9 CFS stands for "Completely Fair Scheduler," and is the new "desktop" process
10 scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. It is the
11 replacement for the previous vanilla scheduler's SCHED_OTHER interactivity
14 80% of CFS's design can be summed up in a single sentence: CFS basically models
15 an "ideal, precise multi-tasking CPU" on real hardware.
17 "Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100% physical
18 power and which can run each task at precise equal speed, in parallel, each at
19 1/nr_running speed. For example: if there are 2 tasks running, then it runs
20 each at 50% physical power --- i.e., actually in parallel.
22 On real hardware, we can run only a single task at once, so we have to
23 introduce the concept of "virtual runtime." The virtual runtime of a task
24 specifies when its next timeslice would start execution on the ideal
25 multi-tasking CPU described above. In practice, the virtual runtime of a task
26 is its actual runtime normalized to the total number of running tasks.
30 2. FEW IMPLEMENTATION DETAILS
31 ==============================
33 In CFS the virtual runtime is expressed and tracked via the per-task
34 p->se.vruntime (nanosec-unit) value. This way, it's possible to accurately
35 timestamp and measure the "expected CPU time" a task should have gotten.
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.
41 CFS's task picking logic is based on this p->se.vruntime value and it is thus
42 very simple: it always tries to run the task with the smallest p->se.vruntime
43 value (i.e., the task which executed least so far). CFS always tries to split
44 up CPU time between runnable tasks as close to "ideal multitasking hardware" as
47 Most of the rest of CFS's design just falls out of this really simple concept,
48 with a few add-on embellishments like nice levels, multiprocessing and various
49 algorithm variants to recognize sleepers.
56 CFS's design is quite radical: it does not use the old data structures for the
57 runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
58 task execution, and thus has no "array switch" artifacts (by which both the
59 previous vanilla scheduler and RSDL/SD are affected).
61 CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
62 increasing value tracking the smallest vruntime among all tasks in the
63 runqueue. The total amount of work done by the system is tracked using
64 min_vruntime; that value is used to place newly activated entities on the left
65 side of the tree as much as possible.
67 The total number of running tasks in the runqueue is accounted through the
68 rq->cfs.load value, which is the sum of the weights of the tasks queued on the
71 CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
72 p->se.vruntime key. CFS picks the "leftmost" task from this tree and sticks to it.
73 As the system progresses forwards, the executed tasks are put into the tree
74 more and more to the right --- slowly but surely giving a chance for every task
75 to become the "leftmost task" and thus get on the CPU within a deterministic
78 Summing up, CFS works like this: it runs a task a bit, and when the task
79 schedules (or a scheduler tick happens) the task's CPU usage is "accounted
80 for": the (small) time it just spent using the physical CPU is added to
81 p->se.vruntime. Once p->se.vruntime gets high enough so that another task
82 becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
83 small amount of "granularity" distance relative to the leftmost task so that we
84 do not over-schedule tasks and trash the cache), then the new leftmost task is
85 picked and the current task is preempted.
89 4. SOME FEATURES OF CFS
90 ========================
92 CFS uses nanosecond granularity accounting and does not rely on any jiffies or
93 other HZ detail. Thus the CFS scheduler has no notion of "timeslices" in the
94 way the previous scheduler had, and has no heuristics whatsoever. There is
95 only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
97 /sys/kernel/debug/sched/base_slice_ns
99 which 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
101 for desktop workloads. SCHED_BATCH is handled by the CFS scheduler module too.
103 In case CONFIG_HZ results in base_slice_ns < TICK_NSEC, the value of
104 base_slice_ns will have little to no impact on the workloads.
106 Due to its design, the CFS scheduler is not prone to any of the "attacks" that
107 exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
108 chew.c, ring-test.c, massive_intr.c all work fine and do not impact
109 interactivity and produce the expected behavior.
111 The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
112 than the previous vanilla scheduler: both types of workloads are isolated much
115 SMP load-balancing has been reworked/sanitized: the runqueue-walking
116 assumptions are gone from the load-balancing code now, and iterators of the
117 scheduling modules are used. The balancing code got quite a bit simpler as a
122 5. Scheduling policies
123 ======================
125 CFS implements three scheduling policies:
127 - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
128 policy that is used for regular tasks.
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
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.
139 SCHED_FIFO/_RR are implemented in sched/rt.c and are as specified by
142 The command chrt from util-linux-ng 2.13.1.1 can set all of these except
147 6. SCHEDULING CLASSES
148 ======================
150 The new CFS scheduler has been designed in such a way to introduce "Scheduling
151 Classes," an extensible hierarchy of scheduler modules. These modules
152 encapsulate scheduling policy details and are handled by the scheduler core
153 without the core code assuming too much about them.
155 sched/fair.c implements the CFS scheduler described above.
157 sched/rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
158 the previous vanilla scheduler did. It uses 100 runqueues (for all 100 RT
159 priority levels, instead of 140 in the previous scheduler) and it needs no
162 Scheduling classes are implemented through the sched_class structure, which
163 contains hooks to functions that must be called whenever an interesting event
166 This is the (partial) list of the hooks:
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.
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.
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.
186 - wakeup_preempt(...)
188 This function checks if a task that entered the runnable state should
189 preempt the currently running task.
191 - pick_next_task(...)
193 This function chooses the most appropriate task eligible to run next.
197 This function is called when a task changes its scheduling class, changes
198 its task group or is scheduled.
202 This function is mostly called from time tick functions; it might lead to
203 process switch. This drives the running preemption.
208 7. GROUP SCHEDULER EXTENSIONS TO CFS
209 =====================================
211 Normally, the scheduler operates on individual tasks and strives to provide
212 fair CPU time to each task. Sometimes, it may be desirable to group tasks and
213 provide fair CPU time to each such task group. For example, it may be
214 desirable to first provide fair CPU time to each user on the system and then to
215 each task belonging to a user.
217 CONFIG_CGROUP_SCHED strives to achieve exactly that. It lets tasks to be
218 grouped and divides CPU time fairly among such groups.
220 CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
223 CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
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.
230 When CONFIG_FAIR_GROUP_SCHED is defined, a "cpu.shares" file is created for each
231 group created using the pseudo filesystem. See example steps below to create
232 task groups and modify their CPU share using the "cgroups" pseudo filesystem::
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
239 # mkdir multimedia # create "multimedia" group of tasks
240 # mkdir browser # create "browser" group of tasks
242 # #Configure the multimedia group to receive twice the CPU bandwidth
243 # #that of browser group
245 # echo 2048 > multimedia/cpu.shares
246 # echo 1024 > browser/cpu.shares
248 # firefox & # Launch firefox and move it to "browser" group
249 # echo <firefox_pid> > browser/tasks
251 # #Launch gmplayer (or your favourite movie player)
252 # echo <movie_player_pid> > multimedia/tasks