xref: /linux/Documentation/trace/hwlat_detector.rst (revision 8fa826b7344d6752f5cfd72380d9fe7bd8c6b928)
1=========================
2Hardware Latency Detector
3=========================
4
5Introduction
6-------------
7
8The tracer hwlat_detector is a special purpose tracer that is used to
9detect large system latencies induced by the behavior of certain underlying
10hardware or firmware, independent of Linux itself. The code was developed
11originally to detect SMIs (System Management Interrupts) on x86 systems,
12however there is nothing x86 specific about this patchset. It was
13originally written for use by the "RT" patch since the Real Time
14kernel is highly latency sensitive.
15
16SMIs are not serviced by the Linux kernel, which means that it does not
17even know that they are occuring. SMIs are instead set up by BIOS code
18and are serviced by BIOS code, usually for "critical" events such as
19management of thermal sensors and fans. Sometimes though, SMIs are used for
20other tasks and those tasks can spend an inordinate amount of time in the
21handler (sometimes measured in milliseconds). Obviously this is a problem if
22you are trying to keep event service latencies down in the microsecond range.
23
24The hardware latency detector works by hogging one of the cpus for configurable
25amounts of time (with interrupts disabled), polling the CPU Time Stamp Counter
26for some period, then looking for gaps in the TSC data. Any gap indicates a
27time when the polling was interrupted and since the interrupts are disabled,
28the only thing that could do that would be an SMI or other hardware hiccup
29(or an NMI, but those can be tracked).
30
31Note that the hwlat detector should *NEVER* be used in a production environment.
32It is intended to be run manually to determine if the hardware platform has a
33problem with long system firmware service routines.
34
35Usage
36------
37
38Write the ASCII text "hwlat" into the current_tracer file of the tracing system
39(mounted at /sys/kernel/tracing or /sys/kernel/tracing). It is possible to
40redefine the threshold in microseconds (us) above which latency spikes will
41be taken into account.
42
43Example::
44
45	# echo hwlat > /sys/kernel/tracing/current_tracer
46	# echo 100 > /sys/kernel/tracing/tracing_thresh
47
48The /sys/kernel/tracing/hwlat_detector interface contains the following files:
49
50  - width - time period to sample with CPUs held (usecs)
51            must be less than the total window size (enforced)
52  - window - total period of sampling, width being inside (usecs)
53
54By default the width is set to 500,000 and window to 1,000,000, meaning that
55for every 1,000,000 usecs (1s) the hwlat detector will spin for 500,000 usecs
56(0.5s). If tracing_thresh contains zero when hwlat tracer is enabled, it will
57change to a default of 10 usecs. If any latencies that exceed the threshold is
58observed then the data will be written to the tracing ring buffer.
59
60The minimum sleep time between periods is 1 millisecond. Even if width
61is less than 1 millisecond apart from window, to allow the system to not
62be totally starved.
63
64If tracing_thresh was zero when hwlat detector was started, it will be set
65back to zero if another tracer is loaded. Note, the last value in
66tracing_thresh that hwlat detector had will be saved and this value will
67be restored in tracing_thresh if it is still zero when hwlat detector is
68started again.
69
70The following tracing directory files are used by the hwlat_detector:
71
72in /sys/kernel/tracing:
73
74 - tracing_threshold	- minimum latency value to be considered (usecs)
75 - tracing_max_latency	- maximum hardware latency actually observed (usecs)
76 - tracing_cpumask	- the CPUs to move the hwlat thread across
77 - hwlat_detector/width	- specified amount of time to spin within window (usecs)
78 - hwlat_detector/window	- amount of time between (width) runs (usecs)
79 - hwlat_detector/mode	- the thread mode
80
81By default, the hwlat detector's kernel thread will migrate across each CPU
82specified in cpumask at the beginning of a new window, in a round-robin
83fashion. This behavior can be changed by changing the thread mode,
84the available options are:
85
86 - none:        do not force migration
87 - round-robin: migrate across each CPU specified in cpumask [default]
88