xref: /linux/Documentation/scheduler/sched-design-CFS.rst (revision 566ab427f827b0256d3e8ce0235d088e6a9c28bd)
1.. _sched_design_CFS:
2
3=============
4CFS Scheduler
5=============
6
7
81.  OVERVIEW
9============
10
11CFS stands for "Completely Fair Scheduler," and is the "desktop" process
12scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. When
13originally merged, it was the replacement for the previous vanilla
14scheduler's SCHED_OTHER interactivity code. Nowadays, CFS is making room
15for EEVDF, for which documentation can be found in
16Documentation/scheduler/sched-eevdf.rst.
17
1880% of CFS's design can be summed up in a single sentence: CFS basically models
19an "ideal, precise multi-tasking CPU" on real hardware.
20
21"Ideal multi-tasking CPU" is a (non-existent  :-)) CPU that has 100% physical
22power and which can run each task at precise equal speed, in parallel, each at
231/nr_running speed.  For example: if there are 2 tasks running, then it runs
24each at 50% physical power --- i.e., actually in parallel.
25
26On real hardware, we can run only a single task at once, so we have to
27introduce the concept of "virtual runtime."  The virtual runtime of a task
28specifies when its next timeslice would start execution on the ideal
29multi-tasking CPU described above.  In practice, the virtual runtime of a task
30is its actual runtime normalized to the total number of running tasks.
31
32
33
342.  FEW IMPLEMENTATION DETAILS
35==============================
36
37In CFS the virtual runtime is expressed and tracked via the per-task
38p->se.vruntime (nanosec-unit) value.  This way, it's possible to accurately
39timestamp and measure the "expected CPU time" a task should have gotten.
40
41   Small detail: on "ideal" hardware, at any time all tasks would have the same
42   p->se.vruntime value --- i.e., tasks would execute simultaneously and no task
43   would ever get "out of balance" from the "ideal" share of CPU time.
44
45CFS's task picking logic is based on this p->se.vruntime value and it is thus
46very simple: it always tries to run the task with the smallest p->se.vruntime
47value (i.e., the task which executed least so far).  CFS always tries to split
48up CPU time between runnable tasks as close to "ideal multitasking hardware" as
49possible.
50
51Most of the rest of CFS's design just falls out of this really simple concept,
52with a few add-on embellishments like nice levels, multiprocessing and various
53algorithm variants to recognize sleepers.
54
55
56
573.  THE RBTREE
58==============
59
60CFS's design is quite radical: it does not use the old data structures for the
61runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
62task execution, and thus has no "array switch" artifacts (by which both the
63previous vanilla scheduler and RSDL/SD are affected).
64
65CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
66increasing value tracking the smallest vruntime among all tasks in the
67runqueue.  The total amount of work done by the system is tracked using
68min_vruntime; that value is used to place newly activated entities on the left
69side of the tree as much as possible.
70
71The total number of running tasks in the runqueue is accounted through the
72rq->cfs.load value, which is the sum of the weights of the tasks queued on the
73runqueue.
74
75CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
76p->se.vruntime key. CFS picks the "leftmost" task from this tree and sticks to it.
77As the system progresses forwards, the executed tasks are put into the tree
78more and more to the right --- slowly but surely giving a chance for every task
79to become the "leftmost task" and thus get on the CPU within a deterministic
80amount of time.
81
82Summing up, CFS works like this: it runs a task a bit, and when the task
83schedules (or a scheduler tick happens) the task's CPU usage is "accounted
84for": the (small) time it just spent using the physical CPU is added to
85p->se.vruntime.  Once p->se.vruntime gets high enough so that another task
86becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
87small amount of "granularity" distance relative to the leftmost task so that we
88do not over-schedule tasks and trash the cache), then the new leftmost task is
89picked and the current task is preempted.
90
91
92
934.  SOME FEATURES OF CFS
94========================
95
96CFS uses nanosecond granularity accounting and does not rely on any jiffies or
97other HZ detail.  Thus the CFS scheduler has no notion of "timeslices" in the
98way the previous scheduler had, and has no heuristics whatsoever.  There is
99only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
100
101   /sys/kernel/debug/sched/base_slice_ns
102
103which can be used to tune the scheduler from "desktop" (i.e., low latencies) to
104"server" (i.e., good batching) workloads.  It defaults to a setting suitable
105for desktop workloads.  SCHED_BATCH is handled by the CFS scheduler module too.
106
107In case CONFIG_HZ results in base_slice_ns < TICK_NSEC, the value of
108base_slice_ns will have little to no impact on the workloads.
109
110Due to its design, the CFS scheduler is not prone to any of the "attacks" that
111exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
112chew.c, ring-test.c, massive_intr.c all work fine and do not impact
113interactivity and produce the expected behavior.
114
115The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
116than the previous vanilla scheduler: both types of workloads are isolated much
117more aggressively.
118
119SMP load-balancing has been reworked/sanitized: the runqueue-walking
120assumptions are gone from the load-balancing code now, and iterators of the
121scheduling modules are used.  The balancing code got quite a bit simpler as a
122result.
123
124
125
1265. Scheduling policies
127======================
128
129CFS implements three scheduling policies:
130
131  - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
132    policy that is used for regular tasks.
133
134  - SCHED_BATCH: Does not preempt nearly as often as regular tasks
135    would, thereby allowing tasks to run longer and make better use of
136    caches but at the cost of interactivity. This is well suited for
137    batch jobs.
138
139  - SCHED_IDLE: This is even weaker than nice 19, but its not a true
140    idle timer scheduler in order to avoid to get into priority
141    inversion problems which would deadlock the machine.
142
143SCHED_FIFO/_RR are implemented in sched/rt.c and are as specified by
144POSIX.
145
146The command chrt from util-linux-ng 2.13.1.1 can set all of these except
147SCHED_IDLE.
148
149
150
1516.  SCHEDULING CLASSES
152======================
153
154The new CFS scheduler has been designed in such a way to introduce "Scheduling
155Classes," an extensible hierarchy of scheduler modules.  These modules
156encapsulate scheduling policy details and are handled by the scheduler core
157without the core code assuming too much about them.
158
159sched/fair.c implements the CFS scheduler described above.
160
161sched/rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
162the previous vanilla scheduler did.  It uses 100 runqueues (for all 100 RT
163priority levels, instead of 140 in the previous scheduler) and it needs no
164expired array.
165
166Scheduling classes are implemented through the sched_class structure, which
167contains hooks to functions that must be called whenever an interesting event
168occurs.
169
170This is the (partial) list of the hooks:
171
172 - enqueue_task(...)
173
174   Called when a task enters a runnable state.
175   It puts the scheduling entity (task) into the red-black tree and
176   increments the nr_running variable.
177
178 - dequeue_task(...)
179
180   When a task is no longer runnable, this function is called to keep the
181   corresponding scheduling entity out of the red-black tree.  It decrements
182   the nr_running variable.
183
184 - yield_task(...)
185
186   This function is basically just a dequeue followed by an enqueue, unless the
187   compat_yield sysctl is turned on; in that case, it places the scheduling
188   entity at the right-most end of the red-black tree.
189
190 - wakeup_preempt(...)
191
192   This function checks if a task that entered the runnable state should
193   preempt the currently running task.
194
195 - pick_next_task(...)
196
197   This function chooses the most appropriate task eligible to run next.
198
199 - set_next_task(...)
200
201   This function is called when a task changes its scheduling class, changes
202   its task group or is scheduled.
203
204 - task_tick(...)
205
206   This function is mostly called from time tick functions; it might lead to
207   process switch.  This drives the running preemption.
208
209
210
211
2127.  GROUP SCHEDULER EXTENSIONS TO CFS
213=====================================
214
215Normally, the scheduler operates on individual tasks and strives to provide
216fair CPU time to each task.  Sometimes, it may be desirable to group tasks and
217provide fair CPU time to each such task group.  For example, it may be
218desirable to first provide fair CPU time to each user on the system and then to
219each task belonging to a user.
220
221CONFIG_CGROUP_SCHED strives to achieve exactly that.  It lets tasks to be
222grouped and divides CPU time fairly among such groups.
223
224CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
225SCHED_RR) tasks.
226
227CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
228SCHED_BATCH) tasks.
229
230   These options need CONFIG_CGROUPS to be defined, and let the administrator
231   create arbitrary groups of tasks, using the "cgroup" pseudo filesystem.  See
232   Documentation/admin-guide/cgroup-v1/cgroups.rst for more information about this filesystem.
233
234When CONFIG_FAIR_GROUP_SCHED is defined, a "cpu.shares" file is created for each
235group created using the pseudo filesystem.  See example steps below to create
236task groups and modify their CPU share using the "cgroups" pseudo filesystem::
237
238	# mount -t tmpfs cgroup_root /sys/fs/cgroup
239	# mkdir /sys/fs/cgroup/cpu
240	# mount -t cgroup -ocpu none /sys/fs/cgroup/cpu
241	# cd /sys/fs/cgroup/cpu
242
243	# mkdir multimedia	# create "multimedia" group of tasks
244	# mkdir browser		# create "browser" group of tasks
245
246	# #Configure the multimedia group to receive twice the CPU bandwidth
247	# #that of browser group
248
249	# echo 2048 > multimedia/cpu.shares
250	# echo 1024 > browser/cpu.shares
251
252	# firefox &	# Launch firefox and move it to "browser" group
253	# echo <firefox_pid> > browser/tasks
254
255	# #Launch gmplayer (or your favourite movie player)
256	# echo <movie_player_pid> > multimedia/tasks
257