sys-lib 0.7.0

System facts for Rust services: OS, CPU, memory, disks, network, host, this process, virtualization, power, and device identity, container-aware by default, plus build info, instance info, and a process guard. No subprocesses, no third-party dependencies by default.
Documentation
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
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
<h1 align="center">
    <img width="99" alt="Rust logo" src="https://raw.githubusercontent.com/jamesgober/rust-collection/72baabd71f00e14aa9184efcb16fa3deddda3a0a/assets/rust-logo.svg">
    <br><b>sys-lib</b><br>
    <sub><sup>API REFERENCE</sup></sub>
</h1>
<div align="center">
    <sup>
        <a href="../README.md" title="Project Home"><b>HOME</b></a>
        <span>&nbsp;│&nbsp;</span>
        <span>API</span>
        <span>&nbsp;│&nbsp;</span>
        <a href="./SOURCES.md" title="Sources"><b>SOURCES</b></a>
        <span>&nbsp;│&nbsp;</span>
        <a href="../dev/PLAN.md" title="Plan"><b>PLAN</b></a>
    </sup>
</div>
<br>

The complete public API of sys-lib 1.0, locked at `0.2.0` and extended at James's review (D-028 to D-031): everything listed here ships before 1.0, nothing waits for a later 1.x. Every item carries the version it arrives in (**Since**); the published crate exposes only what is implemented, so an item marked `0.4` does not exist in `0.3`. Signatures here are the contract: a change after `0.2.0` is recorded in the decision log of [`../dev/PLAN.md`](../dev/PLAN.md). Where an operating system has no such fact, the item says so and returns `Error::Unsupported`, `None`, or an `Unknown` variant; it never guesses. The exact file, system call, or API behind every reading, per OS, is in [`SOURCES.md`](./SOURCES.md).

Costs are estimates (marked [est.]) until they are measured in `0.7.0`; measured numbers then replace them, labelled as measured.

<sub>Copyright &copy; 2026 <strong>James Gober</strong>.</sub>

## Table of contents

- [Setup](#setup)
- [Conventions](#conventions)
  - [Tiers](#tiers)
  - [Three kinds of answers](#three-kinds-of-answers)
  - [Platforms](#platforms)
- [Crate root](#crate-root)
  - [`VERSION`](#version)
  - [`Error` and `Result`](#error-and-result)
  - [`Bytes` and `Units`](#bytes-and-units)
  - [`Percent`](#percent)
  - [`Mhz`](#mhz)
  - [`Rate`](#rate)
  - [`Pressure`](#pressure)
  - [`summary()`](#summary)
- [`build`: build info](#build-build-info)
- [`instance`: instance info](#instance-instance-info)
- [`process`: this process](#process-this-process)
  - [Readings](#readings)
  - [`CpuSampler`](#cpusampler)
  - [Disk I/O of this process](#disk-io-of-this-process)
  - [How the program was started](#how-the-program-was-started)
  - [The process guard](#the-process-guard)
- [`os`: operating system](#os-operating-system)
  - [Kind and identity](#kind-and-identity)
  - [Form factor](#form-factor)
- [`cpu`: processor](#cpu-processor)
  - [Identity and counts](#identity-and-counts)
  - [Usage](#usage)
  - [Load](#load)
  - [CPU pressure](#cpu-pressure)
  - [Features](#features)
  - [Topology, caches, NUMA](#topology-caches-numa)
- [`mem`: memory](#mem-memory)
  - [Pressure and out-of-memory events](#pressure-and-out-of-memory-events)
  - [Swap and paging activity](#swap-and-paging-activity)
  - [Huge pages](#huge-pages)
  - [Memory modules](#memory-modules)
- [`disk`: disks and filesystems](#disk-disks-and-filesystems)
  - [Disk I/O: read and write](#disk-io-read-and-write)
- [`net`: network interfaces and traffic](#net-network-interfaces-and-traffic)
  - [Interfaces](#interfaces)
  - [Traffic: download and upload](#traffic-download-and-upload)
  - [Connections](#connections)
  - [DNS servers](#dns-servers)
  - [Interface options](#interface-options)
- [`host`: the machine and its environment](#host-the-machine-and-its-environment)
  - [System-wide counts](#system-wide-counts)
  - [Firmware and motherboard](#firmware-and-motherboard)
  - [Clock synchronization](#clock-synchronization)
  - [Sessions](#sessions)
- [`virt`: virtualization and containers](#virt-virtualization-and-containers)
- [`power`: battery and AC](#power-battery-and-ac)
- [`sensors`: temperatures](#sensors-temperatures)
- [`gpu`: graphics adapters](#gpu-graphics-adapters)
- [`readiness`: server readiness report](#readiness-server-readiness-report)
- [`bench`: benchmark environment report](#bench-benchmark-environment-report)
- [`security`: security posture](#security-security-posture)
- [`device`: device identity](#device-device-identity)
  - [Tier 1](#tier-1)
  - [`Builder`](#builder)
  - [Stability report](#stability-report)
  - [Fingerprint](#fingerprint)
  - [When not to use it](#when-not-to-use-it)
- [`probe`: explicit roots for monitoring agents](#probe-explicit-roots-for-monitoring-agents)
- [Errors by variant](#errors-by-variant)

## Setup

```toml
[dependencies]
sys = { package = "sys-lib", version = "0.4" }
# Without instance info and the process guard (no dependencies at all):
# sys = { package = "sys-lib", version = "0.4", default-features = false }
# With device identity (from 0.5):
# sys = { package = "sys-lib", version = "0.7", features = ["device-id"] }
```

The package is `sys-lib` on crates.io (`sys` and `system` were taken). Cargo's rename makes it `sys` in your code, so every file writes `sys::os::name()` with no import line. Without the rename (`sys-lib = "0.4"`), add `use sys_lib as sys;` at the top of a file and the code is the same.

| Feature | Default | Enables | Brings |
|---|---|---|---|
| `instance` | yes | `Instance`, `InstanceId`, `instance::current()` | clock-lib, id-forge |
| `guard` | yes | `process::guard`, `Guard`, `GuardBuilder`, `GuardReport`, `GuardWarning` | path-lib |
| `device-id` | no | the `device` module (from 0.5; before that the feature exists and enables nothing) | crypt-io (keyed BLAKE3, zeroize) |

Everything else is always compiled and has no dependencies. No feature brings a third-party crate except `device-id`, whose cryptography comes through crypt-io.

The examples here start with `use sys_lib as sys;` so they compile as written either way; with the Cargo rename that line is not needed.

## Conventions

### Tiers

| Tier | What | Examples |
|---|---|---|
| 1 | free functions, no configuration, a plain value or `Result<plain value>` | `sys::os::name()`, `sys::mem::total()`, `sys::process::guard("site")` |
| 2 | snapshots, samplers, builders, options | `sys::mem::snapshot()`, `sys::cpu::Sampler`, `Guard::builder` |
| 3 | caller buffers and raw counters | `Sampler::sample_per_core_into`, `sys::cpu::times()` |

### Three kinds of answers

| Kind | Cached | Cost after the first call | Examples |
|---|---|---|---|
| compile-time constant | in the binary | 0 | `os::kind()`, `cpu::arch()` |
| static fact | once per process, in a `OnceLock` | about 1 ns [est.] | `os::name()`, `cpu::name()`, `cpu::physical()`, `virt::detect()` |
| live reading | never | one file read or system call, 0.2 to 20 µs [est.] | `mem::used()`, `host::uptime()`, `disk::space()` |
| sampled value | in a caller-owned sampler | one reading and a subtraction | `cpu::Sampler`, `net::Sampler` |

A static fact that failed is cached as failed (the reason is kept, not retried). Live readings that look static but are not: `mem::total()` (cgroup resize, ballooning), `cpu::available()` and `cpu::online()` (affinity, hot-plug, quota), `host::name()`, `host::timezone()`, `net::interfaces()`.

Every public function is safe to call from any thread at any time. All public types are `Send + Sync`; samplers are `Send` and take `&mut self`. sys-lib never starts a thread, never runs another program, and holds no file descriptor or handle after a call returns (the guard's lock file excepted).

### Platforms

Linux, macOS, and Windows are first class (x86_64 and aarch64). Android and iOS are build-checked and best effort: the Linux or Apple code path where it applies unchanged, `Error::Unsupported` elsewhere, as listed per item. Other targets compile and return `Error::Unsupported` from every reading.

Unless an item says otherwise, a function returning `Result` can fail with `Unsupported`, `PermissionDenied`, `NotFound`, `Malformed`, `TooLarge`, or `Os`; see [Errors by variant](#errors-by-variant).

---

## Crate root

### `VERSION`

**Since** 0.1 · ROOT-01

```rust
pub const VERSION: &str;
```

The crate version, populated by Cargo at build time.

```rust
use sys_lib as sys;

println!("sys-lib {}", sys::VERSION);
```

### `Error` and `Result`

**Since** 0.3 · ROOT-02

```rust
#[derive(Debug)]
#[non_exhaustive]
pub enum Error {
    Unsupported { what: &'static str, platform: &'static str },
    PermissionDenied { what: &'static str, hint: &'static str },
    NotFound { what: &'static str, source: &'static str },
    Malformed { what: &'static str, source: &'static str, line: Option<u32> },
    TooLarge { what: &'static str, limit: u64 },
    TooSoon { elapsed: Duration, minimum: Duration },
    Os { what: &'static str, call: &'static str, code: i32 },
    Io { what: &'static str, source: std::io::Error },
    InvalidInput { what: &'static str, reason: &'static str },
    AlreadyRunning { pid: Option<u32>, lock_file: PathBuf },
    RunDirUnavailable { path: PathBuf, source: std::io::Error },
    OpenFilesTooLow { have: u64, need: u64 },
    #[cfg(feature = "device-id")]
    NoIdentifierSource { basis: device::Basis },
}

pub type Result<T, E = Error> = core::result::Result<T, E>;

impl Error {
    /// True for errors that mean "this fact is not available here", which a caller
    /// usually treats as unknown: Unsupported, PermissionDenied, NotFound.
    pub fn is_unavailable(&self) -> bool;
}
```

`Error` is `Send + Sync + 'static` and implements `std::error::Error` (`source()` returns the I/O error for `Io` and `RunDirUnavailable`). `Display` is one sentence saying what failed and what to do. No variant ever carries file content, registry content, or identifier values: `Malformed` names the source (a constant path) and the line number only.

```rust
use sys_lib as sys;

match sys::mem::pressure() {
    Ok(Some(p)) => println!("memory pressure {:?}", p.level),
    Ok(None) => println!("this kernel reports no memory pressure"),
    Err(e) if e.is_unavailable() => println!("unknown here: {e}"),
    Err(e) => return Err(e),
}
# Ok::<(), sys_lib::Error>(())
```

Every variant, its meaning, and what the message tells the caller to do: [Errors by variant](#errors-by-variant).

### `Bytes` and `Units`

**Since** 0.3 · ROOT-03

```rust
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Default)]
pub struct Bytes(/* u64 */);

impl Bytes {
    pub const ZERO: Bytes;
    pub const MAX: Bytes;
    pub const fn new(bytes: u64) -> Bytes;
    pub const fn raw(self) -> u64;
    pub const fn from_kib(n: u64) -> Bytes;           // saturating
    pub const fn from_mib(n: u64) -> Bytes;           // saturating
    pub const fn from_gib(n: u64) -> Bytes;           // saturating
    pub const fn saturating_add(self, other: Bytes) -> Bytes;
    pub const fn saturating_sub(self, other: Bytes) -> Bytes;
    pub const fn checked_add(self, other: Bytes) -> Option<Bytes>;
    pub const fn checked_sub(self, other: Bytes) -> Option<Bytes>;
    pub fn percent_of(self, total: Bytes) -> Percent; // 0% when total is zero; clamped to 100%
    pub fn display(self, units: Units) -> impl fmt::Display;
    pub fn binary(self) -> impl fmt::Display;         // "125.6 GiB"
    pub fn decimal(self) -> impl fmt::Display;        // "134.9 GB"
    pub fn jedec(self) -> impl fmt::Display;          // "125.6 GB" (1024-based, JEDEC labels)
}
impl From<u64> for Bytes;
impl From<Bytes> for u64;
impl fmt::Display for Bytes;  // binary by default, decimal with {:#}
impl fmt::Debug for Bytes;    // Bytes(134876545024)

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Units { Binary, Decimal, Jedec }
```

A byte count with an exact `u64` inside and a `Display` that picks the unit. Formatting is deterministic, allocation-free, and done in integer arithmetic, so it never depends on floating-point printing:

| Rule | Detail |
|---|---|
| `{}` | binary values, IEC labels: `B`, `KiB`, `MiB`, `GiB`, `TiB`, `PiB`, `EiB` |
| `{:#}` | decimal SI: `B`, `kB`, `MB`, `GB`, `TB`, `PB`, `EB` |
| `.jedec()` | binary values with JEDEC labels: `B`, `KB`, `MB`, `GB`, `TB`, `PB`, `EB` (what Windows Explorer shows) |
| unit | the largest unit for which the value is at least 1 |
| precision | one decimal place, trailing `.0` dropped (`128 GiB`, `127.9 GiB`); `{:.0}`, `{:.2}` honoured |
| under 1 KiB | an integer: `0 B`, `1 B`, `512 B` |
| rounding | half away from zero |
| carry | a value that rounds up to 1024.0 of a unit prints as 1 of the next (`1 MiB`, never `1024 KiB`) |
| width, fill, alignment | honoured (`{:>10}`) |
| maximum | `u64::MAX` prints `16 EiB` and, with `{:#}`, `18.4 EB` |

```rust
use sys_lib::{Bytes, Units};

let ram = Bytes::new(134_876_545_024);
assert_eq!(ram.to_string(), "125.6 GiB");
assert_eq!(format!("{ram:#}"), "134.9 GB");
assert_eq!(ram.jedec().to_string(), "125.6 GB");
assert_eq!(format!("{ram:.2}"), "125.61 GiB");
assert_eq!(ram.display(Units::Decimal).to_string(), "134.9 GB");
assert_eq!(ram.raw(), 134_876_545_024);

assert_eq!(Bytes::new(512).to_string(), "512 B");
assert_eq!(Bytes::from_gib(128).to_string(), "128 GiB");
assert_eq!(Bytes::new(1_048_575).to_string(), "1 MiB");      // 1023.999 KiB carries
assert_eq!(Bytes::MAX.to_string(), "16 EiB");
assert_eq!(Bytes::from_gib(u64::MAX), Bytes::MAX);          // saturating
```

```rust
use sys_lib as sys;

let total = sys::mem::total()?;
let used = sys::mem::used()?;
println!("{used} of {total} ({})", used.percent_of(total));  // "41.2 GiB of 125.6 GiB (32.8%)"
# Ok::<(), sys_lib::Error>(())
```

### `Percent`

**Since** 0.3 · ROOT-04

```rust
#[derive(Clone, Copy, Default)]
pub struct Percent(/* f32 */);   // always 0.0..=100.0, never NaN

impl Percent {
    pub const ZERO: Percent;
    pub const FULL: Percent;
    pub fn from_ratio(part: u64, whole: u64) -> Percent;  // clamped; whole == 0 gives ZERO
    pub fn from_value(value: f32) -> Percent;             // clamped; NaN gives ZERO
    pub fn value(self) -> f32;                            // 0.0..=100.0
    pub fn fraction(self) -> f32;                         // 0.0..=1.0
    pub fn complement(self) -> Percent;                   // 100 - value
}
impl PartialEq, Eq, PartialOrd, Ord, Hash for Percent;  // total order: the value is never NaN
impl fmt::Display for Percent;   // "32.8%"; precision honoured
impl fmt::Debug for Percent;     // Percent(32.8)
```

A share that is always within 0 to 100 and never NaN, so it can be sorted, hashed, and compared without surprises. `from_ratio` computes `part * 1000 / whole` in `u128` and converts once, so a percentage of two integers is exact to 0.1%.

```rust
use sys_lib::Percent;

let p = Percent::from_ratio(41, 125);
assert_eq!(p.to_string(), "32.8%");
assert_eq!(format!("{p:.0}"), "33%");
assert_eq!(p.complement().to_string(), "67.2%");
assert_eq!(Percent::from_ratio(5, 0), Percent::ZERO);
assert_eq!(Percent::from_ratio(9, 4), Percent::FULL);    // clamped
assert_eq!(Percent::from_value(f32::NAN), Percent::ZERO);
assert!(Percent::FULL > p);
```

### `Mhz`

**Since** 0.4 · ROOT-05

```rust
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
pub struct Mhz(pub u32);
impl fmt::Display for Mhz;   // "3.0 GHz" at or above 1000 MHz, else "800 MHz"
```

```rust
use sys_lib::Mhz;

assert_eq!(Mhz(5_700).to_string(), "5.7 GHz");
assert_eq!(Mhz(800).to_string(), "800 MHz");
```

### `Rate`

**Since** 0.4 · ROOT-07

```rust
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Default)]
pub struct Rate(/* bytes per second, u64 */);

impl Rate {
    pub const ZERO: Rate;
    pub const fn from_bytes_per_sec(n: u64) -> Rate;
    pub const fn bytes_per_sec(self) -> u64;
    pub const fn bits_per_sec(self) -> u64;           // saturating
    pub fn per(bytes: Bytes, interval: Duration) -> Rate;   // zero for an empty interval
    pub fn bits(self) -> impl fmt::Display;           // "104.9 Mbit/s": how network links are rated
}
impl fmt::Display for Rate;   // "12.5 MiB/s"; {:#} "13.1 MB/s"; the same rules as Bytes
impl fmt::Debug for Rate;     // Rate(13107200/s)
```

A transfer rate, used for network download and upload. It prints like `Bytes` with `/s`, and `.bits()` gives the bit rate network links are sold in.

```rust
use sys_lib::Rate;

let r = Rate::from_bytes_per_sec(13_107_200);
assert_eq!(r.to_string(), "12.5 MiB/s");
assert_eq!(r.bits().to_string(), "104.9 Mbit/s");
```

### `Pressure`

**Since** 0.4 · ROOT-08

```rust
#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct Pressure {
    pub level: PressureLevel,
    pub some: Stall,               // share of time at least one task was stalled
    pub full: Option<Stall>,       // share of time every non-idle task was stalled (None where not reported: host CPU before Linux 5.13)
    pub total: Option<Duration>,   // cumulative stall time
}

impl Pressure {
    pub const WARNING_SOME_AVG10: f32;    // 10.0
    pub const CRITICAL_SOME_AVG10: f32;   // 40.0
    pub const CRITICAL_FULL_AVG10: f32;   // 10.0
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug, Default)]
#[non_exhaustive]
pub struct Stall { pub avg10: Percent, pub avg60: Percent, pub avg300: Percent }

#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
#[non_exhaustive]
pub enum PressureLevel { Normal, Warning, Critical }
```

How much work is stalled waiting for a resource: the best single signal that a machine (or a container) is overloaded, because it measures waiting rather than busyness. A CPU at 100% with no stall is doing its job; a CPU with 30% of tasks stalled is turning work away. `cpu::pressure()`, `mem::pressure()`, and `disk::pressure()` all return it. On Linux it is the kernel's pressure-stall information (PSI, kernel 4.20 and later), read from the process's own cgroup when it has one (so a container sees its own pressure) and from `/proc/pressure/` otherwise. `level` is `Warning` from 10% `some.avg10` and `Critical` from 40% (or any `full.avg10` above 10%), thresholds documented as constants. Where an OS has no such measure the functions return `Ok(None)`.

### `summary()`

**Since** 0.5 · ROOT-06

```rust
pub fn summary() -> Summary;

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Summary {
    pub os: &'static str, pub kernel: &'static str,
    pub cpu: &'static str, pub cpu_counts: cpu::Counts,
    pub memory: Option<mem::Memory>, pub cpu_quota: Option<cpu::CpuQuota>,
    pub virtualization: &'static virt::Virtualization,
    pub form_factor: os::FormFactor,
    pub uptime: Option<Duration>,
    pub load: Option<cpu::Load>,
}
impl Summary {
    pub fn write_kv(&self, out: &mut impl fmt::Write) -> fmt::Result;
    pub fn write_json(&self, out: &mut impl fmt::Write) -> fmt::Result;
}
impl fmt::Display for Summary;   // several lines, for the first log lines and support bundles
```

Every static fact plus one memory snapshot and the uptime, for the start-up log line, a health endpoint, or a support bundle. It never fails: a reading that fails becomes `None` in its field. It deliberately carries **no** host name, user name, MAC address, or device id, because it is meant for output that leaves the machine (threat T13). Cost: the first call reads every static fact; later calls 5 to 20 µs [est.] (memory and uptime are live).

```rust
use sys_lib as sys;

let summary = sys::summary();
println!("{summary}");
// os=Ubuntu 24.04.1 LTS kernel=6.8.0-45-generic cpu="AMD Ryzen 9 7950X" cores=16 threads=32
// available=2 memory.total=256 MiB memory.used=41 MiB limit=cgroup2 virt=docker form=unknown

let mut line = String::new();
summary.write_json(&mut line)?;
# Ok::<(), std::fmt::Error>(())
```

---

## `build`: build info

**Since** 0.3 (BLD-01 to BLD-06; BLD-06 was planned for 0.4 and shipped with the helper, D-032)

Facts fixed when the application was compiled, answering "exactly which build is running?" in logs, crash reports, health checks, and metrics. Zero run-time cost: every field is a `&'static str` in the binary.

```rust
pub mod build {
    pub fn emit();                          // call from the application's build.rs
    pub fn emitter() -> Emitter;            // Tier 2
    #[derive(Clone, Copy, Debug)]
    #[must_use]
    pub struct Emitter { /* private */ }
    impl Emitter {
        pub fn git(self, on: bool) -> Self;        // default true
        pub fn rustc(self, on: bool) -> Self;      // default true
        pub fn features(self, on: bool) -> Self;   // default true
        pub fn emit(self);
    }
}

#[macro_export]
macro_rules! build_info { () => { /* &'static BuildInfo, expanded in the calling crate */ } }

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct BuildInfo {
    pub name: &'static str,          // the application's CARGO_PKG_NAME
    pub version: &'static str,       // the application's CARGO_PKG_VERSION
    pub commit: &'static str,        // full hash, or "unknown"
    pub commit_short: &'static str,  // 7 characters, or "unknown"
    pub dirty: Option<bool>,         // None when unknown
    pub built_at: &'static str,      // RFC 3339 UTC, or "unknown"
    pub profile: &'static str,       // "release", "debug"
    pub target: &'static str,        // "x86_64-unknown-linux-gnu"
    pub rustc: &'static str,         // "1.89.0"
    pub features: &'static str,      // comma-separated, sorted
}
impl BuildInfo {
    pub fn public(&self) -> PublicBuild<'_>;                            // name and version only
    pub fn write_kv(&self, out: &mut impl fmt::Write) -> fmt::Result;   // build.version=0.3.0 ...
    pub fn write_json(&self, out: &mut impl fmt::Write) -> fmt::Result;
}
impl fmt::Display for BuildInfo;   // site 0.3.0 (9db9e36, release, built 2026-10-21T18:02:11Z)
                                   // "9db9e36-dirty" when tracked files differed; unknown parts left out

#[derive(Clone, Copy, Debug)]
pub struct PublicBuild<'a> { /* private */ }
impl fmt::Display for PublicBuild<'_>;   // site 0.3.0
```

| Id | Item | Tier | Cost |
|---|---|---|---|
| BLD-01 | `build::emit()` | 1 | build time only: runs `git` and `$RUSTC`, 20 to 60 ms [est.] |
| BLD-02 | `build_info!()` | 1 | 0 at run time |
| BLD-03 | `BuildInfo` fields, `Display` | 1 | `Display` writes about 80 bytes |
| BLD-04 | `write_kv`, `write_json` | 2 | no allocation |
| BLD-05 | `public()` | 1 | 0 |
| BLD-06 | `emitter()` | 2 | build time only |

**How it works.** The application adds sys-lib as a build-dependency and calls `sys_lib::build::emit()` from its own `build.rs`. At build time the helper runs `git -C $CARGO_MANIFEST_DIR rev-parse HEAD`, `git status --porcelain --untracked-files=no`, and `$RUSTC --version`, then hands the values to the compiler with `cargo:rustc-env=SYS_LIB_BUILD_*`. It prints `cargo:rerun-if-changed` for the git `HEAD` file, the ref it points to (followed through a `.git` file in worktrees), `packed-refs`, and the index (so staging a change refreshes the dirty flag), and for the package's `src/`, `build.rs`, and `Cargo.toml`, plus `cargo:rerun-if-env-changed=SOURCE_DATE_EPOCH`, so a new commit refreshes the values without rebuilding every time. A `HEAD` naming anything but a well-formed ref under `refs/` is ignored, and a path containing a line break is never printed, so a tampered repository cannot add a line to the build script's output (threat T11). The macro expands in the application crate, so it reads the **application's** name and version; it reads the helper's values with `option_env!`, so an application that never called `emit()` still compiles and shows `unknown`.

**Never fails the build.** `git` missing, a crates.io tarball, a shallow clone, a detached HEAD: the value becomes `unknown`. `SOURCE_DATE_EPOCH`, when set, wins over the clock (reproducible builds); a value that is not a non-negative integer is ignored with a `cargo:warning`. Every value is checked (printable ASCII, at most 128 bytes; 512 for the feature list) before it reaches `rustc-env`, because a branch name can contain anything (threat T11). The helper records no build host, user name, or source path.

**At run time sys-lib starts no program.** The helper is the only code in the crate that runs one, and it runs in the application's build script.

```rust,ignore
// The application's build.rs (sys-lib in [build-dependencies] as well as [dependencies])
fn main() {
    sys_lib::build::emit();
}
```

```rust
use sys_lib as sys;

let build: &'static sys::BuildInfo = sys::build_info!();
println!("{build}");                         // site 0.3.0 (9db9e36, release, built 2026-10-21T18:02:11Z)
println!("{}", build.public());              // site 0.3.0: safe for a public health endpoint
if build.dirty == Some(true) {
    println!("warning: built from uncommitted changes");
}
```

```rust,ignore
// Only the commit, no rustc version or feature list.
fn main() {
    sys_lib::build::emitter().rustc(false).features(false).emit();
}
```

---

## `instance`: instance info

**Since** 0.3 · feature `instance` (default)

Facts fixed when the process **started**, answering "which running copy is this?" when several run at once (several servers, or the old and new process during a hand-off restart). Put the instance id in every log line, audit record, and metric so events from two processes are never mixed up.

```rust
pub struct Instance { /* private */ }
impl Instance {
    pub fn start() -> &'static Instance;       // the first call creates it; later calls return the same
    pub fn id(&self) -> InstanceId;
    pub fn started_at(&self) -> SystemTime;
    pub fn uptime(&self) -> Duration;          // monotonic: wall-clock steps do not affect it
    pub fn host(&self) -> &str;                // host name at start, "unknown" if unreadable
    pub fn pid(&self) -> u32;
    pub fn user_at_start(&self) -> &str;       // effective user at start, "unknown" if unreadable
    pub fn is_current_process(&self) -> bool;  // false in a fork()ed child
    pub fn write_kv(&self, out: &mut impl fmt::Write) -> fmt::Result;
    pub fn write_json(&self, out: &mut impl fmt::Write) -> fmt::Result;
}
impl fmt::Display for Instance;   // 01J9ZQ4M8T6Y2K5V3W1X7R0B9C web-01 pid 48213 up 3h12m04s

pub mod instance {
    pub fn current() -> Option<&'static Instance>;   // without creating it
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub struct InstanceId(/* [u8; 16] */);
impl InstanceId {
    pub fn as_bytes(&self) -> &[u8; 16];
    pub fn parse(text: &str) -> Result<InstanceId>;   // exactly 26 Crockford base32 characters
}
impl fmt::Display for InstanceId;   // 26-character ULID
impl fmt::Debug for InstanceId;     // InstanceId(01J9ZQ4M8T6Y2K5V3W1X7R0B9C)
impl FromStr for InstanceId;
```

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| INS-01 | `Instance::start()` | 1 | first call: a ULID (about 50 ns), the host name, and the user (5 to 20 µs); later about 1 ns |
| INS-02 | `instance::current()` | 1 | about 1 ns |
| INS-03 | `id`, `started_at`, `uptime`, `host`, `pid`, `user_at_start` | 1 | about 1 ns; `uptime` one monotonic clock read (about 20 ns) |
| INS-04 | `InstanceId` | 1 | `Display` about 30 ns |
| INS-05 | `write_kv`, `write_json`, `Display` | 2 | no allocation |
| INS-06 | `is_current_process()` | 1 | about 20 ns |

`start()` never fails and is safe to call from many threads at once: exactly one instance is created. The user is recorded **at start**, before a privilege drop; the current effective user is always `process::user()`. A process created with `fork()` and no `exec` inherits the parent's instance; `is_current_process()` says so, and forked workers must not log under the parent's id. `InstanceId::parse` rejects anything but exactly 26 valid characters with `InvalidInput`.

The host name and user name are identifying values. They are in the instance by design (the instance names the running copy); keep them out of output that leaves your control.

```rust
use sys_lib::Instance;

let me = Instance::start();
println!("{me}");                                  // 01J9ZQ4M8T6Y2K5V3W1X7R0B9C web-01 pid 48213 up 0s
let id = me.id();
assert_eq!(id.to_string().len(), 26);
assert_eq!(id.to_string().parse::<sys_lib::InstanceId>()?, id);
assert!(std::ptr::eq(me, Instance::start()));     // one instance per process
# Ok::<(), sys_lib::Error>(())
```

```rust
// Tag every log line.
let instance = sys_lib::Instance::start().id();
eprintln!("instance={instance} event=boot phase=0");
```

---

## `process`: this process

### Readings

**Since** 0.3 (PRC-01, PRC-11, PRC-12), 0.4 (PRC-02 to PRC-10, PRC-13; PRC-05 was planned for 0.5 and ships with `info()`, D-044)

Facts about **this** process. sys-lib does not list or inspect other processes.

```rust
pub mod process {
    pub fn id() -> u32;
    pub fn parent_id() -> Option<u32>;
    pub fn memory() -> Result<Bytes>;                  // resident set size
    pub fn peak_memory() -> Result<Bytes>;             // high-water resident set size
    pub fn memory_detail() -> Result<ProcessMemory>;
    pub fn cpu_time() -> Result<CpuTime>;              // user and system time used so far
    pub fn threads() -> Result<usize>;
    pub fn open_files() -> Result<usize>;              // descriptors (Unix) or handles (Windows)
    pub fn limits() -> Result<Limits>;
    pub fn started_at() -> Result<SystemTime>;
    pub fn user() -> Result<User>;                     // the effective user now
    pub fn is_elevated() -> Result<bool>;              // root, or an elevated Administrator token
    pub fn info() -> Result<ProcessInfo>;              // everything above, in the fewest reads
    pub fn info_with(options: InfoOptions) -> Result<ProcessInfo>;
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct ProcessMemory {
    pub resident: Bytes, pub peak_resident: Bytes, pub virtual_size: Bytes,
    pub private: Option<Bytes>,      // anonymous / private bytes (commit charge on Windows)
    pub shared: Option<Bytes>,
    pub swapped: Option<Bytes>,
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct CpuTime { pub user: Duration, pub system: Duration }
impl CpuTime { pub fn total(&self) -> Duration; }

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Limits {
    pub open_files: Limit, pub processes: Limit, pub address_space: Limit,
    pub core_dump: Limit, pub stack: Limit,
}

#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
#[non_exhaustive]
pub struct Limit { pub soft: Option<u64>, pub hard: Option<u64> }   // None = unlimited or not applicable
                                                                    // since 0.3, with the guard's report

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct User { pub name: String, pub id: Option<u32>, pub is_admin: bool }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct ProcessInfo {
    pub id: u32, pub parent_id: Option<u32>,
    pub memory: ProcessMemory, pub cpu_time: CpuTime,
    pub threads: Option<usize>, pub open_files: Option<usize>,
    pub limits: Limits, pub started_at: Option<SystemTime>,
    pub user: Option<User>, pub is_elevated: Option<bool>,
}

#[derive(Clone, Copy, Debug, Default)]
#[non_exhaustive]
pub struct InfoOptions { pub with_threads: bool }   // Windows: the thread count walks a system-wide snapshot
```

| Id | Item | Since | Linux | macOS | Windows | Cost [est.] |
|---|---|---|---|---|---|---|
| PRC-01 | `id()` | 0.3 | std | std | std | 20 ns |
| PRC-02 | `parent_id()` | 0.4 | `PPid` in `/proc/self/status` | `getppid` | toolhelp snapshot entry for this pid, cached after the first call | 2 µs / 0.1 µs / 1 ms first, then 1 ns |
| PRC-03 | `memory()` | 0.4 | `/proc/self/statm` field 2 × page size | `task_info(MACH_TASK_BASIC_INFO)` | `K32GetProcessMemoryInfo` working set | 2 µs / 0.5 µs / 0.5 µs |
| PRC-04 | `peak_memory()` | 0.4 | `VmHWM` | `getrusage` `ru_maxrss` (bytes on macOS, KiB on Linux) | `PeakWorkingSetSize` | 3 µs / 0.5 µs / 0.5 µs |
| PRC-05 | `memory_detail()` | 0.4 | `/proc/self/status` | `task_info` + `proc_pid_rusage` | `PROCESS_MEMORY_COUNTERS_EX` | 4 µs / 2 µs / 1 µs |
| PRC-06 | `cpu_time()` | 0.4 | `getrusage` | `getrusage` | `GetProcessTimes` | 0.3 µs |
| PRC-07 | `threads()` | 0.4 | `Threads` in `/proc/self/status` | `proc_pidinfo(PROC_PIDTASKINFO)` | toolhelp process snapshot (every process on the machine), this process's thread count | 3 µs / 2 µs / 5 to 60 ms |
| PRC-08 | `open_files()` | 0.4 | entries of `/proc/self/fd` | `proc_pidinfo(PROC_PIDLISTFDS)` | `GetProcessHandleCount` | 1 µs per 100 / 2 µs / 0.3 µs |
| PRC-09 | `limits()` | 0.4 | `getrlimit` × 5 | `getrlimit` × 5 | job-object limits when present; otherwise `None` | 1 µs |
| PRC-10 | `started_at()` | 0.4 | `/proc/self/stat` field 22 + boot time | `proc_pidinfo(PROC_PIDTBSDINFO)` | `GetProcessTimes` creation time | 4 µs / 2 µs / 0.3 µs |
| PRC-11 | `user()` | 0.3 | `geteuid` + `getpwuid_r` | same | `GetUserNameW`; `is_admin` from the token | 5 to 50 µs |
| PRC-12 | `is_elevated()` | 0.3 | `geteuid() == 0` | same | `GetTokenInformation(TokenElevation)` | 0.1 µs / 0.1 µs / 2 µs |
| PRC-13 | `info()`, `info_with()` | 0.4 | each reading once | | | about the sum of the readings |

`memory()`, `peak_memory()`, `cpu_time()`, `threads()` (Linux, macOS), `open_files()`, and `limits()` allocate nothing. `user()` allocates the name. On Linux and macOS, `user()` goes through the system's user database (NSS), which can read `/etc/passwd` or, with LDAP users, wait on the network; that is stated so callers do not put it on a hot path. On Windows, `threads()` walks a system-wide thread snapshot whose cost grows with the machine's total thread count; `info()` skips it unless `InfoOptions { with_threads: true }` is passed.

`is_elevated()` is a check, not a control: privilege drops belong to shield-sec, which calls `is_elevated()` afterwards to verify.

```rust
use sys_lib as sys;

println!("pid {} rss {} peak {}", sys::process::id(), sys::process::memory()?, sys::process::peak_memory()?);
let t = sys::process::cpu_time()?;
println!("cpu {:?} (user {:?}, system {:?})", t.total(), t.user, t.system);
let files = sys::process::limits()?.open_files;
println!("open files {} of {:?}", sys::process::open_files()?, files.soft);
# Ok::<(), sys_lib::Error>(())
```

```rust
use sys_lib as sys;

// After shield-sec dropped privileges:
if sys::process::is_elevated()? {
    return Err("still running elevated; refusing to open the door".into());
}
println!("running as {}", sys::process::user()?.name);
# Ok::<(), Box<dyn std::error::Error>>(())
```

### Disk I/O of this process

**Since** 0.4 · PRC-15

```rust
pub mod process {
    pub fn io() -> Result<ProcessIo>;     // since the process started; no allocation
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct ProcessIo {
    pub read: Bytes,                // from storage
    pub written: Bytes,             // to storage
    pub read_ops: Option<u64>,
    pub write_ops: Option<u64>,
    pub storage_only: bool,         // false where the OS counts all I/O (Windows counts network too)
}
```

| OS | Source | Cost [est.] |
|---|---|---|
| Linux | `/proc/self/io` (`read_bytes`, `write_bytes`: storage; `syscr`, `syscw`) | 3 µs |
| macOS | `proc_pid_rusage(RUSAGE_INFO_V2)` (`ri_diskio_bytesread`, `ri_diskio_byteswritten`) | 2 µs |
| Windows | `GetProcessIoCounters` (all I/O, files and network: `storage_only` is false) | 0.3 µs |

How much this process has read from and written to storage, for a database's own write amplification or a backup tool's throughput. A `process::IoSampler` is not needed: `ProcessIo` values subtract (`Sub` is implemented) and `Rate::per(bytes, interval)` turns a difference into a rate.

### How the program was started

**Since** 0.5 · PRC-24, PRC-25

```rust
pub mod process {
    pub fn launch() -> &'static Launch;    // static: decided at the first call
    pub fn is_service() -> bool;           // started by a service manager or an orchestrator
    pub fn is_interactive() -> bool;       // standard input and output are a terminal (live)
}

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Launch {
    pub by: Launcher,
    pub unit: Option<&'static str>,        // "site.service", the launchd label, the Windows service name
    pub notify_socket: bool,               // systemd's NOTIFY_SOCKET is set: readiness can be reported
    pub evidence: Vec<virt::Evidence>,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Launcher { Systemd, Launchd, WindowsService, ContainerInit, Init, Terminal, Ssh, Cron, Unknown }
```

| Launcher | Linux | macOS | Windows |
|---|---|---|---|
| `Systemd` | `INVOCATION_ID` and the `.service` unit in `/proc/self/cgroup` | | |
| `Launchd` | | parent pid 1 and `XPC_SERVICE_NAME` set to a label (weak evidence) | |
| `WindowsService` | | | session 0 (`ProcessIdToSessionId`) and no console window; a scheduled task running as SYSTEM looks the same, and `unit` is `None` |
| `ContainerInit` | process id 1 inside a container (`virt`) | | |
| `Init` | process id 1 outside a container | | |
| `Ssh` | `SSH_CONNECTION` (weak evidence) | same | |
| `Cron` | `/proc/self/cgroup` in `cron.service` (strong), or the cron environment (weak) | | |
| `Terminal` | standard input or output is a terminal | same | a console is attached and it is interactive |

How the program was started, so a service can choose its behaviour without configuration: plain JSON logs and readiness reporting under systemd, colour and progress output in a terminal, no console tricks in a Windows service. Each answer carries its evidence (environment variables are weak evidence: any parent can set them). sys-lib reports `notify_socket`; sending the readiness message is the service kernel's job.

```rust
use sys_lib as sys;

match sys::process::launch().by {
    sys::process::Launcher::Systemd => { /* JSON logs to the journal; report readiness */ }
    sys::process::Launcher::Terminal => { /* human-readable logs */ }
    _ => {}
}
if sys::process::is_interactive() { println!("press Ctrl-C to stop"); }
```

### `CpuSampler`

**Since** 0.4 · PRC-14 · Tier 2

```rust
pub mod process {
    pub struct CpuSampler { /* previous CpuTime and Instant */ }
    impl CpuSampler {
        pub fn new() -> Result<CpuSampler>;
        pub fn sample(&mut self) -> Result<ProcessUsage>;   // never blocks; no allocation
    }
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct ProcessUsage {
    pub of_one_core: f32,     // 0.0..=100.0 × logical CPUs; 250.0 = two and a half cores busy
    pub of_machine: Percent,  // share of every logical CPU
    pub interval: Duration,
}
```

The share of CPU this process used since the previous sample. Owned by the caller: two monitors keep two samplers and never disturb each other. Returns `TooSoon` when no time passed on the OS clock since the previous sample. Cost: about 0.4 µs per sample [est.].

```rust
use sys_lib as sys;
use std::time::Duration;

let mut cpu = sys::process::CpuSampler::new()?;
std::thread::sleep(Duration::from_millis(200));
let u = cpu.sample()?;
println!("this process: {:.1}% of one core, {} of the machine", u.of_one_core, u.of_machine);
# Ok::<(), sys_lib::Error>(())
```

### The process guard

**Since** 0.3 · PRC-20 to PRC-23 · feature `guard` (default)

What a service does before it touches its first file: make sure it is the only copy running, make every file it creates private by default, and make sure it will not run out of file descriptors under load. One call, secure defaults, refusal with an actionable error when a requirement cannot be met. Hold the returned `Guard` for the life of the process.

```rust
pub mod process {
    pub fn guard(name: &str) -> Result<Guard>;
}

#[must_use = "dropping the guard releases the single-instance lock"]
pub struct Guard { /* lock file, report */ }
impl Guard {
    pub fn builder(name: &str) -> GuardBuilder;
    pub fn report(&self) -> &GuardReport;
    pub fn lock_file(&self) -> Option<&Path>;
    pub fn is_held(&self) -> bool;     // since 0.7: the file at the lock path is still the one locked
}
impl fmt::Debug for Guard;
// Dropping the guard closes the lock file, which releases the lock; it never deletes the file.

#[derive(Clone, Debug)]
pub struct GuardBuilder { /* private */ }
impl GuardBuilder {
    pub fn lock_dir(self, dir: impl Into<PathBuf>) -> Self;   // default: from the user alone, see below
    pub fn lock(self, on: bool) -> Self;                      // default true; false when a supervisor holds the lock
    pub fn allow_network_lock(self, on: bool) -> Self;        // default false
    pub fn umask(self, mask: u32) -> Self;                    // default 0o027; Unix only
    pub fn keep_umask(self) -> Self;                          // leave the umask alone
    pub fn open_files_target(self, n: u64) -> Self;           // default 65_536
    pub fn open_files_minimum(self, n: u64) -> Self;          // default 1_024
    pub fn build(self) -> Result<Guard>;
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct GuardReport {
    pub lock_file: Option<PathBuf>,
    pub lock_filesystem: Option<&'static str>,      // "ext4", "apfs", "NTFS", ...
    pub umask_before: Option<u32>, pub umask_after: Option<u32>,   // None on Windows
    pub open_files_before: Limit, pub open_files_after: Limit,
    pub warnings: Vec<GuardWarning>,
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub enum GuardWarning {
    /// Other threads were running when the umask changed; files they created
    /// meanwhile may have the old permissions. Call the guard before starting threads.
    /// Unix only: Windows has no umask, and its loader starts threads in every process.
    ThreadsAlreadyRunning { threads: usize },
    /// The lock directory is on a filesystem whose locks may not hold across machines,
    /// accepted because `allow_network_lock(true)` was set.
    NetworkLockAllowed { filesystem: &'static str },
    /// `lock(false)`: no single-instance lock was taken.
    LockDisabled,
}
```

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| PRC-20 | `process::guard(name)` | 1 | 10 to 100 µs, once |
| PRC-21 | `Guard::builder` and options | 2 | same |
| PRC-22 | `Guard::report`, `GuardReport` | 1 | 0 |
| PRC-23 | lock directory hardening | 1 | one `stat` and one `statfs` |
| PRC-24 | `Guard::is_held` (0.7) | 1 | one `fstat` and one `lstat` |

The default lock directory depends only on the user the process runs as, never on environment variables (D-106), so a copy started by systemd or launchd and one started from a shell, `sudo`, or cron find the same lock:

| Platform | root | other users |
|---|---|---|
| Linux | `/run/<name>` (`/var/run/<name>` without `/run`) | `<home>/.local/state/<name>/run` |
| macOS | `/var/run/<name>` | `<home>/Library/Application Support/<name>/run` |
| FreeBSD | `/var/run/<name>` | `<home>/.local/state/<name>/run` |
| Windows | the user's local application data known folder, `<name>\run` | the same |

`<home>` is the account database's home directory (`getpwuid_r`), not `$HOME`; a user without one gets `RunDirUnavailable`, telling it to pass `lock_dir`. The lock is per user: copies running as different users do not exclude each other, so **a program guarding shared data passes its data directory** with `lock_dir`, and every copy that could touch the data finds the same lock. Before 0.7 the default followed `$XDG_RUNTIME_DIR`, `$XDG_STATE_HOME`, and `$HOME`, so two copies started with different environments could both run.

What the guard does, in order:

| Step | Linux and macOS | Windows |
|---|---|---|
| 1. open-file limit, checked | `OpenFilesTooLow` if the hard `RLIMIT_NOFILE` is below the minimum, before anything else, so a refusal changes nothing (a second copy started under a low limit gets this error, not `AlreadyRunning`) | not applicable |
| 2. lock directory | the default (above), or an explicit `lock_dir`, created with mode 0700 if missing; refused if it is a symlink, group- or world-writable, owned by another user, or on a network filesystem (`nfs`, `cifs`, `smb2`, `fuse.sshfs`, WSL's `drvfs` and `9p`) unless `allow_network_lock(true)` | refused if it is a reparse point or on a remote drive unless allowed; a missing directory inherits its parent's permissions, and ownership is not checked, so choose a parent other users cannot write to (not `C:\ProgramData` itself) |
| 3. single-instance lock | `<dir>/<name>.lock` opened with `O_NOFOLLOW`, then checked (the opened file must be the one inspected, by device and inode, and owned by this user), locked with `File::try_lock`, then checked again, so a file deleted or replaced between the open and the lock is never the one held (three tries); the OS releases the lock when the process dies, even after `kill -9` | the same file, `File::try_lock`, reparse points refused, opened without delete sharing, so the held file cannot be deleted or renamed |
| 4. pid file | `<dir>/<name>.pid` written after the lock is held, with `path_lib::write_atomic` | same (the locked file cannot be read by another process, so the pid lives beside it) |
| 5. open-file limit, raised | soft limit raised to the target, capped at the hard limit, on Linux at `/proc/sys/fs/nr_open`, and on macOS at `kern.maxfilesperproc` | not applicable (no per-process descriptor count for sockets); reported as unlimited |
| 6. umask | `umask(0o027)`: owner read and write, group read, others nothing | not applicable; `umask_after` is `None` |

The lock file is **never deleted**, not even on a clean exit: deleting it would let a waiting process lock the unlinked file while a third creates a new one, and two copies would run. On Unix something else can still delete it while it is held (a temporary-file cleaner, an operator's `rm`, the removal of a directory above it), and a copy started afterwards would lock a new file; a service that runs for long checks `guard.is_held()` periodically and stops when it is false. On Windows the held file cannot be deleted. The name must be 1 to 64 characters of ASCII letters, digits, `-`, `_`, `.`, not starting with `.`. The umask is process-wide, so call the guard before starting threads; on Unix, if other threads already run, the report says so. Privileges are not dropped here: that is shield-sec's job, after the listening sockets are bound.

Errors, each saying what to do:

| Error | When | The message tells the operator |
|---|---|---|
| `AlreadyRunning { pid, lock_file }` | another copy holds the lock | the lock file, and the pid when the pid file is readable: a hint to check, not proof (just after a copy starts the file can still hold the previous run's pid, which may since belong to another program) |
| `RunDirUnavailable { path, source }` | the directory cannot be created or written, is a symlink, is writable by others, or is on a network filesystem | the path, the reason, and the fix (`chmod 700`, a local directory, `allow_network_lock`) |
| `OpenFilesTooLow { have, need }` | the hard limit is below the minimum | what to raise: `LimitNOFILE=` in the systemd unit, `ulimit -Hn`, or launchd's limits |
| `InvalidInput` | the name failed validation | the rule |

```rust,no_run
use sys_lib as sys;

let guard = sys::process::guard("site")?;      // before any file is opened
println!("lock {:?}, open files now {:?}", guard.lock_file(), guard.report().open_files_after.soft);
// ... run; the lock is held while `guard` lives.
# Ok::<(), sys_lib::Error>(())
```

```rust,no_run
use sys_lib::process::Guard;

// A database locks beside its data, so two copies under different users or run
// directories still exclude each other.
let guard = Guard::builder("myapp")
    .lock_dir("/var/lib/myapp")
    .open_files_target(1_048_576)
    .open_files_minimum(65_536)
    .build()?;
# Ok::<(), sys_lib::Error>(())
```

```rust,no_run
use sys_lib::{process::Guard, Error};

match Guard::builder("site").build() {
    Ok(guard) => { /* run */ drop(guard); }
    Err(Error::AlreadyRunning { pid: Some(pid), .. }) => eprintln!("already running as pid {pid}"),
    Err(e) => eprintln!("cannot start: {e}"),
}
```

---

## `os`: operating system

### Kind and identity

**Since** 0.3 (OS-01 to OS-07), 0.5 (OS-08 to OS-10)

```rust
pub mod os {
    pub const fn kind() -> Kind;
    pub const fn is_windows() -> bool;
    pub const fn is_linux() -> bool;          // the Linux kernel, not Android
    pub const fn is_macos() -> bool;
    pub const fn is_mac() -> bool;            // the same as is_macos
    pub const fn is_android() -> bool;
    pub const fn is_ios() -> bool;
    pub const fn is_unix() -> bool;
    pub const fn is_bsd() -> bool;

    pub fn name() -> &'static str;                  // "Ubuntu 24.04.1 LTS", "macOS 15.1 Sequoia", "Windows 11 Pro"
    pub fn version() -> Option<&'static Version>;
    pub fn kernel() -> &'static str;                // "6.8.0-45-generic", "24.1.0", "10.0.26200.6899"
    pub fn distro() -> &'static str;                // "Ubuntu", "Fedora Linux", "Arch Linux"; "Windows", "macOS", ...
    pub fn distro_id() -> &'static str;             // "ubuntu", "fedora", "arch"; "windows", "macos", ...
    pub fn is_like(id: &str) -> bool;               // is_like("debian") on Ubuntu, Mint, Pop!_OS, Debian
    pub fn os_release() -> Option<&'static OsRelease>;   // every os-release field; Linux only
    pub fn edition() -> Option<&'static str>;       // "Professional", "ServerDatacenter"; Linux VARIANT_ID
    pub fn codename() -> Option<&'static str>;      // "Sequoia", "noble"
    pub fn is_server_edition() -> Option<bool>;     // Windows Server; None where it has no meaning
    pub fn is_64bit() -> bool;                      // the OS, not the process (WOW64, ARM64 emulation)
    pub fn info() -> &'static OsInfo;
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Kind { Linux, Windows, MacOs, Android, Ios, FreeBsd, OpenBsd, NetBsd, Other }
impl Kind { pub const fn name(self) -> &'static str; }   // "linux", "windows", "macos", ...

#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
#[non_exhaustive]
pub struct Version {
    pub major: u32, pub minor: u32, pub patch: u32, pub build: Option<u32>,
    pub text: &'static str,      // the original text: "24.04" stays "24.04"
}
impl fmt::Display for Version;   // prints `text`

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct OsRelease {
    pub id: &'static str, pub id_like: &'static [&'static str], pub name: &'static str,
    pub pretty_name: &'static str, pub version: Option<&'static str>,
    pub version_id: Option<&'static str>, pub codename: Option<&'static str>,
    pub variant_id: Option<&'static str>, pub build_id: Option<&'static str>,
    pub source: Source,          // which file the fields came from
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct OsInfo {
    pub kind: Kind, pub name: &'static str, pub version: Option<&'static Version>,
    pub kernel: &'static str, pub distro: &'static str, pub distro_id: &'static str,
    pub os_release: Option<&'static OsRelease>,
    pub edition: Option<&'static str>, pub codename: Option<&'static str>,   // since 0.5 (OS-08, OS-09)
    pub is_64bit: bool,                                                      // since 0.5 (OS-10)
    pub name_source: Source,     // where `name` came from
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Source { OsRelease, UsrLibOsRelease, LegacyRelease, Sysctl, Registry, SystemProperty, Fallback }
```

| Id | Item | Since | Cost | Android / iOS |
|---|---|---|---|---|
| OS-01 | `kind()` | 0.3 | 0 (const) | yes / yes |
| OS-02 | `is_*` constants | 0.3 | 0 (const) | yes / yes |
| OS-03 | `name()` | 0.3 | about 1 ns; first call 10 to 50 µs [est.] | yes / yes |
| OS-04 | `version()` | 0.3 | about 1 ns | yes / yes |
| OS-05 | `kernel()` | 0.3 | about 1 ns | yes / yes |
| OS-06 | `distro()`, `distro_id()`, `is_like()`, `os_release()` | 0.3 | about 1 ns (`is_like` a few ns) | `"Android"` / `"iOS"`; `os_release()` `None` |
| OS-07 | `info()` | 0.3 | about 1 ns | yes / yes |
| OS-08 | `edition()`, `is_server_edition()` | 0.5 | about 1 ns | `None` |
| OS-09 | `codename()` | 0.5 | about 1 ns | `None` |
| OS-10 | `is_64bit()` | 0.5 | about 1 ns | yes |

**The distribution, on every OS.** `distro()` names the distribution and `distro_id()` gives its lowercase id. On Linux they are the os-release `NAME` and `ID` ("Ubuntu" and `ubuntu`, "Fedora Linux" and `fedora`, "Arch Linux" and `arch`); everywhere else they name the operating system: "Windows" and `windows`, "macOS" and `macos`, "Android" and `android`, "iOS" and `ios`, "FreeBSD" and `freebsd`. They never fail and never return an empty string.

On Linux the answer is the os-release file the system itself ships (os-release(5): `/etc/os-release`, then `/usr/lib/os-release`), parsed with its quoting and escape rules. It is what systemd, package managers, and installers read, so sys-lib agrees with them. A system without os-release falls back to the legacy release files, in this order: `/etc/lsb-release`, `/etc/redhat-release` (and the CentOS, Fedora, Rocky, and AlmaLinux variants), `/etc/SuSE-release`, `/etc/alpine-release`, `/etc/arch-release`, `/etc/gentoo-release`, `/etc/slackware-version`, `/etc/debian_version`; `os_release().source` is then `LegacyRelease`. With none of them (a `scratch` or distroless container), the answer is "Linux" and `linux`: an honest "not known", not a guess. Ids are checked against the os-release rule (lowercase letters, digits, `.`, `_`, `-`); anything else falls through to the next source.

`is_like(id)` is true when the id is the distribution's own `ID` or appears in its `ID_LIKE`: `is_like("debian")` is true on Debian, Ubuntu, Mint, and Pop!_OS, and `is_like("rhel")` on RHEL, Rocky, AlmaLinux, and CentOS. Use it to pick a package manager or a default path by family instead of listing every distribution.

Inside a container the answer is the container image's distribution, which is the userland the program runs on; the host's distribution is not visible and is not guessed. Inside WSL it is the WSL distribution.

Accuracy is tested against a corpus of real os-release and legacy release files from at least 40 distributions and versions (Ubuntu, Debian, Fedora, RHEL, Rocky, AlmaLinux, CentOS Stream, Amazon Linux, Oracle Linux, SUSE, openSUSE Leap and Tumbleweed, Arch, Manjaro, EndeavourOS, Alpine, Gentoo, NixOS, Void, Linux Mint, Pop!_OS, elementary OS, Kali, Raspberry Pi OS, Clear Linux, Photon, Wolfi, Flatcar, Bottlerocket, Azure Linux, Talos, SteamOS, and more), each with its expected `distro()`, `distro_id()`, `version()`, and `is_like()` answers.

`name()` never fails: a minimal container without os-release (`scratch`, distroless) gives `"Linux"`, and `info().name_source` is `Fallback`. On Windows the product comes from the build number, because the registry's `ProductName` still says "Windows 10" on Windows 11; the version comes from `RtlGetVersion`, which reports the real version whatever the executable's manifest declares. On macOS the version comes from `kern.osproductversion`, unaffected by `SYSTEM_VERSION_COMPAT`; a future major without a known codename prints without one. `kind()` maps Android to `Android` (path-lib's `Os` maps it to `OtherUnix`).

```rust
use sys_lib as sys;

println!("{} (kernel {})", sys::os::name(), sys::os::kernel());
if let Some(v) = sys::os::version() {
    if v.major >= 24 { println!("recent enough: {v}"); }
}
println!("{} ({})", sys::os::distro(), sys::os::distro_id());   // "Ubuntu (ubuntu)", "Windows (windows)"
if sys::os::is_like("debian") {
    // apt-based: Debian, Ubuntu, Mint, Pop!_OS, ...
}
if let Some(r) = sys::os::os_release() {
    println!("{} {:?} {:?}", r.pretty_name, r.version_id, r.codename);   // "Ubuntu 24.04.1 LTS" Some("24.04") Some("noble")
}
const ON_WINDOWS: bool = sys::os::is_windows();   // usable in const context
```

### Form factor

**Since** 0.5 · OS-11 to OS-13

```rust
pub mod os {
    pub fn form_factor() -> FormFactor;
    pub fn detect_form_factor() -> &'static Detection<FormFactor>;
    pub fn is_desktop() -> bool;       // Desktop
    pub fn is_laptop() -> bool;        // Laptop or Convertible
    pub fn is_mobile() -> bool;        // Phone, Tablet, Handheld, Wearable
    pub fn is_phone() -> bool;
    pub fn is_tablet() -> bool;
    pub fn is_server() -> bool;        // the chassis or the platform role says server
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum FormFactor {
    Desktop, Laptop, Convertible, Tablet, Phone, Handheld,
    Server, Embedded, Wearable, Tv, Unknown,
}
impl FormFactor { pub const fn name(self) -> &'static str; }   // "laptop"

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct Detection<T> { pub value: T, pub source: &'static str, pub confidence: Confidence }

#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
#[non_exhaustive]
pub enum Confidence { None, Low, Medium, High }
```

What kind of machine this is, from the firmware chassis type, the ACPI power-management profile, the platform role, the model identifier, or a battery, first answer wins (the order per OS is in [`SOURCES.md`](./SOURCES.md)). On Linux, `CHASSIS=` in `/etc/machine-info` overrides the firmware, as systemd honours it. Inside a VM the chassis describes virtual hardware, so the answer is `Unknown` unless an administrator said otherwise: an honest unknown, not a guess. Cost: about 1 ns after the first call; the first call 5 to 500 µs [est.] (an SMBIOS read on Windows).

Detection is advisory. Do not make security decisions on it.

```rust
use sys_lib as sys;

if sys::os::is_laptop() && sys::power::on_ac()? == Some(false) {
    // on battery: back off background work
}
let d = sys::os::detect_form_factor();
println!("{} (from {}, confidence {:?})", d.value.name(), d.source, d.confidence);
# Ok::<(), sys_lib::Error>(())
```

---

## `cpu`: processor

### Identity and counts

**Since** 0.3 (CPU-01 to CPU-06), 0.4 (CPU-07 to CPU-12), 0.5 (CPU-13)

```rust
pub mod cpu {
    pub const fn arch() -> Arch;                    // the architecture this program was compiled for
    pub fn is_64bit() -> bool;                      // the processor is 64-bit
    pub fn is_32bit() -> bool;                      // the processor is 32-bit only
    pub fn native_arch() -> Arch;                   // the hardware (under Rosetta 2, WOW64, ARM64 emulation)
    pub fn is_translated() -> bool;                 // this process runs under emulation

    pub fn name() -> &'static str;                  // cleaned: "AMD Ryzen 9 7950X", "Apple M2 Pro"
    pub fn raw_name() -> Option<&'static str>;      // as reported, for support output
    pub fn vendor() -> Vendor;

    pub fn logical() -> usize;                      // configured hardware threads (static)
    pub fn physical() -> usize;                     // cores (static); logical() if unknown
    pub fn packages() -> usize;                     // sockets (static)
    pub fn available() -> usize;                    // what this process may use now
    pub fn online() -> usize;                       // online now (hot-plug)
    pub fn count() -> Counts;
    pub fn quota() -> Result<Option<CpuQuota>>;     // cgroup cpu.max / job object rate

    pub fn frequency() -> &'static Frequency;        // base and maximum (static)
    pub fn current_frequency() -> Result<Option<Mhz>>;
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Arch { X86_64, X86, Aarch64, Arm, Riscv64, Powerpc64, S390x, Loongarch64, Wasm32, Other }
impl Arch {
    pub const fn name(self) -> &'static str;     // "x86_64", "aarch64"
    pub const fn is_64bit(self) -> bool;
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug, Default)]
#[non_exhaustive]
pub enum Vendor {
    Intel, Amd, Apple, Arm, Qualcomm, Ampere, Nvidia, Samsung, Mediatek, Hygon, Zhaoxin,
    Broadcom, Marvell, Fujitsu, HiSilicon, Microsoft,   // ARM implementers in servers and boards (D-033)
    #[default] Unknown,
}
impl Vendor { pub const fn name(self) -> &'static str; }   // "Intel", "AMD", "ARM", "Unknown"

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Counts {
    pub logical: usize, pub physical: usize, pub packages: usize,
    pub available: usize, pub online: usize,
    pub physical_exact: bool,     // false when the topology was hidden and physical fell back to logical
}

#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct CpuQuota { pub cpus: f32, pub source: LimitSource }   // 1.5 = one and a half CPUs

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Frequency { pub base: Option<Mhz>, pub max: Option<Mhz> }
```

| Id | Item | Since | Cost |
|---|---|---|---|
| CPU-01 | `arch()` (const), `is_64bit()`, `is_32bit()` | 0.3 | 0; `is_64bit` about 1 ns (one check on 32-bit Windows, cached) |
| CPU-02 | `logical()` | 0.3 | about 1 ns |
| CPU-03 | `physical()` | 0.3 | about 1 ns; first call about 200 µs on 32 CPUs (Linux) [est.] |
| CPU-04 | `available()` | 0.3 | Linux 3 to 10 µs, others 0.2 µs [est.] |
| CPU-05 | `name()`, `raw_name()` | 0.3 | about 1 ns |
| CPU-06 | `vendor()` | 0.3 | about 1 ns |
| CPU-07 | `packages()` | 0.4 | about 1 ns |
| CPU-08 | `count()` | 0.4 | `available` and `online` are live |
| CPU-09 | `online()` | 0.4 | 2 µs / 0.2 µs [est.] |
| CPU-10 | `quota()` | 0.4 | 5 to 15 µs [est.] |
| CPU-11 | `frequency()` | 0.4 | about 1 ns |
| CPU-12 | `current_frequency()` | 0.4 | 2 µs per CPU (Linux) / 20 µs (Windows) [est.] |
| CPU-13 | `native_arch()`, `is_translated()` | 0.5 | about 1 ns |

**Three counts, three meanings.** `physical()` is cores; `logical()` is hardware threads, the right size for per-CPU arrays (it counts **present** CPUs, not the `possible` count VMs inflate); `available()` is what **this process** may use right now: the affinity mask, intersected with the cpuset, then limited by the cgroup CPU quota (`ceil(quota / period)`, at least 1), or on Windows the group affinity across every processor group and the job object's CPU rate. Size worker pools by `available()`. On Windows with more than 64 logical processors, every count spans all processor groups. `physical <= logical`, `available <= logical`, `available >= 1`, and `online >= 1` always hold.

**The name is cleaned.** "(R)", "(TM)", trailing frequencies (" CPU @ 3.00GHz"), "CPU" as a word, core-count suffixes ("16-Core Processor"), and repeated spaces are removed; `raw_name()` keeps the original. ARM big.LITTLE systems list their core types largest first ("Cortex-X1 + Cortex-A78 + Cortex-A55"). The name is never empty: it falls back to the vendor and architecture ("ARM aarch64"), then "unknown".

**64-bit or 32-bit.** `is_64bit()` and `is_32bit()` answer for the **processor**, so a 32-bit program on a 64-bit machine still sees `is_64bit() == true` (on Windows through `IsWow64Process2`; a 64-bit build implies a 64-bit processor everywhere). For the program itself, `cpu::arch().is_64bit()` is a compile-time constant. For the operating system, `os::is_64bit()`.

There is no "total frequency": a sum of per-core frequencies is not a meaningful quantity. `frequency()` gives base and maximum; Apple silicon reports neither (`None`).

```rust
use sys_lib as sys;

println!("{} ({:?}, {})", sys::cpu::name(), sys::cpu::vendor(), sys::cpu::arch().name());
println!("{} cores, {} threads, {} usable here", sys::cpu::physical(), sys::cpu::logical(), sys::cpu::available());

// Size a worker pool by what the container allows, not by the host.
let workers = sys::cpu::available();
if let Some(q) = sys::cpu::quota()? {
    println!("CPU quota {:.1} from {:?}", q.cpus, q.source);
}
let f = sys::cpu::frequency();
if let Some(max) = f.max { println!("up to {max}"); }
# Ok::<(), sys_lib::Error>(())
```

### Usage

**Since** 0.4 · CPU-20 to CPU-25

```rust
pub mod cpu {
    pub const DEFAULT_WINDOW: Duration;                     // 100 ms
    pub fn usage() -> Result<Usage>;                        // BLOCKS for DEFAULT_WINDOW
    pub fn usage_over(window: Duration) -> Result<Usage>;   // BLOCKS for `window`, clamped to 10 ms ..= 10 s
    pub fn used() -> Result<Percent>;                       // BLOCKS: usage()?.used()
    pub fn free() -> Result<Percent>;                       // BLOCKS: usage()?.free()
    pub fn times() -> Result<Times>;                        // Tier 3: raw cumulative counters

    pub struct Sampler { /* previous Times and Instant */ }
    impl Sampler {
        pub fn new() -> Result<Sampler>;                                              // the first reading
        pub fn sample(&mut self) -> Result<Usage>;                                    // never blocks; no allocation
        pub fn sample_per_core(&mut self) -> Result<(Usage, Vec<Percent>)>;           // allocates the Vec
        pub fn sample_per_core_into(&mut self, out: &mut [Percent]) -> Result<Usage>; // Tier 3; no allocation
    }
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct Usage {
    pub user: Percent, pub system: Percent, pub idle: Percent,
    pub iowait: Option<Percent>,   // Linux
    pub steal: Option<Percent>,    // Linux VMs: time the hypervisor gave to others
    pub irq: Option<Percent>,
    pub interval: Duration,
}
impl Usage { pub fn used(&self) -> Percent; pub fn free(&self) -> Percent; }

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Times {
    pub user: Duration, pub nice: Duration, pub system: Duration, pub idle: Duration,
    pub iowait: Duration, pub irq: Duration, pub softirq: Duration, pub steal: Duration,
}
```

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| CPU-20 | `usage()`, `used()`, `free()` | 1 | 100 ms of wall time; two readings |
| CPU-21 | `usage_over(window)` | 1 | `window` of wall time |
| CPU-22 | `Sampler::new`, `sample` | 2 | Linux 2 to 4 µs (the first line of `/proc/stat`), macOS 1 to 2 µs, Windows 0.3 µs |
| CPU-23 | `sample_per_core`, `sample_per_core_into` | 2 / 3 | Linux 10 to 60 µs (scales with CPUs), macOS 5 µs, Windows 5 to 20 µs |
| CPU-24 | `Usage` fields | 2 | 0 |
| CPU-25 | `times()` | 3 | one reading |

CPU usage is the share of non-idle time **between two readings** of the OS's cumulative counters; one reading only gives the average since boot. `usage()` takes both readings itself and blocks for 100 ms, which its name and documentation say. A service that reports usage periodically keeps a `Sampler` instead: it never sleeps, each `sample()` measures the time since the previous one, and two samplers never disturb each other (there is no hidden global baseline).

- Steal counts as busy (the CPU was not available to us) and iowait as idle (the CPU was free to run other work); both are also reported separately.
- `TooSoon` when no tick passed on the OS clock since the previous sample (about 10 ms on Linux): never a meaningless 0% or 100%.
- Counters that go backwards (a CPU taken offline and back, a VM migration, a wrap) re-base the sampler: `TooSoon` for the total, 0% for that interval for a core (a slice carries no error per core), never a negative or above-100 value. On Linux per-core values are indexed by CPU number: an offline CPU keeps its slot and reads 0%, and `out` must cover the highest CPU number plus one.
- `sample_per_core_into` returns `TooLarge` without writing anything when `out` is shorter than `cpu::logical()`.
- Windows with more than 64 logical processors, or more than one processor group: the total is correct (system-wide counters); per-core values return `Unsupported` until the per-group call is verified (D-047), and `Sampler::new()` then samples the total only.

```rust
use sys_lib as sys;

println!("CPU {}", sys::cpu::used()?);   // blocks 100 ms: "12.4%"
# Ok::<(), sys_lib::Error>(())
```

```rust,no_run
use sys_lib as sys;
use std::time::Duration;

// A monitor loop: non-blocking, caller-owned.
let mut cpu = sys::cpu::Sampler::new()?;
let mut per_core = vec![sys::Percent::ZERO; sys::cpu::logical()];
loop {
    std::thread::sleep(Duration::from_secs(5));
    let u = cpu.sample_per_core_into(&mut per_core)?;
    println!("used {} (user {}, system {}, steal {:?})", u.used(), u.user, u.system, u.steal);
}
# Ok::<(), sys_lib::Error>(())
```

### Load

**Since** 0.4 · CPU-14

```rust
pub mod cpu {
    pub fn load() -> Result<Option<Load>>;     // Ok(None) on Windows, which keeps no load average
}

#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct Load { pub one: f32, pub five: f32, pub fifteen: f32 }
impl Load {
    pub fn per_cpu(&self) -> Load;   // divided by cpu::logical(): 1.0 means every CPU had work
}
impl fmt::Display for Load;          // "2.15 1.98 1.70"
```

**What load is.** The load average is how many programs, on average, were running or waiting for a CPU over the last 1, 5, and 15 minutes. Compare it with the number of CPUs: on a 16-CPU machine, a load of 4 means a quarter of the machine had work, 16 means every CPU was busy, and 32 means as many programs were waiting as running, so things were slowing down. `per_cpu()` does that division for you: under 1.0 there was spare capacity, over 1.0 there was a queue. The three numbers together show the trend: a 1-minute value well above the 15-minute value means load is rising.

Load differs from `cpu::used()`. Usage says how busy the CPUs were; load says how many programs wanted them, so it shows a queue that a 100% usage figure cannot. Benchmarks record it to show the machine was quiet before a run. On Linux it also counts programs waiting on disk (uninterruptible I/O), so a high load with low CPU usage points at storage.

| OS | Source | Cost [est.] |
|---|---|---|
| Linux | `/proc/loadavg` | 2 µs |
| macOS | `getloadavg` | 0.3 µs |
| Windows | none: Windows keeps no load average, so the answer is `Ok(None)`; nothing is simulated | 0 |

Inside a container the load is the host's (the kernel keeps one per machine); with LXCFS it is the container's.

```rust
use sys_lib as sys;

match sys::cpu::load()? {
    Some(load) => {
        println!("load {load} ({:.2} per CPU)", load.per_cpu().one);   // "load 2.15 1.98 1.70 (0.13 per CPU)"
        if load.per_cpu().one > 1.0 { /* programs are queueing for a CPU */ }
    }
    None => println!("this OS keeps no load average"),
}
# Ok::<(), sys_lib::Error>(())
```

### CPU pressure

**Since** 0.4 · CPU-15

```rust
pub mod cpu {
    pub fn pressure() -> Result<Option<Pressure>>;   // Ok(None) where the OS has no stall measure
}
```

How much runnable work waited for a CPU over the last 10, 60, and 300 seconds ([`Pressure`](#pressure)). It is the direct form of what the load average approximates: load counts programs, pressure measures the time they lost. Linux: the cgroup's `cpu.pressure`, or `/proc/pressure/cpu`. macOS and Windows have no stall measure: `Ok(None)`. Cost: 2 to 5 µs [est.].

### Features

**Since** 0.4 (CPU-30 to CPU-32; CPU-32 `x86_level()` was planned for 0.5 and ships with the features it reads, D-052)

```rust
pub mod cpu {
    pub fn has(feature: Feature) -> bool;          // about 1 ns
    pub fn features() -> &'static Features;
    pub fn x86_level() -> Option<X86Level>;        // x86-64-v1 ..= v4; None off x86-64
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Feature {
    // x86
    Sse2, Sse3, Ssse3, Sse41, Sse42, Popcnt, Avx, Avx2, Fma, F16c, Bmi1, Bmi2, Lzcnt, Movbe,
    Avx512f, Avx512bw, Avx512cd, Avx512dq, Avx512vl, Avx512vnni, Avx512vbmi, AvxVnni,
    Aes, Pclmulqdq, Vaes, Vpclmulqdq, Sha, Rdrand, Rdseed, Adx, Gfni,
    // AArch64
    Neon, Crc32, AesArm, Pmull, Sha2, Sha3, Dotprod, Fp16, Lse, Sve, Sve2, Rcpc, Bf16, I8mm,
}
impl Feature { pub const fn name(self) -> &'static str; }   // rustc's target-feature name: "avx2", "neon"

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug, Default)]
pub struct Features { /* bitset */ }
impl Features {
    pub fn contains(&self, feature: Feature) -> bool;
    pub fn iter(&self) -> impl Iterator<Item = Feature> + '_;
    pub fn len(&self) -> usize;
    pub fn is_empty(&self) -> bool;
}
impl fmt::Display for Features;   // "sse2 sse3 ... avx2 aes": for a log line

#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
#[non_exhaustive]
pub enum X86Level { V1, V2, V3, V4 }
```

Detection goes through std (`is_x86_feature_detected!`, `is_aarch64_feature_detected!`), which checks operating-system support as well as the CPU: a CPU with AVX-512 on an OS that disabled its state reports `false`. That is the same detection the compiler's `#[target_feature]` dispatch trusts, so the two always agree. A feature std cannot detect at the MSRV is left out rather than detected differently; `Sha3` on AArch64 is std's `sha3`, which covers both SHA-3 and SHA-512. Under Rosetta 2, the x86 features are what Rosetta emulates (no AVX).

```rust
use sys_lib::{self as sys, cpu::Feature};

if sys::cpu::has(Feature::Avx2) {
    // pick the AVX2 kernel
}
println!("features: {}", sys::cpu::features());
if let Some(level) = sys::cpu::x86_level() { println!("x86-64-{level:?}"); }
```

### Topology, caches, NUMA

**Since** 0.5 · CPU-40 to CPU-44

```rust
pub mod cpu {
    pub fn topology() -> &'static Topology;
    pub fn caches() -> &'static [Cache];
    pub fn cache_line() -> usize;                 // bytes; 64 when unknown on x86, 128 on Apple silicon
    pub fn numa_nodes() -> &'static [NumaNode];
    pub fn is_hybrid() -> bool;                   // more than one core kind
}

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Topology { pub packages: Vec<Package>, pub cores: Vec<Core> }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Package { pub id: u32, pub cores: Vec<u32> }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Core {
    pub id: u32, pub package: u32,
    pub logical: Vec<u32>,              // logical CPU indexes
    pub kind: CoreKind,
    pub capacity: Option<u16>,          // Linux cpu_capacity, 0..=1024
    pub numa_node: Option<u32>,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum CoreKind { Performance, Efficiency, Uniform, Unknown }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Cache {
    pub level: u8, pub kind: CacheKind, pub size: Bytes, pub line: u32,
    pub shared_by: Vec<u32>,            // logical CPUs sharing this cache instance
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum CacheKind { Data, Instruction, Unified }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct NumaNode { pub id: u32, pub cpus: Vec<u32>, pub memory: Option<Bytes> }
```

Static facts read on first use: about 1 ns afterwards; the first call reads up to a few hundred small files on large Linux hosts (1 to 5 ms on a 256-CPU machine [est.]), which is why they are separate from `cpu::name()`. CPU hot-plug after the first call is not reflected; `cpu::online()` is the live count. Lists are sorted by id. On Linux a core's id is its lowest logical CPU. macOS reports one NUMA node with every CPU, and its logical CPU numbers are positions, efficiency cores first (the OS offers no CPU-to-core mapping). Windows gives a NUMA node's memory on a one-node machine and `None` on several: it offers free memory per node, not attached memory.

```rust
use sys_lib as sys;

let t = sys::cpu::topology();
let fast = t.cores.iter().filter(|c| c.kind == sys::cpu::CoreKind::Performance).count();
println!("{} cores, {fast} performance cores, hybrid: {}", t.cores.len(), sys::cpu::is_hybrid());
for c in sys::cpu::caches() {
    println!("L{} {:?} {} shared by {}", c.level, c.kind, c.size, c.shared_by.len());
}
println!("cache line {} bytes, {} NUMA nodes", sys::cpu::cache_line(), sys::cpu::numa_nodes().len());
```

---

## `mem`: memory

**Since** 0.4 (MEM-01 to MEM-11, MEM-13), 0.5 (MEM-12)

```rust
pub mod mem {
    pub fn total() -> Result<Bytes>;              // effective: min(host, limit)
    pub fn used() -> Result<Bytes>;
    pub fn free() -> Result<Bytes>;               // available to new allocations
    pub fn used_percent() -> Result<Percent>;
    pub fn free_percent() -> Result<Percent>;
    pub fn snapshot() -> Result<Memory>;          // every value from one set of reads
    pub fn host_total() -> Result<Bytes>;         // ignores limits
    pub fn host_free() -> Result<Bytes>;
    pub fn limit() -> Result<Option<MemoryLimit>>;
    pub fn swap() -> Result<Swap>;
    pub fn page_size() -> usize;                  // static
    pub fn pressure() -> Result<Option<Pressure>>;
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct Memory {
    pub total: Bytes, pub used: Bytes, pub free: Bytes,
    pub host_total: Bytes, pub host_free: Bytes,
    pub unused: Option<Bytes>,       // pages doing nothing (the kernel's MemFree)
    pub cached: Option<Bytes>,       // Windows: None (D-102)
    pub limit: Option<MemoryLimit>, pub swap: Swap,
    pub source: MemorySource,
}
impl Memory { pub fn used_percent(&self) -> Percent; pub fn free_percent(&self) -> Percent; }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum MemorySource { Host, Cgroup1, Cgroup2, JobObject, Lxcfs }

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct MemoryLimit { pub hard: Bytes, pub soft: Option<Bytes>, pub swap: Option<Bytes>, pub source: LimitSource }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum LimitSource { Cgroup1, Cgroup2, JobObject }

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct Swap { pub total: Bytes, pub used: Bytes, pub free: Bytes }

```

| Id | Item | Since | Cost per call [est.] |
|---|---|---|---|
| MEM-01 | `total()` | 0.4 | Linux 2 to 4 µs bare, 6 to 15 µs in a cgroup; macOS 0.5 µs; Windows about 5 µs (measured, with the job-object query) |
| MEM-02 | `used()`, `free()` | 0.4 | same |
| MEM-03 | `used_percent()`, `free_percent()` | 0.4 | same |
| MEM-04 | `snapshot()` | 0.4 | the same reads, every value |
| MEM-05 | `host_total()`, `host_free()` | 0.4 | Linux 2 to 4 µs |
| MEM-06 | `limit()` | 0.4 | Linux 4 to 10 µs; Windows 1 µs; macOS 0 (`None`) |
| MEM-07 | `swap()` | 0.4 | Linux 2 to 4 µs; macOS 0.5 µs; Windows 2 µs |
| MEM-08 | `page_size()` | 0.4 | about 1 ns |
| MEM-09 | `pressure()` | 0.4 | Linux 2 µs (`None` on kernels without PSI); macOS 0.5 µs; Windows 1 µs |

**Container-aware by default.** Inside a container or a systemd unit with a memory limit, `total()` is the limit: what this process can use and what the OOM killer enforces. sysinfo and psutil report the host instead; sys-lib gives the host through `host_total()`. The limit is the smallest `memory.max` over the cgroup and its ancestors (cgroup v2), `hierarchical_memory_limit` (cgroup v1), or the job object's process or job memory limit (Windows). "Used" inside a limit is the cgroup's working set: current usage minus inactive file pages, the definition the kubelet uses for eviction. With LXCFS the limit is not applied twice. The limit can change while a process runs (Kubernetes in-place resize, VM ballooning), so `total()` is a live reading.

**"free" means available**: what new allocations can get without swapping, counting reclaimable cache (`MemAvailable` on Linux). The kernel's idle pages (`MemFree`) are `Memory::unused`. On every platform `used + free == total`, `used <= total`, and percentages are within 0 to 100.

Each Tier 1 function performs only the reads it needs. A caller who wants several values calls `snapshot()` and pays for one set of reads. Every scalar reading allocates nothing.

```rust
use sys_lib as sys;

println!("{} of {} used ({})", sys::mem::used()?, sys::mem::total()?, sys::mem::used_percent()?);

let m = sys::mem::snapshot()?;
println!("{} free of {} (host {} of {}), source {:?}", m.free, m.total, m.host_free, m.host_total, m.source);
if let Some(limit) = m.limit {
    println!("limit {} from {:?}, throttles at {:?}", limit.hard, limit.source, limit.soft);
}
# Ok::<(), sys_lib::Error>(())
```

```rust
use sys_lib as sys;

// Size a cache to a quarter of what this process may use.
let cache_bytes = sys::mem::total()?.raw() / 4;
// Shed load before the OOM killer does.
if sys::mem::free_percent()?.value() < 5.0 { /* refuse new work */ }
if let Some(p) = sys::mem::pressure()? {
    if p.level >= sys::PressureLevel::Warning { /* trim caches */ }
}
# Ok::<(), sys_lib::Error>(())
```

`mem::pressure()` returns the shared [`Pressure`](#pressure) type. On macOS it maps `kern.memorystatus_vm_pressure_level` (normal, warn, critical) to `level` with no stall percentages; on Windows a low-memory resource notification gives `Critical`.

### Pressure and out-of-memory events

**Since** 0.4 · MEM-13

```rust
pub mod mem {
    pub fn events() -> Result<Option<MemoryEvents>>;   // counts since the cgroup (or the boot) began
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct MemoryEvents {
    pub oom_kills: u64,        // processes the OOM killer ended
    pub oom: u64,              // times the limit was reached and reclaim failed
    pub max: u64,              // times usage hit the hard limit
    pub high: u64,             // times usage went over the throttle point (memory.high)
    pub scope: EventScope,     // Cgroup or Host
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum EventScope { Cgroup, #[default] Host }   // derives Default
```

Whether this container (or machine) has been running out of memory: a watchdog compares two readings and sheds load or restarts before the next OOM kill. Linux: cgroup v2 `memory.events`, cgroup v1 `memory.oom_control` (`oom_kill`), and outside a cgroup `/proc/vmstat` `oom_kill` (kernel 4.13 and later; `max` and `high` are then 0). macOS and Windows keep no such counters: `Ok(None)`. Cost: 3 µs [est.].

### Swap and paging activity

**Since** 0.4 · MEM-11

```rust
pub mod mem {
    pub fn activity() -> Result<Activity>;                       // BLOCKS for 1 s
    pub struct Sampler { /* previous counters and Instant */ }
    impl Sampler {
        pub fn new() -> Result<Sampler>;
        pub fn sample(&mut self) -> Result<Activity>;            // never blocks; no allocation
    }
}

#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct Activity {
    pub swap_in: Rate, pub swap_out: Rate,        // memory moved between RAM and swap
    pub major_faults_per_sec: Option<f64>,        // page faults that had to read from disk
    pub interval: Duration,
}
```

Whether the machine is thrashing: steady swap-out with major faults means memory is too small for the work, which no memory total can show. Linux: `/proc/vmstat` (`pswpin`, `pswpout`, `pgmajfault`, in pages). macOS: `vm_statistics64` (`swapins`, `swapouts`, `pageins`). Windows: `Unsupported` (the `SYSTEM_PERFORMANCE_INFORMATION` layout cannot be checked against the headers, D-049). Cost: 2 to 5 µs per sample [est.].

### Huge pages

**Since** 0.4 · MEM-10

```rust
pub mod mem {
    pub fn huge_pages() -> Result<Option<HugePages>>;
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct HugePages {
    pub page_size: Bytes,                    // 2 MiB on x86_64
    pub total: u64, pub free: u64, pub reserved: u64,
    pub transparent: Option<TransparentHugePages>,   // Linux THP mode
    pub large_page_minimum: Option<Bytes>,   // Windows GetLargePageMinimum
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum TransparentHugePages { Always, Madvise, Never }
```

Explicit huge pages and the transparent-huge-page mode, which databases and JIT runtimes care about. Linux: `HugePages_*` and `Hugepagesize` in `/proc/meminfo`, `/sys/kernel/mm/transparent_hugepage/enabled`. Windows: `GetLargePageMinimum`. macOS: `Ok(None)` (superpages are not exposed this way).

### Memory modules

**Since** 0.5 · MEM-12

```rust
pub mod mem {
    pub fn modules() -> Result<Vec<MemoryModule>>;
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct MemoryModule {
    pub size: Bytes,
    pub kind: MemoryKind,                 // Ddr4, Ddr5, Lpddr5, ...
    pub form: MemoryForm,                 // Dimm, SoDimm, Soldered, ...
    pub speed_mts: Option<u32>,           // rated, mega-transfers per second
    pub configured_speed_mts: Option<u32>,
    pub locator: Option<String>,          // "DIMM_A1"
    pub manufacturer: Option<String>,
    pub part_number: Option<String>,      // never the serial number
}
#[non_exhaustive] pub enum MemoryKind { Ddr3, Ddr4, Ddr5, Lpddr4, Lpddr5, Lpddr5x, Hbm, Other, Unknown }
#[non_exhaustive] pub enum MemoryForm { Dimm, SoDimm, Soldered, Other, Unknown }
```

The installed memory sticks, from the SMBIOS type 17 records (the same bounded, fuzzed SMBIOS parser device identity uses). Windows reads them unprivileged (`GetSystemFirmwareTable('RSMB')`). Linux keeps the raw table root-only (`/sys/firmware/dmi/tables/DMI`, mode 0400): `PermissionDenied` with a hint unless the process runs as root. macOS: `Unsupported` (Apple silicon has unified memory, one package; no documented source was confirmed for Intel Macs, D-059). Empty slots are left out. Serial numbers are never read into the result.

---

## `disk`: disks and filesystems

**Since** 0.4 (DSK-01 to DSK-03, DSK-06, DSK-07), 0.5 (DSK-04, DSK-05)

```rust
pub mod disk {
    pub fn space(path: impl AsRef<Path>) -> Result<Space>;       // the filesystem holding `path`
    pub fn list() -> Result<Vec<Disk>>;                           // local, real filesystems
    pub fn list_with(options: ListOptions) -> Result<Vec<Disk>>;
    pub fn of(path: impl AsRef<Path>) -> Result<Disk>;            // the mount holding `path`
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Space {
    pub total: Bytes,
    pub free: Bytes,        // every free block
    pub available: Bytes,   // what an unprivileged user can use (excludes the root reserve)
}
impl Space { pub fn used(&self) -> Bytes; pub fn used_percent(&self) -> Percent; }

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Disk {
    pub mount_point: PathBuf,
    pub device: String,          // "/dev/nvme0n1p2", "C:", "\\?\Volume{...}\"
    pub label: Option<String>,   // volume label (Windows, macOS)
    pub filesystem: String,      // "ext4", "btrfs", "apfs", "NTFS", "overlay"
    pub kind: DiskKind,
    pub space: Option<Space>,    // None for media not present (Windows lists it with filesystem "unknown")
    pub read_only: bool,
    pub removable: bool,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum DiskKind { Ssd, Hdd, Removable, Network, Optical, Ram, Virtual, Unknown }

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct ListOptions {
    pub include_network: bool,      // default false: statvfs on a dead NFS mount can hang
    pub include_pseudo: bool,       // proc, sysfs, cgroup, tmpfs, devtmpfs, squashfs; default false
    pub include_duplicates: bool,   // bind mounts and subvolumes of one filesystem; default false
    pub max_mounts: usize,          // default 4_096
}
impl Default for ListOptions;
```

| Id | Item | Since | Cost [est.] |
|---|---|---|---|
| DSK-01 | `space(path)` | 0.4 | 1 to 3 µs; no allocation |
| DSK-02 | `list()` | 0.4 | Linux 20 µs + 3 µs per mount; Windows 50 to 500 µs |
| DSK-03 | `list_with(options)` | 0.4 | as `list()` |
| DSK-04 | `DiskKind` SSD/HDD | 0.5 | 2 to 10 µs per disk |
| DSK-05 | `of(path)` | 0.5 | Linux 20 µs + 3 µs per mount; Windows 110 to 200 µs (measured) |

`list()` never blocks on a dead network mount: network filesystems are skipped unless `include_network` is set, and Windows volume calls run with critical-error dialogs disabled, so an empty card reader never pops up "insert a disk" (threat T8). Mount points appear once per filesystem (deduplicated by device number); escaped characters in Linux mount points (`\040` and friends) are decoded. Inside a container, `list()` shows the container's mounts (the overlay root, volumes): that is correct for the process, and the host's disks are neither visible nor guessed. `space(path)` blocks if the caller points it at a dead network mount; that is the caller's choice and documented. More mounts than `max_mounts` is `TooLarge`, checked before allocation. The list is sorted by mount point.

```rust
use sys_lib as sys;

let s = sys::disk::space("/var/lib/myapp")?;
println!("{} available of {} ({} used)", s.available, s.total, s.used_percent());
if s.available < sys::Bytes::from_gib(10) { /* refuse compaction */ }

for d in sys::disk::list()? {
    println!("{:<20} {:<8} {:?} {:?}", d.mount_point.display(), d.filesystem, d.kind, d.space.map(|s| s.available));
}
# Ok::<(), sys_lib::Error>(())
```

```rust
use sys_lib::disk::{self, ListOptions};

let mut options = ListOptions::default();
options.include_network = true;      // accept the risk of a hang on a dead mount
let all = disk::list_with(options)?;
# Ok::<(), sys_lib::Error>(())
```

### Disk I/O: read and write

**Since** 0.4 · DSK-06, DSK-07

```rust
pub mod disk {
    pub const DEFAULT_WINDOW: Duration;                         // 1 s
    pub fn io() -> Result<DiskIo>;                              // BLOCKS for DEFAULT_WINDOW: every physical disk together
    pub fn io_over(window: Duration) -> Result<DiskIo>;         // BLOCKS for `window`, clamped to 100 ms ..= 60 s
    pub fn io_totals() -> Result<IoTotals>;                     // since boot
    pub fn io_counters() -> Result<Vec<IoCounters>>;            // Tier 2: per physical disk
    pub fn pressure() -> Result<Option<Pressure>>;              // time tasks waited on I/O

    pub struct IoSampler { /* previous counters and Instant */ }
    impl IoSampler {
        pub fn new() -> Result<IoSampler>;
        pub fn sample(&mut self) -> Result<DiskIo>;                        // never blocks; no allocation
        pub fn sample_per_disk(&mut self) -> Result<Vec<DeviceIo>>;        // allocates the Vec
    }
}

#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct DiskIo {
    pub read: Rate, pub write: Rate,
    pub reads_per_sec: f64, pub writes_per_sec: f64,     // operations (IOPS)
    pub busy: Option<Percent>,                           // share of the interval with I/O in flight (the mean over several disks)
    pub average_latency: Option<Duration>,               // time per completed operation
    pub interval: Duration,
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct IoTotals { pub read: Bytes, pub written: Bytes, pub reads: u64, pub writes: u64, pub io_time: Option<Duration> }

#[derive(Clone, Debug, PartialEq)]
#[non_exhaustive]
pub struct DeviceIo { pub device: String, pub io: DiskIo }       // "nvme0n1", "disk0", "PhysicalDrive0"

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct IoCounters { pub device: String, pub totals: IoTotals }
```

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| DSK-06 | `io()`, `io_over()`, `io_totals()`, `io_counters()`, `IoSampler` | 1 / 2 | Linux 10 to 30 µs (`/proc/diskstats`, one read); macOS 50 to 200 µs (IOKit); Windows 20 µs per disk |
| DSK-07 | `pressure()` | 1 | Linux 2 to 5 µs |

How hard the disks are working right now: read and write speed, operations per second, how busy each disk is, and how long each operation takes. Like network traffic, a rate needs two readings: `io()` blocks for one second; a monitor keeps an `IoSampler`. Only physical disks are counted (partitions, device-mapper and RAID members, loop, RAM, and zram devices are left out), so nothing is counted twice.

| OS | Source |
|---|---|
| Linux | `/proc/diskstats`: sectors read and written (512-byte units whatever the device's block size), completed operations, time spent on I/O (`io_ticks`, for `busy`), read and write time (for latency); whole disks only, identified through `/sys/block` |
| macOS | IOKit `IOBlockStorageDriver` `Statistics` (bytes, operations, total time read and written) |
| Windows | `IOCTL_DISK_PERFORMANCE` on each `\\.\PhysicalDriveN` opened with no access rights (`BytesRead`, `BytesWritten`, `ReadCount`, `WriteCount`, `IdleTime`, `QueryTime`); works unprivileged. `QueryTime` is the wall clock, so busy time is counted from a per-process origin |

`disk::pressure()` is the kernel's I/O pressure (the cgroup's `io.pressure`, or `/proc/pressure/io`): the share of time tasks were stalled on I/O. macOS and Windows: `Ok(None)`. A counter that goes backwards (a disk removed and re-attached) re-bases that disk. Inside a container the disks are the host's (the kernel keeps one set); the cgroup's own I/O is `process::io()` for this process.

```rust
use sys_lib as sys;

let io = sys::disk::io()?;            // blocks 1 s
println!("read {} write {} ({:.0} IOPS, busy {:?})", io.read, io.write, io.reads_per_sec + io.writes_per_sec, io.busy);
if let Some(p) = sys::disk::pressure()? {
    if p.level >= sys::PressureLevel::Warning { /* postpone compaction */ }
}
# Ok::<(), sys_lib::Error>(())
```

---

## `net`: network interfaces and traffic

### Interfaces

**Since** 0.4 (NET-01 to NET-03), 0.5 (NET-06)

```rust
pub mod net {
    pub fn interfaces() -> Result<Vec<Interface>>;
    pub fn addresses() -> Result<Vec<IpAddr>>;        // non-loopback, non-link-local addresses of up interfaces
    pub fn primary() -> Result<Option<Interface>>;    // the one holding the default route (IPv4 only on Windows)
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Interface {
    pub name: String,                  // "eth0", "en0", "Ethernet 2"
    pub index: u32,
    pub description: Option<String>,   // Windows friendly description
    pub kind: InterfaceKind,
    pub mac: Option<Mac>,
    pub addresses: Vec<IpNet>,
    pub up: bool,                      // administratively up
    pub running: bool,                 // link up / operational
    pub mtu: Option<u32>,
    pub speed_mbps: Option<u64>,       // None when the driver does not report it
    pub is_physical: bool,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum InterfaceKind { Ethernet, Wifi, Loopback, Tunnel, Bridge, Virtual, Cellular, Other }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub struct Mac { pub bytes: [u8; 6], pub kind: MacKind }
impl fmt::Display for Mac;   // "aa:bb:cc:dd:ee:ff"

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum MacKind { Permanent, Random, Unknown }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
pub struct IpNet { pub addr: IpAddr, pub prefix: u8 }
impl fmt::Display for IpNet;   // "192.168.1.20/24"

```

| Id | Item | Since | Cost [est.] |
|---|---|---|---|
| NET-01 | `interfaces()` | 0.4 | Linux 30 to 150 µs, macOS 20 to 80 µs, Windows about 16 ms (measured, `GetAdaptersAddresses` and `GetIfTable2`) |
| NET-02 | `Mac`, `MacKind` | 0.4 | 0 |
| NET-03 | `addresses()` | 0.4 | Linux 10 to 30 µs |
| NET-06 | `primary()` | 0.5 | 10 to 50 µs |

**Permanent or random MAC.** Phones, Apple devices, Windows, and NetworkManager randomize Wi-Fi addresses. `MacKind::Random` marks them: on Linux from the kernel's `addr_assign_type`, elsewhere from the locally administered bit. A MAC address is an identifier: for device identity use [`device`](#device-device-identity), never a raw MAC.

Interfaces are sorted by index, then name. More than 4 096 interfaces is `TooLarge`. Names from the OS are validated before they are used to build a path.

```rust
use sys_lib as sys;

for i in sys::net::interfaces()? {
    let mac = i.mac.map(|m| format!("{m} ({:?})", m.kind)).unwrap_or_default();
    println!("{} {:?} up={} speed={:?} {mac}", i.name, i.kind, i.running, i.speed_mbps);
    for a in &i.addresses { println!("    {a}"); }
}
println!("addresses: {:?}", sys::net::addresses()?);
# Ok::<(), sys_lib::Error>(())
```

### Traffic: download and upload

**Since** 0.4 · NET-04, NET-05, NET-08

```rust
pub mod net {
    pub const DEFAULT_WINDOW: Duration;                        // 1 s
    pub fn totals() -> Result<Totals>;                         // since boot, every interface but loopback
    pub fn counters() -> Result<Vec<Counters>>;                // Tier 2: per interface, loopback included
    pub fn traffic() -> Result<Traffic>;                       // BLOCKS for DEFAULT_WINDOW
    pub fn traffic_over(window: Duration) -> Result<Traffic>;  // BLOCKS for `window`, clamped to 100 ms ..= 60 s

    pub struct Sampler { /* previous counters and Instant */ }
    impl Sampler {
        pub fn new() -> Result<Sampler>;                                      // the first reading
        pub fn sample(&mut self) -> Result<Traffic>;                          // never blocks; no allocation
        pub fn sample_per_interface(&mut self) -> Result<Vec<InterfaceTraffic>>;  // allocates the Vec
    }
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct Totals {
    pub received: Bytes, pub sent: Bytes,
    pub packets_received: u64, pub packets_sent: u64,
    pub errors: u64, pub dropped: u64,
}

#[derive(Clone, Copy, PartialEq, Debug)]
#[non_exhaustive]
pub struct Traffic {
    pub download: Rate, pub upload: Rate,
    pub packets_in_per_sec: f64, pub packets_out_per_sec: f64,
    pub interval: Duration,
}

#[derive(Clone, Debug, PartialEq)]
#[non_exhaustive]
pub struct InterfaceTraffic { pub interface: String, pub traffic: Traffic }

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Counters {
    pub interface: String,
    pub received: Bytes, pub sent: Bytes,
    pub packets_received: u64, pub packets_sent: u64,
    pub errors_in: u64, pub errors_out: u64, pub dropped_in: u64, pub dropped_out: u64,
}
```

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| NET-04 | `totals()`, `counters()` | 1 / 2 | Linux 5 to 20 µs (`/proc/net/dev`, one read), macOS 10 µs, Windows 10 to 80 ms (`GetIfTable2`, measured) |
| NET-05 | `Sampler` | 2 | as `totals()`; `sample()` allocates nothing |
| NET-08 | `traffic()`, `traffic_over()` | 1 | 1 s of wall time (the window); two readings |

How much this machine is downloading and uploading right now, as `Rate` values ("12.5 MiB/s", or `.bits()` for "104.9 Mbit/s"). Like CPU usage, a rate needs two readings: `traffic()` takes both itself and blocks for one second, which its name and documentation say; a monitor keeps a `Sampler`, which never blocks and measures the time since its previous sample. `totals()` gives the bytes and packets moved since boot, one cheap read, no waiting.

Loopback is left out of the totals (it is the machine talking to itself); on Windows so are the rows that are neither a hardware interface nor connector-present (filter and sub-layer rows, which would count traffic twice). Counters are 64-bit on every platform; macOS's 32-bit counters wrap and are not used. A counter that goes backwards between two samples (an interface removed and re-created, a driver reset) re-bases that interface instead of reporting a negative or huge rate. Inside a container the counters are the container's network namespace, which is the traffic the process can see.

```rust
use sys_lib as sys;

let t = sys::net::traffic()?;            // blocks 1 s
println!("down {} up {}", t.download, t.upload);                   // "down 12.5 MiB/s up 1.2 MiB/s"
println!("down {}", t.download.bits());                           // "down 104.9 Mbit/s"
let total = sys::net::totals()?;
println!("since boot: {} received, {} sent", total.received, total.sent);
# Ok::<(), sys_lib::Error>(())
```

```rust,no_run
use sys_lib as sys;
use std::time::Duration;

// A monitor loop: non-blocking, caller-owned.
let mut net = sys::net::Sampler::new()?;
loop {
    std::thread::sleep(Duration::from_secs(5));
    let t = net.sample()?;
    println!("down {} up {}", t.download, t.upload);
    for i in net.sample_per_interface()? {
        println!("  {}: down {} up {}", i.interface, i.traffic.download, i.traffic.upload);
    }
}
# Ok::<(), sys_lib::Error>(())
```

### Connections

**Since** 0.4 · NET-09

```rust
pub mod net {
    pub fn connections() -> Result<Connections>;
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct Connections {
    pub tcp_established: Option<u64>,   // open TCP connections
    pub tcp_time_wait: Option<u64>,     // closed, waiting out the TIME_WAIT period
    pub tcp_total: Option<u64>,         // every TCP socket, listening ones included
    pub udp: Option<u64>,               // UDP sockets
}
```

How many network connections the machine has open, as counts. Only counts: sys-lib never lists connections, addresses, ports, or which program owns them (that is other people's data and a different tool). A value an OS does not report is `None`.

| OS | Source | Cost [est.] |
|---|---|---|
| Linux | `/proc/net/snmp` (`CurrEstab`), `/proc/net/sockstat` and `sockstat6` (`inuse`, `tw`) | 5 to 10 µs |
| macOS | `net.inet.tcp.pcbcount`, `net.inet.udp.pcbcount` (the active PCB counts: total and UDP; established and time-wait are `None`) | 1 µs |
| Windows | `GetTcpStatisticsEx` (`dwCurrEstab`, `dwNumConns`) for IPv4 and IPv6, `GetUdpStatisticsEx` (`dwNumAddrs`); time-wait is `None` | 5 µs |

Inside a container the counts are the container's network namespace.

```rust
use sys_lib as sys;

let c = sys::net::connections()?;
if let Some(open) = c.tcp_established {
    println!("{open} TCP connections open");
}
# Ok::<(), sys_lib::Error>(())
```

### DNS servers

**Since** 0.5 · NET-10

```rust
pub mod net {
    pub fn dns() -> Result<Dns>;
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Dns {
    pub servers: Vec<IpAddr>,          // what the system resolver uses
    pub upstream: Vec<IpAddr>,         // behind a local stub (systemd-resolved 127.0.0.53), the servers it forwards to
    pub search: Vec<String>,           // search domains
}
```

The DNS servers this machine is configured to use, read from configuration only: sys-lib never sends a DNS query. Linux: `/etc/resolv.conf` (`nameserver`, `search`, `domain`), and when it points at the systemd-resolved stub, `/run/systemd/resolve/resolv.conf` for the upstream servers. macOS: the SystemConfiguration dynamic store `State:/Network/Global/DNS`, falling back to `/etc/resolv.conf`. Windows: the DNS server lists of the up adapters from `GetAdaptersAddresses`, in interface-metric order, without the `fec0:0:0:ffff::` placeholders Windows lists on adapters with no IPv6 DNS; `search` holds the adapters' suffixes. Bounded: at most 64 servers and 64 search domains; anything longer is `TooLarge`.

### Interface options

**Since** 0.4 · NET-07

```rust
pub mod net {
    pub fn interfaces_with(options: InterfaceOptions) -> Result<Vec<Interface>>;
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct InterfaceOptions {
    pub include_down: bool,          // default true
    pub include_loopback: bool,      // default true
    pub details: bool,               // default true; false skips the per-interface reads (speed, MTU, kind; is_physical is then false)
    pub max_interfaces: usize,       // default 4_096
}
impl Default for InterfaceOptions;
```

A Kubernetes node with thousands of `veth` interfaces lists them all in about a millisecond with `details: false`.

---

## `host`: the machine and its environment

**Since** 0.3 (HST-01), 0.4 (HST-02, HST-03, HST-09, HST-10), 0.5 (HST-05 to HST-08, HST-11 to HST-13). The load average is [`cpu::load()`](#load).

```rust
pub mod host {
    pub fn name() -> Result<String>;                      // live
    pub fn uptime() -> Result<Duration>;                  // includes time suspended
    pub fn boot_time() -> Result<SystemTime>;
    pub fn timezone() -> Result<TimeZone>;
    pub fn locale() -> Option<&'static Locale>;          // static: the environment at the first call
    pub fn model() -> Option<&'static Model>;            // static
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct TimeZone {
    pub iana: Option<String>,           // "Europe/London"
    pub windows: Option<String>,        // "GMT Standard Time"
    pub utc_offset_seconds: i32,        // now, including daylight saving
    pub abbreviation: Option<String>,   // "BST"
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Locale { pub tag: &'static str, pub raw: &'static str }   // "en-US", "en_US.UTF-8"

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Model { pub vendor: Option<&'static str>, pub name: &'static str, pub family: Option<&'static str> }
```

| Id | Item | Since | Cost [est.] |
|---|---|---|---|
| HST-01 | `name()` | 0.3 | 2 µs / 0.5 µs / 100 µs once (a new Windows name takes effect at restart), then 0.1 µs |
| HST-02 | `uptime()` | 0.4 | 0.05 to 2 µs |
| HST-03 | `boot_time()` | 0.4 | 2 to 4 µs / 0.5 µs / 0.1 µs |
| HST-05 | `timezone()` | 0.5 | 5 to 15 µs |
| HST-06 | `locale()` | 0.5 | about 1 ns after the first call |
| HST-07 | `model()` | 0.5 | about 1 ns after the first call |

- `name()` is live: DHCP and `hostnamectl` change it while processes run. A name that is not UTF-8 is converted lossily; longer than 255 bytes is truncated.
- `uptime()` is monotonic and counts time suspended; `boot_time()` is the boot instant, the same on every call (it moves only when the wall clock is stepped).
- `timezone()` returns the zone **name** and the current offset, not zone arithmetic (that is a calendar crate's job). `TZ` is used only when it is a valid name (at most 64 bytes, no `..`, no control characters, a leading `:` stripped); it is never opened as a path. When `/etc/localtime` is a copy rather than a link, the name comes from `/etc/timezone` or is `None`: zone files are not compared to guess. On Windows the IANA name comes from an embedded copy of the CLDR Windows-zone table.
- `locale()` is `None` for `C`, `POSIX`, and `C.UTF-8` (no language preference expressed). macOS falls back to the system locale, since an application started from the Finder has no `LANG`.
- `model()` is the hardware model from DMI, SMBIOS, the device tree ("Raspberry Pi 5"), or `hw.model` on macOS ("Mac14,9"; no marketing-name table is shipped).

The host name, time zone, and locale are identifying values; keep them out of output that leaves the machine unless that is intended.

```rust
use sys_lib as sys;

println!("{} up {:?}", sys::host::name()?, sys::host::uptime()?);
let tz = sys::host::timezone()?;
println!("zone {:?}, offset {} s", tz.iana, tz.utc_offset_seconds);
# Ok::<(), sys_lib::Error>(())
```

### System-wide counts

**Since** 0.4 · HST-09, HST-10

```rust
pub mod host {
    pub fn processes() -> Result<Processes>;
    pub fn open_files() -> Result<OpenFiles>;
}

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct Processes { pub total: u64, pub threads: Option<u64>, pub running: Option<u64> }

#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
#[non_exhaustive]
pub struct OpenFiles { pub open: Option<u64>, pub max: Option<u64> }   // Windows: handles, no maximum
```

How many processes and threads the machine runs and how many files are open across it, as totals only: no process is listed or inspected. Inside a container the process count is the container's process namespace.

| OS | Processes | Threads, running | Open files |
|---|---|---|---|
| Linux | the numeric directory names in `/proc` (counted, not opened) | `/proc/loadavg` (running/total), `procs_running` in `/proc/stat` | `/proc/sys/fs/file-nr` (open, max) |
| macOS | `proc_listallpids` with a buffer | `processor_set_statistics` thread count; running `None` | `kern.num_files`, `kern.maxfiles` |
| Windows | `GetPerformanceInfo` `ProcessCount` | `ThreadCount`; running `None` | `HandleCount` (handles), max `None` |

Cost: Linux 20 to 200 µs (scales with processes; bounded at 4 million entries), macOS and Windows 2 to 10 µs [est.].

### Firmware and motherboard

**Since** 0.5 · HST-11, HST-12

```rust
pub mod host {
    pub fn firmware() -> Option<&'static Firmware>;    // static
    pub fn board() -> Option<&'static Board>;          // static
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Firmware {
    pub kind: FirmwareKind,                  // Uefi, Bios, Unknown
    pub vendor: Option<&'static str>,        // "American Megatrends International, LLC.", "Apple Inc."
    pub version: Option<&'static str>,       // "1.40", "11881.41.5"
    pub date: Option<&'static str>,          // "2025-03-14" (normalised from SMBIOS's MM/DD/YYYY)
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct Board { pub vendor: Option<&'static str>, pub name: Option<&'static str>, pub version: Option<&'static str> }

#[non_exhaustive] pub enum FirmwareKind { Uefi, Bios, Unknown }
```

| OS | Firmware | Board |
|---|---|---|
| Linux | `/sys/class/dmi/id/{bios_vendor,bios_version,bios_date}` (mode 0444); UEFI when `/sys/firmware/efi` exists | `board_vendor`, `board_name`, `board_version` (0444) |
| Windows | SMBIOS type 0 from `GetSystemFirmwareTable('RSMB')`; `GetFirmwareType` | SMBIOS type 2 |
| macOS | vendor "Apple Inc."; Apple silicon: version from `system-firmware-version` in the IORegistry, kind `Unknown` (iBoot is not UEFI); Intel: kind `Uefi`, version `None` (D-059) | `board-id` of `IOPlatformExpertDevice` (Intel); Apple silicon: the model identifier |

Placeholder strings ("To be filled by O.E.M.", "Default string") are `None`. Serial numbers are never part of these answers (they are identifiers; see `device`). Inside a VM these describe the virtual firmware, which is correct for the process.

### Clock synchronization

**Since** 0.5 · HST-13

```rust
pub mod host {
    pub fn clock_sync() -> Result<Option<ClockSync>>;
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct ClockSync {
    pub synchronized: bool,                     // the kernel believes the clock is disciplined
    pub max_error: Option<Duration>,            // the kernel's error bound
    pub estimated_error: Option<Duration>,
}
```

Whether the system clock is kept in sync (NTP, chrony, PTP, Windows Time), which leases, distributed locks, and hybrid logical clocks depend on: a cluster node with an unsynchronised clock should not take part. Linux: `adjtimex` in read-only mode (no mode bits set), `TIME_ERROR` or `STA_UNSYNC` meaning not synchronised; a sandbox refusing the call is `Ok(None)`. macOS: `ntp_adjtime` in read-only mode, but the kernel starts with `STA_UNSYNC` set and `timed` may set the clock without clearing it, so a clear flag is `synchronized: true` and a set flag is `Ok(None)`, never a claim that the clock is adrift. Windows and iOS: `Ok(None)`: no unprivileged call tells (D-054). `Ok(None)` means not known, so a cluster should treat it as such and not as unsynchronised. Reading time itself stays in clock-lib.

### Sessions

**Since** 0.5 · HST-08

```rust
pub mod host {
    pub fn sessions() -> Result<Vec<Session>>;
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Session { pub user: String, pub terminal: Option<String>, pub remote: bool, pub since: Option<SystemTime> }
```

Who is logged in: Linux from the utmp file itself (a system that no longer keeps one gives an empty list, D-056), macOS from `getutxent`, Windows from `WTSEnumerateSessionsW` and `WTSQuerySessionInformationW` (wtsapi32, loaded at run time: `Unsupported` where Windows has none, D-055; a session this process may not ask about is left out). iOS: `Unsupported`. The remote host address is not returned (only whether the session is remote). User names are personal data: keep them out of output that leaves the machine.

---

## `virt`: virtualization and containers

**Since** 0.5 · VRT-01 to VRT-04

```rust
pub mod virt {
    pub fn detect() -> &'static Virtualization;
    pub fn is_vm() -> bool;
    pub fn is_container() -> bool;
    pub fn is_wsl() -> bool;
    pub fn is_kubernetes() -> bool;
    pub fn hypervisor() -> Option<Hypervisor>;
    pub fn container() -> Option<Container>;
}

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Virtualization {
    pub hypervisor: Option<Hypervisor>,
    pub container: Option<Container>,
    pub wsl: Option<u8>,           // 1 or 2
    pub kubernetes: bool,
    pub evidence: Vec<Evidence>,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Hypervisor {
    Kvm, Qemu, Vmware, HyperV, Xen, VirtualBox, Parallels, Bhyve, Acrn, Qnx,
    AppleVirtualization, AmazonNitro, GoogleCompute, Bochs, Other,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Container { Docker, Podman, Containerd, Lxc, SystemdNspawn, OpenVz, Other }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub struct Evidence { pub check: &'static str, pub strength: Strength }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Strength { Strong, Weak }
```

| Id | Item | Cost |
|---|---|---|
| VRT-01 | `detect()` | about 1 ns; the first call 20 to 200 µs [est.] |
| VRT-02 | `is_vm`, `is_container`, `is_wsl`, `is_kubernetes` | about 1 ns |
| VRT-03 | `hypervisor()`, `container()` | about 1 ns |
| VRT-04 | `evidence` | 0 |

Layered checks (CPUID hypervisor leaves, DMI strings, hypervisor files, container marker files and cgroup names, the overlay root, the Kubernetes service account, WSL's kernel release) with every check that fired named in `evidence` (constant strings, never raw values). A Windows **host** with Hyper-V, WSL 2, or virtualization-based security enabled sees "Microsoft Hv" in CPUID too; with the root-partition privilege it is not reported as a VM (checked on such a host). Under Rosetta 2 CPUID is not read: translation is not virtualization. Inside WSL 2 both `is_vm()` and `is_wsl()` are true. A container inside a VM sets both `container` and `hypervisor`. Evidence from environment variables (`KUBERNETES_SERVICE_HOST`) is `Weak`: any process can set it.

Detection is advisory. Do not make security decisions on it (threat T14).

```rust
use sys_lib as sys;

let v = sys::virt::detect();
if sys::virt::is_container() {
    println!("in {:?}; kubernetes: {}", v.container, v.kubernetes);
}
for e in &v.evidence { println!("  {} ({:?})", e.check, e.strength); }
```

**Cloud vendor** (VRT-05, since 0.5): `virt::cloud() -> Option<Cloud>` with `Cloud { Aws, Gcp, Azure, Oracle, Alibaba, Hetzner, DigitalOcean, Linode, Vultr, Scaleway, Ovh, Other }`, from the DMI and SMBIOS vendor and product strings and the hypervisor leaves only. sys-lib never contacts a metadata service: instance types, regions, and zones need a network call and belong to a cloud crate.

---

## `power`: battery and AC

**Since** 0.5 · PWR-01 to PWR-03

```rust
pub mod power {
    pub fn has_battery() -> bool;                     // static
    pub fn battery() -> Result<Option<Battery>>;      // None: no system battery
    pub fn on_ac() -> Result<Option<bool>>;           // None: unknown
}

#[derive(Clone, Copy, Debug)]
#[non_exhaustive]
pub struct Battery {
    pub charge: Percent,
    pub state: BatteryState,
    pub time_to_empty: Option<Duration>,
    pub time_to_full: Option<Duration>,
    pub health: Option<Percent>,     // full capacity / design capacity
    pub cycles: Option<u32>,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum BatteryState { Charging, Discharging, Full, NotCharging, Unknown }
```

| Id | Item | Cost [est.] |
|---|---|---|
| PWR-01 | `has_battery()` | about 1 ns after the first call |
| PWR-02 | `battery()` | 10 to 40 µs / 50 to 200 µs / 0.3 µs |
| PWR-03 | `on_ac()` | 5 to 20 µs / 20 µs / 0.3 µs |

The system battery only: wireless mice, keyboards, and headsets report batteries too and are excluded. Several internal batteries are combined by energy (the sum of energy now over the sum of energy when full), not by averaging percentages. A desktop without power information gives `on_ac() == Ok(None)`, not a guess; a Mac desktop says `Some(true)`, because IOKit reports mains power there. Windows reports one combined entry and no health or cycle count through `GetSystemPowerStatus` (`None`); each pack on its own is not verified. Android: best effort through the Linux path (SELinux often denies it: `PermissionDenied`). iOS: `Unsupported`.

```rust
use sys_lib as sys;

if let Some(b) = sys::power::battery()? {
    println!("battery {} {:?}, empty in {:?}", b.charge, b.state, b.time_to_empty);
}
if sys::power::on_ac()? == Some(false) { /* on battery */ }
# Ok::<(), sys_lib::Error>(())
```

**Each battery** (PWR-04, since 0.5): `power::batteries() -> Result<Vec<Battery>>` returns every system battery separately (laptops with two packs, UPS-backed devices that report as system batteries), in the order the OS lists them; `battery()` stays the combined value. Windows gives one combined entry.

---

## `sensors`: temperatures

**Since** 0.6 · SNS-01

```rust
pub mod sensors {
    pub fn temperatures() -> Result<Vec<Temperature>>;
}

#[derive(Clone, Debug, PartialEq)]
#[non_exhaustive]
pub struct Temperature {
    pub label: String,                  // "Package id 0", "Composite", "edge"
    pub kind: SensorKind,               // Cpu, Gpu, Storage, Battery, Ambient, Other
    pub celsius: f32,
    pub high: Option<f32>, pub critical: Option<f32>,
}
```

| OS | Source |
|---|---|
| Linux | `/sys/class/hwmon/hwmon*/temp*_{input,label,max,crit}` (millidegrees); kind from the hwmon `name` (`coretemp`, `k10temp`, `zenpower`, `nvme`, `amdgpu`, `acpitz`); `/sys/class/thermal/thermal_zone*` when hwmon is absent |
| macOS, iOS | `Unsupported`: Apple silicon sensors need private HID event APIs and Intel SMC keys are undocumented; nothing undocumented is used |
| Windows | No CPU entry: the only documented source (`MSAcpi_ThermalZoneTemperature`) needs WMI and Administrator and reports ACPI zones, not the CPU. Drives through `IOCTL_STORAGE_QUERY_PROPERTY` `StorageDeviceTemperatureProperty` on each `\\.\PhysicalDriveN` opened with no access rights (unprivileged), labelled with the product id and drive number; the warning and critical thresholds when the drive gives them (D-071) |
| FreeBSD | `dev.cpu.N.temperature` (with `coretemp` or `amdtemp` loaded) and the ACPI thermal zones `hw.acpi.thermal.tzN.temperature` |

Inside Docker and WSL, hwmon is usually absent: an empty list, not an error.

## `gpu`: graphics adapters

**Since** 0.6 · GPU-01

```rust
pub mod gpu {
    pub fn list() -> Result<Vec<Gpu>>;
}

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Gpu {
    pub name: String, pub vendor: GpuVendor, pub kind: GpuKind,      // Integrated, Discrete, Virtual, Unknown
    pub vendor_id: Option<u16>, pub device_id: Option<u16>,
    pub memory: Option<Bytes>, pub driver: Option<String>,
}
#[non_exhaustive] pub enum GpuVendor { Nvidia, Amd, Intel, Apple, Qualcomm, Arm, Microsoft, Vmware, Other }
```

| OS | Source |
|---|---|
| Linux | `/sys/class/drm/card*/device/{vendor,device,uevent,mem_info_vram_total}`; the name from `pci.ids` when installed (`/usr/share/hwdata/pci.ids`, `/usr/share/misc/pci.ids`; a bounded, fuzzed parser), else "Vendor GPU 0x1234" |
| macOS | IOKit `IOPCIDevice` entries whose `class-code` (little-endian data) has base class 0x03: `model` (data), `vendor-id`, `device-id`, `VRAM,totalMB` or `VRAM,totalsize`. With no PCI adapter (Apple silicon), the `IOAccelerator` services (`AGXAccelerator*`): `model` (a string), `CFBundleIdentifier` as the driver, else the chip from `machdep.cpu.brand_string`. An Apple adapter on PCI, or a `Paravirt` accelerator, is virtual (D-081, D-082) |
| Windows | `EnumDisplayDevicesW` (user32, loaded at run time), one entry per adapter, mirror drivers left out; the adapter's `DeviceKey` registry key gives `HardwareInformation.qwMemorySize`, else `HardwareInformation.MemorySize` (capped near 4 GiB), and `DriverVersion`; the basic display driver and hypervisor adapters are virtual (D-072) |
| iOS, FreeBSD | `Unsupported` (FreeBSD needs `/dev/pci`, which is root-only) |

No GPU usage or memory-in-use: those need vendor libraries (NVML, ROCm SMI, DXGI COM), which sys-lib does not load.

---

## `readiness`: server readiness report

**Since** 0.6 · RDY-01 to RDY-04

```rust
pub mod readiness {
    pub fn report(profile: Profile) -> Report;                              // never fails
    pub fn report_for(profile: Profile, data_dir: impl AsRef<Path>) -> Report;  // adds the data directory's checks
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Profile { General, Database, WebServer, Benchmark }

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Report { pub profile: Profile, pub checks: Vec<Check> }
impl Report {
    pub fn passed(&self) -> bool;                                    // no Fail
    pub fn warnings(&self) -> impl Iterator<Item = &Check>;
    pub fn failures(&self) -> impl Iterator<Item = &Check>;
    pub fn write_kv(&self, out: &mut impl fmt::Write) -> fmt::Result;
    pub fn write_json(&self, out: &mut impl fmt::Write) -> fmt::Result;
}
impl fmt::Display for Report;   // a table: status, check, found, recommended

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Check {
    pub id: &'static str,             // "linux.thp", "limits.nofile", "fs.noatime"
    pub title: &'static str,          // "Transparent huge pages"
    pub status: Status,               // Pass, Warn, Fail, Unknown
    pub found: String,                // "always"
    pub recommended: &'static str,    // "madvise or never"
    pub why: &'static str,            // one sentence
    pub fix: &'static str,            // the command or setting that changes it
}

#[non_exhaustive] pub enum Status { Pass, Warn, Fail, Unknown }
```

A start-up check of the operating-system settings that decide whether a server performs and stays up under load, with the reason and the fix for each, the way enterprise databases warn at boot. **Read-only:** sys-lib reports and never changes a setting; `fix` tells the operator what to run. Every threshold is a documented constant per profile.

| Check | Linux | macOS | Windows | Database profile wants |
|---|---|---|---|---|
| transparent huge pages | `/sys/kernel/mm/transparent_hugepage/{enabled,defrag}` | | | `madvise` or `never` |
| swappiness | `/proc/sys/vm/swappiness` | | | 1 to 10 |
| memory overcommit | `/proc/sys/vm/overcommit_memory`, `overcommit_ratio` | | | 0 or 2, never 1 for databases |
| memory map count | `/proc/sys/vm/max_map_count` | | | at least 262 144 |
| dirty page flushing | `/proc/sys/vm/dirty_ratio`, `dirty_background_ratio` (or `_bytes`) | | | background at most 5%, foreground at most 15% |
| NUMA balancing and zone reclaim | `/proc/sys/kernel/numa_balancing`, `/proc/sys/vm/zone_reclaim_mode` | | | both 0 on multi-socket hosts |
| open-file limits | `RLIMIT_NOFILE` of this process; `/proc/sys/fs/file-max`, `nr_open` | `RLIMIT_NOFILE`; `kern.maxfiles`, `kern.maxfilesperproc` | handles are unlimited: Pass | soft at least 65 536 |
| locked-memory limit | `RLIMIT_MEMLOCK` | same | the lock-pages privilege (`SeLockMemoryPrivilege`) in the token | enough for the buffer pool when it is locked |
| listen backlog | `/proc/sys/net/core/somaxconn`, `net/ipv4/tcp_max_syn_backlog` | `kern.ipc.somaxconn` | | at least 4 096 |
| ephemeral ports | `/proc/sys/net/ipv4/ip_local_port_range` | `net.inet.ip.portrange.first`/`last` | `Unknown`: no documented unprivileged source (netsh keeps the range in the NSI store, D-073) | at least 28 000 ports |
| CPU frequency policy | the `scaling_governor` of each cpufreq policy | | the active power plan (`PowerGetActiveScheme`), High performance and Ultimate Performance matched by GUID or by the plan's personality (D-074) | `performance` / High performance |
| clock source | `/sys/devices/system/clocksource/clocksource0/current_clocksource` | | | `tsc` (`hpet`, `xen`, `acpi_pm` are slow to read) |
| clock synchronized | `host::clock_sync()` | same | same | synchronized |
| swap present | `mem::swap()` | same | pagefile | some swap or zram, small |
| I/O scheduler of the data disk | `/sys/block/<disk>/queue/scheduler` | | | `none` or `mq-deadline` on SSD and NVMe |
| data directory filesystem | `disk::of(data_dir)`: type; access times from the mount's options in `/proc/self/mountinfo` | `statfs` `f_flags`: `MNT_NOATIME`, `MNT_STRICTATIME`; APFS without either updates access times relatime-style (D-083) | `GetVolumeInformationW`; on NTFS, access times from `NtfsDisableLastAccessUpdate` (D-076) | local; ext4 or xfs (Linux), APFS, NTFS or ReFS, ZFS or UFS (FreeBSD); `noatime` or `relatime` |
| data directory space and kind | `disk::space`, `DiskKind` | same | same | SSD; at least 15% free |
| container limits | `mem::limit()`, `cpu::quota()` | | job object | reported (Pass with the values) |

Each check runs only where it applies; elsewhere it is left out (not reported as passing). A value sys-lib cannot read is `Unknown` with the reason, never a guess. A check fails only where the setting loses data or breaks under load: memory overcommit mode 1 for a database, an open-file limit under 1 024 for a server, a data directory on a network filesystem for a database, or under 5% free; everything else warns (D-064). FreeBSD reads `kern.maxfiles`, `kern.maxfilesperproc`, `kern.ipc.soacceptqueue`, the port range, and the timecounter. Cost: under 2 ms on Linux; 10 to 30 ms on Windows, where the power plan, volume, and registry calls dominate (measured on a loaded machine).

```rust
use sys_lib::readiness::{self, Profile};

let report = readiness::report_for(Profile::Database, "/var/lib/myapp");
println!("{report}");
// WARN  linux.thp        Transparent huge pages   always    madvise or never
//       echo madvise > /sys/kernel/mm/transparent_hugepage/enabled
if !report.passed() { /* refuse to start, or log loudly */ }
```

## `bench`: benchmark environment report

**Since** 0.6 · BEN-01 to BEN-03

```rust
pub mod bench {
    pub fn environment() -> Environment;          // never fails; static parts cached, live parts read now
}

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Environment {
    pub governor: Option<String>,          // "performance", "powersave", "schedutil"; Windows power plan name
    pub turbo: Option<bool>,               // boost enabled
    pub smt: Option<bool>,                 // hyper-threading active
    pub invariant_tsc: Option<bool>,       // the CPU's time-stamp counter runs at a constant rate (x86)
    pub clock_source: Option<String>,      // Linux: "tsc"
    pub throttled: Option<bool>,           // thermal or power throttling since boot (Linux counters) or now (Windows MhzLimit)
    pub isolated_cpus: Option<String>,     // Linux isolcpus / nohz_full
    pub aslr: Option<bool>,
    pub on_ac: Option<bool>,
    pub load: Option<cpu::Load>,
    pub virtualized: bool,
    pub warnings: Vec<Warning>,
}
impl Environment {
    pub fn is_reproducible(&self) -> bool;   // no warning that makes results noisy
    pub fn write_kv(&self, out: &mut impl fmt::Write) -> fmt::Result;
    pub fn write_json(&self, out: &mut impl fmt::Write) -> fmt::Result;
}
impl fmt::Display for Environment;

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Warning { PowerSaving, TurboEnabled, Throttled, OnBattery, Busy, Virtualized, UnstableClock }
```

What a benchmark result needs printed next to it so it can be trusted and repeated: whether the CPU frequency could wander (governor, turbo), whether the machine was busy, on battery, throttled, or a VM, and whether the timer is stable. Benchmark harnesses (rust-benchmark, criterion wrappers) print it at the start of a run. Read-only, like the readiness report.

| Item | Linux | macOS | Windows |
|---|---|---|---|
| governor | `cpufreq/scaling_governor` | `None` (no user-visible governor) | active power plan name (`PowerGetActiveScheme`, `PowerReadFriendlyName`) |
| turbo | `intel_pstate/no_turbo`, `cpufreq/boost` | `None` | the plan's PERFBOOSTMODE (not Disabled) with PROCTHROTTLEMAX at 100%, for the current power source: a policy, so `Some(true)` even on a CPU without boost (D-075) |
| SMT | `/sys/devices/system/cpu/smt/active` | `physical != logical` (Apple silicon `Some(false)`) | `LTP_PC_SMT` flag in the core records |
| invariant TSC | CPUID 0x80000007 EDX bit 8 (x86 only) | same on Intel Macs; Apple silicon `None` | same |
| throttled | `thermal_throttle/{core,package}_throttle_count` per CPU | `None` | `CallNtPowerInformation` `MhzLimit < MaxMhz` |
| isolated CPUs | `/sys/devices/system/cpu/{isolated,nohz_full}` | `None` | `None` |

## `security`: security posture

**Since** 0.6 · SEC-01 to SEC-05

```rust
pub mod security {
    pub fn posture() -> Posture;                  // never fails; each field None when unknown
}

#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Posture {
    pub secure_boot: Option<bool>,
    pub tpm: Option<Tpm>,                          // present, and its version
    pub disk_encryption: Option<Encryption>,      // of the volume holding the OS
    pub mandatory_access_control: Option<Mac>,    // SELinux or AppArmor and its mode
    pub aslr: Option<bool>,
    pub kernel_lockdown: Option<Lockdown>,         // Linux lockdown, Windows HVCI/VBS
    pub code_integrity: Option<bool>,             // macOS SIP, Windows HVCI
    pub firewall: Option<bool>,
    pub cpu_vulnerabilities: Vec<Vulnerability>,  // name and status, e.g. "spectre_v2: Mitigated"
    pub elevated: Option<bool>,                   // process::is_elevated()
}
#[non_exhaustive] pub struct Tpm { pub version: TpmVersion }                 // V1_2, V2_0
#[non_exhaustive] pub enum Encryption { Enabled { kind: &'static str }, Disabled }   // "LUKS", "BitLocker", "FileVault"
#[non_exhaustive] pub struct Mac { pub kind: &'static str, pub enforcing: bool }    // "SELinux", "AppArmor"
#[non_exhaustive] pub enum Lockdown { Off, Integrity, Confidentiality }
#[non_exhaustive] pub struct Vulnerability { pub name: String, pub status: VulnStatus, pub detail: String }
#[non_exhaustive] pub enum VulnStatus { NotAffected, Mitigated, Vulnerable, Unknown }
```

The machine's security settings, read-only, for shield-sec, compliance reports, and a service's start-up log. Like detection, it is a report, not a control: a value is what the OS says, and an administrator controls the OS. Raw identifiers (TPM endorsement keys, volume ids) are never read.

| Item | Linux | macOS | Windows |
|---|---|---|---|
| Secure Boot | `/sys/firmware/efi/efivars/SecureBoot-8be4df61-93ca-11d2-aa0d-00e098032b8c` (the fifth byte); without `/sys/firmware/efi`, false (a BIOS boot), or `None` in a container, which may hide it | `None`: no documented unprivileged API (`bputil` and the NVRAM boot-policy variables are a program and private) | `HKLM\SYSTEM\CurrentControlSet\Control\SecureBoot\State` `UEFISecureBootEnabled`; false on a BIOS boot |
| TPM | `/sys/class/tpm/tpm0/tpm_version_major` | `None` (the Secure Enclave is not a TPM) | `Tbsi_GetDeviceInfo` (tbs.dll, unprivileged) |
| disk encryption | the root filesystem's block device and the devices under it (LVM on LUKS): `dm/uuid` starting `CRYPT-` under `/sys/dev/block/<major:minor>` and its `slaves/` | `None`: FileVault's state comes only from `fdesetup`, a program | `None`: the only unprivileged sources are the shell property (COM) and WMI (Administrator) (D-077) |
| SELinux / AppArmor | `/sys/fs/selinux/enforce`; `/sys/module/apparmor/parameters/enabled` | | |
| ASLR | `/proc/sys/kernel/randomize_va_space` | always on: `Some(true)` | `GetProcessMitigationPolicy(ProcessASLRPolicy)` for this process |
| lockdown, code integrity | `/sys/kernel/security/lockdown` | `None`: `csr_get_active_config` is exported, but its header is private to XNU and it is undocumented (D-085) | `HKLM\SYSTEM\CurrentControlSet\Control\DeviceGuard`: `Scenarios\KeyGuard\Status` `IsSecureKernelRunning` (VBS running gives `Integrity`), else the configuration; HVCI `Enabled` while VBS runs gives `code_integrity` (D-078) |
| firewall | `None` (nftables and iptables state is root-only) | `None`: no documented API | each profile's `EnableFirewall`, Group Policy's value first, then the local one under `...\SharedAccess\Parameters\FirewallPolicy`: true or false only when all three profiles agree, else `None` (finding the active profile needs COM, D-079) |
| CPU vulnerabilities | `/sys/devices/system/cpu/vulnerabilities/*` (0444) | `None` | `None` (the documented source needs a driver) |

---

---

## `device`: device identity

**Since** 0.5 · DEV-01 to DEV-07 · feature `device-id` (off by default)

> **Not authentication.** A device id is a claim, not proof: root or Administrator can set the machine id to anything, and hypervisors set firmware UUIDs freely. Never use it as the only check for anything that matters. It is personal data in client software. Read [When not to use it](#when-not-to-use-it) first.

### Tier 1

```rust
pub mod device {
    /// An application-specific device id. `app` is a stable, application-unique string
    /// such as a reverse domain name ("com.example.server"), 1 to 255 bytes.
    pub fn id(app: &str) -> Result<DeviceId>;
    pub fn fingerprint(app: &str) -> Result<Fingerprint>;
    pub fn report() -> StabilityReport;               // sources only, no application needed
    pub fn builder(app: &str) -> Builder;
}

#[derive(Clone, Copy, PartialEq, Eq, Hash)]
pub struct DeviceId(/* [u8; 16] */);
impl DeviceId {
    pub fn as_bytes(&self) -> &[u8; 16];
    pub fn parse(text: &str) -> Result<DeviceId>;     // accepts its own Display format only
    pub fn ct_eq(&self, other: &DeviceId) -> bool;    // constant time
}
impl fmt::Display for DeviceId;   // RFC 9562 UUID text, version 8: "6f1c2a4e-9b3d-8e21-a7c4-..."
impl fmt::Debug for DeviceId;     // DeviceId(6f1c2a4e-...): already a keyed hash, safe to show
impl FromStr for DeviceId;
```

The id is a keyed BLAKE3 hash of **one** identifier the OS keeps (by default the OS installation id: `/etc/machine-id`, the Windows `MachineGuid`, or on macOS the platform UUID), keyed by a key derived from your application id. Two applications on the same machine get unrelated ids, and nobody can turn an id back into the machine id or the serial numbers behind it. The raw identifier is read into a fixed buffer, hashed, and wiped; it never appears in a return value, an error, `Debug`, `Display`, or a log. sys-lib does not cache the id; store it if you need it often.

| Id | Item | Tier | Cost [est.] |
|---|---|---|---|
| DEV-01 | `device::id(app)` | 1 | Linux 5 to 15 µs; macOS 50 to 200 µs (IOKit); Windows 20 to 50 µs (registry); the hash under 1 µs |
| DEV-02 | `DeviceId` | 1 | `Display` about 40 ns |
| DEV-03 | `device::report()` | 1 | every source probed once, 0.1 to 2 ms |
| DEV-04 | `Builder` | 2 | as `id` |
| DEV-05 | `device::fingerprint(app)`, `Fingerprint` | 2 | 0.2 to 3 ms (SMBIOS, interfaces) |
| DEV-06 | `Fingerprint::compare` | 2 | under 1 µs |
| DEV-07 | `to_bytes`, `from_bytes` | 2 | under 1 µs |

```rust
use sys_lib as sys;

let id = sys::device::id("com.example.server")?;
println!("device {id}");                                   // 6f1c2a4e-9b3d-8e21-a7c4-...
let again: sys::device::DeviceId = id.to_string().parse()?;
assert!(id.ct_eq(&again));
# Ok::<(), sys_lib::Error>(())
```

### `Builder`

```rust
pub struct Builder { /* private */ }
impl Builder {
    pub fn basis(self, basis: Basis) -> Self;        // default Basis::Install
    pub fn secret(self, key: [u8; 32]) -> Self;      // an application secret; wiped on drop
    pub fn fallback(self, on: bool) -> Self;         // default false: fail rather than change basis silently
    pub fn id(&self) -> Result<DeviceId>;
    pub fn fingerprint(&self) -> Result<Fingerprint>;
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Basis {
    /// This OS installation. Unprivileged everywhere; changes on reinstall.
    /// macOS has no installation id and uses the platform UUID (the report says so).
    Install,
    /// The hardware: the firmware UUID, then the board serial. Survives reinstalls.
    /// Linux needs root (the DMI identifier files are readable by root only).
    Hardware,
}

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Anchor { InstallId, PlatformUuid, BoardSerial, ProductSerial, CpuSignature, PermanentMac }
```

**With a secret**, nobody without it can compute or correlate the id, even with root on the machine: use one when ids are sent to a server. **Without one**, anyone who knows the raw machine id and your (public) application id can compute the device id, the same property as systemd's application-specific machine id. `fallback(true)` lets `Basis::Hardware` fall back to `Install` when hardware sources are denied; the id is then different from the hardware one, which is why it is off by default.

The derivation is versioned (`sys-lib/device/v1`) and **frozen for all of 1.x**: a given machine and application id give the same device id in every 1.x release. On a dual-boot machine, `Basis::Hardware` gives the same id under Linux (as root) and Windows.

```rust,no_run
use sys_lib::device::{self, Basis};

let secret: [u8; 32] = load_secret();   // the application's own secret, never shipped in clients
let id = device::builder("com.example.licensing")
    .basis(Basis::Hardware)
    .secret(secret)
    .id()?;
# fn load_secret() -> [u8; 32] { [7; 32] }
# Ok::<(), sys_lib::Error>(())
```

### Stability report

```rust
#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct StabilityReport {
    pub sources: Vec<SourceStatus>,       // one per Anchor, in Anchor order
    pub default_basis_available: bool,
    pub clone_risk: Vec<CloneRisk>,       // empty when none detected
    pub confidence: Confidence,           // for Basis::Install
    pub changes_on: Vec<ChangeEvent>,     // for the default basis on this OS
}
impl fmt::Display for StabilityReport;   // a short table for support output; contains no values

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
#[non_exhaustive]
pub struct SourceStatus { pub anchor: Anchor, pub state: SourceState }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum SourceState { Present, Absent, Placeholder, PermissionDenied, Unsupported }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum CloneRisk { Container, VirtualMachine, MachineIdUninitialized, SharedImageLikely }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum ChangeEvent { OsReinstall, BoardReplacement, NicReplacement, CpuUpgrade, VmRecreated, ContainerRecreated, AdminReset }
```

How far to trust the id on this machine, without revealing any value: which sources exist, which are placeholders that manufacturers and hypervisors ship ("To be filled by O.E.M.", all-zero UUIDs, `uninitialized` machine ids) and are treated as absent, which need privileges, whether the machine looks cloned (a container image or VM template sharing one machine id), and what events would change the id.

```rust
use sys_lib::device;

let report = device::report();
println!("{report}");
if !report.clone_risk.is_empty() {
    println!("this id may be shared with other machines: {:?}", report.clone_risk);
}
```

### Fingerprint

```rust
#[derive(Clone, PartialEq, Eq)]
pub struct Fingerprint { /* an app-keyed 16-byte hash per anchor, or absent */ }
impl Fingerprint {
    pub fn compare(&self, stored: &Fingerprint) -> Comparison;
    pub fn to_bytes(&self) -> [u8; FINGERPRINT_LEN];
    pub fn from_bytes(bytes: &[u8]) -> Result<Fingerprint>;
    pub fn anchors(&self) -> impl Iterator<Item = Anchor> + '_;   // which anchors are present
}
impl fmt::Debug for Fingerprint;   // the present anchors only
pub const FINGERPRINT_LEN: usize = 112;

#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct Comparison { pub verdict: Verdict, pub matched: Vec<Anchor>, pub changed: Vec<Anchor>, pub missing: Vec<Anchor> }

#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
#[non_exhaustive]
pub enum Verdict {
    Same,         // every anchor present in both matches, and more than the CPU signature was compared
    LikelySame,   // matched weight >= 5 and at most one strong anchor changed
    Different,
}
```

A fingerprint hashes **every** available anchor separately, so a stored one can tell "this machine, one part changed" (a NIC swap, a reinstall) from "another machine". Concatenating anchors into one id would do the opposite: any one change would change everything. Weights: install id 3, platform UUID 3, board serial 2, product serial 2, permanent MAC 1, CPU signature 1 (identical CPUs share a signature, so it only confirms). Each hash is compared in constant time.

Stored format (112 bytes): magic `SLFP`, format version 1, an anchor bitmap, two reserved zero bytes, six 16-byte anchor hashes, and an 8-byte BLAKE3 checksum. `from_bytes` checks length, magic, version, reserved bytes, bitmap consistency, and checksum, in that order, and returns `Malformed` for anything wrong; the checksum catches storage corruption, not tampering.

```rust
use sys_lib::device::{self, Fingerprint, Verdict};

let now = device::fingerprint("com.example.app")?;
let stored = Fingerprint::from_bytes(&load_stored())?;
match now.compare(&stored).verdict {
    Verdict::Same => {}
    Verdict::LikelySame => { save(now.to_bytes()); }     // one part changed: update the record
    Verdict::Different => { /* treat as a new device */ }
    _ => {}
}
# fn load_stored() -> Vec<u8> { device::fingerprint("com.example.app").unwrap().to_bytes().to_vec() }
# fn save(_: [u8; 112]) {}
# Ok::<(), sys_lib::Error>(())
```

### When not to use it

- **Not authentication.** A device id is a claim, not proof. For proof of a device, use a key held in a TPM or Secure Enclave.
- **Not for tracking across applications or for advertising.** The application-specific key prevents correlation by design; using device identifiers that way breaks platform policies and, in the EU, likely needs consent under the ePrivacy Directive, Article 5(3).
- **Personal data.** Under the GDPR (Recital 30), an identifier from a user's device can identify a person: an application storing a device id needs a lawful basis, a retention period, and a way to delete it. This is a statement of the issue, not legal advice.
- **Not stable enough for licensing alone.** Reinstalls and board replacements happen; licensing should allow re-activation.

---

## `probe`: explicit roots for monitoring agents

**Since** 0.6 · PRB-01 · Linux

```rust
#[cfg(target_os = "linux")]
pub mod probe {
    pub struct Probe { /* root path, budgets */ }
    impl Probe {
        pub fn with_root(root: impl Into<PathBuf>) -> Result<Probe>;   // e.g. "/host", where the host's /proc and /sys are mounted
        pub fn memory(&self) -> Result<mem::Memory>;
        pub fn cpu_sampler(&self) -> Result<cpu::Sampler>;
        pub fn load(&self) -> Result<Option<cpu::Load>>;
        pub fn uptime(&self) -> Result<Duration>;
        pub fn disks(&self) -> Result<Vec<disk::Disk>>;
        pub fn disk_io(&self) -> Result<disk::IoSampler>;
        pub fn net(&self) -> Result<net::Sampler>;
        pub fn processes(&self) -> Result<host::Processes>;
    }
}
```

A monitoring agent in a container that has the host's `/proc` and `/sys` mounted (say at `/host`) reads the host's values through an explicit root. It is the feature gopsutil offers through `HOST_PROC` and `HOST_SYS` environment variables; sys-lib offers it only as a constructor in code, so no deployment can redirect a normal program's readings through its environment (threat T4). The Linux backend already reads through a root prefix (that is how fixture tests run), so the readings are the same code. Mounts and network counters are read through the host's process 1 (`/proc/1/mountinfo`, `/proc/1/net/dev`), which is the host's view when the agent shares the host's process namespace (`hostPID`); otherwise they are the agent's own. CPU, memory (without any cgroup limit), disk I/O, load, and uptime are system-wide; disk space is read through the root (D-065).

## Errors by variant

| Variant | Meaning | What the message tells the caller to do |
|---|---|---|
| `Unsupported { what, platform }` | this reading does not exist on this platform (load average on Windows is `Ok(None)` instead; temperatures on macOS are `Unsupported`) | treat the value as unknown |
| `PermissionDenied { what, hint }` | the OS refused (root-only DMI files, a sandbox, SELinux, seccomp) | run with the permission named in `hint`, or accept the value as unknown |
| `NotFound { what, source }` | a file or key that should exist does not (a container with a reduced `/proc`) | accept the value as unknown; `docs/PLATFORM-NOTES.md` lists known causes |
| `Malformed { what, source, line }` | the OS returned data that does not parse; the source and line, never the content | treat the value as unknown, and report it on the issue tracker with the file named (`summary()` output helps, from 0.5) |
| `TooLarge { what, limit }` | a file or a list exceeded its budget (1 MiB per file; 4 096 mounts or interfaces; 16 384 CPUs), or a caller buffer was too short | treat the value as unknown; when the machine is genuinely that large, raise the budget with the Tier 2 option (from 0.4) or pass a longer buffer, and report it |
| `TooSoon { elapsed, minimum }` | a sampler was asked again before any time passed on the OS clock | sample at least `minimum` apart |
| `Os { what, call, code }` | a system call failed, with its name and OS error code | the code says why; report it if unexpected |
| `Io { what, source }` | file I/O failed outside the cases above | treat the value as unknown; `source` says why; report it if unexpected |
| `InvalidInput { what, reason }` | invalid input from the caller (a guard name with a slash, an empty app id, an id string of the wrong length) | fix the input as `reason` says |
| `AlreadyRunning { pid, lock_file }` | the process guard found another copy holding the lock | stop the other copy (its pid when known) or use another lock directory |
| `RunDirUnavailable { path, source }` | the guard's lock directory cannot be created, written, or trusted | the message names the reason and the fix (`chmod 700`, a local directory, `allow_network_lock`) |
| `OpenFilesTooLow { have, need }` | the hard open-file limit is below the guard's minimum | raise `LimitNOFILE=` in the systemd unit, `ulimit -Hn`, or launchd's limits |
| `NoIdentifierSource { basis }` (feature `device-id`) | no usable identifier for the requested basis | see `device::report()` for the state of each source |

`Error::is_unavailable()` is true for `Unsupported`, `PermissionDenied`, and `NotFound`: the three that mean "this fact is not available here".