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/list.h>
14 * We put the hardirq and softirq counter into the preemption
15 * counter. The bitmask has the following meaning:
17 * - bits 0-7 are the preemption count (max preemption depth: 256)
18 * - bits 8-15 are the softirq count (max # of softirqs: 256)
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.
26 * PREEMPT_MASK: 0x000000ff
27 * SOFTIRQ_MASK: 0x0000ff00
28 * HARDIRQ_MASK: 0x000f0000
29 * NMI_MASK: 0x00f00000
30 * PREEMPT_NEED_RESCHED: 0x80000000
32 #define PREEMPT_BITS 8
33 #define SOFTIRQ_BITS 8
34 #define HARDIRQ_BITS 4
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)
42 #define __IRQ_MASK(x) ((1UL << (x))-1)
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)
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)
54 #define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
56 #define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
59 * Disable preemption until the scheduler is running -- use an unconditional
60 * value so that it also works on !PREEMPT_COUNT kernels.
62 * Reset by start_kernel()->sched_init()->init_idle()->init_idle_preempt_count().
64 #define INIT_PREEMPT_COUNT PREEMPT_OFFSET
67 * Initial preempt_count value; reflects the preempt_count schedule invariant
68 * which states that during context switches:
70 * preempt_count() == 2*PREEMPT_DISABLE_OFFSET
72 * Note: PREEMPT_DISABLE_OFFSET is 0 for !PREEMPT_COUNT kernels.
73 * Note: See finish_task_switch().
75 #define FORK_PREEMPT_COUNT (2*PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
77 /* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
78 #include <asm/preempt.h>
80 #define nmi_count() (preempt_count() & NMI_MASK)
81 #define hardirq_count() (preempt_count() & HARDIRQ_MASK)
82 #ifdef CONFIG_PREEMPT_RT
83 # define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
85 # define softirq_count() (preempt_count() & SOFTIRQ_MASK)
87 #define irq_count() (nmi_count() | hardirq_count() | softirq_count())
90 * Macros to retrieve the current execution context:
92 * in_nmi() - We're in NMI context
93 * in_hardirq() - We're in hard IRQ context
94 * in_serving_softirq() - We're in softirq context
95 * in_task() - We're in task context
97 #define in_nmi() (nmi_count())
98 #define in_hardirq() (hardirq_count())
99 #define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
100 #define in_task() (!(in_nmi() | in_hardirq() | in_serving_softirq()))
103 * The following macros are deprecated and should not be used in new code:
104 * in_irq() - Obsolete version of in_hardirq()
105 * in_softirq() - We have BH disabled, or are processing softirqs
106 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
108 #define in_irq() (hardirq_count())
109 #define in_softirq() (softirq_count())
110 #define in_interrupt() (irq_count())
113 * The preempt_count offset after preempt_disable();
115 #if defined(CONFIG_PREEMPT_COUNT)
116 # define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET
118 # define PREEMPT_DISABLE_OFFSET 0
122 * The preempt_count offset after spin_lock()
124 #if !defined(CONFIG_PREEMPT_RT)
125 #define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
127 /* Locks on RT do not disable preemption */
128 #define PREEMPT_LOCK_OFFSET 0
132 * The preempt_count offset needed for things like:
136 * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
137 * softirqs, such that unlock sequences of:
144 #define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
147 * Are we running in atomic context? WARNING: this macro cannot
148 * always detect atomic context; in particular, it cannot know about
149 * held spinlocks in non-preemptible kernels. Thus it should not be
150 * used in the general case to determine whether sleeping is possible.
151 * Do not use in_atomic() in driver code.
153 #define in_atomic() (preempt_count() != 0)
156 * Check whether we were atomic before we did preempt_disable():
157 * (used by the scheduler)
159 #define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
161 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
162 extern void preempt_count_add(int val);
163 extern void preempt_count_sub(int val);
164 #define preempt_count_dec_and_test() \
165 ({ preempt_count_sub(1); should_resched(0); })
167 #define preempt_count_add(val) __preempt_count_add(val)
168 #define preempt_count_sub(val) __preempt_count_sub(val)
169 #define preempt_count_dec_and_test() __preempt_count_dec_and_test()
172 #define __preempt_count_inc() __preempt_count_add(1)
173 #define __preempt_count_dec() __preempt_count_sub(1)
175 #define preempt_count_inc() preempt_count_add(1)
176 #define preempt_count_dec() preempt_count_sub(1)
178 #ifdef CONFIG_PREEMPT_LAZY
179 #define add_preempt_lazy_count(val) do { preempt_lazy_count() += (val); } while (0)
180 #define sub_preempt_lazy_count(val) do { preempt_lazy_count() -= (val); } while (0)
181 #define inc_preempt_lazy_count() add_preempt_lazy_count(1)
182 #define dec_preempt_lazy_count() sub_preempt_lazy_count(1)
183 #define preempt_lazy_count() (current_thread_info()->preempt_lazy_count)
185 #define add_preempt_lazy_count(val) do { } while (0)
186 #define sub_preempt_lazy_count(val) do { } while (0)
187 #define inc_preempt_lazy_count() do { } while (0)
188 #define dec_preempt_lazy_count() do { } while (0)
189 #define preempt_lazy_count() (0)
192 #ifdef CONFIG_PREEMPT_COUNT
194 #define preempt_disable() \
196 preempt_count_inc(); \
200 #define preempt_lazy_disable() \
202 inc_preempt_lazy_count(); \
206 #define sched_preempt_enable_no_resched() \
209 preempt_count_dec(); \
212 #ifndef CONFIG_PREEMPT_RT
213 # define preempt_enable_no_resched() sched_preempt_enable_no_resched()
214 # define preempt_check_resched_rt() barrier();
216 # define preempt_enable_no_resched() preempt_enable()
217 # define preempt_check_resched_rt() preempt_check_resched()
220 #define preemptible() (preempt_count() == 0 && !irqs_disabled())
222 #ifdef CONFIG_PREEMPTION
223 #define preempt_enable() \
226 if (unlikely(preempt_count_dec_and_test())) \
227 __preempt_schedule(); \
230 #define preempt_enable_notrace() \
233 if (unlikely(__preempt_count_dec_and_test())) \
234 __preempt_schedule_notrace(); \
237 #define preempt_check_resched() \
239 if (should_resched(0)) \
240 __preempt_schedule(); \
244 * open code preempt_check_resched() because it is not exported to modules and
245 * used by local_unlock() or bpf_enable_instrumentation().
247 #define preempt_lazy_enable() \
249 dec_preempt_lazy_count(); \
251 if (should_resched(0)) \
252 __preempt_schedule(); \
255 #else /* !CONFIG_PREEMPTION */
256 #define preempt_enable() \
259 preempt_count_dec(); \
262 #define preempt_lazy_enable() \
264 dec_preempt_lazy_count(); \
268 #define preempt_enable_notrace() \
271 __preempt_count_dec(); \
274 #define preempt_check_resched() do { } while (0)
275 #endif /* CONFIG_PREEMPTION */
277 #define preempt_disable_notrace() \
279 __preempt_count_inc(); \
283 #define preempt_enable_no_resched_notrace() \
286 __preempt_count_dec(); \
289 #else /* !CONFIG_PREEMPT_COUNT */
292 * Even if we don't have any preemption, we need preempt disable/enable
293 * to be barriers, so that we don't have things like get_user/put_user
294 * that can cause faults and scheduling migrate into our preempt-protected
297 #define preempt_disable() barrier()
298 #define sched_preempt_enable_no_resched() barrier()
299 #define preempt_enable_no_resched() barrier()
300 #define preempt_enable() barrier()
301 #define preempt_check_resched() do { } while (0)
303 #define preempt_disable_notrace() barrier()
304 #define preempt_enable_no_resched_notrace() barrier()
305 #define preempt_enable_notrace() barrier()
306 #define preempt_check_resched_rt() barrier()
307 #define preemptible() 0
309 #define preempt_lazy_disable() barrier()
310 #define preempt_lazy_enable() barrier()
312 #endif /* CONFIG_PREEMPT_COUNT */
316 * Modules have no business playing preemption tricks.
318 #undef sched_preempt_enable_no_resched
319 #undef preempt_enable_no_resched
320 #undef preempt_enable_no_resched_notrace
321 #undef preempt_check_resched
324 #define preempt_set_need_resched() \
326 set_preempt_need_resched(); \
328 #define preempt_fold_need_resched() \
330 if (tif_need_resched_now()) \
331 set_preempt_need_resched(); \
334 #ifdef CONFIG_PREEMPT_NOTIFIERS
336 struct preempt_notifier;
339 * preempt_ops - notifiers called when a task is preempted and rescheduled
340 * @sched_in: we're about to be rescheduled:
341 * notifier: struct preempt_notifier for the task being scheduled
342 * cpu: cpu we're scheduled on
343 * @sched_out: we've just been preempted
344 * notifier: struct preempt_notifier for the task being preempted
345 * next: the task that's kicking us out
347 * Please note that sched_in and out are called under different
348 * contexts. sched_out is called with rq lock held and irq disabled
349 * while sched_in is called without rq lock and irq enabled. This
350 * difference is intentional and depended upon by its users.
353 void (*sched_in)(struct preempt_notifier *notifier, int cpu);
354 void (*sched_out)(struct preempt_notifier *notifier,
355 struct task_struct *next);
359 * preempt_notifier - key for installing preemption notifiers
360 * @link: internal use
361 * @ops: defines the notifier functions to be called
363 * Usually used in conjunction with container_of().
365 struct preempt_notifier {
366 struct hlist_node link;
367 struct preempt_ops *ops;
370 void preempt_notifier_inc(void);
371 void preempt_notifier_dec(void);
372 void preempt_notifier_register(struct preempt_notifier *notifier);
373 void preempt_notifier_unregister(struct preempt_notifier *notifier);
375 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
376 struct preempt_ops *ops)
378 INIT_HLIST_NODE(¬ifier->link);
387 * Migrate-Disable and why it is undesired.
389 * When a preempted task becomes elegible to run under the ideal model (IOW it
390 * becomes one of the M highest priority tasks), it might still have to wait
391 * for the preemptee's migrate_disable() section to complete. Thereby suffering
392 * a reduction in bandwidth in the exact duration of the migrate_disable()
395 * Per this argument, the change from preempt_disable() to migrate_disable()
398 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
399 * it would have had to wait for the lower priority task.
401 * - a lower priority tasks; which under preempt_disable() could've instantly
402 * migrated away when another CPU becomes available, is now constrained
403 * by the ability to push the higher priority task away, which might itself be
404 * in a migrate_disable() section, reducing it's available bandwidth.
406 * IOW it trades latency / moves the interference term, but it stays in the
407 * system, and as long as it remains unbounded, the system is not fully
411 * The reason we have it anyway.
413 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
414 * number of primitives into becoming preemptible, they would also allow
415 * migration. This turns out to break a bunch of per-cpu usage. To this end,
416 * all these primitives employ migirate_disable() to restore this implicit
419 * This is a 'temporary' work-around at best. The correct solution is getting
420 * rid of the above assumptions and reworking the code to employ explicit
421 * per-cpu locking or short preempt-disable regions.
423 * The end goal must be to get rid of migrate_disable(), alternatively we need
424 * a schedulability theory that does not depend on abritrary migration.
427 * Notes on the implementation.
429 * The implementation is particularly tricky since existing code patterns
430 * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
431 * This means that it cannot use cpus_read_lock() to serialize against hotplug,
432 * nor can it easily migrate itself into a pending affinity mask change on
436 * Note: even non-work-conserving schedulers like semi-partitioned depends on
437 * migration, so migrate_disable() is not only a problem for
438 * work-conserving schedulers.
441 extern void migrate_disable(void);
442 extern void migrate_enable(void);
446 static inline void migrate_disable(void)
448 preempt_lazy_disable();
451 static inline void migrate_enable(void)
453 preempt_lazy_enable();
456 #endif /* CONFIG_SMP */
458 #endif /* __LINUX_PREEMPT_H */