summaryrefslogtreecommitdiffstats
path: root/src/intel/vulkan/anv_allocator.c
blob: f484b8343d83a85331ad32e2598eb21833fa297a (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
/*
 * Copyright © 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 <stdlib.h>
#include <unistd.h>
#include <limits.h>
#include <assert.h>
#include <sys/mman.h>

#include "anv_private.h"

#include "util/simple_mtx.h"
#include "util/anon_file.h"

#ifdef HAVE_VALGRIND
#define VG_NOACCESS_READ(__ptr) ({                       \
   VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
   __typeof(*(__ptr)) __val = *(__ptr);                  \
   VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
   __val;                                                \
})
#define VG_NOACCESS_WRITE(__ptr, __val) ({                  \
   VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr)));  \
   *(__ptr) = (__val);                                      \
   VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));   \
})
#else
#define VG_NOACCESS_READ(__ptr) (*(__ptr))
#define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
#endif

#ifndef MAP_POPULATE
#define MAP_POPULATE 0
#endif

/* Design goals:
 *
 *  - Lock free (except when resizing underlying bos)
 *
 *  - Constant time allocation with typically only one atomic
 *
 *  - Multiple allocation sizes without fragmentation
 *
 *  - Can grow while keeping addresses and offset of contents stable
 *
 *  - All allocations within one bo so we can point one of the
 *    STATE_BASE_ADDRESS pointers at it.
 *
 * The overall design is a two-level allocator: top level is a fixed size, big
 * block (8k) allocator, which operates out of a bo.  Allocation is done by
 * either pulling a block from the free list or growing the used range of the
 * bo.  Growing the range may run out of space in the bo which we then need to
 * grow.  Growing the bo is tricky in a multi-threaded, lockless environment:
 * we need to keep all pointers and contents in the old map valid.  GEM bos in
 * general can't grow, but we use a trick: we create a memfd and use ftruncate
 * to grow it as necessary.  We mmap the new size and then create a gem bo for
 * it using the new gem userptr ioctl.  Without heavy-handed locking around
 * our allocation fast-path, there isn't really a way to munmap the old mmap,
 * so we just keep it around until garbage collection time.  While the block
 * allocator is lockless for normal operations, we block other threads trying
 * to allocate while we're growing the map.  It sholdn't happen often, and
 * growing is fast anyway.
 *
 * At the next level we can use various sub-allocators.  The state pool is a
 * pool of smaller, fixed size objects, which operates much like the block
 * pool.  It uses a free list for freeing objects, but when it runs out of
 * space it just allocates a new block from the block pool.  This allocator is
 * intended for longer lived state objects such as SURFACE_STATE and most
 * other persistent state objects in the API.  We may need to track more info
 * with these object and a pointer back to the CPU object (eg VkImage).  In
 * those cases we just allocate a slightly bigger object and put the extra
 * state after the GPU state object.
 *
 * The state stream allocator works similar to how the i965 DRI driver streams
 * all its state.  Even with Vulkan, we need to emit transient state (whether
 * surface state base or dynamic state base), and for that we can just get a
 * block and fill it up.  These cases are local to a command buffer and the
 * sub-allocator need not be thread safe.  The streaming allocator gets a new
 * block when it runs out of space and chains them together so they can be
 * easily freed.
 */

/* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
 * We use it to indicate the free list is empty. */
#define EMPTY UINT32_MAX

#define PAGE_SIZE 4096

struct anv_mmap_cleanup {
   void *map;
   size_t size;
};

static inline uint32_t
ilog2_round_up(uint32_t value)
{
   assert(value != 0);
   return 32 - __builtin_clz(value - 1);
}

static inline uint32_t
round_to_power_of_two(uint32_t value)
{
   return 1 << ilog2_round_up(value);
}

struct anv_state_table_cleanup {
   void *map;
   size_t size;
};

#define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
#define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))

static VkResult
anv_state_table_expand_range(struct anv_state_table *table, uint32_t size);

VkResult
anv_state_table_init(struct anv_state_table *table,
                    struct anv_device *device,
                    uint32_t initial_entries)
{
   VkResult result;

   table->device = device;

   /* Just make it 2GB up-front.  The Linux kernel won't actually back it
    * with pages until we either map and fault on one of them or we use
    * userptr and send a chunk of it off to the GPU.
    */
   table->fd = os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE, "state table");
   if (table->fd == -1) {
      result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
      goto fail_fd;
   }

   if (!u_vector_init(&table->cleanups,
                      round_to_power_of_two(sizeof(struct anv_state_table_cleanup)),
                      128)) {
      result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
      goto fail_fd;
   }

   table->state.next = 0;
   table->state.end = 0;
   table->size = 0;

   uint32_t initial_size = initial_entries * ANV_STATE_ENTRY_SIZE;
   result = anv_state_table_expand_range(table, initial_size);
   if (result != VK_SUCCESS)
      goto fail_cleanups;

   return VK_SUCCESS;

 fail_cleanups:
   u_vector_finish(&table->cleanups);
 fail_fd:
   close(table->fd);

   return result;
}

static VkResult
anv_state_table_expand_range(struct anv_state_table *table, uint32_t size)
{
   void *map;
   struct anv_state_table_cleanup *cleanup;

   /* Assert that we only ever grow the pool */
   assert(size >= table->state.end);

   /* Make sure that we don't go outside the bounds of the memfd */
   if (size > BLOCK_POOL_MEMFD_SIZE)
      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);

   cleanup = u_vector_add(&table->cleanups);
   if (!cleanup)
      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);

   *cleanup = ANV_STATE_TABLE_CLEANUP_INIT;

   /* Just leak the old map until we destroy the pool.  We can't munmap it
    * without races or imposing locking on the block allocate fast path. On
    * the whole the leaked maps adds up to less than the size of the
    * current map.  MAP_POPULATE seems like the right thing to do, but we
    * should try to get some numbers.
    */
   map = mmap(NULL, size, PROT_READ | PROT_WRITE,
              MAP_SHARED | MAP_POPULATE, table->fd, 0);
   if (map == MAP_FAILED) {
      return vk_errorf(table->device->instance, table->device,
                       VK_ERROR_OUT_OF_HOST_MEMORY, "mmap failed: %m");
   }

   cleanup->map = map;
   cleanup->size = size;

   table->map = map;
   table->size = size;

   return VK_SUCCESS;
}

static VkResult
anv_state_table_grow(struct anv_state_table *table)
{
   VkResult result = VK_SUCCESS;

   uint32_t used = align_u32(table->state.next * ANV_STATE_ENTRY_SIZE,
                             PAGE_SIZE);
   uint32_t old_size = table->size;

   /* The block pool is always initialized to a nonzero size and this function
    * is always called after initialization.
    */
   assert(old_size > 0);

   uint32_t required = MAX2(used, old_size);
   if (used * 2 <= required) {
      /* If we're in this case then this isn't the firsta allocation and we
       * already have enough space on both sides to hold double what we
       * have allocated.  There's nothing for us to do.
       */
      goto done;
   }

   uint32_t size = old_size * 2;
   while (size < required)
      size *= 2;

   assert(size > table->size);

   result = anv_state_table_expand_range(table, size);

 done:
   return result;
}

void
anv_state_table_finish(struct anv_state_table *table)
{
   struct anv_state_table_cleanup *cleanup;

   u_vector_foreach(cleanup, &table->cleanups) {
      if (cleanup->map)
         munmap(cleanup->map, cleanup->size);
   }

   u_vector_finish(&table->cleanups);

   close(table->fd);
}

VkResult
anv_state_table_add(struct anv_state_table *table, uint32_t *idx,
                    uint32_t count)
{
   struct anv_block_state state, old, new;
   VkResult result;

   assert(idx);

   while(1) {
      state.u64 = __sync_fetch_and_add(&table->state.u64, count);
      if (state.next + count <= state.end) {
         assert(table->map);
         struct anv_free_entry *entry = &table->map[state.next];
         for (int i = 0; i < count; i++) {
            entry[i].state.idx = state.next + i;
         }
         *idx = state.next;
         return VK_SUCCESS;
      } else if (state.next <= state.end) {
         /* We allocated the first block outside the pool so we have to grow
          * the pool.  pool_state->next acts a mutex: threads who try to
          * allocate now will get block indexes above the current limit and
          * hit futex_wait below.
          */
         new.next = state.next + count;
         do {
            result = anv_state_table_grow(table);
            if (result != VK_SUCCESS)
               return result;
            new.end = table->size / ANV_STATE_ENTRY_SIZE;
         } while (new.end < new.next);

         old.u64 = __sync_lock_test_and_set(&table->state.u64, new.u64);
         if (old.next != state.next)
            futex_wake(&table->state.end, INT_MAX);
      } else {
         futex_wait(&table->state.end, state.end, NULL);
         continue;
      }
   }
}

void
anv_free_list_push(union anv_free_list *list,
                   struct anv_state_table *table,
                   uint32_t first, uint32_t count)
{
   union anv_free_list current, old, new;
   uint32_t last = first;

   for (uint32_t i = 1; i < count; i++, last++)
      table->map[last].next = last + 1;

   old = *list;
   do {
      current = old;
      table->map[last].next = current.offset;
      new.offset = first;
      new.count = current.count + 1;
      old.u64 = __sync_val_compare_and_swap(&list->u64, current.u64, new.u64);
   } while (old.u64 != current.u64);
}

struct anv_state *
anv_free_list_pop(union anv_free_list *list,
                  struct anv_state_table *table)
{
   union anv_free_list current, new, old;

   current.u64 = list->u64;
   while (current.offset != EMPTY) {
      __sync_synchronize();
      new.offset = table->map[current.offset].next;
      new.count = current.count + 1;
      old.u64 = __sync_val_compare_and_swap(&list->u64, current.u64, new.u64);
      if (old.u64 == current.u64) {
         struct anv_free_entry *entry = &table->map[current.offset];
         return &entry->state;
      }
      current = old;
   }

   return NULL;
}

static VkResult
anv_block_pool_expand_range(struct anv_block_pool *pool,
                            uint32_t center_bo_offset, uint32_t size);

VkResult
anv_block_pool_init(struct anv_block_pool *pool,
                    struct anv_device *device,
                    uint64_t start_address,
                    uint32_t initial_size)
{
   VkResult result;

   pool->device = device;
   pool->use_softpin = device->instance->physicalDevice.use_softpin;
   pool->nbos = 0;
   pool->size = 0;
   pool->center_bo_offset = 0;
   pool->start_address = gen_canonical_address(start_address);
   pool->map = NULL;

   if (pool->use_softpin) {
      pool->bo = NULL;
      pool->fd = -1;
   } else {
      /* Just make it 2GB up-front.  The Linux kernel won't actually back it
       * with pages until we either map and fault on one of them or we use
       * userptr and send a chunk of it off to the GPU.
       */
      pool->fd = os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE, "block pool");
      if (pool->fd == -1)
         return vk_error(VK_ERROR_INITIALIZATION_FAILED);

      anv_bo_init(&pool->wrapper_bo, 0, 0);
      pool->wrapper_bo.is_wrapper = true;
      pool->bo = &pool->wrapper_bo;
   }

   if (!u_vector_init(&pool->mmap_cleanups,
                      round_to_power_of_two(sizeof(struct anv_mmap_cleanup)),
                      128)) {
      result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
      goto fail_fd;
   }

   pool->state.next = 0;
   pool->state.end = 0;
   pool->back_state.next = 0;
   pool->back_state.end = 0;

   result = anv_block_pool_expand_range(pool, 0, initial_size);
   if (result != VK_SUCCESS)
      goto fail_mmap_cleanups;

   /* Make the entire pool available in the front of the pool.  If back
    * allocation needs to use this space, the "ends" will be re-arranged.
    */
   pool->state.end = pool->size;

   return VK_SUCCESS;

 fail_mmap_cleanups:
   u_vector_finish(&pool->mmap_cleanups);
 fail_fd:
   if (pool->fd >= 0)
      close(pool->fd);

   return result;
}

void
anv_block_pool_finish(struct anv_block_pool *pool)
{
   anv_block_pool_foreach_bo(bo, pool) {
      if (bo->map)
         anv_gem_munmap(bo->map, bo->size);
      anv_gem_close(pool->device, bo->gem_handle);
   }

   struct anv_mmap_cleanup *cleanup;
   u_vector_foreach(cleanup, &pool->mmap_cleanups)
      munmap(cleanup->map, cleanup->size);
   u_vector_finish(&pool->mmap_cleanups);

   if (pool->fd >= 0)
      close(pool->fd);
}

static VkResult
anv_block_pool_expand_range(struct anv_block_pool *pool,
                            uint32_t center_bo_offset, uint32_t size)
{
   /* Assert that we only ever grow the pool */
   assert(center_bo_offset >= pool->back_state.end);
   assert(size - center_bo_offset >= pool->state.end);

   /* Assert that we don't go outside the bounds of the memfd */
   assert(center_bo_offset <= BLOCK_POOL_MEMFD_CENTER);
   assert(pool->use_softpin ||
          size - center_bo_offset <=
          BLOCK_POOL_MEMFD_SIZE - BLOCK_POOL_MEMFD_CENTER);

   /* For state pool BOs we have to be a bit careful about where we place them
    * in the GTT.  There are two documented workarounds for state base address
    * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
    * which state that those two base addresses do not support 48-bit
    * addresses and need to be placed in the bottom 32-bit range.
    * Unfortunately, this is not quite accurate.
    *
    * The real problem is that we always set the size of our state pools in
    * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
    * likely significantly smaller.  We do this because we do not no at the
    * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
    * the pool during command buffer building so we don't actually have a
    * valid final size.  If the address + size, as seen by STATE_BASE_ADDRESS
    * overflows 48 bits, the GPU appears to treat all accesses to the buffer
    * as being out of bounds and returns zero.  For dynamic state, this
    * usually just leads to rendering corruptions, but shaders that are all
    * zero hang the GPU immediately.
    *
    * The easiest solution to do is exactly what the bogus workarounds say to
    * do: restrict these buffers to 32-bit addresses.  We could also pin the
    * BO to some particular location of our choosing, but that's significantly
    * more work than just not setting a flag.  So, we explicitly DO NOT set
    * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
    * hard work for us.  When using softpin, we're in control and the fixed
    * addresses we choose are fine for base addresses.
    */
   enum anv_bo_alloc_flags bo_alloc_flags = 0;
   if (!pool->use_softpin)
      bo_alloc_flags |= ANV_BO_ALLOC_32BIT_ADDRESS;

   uint64_t bo_flags = 0;
   if (pool->device->instance->physicalDevice.has_exec_capture)
      bo_flags |= EXEC_OBJECT_CAPTURE;

   if (pool->use_softpin) {
      uint32_t new_bo_size = size - pool->size;
      struct anv_bo *new_bo;
      VkResult result = anv_device_alloc_bo(pool->device, new_bo_size,
                                            bo_alloc_flags |
                                            ANV_BO_ALLOC_FIXED_ADDRESS |
                                            ANV_BO_ALLOC_MAPPED |
                                            ANV_BO_ALLOC_SNOOPED,
                                            &new_bo);
      if (result != VK_SUCCESS)
         return result;

      assert(center_bo_offset == 0);

      new_bo->offset = pool->start_address + pool->size;
      pool->bos[pool->nbos++] = new_bo;

      /* This pointer will always point to the first BO in the list */
      pool->bo = pool->bos[0];
   } else {
      /* Just leak the old map until we destroy the pool.  We can't munmap it
       * without races or imposing locking on the block allocate fast path. On
       * the whole the leaked maps adds up to less than the size of the
       * current map.  MAP_POPULATE seems like the right thing to do, but we
       * should try to get some numbers.
       */
      void *map = mmap(NULL, size, PROT_READ | PROT_WRITE,
                       MAP_SHARED | MAP_POPULATE, pool->fd,
                       BLOCK_POOL_MEMFD_CENTER - center_bo_offset);
      if (map == MAP_FAILED)
         return vk_errorf(pool->device->instance, pool->device,
                          VK_ERROR_MEMORY_MAP_FAILED, "mmap failed: %m");

      struct anv_bo *new_bo;
      VkResult result = anv_device_import_bo_from_host_ptr(pool->device,
                                                           map, size,
                                                           bo_alloc_flags,
                                                           &new_bo);
      if (result != VK_SUCCESS) {
         munmap(map, size);
         return result;
      }

      struct anv_mmap_cleanup *cleanup = u_vector_add(&pool->mmap_cleanups);
      if (!cleanup) {
         munmap(map, size);
         anv_device_release_bo(pool->device, new_bo);
         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
      }
      cleanup->map = map;
      cleanup->size = size;

      /* Now that we mapped the new memory, we can write the new
       * center_bo_offset back into pool and update pool->map. */
      pool->center_bo_offset = center_bo_offset;
      pool->map = map + center_bo_offset;

      pool->bos[pool->nbos++] = new_bo;
      pool->wrapper_bo.map = new_bo;
   }

   assert(pool->nbos < ANV_MAX_BLOCK_POOL_BOS);
   pool->size = size;

   return VK_SUCCESS;
}

/** Returns current memory map of the block pool.
 *
 * The returned pointer points to the map for the memory at the specified
 * offset. The offset parameter is relative to the "center" of the block pool
 * rather than the start of the block pool BO map.
 */
void*
anv_block_pool_map(struct anv_block_pool *pool, int32_t offset)
{
   if (pool->use_softpin) {
      struct anv_bo *bo = NULL;
      int32_t bo_offset = 0;
      anv_block_pool_foreach_bo(iter_bo, pool) {
         if (offset < bo_offset + iter_bo->size) {
            bo = iter_bo;
            break;
         }
         bo_offset += iter_bo->size;
      }
      assert(bo != NULL);
      assert(offset >= bo_offset);

      return bo->map + (offset - bo_offset);
   } else {
      return pool->map + offset;
   }
}

/** Grows and re-centers the block pool.
 *
 * We grow the block pool in one or both directions in such a way that the
 * following conditions are met:
 *
 *  1) The size of the entire pool is always a power of two.
 *
 *  2) The pool only grows on both ends.  Neither end can get
 *     shortened.
 *
 *  3) At the end of the allocation, we have about twice as much space
 *     allocated for each end as we have used.  This way the pool doesn't
 *     grow too far in one direction or the other.
 *
 *  4) If the _alloc_back() has never been called, then the back portion of
 *     the pool retains a size of zero.  (This makes it easier for users of
 *     the block pool that only want a one-sided pool.)
 *
 *  5) We have enough space allocated for at least one more block in
 *     whichever side `state` points to.
 *
 *  6) The center of the pool is always aligned to both the block_size of
 *     the pool and a 4K CPU page.
 */
static uint32_t
anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state)
{
   VkResult result = VK_SUCCESS;

   pthread_mutex_lock(&pool->device->mutex);

   assert(state == &pool->state || state == &pool->back_state);

   /* Gather a little usage information on the pool.  Since we may have
    * threadsd waiting in queue to get some storage while we resize, it's
    * actually possible that total_used will be larger than old_size.  In
    * particular, block_pool_alloc() increments state->next prior to
    * calling block_pool_grow, so this ensures that we get enough space for
    * which ever side tries to grow the pool.
    *
    * We align to a page size because it makes it easier to do our
    * calculations later in such a way that we state page-aigned.
    */
   uint32_t back_used = align_u32(pool->back_state.next, PAGE_SIZE);
   uint32_t front_used = align_u32(pool->state.next, PAGE_SIZE);
   uint32_t total_used = front_used + back_used;

   assert(state == &pool->state || back_used > 0);

   uint32_t old_size = pool->size;

   /* The block pool is always initialized to a nonzero size and this function
    * is always called after initialization.
    */
   assert(old_size > 0);

   /* The back_used and front_used may actually be smaller than the actual
    * requirement because they are based on the next pointers which are
    * updated prior to calling this function.
    */
   uint32_t back_required = MAX2(back_used, pool->center_bo_offset);
   uint32_t front_required = MAX2(front_used, old_size - pool->center_bo_offset);

   if (back_used * 2 <= back_required && front_used * 2 <= front_required) {
      /* If we're in this case then this isn't the firsta allocation and we
       * already have enough space on both sides to hold double what we
       * have allocated.  There's nothing for us to do.
       */
      goto done;
   }

   uint32_t size = old_size * 2;
   while (size < back_required + front_required)
      size *= 2;

   assert(size > pool->size);

   /* We compute a new center_bo_offset such that, when we double the size
    * of the pool, we maintain the ratio of how much is used by each side.
    * This way things should remain more-or-less balanced.
    */
   uint32_t center_bo_offset;
   if (back_used == 0) {
      /* If we're in this case then we have never called alloc_back().  In
       * this case, we want keep the offset at 0 to make things as simple
       * as possible for users that don't care about back allocations.
       */
      center_bo_offset = 0;
   } else {
      /* Try to "center" the allocation based on how much is currently in
       * use on each side of the center line.
       */
      center_bo_offset = ((uint64_t)size * back_used) / total_used;

      /* Align down to a multiple of the page size */
      center_bo_offset &= ~(PAGE_SIZE - 1);

      assert(center_bo_offset >= back_used);

      /* Make sure we don't shrink the back end of the pool */
      if (center_bo_offset < back_required)
         center_bo_offset = back_required;

      /* Make sure that we don't shrink the front end of the pool */
      if (size - center_bo_offset < front_required)
         center_bo_offset = size - front_required;
   }

   assert(center_bo_offset % PAGE_SIZE == 0);

   result = anv_block_pool_expand_range(pool, center_bo_offset, size);

done:
   pthread_mutex_unlock(&pool->device->mutex);

   if (result == VK_SUCCESS) {
      /* Return the appropriate new size.  This function never actually
       * updates state->next.  Instead, we let the caller do that because it
       * needs to do so in order to maintain its concurrency model.
       */
      if (state == &pool->state) {
         return pool->size - pool->center_bo_offset;
      } else {
         assert(pool->center_bo_offset > 0);
         return pool->center_bo_offset;
      }
   } else {
      return 0;
   }
}

static uint32_t
anv_block_pool_alloc_new(struct anv_block_pool *pool,
                         struct anv_block_state *pool_state,
                         uint32_t block_size, uint32_t *padding)
{
   struct anv_block_state state, old, new;

   /* Most allocations won't generate any padding */
   if (padding)
      *padding = 0;

   while (1) {
      state.u64 = __sync_fetch_and_add(&pool_state->u64, block_size);
      if (state.next + block_size <= state.end) {
         return state.next;
      } else if (state.next <= state.end) {
         if (pool->use_softpin && state.next < state.end) {
            /* We need to grow the block pool, but still have some leftover
             * space that can't be used by that particular allocation. So we
             * add that as a "padding", and return it.
             */
            uint32_t leftover = state.end - state.next;

            /* If there is some leftover space in the pool, the caller must
             * deal with it.
             */
            assert(leftover == 0 || padding);
            if (padding)
               *padding = leftover;
            state.next += leftover;
         }

         /* We allocated the first block outside the pool so we have to grow
          * the pool.  pool_state->next acts a mutex: threads who try to
          * allocate now will get block indexes above the current limit and
          * hit futex_wait below.
          */
         new.next = state.next + block_size;
         do {
            new.end = anv_block_pool_grow(pool, pool_state);
         } while (new.end < new.next);

         old.u64 = __sync_lock_test_and_set(&pool_state->u64, new.u64);
         if (old.next != state.next)
            futex_wake(&pool_state->end, INT_MAX);
         return state.next;
      } else {
         futex_wait(&pool_state->end, state.end, NULL);
         continue;
      }
   }
}

int32_t
anv_block_pool_alloc(struct anv_block_pool *pool,
                     uint32_t block_size, uint32_t *padding)
{
   uint32_t offset;

   offset = anv_block_pool_alloc_new(pool, &pool->state, block_size, padding);

   return offset;
}

/* Allocates a block out of the back of the block pool.
 *
 * This will allocated a block earlier than the "start" of the block pool.
 * The offsets returned from this function will be negative but will still
 * be correct relative to the block pool's map pointer.
 *
 * If you ever use anv_block_pool_alloc_back, then you will have to do
 * gymnastics with the block pool's BO when doing relocations.
 */
int32_t
anv_block_pool_alloc_back(struct anv_block_pool *pool,
                          uint32_t block_size)
{
   int32_t offset = anv_block_pool_alloc_new(pool, &pool->back_state,
                                             block_size, NULL);

   /* The offset we get out of anv_block_pool_alloc_new() is actually the
    * number of bytes downwards from the middle to the end of the block.
    * We need to turn it into a (negative) offset from the middle to the
    * start of the block.
    */
   assert(offset >= 0);
   return -(offset + block_size);
}

VkResult
anv_state_pool_init(struct anv_state_pool *pool,
                    struct anv_device *device,
                    uint64_t start_address,
                    uint32_t block_size)
{
   VkResult result = anv_block_pool_init(&pool->block_pool, device,
                                         start_address,
                                         block_size * 16);
   if (result != VK_SUCCESS)
      return result;

   result = anv_state_table_init(&pool->table, device, 64);
   if (result != VK_SUCCESS) {
      anv_block_pool_finish(&pool->block_pool);
      return result;
   }

   assert(util_is_power_of_two_or_zero(block_size));
   pool->block_size = block_size;
   pool->back_alloc_free_list = ANV_FREE_LIST_EMPTY;
   for (unsigned i = 0; i < ANV_STATE_BUCKETS; i++) {
      pool->buckets[i].free_list = ANV_FREE_LIST_EMPTY;
      pool->buckets[i].block.next = 0;
      pool->buckets[i].block.end = 0;
   }
   VG(VALGRIND_CREATE_MEMPOOL(pool, 0, false));

   return VK_SUCCESS;
}

void
anv_state_pool_finish(struct anv_state_pool *pool)
{
   VG(VALGRIND_DESTROY_MEMPOOL(pool));
   anv_state_table_finish(&pool->table);
   anv_block_pool_finish(&pool->block_pool);
}

static uint32_t
anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool *pool,
                                    struct anv_block_pool *block_pool,
                                    uint32_t state_size,
                                    uint32_t block_size,
                                    uint32_t *padding)
{
   struct anv_block_state block, old, new;
   uint32_t offset;

   /* We don't always use anv_block_pool_alloc(), which would set *padding to
    * zero for us. So if we have a pointer to padding, we must zero it out
    * ourselves here, to make sure we always return some sensible value.
    */
   if (padding)
      *padding = 0;

   /* If our state is large, we don't need any sub-allocation from a block.
    * Instead, we just grab whole (potentially large) blocks.
    */
   if (state_size >= block_size)
      return anv_block_pool_alloc(block_pool, state_size, padding);

 restart:
   block.u64 = __sync_fetch_and_add(&pool->block.u64, state_size);

   if (block.next < block.end) {
      return block.next;
   } else if (block.next == block.end) {
      offset = anv_block_pool_alloc(block_pool, block_size, padding);
      new.next = offset + state_size;
      new.end = offset + block_size;
      old.u64 = __sync_lock_test_and_set(&pool->block.u64, new.u64);
      if (old.next != block.next)
         futex_wake(&pool->block.end, INT_MAX);
      return offset;
   } else {
      futex_wait(&pool->block.end, block.end, NULL);
      goto restart;
   }
}

static uint32_t
anv_state_pool_get_bucket(uint32_t size)
{
   unsigned size_log2 = ilog2_round_up(size);
   assert(size_log2 <= ANV_MAX_STATE_SIZE_LOG2);
   if (size_log2 < ANV_MIN_STATE_SIZE_LOG2)
      size_log2 = ANV_MIN_STATE_SIZE_LOG2;
   return size_log2 - ANV_MIN_STATE_SIZE_LOG2;
}

static uint32_t
anv_state_pool_get_bucket_size(uint32_t bucket)
{
   uint32_t size_log2 = bucket + ANV_MIN_STATE_SIZE_LOG2;
   return 1 << size_log2;
}

/** Helper to push a chunk into the state table.
 *
 * It creates 'count' entries into the state table and update their sizes,
 * offsets and maps, also pushing them as "free" states.
 */
static void
anv_state_pool_return_blocks(struct anv_state_pool *pool,
                             uint32_t chunk_offset, uint32_t count,
                             uint32_t block_size)
{
   /* Disallow returning 0 chunks */
   assert(count != 0);

   /* Make sure we always return chunks aligned to the block_size */
   assert(chunk_offset % block_size == 0);

   uint32_t st_idx;
   UNUSED VkResult result = anv_state_table_add(&pool->table, &st_idx, count);
   assert(result == VK_SUCCESS);
   for (int i = 0; i < count; i++) {
      /* update states that were added back to the state table */
      struct anv_state *state_i = anv_state_table_get(&pool->table,
                                                      st_idx + i);
      state_i->alloc_size = block_size;
      state_i->offset = chunk_offset + block_size * i;
      state_i->map = anv_block_pool_map(&pool->block_pool, state_i->offset);
   }

   uint32_t block_bucket = anv_state_pool_get_bucket(block_size);
   anv_free_list_push(&pool->buckets[block_bucket].free_list,
                      &pool->table, st_idx, count);
}

/** Returns a chunk of memory back to the state pool.
 *
 * Do a two-level split. If chunk_size is bigger than divisor
 * (pool->block_size), we return as many divisor sized blocks as we can, from
 * the end of the chunk.
 *
 * The remaining is then split into smaller blocks (starting at small_size if
 * it is non-zero), with larger blocks always being taken from the end of the
 * chunk.
 */
static void
anv_state_pool_return_chunk(struct anv_state_pool *pool,
                            uint32_t chunk_offset, uint32_t chunk_size,
                            uint32_t small_size)
{
   uint32_t divisor = pool->block_size;
   uint32_t nblocks = chunk_size / divisor;
   uint32_t rest = chunk_size - nblocks * divisor;

   if (nblocks > 0) {
      /* First return divisor aligned and sized chunks. We start returning
       * larger blocks from the end fo the chunk, since they should already be
       * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
       * aligned chunks.
       */
      uint32_t offset = chunk_offset + rest;
      anv_state_pool_return_blocks(pool, offset, nblocks, divisor);
   }

   chunk_size = rest;
   divisor /= 2;

   if (small_size > 0 && small_size < divisor)
      divisor = small_size;

   uint32_t min_size = 1 << ANV_MIN_STATE_SIZE_LOG2;

   /* Just as before, return larger divisor aligned blocks from the end of the
    * chunk first.
    */
   while (chunk_size > 0 && divisor >= min_size) {
      nblocks = chunk_size / divisor;
      rest = chunk_size - nblocks * divisor;
      if (nblocks > 0) {
         anv_state_pool_return_blocks(pool, chunk_offset + rest,
                                      nblocks, divisor);
         chunk_size = rest;
      }
      divisor /= 2;
   }
}

static struct anv_state
anv_state_pool_alloc_no_vg(struct anv_state_pool *pool,
                           uint32_t size, uint32_t align)
{
   uint32_t bucket = anv_state_pool_get_bucket(MAX2(size, align));

   struct anv_state *state;
   uint32_t alloc_size = anv_state_pool_get_bucket_size(bucket);
   int32_t offset;

   /* Try free list first. */
   state = anv_free_list_pop(&pool->buckets[bucket].free_list,
                             &pool->table);
   if (state) {
      assert(state->offset >= 0);
      goto done;
   }

   /* Try to grab a chunk from some larger bucket and split it up */
   for (unsigned b = bucket + 1; b < ANV_STATE_BUCKETS; b++) {
      state = anv_free_list_pop(&pool->buckets[b].free_list, &pool->table);
      if (state) {
         unsigned chunk_size = anv_state_pool_get_bucket_size(b);
         int32_t chunk_offset = state->offset;

         /* First lets update the state we got to its new size. offset and map
          * remain the same.
          */
         state->alloc_size = alloc_size;

         /* Now return the unused part of the chunk back to the pool as free
          * blocks
          *
          * There are a couple of options as to what we do with it:
          *
          *    1) We could fully split the chunk into state.alloc_size sized
          *       pieces.  However, this would mean that allocating a 16B
          *       state could potentially split a 2MB chunk into 512K smaller
          *       chunks.  This would lead to unnecessary fragmentation.
          *
          *    2) The classic "buddy allocator" method would have us split the
          *       chunk in half and return one half.  Then we would split the
          *       remaining half in half and return one half, and repeat as
          *       needed until we get down to the size we want.  However, if
          *       you are allocating a bunch of the same size state (which is
          *       the common case), this means that every other allocation has
          *       to go up a level and every fourth goes up two levels, etc.
          *       This is not nearly as efficient as it could be if we did a
          *       little more work up-front.
          *
          *    3) Split the difference between (1) and (2) by doing a
          *       two-level split.  If it's bigger than some fixed block_size,
          *       we split it into block_size sized chunks and return all but
          *       one of them.  Then we split what remains into
          *       state.alloc_size sized chunks and return them.
          *
          * We choose something close to option (3), which is implemented with
          * anv_state_pool_return_chunk(). That is done by returning the
          * remaining of the chunk, with alloc_size as a hint of the size that
          * we want the smaller chunk split into.
          */
         anv_state_pool_return_chunk(pool, chunk_offset + alloc_size,
                                     chunk_size - alloc_size, alloc_size);
         goto done;
      }
   }

   uint32_t padding;
   offset = anv_fixed_size_state_pool_alloc_new(&pool->buckets[bucket],
                                                &pool->block_pool,
                                                alloc_size,
                                                pool->block_size,
                                                &padding);
   /* Everytime we allocate a new state, add it to the state pool */
   uint32_t idx;
   UNUSED VkResult result = anv_state_table_add(&pool->table, &idx, 1);
   assert(result == VK_SUCCESS);

   state = anv_state_table_get(&pool->table, idx);
   state->offset = offset;
   state->alloc_size = alloc_size;
   state->map = anv_block_pool_map(&pool->block_pool, offset);

   if (padding > 0) {
      uint32_t return_offset = offset - padding;
      anv_state_pool_return_chunk(pool, return_offset, padding, 0);
   }

done:
   return *state;
}

struct anv_state
anv_state_pool_alloc(struct anv_state_pool *pool, uint32_t size, uint32_t align)
{
   if (size == 0)
      return ANV_STATE_NULL;

   struct anv_state state = anv_state_pool_alloc_no_vg(pool, size, align);
   VG(VALGRIND_MEMPOOL_ALLOC(pool, state.map, size));
   return state;
}

struct anv_state
anv_state_pool_alloc_back(struct anv_state_pool *pool)
{
   struct anv_state *state;
   uint32_t alloc_size = pool->block_size;

   state = anv_free_list_pop(&pool->back_alloc_free_list, &pool->table);
   if (state) {
      assert(state->offset < 0);
      goto done;
   }

   int32_t offset;
   offset = anv_block_pool_alloc_back(&pool->block_pool,
                                      pool->block_size);
   uint32_t idx;
   UNUSED VkResult result = anv_state_table_add(&pool->table, &idx, 1);
   assert(result == VK_SUCCESS);

   state = anv_state_table_get(&pool->table, idx);
   state->offset = offset;
   state->alloc_size = alloc_size;
   state->map = anv_block_pool_map(&pool->block_pool, state->offset);

done:
   VG(VALGRIND_MEMPOOL_ALLOC(pool, state->map, state->alloc_size));
   return *state;
}

static void
anv_state_pool_free_no_vg(struct anv_state_pool *pool, struct anv_state state)
{
   assert(util_is_power_of_two_or_zero(state.alloc_size));
   unsigned bucket = anv_state_pool_get_bucket(state.alloc_size);

   if (state.offset < 0) {
      assert(state.alloc_size == pool->block_size);
      anv_free_list_push(&pool->back_alloc_free_list,
                         &pool->table, state.idx, 1);
   } else {
      anv_free_list_push(&pool->buckets[bucket].free_list,
                         &pool->table, state.idx, 1);
   }
}

void
anv_state_pool_free(struct anv_state_pool *pool, struct anv_state state)
{
   if (state.alloc_size == 0)
      return;

   VG(VALGRIND_MEMPOOL_FREE(pool, state.map));
   anv_state_pool_free_no_vg(pool, state);
}

struct anv_state_stream_block {
   struct anv_state block;

   /* The next block */
   struct anv_state_stream_block *next;

#ifdef HAVE_VALGRIND
   /* A pointer to the first user-allocated thing in this block.  This is
    * what valgrind sees as the start of the block.
    */
   void *_vg_ptr;
#endif
};

/* The state stream allocator is a one-shot, single threaded allocator for
 * variable sized blocks.  We use it for allocating dynamic state.
 */
void
anv_state_stream_init(struct anv_state_stream *stream,
                      struct anv_state_pool *state_pool,
                      uint32_t block_size)
{
   stream->state_pool = state_pool;
   stream->block_size = block_size;

   stream->block = ANV_STATE_NULL;

   stream->block_list = NULL;

   /* Ensure that next + whatever > block_size.  This way the first call to
    * state_stream_alloc fetches a new block.
    */
   stream->next = block_size;

   VG(VALGRIND_CREATE_MEMPOOL(stream, 0, false));
}

void
anv_state_stream_finish(struct anv_state_stream *stream)
{
   struct anv_state_stream_block *next = stream->block_list;
   while (next != NULL) {
      struct anv_state_stream_block sb = VG_NOACCESS_READ(next);
      VG(VALGRIND_MEMPOOL_FREE(stream, sb._vg_ptr));
      VG(VALGRIND_MAKE_MEM_UNDEFINED(next, stream->block_size));
      anv_state_pool_free_no_vg(stream->state_pool, sb.block);
      next = sb.next;
   }

   VG(VALGRIND_DESTROY_MEMPOOL(stream));
}

struct anv_state
anv_state_stream_alloc(struct anv_state_stream *stream,
                       uint32_t size, uint32_t alignment)
{
   if (size == 0)
      return ANV_STATE_NULL;

   assert(alignment <= PAGE_SIZE);

   uint32_t offset = align_u32(stream->next, alignment);
   if (offset + size > stream->block.alloc_size) {
      uint32_t block_size = stream->block_size;
      if (block_size < size)
         block_size = round_to_power_of_two(size);

      stream->block = anv_state_pool_alloc_no_vg(stream->state_pool,
                                                 block_size, PAGE_SIZE);

      struct anv_state_stream_block *sb = stream->block.map;
      VG_NOACCESS_WRITE(&sb->block, stream->block);
      VG_NOACCESS_WRITE(&sb->next, stream->block_list);
      stream->block_list = sb;
      VG(VG_NOACCESS_WRITE(&sb->_vg_ptr, NULL));

      VG(VALGRIND_MAKE_MEM_NOACCESS(stream->block.map, stream->block_size));

      /* Reset back to the start plus space for the header */
      stream->next = sizeof(*sb);

      offset = align_u32(stream->next, alignment);
      assert(offset + size <= stream->block.alloc_size);
   }

   struct anv_state state = stream->block;
   state.offset += offset;
   state.alloc_size = size;
   state.map += offset;

   stream->next = offset + size;

#ifdef HAVE_VALGRIND
   struct anv_state_stream_block *sb = stream->block_list;
   void *vg_ptr = VG_NOACCESS_READ(&sb->_vg_ptr);
   if (vg_ptr == NULL) {
      vg_ptr = state.map;
      VG_NOACCESS_WRITE(&sb->_vg_ptr, vg_ptr);
      VALGRIND_MEMPOOL_ALLOC(stream, vg_ptr, size);
   } else {
      void *state_end = state.map + state.alloc_size;
      /* This only updates the mempool.  The newly allocated chunk is still
       * marked as NOACCESS. */
      VALGRIND_MEMPOOL_CHANGE(stream, vg_ptr, vg_ptr, state_end - vg_ptr);
      /* Mark the newly allocated chunk as undefined */
      VALGRIND_MAKE_MEM_UNDEFINED(state.map, state.alloc_size);
   }
#endif

   return state;
}

void
anv_bo_pool_init(struct anv_bo_pool *pool, struct anv_device *device,
                 uint64_t bo_flags)
{
   pool->device = device;
   pool->bo_flags = bo_flags;
   for (unsigned i = 0; i < ARRAY_SIZE(pool->free_list); i++) {
      util_sparse_array_free_list_init(&pool->free_list[i],
                                       &device->bo_cache.bo_map, 0,
                                       offsetof(struct anv_bo, free_index));
   }

   VG(VALGRIND_CREATE_MEMPOOL(pool, 0, false));
}

void
anv_bo_pool_finish(struct anv_bo_pool *pool)
{
   for (unsigned i = 0; i < ARRAY_SIZE(pool->free_list); i++) {
      while (1) {
         struct anv_bo *bo =
            util_sparse_array_free_list_pop_elem(&pool->free_list[i]);
         if (bo == NULL)
            break;

         /* anv_device_release_bo is going to "free" it */
         VG(VALGRIND_MALLOCLIKE_BLOCK(bo->map, bo->size, 0, 1));
         anv_device_release_bo(pool->device, bo);
      }
   }

   VG(VALGRIND_DESTROY_MEMPOOL(pool));
}

VkResult
anv_bo_pool_alloc(struct anv_bo_pool *pool, uint32_t size,
                  struct anv_bo **bo_out)
{
   const unsigned size_log2 = size < 4096 ? 12 : ilog2_round_up(size);
   const unsigned pow2_size = 1 << size_log2;
   const unsigned bucket = size_log2 - 12;
   assert(bucket < ARRAY_SIZE(pool->free_list));

   struct anv_bo *bo =
      util_sparse_array_free_list_pop_elem(&pool->free_list[bucket]);
   if (bo != NULL) {
      VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));
      *bo_out = bo;
      return VK_SUCCESS;
   }

   VkResult result = anv_device_alloc_bo(pool->device,
                                         pow2_size,
                                         ANV_BO_ALLOC_MAPPED |
                                         ANV_BO_ALLOC_SNOOPED,
                                         &bo);
   if (result != VK_SUCCESS)
      return result;

   /* We want it to look like it came from this pool */
   VG(VALGRIND_FREELIKE_BLOCK(bo->map, 0));
   VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));

   *bo_out = bo;

   return VK_SUCCESS;
}

void
anv_bo_pool_free(struct anv_bo_pool *pool, struct anv_bo *bo)
{
   VG(VALGRIND_MEMPOOL_FREE(pool, bo->map));

   assert(util_is_power_of_two_or_zero(bo->size));
   const unsigned size_log2 = ilog2_round_up(bo->size);
   const unsigned bucket = size_log2 - 12;
   assert(bucket < ARRAY_SIZE(pool->free_list));

   assert(util_sparse_array_get(&pool->device->bo_cache.bo_map,
                                bo->gem_handle) == bo);
   util_sparse_array_free_list_push(&pool->free_list[bucket],
                                    &bo->gem_handle, 1);
}

// Scratch pool

void
anv_scratch_pool_init(struct anv_device *device, struct anv_scratch_pool *pool)
{
   memset(pool, 0, sizeof(*pool));
}

void
anv_scratch_pool_finish(struct anv_device *device, struct anv_scratch_pool *pool)
{
   for (unsigned s = 0; s < MESA_SHADER_STAGES; s++) {
      for (unsigned i = 0; i < 16; i++) {
         struct anv_scratch_bo *bo = &pool->bos[i][s];
         if (bo->exists > 0) {
            anv_vma_free(device, &bo->bo);
            anv_gem_close(device, bo->bo.gem_handle);
         }
      }
   }
}

struct anv_bo *
anv_scratch_pool_alloc(struct anv_device *device, struct anv_scratch_pool *pool,
                       gl_shader_stage stage, unsigned per_thread_scratch)
{
   if (per_thread_scratch == 0)
      return NULL;

   unsigned scratch_size_log2 = ffs(per_thread_scratch / 2048);
   assert(scratch_size_log2 < 16);

   struct anv_scratch_bo *bo = &pool->bos[scratch_size_log2][stage];

   /* We can use "exists" to shortcut and ignore the critical section */
   if (bo->exists)
      return &bo->bo;

   pthread_mutex_lock(&device->mutex);

   __sync_synchronize();
   if (bo->exists) {
      pthread_mutex_unlock(&device->mutex);
      return &bo->bo;
   }

   const struct anv_physical_device *physical_device =
      &device->instance->physicalDevice;
   const struct gen_device_info *devinfo = &physical_device->info;

   const unsigned subslices = MAX2(physical_device->subslice_total, 1);

   unsigned scratch_ids_per_subslice;
   if (devinfo->gen >= 11) {
      /* The MEDIA_VFE_STATE docs say:
       *
       *    "Starting with this configuration, the Maximum Number of
       *     Threads must be set to (#EU * 8) for GPGPU dispatches.
       *
       *     Although there are only 7 threads per EU in the configuration,
       *     the FFTID is calculated as if there are 8 threads per EU,
       *     which in turn requires a larger amount of Scratch Space to be
       *     allocated by the driver."
       */
      scratch_ids_per_subslice = 8 * 8;
   } else if (devinfo->is_haswell) {
      /* WaCSScratchSize:hsw
       *
       * Haswell's scratch space address calculation appears to be sparse
       * rather than tightly packed. The Thread ID has bits indicating
       * which subslice, EU within a subslice, and thread within an EU it
       * is. There's a maximum of two slices and two subslices, so these
       * can be stored with a single bit. Even though there are only 10 EUs
       * per subslice, this is stored in 4 bits, so there's an effective
       * maximum value of 16 EUs. Similarly, although there are only 7
       * threads per EU, this is stored in a 3 bit number, giving an
       * effective maximum value of 8 threads per EU.
       *
       * This means that we need to use 16 * 8 instead of 10 * 7 for the
       * number of threads per subslice.
       */
      scratch_ids_per_subslice = 16 * 8;
   } else if (devinfo->is_cherryview) {
      /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
       * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
       * it had 8 EUs.
       */
      scratch_ids_per_subslice = 8 * 7;
   } else {
      scratch_ids_per_subslice = devinfo->max_cs_threads;
   }

   uint32_t max_threads[] = {
      [MESA_SHADER_VERTEX]           = devinfo->max_vs_threads,
      [MESA_SHADER_TESS_CTRL]        = devinfo->max_tcs_threads,
      [MESA_SHADER_TESS_EVAL]        = devinfo->max_tes_threads,
      [MESA_SHADER_GEOMETRY]         = devinfo->max_gs_threads,
      [MESA_SHADER_FRAGMENT]         = devinfo->max_wm_threads,
      [MESA_SHADER_COMPUTE]          = scratch_ids_per_subslice * subslices,
   };

   uint32_t size = per_thread_scratch * max_threads[stage];

   anv_bo_init_new(&bo->bo, device, size);

   /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
    * are still relative to the general state base address.  When we emit
    * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
    * to the maximum (1 page under 4GB).  This allows us to just place the
    * scratch buffers anywhere we wish in the bottom 32 bits of address space
    * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
    * However, in order to do so, we need to ensure that the kernel does not
    * place the scratch BO above the 32-bit boundary.
    *
    * NOTE: Technically, it can't go "anywhere" because the top page is off
    * limits.  However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
    * kernel allocates space using
    *
    *    end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
    *
    * so nothing will ever touch the top page.
    */
   assert(!(bo->bo.flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS));

   if (device->instance->physicalDevice.has_exec_async)
      bo->bo.flags |= EXEC_OBJECT_ASYNC;

   if (device->instance->physicalDevice.use_softpin)
      bo->bo.flags |= EXEC_OBJECT_PINNED;

   anv_vma_alloc(device, &bo->bo);

   /* Set the exists last because it may be read by other threads */
   __sync_synchronize();
   bo->exists = true;

   pthread_mutex_unlock(&device->mutex);

   return &bo->bo;
}

VkResult
anv_bo_cache_init(struct anv_bo_cache *cache)
{
   util_sparse_array_init(&cache->bo_map, sizeof(struct anv_bo), 1024);

   if (pthread_mutex_init(&cache->mutex, NULL)) {
      util_sparse_array_finish(&cache->bo_map);
      return vk_errorf(NULL, NULL, VK_ERROR_OUT_OF_HOST_MEMORY,
                       "pthread_mutex_init failed: %m");
   }

   return VK_SUCCESS;
}

void
anv_bo_cache_finish(struct anv_bo_cache *cache)
{
   util_sparse_array_finish(&cache->bo_map);
   pthread_mutex_destroy(&cache->mutex);
}

#define ANV_BO_CACHE_SUPPORTED_FLAGS \
   (EXEC_OBJECT_WRITE | \
    EXEC_OBJECT_ASYNC | \
    EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
    EXEC_OBJECT_PINNED | \
    EXEC_OBJECT_CAPTURE)

static uint32_t
anv_bo_alloc_flags_to_bo_flags(struct anv_device *device,
                               enum anv_bo_alloc_flags alloc_flags)
{
   struct anv_physical_device *pdevice = &device->instance->physicalDevice;

   uint64_t bo_flags = 0;
   if (!(alloc_flags & ANV_BO_ALLOC_32BIT_ADDRESS) &&
       pdevice->supports_48bit_addresses)
      bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;

   if ((alloc_flags & ANV_BO_ALLOC_CAPTURE) && pdevice->has_exec_capture)
      bo_flags |= EXEC_OBJECT_CAPTURE;

   if (alloc_flags & ANV_BO_ALLOC_IMPLICIT_WRITE) {
      assert(alloc_flags & ANV_BO_ALLOC_IMPLICIT_SYNC);
      bo_flags |= EXEC_OBJECT_WRITE;
   }

   if (!(alloc_flags & ANV_BO_ALLOC_IMPLICIT_SYNC) && pdevice->has_exec_async)
      bo_flags |= EXEC_OBJECT_ASYNC;

   if (pdevice->use_softpin)
      bo_flags |= EXEC_OBJECT_PINNED;

   return bo_flags;
}

VkResult
anv_device_alloc_bo(struct anv_device *device,
                    uint64_t size,
                    enum anv_bo_alloc_flags alloc_flags,
                    struct anv_bo **bo_out)
{
   const uint32_t bo_flags =
      anv_bo_alloc_flags_to_bo_flags(device, alloc_flags);
   assert(bo_flags == (bo_flags & ANV_BO_CACHE_SUPPORTED_FLAGS));

   /* The kernel is going to give us whole pages anyway */
   size = align_u64(size, 4096);

   struct anv_bo new_bo;
   VkResult result = anv_bo_init_new(&new_bo, device, size);
   if (result != VK_SUCCESS)
      return result;

   new_bo.flags = bo_flags;
   new_bo.is_external = (alloc_flags & ANV_BO_ALLOC_EXTERNAL);

   if (alloc_flags & ANV_BO_ALLOC_MAPPED) {
      new_bo.map = anv_gem_mmap(device, new_bo.gem_handle, 0, size, 0);
      if (new_bo.map == MAP_FAILED) {
         anv_gem_close(device, new_bo.gem_handle);
         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
      }
   }

   if (alloc_flags & ANV_BO_ALLOC_SNOOPED) {
      assert(alloc_flags & ANV_BO_ALLOC_MAPPED);
      /* We don't want to change these defaults if it's going to be shared
       * with another process.
       */
      assert(!(alloc_flags & ANV_BO_ALLOC_EXTERNAL));

      /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
       * I915_CACHING_NONE on non-LLC platforms.  For many internal state
       * objects, we'd rather take the snooping overhead than risk forgetting
       * a CLFLUSH somewhere.  Userptr objects are always created as
       * I915_CACHING_CACHED, which on non-LLC means snooped so there's no
       * need to do this there.
       */
      if (!device->info.has_llc) {
         anv_gem_set_caching(device, new_bo.gem_handle,
                             I915_CACHING_CACHED);
      }
   }

   if (alloc_flags & ANV_BO_ALLOC_FIXED_ADDRESS) {
      new_bo.has_fixed_address = true;
   } else {
      if (!anv_vma_alloc(device, &new_bo)) {
         if (new_bo.map)
            anv_gem_munmap(new_bo.map, size);
         anv_gem_close(device, new_bo.gem_handle);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_OUT_OF_DEVICE_MEMORY,
                          "failed to allocate virtual address for BO");
      }
   }

   assert(new_bo.gem_handle);

   /* If we just got this gem_handle from anv_bo_init_new then we know no one
    * else is touching this BO at the moment so we don't need to lock here.
    */
   struct anv_bo *bo = anv_device_lookup_bo(device, new_bo.gem_handle);
   *bo = new_bo;

   *bo_out = bo;

   return VK_SUCCESS;
}

VkResult
anv_device_import_bo_from_host_ptr(struct anv_device *device,
                                   void *host_ptr, uint32_t size,
                                   enum anv_bo_alloc_flags alloc_flags,
                                   struct anv_bo **bo_out)
{
   assert(!(alloc_flags & (ANV_BO_ALLOC_MAPPED |
                           ANV_BO_ALLOC_SNOOPED |
                           ANV_BO_ALLOC_FIXED_ADDRESS)));

   struct anv_bo_cache *cache = &device->bo_cache;
   const uint32_t bo_flags =
      anv_bo_alloc_flags_to_bo_flags(device, alloc_flags);
   assert(bo_flags == (bo_flags & ANV_BO_CACHE_SUPPORTED_FLAGS));

   uint32_t gem_handle = anv_gem_userptr(device, host_ptr, size);
   if (!gem_handle)
      return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);

   pthread_mutex_lock(&cache->mutex);

   struct anv_bo *bo = anv_device_lookup_bo(device, gem_handle);
   if (bo->refcount > 0) {
      /* VK_EXT_external_memory_host doesn't require handling importing the
       * same pointer twice at the same time, but we don't get in the way.  If
       * kernel gives us the same gem_handle, only succeed if the flags match.
       */
      assert(bo->gem_handle == gem_handle);
      if (bo_flags != bo->flags) {
         pthread_mutex_unlock(&cache->mutex);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
                          "same host pointer imported two different ways");
      }
      __sync_fetch_and_add(&bo->refcount, 1);
   } else {
      struct anv_bo new_bo;
      anv_bo_init(&new_bo, gem_handle, size);
      new_bo.map = host_ptr;
      new_bo.flags = bo_flags;
      new_bo.is_external = true;
      new_bo.from_host_ptr = true;

      if (!anv_vma_alloc(device, &new_bo)) {
         anv_gem_close(device, new_bo.gem_handle);
         pthread_mutex_unlock(&cache->mutex);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_OUT_OF_DEVICE_MEMORY,
                          "failed to allocate virtual address for BO");
      }

      *bo = new_bo;
   }

   pthread_mutex_unlock(&cache->mutex);
   *bo_out = bo;

   return VK_SUCCESS;
}

VkResult
anv_device_import_bo(struct anv_device *device,
                     int fd,
                     enum anv_bo_alloc_flags alloc_flags,
                     struct anv_bo **bo_out)
{
   assert(!(alloc_flags & (ANV_BO_ALLOC_MAPPED |
                           ANV_BO_ALLOC_SNOOPED |
                           ANV_BO_ALLOC_FIXED_ADDRESS)));

   struct anv_bo_cache *cache = &device->bo_cache;
   const uint32_t bo_flags =
      anv_bo_alloc_flags_to_bo_flags(device, alloc_flags);
   assert(bo_flags == (bo_flags & ANV_BO_CACHE_SUPPORTED_FLAGS));

   pthread_mutex_lock(&cache->mutex);

   uint32_t gem_handle = anv_gem_fd_to_handle(device, fd);
   if (!gem_handle) {
      pthread_mutex_unlock(&cache->mutex);
      return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
   }

   struct anv_bo *bo = anv_device_lookup_bo(device, gem_handle);
   if (bo->refcount > 0) {
      /* We have to be careful how we combine flags so that it makes sense.
       * Really, though, if we get to this case and it actually matters, the
       * client has imported a BO twice in different ways and they get what
       * they have coming.
       */
      uint64_t new_flags = 0;
      new_flags |= (bo->flags | bo_flags) & EXEC_OBJECT_WRITE;
      new_flags |= (bo->flags & bo_flags) & EXEC_OBJECT_ASYNC;
      new_flags |= (bo->flags & bo_flags) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
      new_flags |= (bo->flags | bo_flags) & EXEC_OBJECT_PINNED;
      new_flags |= (bo->flags | bo_flags) & EXEC_OBJECT_CAPTURE;

      /* It's theoretically possible for a BO to get imported such that it's
       * both pinned and not pinned.  The only way this can happen is if it
       * gets imported as both a semaphore and a memory object and that would
       * be an application error.  Just fail out in that case.
       */
      if ((bo->flags & EXEC_OBJECT_PINNED) !=
          (bo_flags & EXEC_OBJECT_PINNED)) {
         pthread_mutex_unlock(&cache->mutex);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
                          "The same BO was imported two different ways");
      }

      /* It's also theoretically possible that someone could export a BO from
       * one heap and import it into another or to import the same BO into two
       * different heaps.  If this happens, we could potentially end up both
       * allowing and disallowing 48-bit addresses.  There's not much we can
       * do about it if we're pinning so we just throw an error and hope no
       * app is actually that stupid.
       */
      if ((new_flags & EXEC_OBJECT_PINNED) &&
          (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) !=
          (bo_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) {
         pthread_mutex_unlock(&cache->mutex);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
                          "The same BO was imported on two different heaps");
      }

      bo->flags = new_flags;

      __sync_fetch_and_add(&bo->refcount, 1);
   } else {
      off_t size = lseek(fd, 0, SEEK_END);
      if (size == (off_t)-1) {
         anv_gem_close(device, gem_handle);
         pthread_mutex_unlock(&cache->mutex);
         return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
      }

      struct anv_bo new_bo;
      anv_bo_init(&new_bo, gem_handle, size);
      new_bo.flags = bo_flags;
      new_bo.is_external = true;

      if (!anv_vma_alloc(device, &new_bo)) {
         anv_gem_close(device, new_bo.gem_handle);
         pthread_mutex_unlock(&cache->mutex);
         return vk_errorf(device->instance, NULL,
                          VK_ERROR_OUT_OF_DEVICE_MEMORY,
                          "failed to allocate virtual address for BO");
      }

      *bo = new_bo;
   }

   pthread_mutex_unlock(&cache->mutex);
   *bo_out = bo;

   return VK_SUCCESS;
}

VkResult
anv_device_export_bo(struct anv_device *device,
                     struct anv_bo *bo, int *fd_out)
{
   assert(anv_device_lookup_bo(device, bo->gem_handle) == bo);

   /* This BO must have been flagged external in order for us to be able
    * to export it.  This is done based on external options passed into
    * anv_AllocateMemory.
    */
   assert(bo->is_external);

   int fd = anv_gem_handle_to_fd(device, bo->gem_handle);
   if (fd < 0)
      return vk_error(VK_ERROR_TOO_MANY_OBJECTS);

   *fd_out = fd;

   return VK_SUCCESS;
}

static bool
atomic_dec_not_one(uint32_t *counter)
{
   uint32_t old, val;

   val = *counter;
   while (1) {
      if (val == 1)
         return false;

      old = __sync_val_compare_and_swap(counter, val, val - 1);
      if (old == val)
         return true;

      val = old;
   }
}

void
anv_device_release_bo(struct anv_device *device,
                      struct anv_bo *bo)
{
   struct anv_bo_cache *cache = &device->bo_cache;
   assert(anv_device_lookup_bo(device, bo->gem_handle) == bo);

   /* Try to decrement the counter but don't go below one.  If this succeeds
    * then the refcount has been decremented and we are not the last
    * reference.
    */
   if (atomic_dec_not_one(&bo->refcount))
      return;

   pthread_mutex_lock(&cache->mutex);

   /* We are probably the last reference since our attempt to decrement above
    * failed.  However, we can't actually know until we are inside the mutex.
    * Otherwise, someone could import the BO between the decrement and our
    * taking the mutex.
    */
   if (unlikely(__sync_sub_and_fetch(&bo->refcount, 1) > 0)) {
      /* Turns out we're not the last reference.  Unlock and bail. */
      pthread_mutex_unlock(&cache->mutex);
      return;
   }
   assert(bo->refcount == 0);

   if (bo->map && !bo->from_host_ptr)
      anv_gem_munmap(bo->map, bo->size);

   if (!bo->has_fixed_address)
      anv_vma_free(device, bo);

   anv_gem_close(device, bo->gem_handle);

   /* Don't unlock until we've actually closed the BO.  The whole point of
    * the BO cache is to ensure that we correctly handle races with creating
    * and releasing GEM handles and we don't want to let someone import the BO
    * again between mutex unlock and closing the GEM handle.
    */
   pthread_mutex_unlock(&cache->mutex);
}