Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 1 | ============= |
| 2 | CFS Scheduler |
| 3 | ============= |
Ingo Molnar | 5e7eaad | 2007-07-09 18:52:00 +0200 | [diff] [blame] | 4 | |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 5 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 6 | 1. OVERVIEW |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 7 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 8 | CFS stands for "Completely Fair Scheduler," and is the new "desktop" process |
| 9 | scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. It is the |
| 10 | replacement for the previous vanilla scheduler's SCHED_OTHER interactivity |
| 11 | code. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 12 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 13 | 80% of CFS's design can be summed up in a single sentence: CFS basically models |
| 14 | an "ideal, precise multi-tasking CPU" on real hardware. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 15 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 16 | "Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100% physical |
| 17 | power and which can run each task at precise equal speed, in parallel, each at |
| 18 | 1/nr_running speed. For example: if there are 2 tasks running, then it runs |
| 19 | each at 50% physical power --- i.e., actually in parallel. |
| 20 | |
| 21 | On real hardware, we can run only a single task at once, so we have to |
| 22 | introduce the concept of "virtual runtime." The virtual runtime of a task |
| 23 | specifies when its next timeslice would start execution on the ideal |
| 24 | multi-tasking CPU described above. In practice, the virtual runtime of a task |
| 25 | is its actual runtime normalized to the total number of running tasks. |
| 26 | |
| 27 | |
| 28 | |
| 29 | 2. FEW IMPLEMENTATION DETAILS |
| 30 | |
| 31 | In CFS the virtual runtime is expressed and tracked via the per-task |
| 32 | p->se.vruntime (nanosec-unit) value. This way, it's possible to accurately |
| 33 | timestamp and measure the "expected CPU time" a task should have gotten. |
| 34 | |
| 35 | [ small detail: on "ideal" hardware, at any time all tasks would have the same |
| 36 | p->se.vruntime value --- i.e., tasks would execute simultaneously and no task |
| 37 | would ever get "out of balance" from the "ideal" share of CPU time. ] |
| 38 | |
| 39 | CFS's task picking logic is based on this p->se.vruntime value and it is thus |
| 40 | very simple: it always tries to run the task with the smallest p->se.vruntime |
| 41 | value (i.e., the task which executed least so far). CFS always tries to split |
| 42 | up CPU time between runnable tasks as close to "ideal multitasking hardware" as |
| 43 | possible. |
| 44 | |
| 45 | Most of the rest of CFS's design just falls out of this really simple concept, |
| 46 | with a few add-on embellishments like nice levels, multiprocessing and various |
| 47 | algorithm variants to recognize sleepers. |
| 48 | |
| 49 | |
| 50 | |
| 51 | 3. THE RBTREE |
| 52 | |
| 53 | CFS's design is quite radical: it does not use the old data structures for the |
| 54 | runqueues, but it uses a time-ordered rbtree to build a "timeline" of future |
| 55 | task execution, and thus has no "array switch" artifacts (by which both the |
| 56 | previous vanilla scheduler and RSDL/SD are affected). |
| 57 | |
| 58 | CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic |
| 59 | increasing value tracking the smallest vruntime among all tasks in the |
| 60 | runqueue. The total amount of work done by the system is tracked using |
| 61 | min_vruntime; that value is used to place newly activated entities on the left |
| 62 | side of the tree as much as possible. |
| 63 | |
| 64 | The total number of running tasks in the runqueue is accounted through the |
| 65 | rq->cfs.load value, which is the sum of the weights of the tasks queued on the |
| 66 | runqueue. |
| 67 | |
| 68 | CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the |
| 69 | p->se.vruntime key (there is a subtraction using rq->cfs.min_vruntime to |
| 70 | account for possible wraparounds). CFS picks the "leftmost" task from this |
| 71 | tree and sticks to it. |
| 72 | As the system progresses forwards, the executed tasks are put into the tree |
| 73 | more and more to the right --- slowly but surely giving a chance for every task |
| 74 | to become the "leftmost task" and thus get on the CPU within a deterministic |
| 75 | amount of time. |
| 76 | |
| 77 | Summing up, CFS works like this: it runs a task a bit, and when the task |
| 78 | schedules (or a scheduler tick happens) the task's CPU usage is "accounted |
| 79 | for": the (small) time it just spent using the physical CPU is added to |
| 80 | p->se.vruntime. Once p->se.vruntime gets high enough so that another task |
| 81 | becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a |
| 82 | small amount of "granularity" distance relative to the leftmost task so that we |
| 83 | do not over-schedule tasks and trash the cache), then the new leftmost task is |
| 84 | picked and the current task is preempted. |
| 85 | |
| 86 | |
| 87 | |
| 88 | 4. SOME FEATURES OF CFS |
| 89 | |
| 90 | CFS uses nanosecond granularity accounting and does not rely on any jiffies or |
| 91 | other HZ detail. Thus the CFS scheduler has no notion of "timeslices" in the |
| 92 | way the previous scheduler had, and has no heuristics whatsoever. There is |
| 93 | only one central tunable (you have to switch on CONFIG_SCHED_DEBUG): |
| 94 | |
Jiri Kosina | 4078e35 | 2008-10-27 17:41:58 +0100 | [diff] [blame] | 95 | /proc/sys/kernel/sched_min_granularity_ns |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 96 | |
| 97 | which can be used to tune the scheduler from "desktop" (i.e., low latencies) to |
| 98 | "server" (i.e., good batching) workloads. It defaults to a setting suitable |
| 99 | for desktop workloads. SCHED_BATCH is handled by the CFS scheduler module too. |
| 100 | |
| 101 | Due to its design, the CFS scheduler is not prone to any of the "attacks" that |
| 102 | exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c, |
| 103 | chew.c, ring-test.c, massive_intr.c all work fine and do not impact |
| 104 | interactivity and produce the expected behavior. |
| 105 | |
| 106 | The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH |
| 107 | than the previous vanilla scheduler: both types of workloads are isolated much |
| 108 | more aggressively. |
| 109 | |
| 110 | SMP load-balancing has been reworked/sanitized: the runqueue-walking |
| 111 | assumptions are gone from the load-balancing code now, and iterators of the |
| 112 | scheduling modules are used. The balancing code got quite a bit simpler as a |
| 113 | result. |
| 114 | |
| 115 | |
| 116 | |
Martin Steigerwald | 1a73ef6 | 2008-09-23 13:48:44 +0200 | [diff] [blame] | 117 | 5. Scheduling policies |
| 118 | |
| 119 | CFS implements three scheduling policies: |
| 120 | |
| 121 | - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling |
| 122 | policy that is used for regular tasks. |
| 123 | |
| 124 | - SCHED_BATCH: Does not preempt nearly as often as regular tasks |
| 125 | would, thereby allowing tasks to run longer and make better use of |
| 126 | caches but at the cost of interactivity. This is well suited for |
| 127 | batch jobs. |
| 128 | |
| 129 | - SCHED_IDLE: This is even weaker than nice 19, but its not a true |
| 130 | idle timer scheduler in order to avoid to get into priority |
| 131 | inversion problems which would deadlock the machine. |
| 132 | |
| 133 | SCHED_FIFO/_RR are implemented in sched_rt.c and are as specified by |
| 134 | POSIX. |
| 135 | |
| 136 | The command chrt from util-linux-ng 2.13.1.1 can set all of these except |
| 137 | SCHED_IDLE. |
| 138 | |
| 139 | |
| 140 | |
| 141 | 6. SCHEDULING CLASSES |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 142 | |
| 143 | The new CFS scheduler has been designed in such a way to introduce "Scheduling |
| 144 | Classes," an extensible hierarchy of scheduler modules. These modules |
| 145 | encapsulate scheduling policy details and are handled by the scheduler core |
| 146 | without the core code assuming too much about them. |
| 147 | |
| 148 | sched_fair.c implements the CFS scheduler described above. |
| 149 | |
| 150 | sched_rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than |
| 151 | the previous vanilla scheduler did. It uses 100 runqueues (for all 100 RT |
| 152 | priority levels, instead of 140 in the previous scheduler) and it needs no |
| 153 | expired array. |
| 154 | |
| 155 | Scheduling classes are implemented through the sched_class structure, which |
| 156 | contains hooks to functions that must be called whenever an interesting event |
| 157 | occurs. |
| 158 | |
| 159 | This is the (partial) list of the hooks: |
| 160 | |
| 161 | - enqueue_task(...) |
| 162 | |
| 163 | Called when a task enters a runnable state. |
| 164 | It puts the scheduling entity (task) into the red-black tree and |
| 165 | increments the nr_running variable. |
| 166 | |
| 167 | - dequeue_tree(...) |
| 168 | |
| 169 | When a task is no longer runnable, this function is called to keep the |
| 170 | corresponding scheduling entity out of the red-black tree. It decrements |
| 171 | the nr_running variable. |
| 172 | |
| 173 | - yield_task(...) |
| 174 | |
| 175 | This function is basically just a dequeue followed by an enqueue, unless the |
| 176 | compat_yield sysctl is turned on; in that case, it places the scheduling |
| 177 | entity at the right-most end of the red-black tree. |
| 178 | |
| 179 | - check_preempt_curr(...) |
| 180 | |
| 181 | This function checks if a task that entered the runnable state should |
| 182 | preempt the currently running task. |
| 183 | |
| 184 | - pick_next_task(...) |
| 185 | |
| 186 | This function chooses the most appropriate task eligible to run next. |
| 187 | |
| 188 | - set_curr_task(...) |
| 189 | |
| 190 | This function is called when a task changes its scheduling class or changes |
| 191 | its task group. |
| 192 | |
| 193 | - task_tick(...) |
| 194 | |
| 195 | This function is mostly called from time tick functions; it might lead to |
| 196 | process switch. This drives the running preemption. |
| 197 | |
| 198 | - task_new(...) |
| 199 | |
| 200 | The core scheduler gives the scheduling module an opportunity to manage new |
| 201 | task startup. The CFS scheduling module uses it for group scheduling, while |
| 202 | the scheduling module for a real-time task does not use it. |
| 203 | |
| 204 | |
| 205 | |
Martin Steigerwald | 1a73ef6 | 2008-09-23 13:48:44 +0200 | [diff] [blame] | 206 | 7. GROUP SCHEDULER EXTENSIONS TO CFS |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 207 | |
| 208 | Normally, the scheduler operates on individual tasks and strives to provide |
| 209 | fair CPU time to each task. Sometimes, it may be desirable to group tasks and |
| 210 | provide fair CPU time to each such task group. For example, it may be |
| 211 | desirable to first provide fair CPU time to each user on the system and then to |
| 212 | each task belonging to a user. |
| 213 | |
| 214 | CONFIG_GROUP_SCHED strives to achieve exactly that. It lets tasks to be |
| 215 | grouped and divides CPU time fairly among such groups. |
| 216 | |
| 217 | CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and |
| 218 | SCHED_RR) tasks. |
| 219 | |
| 220 | CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and |
| 221 | SCHED_BATCH) tasks. |
| 222 | |
| 223 | At present, there are two (mutually exclusive) mechanisms to group tasks for |
| 224 | CPU bandwidth control purposes: |
| 225 | |
| 226 | - Based on user id (CONFIG_USER_SCHED) |
| 227 | |
| 228 | With this option, tasks are grouped according to their user id. |
| 229 | |
| 230 | - Based on "cgroup" pseudo filesystem (CONFIG_CGROUP_SCHED) |
| 231 | |
| 232 | This options needs CONFIG_CGROUPS to be defined, and lets the administrator |
| 233 | create arbitrary groups of tasks, using the "cgroup" pseudo filesystem. See |
| 234 | Documentation/cgroups.txt for more information about this filesystem. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 235 | |
| 236 | Only one of these options to group tasks can be chosen and not both. |
| 237 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 238 | When CONFIG_USER_SCHED is defined, a directory is created in sysfs for each new |
| 239 | user and a "cpu_share" file is added in that directory. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 240 | |
| 241 | # cd /sys/kernel/uids |
| 242 | # cat 512/cpu_share # Display user 512's CPU share |
| 243 | 1024 |
| 244 | # echo 2048 > 512/cpu_share # Modify user 512's CPU share |
| 245 | # cat 512/cpu_share # Display user 512's CPU share |
| 246 | 2048 |
| 247 | # |
| 248 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 249 | CPU bandwidth between two users is divided in the ratio of their CPU shares. |
| 250 | For example: if you would like user "root" to get twice the bandwidth of user |
| 251 | "guest," then set the cpu_share for both the users such that "root"'s cpu_share |
| 252 | is twice "guest"'s cpu_share. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 253 | |
Claudio Scordino | f58e2c3 | 2008-08-20 15:18:45 +0200 | [diff] [blame] | 254 | When CONFIG_CGROUP_SCHED is defined, a "cpu.shares" file is created for each |
| 255 | group created using the pseudo filesystem. See example steps below to create |
| 256 | task groups and modify their CPU share using the "cgroups" pseudo filesystem. |
Dhaval Giani | 5cb350b | 2007-10-15 17:00:14 +0200 | [diff] [blame] | 257 | |
| 258 | # mkdir /dev/cpuctl |
| 259 | # mount -t cgroup -ocpu none /dev/cpuctl |
| 260 | # cd /dev/cpuctl |
| 261 | |
| 262 | # mkdir multimedia # create "multimedia" group of tasks |
| 263 | # mkdir browser # create "browser" group of tasks |
| 264 | |
| 265 | # #Configure the multimedia group to receive twice the CPU bandwidth |
| 266 | # #that of browser group |
| 267 | |
| 268 | # echo 2048 > multimedia/cpu.shares |
| 269 | # echo 1024 > browser/cpu.shares |
| 270 | |
| 271 | # firefox & # Launch firefox and move it to "browser" group |
| 272 | # echo <firefox_pid> > browser/tasks |
| 273 | |
| 274 | # #Launch gmplayer (or your favourite movie player) |
| 275 | # echo <movie_player_pid> > multimedia/tasks |
Serge E. Hallyn | 94d6a5f | 2008-12-08 15:52:21 -0600 | [diff] [blame^] | 276 | |
| 277 | 8. Implementation note: user namespaces |
| 278 | |
| 279 | User namespaces are intended to be hierarchical. But they are currently |
| 280 | only partially implemented. Each of those has ramifications for CFS. |
| 281 | |
| 282 | First, since user namespaces are hierarchical, the /sys/kernel/uids |
| 283 | presentation is inadequate. Eventually we will likely want to use sysfs |
| 284 | tagging to provide private views of /sys/kernel/uids within each user |
| 285 | namespace. |
| 286 | |
| 287 | Second, the hierarchical nature is intended to support completely |
| 288 | unprivileged use of user namespaces. So if using user groups, then |
| 289 | we want the users in a user namespace to be children of the user |
| 290 | who created it. |
| 291 | |
| 292 | That is currently unimplemented. So instead, every user in a new |
| 293 | user namespace will receive 1024 shares just like any user in the |
| 294 | initial user namespace. Note that at the moment creation of a new |
| 295 | user namespace requires each of CAP_SYS_ADMIN, CAP_SETUID, and |
| 296 | CAP_SETGID. |