projects/linux·blob v6.1 · refs

466 lines · 14.0 KiB

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 */
89static __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#define nmi_count() (preempt_count() & NMI_MASK)
102#define hardirq_count() (preempt_count() & HARDIRQ_MASK)
103#ifdef CONFIG_PREEMPT_RT
104# define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
105#else
106# define softirq_count() (preempt_count() & SOFTIRQ_MASK)
107#endif
108#define irq_count() (nmi_count() | hardirq_count() | softirq_count())
109
110/*
111 * Macros to retrieve the current execution context:
112 *
113 * in_nmi() - We're in NMI context
114 * in_hardirq() - We're in hard IRQ context
115 * in_serving_softirq() - We're in softirq context
116 * in_task() - We're in task context
117 */
118#define in_nmi() (nmi_count())
119#define in_hardirq() (hardirq_count())
120#define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
121#define in_task() (!(in_nmi() | in_hardirq() | in_serving_softirq()))
122
123/*
124 * The following macros are deprecated and should not be used in new code:
125 * in_irq() - Obsolete version of in_hardirq()
126 * in_softirq() - We have BH disabled, or are processing softirqs
127 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
128 */
129#define in_irq() (hardirq_count())
130#define in_softirq() (softirq_count())
131#define in_interrupt() (irq_count())
132
133/*
134 * The preempt_count offset after preempt_disable();
135 */
136#if defined(CONFIG_PREEMPT_COUNT)
137# define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET
138#else
139# define PREEMPT_DISABLE_OFFSET 0
140#endif
141
142/*
143 * The preempt_count offset after spin_lock()
144 */
145#if !defined(CONFIG_PREEMPT_RT)
146#define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
147#else
148/* Locks on RT do not disable preemption */
149#define PREEMPT_LOCK_OFFSET 0
150#endif
151
152/*
153 * The preempt_count offset needed for things like:
154 *
155 * spin_lock_bh()
156 *
157 * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and
158 * softirqs, such that unlock sequences of:
159 *
160 * spin_unlock();
161 * local_bh_enable();
162 *
163 * Work as expected.
164 */
165#define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET)
166
167/*
168 * Are we running in atomic context? WARNING: this macro cannot
169 * always detect atomic context; in particular, it cannot know about
170 * held spinlocks in non-preemptible kernels. Thus it should not be
171 * used in the general case to determine whether sleeping is possible.
172 * Do not use in_atomic() in driver code.
173 */
174#define in_atomic() (preempt_count() != 0)
175
176/*
177 * Check whether we were atomic before we did preempt_disable():
178 * (used by the scheduler)
179 */
180#define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET)
181
182#if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
183extern void preempt_count_add(int val);
184extern void preempt_count_sub(int val);
185#define preempt_count_dec_and_test() \
186 ({ preempt_count_sub(1); should_resched(0); })
187#else
188#define preempt_count_add(val) __preempt_count_add(val)
189#define preempt_count_sub(val) __preempt_count_sub(val)
190#define preempt_count_dec_and_test() __preempt_count_dec_and_test()
191#endif
192
193#define __preempt_count_inc() __preempt_count_add(1)
194#define __preempt_count_dec() __preempt_count_sub(1)
195
196#define preempt_count_inc() preempt_count_add(1)
197#define preempt_count_dec() preempt_count_sub(1)
198
199#ifdef CONFIG_PREEMPT_COUNT
200
201#define preempt_disable() \
202do { \
203 preempt_count_inc(); \
204 barrier(); \
205} while (0)
206
207#define sched_preempt_enable_no_resched() \
208do { \
209 barrier(); \
210 preempt_count_dec(); \
211} while (0)
212
213#define preempt_enable_no_resched() sched_preempt_enable_no_resched()
214
215#define preemptible() (preempt_count() == 0 && !irqs_disabled())
216
217#ifdef CONFIG_PREEMPTION
218#define preempt_enable() \
219do { \
220 barrier(); \
221 if (unlikely(preempt_count_dec_and_test())) \
222 __preempt_schedule(); \
223} while (0)
224
225#define preempt_enable_notrace() \
226do { \
227 barrier(); \
228 if (unlikely(__preempt_count_dec_and_test())) \
229 __preempt_schedule_notrace(); \
230} while (0)
231
232#define preempt_check_resched() \
233do { \
234 if (should_resched(0)) \
235 __preempt_schedule(); \
236} while (0)
237
238#else /* !CONFIG_PREEMPTION */
239#define preempt_enable() \
240do { \
241 barrier(); \
242 preempt_count_dec(); \
243} while (0)
244
245#define preempt_enable_notrace() \
246do { \
247 barrier(); \
248 __preempt_count_dec(); \
249} while (0)
250
251#define preempt_check_resched() do { } while (0)
252#endif /* CONFIG_PREEMPTION */
253
254#define preempt_disable_notrace() \
255do { \
256 __preempt_count_inc(); \
257 barrier(); \
258} while (0)
259
260#define preempt_enable_no_resched_notrace() \
261do { \
262 barrier(); \
263 __preempt_count_dec(); \
264} while (0)
265
266#else /* !CONFIG_PREEMPT_COUNT */
267
268/*
269 * Even if we don't have any preemption, we need preempt disable/enable
270 * to be barriers, so that we don't have things like get_user/put_user
271 * that can cause faults and scheduling migrate into our preempt-protected
272 * region.
273 */
274#define preempt_disable() barrier()
275#define sched_preempt_enable_no_resched() barrier()
276#define preempt_enable_no_resched() barrier()
277#define preempt_enable() barrier()
278#define preempt_check_resched() do { } while (0)
279
280#define preempt_disable_notrace() barrier()
281#define preempt_enable_no_resched_notrace() barrier()
282#define preempt_enable_notrace() barrier()
283#define preemptible() 0
284
285#endif /* CONFIG_PREEMPT_COUNT */
286
287#ifdef MODULE
288/*
289 * Modules have no business playing preemption tricks.
290 */
291#undef sched_preempt_enable_no_resched
292#undef preempt_enable_no_resched
293#undef preempt_enable_no_resched_notrace
294#undef preempt_check_resched
295#endif
296
297#define preempt_set_need_resched() \
298do { \
299 set_preempt_need_resched(); \
300} while (0)
301#define preempt_fold_need_resched() \
302do { \
303 if (tif_need_resched()) \
304 set_preempt_need_resched(); \
305} while (0)
306
307#ifdef CONFIG_PREEMPT_NOTIFIERS
308
309struct preempt_notifier;
310
311/**
312 * preempt_ops - notifiers called when a task is preempted and rescheduled
313 * @sched_in: we're about to be rescheduled:
314 * notifier: struct preempt_notifier for the task being scheduled
315 * cpu: cpu we're scheduled on
316 * @sched_out: we've just been preempted
317 * notifier: struct preempt_notifier for the task being preempted
318 * next: the task that's kicking us out
319 *
320 * Please note that sched_in and out are called under different
321 * contexts. sched_out is called with rq lock held and irq disabled
322 * while sched_in is called without rq lock and irq enabled. This
323 * difference is intentional and depended upon by its users.
324 */
325struct preempt_ops {
326 void (*sched_in)(struct preempt_notifier *notifier, int cpu);
327 void (*sched_out)(struct preempt_notifier *notifier,
328 struct task_struct *next);
329};
330
331/**
332 * preempt_notifier - key for installing preemption notifiers
333 * @link: internal use
334 * @ops: defines the notifier functions to be called
335 *
336 * Usually used in conjunction with container_of().
337 */
338struct preempt_notifier {
339 struct hlist_node link;
340 struct preempt_ops *ops;
341};
342
343void preempt_notifier_inc(void);
344void preempt_notifier_dec(void);
345void preempt_notifier_register(struct preempt_notifier *notifier);
346void preempt_notifier_unregister(struct preempt_notifier *notifier);
347
348static inline void preempt_notifier_init(struct preempt_notifier *notifier,
349 struct preempt_ops *ops)
350{
351 INIT_HLIST_NODE(&notifier->link);
352 notifier->ops = ops;
353}
354
355#endif
356
357#ifdef CONFIG_SMP
358
359/*
360 * Migrate-Disable and why it is undesired.
361 *
362 * When a preempted task becomes elegible to run under the ideal model (IOW it
363 * becomes one of the M highest priority tasks), it might still have to wait
364 * for the preemptee's migrate_disable() section to complete. Thereby suffering
365 * a reduction in bandwidth in the exact duration of the migrate_disable()
366 * section.
367 *
368 * Per this argument, the change from preempt_disable() to migrate_disable()
369 * gets us:
370 *
371 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
372 * it would have had to wait for the lower priority task.
373 *
374 * - a lower priority tasks; which under preempt_disable() could've instantly
375 * migrated away when another CPU becomes available, is now constrained
376 * by the ability to push the higher priority task away, which might itself be
377 * in a migrate_disable() section, reducing it's available bandwidth.
378 *
379 * IOW it trades latency / moves the interference term, but it stays in the
380 * system, and as long as it remains unbounded, the system is not fully
381 * deterministic.
382 *
383 *
384 * The reason we have it anyway.
385 *
386 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
387 * number of primitives into becoming preemptible, they would also allow
388 * migration. This turns out to break a bunch of per-cpu usage. To this end,
389 * all these primitives employ migirate_disable() to restore this implicit
390 * assumption.
391 *
392 * This is a 'temporary' work-around at best. The correct solution is getting
393 * rid of the above assumptions and reworking the code to employ explicit
394 * per-cpu locking or short preempt-disable regions.
395 *
396 * The end goal must be to get rid of migrate_disable(), alternatively we need
397 * a schedulability theory that does not depend on abritrary migration.
398 *
399 *
400 * Notes on the implementation.
401 *
402 * The implementation is particularly tricky since existing code patterns
403 * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
404 * This means that it cannot use cpus_read_lock() to serialize against hotplug,
405 * nor can it easily migrate itself into a pending affinity mask change on
406 * migrate_enable().
407 *
408 *
409 * Note: even non-work-conserving schedulers like semi-partitioned depends on
410 * migration, so migrate_disable() is not only a problem for
411 * work-conserving schedulers.
412 *
413 */
414extern void migrate_disable(void);
415extern void migrate_enable(void);
416
417#else
418
419static inline void migrate_disable(void) { }
420static inline void migrate_enable(void) { }
421
422#endif /* CONFIG_SMP */
423
424/**
425 * preempt_disable_nested - Disable preemption inside a normally preempt disabled section
426 *
427 * Use for code which requires preemption protection inside a critical
428 * section which has preemption disabled implicitly on non-PREEMPT_RT
429 * enabled kernels, by e.g.:
430 * - holding a spinlock/rwlock
431 * - soft interrupt context
432 * - regular interrupt handlers
433 *
434 * On PREEMPT_RT enabled kernels spinlock/rwlock held sections, soft
435 * interrupt context and regular interrupt handlers are preemptible and
436 * only prevent migration. preempt_disable_nested() ensures that preemption
437 * is disabled for cases which require CPU local serialization even on
438 * PREEMPT_RT. For non-PREEMPT_RT kernels this is a NOP.
439 *
440 * The use cases are code sequences which are not serialized by a
441 * particular lock instance, e.g.:
442 * - seqcount write side critical sections where the seqcount is not
443 * associated to a particular lock and therefore the automatic
444 * protection mechanism does not work. This prevents a live lock
445 * against a preempting high priority reader.
446 * - RMW per CPU variable updates like vmstat.
447 */
448/* Macro to avoid header recursion hell vs. lockdep */
449#define preempt_disable_nested() \
450do { \
451 if (IS_ENABLED(CONFIG_PREEMPT_RT)) \
452 preempt_disable(); \
453 else \
454 lockdep_assert_preemption_disabled(); \
455} while (0)
456
457/**
458 * preempt_enable_nested - Undo the effect of preempt_disable_nested()
459 */
460static __always_inline void preempt_enable_nested(void)
461{
462 if (IS_ENABLED(CONFIG_PREEMPT_RT))
463 preempt_enable();
464}
465
466#endif /* __LINUX_PREEMPT_H */