1================
2Delay accounting
3================
4
5Tasks encounter delays in execution when they wait
6for some kernel resource to become available e.g. a
7runnable task may wait for a free CPU to run on.
8
9The per-task delay accounting functionality measures
10the delays experienced by a task while
11
12a) waiting for a CPU (while being runnable)
13b) completion of synchronous block I/O initiated by the task
14c) swapping in pages
15d) memory reclaim
16e) thrashing page cache
17f) direct compact
18
19and makes these statistics available to userspace through
20the taskstats interface.
21
22Such delays provide feedback for setting a task's cpu priority,
23io priority and rss limit values appropriately. Long delays for
24important tasks could be a trigger for raising its corresponding priority.
25
26The functionality, through its use of the taskstats interface, also provides
27delay statistics aggregated for all tasks (or threads) belonging to a
28thread group (corresponding to a traditional Unix process). This is a commonly
29needed aggregation that is more efficiently done by the kernel.
30
31Userspace utilities, particularly resource management applications, can also
32aggregate delay statistics into arbitrary groups. To enable this, delay
33statistics of a task are available both during its lifetime as well as on its
34exit, ensuring continuous and complete monitoring can be done.
35
36
37Interface
38---------
39
40Delay accounting uses the taskstats interface which is described
41in detail in a separate document in this directory. Taskstats returns a
42generic data structure to userspace corresponding to per-pid and per-tgid
43statistics. The delay accounting functionality populates specific fields of
44this structure. See
45
46     include/uapi/linux/taskstats.h
47
48for a description of the fields pertaining to delay accounting.
49It will generally be in the form of counters returning the cumulative
50delay seen for cpu, sync block I/O, swapin, memory reclaim, thrash page
51cache, direct compact etc.
52
53Taking the difference of two successive readings of a given
54counter (say cpu_delay_total) for a task will give the delay
55experienced by the task waiting for the corresponding resource
56in that interval.
57
58When a task exits, records containing the per-task statistics
59are sent to userspace without requiring a command. If it is the last exiting
60task of a thread group, the per-tgid statistics are also sent. More details
61are given in the taskstats interface description.
62
63The getdelays.c userspace utility in tools/accounting directory allows simple
64commands to be run and the corresponding delay statistics to be displayed. It
65also serves as an example of using the taskstats interface.
66
67Usage
68-----
69
70Compile the kernel with::
71
72	CONFIG_TASK_DELAY_ACCT=y
73	CONFIG_TASKSTATS=y
74
75Delay accounting is disabled by default at boot up.
76To enable, add::
77
78   delayacct
79
80to the kernel boot options. The rest of the instructions below assume this has
81been done. Alternatively, use sysctl kernel.task_delayacct to switch the state
82at runtime. Note however that only tasks started after enabling it will have
83delayacct information.
84
85After the system has booted up, use a utility
86similar to  getdelays.c to access the delays
87seen by a given task or a task group (tgid).
88The utility also allows a given command to be
89executed and the corresponding delays to be
90seen.
91
92General format of the getdelays command::
93
94	getdelays [-dilv] [-t tgid] [-p pid]
95
96Get delays, since system boot, for pid 10::
97
98	# ./getdelays -d -p 10
99	(output similar to next case)
100
101Get sum of delays, since system boot, for all pids with tgid 5::
102
103	# ./getdelays -d -t 5
104	print delayacct stats ON
105	TGID	5
106
107
108	CPU             count     real total  virtual total    delay total  delay average
109	                    8        7000000        6872122        3382277          0.423ms
110	IO              count    delay total  delay average
111		            0              0              0ms
112	SWAP            count    delay total  delay average
113	                    0              0              0ms
114	RECLAIM         count    delay total  delay average
115		            0              0              0ms
116	THRASHING       count    delay total  delay average
117	                    0              0              0ms
118	COMPACT         count    delay total  delay average
119	                    0              0              0ms
120
121Get IO accounting for pid 1, it works only with -p::
122
123	# ./getdelays -i -p 1
124	printing IO accounting
125	linuxrc: read=65536, write=0, cancelled_write=0
126
127The above command can be used with -v to get more debug information.
128