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 */
interrupt_context_level(void)90 static __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_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_softirq() (softirq_count())
141 #define in_interrupt() (irq_count())
142
143 /*
144 * The preempt_count offset after preempt_disable();
145 */
146 #if defined(CONFIG_PREEMPT_COUNT)
147 # define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET
148 #else
149 # define PREEMPT_DISABLE_OFFSET 0
150 #endif
151
152 /*
153 * The preempt_count offset after spin_lock()
154 */
155 #if !defined(CONFIG_PREEMPT_RT)
156 #define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
157 #else
158 /* Locks on RT do not disable preemption */
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 struct task_struct;
321
322 /**
323 * preempt_ops - notifiers called when a task is preempted and rescheduled
324 * @sched_in: we're about to be rescheduled:
325 * notifier: struct preempt_notifier for the task being scheduled
326 * cpu: cpu we're scheduled on
327 * @sched_out: we've just been preempted
328 * notifier: struct preempt_notifier for the task being preempted
329 * next: the task that's kicking us out
330 *
331 * Please note that sched_in and out are called under different
332 * contexts. sched_out is called with rq lock held and irq disabled
333 * while sched_in is called without rq lock and irq enabled. This
334 * difference is intentional and depended upon by its users.
335 */
336 struct preempt_ops {
337 void (*sched_in)(struct preempt_notifier *notifier, int cpu);
338 void (*sched_out)(struct preempt_notifier *notifier,
339 struct task_struct *next);
340 };
341
342 /**
343 * preempt_notifier - key for installing preemption notifiers
344 * @link: internal use
345 * @ops: defines the notifier functions to be called
346 *
347 * Usually used in conjunction with container_of().
348 */
349 struct preempt_notifier {
350 struct hlist_node link;
351 struct preempt_ops *ops;
352 };
353
354 void preempt_notifier_inc(void);
355 void preempt_notifier_dec(void);
356 void preempt_notifier_register(struct preempt_notifier *notifier);
357 void preempt_notifier_unregister(struct preempt_notifier *notifier);
358
preempt_notifier_init(struct preempt_notifier * notifier,struct preempt_ops * ops)359 static inline void preempt_notifier_init(struct preempt_notifier *notifier,
360 struct preempt_ops *ops)
361 {
362 /* INIT_HLIST_NODE() open coded, to avoid dependency on list.h */
363 notifier->link.next = NULL;
364 notifier->link.pprev = NULL;
365 notifier->ops = ops;
366 }
367
368 #endif
369
370 /*
371 * Migrate-Disable and why it is undesired.
372 *
373 * When a preempted task becomes eligible to run under the ideal model (IOW it
374 * becomes one of the M highest priority tasks), it might still have to wait
375 * for the preemptee's migrate_disable() section to complete. Thereby suffering
376 * a reduction in bandwidth in the exact duration of the migrate_disable()
377 * section.
378 *
379 * Per this argument, the change from preempt_disable() to migrate_disable()
380 * gets us:
381 *
382 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
383 * it would have had to wait for the lower priority task.
384 *
385 * - a lower priority tasks; which under preempt_disable() could've instantly
386 * migrated away when another CPU becomes available, is now constrained
387 * by the ability to push the higher priority task away, which might itself be
388 * in a migrate_disable() section, reducing its available bandwidth.
389 *
390 * IOW it trades latency / moves the interference term, but it stays in the
391 * system, and as long as it remains unbounded, the system is not fully
392 * deterministic.
393 *
394 *
395 * The reason we have it anyway.
396 *
397 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
398 * number of primitives into becoming preemptible, they would also allow
399 * migration. This turns out to break a bunch of per-cpu usage. To this end,
400 * all these primitives employ migrate_disable() to restore this implicit
401 * assumption.
402 *
403 * This is a 'temporary' work-around at best. The correct solution is getting
404 * rid of the above assumptions and reworking the code to employ explicit
405 * per-cpu locking or short preempt-disable regions.
406 *
407 * The end goal must be to get rid of migrate_disable(), alternatively we need
408 * a schedulability theory that does not depend on arbitrary migration.
409 *
410 *
411 * Notes on the implementation.
412 *
413 * The implementation is particularly tricky since existing code patterns
414 * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
415 * This means that it cannot use cpus_read_lock() to serialize against hotplug,
416 * nor can it easily migrate itself into a pending affinity mask change on
417 * migrate_enable().
418 *
419 *
420 * Note: even non-work-conserving schedulers like semi-partitioned depends on
421 * migration, so migrate_disable() is not only a problem for
422 * work-conserving schedulers.
423 *
424 */
425
426 /**
427 * preempt_disable_nested - Disable preemption inside a normally preempt disabled section
428 *
429 * Use for code which requires preemption protection inside a critical
430 * section which has preemption disabled implicitly on non-PREEMPT_RT
431 * enabled kernels, by e.g.:
432 * - holding a spinlock/rwlock
433 * - soft interrupt context
434 * - regular interrupt handlers
435 *
436 * On PREEMPT_RT enabled kernels spinlock/rwlock held sections, soft
437 * interrupt context and regular interrupt handlers are preemptible and
438 * only prevent migration. preempt_disable_nested() ensures that preemption
439 * is disabled for cases which require CPU local serialization even on
440 * PREEMPT_RT. For non-PREEMPT_RT kernels this is a NOP.
441 *
442 * The use cases are code sequences which are not serialized by a
443 * particular lock instance, e.g.:
444 * - seqcount write side critical sections where the seqcount is not
445 * associated to a particular lock and therefore the automatic
446 * protection mechanism does not work. This prevents a live lock
447 * against a preempting high priority reader.
448 * - RMW per CPU variable updates like vmstat.
449 */
450 /* Macro to avoid header recursion hell vs. lockdep */
451 #define preempt_disable_nested() \
452 do { \
453 if (IS_ENABLED(CONFIG_PREEMPT_RT)) \
454 preempt_disable(); \
455 else \
456 lockdep_assert_preemption_disabled(); \
457 } while (0)
458
459 /**
460 * preempt_enable_nested - Undo the effect of preempt_disable_nested()
461 */
preempt_enable_nested(void)462 static __always_inline void preempt_enable_nested(void)
463 {
464 if (IS_ENABLED(CONFIG_PREEMPT_RT))
465 preempt_enable();
466 }
467
468 DEFINE_LOCK_GUARD_0(preempt, preempt_disable(), preempt_enable())
469 DEFINE_LOCK_GUARD_0(preempt_notrace, preempt_disable_notrace(), preempt_enable_notrace())
470
471 #ifdef CONFIG_PREEMPT_DYNAMIC
472
473 extern bool preempt_model_none(void);
474 extern bool preempt_model_voluntary(void);
475 extern bool preempt_model_full(void);
476 extern bool preempt_model_lazy(void);
477
478 #else
479
480 static inline bool preempt_model_none(void)
481 {
482 return IS_ENABLED(CONFIG_PREEMPT_NONE);
483 }
484 static inline bool preempt_model_voluntary(void)
485 {
486 return IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY);
487 }
488 static inline bool preempt_model_full(void)
489 {
490 return IS_ENABLED(CONFIG_PREEMPT);
491 }
492
493 static inline bool preempt_model_lazy(void)
494 {
495 return IS_ENABLED(CONFIG_PREEMPT_LAZY);
496 }
497
498 #endif
499
preempt_model_rt(void)500 static inline bool preempt_model_rt(void)
501 {
502 return IS_ENABLED(CONFIG_PREEMPT_RT);
503 }
504
505 extern const char *preempt_model_str(void);
506
507 /*
508 * Does the preemption model allow non-cooperative preemption?
509 *
510 * For !CONFIG_PREEMPT_DYNAMIC kernels this is an exact match with
511 * CONFIG_PREEMPTION; for CONFIG_PREEMPT_DYNAMIC this doesn't work as the
512 * kernel is *built* with CONFIG_PREEMPTION=y but may run with e.g. the
513 * PREEMPT_NONE model.
514 */
preempt_model_preemptible(void)515 static inline bool preempt_model_preemptible(void)
516 {
517 return preempt_model_full() || preempt_model_lazy() || preempt_model_rt();
518 }
519
520 #endif /* __LINUX_PREEMPT_H */
521