1=============================
2Per-task statistics interface
3=============================
4
5
6Taskstats is a netlink-based interface for sending per-task and
7per-process statistics from the kernel to userspace.
8
9Taskstats was designed for the following benefits:
10
11- efficiently provide statistics during lifetime of a task and on its exit
12- unified interface for multiple accounting subsystems
13- extensibility for use by future accounting patches
14
15Terminology
16-----------
17
18"pid", "tid" and "task" are used interchangeably and refer to the standard
19Linux task defined by struct task_struct.  per-pid stats are the same as
20per-task stats.
21
22"tgid", "process" and "thread group" are used interchangeably and refer to the
23tasks that share an mm_struct i.e. the traditional Unix process. Despite the
24use of tgid, there is no special treatment for the task that is thread group
25leader - a process is deemed alive as long as it has any task belonging to it.
26
27Usage
28-----
29
30To get statistics during a task's lifetime, userspace opens a unicast netlink
31socket (NETLINK_GENERIC family) and sends commands specifying a pid or a tgid.
32The response contains statistics for a task (if pid is specified) or the sum of
33statistics for all tasks of the process (if tgid is specified).
34
35To obtain statistics for tasks which are exiting, the userspace listener
36sends a register command and specifies a cpumask. Whenever a task exits on
37one of the cpus in the cpumask, its per-pid statistics are sent to the
38registered listener. Using cpumasks allows the data received by one listener
39to be limited and assists in flow control over the netlink interface and is
40explained in more detail below.
41
42If the exiting task is the last thread exiting its thread group,
43an additional record containing the per-tgid stats is also sent to userspace.
44The latter contains the sum of per-pid stats for all threads in the thread
45group, both past and present.
46
47getdelays.c is a simple utility demonstrating usage of the taskstats interface
48for reporting delay accounting statistics. Users can register cpumasks,
49send commands and process responses, listen for per-tid/tgid exit data,
50write the data received to a file and do basic flow control by increasing
51receive buffer sizes.
52
53Interface
54---------
55
56The user-kernel interface is encapsulated in include/linux/taskstats.h
57
58To avoid this documentation becoming obsolete as the interface evolves, only
59an outline of the current version is given. taskstats.h always overrides the
60description here.
61
62struct taskstats is the common accounting structure for both per-pid and
63per-tgid data. It is versioned and can be extended by each accounting subsystem
64that is added to the kernel. The fields and their semantics are defined in the
65taskstats.h file.
66
67The data exchanged between user and kernel space is a netlink message belonging
68to the NETLINK_GENERIC family and using the netlink attributes interface.
69The messages are in the format::
70
71    +----------+- - -+-------------+-------------------+
72    | nlmsghdr | Pad |  genlmsghdr | taskstats payload |
73    +----------+- - -+-------------+-------------------+
74
75
76The taskstats payload is one of the following three kinds:
77
781. Commands: Sent from user to kernel. Commands to get data on
79a pid/tgid consist of one attribute, of type TASKSTATS_CMD_ATTR_PID/TGID,
80containing a u32 pid or tgid in the attribute payload. The pid/tgid denotes
81the task/process for which userspace wants statistics.
82
83Commands to register/deregister interest in exit data from a set of cpus
84consist of one attribute, of type
85TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK and contain a cpumask in the
86attribute payload. The cpumask is specified as an ascii string of
87comma-separated cpu ranges e.g. to listen to exit data from cpus 1,2,3,5,7,8
88the cpumask would be "1-3,5,7-8". If userspace forgets to deregister interest
89in cpus before closing the listening socket, the kernel cleans up its interest
90set over time. However, for the sake of efficiency, an explicit deregistration
91is advisable.
92
932. Response for a command: sent from the kernel in response to a userspace
94command. The payload is a series of three attributes of type:
95
96a) TASKSTATS_TYPE_AGGR_PID/TGID : attribute containing no payload but indicates
97a pid/tgid will be followed by some stats.
98
99b) TASKSTATS_TYPE_PID/TGID: attribute whose payload is the pid/tgid whose stats
100are being returned.
101
102c) TASKSTATS_TYPE_STATS: attribute with a struct taskstats as payload. The
103same structure is used for both per-pid and per-tgid stats.
104
1053. New message sent by kernel whenever a task exits. The payload consists of a
106   series of attributes of the following type:
107
108a) TASKSTATS_TYPE_AGGR_PID: indicates next two attributes will be pid+stats
109b) TASKSTATS_TYPE_PID: contains exiting task's pid
110c) TASKSTATS_TYPE_STATS: contains the exiting task's per-pid stats
111d) TASKSTATS_TYPE_AGGR_TGID: indicates next two attributes will be tgid+stats
112e) TASKSTATS_TYPE_TGID: contains tgid of process to which task belongs
113f) TASKSTATS_TYPE_STATS: contains the per-tgid stats for exiting task's process
114
115
116per-tgid stats
117--------------
118
119Taskstats provides per-process stats, in addition to per-task stats, since
120resource management is often done at a process granularity and aggregating task
121stats in userspace alone is inefficient and potentially inaccurate (due to lack
122of atomicity).
123
124However, maintaining per-process, in addition to per-task stats, within the
125kernel has space and time overheads. To address this, the taskstats code
126accumulates each exiting task's statistics into a process-wide data structure.
127When the last task of a process exits, the process level data accumulated also
128gets sent to userspace (along with the per-task data).
129
130When a user queries to get per-tgid data, the sum of all other live threads in
131the group is added up and added to the accumulated total for previously exited
132threads of the same thread group.
133
134Extending taskstats
135-------------------
136
137There are two ways to extend the taskstats interface to export more
138per-task/process stats as patches to collect them get added to the kernel
139in future:
140
1411. Adding more fields to the end of the existing struct taskstats. Backward
142   compatibility is ensured by the version number within the
143   structure. Userspace will use only the fields of the struct that correspond
144   to the version its using.
145
1462. Defining separate statistic structs and using the netlink attributes
147   interface to return them. Since userspace processes each netlink attribute
148   independently, it can always ignore attributes whose type it does not
149   understand (because it is using an older version of the interface).
150
151
152Choosing between 1. and 2. is a matter of trading off flexibility and
153overhead. If only a few fields need to be added, then 1. is the preferable
154path since the kernel and userspace don't need to incur the overhead of
155processing new netlink attributes. But if the new fields expand the existing
156struct too much, requiring disparate userspace accounting utilities to
157unnecessarily receive large structures whose fields are of no interest, then
158extending the attributes structure would be worthwhile.
159
160Flow control for taskstats
161--------------------------
162
163When the rate of task exits becomes large, a listener may not be able to keep
164up with the kernel's rate of sending per-tid/tgid exit data leading to data
165loss. This possibility gets compounded when the taskstats structure gets
166extended and the number of cpus grows large.
167
168To avoid losing statistics, userspace should do one or more of the following:
169
170- increase the receive buffer sizes for the netlink sockets opened by
171  listeners to receive exit data.
172
173- create more listeners and reduce the number of cpus being listened to by
174  each listener. In the extreme case, there could be one listener for each cpu.
175  Users may also consider setting the cpu affinity of the listener to the subset
176  of cpus to which it listens, especially if they are listening to just one cpu.
177
178Despite these measures, if the userspace receives ENOBUFS error messages
179indicated overflow of receive buffers, it should take measures to handle the
180loss of data.
181