xref: /openbmc/linux/include/linux/preempt.h (revision 64794d6db49730d22f440aef0cf4da98a56a4ea3)
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 #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)
84 #else
85 # define softirq_count()	(preempt_count() & SOFTIRQ_MASK)
86 #endif
87 #define irq_count()	(nmi_count() | hardirq_count() | softirq_count())
88 
89 /*
90  * Macros to retrieve the current execution context:
91  *
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
96  */
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()))
101 
102 /*
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
107  */
108 #define in_irq()		(hardirq_count())
109 #define in_softirq()		(softirq_count())
110 #define in_interrupt()		(irq_count())
111 
112 /*
113  * The preempt_count offset after preempt_disable();
114  */
115 #if defined(CONFIG_PREEMPT_COUNT)
116 # define PREEMPT_DISABLE_OFFSET	PREEMPT_OFFSET
117 #else
118 # define PREEMPT_DISABLE_OFFSET	0
119 #endif
120 
121 /*
122  * The preempt_count offset after spin_lock()
123  */
124 #if !defined(CONFIG_PREEMPT_RT)
125 #define PREEMPT_LOCK_OFFSET	PREEMPT_DISABLE_OFFSET
126 #else
127 #define PREEMPT_LOCK_OFFSET	0
128 #endif
129 
130 /*
131  * The preempt_count offset needed for things like:
132  *
133  *  spin_lock_bh()
134  *
135  * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
136  * softirqs, such that unlock sequences of:
137  *
138  *  spin_unlock();
139  *  local_bh_enable();
140  *
141  * Work as expected.
142  */
143 #define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
144 
145 /*
146  * Are we running in atomic context?  WARNING: this macro cannot
147  * always detect atomic context; in particular, it cannot know about
148  * held spinlocks in non-preemptible kernels.  Thus it should not be
149  * used in the general case to determine whether sleeping is possible.
150  * Do not use in_atomic() in driver code.
151  */
152 #define in_atomic()	(preempt_count() != 0)
153 
154 /*
155  * Check whether we were atomic before we did preempt_disable():
156  * (used by the scheduler)
157  */
158 #define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
159 
160 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
161 extern void preempt_count_add(int val);
162 extern void preempt_count_sub(int val);
163 #define preempt_count_dec_and_test() \
164 	({ preempt_count_sub(1); should_resched(0); })
165 #else
166 #define preempt_count_add(val)	__preempt_count_add(val)
167 #define preempt_count_sub(val)	__preempt_count_sub(val)
168 #define preempt_count_dec_and_test() __preempt_count_dec_and_test()
169 #endif
170 
171 #define __preempt_count_inc() __preempt_count_add(1)
172 #define __preempt_count_dec() __preempt_count_sub(1)
173 
174 #define preempt_count_inc() preempt_count_add(1)
175 #define preempt_count_dec() preempt_count_sub(1)
176 
177 #ifdef CONFIG_PREEMPT_COUNT
178 
179 #define preempt_disable() \
180 do { \
181 	preempt_count_inc(); \
182 	barrier(); \
183 } while (0)
184 
185 #define sched_preempt_enable_no_resched() \
186 do { \
187 	barrier(); \
188 	preempt_count_dec(); \
189 } while (0)
190 
191 #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
192 
193 #define preemptible()	(preempt_count() == 0 && !irqs_disabled())
194 
195 #ifdef CONFIG_PREEMPTION
196 #define preempt_enable() \
197 do { \
198 	barrier(); \
199 	if (unlikely(preempt_count_dec_and_test())) \
200 		__preempt_schedule(); \
201 } while (0)
202 
203 #define preempt_enable_notrace() \
204 do { \
205 	barrier(); \
206 	if (unlikely(__preempt_count_dec_and_test())) \
207 		__preempt_schedule_notrace(); \
208 } while (0)
209 
210 #define preempt_check_resched() \
211 do { \
212 	if (should_resched(0)) \
213 		__preempt_schedule(); \
214 } while (0)
215 
216 #else /* !CONFIG_PREEMPTION */
217 #define preempt_enable() \
218 do { \
219 	barrier(); \
220 	preempt_count_dec(); \
221 } while (0)
222 
223 #define preempt_enable_notrace() \
224 do { \
225 	barrier(); \
226 	__preempt_count_dec(); \
227 } while (0)
228 
229 #define preempt_check_resched() do { } while (0)
230 #endif /* CONFIG_PREEMPTION */
231 
232 #define preempt_disable_notrace() \
233 do { \
234 	__preempt_count_inc(); \
235 	barrier(); \
236 } while (0)
237 
238 #define preempt_enable_no_resched_notrace() \
239 do { \
240 	barrier(); \
241 	__preempt_count_dec(); \
242 } while (0)
243 
244 #else /* !CONFIG_PREEMPT_COUNT */
245 
246 /*
247  * Even if we don't have any preemption, we need preempt disable/enable
248  * to be barriers, so that we don't have things like get_user/put_user
249  * that can cause faults and scheduling migrate into our preempt-protected
250  * region.
251  */
252 #define preempt_disable()			barrier()
253 #define sched_preempt_enable_no_resched()	barrier()
254 #define preempt_enable_no_resched()		barrier()
255 #define preempt_enable()			barrier()
256 #define preempt_check_resched()			do { } while (0)
257 
258 #define preempt_disable_notrace()		barrier()
259 #define preempt_enable_no_resched_notrace()	barrier()
260 #define preempt_enable_notrace()		barrier()
261 #define preemptible()				0
262 
263 #endif /* CONFIG_PREEMPT_COUNT */
264 
265 #ifdef MODULE
266 /*
267  * Modules have no business playing preemption tricks.
268  */
269 #undef sched_preempt_enable_no_resched
270 #undef preempt_enable_no_resched
271 #undef preempt_enable_no_resched_notrace
272 #undef preempt_check_resched
273 #endif
274 
275 #define preempt_set_need_resched() \
276 do { \
277 	set_preempt_need_resched(); \
278 } while (0)
279 #define preempt_fold_need_resched() \
280 do { \
281 	if (tif_need_resched()) \
282 		set_preempt_need_resched(); \
283 } while (0)
284 
285 #ifdef CONFIG_PREEMPT_NOTIFIERS
286 
287 struct preempt_notifier;
288 
289 /**
290  * preempt_ops - notifiers called when a task is preempted and rescheduled
291  * @sched_in: we're about to be rescheduled:
292  *    notifier: struct preempt_notifier for the task being scheduled
293  *    cpu:  cpu we're scheduled on
294  * @sched_out: we've just been preempted
295  *    notifier: struct preempt_notifier for the task being preempted
296  *    next: the task that's kicking us out
297  *
298  * Please note that sched_in and out are called under different
299  * contexts.  sched_out is called with rq lock held and irq disabled
300  * while sched_in is called without rq lock and irq enabled.  This
301  * difference is intentional and depended upon by its users.
302  */
303 struct preempt_ops {
304 	void (*sched_in)(struct preempt_notifier *notifier, int cpu);
305 	void (*sched_out)(struct preempt_notifier *notifier,
306 			  struct task_struct *next);
307 };
308 
309 /**
310  * preempt_notifier - key for installing preemption notifiers
311  * @link: internal use
312  * @ops: defines the notifier functions to be called
313  *
314  * Usually used in conjunction with container_of().
315  */
316 struct preempt_notifier {
317 	struct hlist_node link;
318 	struct preempt_ops *ops;
319 };
320 
321 void preempt_notifier_inc(void);
322 void preempt_notifier_dec(void);
323 void preempt_notifier_register(struct preempt_notifier *notifier);
324 void preempt_notifier_unregister(struct preempt_notifier *notifier);
325 
326 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
327 				     struct preempt_ops *ops)
328 {
329 	INIT_HLIST_NODE(&notifier->link);
330 	notifier->ops = ops;
331 }
332 
333 #endif
334 
335 #ifdef CONFIG_SMP
336 
337 /*
338  * Migrate-Disable and why it is undesired.
339  *
340  * When a preempted task becomes elegible to run under the ideal model (IOW it
341  * becomes one of the M highest priority tasks), it might still have to wait
342  * for the preemptee's migrate_disable() section to complete. Thereby suffering
343  * a reduction in bandwidth in the exact duration of the migrate_disable()
344  * section.
345  *
346  * Per this argument, the change from preempt_disable() to migrate_disable()
347  * gets us:
348  *
349  * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
350  *   it would have had to wait for the lower priority task.
351  *
352  * - a lower priority tasks; which under preempt_disable() could've instantly
353  *   migrated away when another CPU becomes available, is now constrained
354  *   by the ability to push the higher priority task away, which might itself be
355  *   in a migrate_disable() section, reducing it's available bandwidth.
356  *
357  * IOW it trades latency / moves the interference term, but it stays in the
358  * system, and as long as it remains unbounded, the system is not fully
359  * deterministic.
360  *
361  *
362  * The reason we have it anyway.
363  *
364  * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
365  * number of primitives into becoming preemptible, they would also allow
366  * migration. This turns out to break a bunch of per-cpu usage. To this end,
367  * all these primitives employ migirate_disable() to restore this implicit
368  * assumption.
369  *
370  * This is a 'temporary' work-around at best. The correct solution is getting
371  * rid of the above assumptions and reworking the code to employ explicit
372  * per-cpu locking or short preempt-disable regions.
373  *
374  * The end goal must be to get rid of migrate_disable(), alternatively we need
375  * a schedulability theory that does not depend on abritrary migration.
376  *
377  *
378  * Notes on the implementation.
379  *
380  * The implementation is particularly tricky since existing code patterns
381  * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
382  * This means that it cannot use cpus_read_lock() to serialize against hotplug,
383  * nor can it easily migrate itself into a pending affinity mask change on
384  * migrate_enable().
385  *
386  *
387  * Note: even non-work-conserving schedulers like semi-partitioned depends on
388  *       migration, so migrate_disable() is not only a problem for
389  *       work-conserving schedulers.
390  *
391  */
392 extern void migrate_disable(void);
393 extern void migrate_enable(void);
394 
395 #else
396 
397 static inline void migrate_disable(void) { }
398 static inline void migrate_enable(void) { }
399 
400 #endif /* CONFIG_SMP */
401 
402 #endif /* __LINUX_PREEMPT_H */
403