aboutsummaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/i915/i915_request.c
blob: c0df71d7d0ff1499f53d9b98a27f7641fce87a29 (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
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
/*
 * Copyright © 2008-2015 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 */

#include <linux/dma-fence-array.h>
#include <linux/irq_work.h>
#include <linux/prefetch.h>
#include <linux/sched.h>
#include <linux/sched/clock.h>
#include <linux/sched/signal.h>

#include "gem/i915_gem_context.h"
#include "gt/intel_context.h"
#include "gt/intel_ring.h"
#include "gt/intel_rps.h"

#include "i915_active.h"
#include "i915_drv.h"
#include "i915_globals.h"
#include "i915_trace.h"
#include "intel_pm.h"

struct execute_cb {
	struct list_head link;
	struct irq_work work;
	struct i915_sw_fence *fence;
	void (*hook)(struct i915_request *rq, struct dma_fence *signal);
	struct i915_request *signal;
};

static struct i915_global_request {
	struct i915_global base;
	struct kmem_cache *slab_requests;
	struct kmem_cache *slab_execute_cbs;
} global;

static const char *i915_fence_get_driver_name(struct dma_fence *fence)
{
	return dev_name(to_request(fence)->i915->drm.dev);
}

static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
{
	const struct i915_gem_context *ctx;

	/*
	 * The timeline struct (as part of the ppgtt underneath a context)
	 * may be freed when the request is no longer in use by the GPU.
	 * We could extend the life of a context to beyond that of all
	 * fences, possibly keeping the hw resource around indefinitely,
	 * or we just give them a false name. Since
	 * dma_fence_ops.get_timeline_name is a debug feature, the occasional
	 * lie seems justifiable.
	 */
	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
		return "signaled";

	ctx = i915_request_gem_context(to_request(fence));
	if (!ctx)
		return "[" DRIVER_NAME "]";

	return ctx->name;
}

static bool i915_fence_signaled(struct dma_fence *fence)
{
	return i915_request_completed(to_request(fence));
}

static bool i915_fence_enable_signaling(struct dma_fence *fence)
{
	return i915_request_enable_breadcrumb(to_request(fence));
}

static signed long i915_fence_wait(struct dma_fence *fence,
				   bool interruptible,
				   signed long timeout)
{
	return i915_request_wait(to_request(fence),
				 interruptible | I915_WAIT_PRIORITY,
				 timeout);
}

static void i915_fence_release(struct dma_fence *fence)
{
	struct i915_request *rq = to_request(fence);

	/*
	 * The request is put onto a RCU freelist (i.e. the address
	 * is immediately reused), mark the fences as being freed now.
	 * Otherwise the debugobjects for the fences are only marked as
	 * freed when the slab cache itself is freed, and so we would get
	 * caught trying to reuse dead objects.
	 */
	i915_sw_fence_fini(&rq->submit);
	i915_sw_fence_fini(&rq->semaphore);

	kmem_cache_free(global.slab_requests, rq);
}

const struct dma_fence_ops i915_fence_ops = {
	.get_driver_name = i915_fence_get_driver_name,
	.get_timeline_name = i915_fence_get_timeline_name,
	.enable_signaling = i915_fence_enable_signaling,
	.signaled = i915_fence_signaled,
	.wait = i915_fence_wait,
	.release = i915_fence_release,
};

static void irq_execute_cb(struct irq_work *wrk)
{
	struct execute_cb *cb = container_of(wrk, typeof(*cb), work);

	i915_sw_fence_complete(cb->fence);
	kmem_cache_free(global.slab_execute_cbs, cb);
}

static void irq_execute_cb_hook(struct irq_work *wrk)
{
	struct execute_cb *cb = container_of(wrk, typeof(*cb), work);

	cb->hook(container_of(cb->fence, struct i915_request, submit),
		 &cb->signal->fence);
	i915_request_put(cb->signal);

	irq_execute_cb(wrk);
}

static void __notify_execute_cb(struct i915_request *rq)
{
	struct execute_cb *cb;

	lockdep_assert_held(&rq->lock);

	if (list_empty(&rq->execute_cb))
		return;

	list_for_each_entry(cb, &rq->execute_cb, link)
		irq_work_queue(&cb->work);

	/*
	 * XXX Rollback on __i915_request_unsubmit()
	 *
	 * In the future, perhaps when we have an active time-slicing scheduler,
	 * it will be interesting to unsubmit parallel execution and remove
	 * busywaits from the GPU until their master is restarted. This is
	 * quite hairy, we have to carefully rollback the fence and do a
	 * preempt-to-idle cycle on the target engine, all the while the
	 * master execute_cb may refire.
	 */
	INIT_LIST_HEAD(&rq->execute_cb);
}

static inline void
remove_from_client(struct i915_request *request)
{
	struct drm_i915_file_private *file_priv;

	if (!READ_ONCE(request->file_priv))
		return;

	rcu_read_lock();
	file_priv = xchg(&request->file_priv, NULL);
	if (file_priv) {
		spin_lock(&file_priv->mm.lock);
		list_del(&request->client_link);
		spin_unlock(&file_priv->mm.lock);
	}
	rcu_read_unlock();
}

static void free_capture_list(struct i915_request *request)
{
	struct i915_capture_list *capture;

	capture = fetch_and_zero(&request->capture_list);
	while (capture) {
		struct i915_capture_list *next = capture->next;

		kfree(capture);
		capture = next;
	}
}

static void __i915_request_fill(struct i915_request *rq, u8 val)
{
	void *vaddr = rq->ring->vaddr;
	u32 head;

	head = rq->infix;
	if (rq->postfix < head) {
		memset(vaddr + head, val, rq->ring->size - head);
		head = 0;
	}
	memset(vaddr + head, val, rq->postfix - head);
}

static void remove_from_engine(struct i915_request *rq)
{
	struct intel_engine_cs *engine, *locked;

	/*
	 * Virtual engines complicate acquiring the engine timeline lock,
	 * as their rq->engine pointer is not stable until under that
	 * engine lock. The simple ploy we use is to take the lock then
	 * check that the rq still belongs to the newly locked engine.
	 */
	locked = READ_ONCE(rq->engine);
	spin_lock_irq(&locked->active.lock);
	while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
		spin_unlock(&locked->active.lock);
		spin_lock(&engine->active.lock);
		locked = engine;
	}
	list_del_init(&rq->sched.link);
	clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
	clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
	spin_unlock_irq(&locked->active.lock);
}

bool i915_request_retire(struct i915_request *rq)
{
	if (!i915_request_completed(rq))
		return false;

	RQ_TRACE(rq, "\n");

	GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
	trace_i915_request_retire(rq);

	/*
	 * We know the GPU must have read the request to have
	 * sent us the seqno + interrupt, so use the position
	 * of tail of the request to update the last known position
	 * of the GPU head.
	 *
	 * Note this requires that we are always called in request
	 * completion order.
	 */
	GEM_BUG_ON(!list_is_first(&rq->link,
				  &i915_request_timeline(rq)->requests));
	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
		/* Poison before we release our space in the ring */
		__i915_request_fill(rq, POISON_FREE);
	rq->ring->head = rq->postfix;

	/*
	 * We only loosely track inflight requests across preemption,
	 * and so we may find ourselves attempting to retire a _completed_
	 * request that we have removed from the HW and put back on a run
	 * queue.
	 */
	remove_from_engine(rq);

	spin_lock_irq(&rq->lock);
	i915_request_mark_complete(rq);
	if (!i915_request_signaled(rq))
		dma_fence_signal_locked(&rq->fence);
	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags))
		i915_request_cancel_breadcrumb(rq);
	if (i915_request_has_waitboost(rq)) {
		GEM_BUG_ON(!atomic_read(&rq->engine->gt->rps.num_waiters));
		atomic_dec(&rq->engine->gt->rps.num_waiters);
	}
	if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
		set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
		__notify_execute_cb(rq);
	}
	GEM_BUG_ON(!list_empty(&rq->execute_cb));
	spin_unlock_irq(&rq->lock);

	remove_from_client(rq);
	__list_del_entry(&rq->link); /* poison neither prev/next (RCU walks) */

	intel_context_exit(rq->context);
	intel_context_unpin(rq->context);

	free_capture_list(rq);
	i915_sched_node_fini(&rq->sched);
	i915_request_put(rq);

	return true;
}

void i915_request_retire_upto(struct i915_request *rq)
{
	struct intel_timeline * const tl = i915_request_timeline(rq);
	struct i915_request *tmp;

	RQ_TRACE(rq, "\n");

	GEM_BUG_ON(!i915_request_completed(rq));

	do {
		tmp = list_first_entry(&tl->requests, typeof(*tmp), link);
	} while (i915_request_retire(tmp) && tmp != rq);
}

static int
__await_execution(struct i915_request *rq,
		  struct i915_request *signal,
		  void (*hook)(struct i915_request *rq,
			       struct dma_fence *signal),
		  gfp_t gfp)
{
	struct execute_cb *cb;

	if (i915_request_is_active(signal)) {
		if (hook)
			hook(rq, &signal->fence);
		return 0;
	}

	cb = kmem_cache_alloc(global.slab_execute_cbs, gfp);
	if (!cb)
		return -ENOMEM;

	cb->fence = &rq->submit;
	i915_sw_fence_await(cb->fence);
	init_irq_work(&cb->work, irq_execute_cb);

	if (hook) {
		cb->hook = hook;
		cb->signal = i915_request_get(signal);
		cb->work.func = irq_execute_cb_hook;
	}

	spin_lock_irq(&signal->lock);
	if (i915_request_is_active(signal)) {
		if (hook) {
			hook(rq, &signal->fence);
			i915_request_put(signal);
		}
		i915_sw_fence_complete(cb->fence);
		kmem_cache_free(global.slab_execute_cbs, cb);
	} else {
		list_add_tail(&cb->link, &signal->execute_cb);
	}
	spin_unlock_irq(&signal->lock);

	/* Copy across semaphore status as we need the same behaviour */
	rq->sched.flags |= signal->sched.flags;
	return 0;
}

static bool fatal_error(int error)
{
	switch (error) {
	case 0: /* not an error! */
	case -EAGAIN: /* innocent victim of a GT reset (__i915_request_reset) */
	case -ETIMEDOUT: /* waiting for Godot (timer_i915_sw_fence_wake) */
		return false;
	default:
		return true;
	}
}

void __i915_request_skip(struct i915_request *rq)
{
	GEM_BUG_ON(!fatal_error(rq->fence.error));

	if (rq->infix == rq->postfix)
		return;

	/*
	 * As this request likely depends on state from the lost
	 * context, clear out all the user operations leaving the
	 * breadcrumb at the end (so we get the fence notifications).
	 */
	__i915_request_fill(rq, 0);
	rq->infix = rq->postfix;
}

void i915_request_set_error_once(struct i915_request *rq, int error)
{
	int old;

	GEM_BUG_ON(!IS_ERR_VALUE((long)error));

	if (i915_request_signaled(rq))
		return;

	old = READ_ONCE(rq->fence.error);
	do {
		if (fatal_error(old))
			return;
	} while (!try_cmpxchg(&rq->fence.error, &old, error));
}

bool __i915_request_submit(struct i915_request *request)
{
	struct intel_engine_cs *engine = request->engine;
	bool result = false;

	RQ_TRACE(request, "\n");

	GEM_BUG_ON(!irqs_disabled());
	lockdep_assert_held(&engine->active.lock);

	/*
	 * With the advent of preempt-to-busy, we frequently encounter
	 * requests that we have unsubmitted from HW, but left running
	 * until the next ack and so have completed in the meantime. On
	 * resubmission of that completed request, we can skip
	 * updating the payload, and execlists can even skip submitting
	 * the request.
	 *
	 * We must remove the request from the caller's priority queue,
	 * and the caller must only call us when the request is in their
	 * priority queue, under the active.lock. This ensures that the
	 * request has *not* yet been retired and we can safely move
	 * the request into the engine->active.list where it will be
	 * dropped upon retiring. (Otherwise if resubmit a *retired*
	 * request, this would be a horrible use-after-free.)
	 */
	if (i915_request_completed(request))
		goto xfer;

	if (unlikely(intel_context_is_banned(request->context)))
		i915_request_set_error_once(request, -EIO);
	if (unlikely(fatal_error(request->fence.error)))
		__i915_request_skip(request);

	/*
	 * Are we using semaphores when the gpu is already saturated?
	 *
	 * Using semaphores incurs a cost in having the GPU poll a
	 * memory location, busywaiting for it to change. The continual
	 * memory reads can have a noticeable impact on the rest of the
	 * system with the extra bus traffic, stalling the cpu as it too
	 * tries to access memory across the bus (perf stat -e bus-cycles).
	 *
	 * If we installed a semaphore on this request and we only submit
	 * the request after the signaler completed, that indicates the
	 * system is overloaded and using semaphores at this time only
	 * increases the amount of work we are doing. If so, we disable
	 * further use of semaphores until we are idle again, whence we
	 * optimistically try again.
	 */
	if (request->sched.semaphores &&
	    i915_sw_fence_signaled(&request->semaphore))
		engine->saturated |= request->sched.semaphores;

	engine->emit_fini_breadcrumb(request,
				     request->ring->vaddr + request->postfix);

	trace_i915_request_execute(request);
	engine->serial++;
	result = true;

xfer:	/* We may be recursing from the signal callback of another i915 fence */
	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);

	if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)) {
		list_move_tail(&request->sched.link, &engine->active.requests);
		clear_bit(I915_FENCE_FLAG_PQUEUE, &request->fence.flags);
	}

	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) &&
	    !i915_request_enable_breadcrumb(request))
		intel_engine_signal_breadcrumbs(engine);

	__notify_execute_cb(request);

	spin_unlock(&request->lock);

	return result;
}

void i915_request_submit(struct i915_request *request)
{
	struct intel_engine_cs *engine = request->engine;
	unsigned long flags;

	/* Will be called from irq-context when using foreign fences. */
	spin_lock_irqsave(&engine->active.lock, flags);

	__i915_request_submit(request);

	spin_unlock_irqrestore(&engine->active.lock, flags);
}

void __i915_request_unsubmit(struct i915_request *request)
{
	struct intel_engine_cs *engine = request->engine;

	RQ_TRACE(request, "\n");

	GEM_BUG_ON(!irqs_disabled());
	lockdep_assert_held(&engine->active.lock);

	/*
	 * Only unwind in reverse order, required so that the per-context list
	 * is kept in seqno/ring order.
	 */

	/* We may be recursing from the signal callback of another i915 fence */
	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);

	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
		i915_request_cancel_breadcrumb(request);

	GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
	clear_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);

	spin_unlock(&request->lock);

	/* We've already spun, don't charge on resubmitting. */
	if (request->sched.semaphores && i915_request_started(request)) {
		request->sched.attr.priority |= I915_PRIORITY_NOSEMAPHORE;
		request->sched.semaphores = 0;
	}

	/*
	 * We don't need to wake_up any waiters on request->execute, they
	 * will get woken by any other event or us re-adding this request
	 * to the engine timeline (__i915_request_submit()). The waiters
	 * should be quite adapt at finding that the request now has a new
	 * global_seqno to the one they went to sleep on.
	 */
}

void i915_request_unsubmit(struct i915_request *request)
{
	struct intel_engine_cs *engine = request->engine;
	unsigned long flags;

	/* Will be called from irq-context when using foreign fences. */
	spin_lock_irqsave(&engine->active.lock, flags);

	__i915_request_unsubmit(request);

	spin_unlock_irqrestore(&engine->active.lock, flags);
}

static int __i915_sw_fence_call
submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
{
	struct i915_request *request =
		container_of(fence, typeof(*request), submit);

	switch (state) {
	case FENCE_COMPLETE:
		trace_i915_request_submit(request);

		if (unlikely(fence->error))
			i915_request_set_error_once(request, fence->error);

		/*
		 * We need to serialize use of the submit_request() callback
		 * with its hotplugging performed during an emergency
		 * i915_gem_set_wedged().  We use the RCU mechanism to mark the
		 * critical section in order to force i915_gem_set_wedged() to
		 * wait until the submit_request() is completed before
		 * proceeding.
		 */
		rcu_read_lock();
		request->engine->submit_request(request);
		rcu_read_unlock();
		break;

	case FENCE_FREE:
		i915_request_put(request);
		break;
	}

	return NOTIFY_DONE;
}

static void irq_semaphore_cb(struct irq_work *wrk)
{
	struct i915_request *rq =
		container_of(wrk, typeof(*rq), semaphore_work);

	i915_schedule_bump_priority(rq, I915_PRIORITY_NOSEMAPHORE);
	i915_request_put(rq);
}

static int __i915_sw_fence_call
semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
{
	struct i915_request *rq = container_of(fence, typeof(*rq), semaphore);

	switch (state) {
	case FENCE_COMPLETE:
		if (!(READ_ONCE(rq->sched.attr.priority) & I915_PRIORITY_NOSEMAPHORE)) {
			i915_request_get(rq);
			init_irq_work(&rq->semaphore_work, irq_semaphore_cb);
			irq_work_queue(&rq->semaphore_work);
		}
		break;

	case FENCE_FREE:
		i915_request_put(rq);
		break;
	}

	return NOTIFY_DONE;
}

static void retire_requests(struct intel_timeline *tl)
{
	struct i915_request *rq, *rn;

	list_for_each_entry_safe(rq, rn, &tl->requests, link)
		if (!i915_request_retire(rq))
			break;
}

static noinline struct i915_request *
request_alloc_slow(struct intel_timeline *tl, gfp_t gfp)
{
	struct i915_request *rq;

	if (list_empty(&tl->requests))
		goto out;

	if (!gfpflags_allow_blocking(gfp))
		goto out;

	/* Move our oldest request to the slab-cache (if not in use!) */
	rq = list_first_entry(&tl->requests, typeof(*rq), link);
	i915_request_retire(rq);

	rq = kmem_cache_alloc(global.slab_requests,
			      gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
	if (rq)
		return rq;

	/* Ratelimit ourselves to prevent oom from malicious clients */
	rq = list_last_entry(&tl->requests, typeof(*rq), link);
	cond_synchronize_rcu(rq->rcustate);

	/* Retire our old requests in the hope that we free some */
	retire_requests(tl);

out:
	return kmem_cache_alloc(global.slab_requests, gfp);
}

static void __i915_request_ctor(void *arg)
{
	struct i915_request *rq = arg;

	spin_lock_init(&rq->lock);
	i915_sched_node_init(&rq->sched);
	i915_sw_fence_init(&rq->submit, submit_notify);
	i915_sw_fence_init(&rq->semaphore, semaphore_notify);

	dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock, 0, 0);

	rq->file_priv = NULL;
	rq->capture_list = NULL;

	INIT_LIST_HEAD(&rq->execute_cb);
}

struct i915_request *
__i915_request_create(struct intel_context *ce, gfp_t gfp)
{
	struct intel_timeline *tl = ce->timeline;
	struct i915_request *rq;
	u32 seqno;
	int ret;

	might_sleep_if(gfpflags_allow_blocking(gfp));

	/* Check that the caller provided an already pinned context */
	__intel_context_pin(ce);

	/*
	 * Beware: Dragons be flying overhead.
	 *
	 * We use RCU to look up requests in flight. The lookups may
	 * race with the request being allocated from the slab freelist.
	 * That is the request we are writing to here, may be in the process
	 * of being read by __i915_active_request_get_rcu(). As such,
	 * we have to be very careful when overwriting the contents. During
	 * the RCU lookup, we change chase the request->engine pointer,
	 * read the request->global_seqno and increment the reference count.
	 *
	 * The reference count is incremented atomically. If it is zero,
	 * the lookup knows the request is unallocated and complete. Otherwise,
	 * it is either still in use, or has been reallocated and reset
	 * with dma_fence_init(). This increment is safe for release as we
	 * check that the request we have a reference to and matches the active
	 * request.
	 *
	 * Before we increment the refcount, we chase the request->engine
	 * pointer. We must not call kmem_cache_zalloc() or else we set
	 * that pointer to NULL and cause a crash during the lookup. If
	 * we see the request is completed (based on the value of the
	 * old engine and seqno), the lookup is complete and reports NULL.
	 * If we decide the request is not completed (new engine or seqno),
	 * then we grab a reference and double check that it is still the
	 * active request - which it won't be and restart the lookup.
	 *
	 * Do not use kmem_cache_zalloc() here!
	 */
	rq = kmem_cache_alloc(global.slab_requests,
			      gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
	if (unlikely(!rq)) {
		rq = request_alloc_slow(tl, gfp);
		if (!rq) {
			ret = -ENOMEM;
			goto err_unreserve;
		}
	}

	rq->i915 = ce->engine->i915;
	rq->context = ce;
	rq->engine = ce->engine;
	rq->ring = ce->ring;
	rq->execution_mask = ce->engine->mask;

	kref_init(&rq->fence.refcount);
	rq->fence.flags = 0;
	rq->fence.error = 0;
	INIT_LIST_HEAD(&rq->fence.cb_list);

	ret = intel_timeline_get_seqno(tl, rq, &seqno);
	if (ret)
		goto err_free;

	rq->fence.context = tl->fence_context;
	rq->fence.seqno = seqno;

	RCU_INIT_POINTER(rq->timeline, tl);
	RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline);
	rq->hwsp_seqno = tl->hwsp_seqno;
	GEM_BUG_ON(i915_request_completed(rq));

	rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */

	/* We bump the ref for the fence chain */
	i915_sw_fence_reinit(&i915_request_get(rq)->submit);
	i915_sw_fence_reinit(&i915_request_get(rq)->semaphore);

	i915_sched_node_reinit(&rq->sched);

	/* No zalloc, everything must be cleared after use */
	rq->batch = NULL;
	GEM_BUG_ON(rq->file_priv);
	GEM_BUG_ON(rq->capture_list);
	GEM_BUG_ON(!list_empty(&rq->execute_cb));

	/*
	 * Reserve space in the ring buffer for all the commands required to
	 * eventually emit this request. This is to guarantee that the
	 * i915_request_add() call can't fail. Note that the reserve may need
	 * to be redone if the request is not actually submitted straight
	 * away, e.g. because a GPU scheduler has deferred it.
	 *
	 * Note that due to how we add reserved_space to intel_ring_begin()
	 * we need to double our request to ensure that if we need to wrap
	 * around inside i915_request_add() there is sufficient space at
	 * the beginning of the ring as well.
	 */
	rq->reserved_space =
		2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32);

	/*
	 * Record the position of the start of the request so that
	 * should we detect the updated seqno part-way through the
	 * GPU processing the request, we never over-estimate the
	 * position of the head.
	 */
	rq->head = rq->ring->emit;

	ret = rq->engine->request_alloc(rq);
	if (ret)
		goto err_unwind;

	rq->infix = rq->ring->emit; /* end of header; start of user payload */

	intel_context_mark_active(ce);
	list_add_tail_rcu(&rq->link, &tl->requests);

	return rq;

err_unwind:
	ce->ring->emit = rq->head;

	/* Make sure we didn't add ourselves to external state before freeing */
	GEM_BUG_ON(!list_empty(&rq->sched.signalers_list));
	GEM_BUG_ON(!list_empty(&rq->sched.waiters_list));

err_free:
	kmem_cache_free(global.slab_requests, rq);
err_unreserve:
	intel_context_unpin(ce);
	return ERR_PTR(ret);
}

struct i915_request *
i915_request_create(struct intel_context *ce)
{
	struct i915_request *rq;
	struct intel_timeline *tl;

	tl = intel_context_timeline_lock(ce);
	if (IS_ERR(tl))
		return ERR_CAST(tl);

	/* Move our oldest request to the slab-cache (if not in use!) */
	rq = list_first_entry(&tl->requests, typeof(*rq), link);
	if (!list_is_last(&rq->link, &tl->requests))
		i915_request_retire(rq);

	intel_context_enter(ce);
	rq = __i915_request_create(ce, GFP_KERNEL);
	intel_context_exit(ce); /* active reference transferred to request */
	if (IS_ERR(rq))
		goto err_unlock;

	/* Check that we do not interrupt ourselves with a new request */
	rq->cookie = lockdep_pin_lock(&tl->mutex);

	return rq;

err_unlock:
	intel_context_timeline_unlock(tl);
	return rq;
}

static int
i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
{
	struct dma_fence *fence;
	int err;

	if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline))
		return 0;

	if (i915_request_started(signal))
		return 0;

	fence = NULL;
	rcu_read_lock();
	spin_lock_irq(&signal->lock);
	do {
		struct list_head *pos = READ_ONCE(signal->link.prev);
		struct i915_request *prev;

		/* Confirm signal has not been retired, the link is valid */
		if (unlikely(i915_request_started(signal)))
			break;

		/* Is signal the earliest request on its timeline? */
		if (pos == &rcu_dereference(signal->timeline)->requests)
			break;

		/*
		 * Peek at the request before us in the timeline. That
		 * request will only be valid before it is retired, so
		 * after acquiring a reference to it, confirm that it is
		 * still part of the signaler's timeline.
		 */
		prev = list_entry(pos, typeof(*prev), link);
		if (!i915_request_get_rcu(prev))
			break;

		/* After the strong barrier, confirm prev is still attached */
		if (unlikely(READ_ONCE(prev->link.next) != &signal->link)) {
			i915_request_put(prev);
			break;
		}

		fence = &prev->fence;
	} while (0);
	spin_unlock_irq(&signal->lock);
	rcu_read_unlock();
	if (!fence)
		return 0;

	err = 0;
	if (!intel_timeline_sync_is_later(i915_request_timeline(rq), fence))
		err = i915_sw_fence_await_dma_fence(&rq->submit,
						    fence, 0,
						    I915_FENCE_GFP);
	dma_fence_put(fence);

	return err;
}

static intel_engine_mask_t
already_busywaiting(struct i915_request *rq)
{
	/*
	 * Polling a semaphore causes bus traffic, delaying other users of
	 * both the GPU and CPU. We want to limit the impact on others,
	 * while taking advantage of early submission to reduce GPU
	 * latency. Therefore we restrict ourselves to not using more
	 * than one semaphore from each source, and not using a semaphore
	 * if we have detected the engine is saturated (i.e. would not be
	 * submitted early and cause bus traffic reading an already passed
	 * semaphore).
	 *
	 * See the are-we-too-late? check in __i915_request_submit().
	 */
	return rq->sched.semaphores | READ_ONCE(rq->engine->saturated);
}

static int
__emit_semaphore_wait(struct i915_request *to,
		      struct i915_request *from,
		      u32 seqno)
{
	const int has_token = INTEL_GEN(to->i915) >= 12;
	u32 hwsp_offset;
	int len, err;
	u32 *cs;

	GEM_BUG_ON(INTEL_GEN(to->i915) < 8);

	/* We need to pin the signaler's HWSP until we are finished reading. */
	err = intel_timeline_read_hwsp(from, to, &hwsp_offset);
	if (err)
		return err;

	len = 4;
	if (has_token)
		len += 2;

	cs = intel_ring_begin(to, len);
	if (IS_ERR(cs))
		return PTR_ERR(cs);

	/*
	 * Using greater-than-or-equal here means we have to worry
	 * about seqno wraparound. To side step that issue, we swap
	 * the timeline HWSP upon wrapping, so that everyone listening
	 * for the old (pre-wrap) values do not see the much smaller
	 * (post-wrap) values than they were expecting (and so wait
	 * forever).
	 */
	*cs++ = (MI_SEMAPHORE_WAIT |
		 MI_SEMAPHORE_GLOBAL_GTT |
		 MI_SEMAPHORE_POLL |
		 MI_SEMAPHORE_SAD_GTE_SDD) +
		has_token;
	*cs++ = seqno;
	*cs++ = hwsp_offset;
	*cs++ = 0;
	if (has_token) {
		*cs++ = 0;
		*cs++ = MI_NOOP;
	}

	intel_ring_advance(to, cs);
	return 0;
}

static int
emit_semaphore_wait(struct i915_request *to,
		    struct i915_request *from,
		    gfp_t gfp)
{
	const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask;

	if (!intel_context_use_semaphores(to->context))
		goto await_fence;

	if (!rcu_access_pointer(from->hwsp_cacheline))
		goto await_fence;

	/* Just emit the first semaphore we see as request space is limited. */
	if (already_busywaiting(to) & mask)
		goto await_fence;

	if (i915_request_await_start(to, from) < 0)
		goto await_fence;

	/* Only submit our spinner after the signaler is running! */
	if (__await_execution(to, from, NULL, gfp))
		goto await_fence;

	if (__emit_semaphore_wait(to, from, from->fence.seqno))
		goto await_fence;

	to->sched.semaphores |= mask;
	to->sched.flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
	return 0;

await_fence:
	return i915_sw_fence_await_dma_fence(&to->submit,
					     &from->fence, 0,
					     I915_FENCE_GFP);
}

static int
i915_request_await_request(struct i915_request *to, struct i915_request *from)
{
	int ret;

	GEM_BUG_ON(to == from);
	GEM_BUG_ON(to->timeline == from->timeline);

	if (i915_request_completed(from))
		return 0;

	if (to->engine->schedule) {
		ret = i915_sched_node_add_dependency(&to->sched, &from->sched);
		if (ret < 0)
			return ret;
	}

	if (to->engine == from->engine)
		ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
						       &from->submit,
						       I915_FENCE_GFP);
	else
		ret = emit_semaphore_wait(to, from, I915_FENCE_GFP);
	if (ret < 0)
		return ret;

	if (to->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN) {
		ret = i915_sw_fence_await_dma_fence(&to->semaphore,
						    &from->fence, 0,
						    I915_FENCE_GFP);
		if (ret < 0)
			return ret;
	}

	return 0;
}

int
i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
{
	struct dma_fence **child = &fence;
	unsigned int nchild = 1;
	int ret;

	/*
	 * Note that if the fence-array was created in signal-on-any mode,
	 * we should *not* decompose it into its individual fences. However,
	 * we don't currently store which mode the fence-array is operating
	 * in. Fortunately, the only user of signal-on-any is private to
	 * amdgpu and we should not see any incoming fence-array from
	 * sync-file being in signal-on-any mode.
	 */
	if (dma_fence_is_array(fence)) {
		struct dma_fence_array *array = to_dma_fence_array(fence);

		child = array->fences;
		nchild = array->num_fences;
		GEM_BUG_ON(!nchild);
	}

	do {
		fence = *child++;
		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
			i915_sw_fence_set_error_once(&rq->submit, fence->error);
			continue;
		}

		/*
		 * Requests on the same timeline are explicitly ordered, along
		 * with their dependencies, by i915_request_add() which ensures
		 * that requests are submitted in-order through each ring.
		 */
		if (fence->context == rq->fence.context)
			continue;

		/* Squash repeated waits to the same timelines */
		if (fence->context &&
		    intel_timeline_sync_is_later(i915_request_timeline(rq),
						 fence))
			continue;

		if (dma_fence_is_i915(fence))
			ret = i915_request_await_request(rq, to_request(fence));
		else
			ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
							    fence->context ? I915_FENCE_TIMEOUT : 0,
							    I915_FENCE_GFP);
		if (ret < 0)
			return ret;

		/* Record the latest fence used against each timeline */
		if (fence->context)
			intel_timeline_sync_set(i915_request_timeline(rq),
						fence);
	} while (--nchild);

	return 0;
}

static bool intel_timeline_sync_has_start(struct intel_timeline *tl,
					  struct dma_fence *fence)
{
	return __intel_timeline_sync_is_later(tl,
					      fence->context,
					      fence->seqno - 1);
}

static int intel_timeline_sync_set_start(struct intel_timeline *tl,
					 const struct dma_fence *fence)
{
	return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1);
}

static int
__i915_request_await_execution(struct i915_request *to,
			       struct i915_request *from,
			       void (*hook)(struct i915_request *rq,
					    struct dma_fence *signal))
{
	int err;

	GEM_BUG_ON(intel_context_is_barrier(from->context));

	/* Submit both requests at the same time */
	err = __await_execution(to, from, hook, I915_FENCE_GFP);
	if (err)
		return err;

	/* Squash repeated depenendices to the same timelines */
	if (intel_timeline_sync_has_start(i915_request_timeline(to),
					  &from->fence))
		return 0;

	/*
	 * Wait until the start of this request.
	 *
	 * The execution cb fires when we submit the request to HW. But in
	 * many cases this may be long before the request itself is ready to
	 * run (consider that we submit 2 requests for the same context, where
	 * the request of interest is behind an indefinite spinner). So we hook
	 * up to both to reduce our queues and keep the execution lag minimised
	 * in the worst case, though we hope that the await_start is elided.
	 */
	err = i915_request_await_start(to, from);
	if (err < 0)
		return err;

	/*
	 * Ensure both start together [after all semaphores in signal]
	 *
	 * Now that we are queued to the HW at roughly the same time (thanks
	 * to the execute cb) and are ready to run at roughly the same time
	 * (thanks to the await start), our signaler may still be indefinitely
	 * delayed by waiting on a semaphore from a remote engine. If our
	 * signaler depends on a semaphore, so indirectly do we, and we do not
	 * want to start our payload until our signaler also starts theirs.
	 * So we wait.
	 *
	 * However, there is also a second condition for which we need to wait
	 * for the precise start of the signaler. Consider that the signaler
	 * was submitted in a chain of requests following another context
	 * (with just an ordinary intra-engine fence dependency between the
	 * two). In this case the signaler is queued to HW, but not for
	 * immediate execution, and so we must wait until it reaches the
	 * active slot.
	 */
	if (intel_engine_has_semaphores(to->engine)) {
		err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
		if (err < 0)
			return err;
	}

	/* Couple the dependency tree for PI on this exposed to->fence */
	if (to->engine->schedule) {
		err = i915_sched_node_add_dependency(&to->sched, &from->sched);
		if (err < 0)
			return err;
	}

	return intel_timeline_sync_set_start(i915_request_timeline(to),
					     &from->fence);
}

int
i915_request_await_execution(struct i915_request *rq,
			     struct dma_fence *fence,
			     void (*hook)(struct i915_request *rq,
					  struct dma_fence *signal))
{
	struct dma_fence **child = &fence;
	unsigned int nchild = 1;
	int ret;

	if (dma_fence_is_array(fence)) {
		struct dma_fence_array *array = to_dma_fence_array(fence);

		/* XXX Error for signal-on-any fence arrays */

		child = array->fences;
		nchild = array->num_fences;
		GEM_BUG_ON(!nchild);
	}

	do {
		fence = *child++;
		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
			i915_sw_fence_set_error_once(&rq->submit, fence->error);
			continue;
		}

		/*
		 * We don't squash repeated fence dependencies here as we
		 * want to run our callback in all cases.
		 */

		if (dma_fence_is_i915(fence))
			ret = __i915_request_await_execution(rq,
							     to_request(fence),
							     hook);
		else
			ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
							    I915_FENCE_TIMEOUT,
							    GFP_KERNEL);
		if (ret < 0)
			return ret;
	} while (--nchild);

	return 0;
}

/**
 * i915_request_await_object - set this request to (async) wait upon a bo
 * @to: request we are wishing to use
 * @obj: object which may be in use on another ring.
 * @write: whether the wait is on behalf of a writer
 *
 * This code is meant to abstract object synchronization with the GPU.
 * Conceptually we serialise writes between engines inside the GPU.
 * We only allow one engine to write into a buffer at any time, but
 * multiple readers. To ensure each has a coherent view of memory, we must:
 *
 * - If there is an outstanding write request to the object, the new
 *   request must wait for it to complete (either CPU or in hw, requests
 *   on the same ring will be naturally ordered).
 *
 * - If we are a write request (pending_write_domain is set), the new
 *   request must wait for outstanding read requests to complete.
 *
 * Returns 0 if successful, else propagates up the lower layer error.
 */
int
i915_request_await_object(struct i915_request *to,
			  struct drm_i915_gem_object *obj,
			  bool write)
{
	struct dma_fence *excl;
	int ret = 0;

	if (write) {
		struct dma_fence **shared;
		unsigned int count, i;

		ret = dma_resv_get_fences_rcu(obj->base.resv,
							&excl, &count, &shared);
		if (ret)
			return ret;

		for (i = 0; i < count; i++) {
			ret = i915_request_await_dma_fence(to, shared[i]);
			if (ret)
				break;

			dma_fence_put(shared[i]);
		}

		for (; i < count; i++)
			dma_fence_put(shared[i]);
		kfree(shared);
	} else {
		excl = dma_resv_get_excl_rcu(obj->base.resv);
	}

	if (excl) {
		if (ret == 0)
			ret = i915_request_await_dma_fence(to, excl);

		dma_fence_put(excl);
	}

	return ret;
}

static struct i915_request *
__i915_request_add_to_timeline(struct i915_request *rq)
{
	struct intel_timeline *timeline = i915_request_timeline(rq);
	struct i915_request *prev;

	/*
	 * Dependency tracking and request ordering along the timeline
	 * is special cased so that we can eliminate redundant ordering
	 * operations while building the request (we know that the timeline
	 * itself is ordered, and here we guarantee it).
	 *
	 * As we know we will need to emit tracking along the timeline,
	 * we embed the hooks into our request struct -- at the cost of
	 * having to have specialised no-allocation interfaces (which will
	 * be beneficial elsewhere).
	 *
	 * A second benefit to open-coding i915_request_await_request is
	 * that we can apply a slight variant of the rules specialised
	 * for timelines that jump between engines (such as virtual engines).
	 * If we consider the case of virtual engine, we must emit a dma-fence
	 * to prevent scheduling of the second request until the first is
	 * complete (to maximise our greedy late load balancing) and this
	 * precludes optimising to use semaphores serialisation of a single
	 * timeline across engines.
	 */
	prev = to_request(__i915_active_fence_set(&timeline->last_request,
						  &rq->fence));
	if (prev && !i915_request_completed(prev)) {
		/*
		 * The requests are supposed to be kept in order. However,
		 * we need to be wary in case the timeline->last_request
		 * is used as a barrier for external modification to this
		 * context.
		 */
		GEM_BUG_ON(prev->context == rq->context &&
			   i915_seqno_passed(prev->fence.seqno,
					     rq->fence.seqno));

		if (is_power_of_2(READ_ONCE(prev->engine)->mask | rq->engine->mask))
			i915_sw_fence_await_sw_fence(&rq->submit,
						     &prev->submit,
						     &rq->submitq);
		else
			__i915_sw_fence_await_dma_fence(&rq->submit,
							&prev->fence,
							&rq->dmaq);
		if (rq->engine->schedule)
			__i915_sched_node_add_dependency(&rq->sched,
							 &prev->sched,
							 &rq->dep,
							 0);
	}

	/*
	 * Make sure that no request gazumped us - if it was allocated after
	 * our i915_request_alloc() and called __i915_request_add() before
	 * us, the timeline will hold its seqno which is later than ours.
	 */
	GEM_BUG_ON(timeline->seqno != rq->fence.seqno);

	return prev;
}

/*
 * NB: This function is not allowed to fail. Doing so would mean the the
 * request is not being tracked for completion but the work itself is
 * going to happen on the hardware. This would be a Bad Thing(tm).
 */
struct i915_request *__i915_request_commit(struct i915_request *rq)
{
	struct intel_engine_cs *engine = rq->engine;
	struct intel_ring *ring = rq->ring;
	u32 *cs;

	RQ_TRACE(rq, "\n");

	/*
	 * To ensure that this call will not fail, space for its emissions
	 * should already have been reserved in the ring buffer. Let the ring
	 * know that it is time to use that space up.
	 */
	GEM_BUG_ON(rq->reserved_space > ring->space);
	rq->reserved_space = 0;
	rq->emitted_jiffies = jiffies;

	/*
	 * Record the position of the start of the breadcrumb so that
	 * should we detect the updated seqno part-way through the
	 * GPU processing the request, we never over-estimate the
	 * position of the ring's HEAD.
	 */
	cs = intel_ring_begin(rq, engine->emit_fini_breadcrumb_dw);
	GEM_BUG_ON(IS_ERR(cs));
	rq->postfix = intel_ring_offset(rq, cs);

	return __i915_request_add_to_timeline(rq);
}

void __i915_request_queue(struct i915_request *rq,
			  const struct i915_sched_attr *attr)
{
	/*
	 * Let the backend know a new request has arrived that may need
	 * to adjust the existing execution schedule due to a high priority
	 * request - i.e. we may want to preempt the current request in order
	 * to run a high priority dependency chain *before* we can execute this
	 * request.
	 *
	 * This is called before the request is ready to run so that we can
	 * decide whether to preempt the entire chain so that it is ready to
	 * run at the earliest possible convenience.
	 */
	if (attr && rq->engine->schedule)
		rq->engine->schedule(rq, attr);
	i915_sw_fence_commit(&rq->semaphore);
	i915_sw_fence_commit(&rq->submit);
}

void i915_request_add(struct i915_request *rq)
{
	struct intel_timeline * const tl = i915_request_timeline(rq);
	struct i915_sched_attr attr = {};
	struct i915_gem_context *ctx;

	lockdep_assert_held(&tl->mutex);
	lockdep_unpin_lock(&tl->mutex, rq->cookie);

	trace_i915_request_add(rq);
	__i915_request_commit(rq);

	/* XXX placeholder for selftests */
	rcu_read_lock();
	ctx = rcu_dereference(rq->context->gem_context);
	if (ctx)
		attr = ctx->sched;
	rcu_read_unlock();

	if (!(rq->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN))
		attr.priority |= I915_PRIORITY_NOSEMAPHORE;
	if (list_empty(&rq->sched.signalers_list))
		attr.priority |= I915_PRIORITY_WAIT;

	local_bh_disable();
	__i915_request_queue(rq, &attr);
	local_bh_enable(); /* Kick the execlists tasklet if just scheduled */

	mutex_unlock(&tl->mutex);
}

static unsigned long local_clock_ns(unsigned int *cpu)
{
	unsigned long t;

	/*
	 * Cheaply and approximately convert from nanoseconds to microseconds.
	 * The result and subsequent calculations are also defined in the same
	 * approximate microseconds units. The principal source of timing
	 * error here is from the simple truncation.
	 *
	 * Note that local_clock() is only defined wrt to the current CPU;
	 * the comparisons are no longer valid if we switch CPUs. Instead of
	 * blocking preemption for the entire busywait, we can detect the CPU
	 * switch and use that as indicator of system load and a reason to
	 * stop busywaiting, see busywait_stop().
	 */
	*cpu = get_cpu();
	t = local_clock();
	put_cpu();

	return t;
}

static bool busywait_stop(unsigned long timeout, unsigned int cpu)
{
	unsigned int this_cpu;

	if (time_after(local_clock_ns(&this_cpu), timeout))
		return true;

	return this_cpu != cpu;
}

static bool __i915_spin_request(const struct i915_request * const rq, int state)
{
	unsigned long timeout_ns;
	unsigned int cpu;

	/*
	 * Only wait for the request if we know it is likely to complete.
	 *
	 * We don't track the timestamps around requests, nor the average
	 * request length, so we do not have a good indicator that this
	 * request will complete within the timeout. What we do know is the
	 * order in which requests are executed by the context and so we can
	 * tell if the request has been started. If the request is not even
	 * running yet, it is a fair assumption that it will not complete
	 * within our relatively short timeout.
	 */
	if (!i915_request_is_running(rq))
		return false;

	/*
	 * When waiting for high frequency requests, e.g. during synchronous
	 * rendering split between the CPU and GPU, the finite amount of time
	 * required to set up the irq and wait upon it limits the response
	 * rate. By busywaiting on the request completion for a short while we
	 * can service the high frequency waits as quick as possible. However,
	 * if it is a slow request, we want to sleep as quickly as possible.
	 * The tradeoff between waiting and sleeping is roughly the time it
	 * takes to sleep on a request, on the order of a microsecond.
	 */

	timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns);
	timeout_ns += local_clock_ns(&cpu);
	do {
		if (i915_request_completed(rq))
			return true;

		if (signal_pending_state(state, current))
			break;

		if (busywait_stop(timeout_ns, cpu))
			break;

		cpu_relax();
	} while (!need_resched());

	return false;
}

struct request_wait {
	struct dma_fence_cb cb;
	struct task_struct *tsk;
};

static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb)
{
	struct request_wait *wait = container_of(cb, typeof(*wait), cb);

	wake_up_process(wait->tsk);
}

/**
 * i915_request_wait - wait until execution of request has finished
 * @rq: the request to wait upon
 * @flags: how to wait
 * @timeout: how long to wait in jiffies
 *
 * i915_request_wait() waits for the request to be completed, for a
 * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
 * unbounded wait).
 *
 * Returns the remaining time (in jiffies) if the request completed, which may
 * be zero or -ETIME if the request is unfinished after the timeout expires.
 * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
 * pending before the request completes.
 */
long i915_request_wait(struct i915_request *rq,
		       unsigned int flags,
		       long timeout)
{
	const int state = flags & I915_WAIT_INTERRUPTIBLE ?
		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
	struct request_wait wait;

	might_sleep();
	GEM_BUG_ON(timeout < 0);

	if (dma_fence_is_signaled(&rq->fence))
		return timeout;

	if (!timeout)
		return -ETIME;

	trace_i915_request_wait_begin(rq, flags);

	/*
	 * We must never wait on the GPU while holding a lock as we
	 * may need to perform a GPU reset. So while we don't need to
	 * serialise wait/reset with an explicit lock, we do want
	 * lockdep to detect potential dependency cycles.
	 */
	mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_);

	/*
	 * Optimistic spin before touching IRQs.
	 *
	 * We may use a rather large value here to offset the penalty of
	 * switching away from the active task. Frequently, the client will
	 * wait upon an old swapbuffer to throttle itself to remain within a
	 * frame of the gpu. If the client is running in lockstep with the gpu,
	 * then it should not be waiting long at all, and a sleep now will incur
	 * extra scheduler latency in producing the next frame. To try to
	 * avoid adding the cost of enabling/disabling the interrupt to the
	 * short wait, we first spin to see if the request would have completed
	 * in the time taken to setup the interrupt.
	 *
	 * We need upto 5us to enable the irq, and upto 20us to hide the
	 * scheduler latency of a context switch, ignoring the secondary
	 * impacts from a context switch such as cache eviction.
	 *
	 * The scheme used for low-latency IO is called "hybrid interrupt
	 * polling". The suggestion there is to sleep until just before you
	 * expect to be woken by the device interrupt and then poll for its
	 * completion. That requires having a good predictor for the request
	 * duration, which we currently lack.
	 */
	if (IS_ACTIVE(CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT) &&
	    __i915_spin_request(rq, state)) {
		dma_fence_signal(&rq->fence);
		goto out;
	}

	/*
	 * This client is about to stall waiting for the GPU. In many cases
	 * this is undesirable and limits the throughput of the system, as
	 * many clients cannot continue processing user input/output whilst
	 * blocked. RPS autotuning may take tens of milliseconds to respond
	 * to the GPU load and thus incurs additional latency for the client.
	 * We can circumvent that by promoting the GPU frequency to maximum
	 * before we sleep. This makes the GPU throttle up much more quickly
	 * (good for benchmarks and user experience, e.g. window animations),
	 * but at a cost of spending more power processing the workload
	 * (bad for battery).
	 */
	if (flags & I915_WAIT_PRIORITY) {
		if (!i915_request_started(rq) && INTEL_GEN(rq->i915) >= 6)
			intel_rps_boost(rq);
		i915_schedule_bump_priority(rq, I915_PRIORITY_WAIT);
	}

	wait.tsk = current;
	if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake))
		goto out;

	for (;;) {
		set_current_state(state);

		if (i915_request_completed(rq)) {
			dma_fence_signal(&rq->fence);
			break;
		}

		intel_engine_flush_submission(rq->engine);

		if (signal_pending_state(state, current)) {
			timeout = -ERESTARTSYS;
			break;
		}

		if (!timeout) {
			timeout = -ETIME;
			break;
		}

		timeout = io_schedule_timeout(timeout);
	}
	__set_current_state(TASK_RUNNING);

	dma_fence_remove_callback(&rq->fence, &wait.cb);

out:
	mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_);
	trace_i915_request_wait_end(rq);
	return timeout;
}

#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
#include "selftests/mock_request.c"
#include "selftests/i915_request.c"
#endif

static void i915_global_request_shrink(void)
{
	kmem_cache_shrink(global.slab_execute_cbs);
	kmem_cache_shrink(global.slab_requests);
}

static void i915_global_request_exit(void)
{
	kmem_cache_destroy(global.slab_execute_cbs);
	kmem_cache_destroy(global.slab_requests);
}

static struct i915_global_request global = { {
	.shrink = i915_global_request_shrink,
	.exit = i915_global_request_exit,
} };

int __init i915_global_request_init(void)
{
	global.slab_requests =
		kmem_cache_create("i915_request",
				  sizeof(struct i915_request),
				  __alignof__(struct i915_request),
				  SLAB_HWCACHE_ALIGN |
				  SLAB_RECLAIM_ACCOUNT |
				  SLAB_TYPESAFE_BY_RCU,
				  __i915_request_ctor);
	if (!global.slab_requests)
		return -ENOMEM;

	global.slab_execute_cbs = KMEM_CACHE(execute_cb,
					     SLAB_HWCACHE_ALIGN |
					     SLAB_RECLAIM_ACCOUNT |
					     SLAB_TYPESAFE_BY_RCU);
	if (!global.slab_execute_cbs)
		goto err_requests;

	i915_global_register(&global.base);
	return 0;

err_requests:
	kmem_cache_destroy(global.slab_requests);
	return -ENOMEM;
}