xref: /linux-6.15/include/linux/preempt.h (revision 91ebe8bc)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef __LINUX_PREEMPT_H
3 #define __LINUX_PREEMPT_H
4 
5 /*
6  * include/linux/preempt.h - macros for accessing and manipulating
7  * preempt_count (used for kernel preemption, interrupt count, etc.)
8  */
9 
10 #include <linux/linkage.h>
11 #include <linux/list.h>
12 
13 /*
14  * We put the hardirq and softirq counter into the preemption
15  * counter. The bitmask has the following meaning:
16  *
17  * - bits 0-7 are the preemption count (max preemption depth: 256)
18  * - bits 8-15 are the softirq count (max # of softirqs: 256)
19  *
20  * The hardirq count could in theory be the same as the number of
21  * interrupts in the system, but we run all interrupt handlers with
22  * interrupts disabled, so we cannot have nesting interrupts. Though
23  * there are a few palaeontologic drivers which reenable interrupts in
24  * the handler, so we need more than one bit here.
25  *
26  *         PREEMPT_MASK:	0x000000ff
27  *         SOFTIRQ_MASK:	0x0000ff00
28  *         HARDIRQ_MASK:	0x000f0000
29  *             NMI_MASK:	0x00f00000
30  * PREEMPT_NEED_RESCHED:	0x80000000
31  */
32 #define PREEMPT_BITS	8
33 #define SOFTIRQ_BITS	8
34 #define HARDIRQ_BITS	4
35 #define NMI_BITS	4
36 
37 #define PREEMPT_SHIFT	0
38 #define SOFTIRQ_SHIFT	(PREEMPT_SHIFT + PREEMPT_BITS)
39 #define HARDIRQ_SHIFT	(SOFTIRQ_SHIFT + SOFTIRQ_BITS)
40 #define NMI_SHIFT	(HARDIRQ_SHIFT + HARDIRQ_BITS)
41 
42 #define __IRQ_MASK(x)	((1UL << (x))-1)
43 
44 #define PREEMPT_MASK	(__IRQ_MASK(PREEMPT_BITS) << PREEMPT_SHIFT)
45 #define SOFTIRQ_MASK	(__IRQ_MASK(SOFTIRQ_BITS) << SOFTIRQ_SHIFT)
46 #define HARDIRQ_MASK	(__IRQ_MASK(HARDIRQ_BITS) << HARDIRQ_SHIFT)
47 #define NMI_MASK	(__IRQ_MASK(NMI_BITS)     << NMI_SHIFT)
48 
49 #define PREEMPT_OFFSET	(1UL << PREEMPT_SHIFT)
50 #define SOFTIRQ_OFFSET	(1UL << SOFTIRQ_SHIFT)
51 #define HARDIRQ_OFFSET	(1UL << HARDIRQ_SHIFT)
52 #define NMI_OFFSET	(1UL << NMI_SHIFT)
53 
54 #define SOFTIRQ_DISABLE_OFFSET	(2 * SOFTIRQ_OFFSET)
55 
56 #define PREEMPT_DISABLED	(PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
57 
58 /*
59  * Disable preemption until the scheduler is running -- use an unconditional
60  * value so that it also works on !PREEMPT_COUNT kernels.
61  *
62  * Reset by start_kernel()->sched_init()->init_idle()->init_idle_preempt_count().
63  */
64 #define INIT_PREEMPT_COUNT	PREEMPT_OFFSET
65 
66 /*
67  * Initial preempt_count value; reflects the preempt_count schedule invariant
68  * which states that during context switches:
69  *
70  *    preempt_count() == 2*PREEMPT_DISABLE_OFFSET
71  *
72  * Note: PREEMPT_DISABLE_OFFSET is 0 for !PREEMPT_COUNT kernels.
73  * Note: See finish_task_switch().
74  */
75 #define FORK_PREEMPT_COUNT	(2*PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
76 
77 /* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
78 #include <asm/preempt.h>
79 
80 /**
81  * interrupt_context_level - return interrupt context level
82  *
83  * Returns the current interrupt context level.
84  *  0 - normal context
85  *  1 - softirq context
86  *  2 - hardirq context
87  *  3 - NMI context
88  */
89 static __always_inline unsigned char interrupt_context_level(void)
90 {
91 	unsigned long pc = preempt_count();
92 	unsigned char level = 0;
93 
94 	level += !!(pc & (NMI_MASK));
95 	level += !!(pc & (NMI_MASK | HARDIRQ_MASK));
96 	level += !!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET));
97 
98 	return level;
99 }
100 
101 #define nmi_count()	(preempt_count() & NMI_MASK)
102 #define hardirq_count()	(preempt_count() & HARDIRQ_MASK)
103 #ifdef CONFIG_PREEMPT_RT
104 # define softirq_count()	(current->softirq_disable_cnt & SOFTIRQ_MASK)
105 #else
106 # define softirq_count()	(preempt_count() & SOFTIRQ_MASK)
107 #endif
108 #define irq_count()	(nmi_count() | hardirq_count() | softirq_count())
109 
110 /*
111  * Macros to retrieve the current execution context:
112  *
113  * in_nmi()		- We're in NMI context
114  * in_hardirq()		- We're in hard IRQ context
115  * in_serving_softirq()	- We're in softirq context
116  * in_task()		- We're in task context
117  */
118 #define in_nmi()		(nmi_count())
119 #define in_hardirq()		(hardirq_count())
120 #define in_serving_softirq()	(softirq_count() & SOFTIRQ_OFFSET)
121 #define in_task()		(!(in_nmi() | in_hardirq() | in_serving_softirq()))
122 
123 /*
124  * The following macros are deprecated and should not be used in new code:
125  * in_irq()       - Obsolete version of in_hardirq()
126  * in_softirq()   - We have BH disabled, or are processing softirqs
127  * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
128  */
129 #define in_irq()		(hardirq_count())
130 #define in_softirq()		(softirq_count())
131 #define in_interrupt()		(irq_count())
132 
133 /*
134  * The preempt_count offset after preempt_disable();
135  */
136 #if defined(CONFIG_PREEMPT_COUNT)
137 # define PREEMPT_DISABLE_OFFSET	PREEMPT_OFFSET
138 #else
139 # define PREEMPT_DISABLE_OFFSET	0
140 #endif
141 
142 /*
143  * The preempt_count offset after spin_lock()
144  */
145 #if !defined(CONFIG_PREEMPT_RT)
146 #define PREEMPT_LOCK_OFFSET	PREEMPT_DISABLE_OFFSET
147 #else
148 #define PREEMPT_LOCK_OFFSET	0
149 #endif
150 
151 /*
152  * The preempt_count offset needed for things like:
153  *
154  *  spin_lock_bh()
155  *
156  * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
157  * softirqs, such that unlock sequences of:
158  *
159  *  spin_unlock();
160  *  local_bh_enable();
161  *
162  * Work as expected.
163  */
164 #define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
165 
166 /*
167  * Are we running in atomic context?  WARNING: this macro cannot
168  * always detect atomic context; in particular, it cannot know about
169  * held spinlocks in non-preemptible kernels.  Thus it should not be
170  * used in the general case to determine whether sleeping is possible.
171  * Do not use in_atomic() in driver code.
172  */
173 #define in_atomic()	(preempt_count() != 0)
174 
175 /*
176  * Check whether we were atomic before we did preempt_disable():
177  * (used by the scheduler)
178  */
179 #define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
180 
181 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
182 extern void preempt_count_add(int val);
183 extern void preempt_count_sub(int val);
184 #define preempt_count_dec_and_test() \
185 	({ preempt_count_sub(1); should_resched(0); })
186 #else
187 #define preempt_count_add(val)	__preempt_count_add(val)
188 #define preempt_count_sub(val)	__preempt_count_sub(val)
189 #define preempt_count_dec_and_test() __preempt_count_dec_and_test()
190 #endif
191 
192 #define __preempt_count_inc() __preempt_count_add(1)
193 #define __preempt_count_dec() __preempt_count_sub(1)
194 
195 #define preempt_count_inc() preempt_count_add(1)
196 #define preempt_count_dec() preempt_count_sub(1)
197 
198 #ifdef CONFIG_PREEMPT_COUNT
199 
200 #define preempt_disable() \
201 do { \
202 	preempt_count_inc(); \
203 	barrier(); \
204 } while (0)
205 
206 #define sched_preempt_enable_no_resched() \
207 do { \
208 	barrier(); \
209 	preempt_count_dec(); \
210 } while (0)
211 
212 #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
213 
214 #define preemptible()	(preempt_count() == 0 && !irqs_disabled())
215 
216 #ifdef CONFIG_PREEMPTION
217 #define preempt_enable() \
218 do { \
219 	barrier(); \
220 	if (unlikely(preempt_count_dec_and_test())) \
221 		__preempt_schedule(); \
222 } while (0)
223 
224 #define preempt_enable_notrace() \
225 do { \
226 	barrier(); \
227 	if (unlikely(__preempt_count_dec_and_test())) \
228 		__preempt_schedule_notrace(); \
229 } while (0)
230 
231 #define preempt_check_resched() \
232 do { \
233 	if (should_resched(0)) \
234 		__preempt_schedule(); \
235 } while (0)
236 
237 #else /* !CONFIG_PREEMPTION */
238 #define preempt_enable() \
239 do { \
240 	barrier(); \
241 	preempt_count_dec(); \
242 } while (0)
243 
244 #define preempt_enable_notrace() \
245 do { \
246 	barrier(); \
247 	__preempt_count_dec(); \
248 } while (0)
249 
250 #define preempt_check_resched() do { } while (0)
251 #endif /* CONFIG_PREEMPTION */
252 
253 #define preempt_disable_notrace() \
254 do { \
255 	__preempt_count_inc(); \
256 	barrier(); \
257 } while (0)
258 
259 #define preempt_enable_no_resched_notrace() \
260 do { \
261 	barrier(); \
262 	__preempt_count_dec(); \
263 } while (0)
264 
265 #else /* !CONFIG_PREEMPT_COUNT */
266 
267 /*
268  * Even if we don't have any preemption, we need preempt disable/enable
269  * to be barriers, so that we don't have things like get_user/put_user
270  * that can cause faults and scheduling migrate into our preempt-protected
271  * region.
272  */
273 #define preempt_disable()			barrier()
274 #define sched_preempt_enable_no_resched()	barrier()
275 #define preempt_enable_no_resched()		barrier()
276 #define preempt_enable()			barrier()
277 #define preempt_check_resched()			do { } while (0)
278 
279 #define preempt_disable_notrace()		barrier()
280 #define preempt_enable_no_resched_notrace()	barrier()
281 #define preempt_enable_notrace()		barrier()
282 #define preemptible()				0
283 
284 #endif /* CONFIG_PREEMPT_COUNT */
285 
286 #ifdef MODULE
287 /*
288  * Modules have no business playing preemption tricks.
289  */
290 #undef sched_preempt_enable_no_resched
291 #undef preempt_enable_no_resched
292 #undef preempt_enable_no_resched_notrace
293 #undef preempt_check_resched
294 #endif
295 
296 #define preempt_set_need_resched() \
297 do { \
298 	set_preempt_need_resched(); \
299 } while (0)
300 #define preempt_fold_need_resched() \
301 do { \
302 	if (tif_need_resched()) \
303 		set_preempt_need_resched(); \
304 } while (0)
305 
306 #ifdef CONFIG_PREEMPT_NOTIFIERS
307 
308 struct preempt_notifier;
309 
310 /**
311  * preempt_ops - notifiers called when a task is preempted and rescheduled
312  * @sched_in: we're about to be rescheduled:
313  *    notifier: struct preempt_notifier for the task being scheduled
314  *    cpu:  cpu we're scheduled on
315  * @sched_out: we've just been preempted
316  *    notifier: struct preempt_notifier for the task being preempted
317  *    next: the task that's kicking us out
318  *
319  * Please note that sched_in and out are called under different
320  * contexts.  sched_out is called with rq lock held and irq disabled
321  * while sched_in is called without rq lock and irq enabled.  This
322  * difference is intentional and depended upon by its users.
323  */
324 struct preempt_ops {
325 	void (*sched_in)(struct preempt_notifier *notifier, int cpu);
326 	void (*sched_out)(struct preempt_notifier *notifier,
327 			  struct task_struct *next);
328 };
329 
330 /**
331  * preempt_notifier - key for installing preemption notifiers
332  * @link: internal use
333  * @ops: defines the notifier functions to be called
334  *
335  * Usually used in conjunction with container_of().
336  */
337 struct preempt_notifier {
338 	struct hlist_node link;
339 	struct preempt_ops *ops;
340 };
341 
342 void preempt_notifier_inc(void);
343 void preempt_notifier_dec(void);
344 void preempt_notifier_register(struct preempt_notifier *notifier);
345 void preempt_notifier_unregister(struct preempt_notifier *notifier);
346 
347 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
348 				     struct preempt_ops *ops)
349 {
350 	INIT_HLIST_NODE(&notifier->link);
351 	notifier->ops = ops;
352 }
353 
354 #endif
355 
356 #ifdef CONFIG_SMP
357 
358 /*
359  * Migrate-Disable and why it is undesired.
360  *
361  * When a preempted task becomes elegible to run under the ideal model (IOW it
362  * becomes one of the M highest priority tasks), it might still have to wait
363  * for the preemptee's migrate_disable() section to complete. Thereby suffering
364  * a reduction in bandwidth in the exact duration of the migrate_disable()
365  * section.
366  *
367  * Per this argument, the change from preempt_disable() to migrate_disable()
368  * gets us:
369  *
370  * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
371  *   it would have had to wait for the lower priority task.
372  *
373  * - a lower priority tasks; which under preempt_disable() could've instantly
374  *   migrated away when another CPU becomes available, is now constrained
375  *   by the ability to push the higher priority task away, which might itself be
376  *   in a migrate_disable() section, reducing it's available bandwidth.
377  *
378  * IOW it trades latency / moves the interference term, but it stays in the
379  * system, and as long as it remains unbounded, the system is not fully
380  * deterministic.
381  *
382  *
383  * The reason we have it anyway.
384  *
385  * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
386  * number of primitives into becoming preemptible, they would also allow
387  * migration. This turns out to break a bunch of per-cpu usage. To this end,
388  * all these primitives employ migirate_disable() to restore this implicit
389  * assumption.
390  *
391  * This is a 'temporary' work-around at best. The correct solution is getting
392  * rid of the above assumptions and reworking the code to employ explicit
393  * per-cpu locking or short preempt-disable regions.
394  *
395  * The end goal must be to get rid of migrate_disable(), alternatively we need
396  * a schedulability theory that does not depend on abritrary migration.
397  *
398  *
399  * Notes on the implementation.
400  *
401  * The implementation is particularly tricky since existing code patterns
402  * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
403  * This means that it cannot use cpus_read_lock() to serialize against hotplug,
404  * nor can it easily migrate itself into a pending affinity mask change on
405  * migrate_enable().
406  *
407  *
408  * Note: even non-work-conserving schedulers like semi-partitioned depends on
409  *       migration, so migrate_disable() is not only a problem for
410  *       work-conserving schedulers.
411  *
412  */
413 extern void migrate_disable(void);
414 extern void migrate_enable(void);
415 
416 #else
417 
418 static inline void migrate_disable(void) { }
419 static inline void migrate_enable(void) { }
420 
421 #endif /* CONFIG_SMP */
422 
423 #endif /* __LINUX_PREEMPT_H */
424