| | 1 | .\" Copyright (c) 2017, Apple Inc. All rights reserved. |
| | 2 | .\" |
| | 3 | .Dd March 1, 2017 |
| | 4 | .Dt MONOTONIC 9 |
| | 5 | .Os Darwin |
| | 6 | .\" |
| | 7 | .Sh NAME |
| | 8 | .Nm monotonic |
| | 9 | .Nd performance counter access system |
| | 10 | .\" |
| | 11 | .Sh DESCRIPTION |
| | 12 | .Nm |
| | 13 | allows kernel and user space clients to configure and read hardware performance |
| | 14 | counters. The hardware counters can be virtualized to count per-thread and |
| | 15 | per-process. |
| | 16 | .Nm |
| | 17 | is split into three major layers: |
| | 18 | .Bl -dash |
| | 19 | .It |
| | 20 | The machine-dependent implementations manipulate hardware registers to configure |
| | 21 | and access the counters. This layer provides a machine-independent interface |
| | 22 | that can be used by the next layer. |
| | 23 | .It |
| | 24 | A set of hooks and kernel routines manage the counters and provide higher-level |
| | 25 | abstractions, like 64-bit counters and counting only events that occurred on a |
| | 26 | thread or in a process. |
| | 27 | .It |
| | 28 | A user space interface that is presented as device nodes under |
| | 29 | .Pa /dev/monotonic . |
| | 30 | See |
| | 31 | .Xr monotonic 4 . |
| | 32 | Mach thread and task ports are used for the per-thread and per-process counts, |
| | 33 | with special inspection routines. Some counter values are also reflected into |
| | 34 | .Fn getrusage , |
| | 35 | for use after a process has exited. See |
| | 36 | .Xr getrusage 2 . |
| | 37 | .El |
| | 38 | .Pp |
| | 39 | .\".Sh DIAGNOSTICS |
| | 40 | .\" |
| | 41 | .Sh SEE ALSO |
| | 42 | .Xr count 1 , |
| | 43 | .Xr mperf 1 , |
| | 44 | .Xr easyperf 1 , |
| | 45 | .Xr getrusage 2 , |
| | 46 | .Xr monotonic 4 , |
| | 47 | .Xr perf 1 |
| | 48 | .\" |
| | 49 | .Sh HISTORY |
| | 50 | .Nm |
| | 51 | replaces the kernel performance counter system, kpc. For the time being, |
| | 52 | .Nm |
| | 53 | backs portions of the existing kpc |
| | 54 | .Fn sysctl |
| | 55 | interface. Prior to kpc, the AppleProfileFamily kernel extensions provided |
| | 56 | performance counter interfaces. The kernel extensions themselves expanded upon |
| | 57 | functionality provided for PowerPC by CHUD. |
| | 58 | .\".Sh CAVEATS |
| | 59 | .\" |
| | 60 | .Sh SECURITY CONSIDERATIONS |
| | 61 | .Pp |
| | 62 | .Pp |
| | 63 | .Bl -dash |
| | 64 | .It |
| | 65 | Hardware performance counters are an ideal tool for side-channel attacks. By |
| | 66 | observing how the counters are affected by an otherwise opaque process, an |
| | 67 | attacker can obtain sensitive data or key material. |
| | 68 | .Pp |
| | 69 | For this reason, the hardware performance counters cannot be queried directly |
| | 70 | from user space. Instead, all processes, including those owned by root, can |
| | 71 | only query the thread and process counters if they have the corresponding Mach |
| | 72 | thread or task port. |
| | 73 | .It |
| | 74 | When used in sampling mode, hardware performance counters can induce interrupt |
| | 75 | storms that translate to denial-of-service attacks on a system. Even a careless |
| | 76 | user can stumble over this issue, since reasonable periods for some events are |
| | 77 | far too aggressive for others. |
| | 78 | .Pp |
| | 79 | If a hardware performance counter takes too many interrupts in a short amount of |
| | 80 | time, it will be disabled. |
| | 81 | .El |
projects / apple / xnu.git / blame_incremental