1 // SPDX-License-Identifier: GPL-2.0-only 1
2 /*
3 * Context tracking: Probe on high level conte
4 * userspace, guest or idle.
5 *
6 * This is used by RCU to remove its dependenc
7 * runs in idle, userspace or guest mode.
8 *
9 * User/guest tracking started by Frederic Wei
10 *
11 * Copyright (C) 2012 Red Hat, Inc., Frederic
12 *
13 * Many thanks to Gilad Ben-Yossef, Paul McKen
14 * Steven Rostedt, Peter Zijlstra for suggesti
15 *
16 * RCU extended quiescent state bits imported
17 * where the relevant authorship may be found.
18 */
19
20 #include <linux/context_tracking.h>
21 #include <linux/rcupdate.h>
22 #include <linux/sched.h>
23 #include <linux/hardirq.h>
24 #include <linux/export.h>
25 #include <linux/kprobes.h>
26 #include <trace/events/rcu.h>
27
28
29 DEFINE_PER_CPU(struct context_tracking, contex
30 #ifdef CONFIG_CONTEXT_TRACKING_IDLE
31 .dynticks_nesting = 1,
32 .dynticks_nmi_nesting = DYNTICK_IRQ_NO
33 #endif
34 .state = ATOMIC_INIT(RCU_DYNTICKS_IDX)
35 };
36 EXPORT_SYMBOL_GPL(context_tracking);
37
38 #ifdef CONFIG_CONTEXT_TRACKING_IDLE
39 #define TPS(x) tracepoint_string(x)
40
41 /* Record the current task on dyntick-idle ent
42 static __always_inline void rcu_dynticks_task_
43 {
44 #if defined(CONFIG_TASKS_RCU) && defined(CONFI
45 WRITE_ONCE(current->rcu_tasks_idle_cpu
46 #endif /* #if defined(CONFIG_TASKS_RCU) && def
47 }
48
49 /* Record no current task on dyntick-idle exit
50 static __always_inline void rcu_dynticks_task_
51 {
52 #if defined(CONFIG_TASKS_RCU) && defined(CONFI
53 WRITE_ONCE(current->rcu_tasks_idle_cpu
54 #endif /* #if defined(CONFIG_TASKS_RCU) && def
55 }
56
57 /* Turn on heavyweight RCU tasks trace readers
58 static __always_inline void rcu_dynticks_task_
59 {
60 #ifdef CONFIG_TASKS_TRACE_RCU
61 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_
62 current->trc_reader_special.b.
63 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
64 }
65
66 /* Turn off heavyweight RCU tasks trace reader
67 static __always_inline void rcu_dynticks_task_
68 {
69 #ifdef CONFIG_TASKS_TRACE_RCU
70 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_
71 current->trc_reader_special.b.
72 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
73 }
74
75 /*
76 * Record entry into an extended quiescent sta
77 * called when not already in an extended quie
78 * RCU is watching prior to the call to this f
79 * watching upon return.
80 */
81 static noinstr void ct_kernel_exit_state(int o
82 {
83 int seq;
84
85 /*
86 * CPUs seeing atomic_add_return() mus
87 * critical sections, and we also must
88 * next idle sojourn.
89 */
90 rcu_dynticks_task_trace_enter(); // B
91 seq = ct_state_inc(offset);
92 // RCU is no longer watching. Better
93 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
94 }
95
96 /*
97 * Record exit from an extended quiescent stat
98 * called from an extended quiescent state, th
99 * prior to the call to this function and is w
100 */
101 static noinstr void ct_kernel_enter_state(int
102 {
103 int seq;
104
105 /*
106 * CPUs seeing atomic_add_return() mus
107 * and we also must force ordering wit
108 * critical section.
109 */
110 seq = ct_state_inc(offset);
111 // RCU is now watching. Better not be
112 rcu_dynticks_task_trace_exit(); // Af
113 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
114 }
115
116 /*
117 * Enter an RCU extended quiescent state, whic
118 * idle loop or adaptive-tickless usermode exe
119 *
120 * We crowbar the ->dynticks_nmi_nesting field
121 * the possibility of usermode upcalls having
122 * of interrupt nesting level during the prior
123 */
124 static void noinstr ct_kernel_exit(bool user,
125 {
126 struct context_tracking *ct = this_cpu
127
128 WARN_ON_ONCE(ct_dynticks_nmi_nesting()
129 WRITE_ONCE(ct->dynticks_nmi_nesting, 0
130 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
131 ct_dynticks_nesting() ==
132 if (ct_dynticks_nesting() != 1) {
133 // RCU will still be watching,
134 ct->dynticks_nesting--;
135 return;
136 }
137
138 instrumentation_begin();
139 lockdep_assert_irqs_disabled();
140 trace_rcu_dyntick(TPS("Start"), ct_dyn
141 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
142 rcu_preempt_deferred_qs(current);
143
144 // instrumentation for the noinstr ct_
145 instrument_atomic_write(&ct->state, si
146
147 instrumentation_end();
148 WRITE_ONCE(ct->dynticks_nesting, 0); /
149 // RCU is watching here ...
150 ct_kernel_exit_state(offset);
151 // ... but is no longer watching here.
152 rcu_dynticks_task_enter();
153 }
154
155 /*
156 * Exit an RCU extended quiescent state, which
157 * idle loop or adaptive-tickless usermode exe
158 *
159 * We crowbar the ->dynticks_nmi_nesting field
160 * allow for the possibility of usermode upcal
161 * interrupt nesting level during the busy per
162 */
163 static void noinstr ct_kernel_enter(bool user,
164 {
165 struct context_tracking *ct = this_cpu
166 long oldval;
167
168 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
169 oldval = ct_dynticks_nesting();
170 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
171 if (oldval) {
172 // RCU was already watching, s
173 ct->dynticks_nesting++;
174 return;
175 }
176 rcu_dynticks_task_exit();
177 // RCU is not watching here ...
178 ct_kernel_enter_state(offset);
179 // ... but is watching here.
180 instrumentation_begin();
181
182 // instrumentation for the noinstr ct_
183 instrument_atomic_write(&ct->state, si
184
185 trace_rcu_dyntick(TPS("End"), ct_dynti
186 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
187 WRITE_ONCE(ct->dynticks_nesting, 1);
188 WARN_ON_ONCE(ct_dynticks_nmi_nesting()
189 WRITE_ONCE(ct->dynticks_nmi_nesting, D
190 instrumentation_end();
191 }
192
193 /**
194 * ct_nmi_exit - inform RCU of exit from NMI c
195 *
196 * If we are returning from the outermost NMI
197 * RCU-idle period, update ct->state and ct->d
198 * to let the RCU grace-period handling know t
199 * being RCU-idle.
200 *
201 * If you add or remove a call to ct_nmi_exit(
202 * with CONFIG_RCU_EQS_DEBUG=y.
203 */
204 void noinstr ct_nmi_exit(void)
205 {
206 struct context_tracking *ct = this_cpu
207
208 instrumentation_begin();
209 /*
210 * Check for ->dynticks_nmi_nesting un
211 * (We are exiting an NMI handler, so
212 * to us!)
213 */
214 WARN_ON_ONCE(ct_dynticks_nmi_nesting()
215 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_
216
217 /*
218 * If the nesting level is not 1, the
219 * leave it in non-RCU-idle state.
220 */
221 if (ct_dynticks_nmi_nesting() != 1) {
222 trace_rcu_dyntick(TPS("--="),
223 ct_dynticks(
224 WRITE_ONCE(ct->dynticks_nmi_ne
225 ct_dynticks_nmi_nes
226 instrumentation_end();
227 return;
228 }
229
230 /* This NMI interrupted an RCU-idle CP
231 trace_rcu_dyntick(TPS("Startirq"), ct_
232 WRITE_ONCE(ct->dynticks_nmi_nesting, 0
233
234 // instrumentation for the noinstr ct_
235 instrument_atomic_write(&ct->state, si
236 instrumentation_end();
237
238 // RCU is watching here ...
239 ct_kernel_exit_state(RCU_DYNTICKS_IDX)
240 // ... but is no longer watching here.
241
242 if (!in_nmi())
243 rcu_dynticks_task_enter();
244 }
245
246 /**
247 * ct_nmi_enter - inform RCU of entry to NMI c
248 *
249 * If the CPU was idle from RCU's viewpoint, u
250 * ct->dynticks_nmi_nesting to let the RCU gra
251 * that the CPU is active. This implementatio
252 * long as the nesting level does not overflow
253 * run out of stack space first.)
254 *
255 * If you add or remove a call to ct_nmi_enter
256 * with CONFIG_RCU_EQS_DEBUG=y.
257 */
258 void noinstr ct_nmi_enter(void)
259 {
260 long incby = 2;
261 struct context_tracking *ct = this_cpu
262
263 /* Complain about underflow. */
264 WARN_ON_ONCE(ct_dynticks_nmi_nesting()
265
266 /*
267 * If idle from RCU viewpoint, atomica
268 * to mark non-idle and increment ->dy
269 * Otherwise, increment ->dynticks_nmi
270 * if ->dynticks_nmi_nesting is equal
271 * to be in the outermost NMI handler
272 * period (observation due to Andy Lut
273 */
274 if (rcu_dynticks_curr_cpu_in_eqs()) {
275
276 if (!in_nmi())
277 rcu_dynticks_task_exit
278
279 // RCU is not watching here ..
280 ct_kernel_enter_state(RCU_DYNT
281 // ... but is watching here.
282
283 instrumentation_begin();
284 // instrumentation for the noi
285 instrument_atomic_read(&ct->st
286 // instrumentation for the noi
287 instrument_atomic_write(&ct->s
288
289 incby = 1;
290 } else if (!in_nmi()) {
291 instrumentation_begin();
292 rcu_irq_enter_check_tick();
293 } else {
294 instrumentation_begin();
295 }
296
297 trace_rcu_dyntick(incby == 1 ? TPS("En
298 ct_dynticks_nmi_nest
299 ct_dynticks_nmi_nest
300 instrumentation_end();
301 WRITE_ONCE(ct->dynticks_nmi_nesting, /
302 ct_dynticks_nmi_nesting() +
303 barrier();
304 }
305
306 /**
307 * ct_idle_enter - inform RCU that current CPU
308 *
309 * Enter idle mode, in other words, -leave- th
310 * read-side critical sections can occur. (Th
311 * critical sections can occur in irq handlers
312 * handled by irq_enter() and irq_exit().)
313 *
314 * If you add or remove a call to ct_idle_ente
315 * CONFIG_RCU_EQS_DEBUG=y.
316 */
317 void noinstr ct_idle_enter(void)
318 {
319 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS
320 ct_kernel_exit(false, RCU_DYNTICKS_IDX
321 }
322 EXPORT_SYMBOL_GPL(ct_idle_enter);
323
324 /**
325 * ct_idle_exit - inform RCU that current CPU
326 *
327 * Exit idle mode, in other words, -enter- the
328 * read-side critical sections can occur.
329 *
330 * If you add or remove a call to ct_idle_exit
331 * CONFIG_RCU_EQS_DEBUG=y.
332 */
333 void noinstr ct_idle_exit(void)
334 {
335 unsigned long flags;
336
337 raw_local_irq_save(flags);
338 ct_kernel_enter(false, RCU_DYNTICKS_ID
339 raw_local_irq_restore(flags);
340 }
341 EXPORT_SYMBOL_GPL(ct_idle_exit);
342
343 /**
344 * ct_irq_enter - inform RCU that current CPU
345 *
346 * Enter an interrupt handler, which might pos
347 * idle mode, in other words, entering the mod
348 * sections can occur. The caller must have d
349 *
350 * Note that the Linux kernel is fully capable
351 * handler that it never exits, for example wh
352 * This code assumes that the idle loop never
353 * If your architecture's idle loop does do up
354 * anything else that results in unbalanced ca
355 * irq_exit() functions), RCU will give you wh
356 * But very infrequently and irreproducibly.
357 *
358 * Use things like work queues to work around
359 *
360 * You have been warned.
361 *
362 * If you add or remove a call to ct_irq_enter
363 * CONFIG_RCU_EQS_DEBUG=y.
364 */
365 noinstr void ct_irq_enter(void)
366 {
367 lockdep_assert_irqs_disabled();
368 ct_nmi_enter();
369 }
370
371 /**
372 * ct_irq_exit - inform RCU that current CPU i
373 *
374 * Exit from an interrupt handler, which might
375 * idle mode, in other words, leaving the mode
376 * sections can occur. The caller must have d
377 *
378 * This code assumes that the idle loop never
379 * result in unbalanced calls to irq_enter() a
380 * architecture's idle loop violates this assu
381 * you deserve, good and hard. But very infre
382 *
383 * Use things like work queues to work around
384 *
385 * You have been warned.
386 *
387 * If you add or remove a call to ct_irq_exit(
388 * CONFIG_RCU_EQS_DEBUG=y.
389 */
390 noinstr void ct_irq_exit(void)
391 {
392 lockdep_assert_irqs_disabled();
393 ct_nmi_exit();
394 }
395
396 /*
397 * Wrapper for ct_irq_enter() where interrupts
398 *
399 * If you add or remove a call to ct_irq_enter
400 * with CONFIG_RCU_EQS_DEBUG=y.
401 */
402 void ct_irq_enter_irqson(void)
403 {
404 unsigned long flags;
405
406 local_irq_save(flags);
407 ct_irq_enter();
408 local_irq_restore(flags);
409 }
410
411 /*
412 * Wrapper for ct_irq_exit() where interrupts
413 *
414 * If you add or remove a call to ct_irq_exit_
415 * with CONFIG_RCU_EQS_DEBUG=y.
416 */
417 void ct_irq_exit_irqson(void)
418 {
419 unsigned long flags;
420
421 local_irq_save(flags);
422 ct_irq_exit();
423 local_irq_restore(flags);
424 }
425 #else
426 static __always_inline void ct_kernel_exit(boo
427 static __always_inline void ct_kernel_enter(bo
428 #endif /* #ifdef CONFIG_CONTEXT_TRACKING_IDLE
429
430 #ifdef CONFIG_CONTEXT_TRACKING_USER
431
432 #define CREATE_TRACE_POINTS
433 #include <trace/events/context_tracking.h>
434
435 DEFINE_STATIC_KEY_FALSE_RO(context_tracking_ke
436 EXPORT_SYMBOL_GPL(context_tracking_key);
437
438 static noinstr bool context_tracking_recursion
439 {
440 int recursion;
441
442 recursion = __this_cpu_inc_return(cont
443 if (recursion == 1)
444 return true;
445
446 WARN_ONCE((recursion < 1), "Invalid co
447 __this_cpu_dec(context_tracking.recurs
448
449 return false;
450 }
451
452 static __always_inline void context_tracking_r
453 {
454 __this_cpu_dec(context_tracking.recurs
455 }
456
457 /**
458 * __ct_user_enter - Inform the context tracki
459 * to enter user or guest sp
460 *
461 * @state: userspace context-tracking state to
462 *
463 * This function must be called right before w
464 * to user or guest space, when it's guarantee
465 * instructions to execute won't use any RCU r
466 * because this function sets RCU in extended
467 */
468 void noinstr __ct_user_enter(enum ctx_state st
469 {
470 struct context_tracking *ct = this_cpu
471 lockdep_assert_irqs_disabled();
472
473 /* Kernel threads aren't supposed to g
474 WARN_ON_ONCE(!current->mm);
475
476 if (!context_tracking_recursion_enter(
477 return;
478
479 if (__ct_state() != state) {
480 if (ct->active) {
481 /*
482 * At this stage, only
483 * then we'll run in u
484 * any RCU read-side c
485 * user_exit() or ct_i
486 * on the tick.
487 */
488 if (state == CONTEXT_U
489 instrumentatio
490 trace_user_ent
491 vtime_user_ent
492 instrumentatio
493 }
494 /*
495 * Other than generic
496 * rescheduling opport
497 * that will fire and
498 */
499 rcu_irq_work_resched()
500
501 /*
502 * Enter RCU idle mode
503 * is permitted betwee
504 * CPU doesn't need to
505 * when the CPU runs i
506 */
507 ct_kernel_exit(true, R
508
509 /*
510 * Special case if we
511 * cputime accounting
512 * In this we case we
513 */
514 if (!IS_ENABLED(CONFIG
515 raw_atomic_set
516 } else {
517 /*
518 * Even if context tra
519 * the full dynticks m
520 * context transitions
521 * other CPUs.
522 * If a task triggers
523 * handler and then mi
524 * the exception retur
525 * This information ca
526 * exception_enter().
527 * OTOH we can spare t
528 * is false because we
529 */
530 if (!IS_ENABLED(CONFIG
531 /* Tracking fo
532 raw_atomic_set
533 } else {
534 /*
535 * Tracking fo
536 * with NMIs.
537 * RCU only re
538 * ordered.
539 */
540 raw_atomic_add
541 }
542 }
543 }
544 context_tracking_recursion_exit();
545 }
546 EXPORT_SYMBOL_GPL(__ct_user_enter);
547
548 /*
549 * OBSOLETE:
550 * This function should be noinstr but the bel
551 * unsafe because it involves illegal RCU uses
552 * This is unlikely to be fixed as this functi
553 * way is to call __context_tracking_enter() t
554 * or context_tracking_guest_enter(). It shoul
555 * responsibility to call into context trackin
556 */
557 void ct_user_enter(enum ctx_state state)
558 {
559 unsigned long flags;
560
561 /*
562 * Some contexts may involve an except
563 * leading to that nesting:
564 * ct_irq_enter() rcu_eqs_exit(true) r
565 * This would mess up the dyntick_nest
566 * helpers are enough to protect RCU u
567 * just return immediately if we detec
568 */
569 if (in_interrupt())
570 return;
571
572 local_irq_save(flags);
573 __ct_user_enter(state);
574 local_irq_restore(flags);
575 }
576 NOKPROBE_SYMBOL(ct_user_enter);
577 EXPORT_SYMBOL_GPL(ct_user_enter);
578
579 /**
580 * user_enter_callable() - Unfortunate ASM cal
581 * archs that didn't m
582 * static key from low
583 *
584 * This OBSOLETE function should be noinstr bu
585 * local_irq_restore(), involving illegal RCU
586 * This is unlikely to be fixed as this functi
587 * way is to call user_enter_irqoff(). It shou
588 * responsibility to call into context trackin
589 */
590 void user_enter_callable(void)
591 {
592 user_enter();
593 }
594 NOKPROBE_SYMBOL(user_enter_callable);
595
596 /**
597 * __ct_user_exit - Inform the context trackin
598 * exiting user or guest mode
599 *
600 * @state: userspace context-tracking state be
601 *
602 * This function must be called after we enter
603 * guest space before any use of RCU read side
604 * potentially include any high level kernel c
605 * signal handling, etc...
606 *
607 * This call supports re-entrancy. This way it
608 * handler without needing to know if we came
609 */
610 void noinstr __ct_user_exit(enum ctx_state sta
611 {
612 struct context_tracking *ct = this_cpu
613
614 if (!context_tracking_recursion_enter(
615 return;
616
617 if (__ct_state() == state) {
618 if (ct->active) {
619 /*
620 * Exit RCU idle mode
621 * run a RCU read side
622 */
623 ct_kernel_enter(true,
624 if (state == CONTEXT_U
625 instrumentatio
626 vtime_user_exi
627 trace_user_exi
628 instrumentatio
629 }
630
631 /*
632 * Special case if we
633 * cputime accounting
634 * In this we case we
635 */
636 if (!IS_ENABLED(CONFIG
637 raw_atomic_set
638
639 } else {
640 if (!IS_ENABLED(CONFIG
641 /* Tracking fo
642 raw_atomic_set
643 } else {
644 /*
645 * Tracking fo
646 * with NMIs.
647 * RCU only re
648 * ordered.
649 */
650 raw_atomic_sub
651 }
652 }
653 }
654 context_tracking_recursion_exit();
655 }
656 EXPORT_SYMBOL_GPL(__ct_user_exit);
657
658 /*
659 * OBSOLETE:
660 * This function should be noinstr but the bel
661 * unsafe because it involves illegal RCU uses
662 * This is unlikely to be fixed as this functi
663 * way is to call __context_tracking_exit() th
664 * or context_tracking_guest_exit(). It should
665 * responsibility to call into context trackin
666 */
667 void ct_user_exit(enum ctx_state state)
668 {
669 unsigned long flags;
670
671 if (in_interrupt())
672 return;
673
674 local_irq_save(flags);
675 __ct_user_exit(state);
676 local_irq_restore(flags);
677 }
678 NOKPROBE_SYMBOL(ct_user_exit);
679 EXPORT_SYMBOL_GPL(ct_user_exit);
680
681 /**
682 * user_exit_callable() - Unfortunate ASM call
683 * archs that didn't ma
684 * static key from low
685 *
686 * This OBSOLETE function should be noinstr bu
687 * involving illegal RCU uses through tracing
688 * to be fixed as this function is obsolete. T
689 * user_exit_irqoff(). It should be the arch e
690 * call into context tracking with IRQs disabl
691 */
692 void user_exit_callable(void)
693 {
694 user_exit();
695 }
696 NOKPROBE_SYMBOL(user_exit_callable);
697
698 void __init ct_cpu_track_user(int cpu)
699 {
700 static __initdata bool initialized = f
701
702 if (!per_cpu(context_tracking.active,
703 per_cpu(context_tracking.activ
704 static_branch_inc(&context_tra
705 }
706
707 if (initialized)
708 return;
709
710 #ifdef CONFIG_HAVE_TIF_NOHZ
711 /*
712 * Set TIF_NOHZ to init/0 and let it p
713 * This assumes that init is the only
714 */
715 set_tsk_thread_flag(&init_task, TIF_NO
716 #endif
717 WARN_ON_ONCE(!tasklist_empty());
718
719 initialized = true;
720 }
721
722 #ifdef CONFIG_CONTEXT_TRACKING_USER_FORCE
723 void __init context_tracking_init(void)
724 {
725 int cpu;
726
727 for_each_possible_cpu(cpu)
728 ct_cpu_track_user(cpu);
729 }
730 #endif
731
732 #endif /* #ifdef CONFIG_CONTEXT_TRACKING_USER
733
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.