xref: /linux/Documentation/scheduler/sched-ext.rst (revision fcb3ad4366b9c810cbb9da34c076a9a52d8aa1e0)
1==========================
2Extensible Scheduler Class
3==========================
4
5sched_ext is a scheduler class whose behavior can be defined by a set of BPF
6programs - the BPF scheduler.
7
8* sched_ext exports a full scheduling interface so that any scheduling
9  algorithm can be implemented on top.
10
11* The BPF scheduler can group CPUs however it sees fit and schedule them
12  together, as tasks aren't tied to specific CPUs at the time of wakeup.
13
14* The BPF scheduler can be turned on and off dynamically anytime.
15
16* The system integrity is maintained no matter what the BPF scheduler does.
17  The default scheduling behavior is restored anytime an error is detected,
18  a runnable task stalls, or on invoking the SysRq key sequence
19  :kbd:`SysRq-S`.
20
21* When the BPF scheduler triggers an error, debug information is dumped to
22  aid debugging. The debug dump is passed to and printed out by the
23  scheduler binary. The debug dump can also be accessed through the
24  `sched_ext_dump` tracepoint. The SysRq key sequence :kbd:`SysRq-D`
25  triggers a debug dump. This doesn't terminate the BPF scheduler and can
26  only be read through the tracepoint.
27
28Switching to and from sched_ext
29===============================
30
31``CONFIG_SCHED_CLASS_EXT`` is the config option to enable sched_ext and
32``tools/sched_ext`` contains the example schedulers. The following config
33options should be enabled to use sched_ext:
34
35.. code-block:: none
36
37    CONFIG_BPF=y
38    CONFIG_SCHED_CLASS_EXT=y
39    CONFIG_BPF_SYSCALL=y
40    CONFIG_BPF_JIT=y
41    CONFIG_DEBUG_INFO_BTF=y
42    CONFIG_BPF_JIT_ALWAYS_ON=y
43    CONFIG_BPF_JIT_DEFAULT_ON=y
44    CONFIG_PAHOLE_HAS_SPLIT_BTF=y
45    CONFIG_PAHOLE_HAS_BTF_TAG=y
46
47sched_ext is used only when the BPF scheduler is loaded and running.
48
49If a task explicitly sets its scheduling policy to ``SCHED_EXT``, it will be
50treated as ``SCHED_NORMAL`` and scheduled by CFS until the BPF scheduler is
51loaded.
52
53When the BPF scheduler is loaded and ``SCX_OPS_SWITCH_PARTIAL`` is not set
54in ``ops->flags``, all ``SCHED_NORMAL``, ``SCHED_BATCH``, ``SCHED_IDLE``, and
55``SCHED_EXT`` tasks are scheduled by sched_ext.
56
57However, when the BPF scheduler is loaded and ``SCX_OPS_SWITCH_PARTIAL`` is
58set in ``ops->flags``, only tasks with the ``SCHED_EXT`` policy are scheduled
59by sched_ext, while tasks with ``SCHED_NORMAL``, ``SCHED_BATCH`` and
60``SCHED_IDLE`` policies are scheduled by CFS.
61
62Terminating the sched_ext scheduler program, triggering :kbd:`SysRq-S`, or
63detection of any internal error including stalled runnable tasks aborts the
64BPF scheduler and reverts all tasks back to CFS.
65
66.. code-block:: none
67
68    # make -j16 -C tools/sched_ext
69    # tools/sched_ext/build/bin/scx_simple
70    local=0 global=3
71    local=5 global=24
72    local=9 global=44
73    local=13 global=56
74    local=17 global=72
75    ^CEXIT: BPF scheduler unregistered
76
77The current status of the BPF scheduler can be determined as follows:
78
79.. code-block:: none
80
81    # cat /sys/kernel/sched_ext/state
82    enabled
83    # cat /sys/kernel/sched_ext/root/ops
84    simple
85
86You can check if any BPF scheduler has ever been loaded since boot by examining
87this monotonically incrementing counter (a value of zero indicates that no BPF
88scheduler has been loaded):
89
90.. code-block:: none
91
92    # cat /sys/kernel/sched_ext/enable_seq
93    1
94
95``tools/sched_ext/scx_show_state.py`` is a drgn script which shows more
96detailed information:
97
98.. code-block:: none
99
100    # tools/sched_ext/scx_show_state.py
101    ops           : simple
102    enabled       : 1
103    switching_all : 1
104    switched_all  : 1
105    enable_state  : enabled (2)
106    bypass_depth  : 0
107    nr_rejected   : 0
108    enable_seq    : 1
109
110If ``CONFIG_SCHED_DEBUG`` is set, whether a given task is on sched_ext can
111be determined as follows:
112
113.. code-block:: none
114
115    # grep ext /proc/self/sched
116    ext.enabled                                  :                    1
117
118The Basics
119==========
120
121Userspace can implement an arbitrary BPF scheduler by loading a set of BPF
122programs that implement ``struct sched_ext_ops``. The only mandatory field
123is ``ops.name`` which must be a valid BPF object name. All operations are
124optional. The following modified excerpt is from
125``tools/sched_ext/scx_simple.bpf.c`` showing a minimal global FIFO scheduler.
126
127.. code-block:: c
128
129    /*
130     * Decide which CPU a task should be migrated to before being
131     * enqueued (either at wakeup, fork time, or exec time). If an
132     * idle core is found by the default ops.select_cpu() implementation,
133     * then insert the task directly into SCX_DSQ_LOCAL and skip the
134     * ops.enqueue() callback.
135     *
136     * Note that this implementation has exactly the same behavior as the
137     * default ops.select_cpu implementation. The behavior of the scheduler
138     * would be exactly same if the implementation just didn't define the
139     * simple_select_cpu() struct_ops prog.
140     */
141    s32 BPF_STRUCT_OPS(simple_select_cpu, struct task_struct *p,
142                       s32 prev_cpu, u64 wake_flags)
143    {
144            s32 cpu;
145            /* Need to initialize or the BPF verifier will reject the program */
146            bool direct = false;
147
148            cpu = scx_bpf_select_cpu_dfl(p, prev_cpu, wake_flags, &direct);
149
150            if (direct)
151                    scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, SCX_SLICE_DFL, 0);
152
153            return cpu;
154    }
155
156    /*
157     * Do a direct insertion of a task to the global DSQ. This ops.enqueue()
158     * callback will only be invoked if we failed to find a core to insert
159     * into in ops.select_cpu() above.
160     *
161     * Note that this implementation has exactly the same behavior as the
162     * default ops.enqueue implementation, which just dispatches the task
163     * to SCX_DSQ_GLOBAL. The behavior of the scheduler would be exactly same
164     * if the implementation just didn't define the simple_enqueue struct_ops
165     * prog.
166     */
167    void BPF_STRUCT_OPS(simple_enqueue, struct task_struct *p, u64 enq_flags)
168    {
169            scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL, enq_flags);
170    }
171
172    s32 BPF_STRUCT_OPS_SLEEPABLE(simple_init)
173    {
174            /*
175             * By default, all SCHED_EXT, SCHED_OTHER, SCHED_IDLE, and
176             * SCHED_BATCH tasks should use sched_ext.
177             */
178            return 0;
179    }
180
181    void BPF_STRUCT_OPS(simple_exit, struct scx_exit_info *ei)
182    {
183            exit_type = ei->type;
184    }
185
186    SEC(".struct_ops")
187    struct sched_ext_ops simple_ops = {
188            .select_cpu             = (void *)simple_select_cpu,
189            .enqueue                = (void *)simple_enqueue,
190            .init                   = (void *)simple_init,
191            .exit                   = (void *)simple_exit,
192            .name                   = "simple",
193    };
194
195Dispatch Queues
196---------------
197
198To match the impedance between the scheduler core and the BPF scheduler,
199sched_ext uses DSQs (dispatch queues) which can operate as both a FIFO and a
200priority queue. By default, there is one global FIFO (``SCX_DSQ_GLOBAL``),
201and one local dsq per CPU (``SCX_DSQ_LOCAL``). The BPF scheduler can manage
202an arbitrary number of dsq's using ``scx_bpf_create_dsq()`` and
203``scx_bpf_destroy_dsq()``.
204
205A CPU always executes a task from its local DSQ. A task is "inserted" into a
206DSQ. A task in a non-local DSQ is "move"d into the target CPU's local DSQ.
207
208When a CPU is looking for the next task to run, if the local DSQ is not
209empty, the first task is picked. Otherwise, the CPU tries to move a task
210from the global DSQ. If that doesn't yield a runnable task either,
211``ops.dispatch()`` is invoked.
212
213Scheduling Cycle
214----------------
215
216The following briefly shows how a waking task is scheduled and executed.
217
2181. When a task is waking up, ``ops.select_cpu()`` is the first operation
219   invoked. This serves two purposes. First, CPU selection optimization
220   hint. Second, waking up the selected CPU if idle.
221
222   The CPU selected by ``ops.select_cpu()`` is an optimization hint and not
223   binding. The actual decision is made at the last step of scheduling.
224   However, there is a small performance gain if the CPU
225   ``ops.select_cpu()`` returns matches the CPU the task eventually runs on.
226
227   A side-effect of selecting a CPU is waking it up from idle. While a BPF
228   scheduler can wake up any cpu using the ``scx_bpf_kick_cpu()`` helper,
229   using ``ops.select_cpu()`` judiciously can be simpler and more efficient.
230
231   A task can be immediately inserted into a DSQ from ``ops.select_cpu()``
232   by calling ``scx_bpf_dsq_insert()``. If the task is inserted into
233   ``SCX_DSQ_LOCAL`` from ``ops.select_cpu()``, it will be inserted into the
234   local DSQ of whichever CPU is returned from ``ops.select_cpu()``.
235   Additionally, inserting directly from ``ops.select_cpu()`` will cause the
236   ``ops.enqueue()`` callback to be skipped.
237
238   Note that the scheduler core will ignore an invalid CPU selection, for
239   example, if it's outside the allowed cpumask of the task.
240
2412. Once the target CPU is selected, ``ops.enqueue()`` is invoked (unless the
242   task was inserted directly from ``ops.select_cpu()``). ``ops.enqueue()``
243   can make one of the following decisions:
244
245   * Immediately insert the task into either the global or local DSQ by
246     calling ``scx_bpf_dsq_insert()`` with ``SCX_DSQ_GLOBAL`` or
247     ``SCX_DSQ_LOCAL``, respectively.
248
249   * Immediately insert the task into a custom DSQ by calling
250     ``scx_bpf_dsq_insert()`` with a DSQ ID which is smaller than 2^63.
251
252   * Queue the task on the BPF side.
253
2543. When a CPU is ready to schedule, it first looks at its local DSQ. If
255   empty, it then looks at the global DSQ. If there still isn't a task to
256   run, ``ops.dispatch()`` is invoked which can use the following two
257   functions to populate the local DSQ.
258
259   * ``scx_bpf_dsq_insert()`` inserts a task to a DSQ. Any target DSQ can be
260     used - ``SCX_DSQ_LOCAL``, ``SCX_DSQ_LOCAL_ON | cpu``,
261     ``SCX_DSQ_GLOBAL`` or a custom DSQ. While ``scx_bpf_dsq_insert()``
262     currently can't be called with BPF locks held, this is being worked on
263     and will be supported. ``scx_bpf_dsq_insert()`` schedules insertion
264     rather than performing them immediately. There can be up to
265     ``ops.dispatch_max_batch`` pending tasks.
266
267   * ``scx_bpf_move_to_local()`` moves a task from the specified non-local
268     DSQ to the dispatching DSQ. This function cannot be called with any BPF
269     locks held. ``scx_bpf_move_to_local()`` flushes the pending insertions
270     tasks before trying to move from the specified DSQ.
271
2724. After ``ops.dispatch()`` returns, if there are tasks in the local DSQ,
273   the CPU runs the first one. If empty, the following steps are taken:
274
275   * Try to move from the global DSQ. If successful, run the task.
276
277   * If ``ops.dispatch()`` has dispatched any tasks, retry #3.
278
279   * If the previous task is an SCX task and still runnable, keep executing
280     it (see ``SCX_OPS_ENQ_LAST``).
281
282   * Go idle.
283
284Note that the BPF scheduler can always choose to dispatch tasks immediately
285in ``ops.enqueue()`` as illustrated in the above simple example. If only the
286built-in DSQs are used, there is no need to implement ``ops.dispatch()`` as
287a task is never queued on the BPF scheduler and both the local and global
288DSQs are executed automatically.
289
290``scx_bpf_dsq_insert()`` inserts the task on the FIFO of the target DSQ. Use
291``scx_bpf_dsq_insert_vtime()`` for the priority queue. Internal DSQs such as
292``SCX_DSQ_LOCAL`` and ``SCX_DSQ_GLOBAL`` do not support priority-queue
293dispatching, and must be dispatched to with ``scx_bpf_dsq_insert()``. See
294the function documentation and usage in ``tools/sched_ext/scx_simple.bpf.c``
295for more information.
296
297Where to Look
298=============
299
300* ``include/linux/sched/ext.h`` defines the core data structures, ops table
301  and constants.
302
303* ``kernel/sched/ext.c`` contains sched_ext core implementation and helpers.
304  The functions prefixed with ``scx_bpf_`` can be called from the BPF
305  scheduler.
306
307* ``tools/sched_ext/`` hosts example BPF scheduler implementations.
308
309  * ``scx_simple[.bpf].c``: Minimal global FIFO scheduler example using a
310    custom DSQ.
311
312  * ``scx_qmap[.bpf].c``: A multi-level FIFO scheduler supporting five
313    levels of priority implemented with ``BPF_MAP_TYPE_QUEUE``.
314
315ABI Instability
316===============
317
318The APIs provided by sched_ext to BPF schedulers programs have no stability
319guarantees. This includes the ops table callbacks and constants defined in
320``include/linux/sched/ext.h``, as well as the ``scx_bpf_`` kfuncs defined in
321``kernel/sched/ext.c``.
322
323While we will attempt to provide a relatively stable API surface when
324possible, they are subject to change without warning between kernel
325versions.
326