summaryrefslogtreecommitdiffstats
path: root/Documentation/accounting
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/accounting')
-rw-r--r--Documentation/accounting/cgroupstats.rst31
-rw-r--r--Documentation/accounting/delay-accounting.rst130
-rw-r--r--Documentation/accounting/index.rst14
-rw-r--r--Documentation/accounting/psi.rst188
-rw-r--r--Documentation/accounting/taskstats-struct.rst199
-rw-r--r--Documentation/accounting/taskstats.rst180
6 files changed, 742 insertions, 0 deletions
diff --git a/Documentation/accounting/cgroupstats.rst b/Documentation/accounting/cgroupstats.rst
new file mode 100644
index 000000000..85186e7d4
--- /dev/null
+++ b/Documentation/accounting/cgroupstats.rst
@@ -0,0 +1,31 @@
+==================
+Control Groupstats
+==================
+
+Control Groupstats is inspired by the discussion at
+https://lore.kernel.org/r/461CF883.2030308@sw.ru and implements per cgroup statistics as
+suggested by Andrew Morton in https://lore.kernel.org/r/20070411114927.1277d7c9.akpm@linux-foundation.org.
+
+Per cgroup statistics infrastructure re-uses code from the taskstats
+interface. A new set of cgroup operations are registered with commands
+and attributes specific to cgroups. It should be very easy to
+extend per cgroup statistics, by adding members to the cgroupstats
+structure.
+
+The current model for cgroupstats is a pull, a push model (to post
+statistics on interesting events), should be very easy to add. Currently
+user space requests for statistics by passing the cgroup path.
+Statistics about the state of all the tasks in the cgroup is returned to
+user space.
+
+NOTE: We currently rely on delay accounting for extracting information
+about tasks blocked on I/O. If CONFIG_TASK_DELAY_ACCT is disabled, this
+information will not be available.
+
+To extract cgroup statistics a utility very similar to getdelays.c
+has been developed, the sample output of the utility is shown below::
+
+ ~/balbir/cgroupstats # ./getdelays -C "/sys/fs/cgroup/a"
+ sleeping 1, blocked 0, running 1, stopped 0, uninterruptible 0
+ ~/balbir/cgroupstats # ./getdelays -C "/sys/fs/cgroup"
+ sleeping 155, blocked 0, running 1, stopped 0, uninterruptible 2
diff --git a/Documentation/accounting/delay-accounting.rst b/Documentation/accounting/delay-accounting.rst
new file mode 100644
index 000000000..7103b62ba
--- /dev/null
+++ b/Documentation/accounting/delay-accounting.rst
@@ -0,0 +1,130 @@
+================
+Delay accounting
+================
+
+Tasks encounter delays in execution when they wait
+for some kernel resource to become available e.g. a
+runnable task may wait for a free CPU to run on.
+
+The per-task delay accounting functionality measures
+the delays experienced by a task while
+
+a) waiting for a CPU (while being runnable)
+b) completion of synchronous block I/O initiated by the task
+c) swapping in pages
+d) memory reclaim
+e) thrashing
+f) direct compact
+g) write-protect copy
+
+and makes these statistics available to userspace through
+the taskstats interface.
+
+Such delays provide feedback for setting a task's cpu priority,
+io priority and rss limit values appropriately. Long delays for
+important tasks could be a trigger for raising its corresponding priority.
+
+The functionality, through its use of the taskstats interface, also provides
+delay statistics aggregated for all tasks (or threads) belonging to a
+thread group (corresponding to a traditional Unix process). This is a commonly
+needed aggregation that is more efficiently done by the kernel.
+
+Userspace utilities, particularly resource management applications, can also
+aggregate delay statistics into arbitrary groups. To enable this, delay
+statistics of a task are available both during its lifetime as well as on its
+exit, ensuring continuous and complete monitoring can be done.
+
+
+Interface
+---------
+
+Delay accounting uses the taskstats interface which is described
+in detail in a separate document in this directory. Taskstats returns a
+generic data structure to userspace corresponding to per-pid and per-tgid
+statistics. The delay accounting functionality populates specific fields of
+this structure. See
+
+ include/uapi/linux/taskstats.h
+
+for a description of the fields pertaining to delay accounting.
+It will generally be in the form of counters returning the cumulative
+delay seen for cpu, sync block I/O, swapin, memory reclaim, thrash page
+cache, direct compact, write-protect copy etc.
+
+Taking the difference of two successive readings of a given
+counter (say cpu_delay_total) for a task will give the delay
+experienced by the task waiting for the corresponding resource
+in that interval.
+
+When a task exits, records containing the per-task statistics
+are sent to userspace without requiring a command. If it is the last exiting
+task of a thread group, the per-tgid statistics are also sent. More details
+are given in the taskstats interface description.
+
+The getdelays.c userspace utility in tools/accounting directory allows simple
+commands to be run and the corresponding delay statistics to be displayed. It
+also serves as an example of using the taskstats interface.
+
+Usage
+-----
+
+Compile the kernel with::
+
+ CONFIG_TASK_DELAY_ACCT=y
+ CONFIG_TASKSTATS=y
+
+Delay accounting is disabled by default at boot up.
+To enable, add::
+
+ delayacct
+
+to the kernel boot options. The rest of the instructions below assume this has
+been done. Alternatively, use sysctl kernel.task_delayacct to switch the state
+at runtime. Note however that only tasks started after enabling it will have
+delayacct information.
+
+After the system has booted up, use a utility
+similar to getdelays.c to access the delays
+seen by a given task or a task group (tgid).
+The utility also allows a given command to be
+executed and the corresponding delays to be
+seen.
+
+General format of the getdelays command::
+
+ getdelays [-dilv] [-t tgid] [-p pid]
+
+Get delays, since system boot, for pid 10::
+
+ # ./getdelays -d -p 10
+ (output similar to next case)
+
+Get sum of delays, since system boot, for all pids with tgid 5::
+
+ # ./getdelays -d -t 5
+ print delayacct stats ON
+ TGID 5
+
+
+ CPU count real total virtual total delay total delay average
+ 8 7000000 6872122 3382277 0.423ms
+ IO count delay total delay average
+ 0 0 0ms
+ SWAP count delay total delay average
+ 0 0 0ms
+ RECLAIM count delay total delay average
+ 0 0 0ms
+ THRASHING count delay total delay average
+ 0 0 0ms
+ COMPACT count delay total delay average
+ 0 0 0ms
+ WPCOPY count delay total delay average
+ 0 0 0ms
+
+Get IO accounting for pid 1, it works only with -p::
+
+ # ./getdelays -i -p 1
+ printing IO accounting
+ linuxrc: read=65536, write=0, cancelled_write=0
+
+The above command can be used with -v to get more debug information.
diff --git a/Documentation/accounting/index.rst b/Documentation/accounting/index.rst
new file mode 100644
index 000000000..9369d8bf3
--- /dev/null
+++ b/Documentation/accounting/index.rst
@@ -0,0 +1,14 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+==========
+Accounting
+==========
+
+.. toctree::
+ :maxdepth: 1
+
+ cgroupstats
+ delay-accounting
+ psi
+ taskstats
+ taskstats-struct
diff --git a/Documentation/accounting/psi.rst b/Documentation/accounting/psi.rst
new file mode 100644
index 000000000..df6062eb3
--- /dev/null
+++ b/Documentation/accounting/psi.rst
@@ -0,0 +1,188 @@
+.. _psi:
+
+================================
+PSI - Pressure Stall Information
+================================
+
+:Date: April, 2018
+:Author: Johannes Weiner <hannes@cmpxchg.org>
+
+When CPU, memory or IO devices are contended, workloads experience
+latency spikes, throughput losses, and run the risk of OOM kills.
+
+Without an accurate measure of such contention, users are forced to
+either play it safe and under-utilize their hardware resources, or
+roll the dice and frequently suffer the disruptions resulting from
+excessive overcommit.
+
+The psi feature identifies and quantifies the disruptions caused by
+such resource crunches and the time impact it has on complex workloads
+or even entire systems.
+
+Having an accurate measure of productivity losses caused by resource
+scarcity aids users in sizing workloads to hardware--or provisioning
+hardware according to workload demand.
+
+As psi aggregates this information in realtime, systems can be managed
+dynamically using techniques such as load shedding, migrating jobs to
+other systems or data centers, or strategically pausing or killing low
+priority or restartable batch jobs.
+
+This allows maximizing hardware utilization without sacrificing
+workload health or risking major disruptions such as OOM kills.
+
+Pressure interface
+==================
+
+Pressure information for each resource is exported through the
+respective file in /proc/pressure/ -- cpu, memory, and io.
+
+The format is as such::
+
+ some avg10=0.00 avg60=0.00 avg300=0.00 total=0
+ full avg10=0.00 avg60=0.00 avg300=0.00 total=0
+
+The "some" line indicates the share of time in which at least some
+tasks are stalled on a given resource.
+
+The "full" line indicates the share of time in which all non-idle
+tasks are stalled on a given resource simultaneously. In this state
+actual CPU cycles are going to waste, and a workload that spends
+extended time in this state is considered to be thrashing. This has
+severe impact on performance, and it's useful to distinguish this
+situation from a state where some tasks are stalled but the CPU is
+still doing productive work. As such, time spent in this subset of the
+stall state is tracked separately and exported in the "full" averages.
+
+CPU full is undefined at the system level, but has been reported
+since 5.13, so it is set to zero for backward compatibility.
+
+The ratios (in %) are tracked as recent trends over ten, sixty, and
+three hundred second windows, which gives insight into short term events
+as well as medium and long term trends. The total absolute stall time
+(in us) is tracked and exported as well, to allow detection of latency
+spikes which wouldn't necessarily make a dent in the time averages,
+or to average trends over custom time frames.
+
+Monitoring for pressure thresholds
+==================================
+
+Users can register triggers and use poll() to be woken up when resource
+pressure exceeds certain thresholds.
+
+A trigger describes the maximum cumulative stall time over a specific
+time window, e.g. 100ms of total stall time within any 500ms window to
+generate a wakeup event.
+
+To register a trigger user has to open psi interface file under
+/proc/pressure/ representing the resource to be monitored and write the
+desired threshold and time window. The open file descriptor should be
+used to wait for trigger events using select(), poll() or epoll().
+The following format is used::
+
+ <some|full> <stall amount in us> <time window in us>
+
+For example writing "some 150000 1000000" into /proc/pressure/memory
+would add 150ms threshold for partial memory stall measured within
+1sec time window. Writing "full 50000 1000000" into /proc/pressure/io
+would add 50ms threshold for full io stall measured within 1sec time window.
+
+Triggers can be set on more than one psi metric and more than one trigger
+for the same psi metric can be specified. However for each trigger a separate
+file descriptor is required to be able to poll it separately from others,
+therefore for each trigger a separate open() syscall should be made even
+when opening the same psi interface file. Write operations to a file descriptor
+with an already existing psi trigger will fail with EBUSY.
+
+Monitors activate only when system enters stall state for the monitored
+psi metric and deactivates upon exit from the stall state. While system is
+in the stall state psi signal growth is monitored at a rate of 10 times per
+tracking window.
+
+The kernel accepts window sizes ranging from 500ms to 10s, therefore min
+monitoring update interval is 50ms and max is 1s. Min limit is set to
+prevent overly frequent polling. Max limit is chosen as a high enough number
+after which monitors are most likely not needed and psi averages can be used
+instead.
+
+Unprivileged users can also create monitors, with the only limitation that the
+window size must be a multiple of 2s, in order to prevent excessive resource
+usage.
+
+When activated, psi monitor stays active for at least the duration of one
+tracking window to avoid repeated activations/deactivations when system is
+bouncing in and out of the stall state.
+
+Notifications to the userspace are rate-limited to one per tracking window.
+
+The trigger will de-register when the file descriptor used to define the
+trigger is closed.
+
+Userspace monitor usage example
+===============================
+
+::
+
+ #include <errno.h>
+ #include <fcntl.h>
+ #include <stdio.h>
+ #include <poll.h>
+ #include <string.h>
+ #include <unistd.h>
+
+ /*
+ * Monitor memory partial stall with 1s tracking window size
+ * and 150ms threshold.
+ */
+ int main() {
+ const char trig[] = "some 150000 1000000";
+ struct pollfd fds;
+ int n;
+
+ fds.fd = open("/proc/pressure/memory", O_RDWR | O_NONBLOCK);
+ if (fds.fd < 0) {
+ printf("/proc/pressure/memory open error: %s\n",
+ strerror(errno));
+ return 1;
+ }
+ fds.events = POLLPRI;
+
+ if (write(fds.fd, trig, strlen(trig) + 1) < 0) {
+ printf("/proc/pressure/memory write error: %s\n",
+ strerror(errno));
+ return 1;
+ }
+
+ printf("waiting for events...\n");
+ while (1) {
+ n = poll(&fds, 1, -1);
+ if (n < 0) {
+ printf("poll error: %s\n", strerror(errno));
+ return 1;
+ }
+ if (fds.revents & POLLERR) {
+ printf("got POLLERR, event source is gone\n");
+ return 0;
+ }
+ if (fds.revents & POLLPRI) {
+ printf("event triggered!\n");
+ } else {
+ printf("unknown event received: 0x%x\n", fds.revents);
+ return 1;
+ }
+ }
+
+ return 0;
+ }
+
+Cgroup2 interface
+=================
+
+In a system with a CONFIG_CGROUP=y kernel and the cgroup2 filesystem
+mounted, pressure stall information is also tracked for tasks grouped
+into cgroups. Each subdirectory in the cgroupfs mountpoint contains
+cpu.pressure, memory.pressure, and io.pressure files; the format is
+the same as the /proc/pressure/ files.
+
+Per-cgroup psi monitors can be specified and used the same way as
+system-wide ones.
diff --git a/Documentation/accounting/taskstats-struct.rst b/Documentation/accounting/taskstats-struct.rst
new file mode 100644
index 000000000..ca90fd489
--- /dev/null
+++ b/Documentation/accounting/taskstats-struct.rst
@@ -0,0 +1,199 @@
+====================
+The struct taskstats
+====================
+
+This document contains an explanation of the struct taskstats fields.
+
+There are three different groups of fields in the struct taskstats:
+
+1) Common and basic accounting fields
+ If CONFIG_TASKSTATS is set, the taskstats interface is enabled and
+ the common fields and basic accounting fields are collected for
+ delivery at do_exit() of a task.
+2) Delay accounting fields
+ These fields are placed between::
+
+ /* Delay accounting fields start */
+
+ and::
+
+ /* Delay accounting fields end */
+
+ Their values are collected if CONFIG_TASK_DELAY_ACCT is set.
+3) Extended accounting fields
+ These fields are placed between::
+
+ /* Extended accounting fields start */
+
+ and::
+
+ /* Extended accounting fields end */
+
+ Their values are collected if CONFIG_TASK_XACCT is set.
+
+4) Per-task and per-thread context switch count statistics
+
+5) Time accounting for SMT machines
+
+6) Extended delay accounting fields for memory reclaim
+
+Future extension should add fields to the end of the taskstats struct, and
+should not change the relative position of each field within the struct.
+
+::
+
+ struct taskstats {
+
+1) Common and basic accounting fields::
+
+ /* The version number of this struct. This field is always set to
+ * TAKSTATS_VERSION, which is defined in <linux/taskstats.h>.
+ * Each time the struct is changed, the value should be incremented.
+ */
+ __u16 version;
+
+ /* The exit code of a task. */
+ __u32 ac_exitcode; /* Exit status */
+
+ /* The accounting flags of a task as defined in <linux/acct.h>
+ * Defined values are AFORK, ASU, ACOMPAT, ACORE, and AXSIG.
+ */
+ __u8 ac_flag; /* Record flags */
+
+ /* The value of task_nice() of a task. */
+ __u8 ac_nice; /* task_nice */
+
+ /* The name of the command that started this task. */
+ char ac_comm[TS_COMM_LEN]; /* Command name */
+
+ /* The scheduling discipline as set in task->policy field. */
+ __u8 ac_sched; /* Scheduling discipline */
+
+ __u8 ac_pad[3];
+ __u32 ac_uid; /* User ID */
+ __u32 ac_gid; /* Group ID */
+ __u32 ac_pid; /* Process ID */
+ __u32 ac_ppid; /* Parent process ID */
+
+ /* The time when a task begins, in [secs] since 1970. */
+ __u32 ac_btime; /* Begin time [sec since 1970] */
+
+ /* The elapsed time of a task, in [usec]. */
+ __u64 ac_etime; /* Elapsed time [usec] */
+
+ /* The user CPU time of a task, in [usec]. */
+ __u64 ac_utime; /* User CPU time [usec] */
+
+ /* The system CPU time of a task, in [usec]. */
+ __u64 ac_stime; /* System CPU time [usec] */
+
+ /* The minor page fault count of a task, as set in task->min_flt. */
+ __u64 ac_minflt; /* Minor Page Fault Count */
+
+ /* The major page fault count of a task, as set in task->maj_flt. */
+ __u64 ac_majflt; /* Major Page Fault Count */
+
+
+2) Delay accounting fields::
+
+ /* Delay accounting fields start
+ *
+ * All values, until the comment "Delay accounting fields end" are
+ * available only if delay accounting is enabled, even though the last
+ * few fields are not delays
+ *
+ * xxx_count is the number of delay values recorded
+ * xxx_delay_total is the corresponding cumulative delay in nanoseconds
+ *
+ * xxx_delay_total wraps around to zero on overflow
+ * xxx_count incremented regardless of overflow
+ */
+
+ /* Delay waiting for cpu, while runnable
+ * count, delay_total NOT updated atomically
+ */
+ __u64 cpu_count;
+ __u64 cpu_delay_total;
+
+ /* Following four fields atomically updated using task->delays->lock */
+
+ /* Delay waiting for synchronous block I/O to complete
+ * does not account for delays in I/O submission
+ */
+ __u64 blkio_count;
+ __u64 blkio_delay_total;
+
+ /* Delay waiting for page fault I/O (swap in only) */
+ __u64 swapin_count;
+ __u64 swapin_delay_total;
+
+ /* cpu "wall-clock" running time
+ * On some architectures, value will adjust for cpu time stolen
+ * from the kernel in involuntary waits due to virtualization.
+ * Value is cumulative, in nanoseconds, without a corresponding count
+ * and wraps around to zero silently on overflow
+ */
+ __u64 cpu_run_real_total;
+
+ /* cpu "virtual" running time
+ * Uses time intervals seen by the kernel i.e. no adjustment
+ * for kernel's involuntary waits due to virtualization.
+ * Value is cumulative, in nanoseconds, without a corresponding count
+ * and wraps around to zero silently on overflow
+ */
+ __u64 cpu_run_virtual_total;
+ /* Delay accounting fields end */
+ /* version 1 ends here */
+
+
+3) Extended accounting fields::
+
+ /* Extended accounting fields start */
+
+ /* Accumulated RSS usage in duration of a task, in MBytes-usecs.
+ * The current rss usage is added to this counter every time
+ * a tick is charged to a task's system time. So, at the end we
+ * will have memory usage multiplied by system time. Thus an
+ * average usage per system time unit can be calculated.
+ */
+ __u64 coremem; /* accumulated RSS usage in MB-usec */
+
+ /* Accumulated virtual memory usage in duration of a task.
+ * Same as acct_rss_mem1 above except that we keep track of VM usage.
+ */
+ __u64 virtmem; /* accumulated VM usage in MB-usec */
+
+ /* High watermark of RSS usage in duration of a task, in KBytes. */
+ __u64 hiwater_rss; /* High-watermark of RSS usage */
+
+ /* High watermark of VM usage in duration of a task, in KBytes. */
+ __u64 hiwater_vm; /* High-water virtual memory usage */
+
+ /* The following four fields are I/O statistics of a task. */
+ __u64 read_char; /* bytes read */
+ __u64 write_char; /* bytes written */
+ __u64 read_syscalls; /* read syscalls */
+ __u64 write_syscalls; /* write syscalls */
+
+ /* Extended accounting fields end */
+
+4) Per-task and per-thread statistics::
+
+ __u64 nvcsw; /* Context voluntary switch counter */
+ __u64 nivcsw; /* Context involuntary switch counter */
+
+5) Time accounting for SMT machines::
+
+ __u64 ac_utimescaled; /* utime scaled on frequency etc */
+ __u64 ac_stimescaled; /* stime scaled on frequency etc */
+ __u64 cpu_scaled_run_real_total; /* scaled cpu_run_real_total */
+
+6) Extended delay accounting fields for memory reclaim::
+
+ /* Delay waiting for memory reclaim */
+ __u64 freepages_count;
+ __u64 freepages_delay_total;
+
+::
+
+ }
diff --git a/Documentation/accounting/taskstats.rst b/Documentation/accounting/taskstats.rst
new file mode 100644
index 000000000..2a28b7f55
--- /dev/null
+++ b/Documentation/accounting/taskstats.rst
@@ -0,0 +1,180 @@
+=============================
+Per-task statistics interface
+=============================
+
+
+Taskstats is a netlink-based interface for sending per-task and
+per-process statistics from the kernel to userspace.
+
+Taskstats was designed for the following benefits:
+
+- efficiently provide statistics during lifetime of a task and on its exit
+- unified interface for multiple accounting subsystems
+- extensibility for use by future accounting patches
+
+Terminology
+-----------
+
+"pid", "tid" and "task" are used interchangeably and refer to the standard
+Linux task defined by struct task_struct. per-pid stats are the same as
+per-task stats.
+
+"tgid", "process" and "thread group" are used interchangeably and refer to the
+tasks that share an mm_struct i.e. the traditional Unix process. Despite the
+use of tgid, there is no special treatment for the task that is thread group
+leader - a process is deemed alive as long as it has any task belonging to it.
+
+Usage
+-----
+
+To get statistics during a task's lifetime, userspace opens a unicast netlink
+socket (NETLINK_GENERIC family) and sends commands specifying a pid or a tgid.
+The response contains statistics for a task (if pid is specified) or the sum of
+statistics for all tasks of the process (if tgid is specified).
+
+To obtain statistics for tasks which are exiting, the userspace listener
+sends a register command and specifies a cpumask. Whenever a task exits on
+one of the cpus in the cpumask, its per-pid statistics are sent to the
+registered listener. Using cpumasks allows the data received by one listener
+to be limited and assists in flow control over the netlink interface and is
+explained in more detail below.
+
+If the exiting task is the last thread exiting its thread group,
+an additional record containing the per-tgid stats is also sent to userspace.
+The latter contains the sum of per-pid stats for all threads in the thread
+group, both past and present.
+
+getdelays.c is a simple utility demonstrating usage of the taskstats interface
+for reporting delay accounting statistics. Users can register cpumasks,
+send commands and process responses, listen for per-tid/tgid exit data,
+write the data received to a file and do basic flow control by increasing
+receive buffer sizes.
+
+Interface
+---------
+
+The user-kernel interface is encapsulated in include/linux/taskstats.h
+
+To avoid this documentation becoming obsolete as the interface evolves, only
+an outline of the current version is given. taskstats.h always overrides the
+description here.
+
+struct taskstats is the common accounting structure for both per-pid and
+per-tgid data. It is versioned and can be extended by each accounting subsystem
+that is added to the kernel. The fields and their semantics are defined in the
+taskstats.h file.
+
+The data exchanged between user and kernel space is a netlink message belonging
+to the NETLINK_GENERIC family and using the netlink attributes interface.
+The messages are in the format::
+
+ +----------+- - -+-------------+-------------------+
+ | nlmsghdr | Pad | genlmsghdr | taskstats payload |
+ +----------+- - -+-------------+-------------------+
+
+
+The taskstats payload is one of the following three kinds:
+
+1. Commands: Sent from user to kernel. Commands to get data on
+a pid/tgid consist of one attribute, of type TASKSTATS_CMD_ATTR_PID/TGID,
+containing a u32 pid or tgid in the attribute payload. The pid/tgid denotes
+the task/process for which userspace wants statistics.
+
+Commands to register/deregister interest in exit data from a set of cpus
+consist of one attribute, of type
+TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK and contain a cpumask in the
+attribute payload. The cpumask is specified as an ascii string of
+comma-separated cpu ranges e.g. to listen to exit data from cpus 1,2,3,5,7,8
+the cpumask would be "1-3,5,7-8". If userspace forgets to deregister interest
+in cpus before closing the listening socket, the kernel cleans up its interest
+set over time. However, for the sake of efficiency, an explicit deregistration
+is advisable.
+
+2. Response for a command: sent from the kernel in response to a userspace
+command. The payload is a series of three attributes of type:
+
+a) TASKSTATS_TYPE_AGGR_PID/TGID : attribute containing no payload but indicates
+a pid/tgid will be followed by some stats.
+
+b) TASKSTATS_TYPE_PID/TGID: attribute whose payload is the pid/tgid whose stats
+are being returned.
+
+c) TASKSTATS_TYPE_STATS: attribute with a struct taskstats as payload. The
+same structure is used for both per-pid and per-tgid stats.
+
+3. New message sent by kernel whenever a task exits. The payload consists of a
+ series of attributes of the following type:
+
+a) TASKSTATS_TYPE_AGGR_PID: indicates next two attributes will be pid+stats
+b) TASKSTATS_TYPE_PID: contains exiting task's pid
+c) TASKSTATS_TYPE_STATS: contains the exiting task's per-pid stats
+d) TASKSTATS_TYPE_AGGR_TGID: indicates next two attributes will be tgid+stats
+e) TASKSTATS_TYPE_TGID: contains tgid of process to which task belongs
+f) TASKSTATS_TYPE_STATS: contains the per-tgid stats for exiting task's process
+
+
+per-tgid stats
+--------------
+
+Taskstats provides per-process stats, in addition to per-task stats, since
+resource management is often done at a process granularity and aggregating task
+stats in userspace alone is inefficient and potentially inaccurate (due to lack
+of atomicity).
+
+However, maintaining per-process, in addition to per-task stats, within the
+kernel has space and time overheads. To address this, the taskstats code
+accumulates each exiting task's statistics into a process-wide data structure.
+When the last task of a process exits, the process level data accumulated also
+gets sent to userspace (along with the per-task data).
+
+When a user queries to get per-tgid data, the sum of all other live threads in
+the group is added up and added to the accumulated total for previously exited
+threads of the same thread group.
+
+Extending taskstats
+-------------------
+
+There are two ways to extend the taskstats interface to export more
+per-task/process stats as patches to collect them get added to the kernel
+in future:
+
+1. Adding more fields to the end of the existing struct taskstats. Backward
+ compatibility is ensured by the version number within the
+ structure. Userspace will use only the fields of the struct that correspond
+ to the version its using.
+
+2. Defining separate statistic structs and using the netlink attributes
+ interface to return them. Since userspace processes each netlink attribute
+ independently, it can always ignore attributes whose type it does not
+ understand (because it is using an older version of the interface).
+
+
+Choosing between 1. and 2. is a matter of trading off flexibility and
+overhead. If only a few fields need to be added, then 1. is the preferable
+path since the kernel and userspace don't need to incur the overhead of
+processing new netlink attributes. But if the new fields expand the existing
+struct too much, requiring disparate userspace accounting utilities to
+unnecessarily receive large structures whose fields are of no interest, then
+extending the attributes structure would be worthwhile.
+
+Flow control for taskstats
+--------------------------
+
+When the rate of task exits becomes large, a listener may not be able to keep
+up with the kernel's rate of sending per-tid/tgid exit data leading to data
+loss. This possibility gets compounded when the taskstats structure gets
+extended and the number of cpus grows large.
+
+To avoid losing statistics, userspace should do one or more of the following:
+
+- increase the receive buffer sizes for the netlink sockets opened by
+ listeners to receive exit data.
+
+- create more listeners and reduce the number of cpus being listened to by
+ each listener. In the extreme case, there could be one listener for each cpu.
+ Users may also consider setting the cpu affinity of the listener to the subset
+ of cpus to which it listens, especially if they are listening to just one cpu.
+
+Despite these measures, if the userspace receives ENOBUFS error messages
+indicated overflow of receive buffers, it should take measures to handle the
+loss of data.