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/cleanup.h>
12#include <linux/types.h>
13
14/*
15 * We put the hardirq and softirq counter into the preemption
16 * counter. The bitmask has the following meaning:
17 *
18 * - bits 0-7 are the preemption count (max preemption depth: 256)
19 * - bits 8-15 are the softirq count (max # of softirqs: 256)
20 *
21 * The hardirq count could in theory be the same as the number of
22 * interrupts in the system, but we run all interrupt handlers with
23 * interrupts disabled, so we cannot have nesting interrupts. Though
24 * there are a few palaeontologic drivers which reenable interrupts in
25 * the handler, so we need more than one bit here.
26 *
27 * PREEMPT_MASK: 0x000000ff
28 * SOFTIRQ_MASK: 0x0000ff00
29 * HARDIRQ_MASK: 0x000f0000
30 * NMI_MASK: 0x00f00000
31 * PREEMPT_NEED_RESCHED: 0x80000000
32 */
33#define PREEMPT_BITS 8
34#define SOFTIRQ_BITS 8
35#define HARDIRQ_BITS 4
36#define NMI_BITS 4
37
38#define PREEMPT_SHIFT 0
39#define SOFTIRQ_SHIFT (PREEMPT_SHIFT + PREEMPT_BITS)
40#define HARDIRQ_SHIFT (SOFTIRQ_SHIFT + SOFTIRQ_BITS)
41#define NMI_SHIFT (HARDIRQ_SHIFT + HARDIRQ_BITS)
42
43#define __IRQ_MASK(x) ((1UL << (x))-1)
44
45#define PREEMPT_MASK (__IRQ_MASK(PREEMPT_BITS) << PREEMPT_SHIFT)
46#define SOFTIRQ_MASK (__IRQ_MASK(SOFTIRQ_BITS) << SOFTIRQ_SHIFT)
47#define HARDIRQ_MASK (__IRQ_MASK(HARDIRQ_BITS) << HARDIRQ_SHIFT)
48#define NMI_MASK (__IRQ_MASK(NMI_BITS) << NMI_SHIFT)
49
50#define PREEMPT_OFFSET (1UL << PREEMPT_SHIFT)
51#define SOFTIRQ_OFFSET (1UL << SOFTIRQ_SHIFT)
52#define HARDIRQ_OFFSET (1UL << HARDIRQ_SHIFT)
53#define NMI_OFFSET (1UL << NMI_SHIFT)
54
55#define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
56
57#define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
58
59/*
60 * Disable preemption until the scheduler is running -- use an unconditional
61 * value so that it also works on !PREEMPT_COUNT kernels.
62 *
63 * Reset by start_kernel()->sched_init()->init_idle()->init_idle_preempt_count().
64 */
65#define INIT_PREEMPT_COUNT PREEMPT_OFFSET
66
67/*
68 * Initial preempt_count value; reflects the preempt_count schedule invariant
69 * which states that during context switches:
70 *
71 * preempt_count() == 2*PREEMPT_DISABLE_OFFSET
72 *
73 * Note: PREEMPT_DISABLE_OFFSET is 0 for !PREEMPT_COUNT kernels.
74 * Note: See finish_task_switch().
75 */
76#define FORK_PREEMPT_COUNT (2*PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
77
78/* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
79#include <asm/preempt.h>
80
81/**
82 * interrupt_context_level - return interrupt context level
83 *
84 * Returns the current interrupt context level.
85 * 0 - normal context
86 * 1 - softirq context
87 * 2 - hardirq context
88 * 3 - NMI context
89 */
90static __always_inline unsigned char interrupt_context_level(void)
91{
92 unsigned long pc = preempt_count();
93 unsigned char level = 0;
94
95 level += !!(pc & (NMI_MASK));
96 level += !!(pc & (NMI_MASK | HARDIRQ_MASK));
97 level += !!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET));
98
99 return level;
100}
101
102/*
103 * These macro definitions avoid redundant invocations of preempt_count()
104 * because such invocations would result in redundant loads given that
105 * preempt_count() is commonly implemented with READ_ONCE().
106 */
107
108#define nmi_count() (preempt_count() & NMI_MASK)
109#define hardirq_count() (preempt_count() & HARDIRQ_MASK)
110#ifdef CONFIG_PREEMPT_RT
111# define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
112# define irq_count() ((preempt_count() & (NMI_MASK | HARDIRQ_MASK)) | softirq_count())
113#else
114# define softirq_count() (preempt_count() & SOFTIRQ_MASK)
115# define irq_count() (preempt_count() & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_MASK))
116#endif
117
118/*
119 * Macros to retrieve the current execution context:
120 *
121 * in_nmi() - We're in NMI context
122 * in_hardirq() - We're in hard IRQ context
123 * in_serving_softirq() - We're in softirq context
124 * in_task() - We're in task context
125 */
126#define in_nmi() (nmi_count())
127#define in_hardirq() (hardirq_count())
128#define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
129#ifdef CONFIG_PREEMPT_RT
130# define in_task() (!((preempt_count() & (NMI_MASK | HARDIRQ_MASK)) | in_serving_softirq()))
131#else
132# define in_task() (!(preempt_count() & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
133#endif
134
135/*
136 * The following macros are deprecated and should not be used in new code:
137 * in_irq() - Obsolete version of in_hardirq()
138 * in_softirq() - We have BH disabled, or are processing softirqs
139 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
140 */
141#define in_irq() (hardirq_count())
142#define in_softirq() (softirq_count())
143#define in_interrupt() (irq_count())
144
145/*
146 * The preempt_count offset after preempt_disable();
147 */
148#if defined(CONFIG_PREEMPT_COUNT)
149# define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET
150#else
151# define PREEMPT_DISABLE_OFFSET 0
152#endif
153
154/*
155 * The preempt_count offset after spin_lock()
156 */
157#if !defined(CONFIG_PREEMPT_RT)
158#define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
159#else
160/* Locks on RT do not disable preemption */
161#define PREEMPT_LOCK_OFFSET 0
162#endif
163
164/*
165 * The preempt_count offset needed for things like:
166 *
167 * spin_lock_bh()
168 *
169 * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
170 * softirqs, such that unlock sequences of:
171 *
172 * spin_unlock();
173 * local_bh_enable();
174 *
175 * Work as expected.
176 */
177#define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
178
179/*
180 * Are we running in atomic context? WARNING: this macro cannot
181 * always detect atomic context; in particular, it cannot know about
182 * held spinlocks in non-preemptible kernels. Thus it should not be
183 * used in the general case to determine whether sleeping is possible.
184 * Do not use in_atomic() in driver code.
185 */
186#define in_atomic() (preempt_count() != 0)
187
188/*
189 * Check whether we were atomic before we did preempt_disable():
190 * (used by the scheduler)
191 */
192#define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
193
194#if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
195extern void preempt_count_add(int val);
196extern void preempt_count_sub(int val);
197#define preempt_count_dec_and_test() \
198 ({ preempt_count_sub(1); should_resched(0); })
199#else
200#define preempt_count_add(val) __preempt_count_add(val)
201#define preempt_count_sub(val) __preempt_count_sub(val)
202#define preempt_count_dec_and_test() __preempt_count_dec_and_test()
203#endif
204
205#define __preempt_count_inc() __preempt_count_add(1)
206#define __preempt_count_dec() __preempt_count_sub(1)
207
208#define preempt_count_inc() preempt_count_add(1)
209#define preempt_count_dec() preempt_count_sub(1)
210
211#ifdef CONFIG_PREEMPT_COUNT
212
213#define preempt_disable() \
214do { \
215 preempt_count_inc(); \
216 barrier(); \
217} while (0)
218
219#define sched_preempt_enable_no_resched() \
220do { \
221 barrier(); \
222 preempt_count_dec(); \
223} while (0)
224
225#define preempt_enable_no_resched() sched_preempt_enable_no_resched()
226
227#define preemptible() (preempt_count() == 0 && !irqs_disabled())
228
229#ifdef CONFIG_PREEMPTION
230#define preempt_enable() \
231do { \
232 barrier(); \
233 if (unlikely(preempt_count_dec_and_test())) \
234 __preempt_schedule(); \
235} while (0)
236
237#define preempt_enable_notrace() \
238do { \
239 barrier(); \
240 if (unlikely(__preempt_count_dec_and_test())) \
241 __preempt_schedule_notrace(); \
242} while (0)
243
244#define preempt_check_resched() \
245do { \
246 if (should_resched(0)) \
247 __preempt_schedule(); \
248} while (0)
249
250#else /* !CONFIG_PREEMPTION */
251#define preempt_enable() \
252do { \
253 barrier(); \
254 preempt_count_dec(); \
255} while (0)
256
257#define preempt_enable_notrace() \
258do { \
259 barrier(); \
260 __preempt_count_dec(); \
261} while (0)
262
263#define preempt_check_resched() do { } while (0)
264#endif /* CONFIG_PREEMPTION */
265
266#define preempt_disable_notrace() \
267do { \
268 __preempt_count_inc(); \
269 barrier(); \
270} while (0)
271
272#define preempt_enable_no_resched_notrace() \
273do { \
274 barrier(); \
275 __preempt_count_dec(); \
276} while (0)
277
278#else /* !CONFIG_PREEMPT_COUNT */
279
280/*
281 * Even if we don't have any preemption, we need preempt disable/enable
282 * to be barriers, so that we don't have things like get_user/put_user
283 * that can cause faults and scheduling migrate into our preempt-protected
284 * region.
285 */
286#define preempt_disable() barrier()
287#define sched_preempt_enable_no_resched() barrier()
288#define preempt_enable_no_resched() barrier()
289#define preempt_enable() barrier()
290#define preempt_check_resched() do { } while (0)
291
292#define preempt_disable_notrace() barrier()
293#define preempt_enable_no_resched_notrace() barrier()
294#define preempt_enable_notrace() barrier()
295#define preemptible() 0
296
297#endif /* CONFIG_PREEMPT_COUNT */
298
299#ifdef MODULE
300/*
301 * Modules have no business playing preemption tricks.
302 */
303#undef sched_preempt_enable_no_resched
304#undef preempt_enable_no_resched
305#undef preempt_enable_no_resched_notrace
306#undef preempt_check_resched
307#endif
308
309#define preempt_set_need_resched() \
310do { \
311 set_preempt_need_resched(); \
312} while (0)
313#define preempt_fold_need_resched() \
314do { \
315 if (tif_need_resched()) \
316 set_preempt_need_resched(); \
317} while (0)
318
319#ifdef CONFIG_PREEMPT_NOTIFIERS
320
321struct preempt_notifier;
322struct task_struct;
323
324/**
325 * preempt_ops - notifiers called when a task is preempted and rescheduled
326 * @sched_in: we're about to be rescheduled:
327 * notifier: struct preempt_notifier for the task being scheduled
328 * cpu: cpu we're scheduled on
329 * @sched_out: we've just been preempted
330 * notifier: struct preempt_notifier for the task being preempted
331 * next: the task that's kicking us out
332 *
333 * Please note that sched_in and out are called under different
334 * contexts. sched_out is called with rq lock held and irq disabled
335 * while sched_in is called without rq lock and irq enabled. This
336 * difference is intentional and depended upon by its users.
337 */
338struct preempt_ops {
339 void (*sched_in)(struct preempt_notifier *notifier, int cpu);
340 void (*sched_out)(struct preempt_notifier *notifier,
341 struct task_struct *next);
342};
343
344/**
345 * preempt_notifier - key for installing preemption notifiers
346 * @link: internal use
347 * @ops: defines the notifier functions to be called
348 *
349 * Usually used in conjunction with container_of().
350 */
351struct preempt_notifier {
352 struct hlist_node link;
353 struct preempt_ops *ops;
354};
355
356void preempt_notifier_inc(void);
357void preempt_notifier_dec(void);
358void preempt_notifier_register(struct preempt_notifier *notifier);
359void preempt_notifier_unregister(struct preempt_notifier *notifier);
360
361static inline void preempt_notifier_init(struct preempt_notifier *notifier,
362 struct preempt_ops *ops)
363{
364 /* INIT_HLIST_NODE() open coded, to avoid dependency on list.h */
365 notifier->link.next = NULL;
366 notifier->link.pprev = NULL;
367 notifier->ops = ops;
368}
369
370#endif
371
372/*
373 * Migrate-Disable and why it is undesired.
374 *
375 * When a preempted task becomes eligible to run under the ideal model (IOW it
376 * becomes one of the M highest priority tasks), it might still have to wait
377 * for the preemptee's migrate_disable() section to complete. Thereby suffering
378 * a reduction in bandwidth in the exact duration of the migrate_disable()
379 * section.
380 *
381 * Per this argument, the change from preempt_disable() to migrate_disable()
382 * gets us:
383 *
384 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
385 * it would have had to wait for the lower priority task.
386 *
387 * - a lower priority tasks; which under preempt_disable() could've instantly
388 * migrated away when another CPU becomes available, is now constrained
389 * by the ability to push the higher priority task away, which might itself be
390 * in a migrate_disable() section, reducing its available bandwidth.
391 *
392 * IOW it trades latency / moves the interference term, but it stays in the
393 * system, and as long as it remains unbounded, the system is not fully
394 * deterministic.
395 *
396 *
397 * The reason we have it anyway.
398 *
399 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
400 * number of primitives into becoming preemptible, they would also allow
401 * migration. This turns out to break a bunch of per-cpu usage. To this end,
402 * all these primitives employ migrate_disable() to restore this implicit
403 * assumption.
404 *
405 * This is a 'temporary' work-around at best. The correct solution is getting
406 * rid of the above assumptions and reworking the code to employ explicit
407 * per-cpu locking or short preempt-disable regions.
408 *
409 * The end goal must be to get rid of migrate_disable(), alternatively we need
410 * a schedulability theory that does not depend on arbitrary migration.
411 *
412 *
413 * Notes on the implementation.
414 *
415 * The implementation is particularly tricky since existing code patterns
416 * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
417 * This means that it cannot use cpus_read_lock() to serialize against hotplug,
418 * nor can it easily migrate itself into a pending affinity mask change on
419 * migrate_enable().
420 *
421 *
422 * Note: even non-work-conserving schedulers like semi-partitioned depends on
423 * migration, so migrate_disable() is not only a problem for
424 * work-conserving schedulers.
425 *
426 */
427
428/**
429 * preempt_disable_nested - Disable preemption inside a normally preempt disabled section
430 *
431 * Use for code which requires preemption protection inside a critical
432 * section which has preemption disabled implicitly on non-PREEMPT_RT
433 * enabled kernels, by e.g.:
434 * - holding a spinlock/rwlock
435 * - soft interrupt context
436 * - regular interrupt handlers
437 *
438 * On PREEMPT_RT enabled kernels spinlock/rwlock held sections, soft
439 * interrupt context and regular interrupt handlers are preemptible and
440 * only prevent migration. preempt_disable_nested() ensures that preemption
441 * is disabled for cases which require CPU local serialization even on
442 * PREEMPT_RT. For non-PREEMPT_RT kernels this is a NOP.
443 *
444 * The use cases are code sequences which are not serialized by a
445 * particular lock instance, e.g.:
446 * - seqcount write side critical sections where the seqcount is not
447 * associated to a particular lock and therefore the automatic
448 * protection mechanism does not work. This prevents a live lock
449 * against a preempting high priority reader.
450 * - RMW per CPU variable updates like vmstat.
451 */
452/* Macro to avoid header recursion hell vs. lockdep */
453#define preempt_disable_nested() \
454do { \
455 if (IS_ENABLED(CONFIG_PREEMPT_RT)) \
456 preempt_disable(); \
457 else \
458 lockdep_assert_preemption_disabled(); \
459} while (0)
460
461/**
462 * preempt_enable_nested - Undo the effect of preempt_disable_nested()
463 */
464static __always_inline void preempt_enable_nested(void)
465{
466 if (IS_ENABLED(CONFIG_PREEMPT_RT))
467 preempt_enable();
468}
469
470DEFINE_LOCK_GUARD_0(preempt, preempt_disable(), preempt_enable())
471DEFINE_LOCK_GUARD_0(preempt_notrace, preempt_disable_notrace(), preempt_enable_notrace())
472
473#ifdef CONFIG_PREEMPT_DYNAMIC
474
475extern bool preempt_model_none(void);
476extern bool preempt_model_voluntary(void);
477extern bool preempt_model_full(void);
478extern bool preempt_model_lazy(void);
479
480#else
481
482static inline bool preempt_model_none(void)
483{
484 return IS_ENABLED(CONFIG_PREEMPT_NONE);
485}
486static inline bool preempt_model_voluntary(void)
487{
488 return IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY);
489}
490static inline bool preempt_model_full(void)
491{
492 return IS_ENABLED(CONFIG_PREEMPT);
493}
494
495static inline bool preempt_model_lazy(void)
496{
497 return IS_ENABLED(CONFIG_PREEMPT_LAZY);
498}
499
500#endif
501
502static inline bool preempt_model_rt(void)
503{
504 return IS_ENABLED(CONFIG_PREEMPT_RT);
505}
506
507extern const char *preempt_model_str(void);
508
509/*
510 * Does the preemption model allow non-cooperative preemption?
511 *
512 * For !CONFIG_PREEMPT_DYNAMIC kernels this is an exact match with
513 * CONFIG_PREEMPTION; for CONFIG_PREEMPT_DYNAMIC this doesn't work as the
514 * kernel is *built* with CONFIG_PREEMPTION=y but may run with e.g. the
515 * PREEMPT_NONE model.
516 */
517static inline bool preempt_model_preemptible(void)
518{
519 return preempt_model_full() || preempt_model_lazy() || preempt_model_rt();
520}
521
522#endif /* __LINUX_PREEMPT_H */
523