summaryrefslogtreecommitdiffstats
path: root/drivers/infiniband/hw/mlx5/odp.c
blob: 77dca1e05bba8db7dc40c6346855c84386302f4b (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
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
/*
 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * 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 <rdma/ib_umem.h>
#include <rdma/ib_umem_odp.h>
#include <linux/kernel.h>

#include "mlx5_ib.h"
#include "cmd.h"
#include "qp.h"

#include <linux/mlx5/eq.h>

/* Contains the details of a pagefault. */
struct mlx5_pagefault {
	u32			bytes_committed;
	u32			token;
	u8			event_subtype;
	u8			type;
	union {
		/* Initiator or send message responder pagefault details. */
		struct {
			/* Received packet size, only valid for responders. */
			u32	packet_size;
			/*
			 * Number of resource holding WQE, depends on type.
			 */
			u32	wq_num;
			/*
			 * WQE index. Refers to either the send queue or
			 * receive queue, according to event_subtype.
			 */
			u16	wqe_index;
		} wqe;
		/* RDMA responder pagefault details */
		struct {
			u32	r_key;
			/*
			 * Received packet size, minimal size page fault
			 * resolution required for forward progress.
			 */
			u32	packet_size;
			u32	rdma_op_len;
			u64	rdma_va;
		} rdma;
	};

	struct mlx5_ib_pf_eq	*eq;
	struct work_struct	work;
};

#define MAX_PREFETCH_LEN (4*1024*1024U)

/* Timeout in ms to wait for an active mmu notifier to complete when handling
 * a pagefault. */
#define MMU_NOTIFIER_TIMEOUT 1000

#define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
#define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
#define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
#define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
#define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))

#define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT

static u64 mlx5_imr_ksm_entries;

static void populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
			struct mlx5_ib_mr *imr, int flags)
{
	struct mlx5_klm *end = pklm + nentries;

	if (flags & MLX5_IB_UPD_XLT_ZAP) {
		for (; pklm != end; pklm++, idx++) {
			pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
			pklm->key = cpu_to_be32(imr->dev->null_mkey);
			pklm->va = 0;
		}
		return;
	}

	/*
	 * The locking here is pretty subtle. Ideally the implicit_children
	 * xarray would be protected by the umem_mutex, however that is not
	 * possible. Instead this uses a weaker update-then-lock pattern:
	 *
	 *  srcu_read_lock()
	 *    xa_store()
	 *    mutex_lock(umem_mutex)
	 *     mlx5_ib_update_xlt()
	 *    mutex_unlock(umem_mutex)
	 *    destroy lkey
	 *
	 * ie any change the xarray must be followed by the locked update_xlt
	 * before destroying.
	 *
	 * The umem_mutex provides the acquire/release semantic needed to make
	 * the xa_store() visible to a racing thread. While SRCU is not
	 * technically required, using it gives consistent use of the SRCU
	 * locking around the xarray.
	 */
	lockdep_assert_held(&to_ib_umem_odp(imr->umem)->umem_mutex);
	lockdep_assert_held(&imr->dev->odp_srcu);

	for (; pklm != end; pklm++, idx++) {
		struct mlx5_ib_mr *mtt = xa_load(&imr->implicit_children, idx);

		pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
		if (mtt) {
			pklm->key = cpu_to_be32(mtt->ibmr.lkey);
			pklm->va = cpu_to_be64(idx * MLX5_IMR_MTT_SIZE);
		} else {
			pklm->key = cpu_to_be32(imr->dev->null_mkey);
			pklm->va = 0;
		}
	}
}

static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
{
	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;

	if (umem_dma & ODP_READ_ALLOWED_BIT)
		mtt_entry |= MLX5_IB_MTT_READ;
	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
		mtt_entry |= MLX5_IB_MTT_WRITE;

	return mtt_entry;
}

static void populate_mtt(__be64 *pas, size_t idx, size_t nentries,
			 struct mlx5_ib_mr *mr, int flags)
{
	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
	dma_addr_t pa;
	size_t i;

	if (flags & MLX5_IB_UPD_XLT_ZAP)
		return;

	for (i = 0; i < nentries; i++) {
		pa = odp->dma_list[idx + i];
		pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
	}
}

void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
			   struct mlx5_ib_mr *mr, int flags)
{
	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
		populate_klm(xlt, idx, nentries, mr, flags);
	} else {
		populate_mtt(xlt, idx, nentries, mr, flags);
	}
}

static void dma_fence_odp_mr(struct mlx5_ib_mr *mr)
{
	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);

	/* Ensure mlx5_ib_invalidate_range() will not touch the MR any more */
	mutex_lock(&odp->umem_mutex);
	if (odp->npages) {
		mlx5_mr_cache_invalidate(mr);
		ib_umem_odp_unmap_dma_pages(odp, ib_umem_start(odp),
					    ib_umem_end(odp));
		WARN_ON(odp->npages);
	}
	odp->private = NULL;
	mutex_unlock(&odp->umem_mutex);

	if (!mr->cache_ent) {
		mlx5_core_destroy_mkey(mr->dev->mdev, &mr->mmkey);
		WARN_ON(mr->descs);
	}
}

/*
 * This must be called after the mr has been removed from implicit_children
 * and the SRCU synchronized.  NOTE: The MR does not necessarily have to be
 * empty here, parallel page faults could have raced with the free process and
 * added pages to it.
 */
static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt)
{
	struct mlx5_ib_mr *imr = mr->parent;
	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
	int srcu_key;

	/* implicit_child_mr's are not allowed to have deferred work */
	WARN_ON(atomic_read(&mr->num_deferred_work));

	if (need_imr_xlt) {
		srcu_key = srcu_read_lock(&mr->dev->odp_srcu);
		mutex_lock(&odp_imr->umem_mutex);
		mlx5_ib_update_xlt(mr->parent, idx, 1, 0,
				   MLX5_IB_UPD_XLT_INDIRECT |
				   MLX5_IB_UPD_XLT_ATOMIC);
		mutex_unlock(&odp_imr->umem_mutex);
		srcu_read_unlock(&mr->dev->odp_srcu, srcu_key);
	}

	dma_fence_odp_mr(mr);

	mr->parent = NULL;
	mlx5_mr_cache_free(mr->dev, mr);
	ib_umem_odp_release(odp);
	if (atomic_dec_and_test(&imr->num_deferred_work))
		wake_up(&imr->q_deferred_work);
}

static void free_implicit_child_mr_work(struct work_struct *work)
{
	struct mlx5_ib_mr *mr =
		container_of(work, struct mlx5_ib_mr, odp_destroy.work);

	free_implicit_child_mr(mr, true);
}

static void free_implicit_child_mr_rcu(struct rcu_head *head)
{
	struct mlx5_ib_mr *mr =
		container_of(head, struct mlx5_ib_mr, odp_destroy.rcu);

	/* Freeing a MR is a sleeping operation, so bounce to a work queue */
	INIT_WORK(&mr->odp_destroy.work, free_implicit_child_mr_work);
	queue_work(system_unbound_wq, &mr->odp_destroy.work);
}

static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
{
	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
	unsigned long idx = ib_umem_start(odp) >> MLX5_IMR_MTT_SHIFT;
	struct mlx5_ib_mr *imr = mr->parent;

	xa_lock(&imr->implicit_children);
	/*
	 * This can race with mlx5_ib_free_implicit_mr(), the first one to
	 * reach the xa lock wins the race and destroys the MR.
	 */
	if (__xa_cmpxchg(&imr->implicit_children, idx, mr, NULL, GFP_ATOMIC) !=
	    mr)
		goto out_unlock;

	atomic_inc(&imr->num_deferred_work);
	call_srcu(&mr->dev->odp_srcu, &mr->odp_destroy.rcu,
		  free_implicit_child_mr_rcu);

out_unlock:
	xa_unlock(&imr->implicit_children);
}

static bool mlx5_ib_invalidate_range(struct mmu_interval_notifier *mni,
				     const struct mmu_notifier_range *range,
				     unsigned long cur_seq)
{
	struct ib_umem_odp *umem_odp =
		container_of(mni, struct ib_umem_odp, notifier);
	struct mlx5_ib_mr *mr;
	const u64 umr_block_mask = (MLX5_UMR_MTT_ALIGNMENT /
				    sizeof(struct mlx5_mtt)) - 1;
	u64 idx = 0, blk_start_idx = 0;
	u64 invalidations = 0;
	unsigned long start;
	unsigned long end;
	int in_block = 0;
	u64 addr;

	if (!mmu_notifier_range_blockable(range))
		return false;

	mutex_lock(&umem_odp->umem_mutex);
	mmu_interval_set_seq(mni, cur_seq);
	/*
	 * If npages is zero then umem_odp->private may not be setup yet. This
	 * does not complete until after the first page is mapped for DMA.
	 */
	if (!umem_odp->npages)
		goto out;
	mr = umem_odp->private;

	start = max_t(u64, ib_umem_start(umem_odp), range->start);
	end = min_t(u64, ib_umem_end(umem_odp), range->end);

	/*
	 * Iteration one - zap the HW's MTTs. The notifiers_count ensures that
	 * while we are doing the invalidation, no page fault will attempt to
	 * overwrite the same MTTs.  Concurent invalidations might race us,
	 * but they will write 0s as well, so no difference in the end result.
	 */
	for (addr = start; addr < end; addr += BIT(umem_odp->page_shift)) {
		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
		/*
		 * Strive to write the MTTs in chunks, but avoid overwriting
		 * non-existing MTTs. The huristic here can be improved to
		 * estimate the cost of another UMR vs. the cost of bigger
		 * UMR.
		 */
		if (umem_odp->dma_list[idx] &
		    (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) {
			if (!in_block) {
				blk_start_idx = idx;
				in_block = 1;
			}

			/* Count page invalidations */
			invalidations += idx - blk_start_idx + 1;
		} else {
			u64 umr_offset = idx & umr_block_mask;

			if (in_block && umr_offset == 0) {
				mlx5_ib_update_xlt(mr, blk_start_idx,
						   idx - blk_start_idx, 0,
						   MLX5_IB_UPD_XLT_ZAP |
						   MLX5_IB_UPD_XLT_ATOMIC);
				in_block = 0;
			}
		}
	}
	if (in_block)
		mlx5_ib_update_xlt(mr, blk_start_idx,
				   idx - blk_start_idx + 1, 0,
				   MLX5_IB_UPD_XLT_ZAP |
				   MLX5_IB_UPD_XLT_ATOMIC);

	mlx5_update_odp_stats(mr, invalidations, invalidations);

	/*
	 * We are now sure that the device will not access the
	 * memory. We can safely unmap it, and mark it as dirty if
	 * needed.
	 */

	ib_umem_odp_unmap_dma_pages(umem_odp, start, end);

	if (unlikely(!umem_odp->npages && mr->parent))
		destroy_unused_implicit_child_mr(mr);
out:
	mutex_unlock(&umem_odp->umem_mutex);
	return true;
}

const struct mmu_interval_notifier_ops mlx5_mn_ops = {
	.invalidate = mlx5_ib_invalidate_range,
};

void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
{
	struct ib_odp_caps *caps = &dev->odp_caps;

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

	if (!MLX5_CAP_GEN(dev->mdev, pg) ||
	    !mlx5_ib_can_use_umr(dev, true, 0))
		return;

	caps->general_caps = IB_ODP_SUPPORT;

	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
		dev->odp_max_size = U64_MAX;
	else
		dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT);

	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send))
		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;

	if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive))
		caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;

	if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive))
		caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;

	if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive))
		caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;

	if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
	    MLX5_CAP_GEN(dev->mdev, null_mkey) &&
	    MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset) &&
	    !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled))
		caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
}

static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
				      struct mlx5_pagefault *pfault,
				      int error)
{
	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
		     pfault->wqe.wq_num : pfault->token;
	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {};
	int err;

	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
	MLX5_SET(page_fault_resume_in, in, error, !!error);

	err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in);
	if (err)
		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
			    wq_num, err);
}

static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
						unsigned long idx)
{
	struct ib_umem_odp *odp;
	struct mlx5_ib_mr *mr;
	struct mlx5_ib_mr *ret;
	int err;

	odp = ib_umem_odp_alloc_child(to_ib_umem_odp(imr->umem),
				      idx * MLX5_IMR_MTT_SIZE,
				      MLX5_IMR_MTT_SIZE, &mlx5_mn_ops);
	if (IS_ERR(odp))
		return ERR_CAST(odp);

	ret = mr = mlx5_mr_cache_alloc(imr->dev, MLX5_IMR_MTT_CACHE_ENTRY);
	if (IS_ERR(mr))
		goto out_umem;

	mr->ibmr.pd = imr->ibmr.pd;
	mr->access_flags = imr->access_flags;
	mr->umem = &odp->umem;
	mr->ibmr.lkey = mr->mmkey.key;
	mr->ibmr.rkey = mr->mmkey.key;
	mr->mmkey.iova = idx * MLX5_IMR_MTT_SIZE;
	mr->parent = imr;
	odp->private = mr;

	err = mlx5_ib_update_xlt(mr, 0,
				 MLX5_IMR_MTT_ENTRIES,
				 PAGE_SHIFT,
				 MLX5_IB_UPD_XLT_ZAP |
				 MLX5_IB_UPD_XLT_ENABLE);
	if (err) {
		ret = ERR_PTR(err);
		goto out_mr;
	}

	/*
	 * Once the store to either xarray completes any error unwind has to
	 * use synchronize_srcu(). Avoid this with xa_reserve()
	 */
	ret = xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
			 GFP_KERNEL);
	if (unlikely(ret)) {
		if (xa_is_err(ret)) {
			ret = ERR_PTR(xa_err(ret));
			goto out_mr;
		}
		/*
		 * Another thread beat us to creating the child mr, use
		 * theirs.
		 */
		goto out_mr;
	}

	mlx5_ib_dbg(imr->dev, "key %x mr %p\n", mr->mmkey.key, mr);
	return mr;

out_mr:
	mlx5_mr_cache_free(imr->dev, mr);
out_umem:
	ib_umem_odp_release(odp);
	return ret;
}

struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
					     struct ib_udata *udata,
					     int access_flags)
{
	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
	struct ib_umem_odp *umem_odp;
	struct mlx5_ib_mr *imr;
	int err;

	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
	if (IS_ERR(umem_odp))
		return ERR_CAST(umem_odp);

	imr = mlx5_mr_cache_alloc(dev, MLX5_IMR_KSM_CACHE_ENTRY);
	if (IS_ERR(imr)) {
		err = PTR_ERR(imr);
		goto out_umem;
	}

	imr->ibmr.pd = &pd->ibpd;
	imr->access_flags = access_flags;
	imr->mmkey.iova = 0;
	imr->umem = &umem_odp->umem;
	imr->ibmr.lkey = imr->mmkey.key;
	imr->ibmr.rkey = imr->mmkey.key;
	imr->umem = &umem_odp->umem;
	imr->is_odp_implicit = true;
	atomic_set(&imr->num_deferred_work, 0);
	init_waitqueue_head(&imr->q_deferred_work);
	xa_init(&imr->implicit_children);

	err = mlx5_ib_update_xlt(imr, 0,
				 mlx5_imr_ksm_entries,
				 MLX5_KSM_PAGE_SHIFT,
				 MLX5_IB_UPD_XLT_INDIRECT |
				 MLX5_IB_UPD_XLT_ZAP |
				 MLX5_IB_UPD_XLT_ENABLE);
	if (err)
		goto out_mr;

	err = xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key),
			      &imr->mmkey, GFP_KERNEL));
	if (err)
		goto out_mr;

	mlx5_ib_dbg(dev, "key %x mr %p\n", imr->mmkey.key, imr);
	return imr;
out_mr:
	mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
	mlx5_mr_cache_free(dev, imr);
out_umem:
	ib_umem_odp_release(umem_odp);
	return ERR_PTR(err);
}

void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
{
	struct ib_umem_odp *odp_imr = to_ib_umem_odp(imr->umem);
	struct mlx5_ib_dev *dev = imr->dev;
	struct list_head destroy_list;
	struct mlx5_ib_mr *mtt;
	struct mlx5_ib_mr *tmp;
	unsigned long idx;

	INIT_LIST_HEAD(&destroy_list);

	xa_erase(&dev->odp_mkeys, mlx5_base_mkey(imr->mmkey.key));
	/*
	 * This stops the SRCU protected page fault path from touching either
	 * the imr or any children. The page fault path can only reach the
	 * children xarray via the imr.
	 */
	synchronize_srcu(&dev->odp_srcu);

	/*
	 * All work on the prefetch list must be completed, xa_erase() prevented
	 * new work from being created.
	 */
	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));

	/*
	 * At this point it is forbidden for any other thread to enter
	 * pagefault_mr() on this imr. It is already forbidden to call
	 * pagefault_mr() on an implicit child. Due to this additions to
	 * implicit_children are prevented.
	 */

	/*
	 * Block destroy_unused_implicit_child_mr() from incrementing
	 * num_deferred_work.
	 */
	xa_lock(&imr->implicit_children);
	xa_for_each (&imr->implicit_children, idx, mtt) {
		__xa_erase(&imr->implicit_children, idx);
		list_add(&mtt->odp_destroy.elm, &destroy_list);
	}
	xa_unlock(&imr->implicit_children);

	/*
	 * Wait for any concurrent destroy_unused_implicit_child_mr() to
	 * complete.
	 */
	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));

	/*
	 * Fence the imr before we destroy the children. This allows us to
	 * skip updating the XLT of the imr during destroy of the child mkey
	 * the imr points to.
	 */
	mlx5_mr_cache_invalidate(imr);

	list_for_each_entry_safe (mtt, tmp, &destroy_list, odp_destroy.elm)
		free_implicit_child_mr(mtt, false);

	mlx5_mr_cache_free(dev, imr);
	ib_umem_odp_release(odp_imr);
}

/**
 * mlx5_ib_fence_odp_mr - Stop all access to the ODP MR
 * @mr: to fence
 *
 * On return no parallel threads will be touching this MR and no DMA will be
 * active.
 */
void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr)
{
	/* Prevent new page faults and prefetch requests from succeeding */
	xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));

	/* Wait for all running page-fault handlers to finish. */
	synchronize_srcu(&mr->dev->odp_srcu);

	wait_event(mr->q_deferred_work, !atomic_read(&mr->num_deferred_work));

	dma_fence_odp_mr(mr);
}

#define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
			     u64 user_va, size_t bcnt, u32 *bytes_mapped,
			     u32 flags)
{
	int page_shift, ret, np;
	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
	unsigned long current_seq;
	u64 access_mask;
	u64 start_idx;

	page_shift = odp->page_shift;
	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
	access_mask = ODP_READ_ALLOWED_BIT;

	if (odp->umem.writable && !downgrade)
		access_mask |= ODP_WRITE_ALLOWED_BIT;

	current_seq = mmu_interval_read_begin(&odp->notifier);

	np = ib_umem_odp_map_dma_pages(odp, user_va, bcnt, access_mask,
				       current_seq);
	if (np < 0)
		return np;

	mutex_lock(&odp->umem_mutex);
	if (!mmu_interval_read_retry(&odp->notifier, current_seq)) {
		/*
		 * No need to check whether the MTTs really belong to
		 * this MR, since ib_umem_odp_map_dma_pages already
		 * checks this.
		 */
		ret = mlx5_ib_update_xlt(mr, start_idx, np,
					 page_shift, MLX5_IB_UPD_XLT_ATOMIC);
	} else {
		ret = -EAGAIN;
	}
	mutex_unlock(&odp->umem_mutex);

	if (ret < 0) {
		if (ret != -EAGAIN)
			mlx5_ib_err(mr->dev,
				    "Failed to update mkey page tables\n");
		goto out;
	}

	if (bytes_mapped) {
		u32 new_mappings = (np << page_shift) -
			(user_va - round_down(user_va, 1 << page_shift));

		*bytes_mapped += min_t(u32, new_mappings, bcnt);
	}

	return np << (page_shift - PAGE_SHIFT);

out:
	return ret;
}

static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
				 struct ib_umem_odp *odp_imr, u64 user_va,
				 size_t bcnt, u32 *bytes_mapped, u32 flags)
{
	unsigned long end_idx = (user_va + bcnt - 1) >> MLX5_IMR_MTT_SHIFT;
	unsigned long upd_start_idx = end_idx + 1;
	unsigned long upd_len = 0;
	unsigned long npages = 0;
	int err;
	int ret;

	if (unlikely(user_va >= mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE ||
		     mlx5_imr_ksm_entries * MLX5_IMR_MTT_SIZE - user_va < bcnt))
		return -EFAULT;

	/* Fault each child mr that intersects with our interval. */
	while (bcnt) {
		unsigned long idx = user_va >> MLX5_IMR_MTT_SHIFT;
		struct ib_umem_odp *umem_odp;
		struct mlx5_ib_mr *mtt;
		u64 len;

		mtt = xa_load(&imr->implicit_children, idx);
		if (unlikely(!mtt)) {
			mtt = implicit_get_child_mr(imr, idx);
			if (IS_ERR(mtt)) {
				ret = PTR_ERR(mtt);
				goto out;
			}
			upd_start_idx = min(upd_start_idx, idx);
			upd_len = idx - upd_start_idx + 1;
		}

		umem_odp = to_ib_umem_odp(mtt->umem);
		len = min_t(u64, user_va + bcnt, ib_umem_end(umem_odp)) -
		      user_va;

		ret = pagefault_real_mr(mtt, umem_odp, user_va, len,
					bytes_mapped, flags);
		if (ret < 0)
			goto out;
		user_va += len;
		bcnt -= len;
		npages += ret;
	}

	ret = npages;

	/*
	 * Any time the implicit_children are changed we must perform an
	 * update of the xlt before exiting to ensure the HW and the
	 * implicit_children remains synchronized.
	 */
out:
	if (likely(!upd_len))
		return ret;

	/*
	 * Notice this is not strictly ordered right, the KSM is updated after
	 * the implicit_children is updated, so a parallel page fault could
	 * see a MR that is not yet visible in the KSM.  This is similar to a
	 * parallel page fault seeing a MR that is being concurrently removed
	 * from the KSM. Both of these improbable situations are resolved
	 * safely by resuming the HW and then taking another page fault. The
	 * next pagefault handler will see the new information.
	 */
	mutex_lock(&odp_imr->umem_mutex);
	err = mlx5_ib_update_xlt(imr, upd_start_idx, upd_len, 0,
				 MLX5_IB_UPD_XLT_INDIRECT |
					 MLX5_IB_UPD_XLT_ATOMIC);
	mutex_unlock(&odp_imr->umem_mutex);
	if (err) {
		mlx5_ib_err(imr->dev, "Failed to update PAS\n");
		return err;
	}
	return ret;
}

/*
 * Returns:
 *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
 *           not accessible, or the MR is no longer valid.
 *  -EAGAIN/-ENOMEM: The operation should be retried
 *
 *  -EINVAL/others: General internal malfunction
 *  >0: Number of pages mapped
 */
static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
			u32 *bytes_mapped, u32 flags)
{
	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);

	if (unlikely(io_virt < mr->mmkey.iova))
		return -EFAULT;

	if (!odp->is_implicit_odp) {
		u64 user_va;

		if (check_add_overflow(io_virt - mr->mmkey.iova,
				       (u64)odp->umem.address, &user_va))
			return -EFAULT;
		if (unlikely(user_va >= ib_umem_end(odp) ||
			     ib_umem_end(odp) - user_va < bcnt))
			return -EFAULT;
		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
					 flags);
	}
	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
				     flags);
}

struct pf_frame {
	struct pf_frame *next;
	u32 key;
	u64 io_virt;
	size_t bcnt;
	int depth;
};

static bool mkey_is_eq(struct mlx5_core_mkey *mmkey, u32 key)
{
	if (!mmkey)
		return false;
	if (mmkey->type == MLX5_MKEY_MW)
		return mlx5_base_mkey(mmkey->key) == mlx5_base_mkey(key);
	return mmkey->key == key;
}

static int get_indirect_num_descs(struct mlx5_core_mkey *mmkey)
{
	struct mlx5_ib_mw *mw;
	struct mlx5_ib_devx_mr *devx_mr;

	if (mmkey->type == MLX5_MKEY_MW) {
		mw = container_of(mmkey, struct mlx5_ib_mw, mmkey);
		return mw->ndescs;
	}

	devx_mr = container_of(mmkey, struct mlx5_ib_devx_mr,
			       mmkey);
	return devx_mr->ndescs;
}

/*
 * Handle a single data segment in a page-fault WQE or RDMA region.
 *
 * Returns number of OS pages retrieved on success. The caller may continue to
 * the next data segment.
 * Can return the following error codes:
 * -EAGAIN to designate a temporary error. The caller will abort handling the
 *  page fault and resolve it.
 * -EFAULT when there's an error mapping the requested pages. The caller will
 *  abort the page fault handling.
 */
static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
					 struct ib_pd *pd, u32 key,
					 u64 io_virt, size_t bcnt,
					 u32 *bytes_committed,
					 u32 *bytes_mapped)
{
	int npages = 0, srcu_key, ret, i, outlen, cur_outlen = 0, depth = 0;
	struct pf_frame *head = NULL, *frame;
	struct mlx5_core_mkey *mmkey;
	struct mlx5_ib_mr *mr;
	struct mlx5_klm *pklm;
	u32 *out = NULL;
	size_t offset;
	int ndescs;

	srcu_key = srcu_read_lock(&dev->odp_srcu);

	io_virt += *bytes_committed;
	bcnt -= *bytes_committed;

next_mr:
	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(key));
	if (!mmkey) {
		mlx5_ib_dbg(
			dev,
			"skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
			key);
		if (bytes_mapped)
			*bytes_mapped += bcnt;
		/*
		 * The user could specify a SGL with multiple lkeys and only
		 * some of them are ODP. Treat the non-ODP ones as fully
		 * faulted.
		 */
		ret = 0;
		goto srcu_unlock;
	}
	if (!mkey_is_eq(mmkey, key)) {
		mlx5_ib_dbg(dev, "failed to find mkey %x\n", key);
		ret = -EFAULT;
		goto srcu_unlock;
	}

	switch (mmkey->type) {
	case MLX5_MKEY_MR:
		mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);

		ret = pagefault_mr(mr, io_virt, bcnt, bytes_mapped, 0);
		if (ret < 0)
			goto srcu_unlock;

		/*
		 * When prefetching a page, page fault is generated
		 * in order to bring the page to the main memory.
		 * In the current flow, page faults are being counted.
		 */
		mlx5_update_odp_stats(mr, faults, ret);

		npages += ret;
		ret = 0;
		break;

	case MLX5_MKEY_MW:
	case MLX5_MKEY_INDIRECT_DEVX:
		ndescs = get_indirect_num_descs(mmkey);

		if (depth >= MLX5_CAP_GEN(dev->mdev, max_indirection)) {
			mlx5_ib_dbg(dev, "indirection level exceeded\n");
			ret = -EFAULT;
			goto srcu_unlock;
		}

		outlen = MLX5_ST_SZ_BYTES(query_mkey_out) +
			sizeof(*pklm) * (ndescs - 2);

		if (outlen > cur_outlen) {
			kfree(out);
			out = kzalloc(outlen, GFP_KERNEL);
			if (!out) {
				ret = -ENOMEM;
				goto srcu_unlock;
			}
			cur_outlen = outlen;
		}

		pklm = (struct mlx5_klm *)MLX5_ADDR_OF(query_mkey_out, out,
						       bsf0_klm0_pas_mtt0_1);

		ret = mlx5_core_query_mkey(dev->mdev, mmkey, out, outlen);
		if (ret)
			goto srcu_unlock;

		offset = io_virt - MLX5_GET64(query_mkey_out, out,
					      memory_key_mkey_entry.start_addr);

		for (i = 0; bcnt && i < ndescs; i++, pklm++) {
			if (offset >= be32_to_cpu(pklm->bcount)) {
				offset -= be32_to_cpu(pklm->bcount);
				continue;
			}

			frame = kzalloc(sizeof(*frame), GFP_KERNEL);
			if (!frame) {
				ret = -ENOMEM;
				goto srcu_unlock;
			}

			frame->key = be32_to_cpu(pklm->key);
			frame->io_virt = be64_to_cpu(pklm->va) + offset;
			frame->bcnt = min_t(size_t, bcnt,
					    be32_to_cpu(pklm->bcount) - offset);
			frame->depth = depth + 1;
			frame->next = head;
			head = frame;

			bcnt -= frame->bcnt;
			offset = 0;
		}
		break;

	default:
		mlx5_ib_dbg(dev, "wrong mkey type %d\n", mmkey->type);
		ret = -EFAULT;
		goto srcu_unlock;
	}

	if (head) {
		frame = head;
		head = frame->next;

		key = frame->key;
		io_virt = frame->io_virt;
		bcnt = frame->bcnt;
		depth = frame->depth;
		kfree(frame);

		goto next_mr;
	}

srcu_unlock:
	while (head) {
		frame = head;
		head = frame->next;
		kfree(frame);
	}
	kfree(out);

	srcu_read_unlock(&dev->odp_srcu, srcu_key);
	*bytes_committed = 0;
	return ret ? ret : npages;
}

/**
 * Parse a series of data segments for page fault handling.
 *
 * @pfault contains page fault information.
 * @wqe points at the first data segment in the WQE.
 * @wqe_end points after the end of the WQE.
 * @bytes_mapped receives the number of bytes that the function was able to
 *               map. This allows the caller to decide intelligently whether
 *               enough memory was mapped to resolve the page fault
 *               successfully (e.g. enough for the next MTU, or the entire
 *               WQE).
 * @total_wqe_bytes receives the total data size of this WQE in bytes (minus
 *                  the committed bytes).
 *
 * Returns the number of pages loaded if positive, zero for an empty WQE, or a
 * negative error code.
 */
static int pagefault_data_segments(struct mlx5_ib_dev *dev,
				   struct mlx5_pagefault *pfault,
				   void *wqe,
				   void *wqe_end, u32 *bytes_mapped,
				   u32 *total_wqe_bytes, bool receive_queue)
{
	int ret = 0, npages = 0;
	u64 io_virt;
	u32 key;
	u32 byte_count;
	size_t bcnt;
	int inline_segment;

	if (bytes_mapped)
		*bytes_mapped = 0;
	if (total_wqe_bytes)
		*total_wqe_bytes = 0;

	while (wqe < wqe_end) {
		struct mlx5_wqe_data_seg *dseg = wqe;

		io_virt = be64_to_cpu(dseg->addr);
		key = be32_to_cpu(dseg->lkey);
		byte_count = be32_to_cpu(dseg->byte_count);
		inline_segment = !!(byte_count &  MLX5_INLINE_SEG);
		bcnt	       = byte_count & ~MLX5_INLINE_SEG;

		if (inline_segment) {
			bcnt = bcnt & MLX5_WQE_INLINE_SEG_BYTE_COUNT_MASK;
			wqe += ALIGN(sizeof(struct mlx5_wqe_inline_seg) + bcnt,
				     16);
		} else {
			wqe += sizeof(*dseg);
		}

		/* receive WQE end of sg list. */
		if (receive_queue && bcnt == 0 && key == MLX5_INVALID_LKEY &&
		    io_virt == 0)
			break;

		if (!inline_segment && total_wqe_bytes) {
			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
					pfault->bytes_committed);
		}

		/* A zero length data segment designates a length of 2GB. */
		if (bcnt == 0)
			bcnt = 1U << 31;

		if (inline_segment || bcnt <= pfault->bytes_committed) {
			pfault->bytes_committed -=
				min_t(size_t, bcnt,
				      pfault->bytes_committed);
			continue;
		}

		ret = pagefault_single_data_segment(dev, NULL, key,
						    io_virt, bcnt,
						    &pfault->bytes_committed,
						    bytes_mapped);
		if (ret < 0)
			break;
		npages += ret;
	}

	return ret < 0 ? ret : npages;
}

/*
 * Parse initiator WQE. Advances the wqe pointer to point at the
 * scatter-gather list, and set wqe_end to the end of the WQE.
 */
static int mlx5_ib_mr_initiator_pfault_handler(
	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
{
	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
	u16 wqe_index = pfault->wqe.wqe_index;
	struct mlx5_base_av *av;
	unsigned ds, opcode;
	u32 qpn = qp->trans_qp.base.mqp.qpn;

	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
	if (ds * MLX5_WQE_DS_UNITS > wqe_length) {
		mlx5_ib_err(dev, "Unable to read the complete WQE. ds = 0x%x, ret = 0x%x\n",
			    ds, wqe_length);
		return -EFAULT;
	}

	if (ds == 0) {
		mlx5_ib_err(dev, "Got WQE with zero DS. wqe_index=%x, qpn=%x\n",
			    wqe_index, qpn);
		return -EFAULT;
	}

	*wqe_end = *wqe + ds * MLX5_WQE_DS_UNITS;
	*wqe += sizeof(*ctrl);

	opcode = be32_to_cpu(ctrl->opmod_idx_opcode) &
		 MLX5_WQE_CTRL_OPCODE_MASK;

	if (qp->ibqp.qp_type == IB_QPT_XRC_INI)
		*wqe += sizeof(struct mlx5_wqe_xrc_seg);

	if (qp->type == IB_QPT_UD || qp->type == MLX5_IB_QPT_DCI) {
		av = *wqe;
		if (av->dqp_dct & cpu_to_be32(MLX5_EXTENDED_UD_AV))
			*wqe += sizeof(struct mlx5_av);
		else
			*wqe += sizeof(struct mlx5_base_av);
	}

	switch (opcode) {
	case MLX5_OPCODE_RDMA_WRITE:
	case MLX5_OPCODE_RDMA_WRITE_IMM:
	case MLX5_OPCODE_RDMA_READ:
		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
		break;
	case MLX5_OPCODE_ATOMIC_CS:
	case MLX5_OPCODE_ATOMIC_FA:
		*wqe += sizeof(struct mlx5_wqe_raddr_seg);
		*wqe += sizeof(struct mlx5_wqe_atomic_seg);
		break;
	}

	return 0;
}

/*
 * Parse responder WQE and set wqe_end to the end of the WQE.
 */
static int mlx5_ib_mr_responder_pfault_handler_srq(struct mlx5_ib_dev *dev,
						   struct mlx5_ib_srq *srq,
						   void **wqe, void **wqe_end,
						   int wqe_length)
{
	int wqe_size = 1 << srq->msrq.wqe_shift;

	if (wqe_size > wqe_length) {
		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
		return -EFAULT;
	}

	*wqe_end = *wqe + wqe_size;
	*wqe += sizeof(struct mlx5_wqe_srq_next_seg);

	return 0;
}

static int mlx5_ib_mr_responder_pfault_handler_rq(struct mlx5_ib_dev *dev,
						  struct mlx5_ib_qp *qp,
						  void *wqe, void **wqe_end,
						  int wqe_length)
{
	struct mlx5_ib_wq *wq = &qp->rq;
	int wqe_size = 1 << wq->wqe_shift;

	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE) {
		mlx5_ib_err(dev, "ODP fault with WQE signatures is not supported\n");
		return -EFAULT;
	}

	if (wqe_size > wqe_length) {
		mlx5_ib_err(dev, "Couldn't read all of the receive WQE's content\n");
		return -EFAULT;
	}

	*wqe_end = wqe + wqe_size;

	return 0;
}

static inline struct mlx5_core_rsc_common *odp_get_rsc(struct mlx5_ib_dev *dev,
						       u32 wq_num, int pf_type)
{
	struct mlx5_core_rsc_common *common = NULL;
	struct mlx5_core_srq *srq;

	switch (pf_type) {
	case MLX5_WQE_PF_TYPE_RMP:
		srq = mlx5_cmd_get_srq(dev, wq_num);
		if (srq)
			common = &srq->common;
		break;
	case MLX5_WQE_PF_TYPE_REQ_SEND_OR_WRITE:
	case MLX5_WQE_PF_TYPE_RESP:
	case MLX5_WQE_PF_TYPE_REQ_READ_OR_ATOMIC:
		common = mlx5_core_res_hold(dev, wq_num, MLX5_RES_QP);
		break;
	default:
		break;
	}

	return common;
}

static inline struct mlx5_ib_qp *res_to_qp(struct mlx5_core_rsc_common *res)
{
	struct mlx5_core_qp *mqp = (struct mlx5_core_qp *)res;

	return to_mibqp(mqp);
}

static inline struct mlx5_ib_srq *res_to_srq(struct mlx5_core_rsc_common *res)
{
	struct mlx5_core_srq *msrq =
		container_of(res, struct mlx5_core_srq, common);

	return to_mibsrq(msrq);
}

static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
					  struct mlx5_pagefault *pfault)
{
	bool sq = pfault->type & MLX5_PFAULT_REQUESTOR;
	u16 wqe_index = pfault->wqe.wqe_index;
	void *wqe, *wqe_start = NULL, *wqe_end = NULL;
	u32 bytes_mapped, total_wqe_bytes;
	struct mlx5_core_rsc_common *res;
	int resume_with_error = 1;
	struct mlx5_ib_qp *qp;
	size_t bytes_copied;
	int ret = 0;

	res = odp_get_rsc(dev, pfault->wqe.wq_num, pfault->type);
	if (!res) {
		mlx5_ib_dbg(dev, "wqe page fault for missing resource %d\n", pfault->wqe.wq_num);
		return;
	}

	if (res->res != MLX5_RES_QP && res->res != MLX5_RES_SRQ &&
	    res->res != MLX5_RES_XSRQ) {
		mlx5_ib_err(dev, "wqe page fault for unsupported type %d\n",
			    pfault->type);
		goto resolve_page_fault;
	}

	wqe_start = (void *)__get_free_page(GFP_KERNEL);
	if (!wqe_start) {
		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
		goto resolve_page_fault;
	}

	wqe = wqe_start;
	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
	if (qp && sq) {
		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
					  &bytes_copied);
		if (ret)
			goto read_user;
		ret = mlx5_ib_mr_initiator_pfault_handler(
			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
	} else if (qp && !sq) {
		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
					  &bytes_copied);
		if (ret)
			goto read_user;
		ret = mlx5_ib_mr_responder_pfault_handler_rq(
			dev, qp, wqe, &wqe_end, bytes_copied);
	} else if (!qp) {
		struct mlx5_ib_srq *srq = res_to_srq(res);

		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
					   &bytes_copied);
		if (ret)
			goto read_user;
		ret = mlx5_ib_mr_responder_pfault_handler_srq(
			dev, srq, &wqe, &wqe_end, bytes_copied);
	}

	if (ret < 0 || wqe >= wqe_end)
		goto resolve_page_fault;

	ret = pagefault_data_segments(dev, pfault, wqe, wqe_end, &bytes_mapped,
				      &total_wqe_bytes, !sq);
	if (ret == -EAGAIN)
		goto out;

	if (ret < 0 || total_wqe_bytes > bytes_mapped)
		goto resolve_page_fault;

out:
	ret = 0;
	resume_with_error = 0;

read_user:
	if (ret)
		mlx5_ib_err(
			dev,
			"Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n",
			ret, wqe_index, pfault->token);

resolve_page_fault:
	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
		    pfault->wqe.wq_num, resume_with_error,
		    pfault->type);
	mlx5_core_res_put(res);
	free_page((unsigned long)wqe_start);
}

static int pages_in_range(u64 address, u32 length)
{
	return (ALIGN(address + length, PAGE_SIZE) -
		(address & PAGE_MASK)) >> PAGE_SHIFT;
}

static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
					   struct mlx5_pagefault *pfault)
{
	u64 address;
	u32 length;
	u32 prefetch_len = pfault->bytes_committed;
	int prefetch_activated = 0;
	u32 rkey = pfault->rdma.r_key;
	int ret;

	/* The RDMA responder handler handles the page fault in two parts.
	 * First it brings the necessary pages for the current packet
	 * (and uses the pfault context), and then (after resuming the QP)
	 * prefetches more pages. The second operation cannot use the pfault
	 * context and therefore uses the dummy_pfault context allocated on
	 * the stack */
	pfault->rdma.rdma_va += pfault->bytes_committed;
	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
					 pfault->rdma.rdma_op_len);
	pfault->bytes_committed = 0;

	address = pfault->rdma.rdma_va;
	length  = pfault->rdma.rdma_op_len;

	/* For some operations, the hardware cannot tell the exact message
	 * length, and in those cases it reports zero. Use prefetch
	 * logic. */
	if (length == 0) {
		prefetch_activated = 1;
		length = pfault->rdma.packet_size;
		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
	}

	ret = pagefault_single_data_segment(dev, NULL, rkey, address, length,
					    &pfault->bytes_committed, NULL);
	if (ret == -EAGAIN) {
		/* We're racing with an invalidation, don't prefetch */
		prefetch_activated = 0;
	} else if (ret < 0 || pages_in_range(address, length) > ret) {
		mlx5_ib_page_fault_resume(dev, pfault, 1);
		if (ret != -ENOENT)
			mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
				    ret, pfault->token, pfault->type);
		return;
	}

	mlx5_ib_page_fault_resume(dev, pfault, 0);
	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
		    pfault->token, pfault->type,
		    prefetch_activated);

	/* At this point, there might be a new pagefault already arriving in
	 * the eq, switch to the dummy pagefault for the rest of the
	 * processing. We're still OK with the objects being alive as the
	 * work-queue is being fenced. */

	if (prefetch_activated) {
		u32 bytes_committed = 0;

		ret = pagefault_single_data_segment(dev, NULL, rkey, address,
						    prefetch_len,
						    &bytes_committed, NULL);
		if (ret < 0 && ret != -EAGAIN) {
			mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
				    ret, pfault->token, address, prefetch_len);
		}
	}
}

static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
{
	u8 event_subtype = pfault->event_subtype;

	switch (event_subtype) {
	case MLX5_PFAULT_SUBTYPE_WQE:
		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
		break;
	case MLX5_PFAULT_SUBTYPE_RDMA:
		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
		break;
	default:
		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
			    event_subtype);
		mlx5_ib_page_fault_resume(dev, pfault, 1);
	}
}

static void mlx5_ib_eqe_pf_action(struct work_struct *work)
{
	struct mlx5_pagefault *pfault = container_of(work,
						     struct mlx5_pagefault,
						     work);
	struct mlx5_ib_pf_eq *eq = pfault->eq;

	mlx5_ib_pfault(eq->dev, pfault);
	mempool_free(pfault, eq->pool);
}

static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
{
	struct mlx5_eqe_page_fault *pf_eqe;
	struct mlx5_pagefault *pfault;
	struct mlx5_eqe *eqe;
	int cc = 0;

	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
		if (!pfault) {
			schedule_work(&eq->work);
			break;
		}

		pf_eqe = &eqe->data.page_fault;
		pfault->event_subtype = eqe->sub_type;
		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);

		mlx5_ib_dbg(eq->dev,
			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
			    eqe->sub_type, pfault->bytes_committed);

		switch (eqe->sub_type) {
		case MLX5_PFAULT_SUBTYPE_RDMA:
			/* RDMA based event */
			pfault->type =
				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
			pfault->token =
				be32_to_cpu(pf_eqe->rdma.pftype_token) &
				MLX5_24BIT_MASK;
			pfault->rdma.r_key =
				be32_to_cpu(pf_eqe->rdma.r_key);
			pfault->rdma.packet_size =
				be16_to_cpu(pf_eqe->rdma.packet_length);
			pfault->rdma.rdma_op_len =
				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
			pfault->rdma.rdma_va =
				be64_to_cpu(pf_eqe->rdma.rdma_va);
			mlx5_ib_dbg(eq->dev,
				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
				    pfault->type, pfault->token,
				    pfault->rdma.r_key);
			mlx5_ib_dbg(eq->dev,
				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
				    pfault->rdma.rdma_op_len,
				    pfault->rdma.rdma_va);
			break;

		case MLX5_PFAULT_SUBTYPE_WQE:
			/* WQE based event */
			pfault->type =
				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
			pfault->token =
				be32_to_cpu(pf_eqe->wqe.token);
			pfault->wqe.wq_num =
				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
				MLX5_24BIT_MASK;
			pfault->wqe.wqe_index =
				be16_to_cpu(pf_eqe->wqe.wqe_index);
			pfault->wqe.packet_size =
				be16_to_cpu(pf_eqe->wqe.packet_length);
			mlx5_ib_dbg(eq->dev,
				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
				    pfault->type, pfault->token,
				    pfault->wqe.wq_num,
				    pfault->wqe.wqe_index);
			break;

		default:
			mlx5_ib_warn(eq->dev,
				     "Unsupported page fault event sub-type: 0x%02hhx\n",
				     eqe->sub_type);
			/* Unsupported page faults should still be
			 * resolved by the page fault handler
			 */
		}

		pfault->eq = eq;
		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
		queue_work(eq->wq, &pfault->work);

		cc = mlx5_eq_update_cc(eq->core, ++cc);
	}

	mlx5_eq_update_ci(eq->core, cc, 1);
}

static int mlx5_ib_eq_pf_int(struct notifier_block *nb, unsigned long type,
			     void *data)
{
	struct mlx5_ib_pf_eq *eq =
		container_of(nb, struct mlx5_ib_pf_eq, irq_nb);
	unsigned long flags;

	if (spin_trylock_irqsave(&eq->lock, flags)) {
		mlx5_ib_eq_pf_process(eq);
		spin_unlock_irqrestore(&eq->lock, flags);
	} else {
		schedule_work(&eq->work);
	}

	return IRQ_HANDLED;
}

/* mempool_refill() was proposed but unfortunately wasn't accepted
 * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
 * Cheap workaround.
 */
static void mempool_refill(mempool_t *pool)
{
	while (pool->curr_nr < pool->min_nr)
		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
}

static void mlx5_ib_eq_pf_action(struct work_struct *work)
{
	struct mlx5_ib_pf_eq *eq =
		container_of(work, struct mlx5_ib_pf_eq, work);

	mempool_refill(eq->pool);

	spin_lock_irq(&eq->lock);
	mlx5_ib_eq_pf_process(eq);
	spin_unlock_irq(&eq->lock);
}

enum {
	MLX5_IB_NUM_PF_EQE	= 0x1000,
	MLX5_IB_NUM_PF_DRAIN	= 64,
};

static int
mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
{
	struct mlx5_eq_param param = {};
	int err;

	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
	spin_lock_init(&eq->lock);
	eq->dev = dev;

	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
					       sizeof(struct mlx5_pagefault));
	if (!eq->pool)
		return -ENOMEM;

	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
				 MLX5_NUM_CMD_EQE);
	if (!eq->wq) {
		err = -ENOMEM;
		goto err_mempool;
	}

	eq->irq_nb.notifier_call = mlx5_ib_eq_pf_int;
	param = (struct mlx5_eq_param) {
		.irq_index = 0,
		.nent = MLX5_IB_NUM_PF_EQE,
	};
	param.mask[0] = 1ull << MLX5_EVENT_TYPE_PAGE_FAULT;
	eq->core = mlx5_eq_create_generic(dev->mdev, &param);
	if (IS_ERR(eq->core)) {
		err = PTR_ERR(eq->core);
		goto err_wq;
	}
	err = mlx5_eq_enable(dev->mdev, eq->core, &eq->irq_nb);
	if (err) {
		mlx5_ib_err(dev, "failed to enable odp EQ %d\n", err);
		goto err_eq;
	}

	return 0;
err_eq:
	mlx5_eq_destroy_generic(dev->mdev, eq->core);
err_wq:
	destroy_workqueue(eq->wq);
err_mempool:
	mempool_destroy(eq->pool);
	return err;
}

static int
mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
{
	int err;

	mlx5_eq_disable(dev->mdev, eq->core, &eq->irq_nb);
	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
	cancel_work_sync(&eq->work);
	destroy_workqueue(eq->wq);
	mempool_destroy(eq->pool);

	return err;
}

void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
{
	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
		return;

	switch (ent->order - 2) {
	case MLX5_IMR_MTT_CACHE_ENTRY:
		ent->page = PAGE_SHIFT;
		ent->xlt = MLX5_IMR_MTT_ENTRIES *
			   sizeof(struct mlx5_mtt) /
			   MLX5_IB_UMR_OCTOWORD;
		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
		ent->limit = 0;
		break;

	case MLX5_IMR_KSM_CACHE_ENTRY:
		ent->page = MLX5_KSM_PAGE_SHIFT;
		ent->xlt = mlx5_imr_ksm_entries *
			   sizeof(struct mlx5_klm) /
			   MLX5_IB_UMR_OCTOWORD;
		ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
		ent->limit = 0;
		break;
	}
}

static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
	.advise_mr = mlx5_ib_advise_mr,
};

int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
{
	int ret = 0;

	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
		return ret;

	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_odp_ops);

	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
		if (ret) {
			mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
			return ret;
		}
	}

	ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);

	return ret;
}

void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
{
	if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT))
		return;

	mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
}

int mlx5_ib_odp_init(void)
{
	mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
				       MLX5_IMR_MTT_BITS);

	return 0;
}

struct prefetch_mr_work {
	struct work_struct work;
	u32 pf_flags;
	u32 num_sge;
	struct {
		u64 io_virt;
		struct mlx5_ib_mr *mr;
		size_t length;
	} frags[];
};

static void destroy_prefetch_work(struct prefetch_mr_work *work)
{
	u32 i;

	for (i = 0; i < work->num_sge; ++i)
		if (atomic_dec_and_test(&work->frags[i].mr->num_deferred_work))
			wake_up(&work->frags[i].mr->q_deferred_work);
	kvfree(work);
}

static struct mlx5_ib_mr *
get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
		    u32 lkey)
{
	struct mlx5_ib_dev *dev = to_mdev(pd->device);
	struct mlx5_core_mkey *mmkey;
	struct ib_umem_odp *odp;
	struct mlx5_ib_mr *mr;

	lockdep_assert_held(&dev->odp_srcu);

	mmkey = xa_load(&dev->odp_mkeys, mlx5_base_mkey(lkey));
	if (!mmkey || mmkey->key != lkey || mmkey->type != MLX5_MKEY_MR)
		return NULL;

	mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);

	if (mr->ibmr.pd != pd)
		return NULL;

	odp = to_ib_umem_odp(mr->umem);

	/* prefetch with write-access must be supported by the MR */
	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
	    !odp->umem.writable)
		return NULL;

	return mr;
}

static void mlx5_ib_prefetch_mr_work(struct work_struct *w)
{
	struct prefetch_mr_work *work =
		container_of(w, struct prefetch_mr_work, work);
	u32 bytes_mapped = 0;
	u32 i;

	for (i = 0; i < work->num_sge; ++i)
		pagefault_mr(work->frags[i].mr, work->frags[i].io_virt,
			     work->frags[i].length, &bytes_mapped,
			     work->pf_flags);

	destroy_prefetch_work(work);
}

static bool init_prefetch_work(struct ib_pd *pd,
			       enum ib_uverbs_advise_mr_advice advice,
			       u32 pf_flags, struct prefetch_mr_work *work,
			       struct ib_sge *sg_list, u32 num_sge)
{
	u32 i;

	INIT_WORK(&work->work, mlx5_ib_prefetch_mr_work);
	work->pf_flags = pf_flags;

	for (i = 0; i < num_sge; ++i) {
		work->frags[i].io_virt = sg_list[i].addr;
		work->frags[i].length = sg_list[i].length;
		work->frags[i].mr =
			get_prefetchable_mr(pd, advice, sg_list[i].lkey);
		if (!work->frags[i].mr) {
			work->num_sge = i;
			return false;
		}

		/* Keep the MR pointer will valid outside the SRCU */
		atomic_inc(&work->frags[i].mr->num_deferred_work);
	}
	work->num_sge = num_sge;
	return true;
}

static int mlx5_ib_prefetch_sg_list(struct ib_pd *pd,
				    enum ib_uverbs_advise_mr_advice advice,
				    u32 pf_flags, struct ib_sge *sg_list,
				    u32 num_sge)
{
	struct mlx5_ib_dev *dev = to_mdev(pd->device);
	u32 bytes_mapped = 0;
	int srcu_key;
	int ret = 0;
	u32 i;

	srcu_key = srcu_read_lock(&dev->odp_srcu);
	for (i = 0; i < num_sge; ++i) {
		struct mlx5_ib_mr *mr;

		mr = get_prefetchable_mr(pd, advice, sg_list[i].lkey);
		if (!mr) {
			ret = -ENOENT;
			goto out;
		}
		ret = pagefault_mr(mr, sg_list[i].addr, sg_list[i].length,
				   &bytes_mapped, pf_flags);
		if (ret < 0)
			goto out;
	}
	ret = 0;

out:
	srcu_read_unlock(&dev->odp_srcu, srcu_key);
	return ret;
}

int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
			       enum ib_uverbs_advise_mr_advice advice,
			       u32 flags, struct ib_sge *sg_list, u32 num_sge)
{
	struct mlx5_ib_dev *dev = to_mdev(pd->device);
	u32 pf_flags = 0;
	struct prefetch_mr_work *work;
	int srcu_key;

	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
		pf_flags |= MLX5_PF_FLAGS_DOWNGRADE;

	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
		return mlx5_ib_prefetch_sg_list(pd, advice, pf_flags, sg_list,
						num_sge);

	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
	if (!work)
		return -ENOMEM;

	srcu_key = srcu_read_lock(&dev->odp_srcu);
	if (!init_prefetch_work(pd, advice, pf_flags, work, sg_list, num_sge)) {
		srcu_read_unlock(&dev->odp_srcu, srcu_key);
		destroy_prefetch_work(work);
		return -EINVAL;
	}
	queue_work(system_unbound_wq, &work->work);
	srcu_read_unlock(&dev->odp_srcu, srcu_key);
	return 0;
}