1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef __LINUX_PREEMPT_H
3 #define __LINUX_PREEMPT_H
6 * include/linux/preempt.h - macros for accessing and manipulating
7 * preempt_count (used for kernel preemption, interrupt count, etc.)
10 #include <linux/linkage.h>
11 #include <linux/cleanup.h>
12 #include <linux/list.h>
15 * We put the hardirq and softirq counter into the preemption
16 * counter. The bitmask has the following meaning:
18 * - bits 0-7 are the preemption count (max preemption depth: 256)
19 * - bits 8-15 are the softirq count (max # of softirqs: 256)
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.
27 * PREEMPT_MASK: 0x000000ff
28 * SOFTIRQ_MASK: 0x0000ff00
29 * HARDIRQ_MASK: 0x000f0000
30 * NMI_MASK: 0x00f00000
31 * PREEMPT_NEED_RESCHED: 0x80000000
33 #define PREEMPT_BITS 8
34 #define SOFTIRQ_BITS 8
35 #define HARDIRQ_BITS 4
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)
43 #define __IRQ_MASK(x) ((1UL << (x))-1)
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)
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)
55 #define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
57 #define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
60 * Disable preemption until the scheduler is running -- use an unconditional
61 * value so that it also works on !PREEMPT_COUNT kernels.
63 * Reset by start_kernel()->sched_init()->init_idle()->init_idle_preempt_count().
65 #define INIT_PREEMPT_COUNT PREEMPT_OFFSET
68 * Initial preempt_count value; reflects the preempt_count schedule invariant
69 * which states that during context switches:
71 * preempt_count() == 2*PREEMPT_DISABLE_OFFSET
73 * Note: PREEMPT_DISABLE_OFFSET is 0 for !PREEMPT_COUNT kernels.
74 * Note: See finish_task_switch().
76 #define FORK_PREEMPT_COUNT (2*PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
78 /* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
79 #include <asm/preempt.h>
82 * interrupt_context_level - return interrupt context level
84 * Returns the current interrupt context level.
90 static __always_inline unsigned char interrupt_context_level(void)
92 unsigned long pc = preempt_count();
93 unsigned char level = 0;
95 level += !!(pc & (NMI_MASK));
96 level += !!(pc & (NMI_MASK | HARDIRQ_MASK));
97 level += !!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET));
102 #define nmi_count() (preempt_count() & NMI_MASK)
103 #define hardirq_count() (preempt_count() & HARDIRQ_MASK)
104 #ifdef CONFIG_PREEMPT_RT
105 # define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
107 # define softirq_count() (preempt_count() & SOFTIRQ_MASK)
109 #define irq_count() (nmi_count() | hardirq_count() | softirq_count())
112 * Macros to retrieve the current execution context:
114 * in_nmi() - We're in NMI context
115 * in_hardirq() - We're in hard IRQ context
116 * in_serving_softirq() - We're in softirq context
117 * in_task() - We're in task context
119 #define in_nmi() (nmi_count())
120 #define in_hardirq() (hardirq_count())
121 #define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
122 #define in_task() (!(in_nmi() | in_hardirq() | in_serving_softirq()))
125 * The following macros are deprecated and should not be used in new code:
126 * in_irq() - Obsolete version of in_hardirq()
127 * in_softirq() - We have BH disabled, or are processing softirqs
128 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
130 #define in_irq() (hardirq_count())
131 #define in_softirq() (softirq_count())
132 #define in_interrupt() (irq_count())
135 * The preempt_count offset after preempt_disable();
137 #if defined(CONFIG_PREEMPT_COUNT)
138 # define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET
140 # define PREEMPT_DISABLE_OFFSET 0
144 * The preempt_count offset after spin_lock()
146 #if !defined(CONFIG_PREEMPT_RT)
147 #define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
149 /* Locks on RT do not disable preemption */
150 #define PREEMPT_LOCK_OFFSET 0
154 * The preempt_count offset needed for things like:
158 * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
159 * softirqs, such that unlock sequences of:
166 #define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
169 * Are we running in atomic context? WARNING: this macro cannot
170 * always detect atomic context; in particular, it cannot know about
171 * held spinlocks in non-preemptible kernels. Thus it should not be
172 * used in the general case to determine whether sleeping is possible.
173 * Do not use in_atomic() in driver code.
175 #define in_atomic() (preempt_count() != 0)
178 * Check whether we were atomic before we did preempt_disable():
179 * (used by the scheduler)
181 #define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
183 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
184 extern void preempt_count_add(int val);
185 extern void preempt_count_sub(int val);
186 #define preempt_count_dec_and_test() \
187 ({ preempt_count_sub(1); should_resched(0); })
189 #define preempt_count_add(val) __preempt_count_add(val)
190 #define preempt_count_sub(val) __preempt_count_sub(val)
191 #define preempt_count_dec_and_test() __preempt_count_dec_and_test()
194 #define __preempt_count_inc() __preempt_count_add(1)
195 #define __preempt_count_dec() __preempt_count_sub(1)
197 #define preempt_count_inc() preempt_count_add(1)
198 #define preempt_count_dec() preempt_count_sub(1)
200 #ifdef CONFIG_PREEMPT_COUNT
202 #define preempt_disable() \
204 preempt_count_inc(); \
208 #define sched_preempt_enable_no_resched() \
211 preempt_count_dec(); \
214 #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
216 #define preemptible() (preempt_count() == 0 && !irqs_disabled())
218 #ifdef CONFIG_PREEMPTION
219 #define preempt_enable() \
222 if (unlikely(preempt_count_dec_and_test())) \
223 __preempt_schedule(); \
226 #define preempt_enable_notrace() \
229 if (unlikely(__preempt_count_dec_and_test())) \
230 __preempt_schedule_notrace(); \
233 #define preempt_check_resched() \
235 if (should_resched(0)) \
236 __preempt_schedule(); \
239 #else /* !CONFIG_PREEMPTION */
240 #define preempt_enable() \
243 preempt_count_dec(); \
246 #define preempt_enable_notrace() \
249 __preempt_count_dec(); \
252 #define preempt_check_resched() do { } while (0)
253 #endif /* CONFIG_PREEMPTION */
255 #define preempt_disable_notrace() \
257 __preempt_count_inc(); \
261 #define preempt_enable_no_resched_notrace() \
264 __preempt_count_dec(); \
267 #else /* !CONFIG_PREEMPT_COUNT */
270 * Even if we don't have any preemption, we need preempt disable/enable
271 * to be barriers, so that we don't have things like get_user/put_user
272 * that can cause faults and scheduling migrate into our preempt-protected
275 #define preempt_disable() barrier()
276 #define sched_preempt_enable_no_resched() barrier()
277 #define preempt_enable_no_resched() barrier()
278 #define preempt_enable() barrier()
279 #define preempt_check_resched() do { } while (0)
281 #define preempt_disable_notrace() barrier()
282 #define preempt_enable_no_resched_notrace() barrier()
283 #define preempt_enable_notrace() barrier()
284 #define preemptible() 0
286 #endif /* CONFIG_PREEMPT_COUNT */
290 * Modules have no business playing preemption tricks.
292 #undef sched_preempt_enable_no_resched
293 #undef preempt_enable_no_resched
294 #undef preempt_enable_no_resched_notrace
295 #undef preempt_check_resched
298 #define preempt_set_need_resched() \
300 set_preempt_need_resched(); \
302 #define preempt_fold_need_resched() \
304 if (tif_need_resched()) \
305 set_preempt_need_resched(); \
308 #ifdef CONFIG_PREEMPT_NOTIFIERS
310 struct preempt_notifier;
313 * preempt_ops - notifiers called when a task is preempted and rescheduled
314 * @sched_in: we're about to be rescheduled:
315 * notifier: struct preempt_notifier for the task being scheduled
316 * cpu: cpu we're scheduled on
317 * @sched_out: we've just been preempted
318 * notifier: struct preempt_notifier for the task being preempted
319 * next: the task that's kicking us out
321 * Please note that sched_in and out are called under different
322 * contexts. sched_out is called with rq lock held and irq disabled
323 * while sched_in is called without rq lock and irq enabled. This
324 * difference is intentional and depended upon by its users.
327 void (*sched_in)(struct preempt_notifier *notifier, int cpu);
328 void (*sched_out)(struct preempt_notifier *notifier,
329 struct task_struct *next);
333 * preempt_notifier - key for installing preemption notifiers
334 * @link: internal use
335 * @ops: defines the notifier functions to be called
337 * Usually used in conjunction with container_of().
339 struct preempt_notifier {
340 struct hlist_node link;
341 struct preempt_ops *ops;
344 void preempt_notifier_inc(void);
345 void preempt_notifier_dec(void);
346 void preempt_notifier_register(struct preempt_notifier *notifier);
347 void preempt_notifier_unregister(struct preempt_notifier *notifier);
349 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
350 struct preempt_ops *ops)
352 INIT_HLIST_NODE(¬ifier->link);
361 * Migrate-Disable and why it is undesired.
363 * When a preempted task becomes elegible to run under the ideal model (IOW it
364 * becomes one of the M highest priority tasks), it might still have to wait
365 * for the preemptee's migrate_disable() section to complete. Thereby suffering
366 * a reduction in bandwidth in the exact duration of the migrate_disable()
369 * Per this argument, the change from preempt_disable() to migrate_disable()
372 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
373 * it would have had to wait for the lower priority task.
375 * - a lower priority tasks; which under preempt_disable() could've instantly
376 * migrated away when another CPU becomes available, is now constrained
377 * by the ability to push the higher priority task away, which might itself be
378 * in a migrate_disable() section, reducing it's available bandwidth.
380 * IOW it trades latency / moves the interference term, but it stays in the
381 * system, and as long as it remains unbounded, the system is not fully
385 * The reason we have it anyway.
387 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
388 * number of primitives into becoming preemptible, they would also allow
389 * migration. This turns out to break a bunch of per-cpu usage. To this end,
390 * all these primitives employ migirate_disable() to restore this implicit
393 * This is a 'temporary' work-around at best. The correct solution is getting
394 * rid of the above assumptions and reworking the code to employ explicit
395 * per-cpu locking or short preempt-disable regions.
397 * The end goal must be to get rid of migrate_disable(), alternatively we need
398 * a schedulability theory that does not depend on abritrary migration.
401 * Notes on the implementation.
403 * The implementation is particularly tricky since existing code patterns
404 * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
405 * This means that it cannot use cpus_read_lock() to serialize against hotplug,
406 * nor can it easily migrate itself into a pending affinity mask change on
410 * Note: even non-work-conserving schedulers like semi-partitioned depends on
411 * migration, so migrate_disable() is not only a problem for
412 * work-conserving schedulers.
415 extern void migrate_disable(void);
416 extern void migrate_enable(void);
420 static inline void migrate_disable(void) { }
421 static inline void migrate_enable(void) { }
423 #endif /* CONFIG_SMP */
426 * preempt_disable_nested - Disable preemption inside a normally preempt disabled section
428 * Use for code which requires preemption protection inside a critical
429 * section which has preemption disabled implicitly on non-PREEMPT_RT
430 * enabled kernels, by e.g.:
431 * - holding a spinlock/rwlock
432 * - soft interrupt context
433 * - regular interrupt handlers
435 * On PREEMPT_RT enabled kernels spinlock/rwlock held sections, soft
436 * interrupt context and regular interrupt handlers are preemptible and
437 * only prevent migration. preempt_disable_nested() ensures that preemption
438 * is disabled for cases which require CPU local serialization even on
439 * PREEMPT_RT. For non-PREEMPT_RT kernels this is a NOP.
441 * The use cases are code sequences which are not serialized by a
442 * particular lock instance, e.g.:
443 * - seqcount write side critical sections where the seqcount is not
444 * associated to a particular lock and therefore the automatic
445 * protection mechanism does not work. This prevents a live lock
446 * against a preempting high priority reader.
447 * - RMW per CPU variable updates like vmstat.
449 /* Macro to avoid header recursion hell vs. lockdep */
450 #define preempt_disable_nested() \
452 if (IS_ENABLED(CONFIG_PREEMPT_RT)) \
455 lockdep_assert_preemption_disabled(); \
459 * preempt_enable_nested - Undo the effect of preempt_disable_nested()
461 static __always_inline void preempt_enable_nested(void)
463 if (IS_ENABLED(CONFIG_PREEMPT_RT))
467 DEFINE_LOCK_GUARD_0(preempt, preempt_disable(), preempt_enable())
468 DEFINE_LOCK_GUARD_0(preempt_notrace, preempt_disable_notrace(), preempt_enable_notrace())
469 DEFINE_LOCK_GUARD_0(migrate, migrate_disable(), migrate_enable())
471 #endif /* __LINUX_PREEMPT_H */