• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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  */
interrupt_context_level(void)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 /*
102  * These macro definitions avoid redundant invocations of preempt_count()
103  * because such invocations would result in redundant loads given that
104  * preempt_count() is commonly implemented with READ_ONCE().
105  */
106 
107 #define nmi_count()	(preempt_count() & NMI_MASK)
108 #define hardirq_count()	(preempt_count() & HARDIRQ_MASK)
109 #ifdef CONFIG_PREEMPT_RT
110 # define softirq_count()	(current->softirq_disable_cnt & SOFTIRQ_MASK)
111 # define irq_count()		((preempt_count() & (NMI_MASK | HARDIRQ_MASK)) | softirq_count())
112 #else
113 # define softirq_count()	(preempt_count() & SOFTIRQ_MASK)
114 # define irq_count()		(preempt_count() & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_MASK))
115 #endif
116 
117 /*
118  * Macros to retrieve the current execution context:
119  *
120  * in_nmi()		- We're in NMI context
121  * in_hardirq()		- We're in hard IRQ context
122  * in_serving_softirq()	- We're in softirq context
123  * in_task()		- We're in task context
124  */
125 #define in_nmi()		(nmi_count())
126 #define in_hardirq()		(hardirq_count())
127 #define in_serving_softirq()	(softirq_count() & SOFTIRQ_OFFSET)
128 #ifdef CONFIG_PREEMPT_RT
129 # define in_task()		(!((preempt_count() & (NMI_MASK | HARDIRQ_MASK)) | in_serving_softirq()))
130 #else
131 # define in_task()		(!(preempt_count() & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
132 #endif
133 
134 /*
135  * The following macros are deprecated and should not be used in new code:
136  * in_irq()       - Obsolete version of in_hardirq()
137  * in_softirq()   - We have BH disabled, or are processing softirqs
138  * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
139  */
140 #define in_irq()		(hardirq_count())
141 #define in_softirq()		(softirq_count())
142 #define in_interrupt()		(irq_count())
143 
144 /*
145  * The preempt_count offset after preempt_disable();
146  */
147 #if defined(CONFIG_PREEMPT_COUNT)
148 # define PREEMPT_DISABLE_OFFSET	PREEMPT_OFFSET
149 #else
150 # define PREEMPT_DISABLE_OFFSET	0
151 #endif
152 
153 /*
154  * The preempt_count offset after spin_lock()
155  */
156 #if !defined(CONFIG_PREEMPT_RT)
157 #define PREEMPT_LOCK_OFFSET	PREEMPT_DISABLE_OFFSET
158 #else
159 #define PREEMPT_LOCK_OFFSET	0
160 #endif
161 
162 /*
163  * The preempt_count offset needed for things like:
164  *
165  *  spin_lock_bh()
166  *
167  * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
168  * softirqs, such that unlock sequences of:
169  *
170  *  spin_unlock();
171  *  local_bh_enable();
172  *
173  * Work as expected.
174  */
175 #define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
176 
177 /*
178  * Are we running in atomic context?  WARNING: this macro cannot
179  * always detect atomic context; in particular, it cannot know about
180  * held spinlocks in non-preemptible kernels.  Thus it should not be
181  * used in the general case to determine whether sleeping is possible.
182  * Do not use in_atomic() in driver code.
183  */
184 #define in_atomic()	(preempt_count() != 0)
185 
186 /*
187  * Check whether we were atomic before we did preempt_disable():
188  * (used by the scheduler)
189  */
190 #define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
191 
192 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
193 extern void preempt_count_add(int val);
194 extern void preempt_count_sub(int val);
195 #define preempt_count_dec_and_test() \
196 	({ preempt_count_sub(1); should_resched(0); })
197 #else
198 #define preempt_count_add(val)	__preempt_count_add(val)
199 #define preempt_count_sub(val)	__preempt_count_sub(val)
200 #define preempt_count_dec_and_test() __preempt_count_dec_and_test()
201 #endif
202 
203 #define __preempt_count_inc() __preempt_count_add(1)
204 #define __preempt_count_dec() __preempt_count_sub(1)
205 
206 #define preempt_count_inc() preempt_count_add(1)
207 #define preempt_count_dec() preempt_count_sub(1)
208 
209 #ifdef CONFIG_PREEMPT_COUNT
210 
211 #define preempt_disable() \
212 do { \
213 	preempt_count_inc(); \
214 	barrier(); \
215 } while (0)
216 
217 #define sched_preempt_enable_no_resched() \
218 do { \
219 	barrier(); \
220 	preempt_count_dec(); \
221 } while (0)
222 
223 #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
224 
225 #define preemptible()	(preempt_count() == 0 && !irqs_disabled())
226 
227 #ifdef CONFIG_PREEMPTION
228 #define preempt_enable() \
229 do { \
230 	barrier(); \
231 	if (unlikely(preempt_count_dec_and_test())) \
232 		__preempt_schedule(); \
233 } while (0)
234 
235 #define preempt_enable_notrace() \
236 do { \
237 	barrier(); \
238 	if (unlikely(__preempt_count_dec_and_test())) \
239 		__preempt_schedule_notrace(); \
240 } while (0)
241 
242 #define preempt_check_resched() \
243 do { \
244 	if (should_resched(0)) \
245 		__preempt_schedule(); \
246 } while (0)
247 
248 #else /* !CONFIG_PREEMPTION */
249 #define preempt_enable() \
250 do { \
251 	barrier(); \
252 	preempt_count_dec(); \
253 } while (0)
254 
255 #define preempt_enable_notrace() \
256 do { \
257 	barrier(); \
258 	__preempt_count_dec(); \
259 } while (0)
260 
261 #define preempt_check_resched() do { } while (0)
262 #endif /* CONFIG_PREEMPTION */
263 
264 #define preempt_disable_notrace() \
265 do { \
266 	__preempt_count_inc(); \
267 	barrier(); \
268 } while (0)
269 
270 #define preempt_enable_no_resched_notrace() \
271 do { \
272 	barrier(); \
273 	__preempt_count_dec(); \
274 } while (0)
275 
276 #else /* !CONFIG_PREEMPT_COUNT */
277 
278 /*
279  * Even if we don't have any preemption, we need preempt disable/enable
280  * to be barriers, so that we don't have things like get_user/put_user
281  * that can cause faults and scheduling migrate into our preempt-protected
282  * region.
283  */
284 #define preempt_disable()			barrier()
285 #define sched_preempt_enable_no_resched()	barrier()
286 #define preempt_enable_no_resched()		barrier()
287 #define preempt_enable()			barrier()
288 #define preempt_check_resched()			do { } while (0)
289 
290 #define preempt_disable_notrace()		barrier()
291 #define preempt_enable_no_resched_notrace()	barrier()
292 #define preempt_enable_notrace()		barrier()
293 #define preemptible()				0
294 
295 #endif /* CONFIG_PREEMPT_COUNT */
296 
297 #ifdef MODULE
298 /*
299  * Modules have no business playing preemption tricks.
300  */
301 #undef sched_preempt_enable_no_resched
302 #undef preempt_enable_no_resched
303 #undef preempt_enable_no_resched_notrace
304 #undef preempt_check_resched
305 #endif
306 
307 #define preempt_set_need_resched() \
308 do { \
309 	set_preempt_need_resched(); \
310 } while (0)
311 #define preempt_fold_need_resched() \
312 do { \
313 	if (tif_need_resched()) \
314 		set_preempt_need_resched(); \
315 } while (0)
316 
317 #ifdef CONFIG_PREEMPT_NOTIFIERS
318 
319 struct preempt_notifier;
320 
321 /**
322  * preempt_ops - notifiers called when a task is preempted and rescheduled
323  * @sched_in: we're about to be rescheduled:
324  *    notifier: struct preempt_notifier for the task being scheduled
325  *    cpu:  cpu we're scheduled on
326  * @sched_out: we've just been preempted
327  *    notifier: struct preempt_notifier for the task being preempted
328  *    next: the task that's kicking us out
329  *
330  * Please note that sched_in and out are called under different
331  * contexts.  sched_out is called with rq lock held and irq disabled
332  * while sched_in is called without rq lock and irq enabled.  This
333  * difference is intentional and depended upon by its users.
334  */
335 struct preempt_ops {
336 	void (*sched_in)(struct preempt_notifier *notifier, int cpu);
337 	void (*sched_out)(struct preempt_notifier *notifier,
338 			  struct task_struct *next);
339 };
340 
341 /**
342  * preempt_notifier - key for installing preemption notifiers
343  * @link: internal use
344  * @ops: defines the notifier functions to be called
345  *
346  * Usually used in conjunction with container_of().
347  */
348 struct preempt_notifier {
349 	struct hlist_node link;
350 	struct preempt_ops *ops;
351 };
352 
353 void preempt_notifier_inc(void);
354 void preempt_notifier_dec(void);
355 void preempt_notifier_register(struct preempt_notifier *notifier);
356 void preempt_notifier_unregister(struct preempt_notifier *notifier);
357 
preempt_notifier_init(struct preempt_notifier * notifier,struct preempt_ops * ops)358 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
359 				     struct preempt_ops *ops)
360 {
361 	INIT_HLIST_NODE(&notifier->link);
362 	notifier->ops = ops;
363 }
364 
365 #endif
366 
367 #ifdef CONFIG_SMP
368 
369 /*
370  * Migrate-Disable and why it is undesired.
371  *
372  * When a preempted task becomes elegible to run under the ideal model (IOW it
373  * becomes one of the M highest priority tasks), it might still have to wait
374  * for the preemptee's migrate_disable() section to complete. Thereby suffering
375  * a reduction in bandwidth in the exact duration of the migrate_disable()
376  * section.
377  *
378  * Per this argument, the change from preempt_disable() to migrate_disable()
379  * gets us:
380  *
381  * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
382  *   it would have had to wait for the lower priority task.
383  *
384  * - a lower priority tasks; which under preempt_disable() could've instantly
385  *   migrated away when another CPU becomes available, is now constrained
386  *   by the ability to push the higher priority task away, which might itself be
387  *   in a migrate_disable() section, reducing it's available bandwidth.
388  *
389  * IOW it trades latency / moves the interference term, but it stays in the
390  * system, and as long as it remains unbounded, the system is not fully
391  * deterministic.
392  *
393  *
394  * The reason we have it anyway.
395  *
396  * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
397  * number of primitives into becoming preemptible, they would also allow
398  * migration. This turns out to break a bunch of per-cpu usage. To this end,
399  * all these primitives employ migirate_disable() to restore this implicit
400  * assumption.
401  *
402  * This is a 'temporary' work-around at best. The correct solution is getting
403  * rid of the above assumptions and reworking the code to employ explicit
404  * per-cpu locking or short preempt-disable regions.
405  *
406  * The end goal must be to get rid of migrate_disable(), alternatively we need
407  * a schedulability theory that does not depend on abritrary migration.
408  *
409  *
410  * Notes on the implementation.
411  *
412  * The implementation is particularly tricky since existing code patterns
413  * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
414  * This means that it cannot use cpus_read_lock() to serialize against hotplug,
415  * nor can it easily migrate itself into a pending affinity mask change on
416  * migrate_enable().
417  *
418  *
419  * Note: even non-work-conserving schedulers like semi-partitioned depends on
420  *       migration, so migrate_disable() is not only a problem for
421  *       work-conserving schedulers.
422  *
423  */
424 extern void migrate_disable(void);
425 extern void migrate_enable(void);
426 
427 #else
428 
migrate_disable(void)429 static inline void migrate_disable(void) { }
migrate_enable(void)430 static inline void migrate_enable(void) { }
431 
432 #endif /* CONFIG_SMP */
433 
434 #endif /* __LINUX_PREEMPT_H */
435