blob: 3207717a0d1ac4c6e251f091e1f7d2dda3056aef [file] [log] [blame]
Jon Mastersc850ed32015-04-10 14:57:46 -04001Introduction:
2-------------
3
4The tracer hwlat_detector is a special purpose tracer that is used to
5detect large system latencies induced by the behavior of certain underlying
6hardware or firmware, independent of Linux itself. The code was developed
7originally to detect SMIs (System Management Interrupts) on x86 systems,
8however there is nothing x86 specific about this patchset. It was
9originally written for use by the "RT" patch since the Real Time
10kernel is highly latency sensitive.
11
12SMIs are not serviced by the Linux kernel, which means that it does not
13even know that they are occuring. SMIs are instead set up by BIOS code
14and are serviced by BIOS code, usually for "critical" events such as
15management of thermal sensors and fans. Sometimes though, SMIs are used for
16other tasks and those tasks can spend an inordinate amount of time in the
17handler (sometimes measured in milliseconds). Obviously this is a problem if
18you are trying to keep event service latencies down in the microsecond range.
19
20The hardware latency detector works by hogging one of the cpus for configurable
21amounts of time (with interrupts disabled), polling the CPU Time Stamp Counter
22for some period, then looking for gaps in the TSC data. Any gap indicates a
23time when the polling was interrupted and since the interrupts are disabled,
24the only thing that could do that would be an SMI or other hardware hiccup
25(or an NMI, but those can be tracked).
26
27Note that the hwlat detector should *NEVER* be used in a production environment.
28It is intended to be run manually to determine if the hardware platform has a
29problem with long system firmware service routines.
30
31Usage:
32------
33
34Write the ASCII text "hwlat" into the current_tracer file of the tracing system
35(mounted at /sys/kernel/tracing or /sys/kernel/tracing). It is possible to
36redefine the threshold in microseconds (us) above which latency spikes will
37be taken into account.
38
39Example:
40
41 # echo hwlat > /sys/kernel/tracing/current_tracer
42 # echo 100 > /sys/kernel/tracing/tracing_thresh
43
44The /sys/kernel/tracing/hwlat_detector interface contains the following files:
45
46width - time period to sample with CPUs held (usecs)
47 must be less than the total window size (enforced)
48window - total period of sampling, width being inside (usecs)
49
50By default the width is set to 500,000 and window to 1,000,000, meaning that
51for every 1,000,000 usecs (1s) the hwlat detector will spin for 500,000 usecs
52(0.5s). If tracing_thresh contains zero when hwlat tracer is enabled, it will
53change to a default of 10 usecs. If any latencies that exceed the threshold is
54observed then the data will be written to the tracing ring buffer.
55
56The minimum sleep time between periods is 1 millisecond. Even if width
57is less than 1 millisecond apart from window, to allow the system to not
58be totally starved.
59
60If tracing_thresh was zero when hwlat detector was started, it will be set
61back to zero if another tracer is loaded. Note, the last value in
62tracing_thresh that hwlat detector had will be saved and this value will
63be restored in tracing_thresh if it is still zero when hwlat detector is
64started again.
65
66The following tracing directory files are used by the hwlat_detector:
67
68in /sys/kernel/tracing:
69
70 tracing_threshold - minimum latency value to be considered (usecs)
71 tracing_max_latency - maximum hardware latency actually observed (usecs)
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -040072 tracing_cpumask - the CPUs to move the hwlat thread across
Jon Mastersc850ed32015-04-10 14:57:46 -040073 hwlat_detector/width - specified amount of time to spin within window (usecs)
74 hwlat_detector/window - amount of time between (width) runs (usecs)
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -040075
76The hwlat detector's kernel thread will migrate across each CPU specified in
77tracing_cpumask between each window. To limit the migration, either modify
78tracing_cpumask, or modify the hwlat kernel thread (named [hwlatd]) CPU
79affinity directly, and the migration will stop.