CB is created only if a user CB was allocated by driver and
	 * MMU is disabled. If MMU is enabled, the user CB should be used
	 * instead. If the user CB wasn't allocated by driver, assume that it
	 * holds an address.
	 */
	if (job->patched_cb)
		ptr = job->patched_cb->bus_address;
	else if (job->is_kernel_allocated_cb)
		ptr = job->user_cb->bus_address;
	else
		ptr = (u64) (uintptr_t) job->user_cb;

	hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
}

static int init_signal_cs(struct hl_device *hdev,
		struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
{
	struct hl_sync_stream_properties *prop;
	struct hl_hw_sob *hw_sob;
	u32 q_idx;
	int rc = 0;

	q_idx = job->hw_queue_id;
	prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
	hw_sob = &prop->hw_sob[prop->curr_sob_offset];

	cs_cmpl->hw_sob = hw_sob;
	cs_cmpl->sob_val = prop->next_sob_val;

	dev_dbg(hdev->dev,
		"generate signal CB, sob_id: %d, sob val: %u, q_idx: %d, seq: %llu\n",
		cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx,
		cs_cmpl->cs_seq);

	/* we set an EB since we must make sure all oeprations are done
	 * when sending the signal
	 */
	hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb,
				cs_cmpl->hw_sob->sob_id, 0, true);

	rc = hl_cs_signal_sob_wraparound_handler(hdev, q_idx, &hw_sob, 1,
								false);

	job->cs->sob_addr_offset = hw_sob->sob_addr;
	job->cs->initial_sob_count = prop->next_sob_val - 1;

	return rc;
}

void hl_hw_queue_encaps_sig_set_sob_info(struct hl_device *hdev,
			struct hl_cs *cs, struct hl_cs_job *job,
			struct hl_cs_compl *cs_cmpl)
{
	struct hl_cs_encaps_sig_handle *handle = cs->encaps_sig_hdl;
	u32 offset = 0;

	cs_cmpl->hw_sob = handle->hw_sob;

	/* Note that encaps_sig_wait_offset was validated earlier in the flow
	 * for offset value which exceeds the max reserved signal count.
	 * always decrement 1 of the offset since when the user
	 * set offset 1 for example he mean to wait only for the first
	 * signal only, which will be pre_sob_val, and if he set offset 2
	 * then the value required is (pre_sob_val + 1) and so on...
	 * if user set wait offset to 0, then treat it as legacy wait cs,
	 * wait for the next signal.
	 */
	if (job->encaps_sig_wait_offset)
		offset = job->encaps_sig_wait_offset - 1;

	cs_cmpl->sob_val = handle->pre_sob_val + offset;
}

static int init_wait_cs(struct hl_device *hdev, struct hl_cs *cs,
		struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
{
	struct hl_gen_wait_properties wait_prop;
	struct hl_sync_stream_properties *prop;
	struct hl_cs_compl *signal_cs_cmpl;
	u32 q_idx;

	q_idx = job->hw_queue_id;
	prop = &hdev->kernel_queues[q_idx].sync_stream_prop;

	signal_cs_cmpl = container_of(cs->signal_fence,
					struct hl_cs_compl,
					base_fence);

	if (cs->encaps_signals) {
		/* use the encaps signal handle stored earlier in the flow
		 * and set the SOB information from the encaps
		 * signals handle
		 */
		hl_hw_queue_encaps_sig_set_sob_info(hdev, cs, job, cs_cmpl);

		dev_dbg(hdev->dev, "Wait for encaps signals handle, qidx(%u), CS sequence(%llu), sob val: 0x%x, offset: %u\n",
				cs->encaps_sig_hdl->q_idx,
				cs->encaps_sig_hdl->cs_seq,
				cs_cmpl->sob_val,
				job->encaps_sig_wait_offset);
	} else {
		/* Copy the SOB id and value of the signal CS */
		cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
		cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
	}

	/* check again if the signal cs already completed.
	 * if yes then don't send any wait cs since the hw_sob
	 * could be in reset already. if signal is not completed
	 * then get refcount to hw_sob to prevent resetting the sob
	 * while wait cs is not submitted.
	 * note that this check is protected by two locks,
	 * hw queue lock and completion object lock,
	 * and the same completion object lock also protects
	 * the hw_sob reset handler function.
	 * The hw_queue lock prevent out of sync of hw_sob
	 * refcount value, changed by signal/wait flows.
	 */
	spin_lock(&signal_cs_cmpl->lock);

	if (completion_done(&cs->signal_fence->completion)) {
		spin_unlock(&signal_cs_cmpl->lock);
		return -EINVAL;
	}

	kref_get(&cs_cmpl->hw_sob->kref);

	spin_unlock(&signal_cs_cmpl->lock);

	dev_dbg(hdev->dev,
		"generate wait CB, sob_id: %d, sob_val: 0x%x, mon_id: %d, q_idx: %d, seq: %llu\n",
		cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val,
		prop->base_mon_id, q_idx, cs->sequence);

	wait_prop.data = (void *) job->patched_cb;
	wait_prop.sob_base = cs_cmpl->hw_sob->sob_id;
	wait_prop.sob_mask = 0x1;
	wait_prop.sob_val = cs_cmpl->sob_val;
	wait_prop.mon_id = prop->base_mon_id;
	wait_prop.q_idx = q_idx;
	wait_prop.size = 0;

	hdev->asic_funcs->gen_wait_cb(hdev, &wait_prop);

	mb();
	hl_fence_put(cs->signal_fence);
	cs->signal_fence = NULL;

	return 0;
}

/*
 * init_signal_wait_cs - initialize a signal/wait CS
 * @cs: pointer to the signal/wait CS
 *
 * H/W queues spinlock should be taken before calling this function
 */
static int init_signal_wait_cs(struct hl_cs *cs)
{
	struct hl_ctx *ctx = cs->ctx;
	struct hl_device *hdev = ctx->hdev;
	struct hl_cs_job *job;
	struct hl_cs_compl *cs_cmpl =
			container_of(cs->fence, struct hl_cs_compl, base_fence);
	int rc = 0;

	/* There is only one job in a signal/wait CS */
	job = list_first_entry(&cs->job_list, struct hl_cs_job,
				cs_node);

	if (cs->type & CS_TYPE_SIGNAL)
		rc = init_signal_cs(hdev, job, cs_cmpl);
	else if (cs->type & CS_TYPE_WAIT)
		rc = init_wait_cs(hdev, cs, job, cs_cmpl);

	return rc;
}

static int encaps_sig_first_staged_cs_handler
			(struct hl_device *hdev, struct hl_cs *cs)
{
	struct hl_cs_compl *cs_cmpl =
			container_of(cs->fence,
					struct hl_cs_compl, base_fence);
	struct hl_cs_encaps_sig_handle *encaps_sig_hdl;
	struct hl_encaps_signals_mgr *mgr;
	int rc = 0;

	mgr = &cs->ctx->sig_mgr;

	spin_lock(&mgr->lock);
	encaps_sig_hdl = idr_find(&mgr->handles, cs->encaps_sig_hdl_id);
	if (encaps_sig_hdl) {
		/*
		 * Set handler CS sequence,
		 * the CS which contains the encapsulated signals.
		 */
		encaps_sig_hdl->cs_seq = cs->sequence;
		/* store the handle and set encaps signal indication,
		 * to be used later in cs_do_release to put the last
		 * reference to encaps signals handlers.
		 */
		cs_cmpl->encaps_signals = true;
		cs_cmpl->encaps_sig_hdl = encaps_sig_hdl;

		/* set hw_sob pointer in completion object
		 * since it's used in cs_do_release flow to put
		 * refcount to sob
		 */
		cs_cmpl->hw_sob = encaps_sig_hdl->hw_sob;
		cs_cmpl->sob_val = encaps_sig_hdl->pre_sob_val +
						encaps_sig_hdl->count;

		dev_dbg(hdev->dev, "CS seq (%llu) added to encaps signal handler id (%u), count(%u), qidx(%u), sob(%u), val(%u)\n",
				cs->sequence, encaps_sig_hdl->id,
				encaps_sig_hdl->count,
				encaps_sig_hdl->q_idx,
				cs_cmpl->hw_sob->sob_id,
				cs_cmpl->sob_val);

	} else {
		dev_err(hdev->dev, "encaps handle id(%u) wasn't found!\n",
				cs->encaps_sig_hdl_id);
		rc = -EINVAL;
	}

	spin_unlock(&mgr->lock);

	return rc;
}

/*
 * hl_hw_queue_schedule_cs - schedule a command submission
 * @cs: pointer to the CS
 */
int hl_hw_queue_schedule_cs(struct hl_cs *cs)
{
	enum hl_device_status status;
	struct hl_cs_counters_atomic *cntr;
	struct hl_ctx *ctx = cs->ctx;
	struct hl_device *hdev = ctx->hdev;
	struct hl_cs_job *job, *tmp;
	struct hl_hw_queue *q;
	int rc = 0, i, cq_cnt;
	bool first_entry;
	u32 max_queues;

	cntr = &hdev->aggregated_cs_counters;

	hdev->asic_funcs->hw_queues_lock(hdev);

	if (!hl_device_operational(hdev, &status)) {
		atomic64_inc(&cntr->device_in_reset_drop_cnt);
		atomic64_inc(&ctx->cs_counters.device_in_reset_drop_cnt);
		dev_err(hdev->dev,
			"device is %s, CS rejected!\n", hdev->status[status]);
		rc = -EPERM;
		goto out;
	}

	max_queues = hdev->asic_prop.max_queues;

	q = &hdev->kernel_queues[0];
	for (i = 0, cq_cnt = 0 ; i < max_queues ; i++, q++) {
		if (cs->jobs_in_queue_cnt[i]) {
			switch (q->queue_type) {
			case QUEUE_TYPE_EXT:
				rc = ext_queue_sanity_checks(hdev, q,
						cs->jobs_in_queue_cnt[i],
						cs_needs_completion(cs) ?
								true : false);
				break;
			case QUEUE_TYPE_INT:
				rc = int_queue_sanity_checks(hdev, q,
						cs->jobs_in_queue_cnt[i]);
				break;
			case QUEUE_TYPE_HW:
				rc = hw_queue_sanity_checks(hdev, q,
						cs->jobs_in_queue_cnt[i]);
				break;
			default:
				dev_err(hdev->dev, "Queue type %d is invalid\n",
					q->queue_type);
				rc = -EINVAL;
				break;
			}

			if (rc) {
				atomic64_inc(
					&ctx->cs_counters.queue_full_drop_cnt);
				atomic64_inc(&cntr->queue_full_drop_cnt);
				goto unroll_cq_resv;
			}

			if (q->queue_type == QUEUE_TYPE_EXT)
				cq_cnt++;
		}
	}

	if ((cs->type == CS_TYPE_SIGNAL) || (cs->type == CS_TYPE_WAIT)) {
		rc = init_signal_wait_cs(cs);
		if (rc)
			goto unroll_cq_resv;
	} else if (cs->type == CS_TYPE_COLLECTIVE_WAIT) {
		rc = hdev->asic_funcs->collective_wait_init_cs(cs);
		if (rc)
			goto unroll_cq_resv;
	}

	rc = hdev->asic_funcs->pre_schedule_cs(cs);
	if (rc) {
		dev_err(hdev->dev,
			"Failed in pre-submission operations of CS %d.%llu\n",
			ctx->asid, cs->sequence);
		goto unroll_cq_resv;
	}

	hdev->shadow_cs_queue[cs->sequence &
				(hdev->asic_prop.max_pending_cs - 1)] = cs;

	if (cs->encaps_signals && cs->staged_first) {
		rc = encaps_sig_first_staged_cs_handler(hdev, cs);
		if (rc)
			goto unroll_cq_resv;
	}

	spin_lock(&hdev->cs_mirror_lock);

	/* Verify staged CS exists and add to the staged list */
	if (cs->staged_cs && !cs->staged_first) {
		struct hl_cs *staged_cs;

		staged_cs = hl_staged_cs_find_first(hdev, cs->staged_sequence);
		if (!staged_cs) {
			dev_err(hdev->dev,
				"Cannot find staged submission sequence %llu",
				cs->staged_sequence);
			rc = -EINVAL;
			goto unlock_cs_mirror;
		}

		if (is_staged_cs_last_exists(hdev, staged_cs)) {
			dev_err(hdev->dev,
				"Staged submission sequence %llu already submitted",
				cs->staged_sequence);
			rc = -EINVAL;
			goto unlock_cs_mirror;
		}

		list_add_tail(&cs->staged_cs_node, &staged_cs->staged_cs_node);

		/* update stream map of the first CS */
		if (hdev->supports_wait_for_multi_cs)
			staged_cs->fence->stream_master_qid_map |=
					cs->fence->stream_master_qid_map;
	}

	list_add_tail(&cs->mirror_node, &hdev->cs_mirror_list);

	/* Queue TDR if the CS is the first entry and if timeout is wanted */
	first_entry = list_first_entry(&hdev->cs_mirror_list,
					struct hl_cs, mirror_node) == cs;
	if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
				first_entry && cs_needs_timeout(cs)) {
		cs->tdr_active = true;
		schedule_delayed_work(&cs->work_tdr, cs->timeout_jiffies);

	}

	spin_unlock(&hdev->cs_mirror_lock);

	list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
		switch (job->queue_type) {
		case QUEUE_TYPE_EXT:
			ext_queue_schedule_job(job);
			break;
		case QUEUE_TYPE_INT:
			int_queue_schedule_job(job);
			break;
		case QUEUE_TYPE_HW:
			hw_queue_schedule_job(job);
			break;
		default:
			break;
		}

	cs->submitted = true;

	goto out;

unlock_cs_mirror:
	spin_unlock(&hdev->cs_mirror_lock);
unroll_cq_resv:
	q = &hdev->kernel_queues[0];
	for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
		if ((q->queue_type == QUEUE_TYPE_EXT) &&
						(cs->jobs_in_queue_cnt[i])) {
			atomic_t *free_slots =
				&hdev->completion_queue[i].free_slots_cnt;
			atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
			cq_cnt--;
		}
	}

out:
	hdev->asic_funcs->hw_queues_unlock(hdev);

	return rc;
}

/*
 * hl_hw_queue_inc_ci_kernel - increment ci for kernel's queue
 *
 * @hdev: pointer to hl_device structure
 * @hw_queue_id: which queue to increment its ci
 */
void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
{
	struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];

	atomic_inc(&q->ci);
}

static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
					bool is_cpu_queue)
{
	void *p;
	int rc;

	if (is_cpu_queue)
		p = hl_cpu_accessible_dma_pool_alloc(hdev, HL_QUEUE_SIZE_IN_BYTES, &q->bus_address);
	else
		p = hl_asic_dma_alloc_coherent(hdev, HL_QUEUE_SIZE_IN_BYTES, &q->bus_address,
						GFP_KERNEL | __GFP_ZERO);
	if (!p)
		return -ENOMEM;

	q->kernel_address = p;

	q->shadow_queue = kmalloc_array(HL_QUEUE_LENGTH, sizeof(struct hl_cs_job *), GFP_KERNEL);
	if (!q->shadow_queue) {
		dev_err(hdev->dev,
			"Failed to allocate shadow queue for H/W queue %d\n",
			q->hw_queue_id);
		rc = -ENOMEM;
		goto free_queue;
	}

	/* Make sure read/write pointers are initialized to start of queue */
	atomic_set(&q->ci, 0);
	q->pi = 0;

	return 0;

free_queue:
	if (is_cpu_queue)
		hl_cpu_accessible_dma_pool_free(hdev, HL_QUEUE_SIZE_IN_BYTES, q->kernel_address);
	else
		hl_asic_dma_free_coherent(hdev, HL_QUEUE_SIZE_IN_BYTES, q->kernel_address,
						q->bus_address);

	return rc;
}

static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
{
	void *p;

	p = hdev->asic_funcs->get_int_queue_base(hdev, q->hw_queue_id,
					&q->bus_address, &q->int_queue_len);
	if (!p) {
		dev_err(hdev->dev,
			"Failed to get base address for internal queue %d\n",
			q->hw_queue_id);
		return -EFAULT;
	}

	q->kernel_address = p;
	q->pi = 0;
	atomic_set(&q->ci, 0);

	return 0;
}

static int cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
{
	return ext_and_cpu_queue_init(hdev, q, true);
}

static int ext_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
{
	return ext_and_cpu_queue_init(hdev, q, false);
}

static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
{
	void *p;

	p = hl_asic_dma_alloc_coherent(hdev, HL_QUEUE_SIZE_IN_BYTES, &q->bus_address,
					GFP_KERNEL | __GFP_ZERO);
	if (!p)
		return -ENOMEM;

	q->kernel_address = p;

	/* Make sure read/write pointers are initialized to start of queue */
	atomic_set(&q->ci, 0);
	q->pi = 0;

	return 0;
}

static void sync_stream_queue_init(struct hl_device *hdev, u32 q_idx)
{
	struct hl_sync_stream_properties *sync_stream_prop;
	struct asic_fixed_properties *prop = &hdev->asic_prop;
	struct hl_hw_sob *hw_sob;
	int sob, reserved_mon_idx, queue_idx;

	sync_stream_prop = &hdev->kernel_queues[q_idx].sync_stream_prop;

	/* We use 'collective_mon_idx' as a running index in order to reserve
	 * monitors for collective master/slave queues.
	 * collective master queue gets 2 reserved monitors
	 * collective slave queue gets 1 reserved monitor
	 */
	if (hdev->kernel_queues[q_idx].collective_mode ==
			HL_COLLECTIVE_MASTER) {
		reserved_mon_idx = hdev->collective_mon_idx;

		/* reserve the first monitor for collective master queue */
		sync_stream_prop->collective_mstr_mon_id[0] =
			prop->collective_first_mon + reserved_mon_idx;

		/* reserve the second monitor for collective master queue */
		sync_stream_prop->collective_mstr_mon_id[1] =
			prop->collective_first_mon + reserved_mon_idx + 1;

		hdev->collective_mon_idx += HL_COLLECTIVE_RSVD_MSTR_MONS;
	} else if (hdev->kernel_queues[q_idx].collective_mode ==
			HL_COLLECTIVE_SLAVE) {
		reserved_mon_idx = hdev->collective_mon_idx++;

		/* reserve a monitor for collective slave queue */
		sync_stream_prop->collective_slave_mon_id =
			prop->collective_first_mon + reserved_mon_idx;
	}

	if (!hdev->kernel_queues[q_idx].supports_sync_stream)
		return;

	queue_idx = hdev->sync_stream_queue_idx++;

	sync_stream_prop->base_sob_id = prop->sync_stream_first_sob +
			(queue_idx * HL_RSVD_SOBS);
	sync_stream_prop->base_mon_id = prop->sync_stream_first_mon +
			(queue_idx * HL_RSVD_MONS);
	sync_stream_prop->next_sob_val = 1;
	sync_stream_prop->curr_sob_offset = 0;

	for (sob = 0 ; sob < HL_RSVD_SOBS ; sob++) {
		hw_sob = &sync_stream_prop->hw_sob[sob];
		hw_sob->hdev = hdev;
		hw_sob->sob_id = sync_stream_prop->base_sob_id + sob;
		hw_sob->sob_addr =
			hdev->asic_funcs->get_sob_addr(hdev, hw_sob->sob_id);
		hw_sob->q_idx = q_idx;
		kref_init(&hw_sob->kref);
	}
}

static void sync_stream_queue_reset(struct hl_device *hdev, u32 q_idx)
{
	struct hl_sync_stream_properties *prop =
			&hdev->kernel_queues[q_idx].sync_stream_prop;

	/*
	 * In case we got here due to a stuck CS, the refcnt might be bigger
	 * than 1 and therefore we reset it.
	 */
	kref_init(&prop->hw_sob[prop->curr_sob_offset].kref);
	prop->curr_sob_offset = 0;
	prop->next_sob_val = 1;
}

/*
 * queue_init - main initialization function for H/W queue object
 *
 * @hdev: pointer to hl_device device structure
 * @q: pointer to hl_hw_queue queue structure
 * @hw_queue_id: The id of the H/W queue
 *
 * Allocate dma-able memory for the queue and initialize fields
 * Returns 0 on success
 */
static int queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
			u32 hw_queue_id)
{
	int rc;

	q->hw_queue_id = hw_queue_id;

	switch (q->queue_type) {
	case QUEUE_TYPE_EXT:
		rc = ext_queue_init(hdev, q);
		break;
	case QUEUE_TYPE_INT:
		rc = int_queue_init(hdev, q);
		break;
	case QUEUE_TYPE_CPU:
		rc = cpu_queue_init(hdev, q);
		break;
	case QUEUE_TYPE_HW:
		rc = hw_queue_init(hdev, q);
		break;
	case QUEUE_TYPE_NA:
		q->valid = 0;
		return 0;
	default:
		dev_crit(hdev->dev, "wrong queue type %d during init\n",
			q->queue_type);
		rc = -EINVAL;
		break;
	}

	sync_stream_queue_init(hdev, q->hw_queue_id);

	if (rc)
		return rc;

	q->valid = 1;

	return 0;
}

/*
 * hw_queue_fini - destroy queue
 *
 * @hdev: pointer to hl_device device structure
 * @q: pointer to hl_hw_queue queue structure
 *
 * Free the queue memory
 */
static void queue_fini(struct hl_device *hdev, struct hl_hw_queue *q)
{
	if (!q->valid)
		return;

	/*
	 * If we arrived here, there are no jobs waiting on this queue
	 * so we can safely remove it.
	 * This is because this function can only called when:
	 * 1. Either a context is deleted, which only can occur if all its
	 *    jobs were finished
	 * 2. A context wasn't able to be created due to failure or timeout,
	 *    which means there are no jobs on the queue yet
	 *
	 * The only exception are the queues of the kernel context, but
	 * if they are being destroyed, it means that the entire module is
	 * being removed. If the module is removed, it means there is no open
	 * user context. It also means that if a job was submitted by
	 * the kernel driver (e.g. context creation), the job itself was
	 * released by the kernel driver when a timeout occurred on its
	 * Completion. Thus, we don't need to release it again.
	 */

	if (q->queue_type == QUEUE_TYPE_INT)
		return;

	kfree(q->shadow_queue);

	if (q->queue_type == QUEUE_TYPE_CPU)
		hl_cpu_accessible_dma_pool_free(hdev, HL_QUEUE_SIZE_IN_BYTES, q->kernel_address);
	else
		hl_asic_dma_free_coherent(hdev, HL_QUEUE_SIZE_IN_BYTES, q->kernel_address,
						q->bus_address);
}

int hl_hw_queues_create(struct hl_device *hdev)
{
	struct asic_fixed_properties *asic = &hdev->asic_prop;
	struct hl_hw_queue *q;
	int i, rc, q_ready_cnt;

	hdev->kernel_queues = kcalloc(asic->max_queues,
				sizeof(*hdev->kernel_queues), GFP_KERNEL);

	if (!hdev->kernel_queues) {
		dev_err(hdev->dev, "Not enough memory for H/W queues\n");
		return -ENOMEM;
	}

	/* Initialize the H/W queues */
	for (i = 0, q_ready_cnt = 0, q = hdev->kernel_queues;
			i < asic->max_queues ; i++, q_ready_cnt++, q++) {

		q->queue_type = asic->hw_queues_props[i].type;
		q->supports_sync_stream =
				asic->hw_queues_props[i].supports_sync_stream;
		q->collective_mode = asic->hw_queues_props[i].collective_mode;
		rc = queue_init(hdev, q, i);
		if (rc) {
			dev_err(hdev->dev,
				"failed to initialize queue %d\n", i);
			goto release_queues;
		}
	}

	return 0;

release_queues:
	for (i = 0, q = hdev->kernel_queues ; i < q_ready_cnt ; i++, q++)
		queue_fini(hdev, q);

	kfree(hdev->kernel_queues);

	return rc;
}

void hl_hw_queues_destroy(struct hl_device *hdev)
{
	struct hl_hw_queue *q;
	u32 max_queues = hdev->asic_prop.max_queues;
	int i;

	for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++)
		queue_fini(hdev, q);

	kfree(hdev->kernel_queues);
}

void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
{
	struct hl_hw_queue *q;
	u32 max_queues = hdev->asic_prop.max_queues;
	int i;

	for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++) {
		if ((!q->valid) ||
			((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
			continue;
		q->pi = 0;
		atomic_set(&q->ci, 0);

		if (q->supports_sync_stream)
			sync_stream_queue_reset(hdev, q->hw_queue_id);
	}
}