xref: /linux/kernel/sched/ext_internal.h (revision c7e739746dbde9ae401fd88824f5656c5e2361fc)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
4  *
5  * Copyright (c) 2025 Meta Platforms, Inc. and affiliates.
6  * Copyright (c) 2025 Tejun Heo <tj@kernel.org>
7  */
8 #define SCX_OP_IDX(op)		(offsetof(struct sched_ext_ops, op) / sizeof(void (*)(void)))
9 
10 enum scx_consts {
11 	SCX_DSP_DFL_MAX_BATCH		= 32,
12 	SCX_DSP_MAX_LOOPS		= 32,
13 	SCX_WATCHDOG_MAX_TIMEOUT	= 30 * HZ,
14 
15 	SCX_EXIT_BT_LEN			= 64,
16 	SCX_EXIT_MSG_LEN		= 1024,
17 	SCX_EXIT_DUMP_DFL_LEN		= 32768,
18 
19 	SCX_CPUPERF_ONE			= SCHED_CAPACITY_SCALE,
20 
21 	/*
22 	 * Iterating all tasks may take a while. Periodically drop
23 	 * scx_tasks_lock to avoid causing e.g. CSD and RCU stalls.
24 	 */
25 	SCX_TASK_ITER_BATCH		= 32,
26 };
27 
28 enum scx_exit_kind {
29 	SCX_EXIT_NONE,
30 	SCX_EXIT_DONE,
31 
32 	SCX_EXIT_UNREG = 64,	/* user-space initiated unregistration */
33 	SCX_EXIT_UNREG_BPF,	/* BPF-initiated unregistration */
34 	SCX_EXIT_UNREG_KERN,	/* kernel-initiated unregistration */
35 	SCX_EXIT_SYSRQ,		/* requested by 'S' sysrq */
36 
37 	SCX_EXIT_ERROR = 1024,	/* runtime error, error msg contains details */
38 	SCX_EXIT_ERROR_BPF,	/* ERROR but triggered through scx_bpf_error() */
39 	SCX_EXIT_ERROR_STALL,	/* watchdog detected stalled runnable tasks */
40 };
41 
42 /*
43  * An exit code can be specified when exiting with scx_bpf_exit() or scx_exit(),
44  * corresponding to exit_kind UNREG_BPF and UNREG_KERN respectively. The codes
45  * are 64bit of the format:
46  *
47  *   Bits: [63  ..  48 47   ..  32 31 .. 0]
48  *         [ SYS ACT ] [ SYS RSN ] [ USR  ]
49  *
50  *   SYS ACT: System-defined exit actions
51  *   SYS RSN: System-defined exit reasons
52  *   USR    : User-defined exit codes and reasons
53  *
54  * Using the above, users may communicate intention and context by ORing system
55  * actions and/or system reasons with a user-defined exit code.
56  */
57 enum scx_exit_code {
58 	/* Reasons */
59 	SCX_ECODE_RSN_HOTPLUG	= 1LLU << 32,
60 
61 	/* Actions */
62 	SCX_ECODE_ACT_RESTART	= 1LLU << 48,
63 };
64 
65 /*
66  * scx_exit_info is passed to ops.exit() to describe why the BPF scheduler is
67  * being disabled.
68  */
69 struct scx_exit_info {
70 	/* %SCX_EXIT_* - broad category of the exit reason */
71 	enum scx_exit_kind	kind;
72 
73 	/* exit code if gracefully exiting */
74 	s64			exit_code;
75 
76 	/* textual representation of the above */
77 	const char		*reason;
78 
79 	/* backtrace if exiting due to an error */
80 	unsigned long		*bt;
81 	u32			bt_len;
82 
83 	/* informational message */
84 	char			*msg;
85 
86 	/* debug dump */
87 	char			*dump;
88 };
89 
90 /* sched_ext_ops.flags */
91 enum scx_ops_flags {
92 	/*
93 	 * Keep built-in idle tracking even if ops.update_idle() is implemented.
94 	 */
95 	SCX_OPS_KEEP_BUILTIN_IDLE	= 1LLU << 0,
96 
97 	/*
98 	 * By default, if there are no other task to run on the CPU, ext core
99 	 * keeps running the current task even after its slice expires. If this
100 	 * flag is specified, such tasks are passed to ops.enqueue() with
101 	 * %SCX_ENQ_LAST. See the comment above %SCX_ENQ_LAST for more info.
102 	 */
103 	SCX_OPS_ENQ_LAST		= 1LLU << 1,
104 
105 	/*
106 	 * An exiting task may schedule after PF_EXITING is set. In such cases,
107 	 * bpf_task_from_pid() may not be able to find the task and if the BPF
108 	 * scheduler depends on pid lookup for dispatching, the task will be
109 	 * lost leading to various issues including RCU grace period stalls.
110 	 *
111 	 * To mask this problem, by default, unhashed tasks are automatically
112 	 * dispatched to the local DSQ on enqueue. If the BPF scheduler doesn't
113 	 * depend on pid lookups and wants to handle these tasks directly, the
114 	 * following flag can be used.
115 	 */
116 	SCX_OPS_ENQ_EXITING		= 1LLU << 2,
117 
118 	/*
119 	 * If set, only tasks with policy set to SCHED_EXT are attached to
120 	 * sched_ext. If clear, SCHED_NORMAL tasks are also included.
121 	 */
122 	SCX_OPS_SWITCH_PARTIAL		= 1LLU << 3,
123 
124 	/*
125 	 * A migration disabled task can only execute on its current CPU. By
126 	 * default, such tasks are automatically put on the CPU's local DSQ with
127 	 * the default slice on enqueue. If this ops flag is set, they also go
128 	 * through ops.enqueue().
129 	 *
130 	 * A migration disabled task never invokes ops.select_cpu() as it can
131 	 * only select the current CPU. Also, p->cpus_ptr will only contain its
132 	 * current CPU while p->nr_cpus_allowed keeps tracking p->user_cpus_ptr
133 	 * and thus may disagree with cpumask_weight(p->cpus_ptr).
134 	 */
135 	SCX_OPS_ENQ_MIGRATION_DISABLED	= 1LLU << 4,
136 
137 	/*
138 	 * Queued wakeup (ttwu_queue) is a wakeup optimization that invokes
139 	 * ops.enqueue() on the ops.select_cpu() selected or the wakee's
140 	 * previous CPU via IPI (inter-processor interrupt) to reduce cacheline
141 	 * transfers. When this optimization is enabled, ops.select_cpu() is
142 	 * skipped in some cases (when racing against the wakee switching out).
143 	 * As the BPF scheduler may depend on ops.select_cpu() being invoked
144 	 * during wakeups, queued wakeup is disabled by default.
145 	 *
146 	 * If this ops flag is set, queued wakeup optimization is enabled and
147 	 * the BPF scheduler must be able to handle ops.enqueue() invoked on the
148 	 * wakee's CPU without preceding ops.select_cpu() even for tasks which
149 	 * may be executed on multiple CPUs.
150 	 */
151 	SCX_OPS_ALLOW_QUEUED_WAKEUP	= 1LLU << 5,
152 
153 	/*
154 	 * If set, enable per-node idle cpumasks. If clear, use a single global
155 	 * flat idle cpumask.
156 	 */
157 	SCX_OPS_BUILTIN_IDLE_PER_NODE	= 1LLU << 6,
158 
159 	/*
160 	 * CPU cgroup support flags
161 	 */
162 	SCX_OPS_HAS_CGROUP_WEIGHT	= 1LLU << 16,	/* DEPRECATED, will be removed on 6.18 */
163 
164 	SCX_OPS_ALL_FLAGS		= SCX_OPS_KEEP_BUILTIN_IDLE |
165 					  SCX_OPS_ENQ_LAST |
166 					  SCX_OPS_ENQ_EXITING |
167 					  SCX_OPS_ENQ_MIGRATION_DISABLED |
168 					  SCX_OPS_ALLOW_QUEUED_WAKEUP |
169 					  SCX_OPS_SWITCH_PARTIAL |
170 					  SCX_OPS_BUILTIN_IDLE_PER_NODE |
171 					  SCX_OPS_HAS_CGROUP_WEIGHT,
172 
173 	/* high 8 bits are internal, don't include in SCX_OPS_ALL_FLAGS */
174 	__SCX_OPS_INTERNAL_MASK		= 0xffLLU << 56,
175 
176 	SCX_OPS_HAS_CPU_PREEMPT		= 1LLU << 56,
177 };
178 
179 /* argument container for ops.init_task() */
180 struct scx_init_task_args {
181 	/*
182 	 * Set if ops.init_task() is being invoked on the fork path, as opposed
183 	 * to the scheduler transition path.
184 	 */
185 	bool			fork;
186 #ifdef CONFIG_EXT_GROUP_SCHED
187 	/* the cgroup the task is joining */
188 	struct cgroup		*cgroup;
189 #endif
190 };
191 
192 /* argument container for ops.exit_task() */
193 struct scx_exit_task_args {
194 	/* Whether the task exited before running on sched_ext. */
195 	bool cancelled;
196 };
197 
198 /* argument container for ops->cgroup_init() */
199 struct scx_cgroup_init_args {
200 	/* the weight of the cgroup [1..10000] */
201 	u32			weight;
202 
203 	/* bandwidth control parameters from cpu.max and cpu.max.burst */
204 	u64			bw_period_us;
205 	u64			bw_quota_us;
206 	u64			bw_burst_us;
207 };
208 
209 enum scx_cpu_preempt_reason {
210 	/* next task is being scheduled by &sched_class_rt */
211 	SCX_CPU_PREEMPT_RT,
212 	/* next task is being scheduled by &sched_class_dl */
213 	SCX_CPU_PREEMPT_DL,
214 	/* next task is being scheduled by &sched_class_stop */
215 	SCX_CPU_PREEMPT_STOP,
216 	/* unknown reason for SCX being preempted */
217 	SCX_CPU_PREEMPT_UNKNOWN,
218 };
219 
220 /*
221  * Argument container for ops->cpu_acquire(). Currently empty, but may be
222  * expanded in the future.
223  */
224 struct scx_cpu_acquire_args {};
225 
226 /* argument container for ops->cpu_release() */
227 struct scx_cpu_release_args {
228 	/* the reason the CPU was preempted */
229 	enum scx_cpu_preempt_reason reason;
230 
231 	/* the task that's going to be scheduled on the CPU */
232 	struct task_struct	*task;
233 };
234 
235 /*
236  * Informational context provided to dump operations.
237  */
238 struct scx_dump_ctx {
239 	enum scx_exit_kind	kind;
240 	s64			exit_code;
241 	const char		*reason;
242 	u64			at_ns;
243 	u64			at_jiffies;
244 };
245 
246 /**
247  * struct sched_ext_ops - Operation table for BPF scheduler implementation
248  *
249  * A BPF scheduler can implement an arbitrary scheduling policy by
250  * implementing and loading operations in this table. Note that a userland
251  * scheduling policy can also be implemented using the BPF scheduler
252  * as a shim layer.
253  */
254 struct sched_ext_ops {
255 	/**
256 	 * @select_cpu: Pick the target CPU for a task which is being woken up
257 	 * @p: task being woken up
258 	 * @prev_cpu: the cpu @p was on before sleeping
259 	 * @wake_flags: SCX_WAKE_*
260 	 *
261 	 * Decision made here isn't final. @p may be moved to any CPU while it
262 	 * is getting dispatched for execution later. However, as @p is not on
263 	 * the rq at this point, getting the eventual execution CPU right here
264 	 * saves a small bit of overhead down the line.
265 	 *
266 	 * If an idle CPU is returned, the CPU is kicked and will try to
267 	 * dispatch. While an explicit custom mechanism can be added,
268 	 * select_cpu() serves as the default way to wake up idle CPUs.
269 	 *
270 	 * @p may be inserted into a DSQ directly by calling
271 	 * scx_bpf_dsq_insert(). If so, the ops.enqueue() will be skipped.
272 	 * Directly inserting into %SCX_DSQ_LOCAL will put @p in the local DSQ
273 	 * of the CPU returned by this operation.
274 	 *
275 	 * Note that select_cpu() is never called for tasks that can only run
276 	 * on a single CPU or tasks with migration disabled, as they don't have
277 	 * the option to select a different CPU. See select_task_rq() for
278 	 * details.
279 	 */
280 	s32 (*select_cpu)(struct task_struct *p, s32 prev_cpu, u64 wake_flags);
281 
282 	/**
283 	 * @enqueue: Enqueue a task on the BPF scheduler
284 	 * @p: task being enqueued
285 	 * @enq_flags: %SCX_ENQ_*
286 	 *
287 	 * @p is ready to run. Insert directly into a DSQ by calling
288 	 * scx_bpf_dsq_insert() or enqueue on the BPF scheduler. If not directly
289 	 * inserted, the bpf scheduler owns @p and if it fails to dispatch @p,
290 	 * the task will stall.
291 	 *
292 	 * If @p was inserted into a DSQ from ops.select_cpu(), this callback is
293 	 * skipped.
294 	 */
295 	void (*enqueue)(struct task_struct *p, u64 enq_flags);
296 
297 	/**
298 	 * @dequeue: Remove a task from the BPF scheduler
299 	 * @p: task being dequeued
300 	 * @deq_flags: %SCX_DEQ_*
301 	 *
302 	 * Remove @p from the BPF scheduler. This is usually called to isolate
303 	 * the task while updating its scheduling properties (e.g. priority).
304 	 *
305 	 * The ext core keeps track of whether the BPF side owns a given task or
306 	 * not and can gracefully ignore spurious dispatches from BPF side,
307 	 * which makes it safe to not implement this method. However, depending
308 	 * on the scheduling logic, this can lead to confusing behaviors - e.g.
309 	 * scheduling position not being updated across a priority change.
310 	 */
311 	void (*dequeue)(struct task_struct *p, u64 deq_flags);
312 
313 	/**
314 	 * @dispatch: Dispatch tasks from the BPF scheduler and/or user DSQs
315 	 * @cpu: CPU to dispatch tasks for
316 	 * @prev: previous task being switched out
317 	 *
318 	 * Called when a CPU's local dsq is empty. The operation should dispatch
319 	 * one or more tasks from the BPF scheduler into the DSQs using
320 	 * scx_bpf_dsq_insert() and/or move from user DSQs into the local DSQ
321 	 * using scx_bpf_dsq_move_to_local().
322 	 *
323 	 * The maximum number of times scx_bpf_dsq_insert() can be called
324 	 * without an intervening scx_bpf_dsq_move_to_local() is specified by
325 	 * ops.dispatch_max_batch. See the comments on top of the two functions
326 	 * for more details.
327 	 *
328 	 * When not %NULL, @prev is an SCX task with its slice depleted. If
329 	 * @prev is still runnable as indicated by set %SCX_TASK_QUEUED in
330 	 * @prev->scx.flags, it is not enqueued yet and will be enqueued after
331 	 * ops.dispatch() returns. To keep executing @prev, return without
332 	 * dispatching or moving any tasks. Also see %SCX_OPS_ENQ_LAST.
333 	 */
334 	void (*dispatch)(s32 cpu, struct task_struct *prev);
335 
336 	/**
337 	 * @tick: Periodic tick
338 	 * @p: task running currently
339 	 *
340 	 * This operation is called every 1/HZ seconds on CPUs which are
341 	 * executing an SCX task. Setting @p->scx.slice to 0 will trigger an
342 	 * immediate dispatch cycle on the CPU.
343 	 */
344 	void (*tick)(struct task_struct *p);
345 
346 	/**
347 	 * @runnable: A task is becoming runnable on its associated CPU
348 	 * @p: task becoming runnable
349 	 * @enq_flags: %SCX_ENQ_*
350 	 *
351 	 * This and the following three functions can be used to track a task's
352 	 * execution state transitions. A task becomes ->runnable() on a CPU,
353 	 * and then goes through one or more ->running() and ->stopping() pairs
354 	 * as it runs on the CPU, and eventually becomes ->quiescent() when it's
355 	 * done running on the CPU.
356 	 *
357 	 * @p is becoming runnable on the CPU because it's
358 	 *
359 	 * - waking up (%SCX_ENQ_WAKEUP)
360 	 * - being moved from another CPU
361 	 * - being restored after temporarily taken off the queue for an
362 	 *   attribute change.
363 	 *
364 	 * This and ->enqueue() are related but not coupled. This operation
365 	 * notifies @p's state transition and may not be followed by ->enqueue()
366 	 * e.g. when @p is being dispatched to a remote CPU, or when @p is
367 	 * being enqueued on a CPU experiencing a hotplug event. Likewise, a
368 	 * task may be ->enqueue()'d without being preceded by this operation
369 	 * e.g. after exhausting its slice.
370 	 */
371 	void (*runnable)(struct task_struct *p, u64 enq_flags);
372 
373 	/**
374 	 * @running: A task is starting to run on its associated CPU
375 	 * @p: task starting to run
376 	 *
377 	 * Note that this callback may be called from a CPU other than the
378 	 * one the task is going to run on. This can happen when a task
379 	 * property is changed (i.e., affinity), since scx_next_task_scx(),
380 	 * which triggers this callback, may run on a CPU different from
381 	 * the task's assigned CPU.
382 	 *
383 	 * Therefore, always use scx_bpf_task_cpu(@p) to determine the
384 	 * target CPU the task is going to use.
385 	 *
386 	 * See ->runnable() for explanation on the task state notifiers.
387 	 */
388 	void (*running)(struct task_struct *p);
389 
390 	/**
391 	 * @stopping: A task is stopping execution
392 	 * @p: task stopping to run
393 	 * @runnable: is task @p still runnable?
394 	 *
395 	 * Note that this callback may be called from a CPU other than the
396 	 * one the task was running on. This can happen when a task
397 	 * property is changed (i.e., affinity), since dequeue_task_scx(),
398 	 * which triggers this callback, may run on a CPU different from
399 	 * the task's assigned CPU.
400 	 *
401 	 * Therefore, always use scx_bpf_task_cpu(@p) to retrieve the CPU
402 	 * the task was running on.
403 	 *
404 	 * See ->runnable() for explanation on the task state notifiers. If
405 	 * !@runnable, ->quiescent() will be invoked after this operation
406 	 * returns.
407 	 */
408 	void (*stopping)(struct task_struct *p, bool runnable);
409 
410 	/**
411 	 * @quiescent: A task is becoming not runnable on its associated CPU
412 	 * @p: task becoming not runnable
413 	 * @deq_flags: %SCX_DEQ_*
414 	 *
415 	 * See ->runnable() for explanation on the task state notifiers.
416 	 *
417 	 * @p is becoming quiescent on the CPU because it's
418 	 *
419 	 * - sleeping (%SCX_DEQ_SLEEP)
420 	 * - being moved to another CPU
421 	 * - being temporarily taken off the queue for an attribute change
422 	 *   (%SCX_DEQ_SAVE)
423 	 *
424 	 * This and ->dequeue() are related but not coupled. This operation
425 	 * notifies @p's state transition and may not be preceded by ->dequeue()
426 	 * e.g. when @p is being dispatched to a remote CPU.
427 	 */
428 	void (*quiescent)(struct task_struct *p, u64 deq_flags);
429 
430 	/**
431 	 * @yield: Yield CPU
432 	 * @from: yielding task
433 	 * @to: optional yield target task
434 	 *
435 	 * If @to is NULL, @from is yielding the CPU to other runnable tasks.
436 	 * The BPF scheduler should ensure that other available tasks are
437 	 * dispatched before the yielding task. Return value is ignored in this
438 	 * case.
439 	 *
440 	 * If @to is not-NULL, @from wants to yield the CPU to @to. If the bpf
441 	 * scheduler can implement the request, return %true; otherwise, %false.
442 	 */
443 	bool (*yield)(struct task_struct *from, struct task_struct *to);
444 
445 	/**
446 	 * @core_sched_before: Task ordering for core-sched
447 	 * @a: task A
448 	 * @b: task B
449 	 *
450 	 * Used by core-sched to determine the ordering between two tasks. See
451 	 * Documentation/admin-guide/hw-vuln/core-scheduling.rst for details on
452 	 * core-sched.
453 	 *
454 	 * Both @a and @b are runnable and may or may not currently be queued on
455 	 * the BPF scheduler. Should return %true if @a should run before @b.
456 	 * %false if there's no required ordering or @b should run before @a.
457 	 *
458 	 * If not specified, the default is ordering them according to when they
459 	 * became runnable.
460 	 */
461 	bool (*core_sched_before)(struct task_struct *a, struct task_struct *b);
462 
463 	/**
464 	 * @set_weight: Set task weight
465 	 * @p: task to set weight for
466 	 * @weight: new weight [1..10000]
467 	 *
468 	 * Update @p's weight to @weight.
469 	 */
470 	void (*set_weight)(struct task_struct *p, u32 weight);
471 
472 	/**
473 	 * @set_cpumask: Set CPU affinity
474 	 * @p: task to set CPU affinity for
475 	 * @cpumask: cpumask of cpus that @p can run on
476 	 *
477 	 * Update @p's CPU affinity to @cpumask.
478 	 */
479 	void (*set_cpumask)(struct task_struct *p,
480 			    const struct cpumask *cpumask);
481 
482 	/**
483 	 * @update_idle: Update the idle state of a CPU
484 	 * @cpu: CPU to update the idle state for
485 	 * @idle: whether entering or exiting the idle state
486 	 *
487 	 * This operation is called when @rq's CPU goes or leaves the idle
488 	 * state. By default, implementing this operation disables the built-in
489 	 * idle CPU tracking and the following helpers become unavailable:
490 	 *
491 	 * - scx_bpf_select_cpu_dfl()
492 	 * - scx_bpf_select_cpu_and()
493 	 * - scx_bpf_test_and_clear_cpu_idle()
494 	 * - scx_bpf_pick_idle_cpu()
495 	 *
496 	 * The user also must implement ops.select_cpu() as the default
497 	 * implementation relies on scx_bpf_select_cpu_dfl().
498 	 *
499 	 * Specify the %SCX_OPS_KEEP_BUILTIN_IDLE flag to keep the built-in idle
500 	 * tracking.
501 	 */
502 	void (*update_idle)(s32 cpu, bool idle);
503 
504 	/**
505 	 * @cpu_acquire: A CPU is becoming available to the BPF scheduler
506 	 * @cpu: The CPU being acquired by the BPF scheduler.
507 	 * @args: Acquire arguments, see the struct definition.
508 	 *
509 	 * A CPU that was previously released from the BPF scheduler is now once
510 	 * again under its control.
511 	 */
512 	void (*cpu_acquire)(s32 cpu, struct scx_cpu_acquire_args *args);
513 
514 	/**
515 	 * @cpu_release: A CPU is taken away from the BPF scheduler
516 	 * @cpu: The CPU being released by the BPF scheduler.
517 	 * @args: Release arguments, see the struct definition.
518 	 *
519 	 * The specified CPU is no longer under the control of the BPF
520 	 * scheduler. This could be because it was preempted by a higher
521 	 * priority sched_class, though there may be other reasons as well. The
522 	 * caller should consult @args->reason to determine the cause.
523 	 */
524 	void (*cpu_release)(s32 cpu, struct scx_cpu_release_args *args);
525 
526 	/**
527 	 * @init_task: Initialize a task to run in a BPF scheduler
528 	 * @p: task to initialize for BPF scheduling
529 	 * @args: init arguments, see the struct definition
530 	 *
531 	 * Either we're loading a BPF scheduler or a new task is being forked.
532 	 * Initialize @p for BPF scheduling. This operation may block and can
533 	 * be used for allocations, and is called exactly once for a task.
534 	 *
535 	 * Return 0 for success, -errno for failure. An error return while
536 	 * loading will abort loading of the BPF scheduler. During a fork, it
537 	 * will abort that specific fork.
538 	 */
539 	s32 (*init_task)(struct task_struct *p, struct scx_init_task_args *args);
540 
541 	/**
542 	 * @exit_task: Exit a previously-running task from the system
543 	 * @p: task to exit
544 	 * @args: exit arguments, see the struct definition
545 	 *
546 	 * @p is exiting or the BPF scheduler is being unloaded. Perform any
547 	 * necessary cleanup for @p.
548 	 */
549 	void (*exit_task)(struct task_struct *p, struct scx_exit_task_args *args);
550 
551 	/**
552 	 * @enable: Enable BPF scheduling for a task
553 	 * @p: task to enable BPF scheduling for
554 	 *
555 	 * Enable @p for BPF scheduling. enable() is called on @p any time it
556 	 * enters SCX, and is always paired with a matching disable().
557 	 */
558 	void (*enable)(struct task_struct *p);
559 
560 	/**
561 	 * @disable: Disable BPF scheduling for a task
562 	 * @p: task to disable BPF scheduling for
563 	 *
564 	 * @p is exiting, leaving SCX or the BPF scheduler is being unloaded.
565 	 * Disable BPF scheduling for @p. A disable() call is always matched
566 	 * with a prior enable() call.
567 	 */
568 	void (*disable)(struct task_struct *p);
569 
570 	/**
571 	 * @dump: Dump BPF scheduler state on error
572 	 * @ctx: debug dump context
573 	 *
574 	 * Use scx_bpf_dump() to generate BPF scheduler specific debug dump.
575 	 */
576 	void (*dump)(struct scx_dump_ctx *ctx);
577 
578 	/**
579 	 * @dump_cpu: Dump BPF scheduler state for a CPU on error
580 	 * @ctx: debug dump context
581 	 * @cpu: CPU to generate debug dump for
582 	 * @idle: @cpu is currently idle without any runnable tasks
583 	 *
584 	 * Use scx_bpf_dump() to generate BPF scheduler specific debug dump for
585 	 * @cpu. If @idle is %true and this operation doesn't produce any
586 	 * output, @cpu is skipped for dump.
587 	 */
588 	void (*dump_cpu)(struct scx_dump_ctx *ctx, s32 cpu, bool idle);
589 
590 	/**
591 	 * @dump_task: Dump BPF scheduler state for a runnable task on error
592 	 * @ctx: debug dump context
593 	 * @p: runnable task to generate debug dump for
594 	 *
595 	 * Use scx_bpf_dump() to generate BPF scheduler specific debug dump for
596 	 * @p.
597 	 */
598 	void (*dump_task)(struct scx_dump_ctx *ctx, struct task_struct *p);
599 
600 #ifdef CONFIG_EXT_GROUP_SCHED
601 	/**
602 	 * @cgroup_init: Initialize a cgroup
603 	 * @cgrp: cgroup being initialized
604 	 * @args: init arguments, see the struct definition
605 	 *
606 	 * Either the BPF scheduler is being loaded or @cgrp created, initialize
607 	 * @cgrp for sched_ext. This operation may block.
608 	 *
609 	 * Return 0 for success, -errno for failure. An error return while
610 	 * loading will abort loading of the BPF scheduler. During cgroup
611 	 * creation, it will abort the specific cgroup creation.
612 	 */
613 	s32 (*cgroup_init)(struct cgroup *cgrp,
614 			   struct scx_cgroup_init_args *args);
615 
616 	/**
617 	 * @cgroup_exit: Exit a cgroup
618 	 * @cgrp: cgroup being exited
619 	 *
620 	 * Either the BPF scheduler is being unloaded or @cgrp destroyed, exit
621 	 * @cgrp for sched_ext. This operation my block.
622 	 */
623 	void (*cgroup_exit)(struct cgroup *cgrp);
624 
625 	/**
626 	 * @cgroup_prep_move: Prepare a task to be moved to a different cgroup
627 	 * @p: task being moved
628 	 * @from: cgroup @p is being moved from
629 	 * @to: cgroup @p is being moved to
630 	 *
631 	 * Prepare @p for move from cgroup @from to @to. This operation may
632 	 * block and can be used for allocations.
633 	 *
634 	 * Return 0 for success, -errno for failure. An error return aborts the
635 	 * migration.
636 	 */
637 	s32 (*cgroup_prep_move)(struct task_struct *p,
638 				struct cgroup *from, struct cgroup *to);
639 
640 	/**
641 	 * @cgroup_move: Commit cgroup move
642 	 * @p: task being moved
643 	 * @from: cgroup @p is being moved from
644 	 * @to: cgroup @p is being moved to
645 	 *
646 	 * Commit the move. @p is dequeued during this operation.
647 	 */
648 	void (*cgroup_move)(struct task_struct *p,
649 			    struct cgroup *from, struct cgroup *to);
650 
651 	/**
652 	 * @cgroup_cancel_move: Cancel cgroup move
653 	 * @p: task whose cgroup move is being canceled
654 	 * @from: cgroup @p was being moved from
655 	 * @to: cgroup @p was being moved to
656 	 *
657 	 * @p was cgroup_prep_move()'d but failed before reaching cgroup_move().
658 	 * Undo the preparation.
659 	 */
660 	void (*cgroup_cancel_move)(struct task_struct *p,
661 				   struct cgroup *from, struct cgroup *to);
662 
663 	/**
664 	 * @cgroup_set_weight: A cgroup's weight is being changed
665 	 * @cgrp: cgroup whose weight is being updated
666 	 * @weight: new weight [1..10000]
667 	 *
668 	 * Update @cgrp's weight to @weight.
669 	 */
670 	void (*cgroup_set_weight)(struct cgroup *cgrp, u32 weight);
671 
672 	/**
673 	 * @cgroup_set_bandwidth: A cgroup's bandwidth is being changed
674 	 * @cgrp: cgroup whose bandwidth is being updated
675 	 * @period_us: bandwidth control period
676 	 * @quota_us: bandwidth control quota
677 	 * @burst_us: bandwidth control burst
678 	 *
679 	 * Update @cgrp's bandwidth control parameters. This is from the cpu.max
680 	 * cgroup interface.
681 	 *
682 	 * @quota_us / @period_us determines the CPU bandwidth @cgrp is entitled
683 	 * to. For example, if @period_us is 1_000_000 and @quota_us is
684 	 * 2_500_000. @cgrp is entitled to 2.5 CPUs. @burst_us can be
685 	 * interpreted in the same fashion and specifies how much @cgrp can
686 	 * burst temporarily. The specific control mechanism and thus the
687 	 * interpretation of @period_us and burstiness is upto to the BPF
688 	 * scheduler.
689 	 */
690 	void (*cgroup_set_bandwidth)(struct cgroup *cgrp,
691 				     u64 period_us, u64 quota_us, u64 burst_us);
692 
693 #endif	/* CONFIG_EXT_GROUP_SCHED */
694 
695 	/*
696 	 * All online ops must come before ops.cpu_online().
697 	 */
698 
699 	/**
700 	 * @cpu_online: A CPU became online
701 	 * @cpu: CPU which just came up
702 	 *
703 	 * @cpu just came online. @cpu will not call ops.enqueue() or
704 	 * ops.dispatch(), nor run tasks associated with other CPUs beforehand.
705 	 */
706 	void (*cpu_online)(s32 cpu);
707 
708 	/**
709 	 * @cpu_offline: A CPU is going offline
710 	 * @cpu: CPU which is going offline
711 	 *
712 	 * @cpu is going offline. @cpu will not call ops.enqueue() or
713 	 * ops.dispatch(), nor run tasks associated with other CPUs afterwards.
714 	 */
715 	void (*cpu_offline)(s32 cpu);
716 
717 	/*
718 	 * All CPU hotplug ops must come before ops.init().
719 	 */
720 
721 	/**
722 	 * @init: Initialize the BPF scheduler
723 	 */
724 	s32 (*init)(void);
725 
726 	/**
727 	 * @exit: Clean up after the BPF scheduler
728 	 * @info: Exit info
729 	 *
730 	 * ops.exit() is also called on ops.init() failure, which is a bit
731 	 * unusual. This is to allow rich reporting through @info on how
732 	 * ops.init() failed.
733 	 */
734 	void (*exit)(struct scx_exit_info *info);
735 
736 	/**
737 	 * @dispatch_max_batch: Max nr of tasks that dispatch() can dispatch
738 	 */
739 	u32 dispatch_max_batch;
740 
741 	/**
742 	 * @flags: %SCX_OPS_* flags
743 	 */
744 	u64 flags;
745 
746 	/**
747 	 * @timeout_ms: The maximum amount of time, in milliseconds, that a
748 	 * runnable task should be able to wait before being scheduled. The
749 	 * maximum timeout may not exceed the default timeout of 30 seconds.
750 	 *
751 	 * Defaults to the maximum allowed timeout value of 30 seconds.
752 	 */
753 	u32 timeout_ms;
754 
755 	/**
756 	 * @exit_dump_len: scx_exit_info.dump buffer length. If 0, the default
757 	 * value of 32768 is used.
758 	 */
759 	u32 exit_dump_len;
760 
761 	/**
762 	 * @hotplug_seq: A sequence number that may be set by the scheduler to
763 	 * detect when a hotplug event has occurred during the loading process.
764 	 * If 0, no detection occurs. Otherwise, the scheduler will fail to
765 	 * load if the sequence number does not match @scx_hotplug_seq on the
766 	 * enable path.
767 	 */
768 	u64 hotplug_seq;
769 
770 	/**
771 	 * @name: BPF scheduler's name
772 	 *
773 	 * Must be a non-zero valid BPF object name including only isalnum(),
774 	 * '_' and '.' chars. Shows up in kernel.sched_ext_ops sysctl while the
775 	 * BPF scheduler is enabled.
776 	 */
777 	char name[SCX_OPS_NAME_LEN];
778 
779 	/* internal use only, must be NULL */
780 	void *priv;
781 };
782 
783 enum scx_opi {
784 	SCX_OPI_BEGIN			= 0,
785 	SCX_OPI_NORMAL_BEGIN		= 0,
786 	SCX_OPI_NORMAL_END		= SCX_OP_IDX(cpu_online),
787 	SCX_OPI_CPU_HOTPLUG_BEGIN	= SCX_OP_IDX(cpu_online),
788 	SCX_OPI_CPU_HOTPLUG_END		= SCX_OP_IDX(init),
789 	SCX_OPI_END			= SCX_OP_IDX(init),
790 };
791 
792 /*
793  * Collection of event counters. Event types are placed in descending order.
794  */
795 struct scx_event_stats {
796 	/*
797 	 * If ops.select_cpu() returns a CPU which can't be used by the task,
798 	 * the core scheduler code silently picks a fallback CPU.
799 	 */
800 	s64		SCX_EV_SELECT_CPU_FALLBACK;
801 
802 	/*
803 	 * When dispatching to a local DSQ, the CPU may have gone offline in
804 	 * the meantime. In this case, the task is bounced to the global DSQ.
805 	 */
806 	s64		SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE;
807 
808 	/*
809 	 * If SCX_OPS_ENQ_LAST is not set, the number of times that a task
810 	 * continued to run because there were no other tasks on the CPU.
811 	 */
812 	s64		SCX_EV_DISPATCH_KEEP_LAST;
813 
814 	/*
815 	 * If SCX_OPS_ENQ_EXITING is not set, the number of times that a task
816 	 * is dispatched to a local DSQ when exiting.
817 	 */
818 	s64		SCX_EV_ENQ_SKIP_EXITING;
819 
820 	/*
821 	 * If SCX_OPS_ENQ_MIGRATION_DISABLED is not set, the number of times a
822 	 * migration disabled task skips ops.enqueue() and is dispatched to its
823 	 * local DSQ.
824 	 */
825 	s64		SCX_EV_ENQ_SKIP_MIGRATION_DISABLED;
826 
827 	/*
828 	 * Total number of times a task's time slice was refilled with the
829 	 * default value (SCX_SLICE_DFL).
830 	 */
831 	s64		SCX_EV_REFILL_SLICE_DFL;
832 
833 	/*
834 	 * The total duration of bypass modes in nanoseconds.
835 	 */
836 	s64		SCX_EV_BYPASS_DURATION;
837 
838 	/*
839 	 * The number of tasks dispatched in the bypassing mode.
840 	 */
841 	s64		SCX_EV_BYPASS_DISPATCH;
842 
843 	/*
844 	 * The number of times the bypassing mode has been activated.
845 	 */
846 	s64		SCX_EV_BYPASS_ACTIVATE;
847 };
848 
849 struct scx_sched_pcpu {
850 	/*
851 	 * The event counters are in a per-CPU variable to minimize the
852 	 * accounting overhead. A system-wide view on the event counter is
853 	 * constructed when requested by scx_bpf_events().
854 	 */
855 	struct scx_event_stats	event_stats;
856 };
857 
858 struct scx_sched {
859 	struct sched_ext_ops	ops;
860 	DECLARE_BITMAP(has_op, SCX_OPI_END);
861 
862 	/*
863 	 * Dispatch queues.
864 	 *
865 	 * The global DSQ (%SCX_DSQ_GLOBAL) is split per-node for scalability.
866 	 * This is to avoid live-locking in bypass mode where all tasks are
867 	 * dispatched to %SCX_DSQ_GLOBAL and all CPUs consume from it. If
868 	 * per-node split isn't sufficient, it can be further split.
869 	 */
870 	struct rhashtable	dsq_hash;
871 	struct scx_dispatch_q	**global_dsqs;
872 	struct scx_sched_pcpu __percpu *pcpu;
873 
874 	bool			warned_zero_slice:1;
875 	bool			warned_deprecated_rq:1;
876 
877 	atomic_t		exit_kind;
878 	struct scx_exit_info	*exit_info;
879 
880 	struct kobject		kobj;
881 
882 	struct kthread_worker	*helper;
883 	struct irq_work		error_irq_work;
884 	struct kthread_work	disable_work;
885 	struct rcu_work		rcu_work;
886 };
887 
888 enum scx_wake_flags {
889 	/* expose select WF_* flags as enums */
890 	SCX_WAKE_FORK		= WF_FORK,
891 	SCX_WAKE_TTWU		= WF_TTWU,
892 	SCX_WAKE_SYNC		= WF_SYNC,
893 };
894 
895 enum scx_enq_flags {
896 	/* expose select ENQUEUE_* flags as enums */
897 	SCX_ENQ_WAKEUP		= ENQUEUE_WAKEUP,
898 	SCX_ENQ_HEAD		= ENQUEUE_HEAD,
899 	SCX_ENQ_CPU_SELECTED	= ENQUEUE_RQ_SELECTED,
900 
901 	/* high 32bits are SCX specific */
902 
903 	/*
904 	 * Set the following to trigger preemption when calling
905 	 * scx_bpf_dsq_insert() with a local dsq as the target. The slice of the
906 	 * current task is cleared to zero and the CPU is kicked into the
907 	 * scheduling path. Implies %SCX_ENQ_HEAD.
908 	 */
909 	SCX_ENQ_PREEMPT		= 1LLU << 32,
910 
911 	/*
912 	 * The task being enqueued was previously enqueued on the current CPU's
913 	 * %SCX_DSQ_LOCAL, but was removed from it in a call to the
914 	 * scx_bpf_reenqueue_local() kfunc. If scx_bpf_reenqueue_local() was
915 	 * invoked in a ->cpu_release() callback, and the task is again
916 	 * dispatched back to %SCX_LOCAL_DSQ by this current ->enqueue(), the
917 	 * task will not be scheduled on the CPU until at least the next invocation
918 	 * of the ->cpu_acquire() callback.
919 	 */
920 	SCX_ENQ_REENQ		= 1LLU << 40,
921 
922 	/*
923 	 * The task being enqueued is the only task available for the cpu. By
924 	 * default, ext core keeps executing such tasks but when
925 	 * %SCX_OPS_ENQ_LAST is specified, they're ops.enqueue()'d with the
926 	 * %SCX_ENQ_LAST flag set.
927 	 *
928 	 * The BPF scheduler is responsible for triggering a follow-up
929 	 * scheduling event. Otherwise, Execution may stall.
930 	 */
931 	SCX_ENQ_LAST		= 1LLU << 41,
932 
933 	/* high 8 bits are internal */
934 	__SCX_ENQ_INTERNAL_MASK	= 0xffLLU << 56,
935 
936 	SCX_ENQ_CLEAR_OPSS	= 1LLU << 56,
937 	SCX_ENQ_DSQ_PRIQ	= 1LLU << 57,
938 };
939 
940 enum scx_deq_flags {
941 	/* expose select DEQUEUE_* flags as enums */
942 	SCX_DEQ_SLEEP		= DEQUEUE_SLEEP,
943 
944 	/* high 32bits are SCX specific */
945 
946 	/*
947 	 * The generic core-sched layer decided to execute the task even though
948 	 * it hasn't been dispatched yet. Dequeue from the BPF side.
949 	 */
950 	SCX_DEQ_CORE_SCHED_EXEC	= 1LLU << 32,
951 };
952 
953 enum scx_pick_idle_cpu_flags {
954 	SCX_PICK_IDLE_CORE	= 1LLU << 0,	/* pick a CPU whose SMT siblings are also idle */
955 	SCX_PICK_IDLE_IN_NODE	= 1LLU << 1,	/* pick a CPU in the same target NUMA node */
956 };
957 
958 enum scx_kick_flags {
959 	/*
960 	 * Kick the target CPU if idle. Guarantees that the target CPU goes
961 	 * through at least one full scheduling cycle before going idle. If the
962 	 * target CPU can be determined to be currently not idle and going to go
963 	 * through a scheduling cycle before going idle, noop.
964 	 */
965 	SCX_KICK_IDLE		= 1LLU << 0,
966 
967 	/*
968 	 * Preempt the current task and execute the dispatch path. If the
969 	 * current task of the target CPU is an SCX task, its ->scx.slice is
970 	 * cleared to zero before the scheduling path is invoked so that the
971 	 * task expires and the dispatch path is invoked.
972 	 */
973 	SCX_KICK_PREEMPT	= 1LLU << 1,
974 
975 	/*
976 	 * Wait for the CPU to be rescheduled. The scx_bpf_kick_cpu() call will
977 	 * return after the target CPU finishes picking the next task.
978 	 */
979 	SCX_KICK_WAIT		= 1LLU << 2,
980 };
981 
982 enum scx_tg_flags {
983 	SCX_TG_ONLINE		= 1U << 0,
984 	SCX_TG_INITED		= 1U << 1,
985 };
986 
987 enum scx_enable_state {
988 	SCX_ENABLING,
989 	SCX_ENABLED,
990 	SCX_DISABLING,
991 	SCX_DISABLED,
992 };
993 
994 static const char *scx_enable_state_str[] = {
995 	[SCX_ENABLING]		= "enabling",
996 	[SCX_ENABLED]		= "enabled",
997 	[SCX_DISABLING]		= "disabling",
998 	[SCX_DISABLED]		= "disabled",
999 };
1000 
1001 /*
1002  * sched_ext_entity->ops_state
1003  *
1004  * Used to track the task ownership between the SCX core and the BPF scheduler.
1005  * State transitions look as follows:
1006  *
1007  * NONE -> QUEUEING -> QUEUED -> DISPATCHING
1008  *   ^              |                 |
1009  *   |              v                 v
1010  *   \-------------------------------/
1011  *
1012  * QUEUEING and DISPATCHING states can be waited upon. See wait_ops_state() call
1013  * sites for explanations on the conditions being waited upon and why they are
1014  * safe. Transitions out of them into NONE or QUEUED must store_release and the
1015  * waiters should load_acquire.
1016  *
1017  * Tracking scx_ops_state enables sched_ext core to reliably determine whether
1018  * any given task can be dispatched by the BPF scheduler at all times and thus
1019  * relaxes the requirements on the BPF scheduler. This allows the BPF scheduler
1020  * to try to dispatch any task anytime regardless of its state as the SCX core
1021  * can safely reject invalid dispatches.
1022  */
1023 enum scx_ops_state {
1024 	SCX_OPSS_NONE,		/* owned by the SCX core */
1025 	SCX_OPSS_QUEUEING,	/* in transit to the BPF scheduler */
1026 	SCX_OPSS_QUEUED,	/* owned by the BPF scheduler */
1027 	SCX_OPSS_DISPATCHING,	/* in transit back to the SCX core */
1028 
1029 	/*
1030 	 * QSEQ brands each QUEUED instance so that, when dispatch races
1031 	 * dequeue/requeue, the dispatcher can tell whether it still has a claim
1032 	 * on the task being dispatched.
1033 	 *
1034 	 * As some 32bit archs can't do 64bit store_release/load_acquire,
1035 	 * p->scx.ops_state is atomic_long_t which leaves 30 bits for QSEQ on
1036 	 * 32bit machines. The dispatch race window QSEQ protects is very narrow
1037 	 * and runs with IRQ disabled. 30 bits should be sufficient.
1038 	 */
1039 	SCX_OPSS_QSEQ_SHIFT	= 2,
1040 };
1041 
1042 /* Use macros to ensure that the type is unsigned long for the masks */
1043 #define SCX_OPSS_STATE_MASK	((1LU << SCX_OPSS_QSEQ_SHIFT) - 1)
1044 #define SCX_OPSS_QSEQ_MASK	(~SCX_OPSS_STATE_MASK)
1045 
1046 DECLARE_PER_CPU(struct rq *, scx_locked_rq_state);
1047 
1048 /*
1049  * Return the rq currently locked from an scx callback, or NULL if no rq is
1050  * locked.
1051  */
1052 static inline struct rq *scx_locked_rq(void)
1053 {
1054 	return __this_cpu_read(scx_locked_rq_state);
1055 }
1056 
1057 static inline bool scx_kf_allowed_if_unlocked(void)
1058 {
1059 	return !current->scx.kf_mask;
1060 }
1061 
1062 static inline bool scx_rq_bypassing(struct rq *rq)
1063 {
1064 	return unlikely(rq->scx.flags & SCX_RQ_BYPASSING);
1065 }
1066