xref: /linux/Documentation/scheduler/sched-design-CFS.rst (revision 4b660dbd9ee2059850fd30e0df420ca7a38a1856)
1=============
2CFS Scheduler
3=============
4
5
61.  OVERVIEW
7============
8
9CFS stands for "Completely Fair Scheduler," and is the new "desktop" process
10scheduler implemented by Ingo Molnar and merged in Linux 2.6.23.  It is the
11replacement for the previous vanilla scheduler's SCHED_OTHER interactivity
12code.
13
1480% of CFS's design can be summed up in a single sentence: CFS basically models
15an "ideal, precise multi-tasking CPU" on real hardware.
16
17"Ideal multi-tasking CPU" is a (non-existent  :-)) CPU that has 100% physical
18power and which can run each task at precise equal speed, in parallel, each at
191/nr_running speed.  For example: if there are 2 tasks running, then it runs
20each at 50% physical power --- i.e., actually in parallel.
21
22On real hardware, we can run only a single task at once, so we have to
23introduce the concept of "virtual runtime."  The virtual runtime of a task
24specifies when its next timeslice would start execution on the ideal
25multi-tasking CPU described above.  In practice, the virtual runtime of a task
26is its actual runtime normalized to the total number of running tasks.
27
28
29
302.  FEW IMPLEMENTATION DETAILS
31==============================
32
33In CFS the virtual runtime is expressed and tracked via the per-task
34p->se.vruntime (nanosec-unit) value.  This way, it's possible to accurately
35timestamp and measure the "expected CPU time" a task should have gotten.
36
37   Small detail: on "ideal" hardware, at any time all tasks would have the same
38   p->se.vruntime value --- i.e., tasks would execute simultaneously and no task
39   would ever get "out of balance" from the "ideal" share of CPU time.
40
41CFS's task picking logic is based on this p->se.vruntime value and it is thus
42very simple: it always tries to run the task with the smallest p->se.vruntime
43value (i.e., the task which executed least so far).  CFS always tries to split
44up CPU time between runnable tasks as close to "ideal multitasking hardware" as
45possible.
46
47Most of the rest of CFS's design just falls out of this really simple concept,
48with a few add-on embellishments like nice levels, multiprocessing and various
49algorithm variants to recognize sleepers.
50
51
52
533.  THE RBTREE
54==============
55
56CFS's design is quite radical: it does not use the old data structures for the
57runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
58task execution, and thus has no "array switch" artifacts (by which both the
59previous vanilla scheduler and RSDL/SD are affected).
60
61CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
62increasing value tracking the smallest vruntime among all tasks in the
63runqueue.  The total amount of work done by the system is tracked using
64min_vruntime; that value is used to place newly activated entities on the left
65side of the tree as much as possible.
66
67The total number of running tasks in the runqueue is accounted through the
68rq->cfs.load value, which is the sum of the weights of the tasks queued on the
69runqueue.
70
71CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
72p->se.vruntime key. CFS picks the "leftmost" task from this tree and sticks to it.
73As the system progresses forwards, the executed tasks are put into the tree
74more and more to the right --- slowly but surely giving a chance for every task
75to become the "leftmost task" and thus get on the CPU within a deterministic
76amount of time.
77
78Summing up, CFS works like this: it runs a task a bit, and when the task
79schedules (or a scheduler tick happens) the task's CPU usage is "accounted
80for": the (small) time it just spent using the physical CPU is added to
81p->se.vruntime.  Once p->se.vruntime gets high enough so that another task
82becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
83small amount of "granularity" distance relative to the leftmost task so that we
84do not over-schedule tasks and trash the cache), then the new leftmost task is
85picked and the current task is preempted.
86
87
88
894.  SOME FEATURES OF CFS
90========================
91
92CFS uses nanosecond granularity accounting and does not rely on any jiffies or
93other HZ detail.  Thus the CFS scheduler has no notion of "timeslices" in the
94way the previous scheduler had, and has no heuristics whatsoever.  There is
95only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
96
97   /sys/kernel/debug/sched/base_slice_ns
98
99which can be used to tune the scheduler from "desktop" (i.e., low latencies) to
100"server" (i.e., good batching) workloads.  It defaults to a setting suitable
101for desktop workloads.  SCHED_BATCH is handled by the CFS scheduler module too.
102
103In case CONFIG_HZ results in base_slice_ns < TICK_NSEC, the value of
104base_slice_ns will have little to no impact on the workloads.
105
106Due to its design, the CFS scheduler is not prone to any of the "attacks" that
107exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
108chew.c, ring-test.c, massive_intr.c all work fine and do not impact
109interactivity and produce the expected behavior.
110
111The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
112than the previous vanilla scheduler: both types of workloads are isolated much
113more aggressively.
114
115SMP load-balancing has been reworked/sanitized: the runqueue-walking
116assumptions are gone from the load-balancing code now, and iterators of the
117scheduling modules are used.  The balancing code got quite a bit simpler as a
118result.
119
120
121
1225. Scheduling policies
123======================
124
125CFS implements three scheduling policies:
126
127  - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
128    policy that is used for regular tasks.
129
130  - SCHED_BATCH: Does not preempt nearly as often as regular tasks
131    would, thereby allowing tasks to run longer and make better use of
132    caches but at the cost of interactivity. This is well suited for
133    batch jobs.
134
135  - SCHED_IDLE: This is even weaker than nice 19, but its not a true
136    idle timer scheduler in order to avoid to get into priority
137    inversion problems which would deadlock the machine.
138
139SCHED_FIFO/_RR are implemented in sched/rt.c and are as specified by
140POSIX.
141
142The command chrt from util-linux-ng 2.13.1.1 can set all of these except
143SCHED_IDLE.
144
145
146
1476.  SCHEDULING CLASSES
148======================
149
150The new CFS scheduler has been designed in such a way to introduce "Scheduling
151Classes," an extensible hierarchy of scheduler modules.  These modules
152encapsulate scheduling policy details and are handled by the scheduler core
153without the core code assuming too much about them.
154
155sched/fair.c implements the CFS scheduler described above.
156
157sched/rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
158the previous vanilla scheduler did.  It uses 100 runqueues (for all 100 RT
159priority levels, instead of 140 in the previous scheduler) and it needs no
160expired array.
161
162Scheduling classes are implemented through the sched_class structure, which
163contains hooks to functions that must be called whenever an interesting event
164occurs.
165
166This is the (partial) list of the hooks:
167
168 - enqueue_task(...)
169
170   Called when a task enters a runnable state.
171   It puts the scheduling entity (task) into the red-black tree and
172   increments the nr_running variable.
173
174 - dequeue_task(...)
175
176   When a task is no longer runnable, this function is called to keep the
177   corresponding scheduling entity out of the red-black tree.  It decrements
178   the nr_running variable.
179
180 - yield_task(...)
181
182   This function is basically just a dequeue followed by an enqueue, unless the
183   compat_yield sysctl is turned on; in that case, it places the scheduling
184   entity at the right-most end of the red-black tree.
185
186 - wakeup_preempt(...)
187
188   This function checks if a task that entered the runnable state should
189   preempt the currently running task.
190
191 - pick_next_task(...)
192
193   This function chooses the most appropriate task eligible to run next.
194
195 - set_next_task(...)
196
197   This function is called when a task changes its scheduling class, changes
198   its task group or is scheduled.
199
200 - task_tick(...)
201
202   This function is mostly called from time tick functions; it might lead to
203   process switch.  This drives the running preemption.
204
205
206
207
2087.  GROUP SCHEDULER EXTENSIONS TO CFS
209=====================================
210
211Normally, the scheduler operates on individual tasks and strives to provide
212fair CPU time to each task.  Sometimes, it may be desirable to group tasks and
213provide fair CPU time to each such task group.  For example, it may be
214desirable to first provide fair CPU time to each user on the system and then to
215each task belonging to a user.
216
217CONFIG_CGROUP_SCHED strives to achieve exactly that.  It lets tasks to be
218grouped and divides CPU time fairly among such groups.
219
220CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
221SCHED_RR) tasks.
222
223CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
224SCHED_BATCH) tasks.
225
226   These options need CONFIG_CGROUPS to be defined, and let the administrator
227   create arbitrary groups of tasks, using the "cgroup" pseudo filesystem.  See
228   Documentation/admin-guide/cgroup-v1/cgroups.rst for more information about this filesystem.
229
230When CONFIG_FAIR_GROUP_SCHED is defined, a "cpu.shares" file is created for each
231group created using the pseudo filesystem.  See example steps below to create
232task groups and modify their CPU share using the "cgroups" pseudo filesystem::
233
234	# mount -t tmpfs cgroup_root /sys/fs/cgroup
235	# mkdir /sys/fs/cgroup/cpu
236	# mount -t cgroup -ocpu none /sys/fs/cgroup/cpu
237	# cd /sys/fs/cgroup/cpu
238
239	# mkdir multimedia	# create "multimedia" group of tasks
240	# mkdir browser		# create "browser" group of tasks
241
242	# #Configure the multimedia group to receive twice the CPU bandwidth
243	# #that of browser group
244
245	# echo 2048 > multimedia/cpu.shares
246	# echo 1024 > browser/cpu.shares
247
248	# firefox &	# Launch firefox and move it to "browser" group
249	# echo <firefox_pid> > browser/tasks
250
251	# #Launch gmplayer (or your favourite movie player)
252	# echo <movie_player_pid> > multimedia/tasks
253