Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Context tracking: Probe on high level context boundaries such as kernel,
4 * userspace, guest or idle.
5 *
6 * This is used by RCU to remove its dependency on the timer tick while a CPU
7 * runs in idle, userspace or guest mode.
8 *
9 * User/guest tracking started by Frederic Weisbecker:
10 *
11 * Copyright (C) 2012 Red Hat, Inc., Frederic Weisbecker
12 *
13 * Many thanks to Gilad Ben-Yossef, Paul McKenney, Ingo Molnar, Andrew Morton,
14 * Steven Rostedt, Peter Zijlstra for suggestions and improvements.
15 *
16 * RCU extended quiescent state bits imported from kernel/rcu/tree.c
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
29DEFINE_PER_CPU(struct context_tracking, context_tracking) = {
30#ifdef CONFIG_CONTEXT_TRACKING_IDLE
31 .nesting = 1,
32 .nmi_nesting = CT_NESTING_IRQ_NONIDLE,
33#endif
34 .state = ATOMIC_INIT(CT_RCU_WATCHING),
35};
36EXPORT_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 exiting RCU-tasks (dyntick-idle entry). */
42static __always_inline void rcu_task_exit(void)
43{
44#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
45 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
46#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
47}
48
49/* Record no current task on entering RCU-tasks (dyntick-idle exit). */
50static __always_inline void rcu_task_enter(void)
51{
52#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
53 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
54#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
55}
56
57/* Turn on heavyweight RCU tasks trace readers on kernel exit. */
58static __always_inline void rcu_task_trace_heavyweight_enter(void)
59{
60#ifdef CONFIG_TASKS_TRACE_RCU
61 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
62 current->trc_reader_special.b.need_mb = true;
63#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
64}
65
66/* Turn off heavyweight RCU tasks trace readers on kernel entry. */
67static __always_inline void rcu_task_trace_heavyweight_exit(void)
68{
69#ifdef CONFIG_TASKS_TRACE_RCU
70 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
71 current->trc_reader_special.b.need_mb = false;
72#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
73}
74
75/*
76 * Record entry into an extended quiescent state. This is only to be
77 * called when not already in an extended quiescent state, that is,
78 * RCU is watching prior to the call to this function and is no longer
79 * watching upon return.
80 */
81static noinstr void ct_kernel_exit_state(int offset)
82{
83 int seq;
84
85 /*
86 * CPUs seeing atomic_add_return() must see prior RCU read-side
87 * critical sections, and we also must force ordering with the
88 * next idle sojourn.
89 */
90 rcu_task_trace_heavyweight_enter(); // Before CT state update!
91 seq = ct_state_inc(offset);
92 // RCU is no longer watching. Better be in extended quiescent state!
93 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && (seq & CT_RCU_WATCHING));
94}
95
96/*
97 * Record exit from an extended quiescent state. This is only to be
98 * called from an extended quiescent state, that is, RCU is not watching
99 * prior to the call to this function and is watching upon return.
100 */
101static noinstr void ct_kernel_enter_state(int offset)
102{
103 int seq;
104
105 /*
106 * CPUs seeing atomic_add_return() must see prior idle sojourns,
107 * and we also must force ordering with the next RCU read-side
108 * critical section.
109 */
110 seq = ct_state_inc(offset);
111 // RCU is now watching. Better not be in an extended quiescent state!
112 rcu_task_trace_heavyweight_exit(); // After CT state update!
113 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !(seq & CT_RCU_WATCHING));
114}
115
116/*
117 * Enter an RCU extended quiescent state, which can be either the
118 * idle loop or adaptive-tickless usermode execution.
119 *
120 * We crowbar the ->nmi_nesting field to zero to allow for
121 * the possibility of usermode upcalls having messed up our count
122 * of interrupt nesting level during the prior busy period.
123 */
124static void noinstr ct_kernel_exit(bool user, int offset)
125{
126 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
127
128 WARN_ON_ONCE(ct_nmi_nesting() != CT_NESTING_IRQ_NONIDLE);
129 WRITE_ONCE(ct->nmi_nesting, 0);
130 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
131 ct_nesting() == 0);
132 if (ct_nesting() != 1) {
133 // RCU will still be watching, so just do accounting and leave.
134 ct->nesting--;
135 return;
136 }
137
138 instrumentation_begin();
139 lockdep_assert_irqs_disabled();
140 trace_rcu_watching(TPS("End"), ct_nesting(), 0, ct_rcu_watching());
141 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
142 rcu_preempt_deferred_qs(current);
143
144 // instrumentation for the noinstr ct_kernel_exit_state()
145 instrument_atomic_write(&ct->state, sizeof(ct->state));
146
147 instrumentation_end();
148 WRITE_ONCE(ct->nesting, 0); /* Avoid irq-access tearing. */
149 // RCU is watching here ...
150 ct_kernel_exit_state(offset);
151 // ... but is no longer watching here.
152 rcu_task_exit();
153}
154
155/*
156 * Exit an RCU extended quiescent state, which can be either the
157 * idle loop or adaptive-tickless usermode execution.
158 *
159 * We crowbar the ->nmi_nesting field to CT_NESTING_IRQ_NONIDLE to
160 * allow for the possibility of usermode upcalls messing up our count of
161 * interrupt nesting level during the busy period that is just now starting.
162 */
163static void noinstr ct_kernel_enter(bool user, int offset)
164{
165 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
166 long oldval;
167
168 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
169 oldval = ct_nesting();
170 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
171 if (oldval) {
172 // RCU was already watching, so just do accounting and leave.
173 ct->nesting++;
174 return;
175 }
176 rcu_task_enter();
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_kernel_enter_state()
183 instrument_atomic_write(&ct->state, sizeof(ct->state));
184
185 trace_rcu_watching(TPS("Start"), ct_nesting(), 1, ct_rcu_watching());
186 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
187 WRITE_ONCE(ct->nesting, 1);
188 WARN_ON_ONCE(ct_nmi_nesting());
189 WRITE_ONCE(ct->nmi_nesting, CT_NESTING_IRQ_NONIDLE);
190 instrumentation_end();
191}
192
193/**
194 * ct_nmi_exit - inform RCU of exit from NMI context
195 *
196 * If we are returning from the outermost NMI handler that interrupted an
197 * RCU-idle period, update ct->state and ct->nmi_nesting
198 * to let the RCU grace-period handling know that the CPU is back to
199 * being RCU-idle.
200 *
201 * If you add or remove a call to ct_nmi_exit(), be sure to test
202 * with CONFIG_RCU_EQS_DEBUG=y.
203 */
204void noinstr ct_nmi_exit(void)
205{
206 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
207
208 instrumentation_begin();
209 /*
210 * Check for ->nmi_nesting underflow and bad CT state.
211 * (We are exiting an NMI handler, so RCU better be paying attention
212 * to us!)
213 */
214 WARN_ON_ONCE(ct_nmi_nesting() <= 0);
215 WARN_ON_ONCE(!rcu_is_watching_curr_cpu());
216
217 /*
218 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
219 * leave it in non-RCU-idle state.
220 */
221 if (ct_nmi_nesting() != 1) {
222 trace_rcu_watching(TPS("--="), ct_nmi_nesting(), ct_nmi_nesting() - 2,
223 ct_rcu_watching());
224 WRITE_ONCE(ct->nmi_nesting, /* No store tearing. */
225 ct_nmi_nesting() - 2);
226 instrumentation_end();
227 return;
228 }
229
230 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
231 trace_rcu_watching(TPS("Endirq"), ct_nmi_nesting(), 0, ct_rcu_watching());
232 WRITE_ONCE(ct->nmi_nesting, 0); /* Avoid store tearing. */
233
234 // instrumentation for the noinstr ct_kernel_exit_state()
235 instrument_atomic_write(&ct->state, sizeof(ct->state));
236 instrumentation_end();
237
238 // RCU is watching here ...
239 ct_kernel_exit_state(CT_RCU_WATCHING);
240 // ... but is no longer watching here.
241
242 if (!in_nmi())
243 rcu_task_exit();
244}
245
246/**
247 * ct_nmi_enter - inform RCU of entry to NMI context
248 *
249 * If the CPU was idle from RCU's viewpoint, update ct->state and
250 * ct->nmi_nesting to let the RCU grace-period handling know
251 * that the CPU is active. This implementation permits nested NMIs, as
252 * long as the nesting level does not overflow an int. (You will probably
253 * run out of stack space first.)
254 *
255 * If you add or remove a call to ct_nmi_enter(), be sure to test
256 * with CONFIG_RCU_EQS_DEBUG=y.
257 */
258void noinstr ct_nmi_enter(void)
259{
260 long incby = 2;
261 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
262
263 /* Complain about underflow. */
264 WARN_ON_ONCE(ct_nmi_nesting() < 0);
265
266 /*
267 * If idle from RCU viewpoint, atomically increment CT state
268 * to mark non-idle and increment ->nmi_nesting by one.
269 * Otherwise, increment ->nmi_nesting by two. This means
270 * if ->nmi_nesting is equal to one, we are guaranteed
271 * to be in the outermost NMI handler that interrupted an RCU-idle
272 * period (observation due to Andy Lutomirski).
273 */
274 if (!rcu_is_watching_curr_cpu()) {
275
276 if (!in_nmi())
277 rcu_task_enter();
278
279 // RCU is not watching here ...
280 ct_kernel_enter_state(CT_RCU_WATCHING);
281 // ... but is watching here.
282
283 instrumentation_begin();
284 // instrumentation for the noinstr rcu_is_watching_curr_cpu()
285 instrument_atomic_read(&ct->state, sizeof(ct->state));
286 // instrumentation for the noinstr ct_kernel_enter_state()
287 instrument_atomic_write(&ct->state, sizeof(ct->state));
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_watching(incby == 1 ? TPS("Startirq") : TPS("++="),
298 ct_nmi_nesting(),
299 ct_nmi_nesting() + incby, ct_rcu_watching());
300 instrumentation_end();
301 WRITE_ONCE(ct->nmi_nesting, /* Prevent store tearing. */
302 ct_nmi_nesting() + incby);
303 barrier();
304}
305
306/**
307 * ct_idle_enter - inform RCU that current CPU is entering idle
308 *
309 * Enter idle mode, in other words, -leave- the mode in which RCU
310 * read-side critical sections can occur. (Though RCU read-side
311 * critical sections can occur in irq handlers in idle, a possibility
312 * handled by irq_enter() and irq_exit().)
313 *
314 * If you add or remove a call to ct_idle_enter(), be sure to test with
315 * CONFIG_RCU_EQS_DEBUG=y.
316 */
317void noinstr ct_idle_enter(void)
318{
319 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
320 ct_kernel_exit(false, CT_RCU_WATCHING + CT_STATE_IDLE);
321}
322EXPORT_SYMBOL_GPL(ct_idle_enter);
323
324/**
325 * ct_idle_exit - inform RCU that current CPU is leaving idle
326 *
327 * Exit idle mode, in other words, -enter- the mode in which RCU
328 * read-side critical sections can occur.
329 *
330 * If you add or remove a call to ct_idle_exit(), be sure to test with
331 * CONFIG_RCU_EQS_DEBUG=y.
332 */
333void noinstr ct_idle_exit(void)
334{
335 unsigned long flags;
336
337 raw_local_irq_save(flags);
338 ct_kernel_enter(false, CT_RCU_WATCHING - CT_STATE_IDLE);
339 raw_local_irq_restore(flags);
340}
341EXPORT_SYMBOL_GPL(ct_idle_exit);
342
343/**
344 * ct_irq_enter - inform RCU that current CPU is entering irq away from idle
345 *
346 * Enter an interrupt handler, which might possibly result in exiting
347 * idle mode, in other words, entering the mode in which read-side critical
348 * sections can occur. The caller must have disabled interrupts.
349 *
350 * Note that the Linux kernel is fully capable of entering an interrupt
351 * handler that it never exits, for example when doing upcalls to user mode!
352 * This code assumes that the idle loop never does upcalls to user mode.
353 * If your architecture's idle loop does do upcalls to user mode (or does
354 * anything else that results in unbalanced calls to the irq_enter() and
355 * irq_exit() functions), RCU will give you what you deserve, good and hard.
356 * But very infrequently and irreproducibly.
357 *
358 * Use things like work queues to work around this limitation.
359 *
360 * You have been warned.
361 *
362 * If you add or remove a call to ct_irq_enter(), be sure to test with
363 * CONFIG_RCU_EQS_DEBUG=y.
364 */
365noinstr 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 is exiting irq towards idle
373 *
374 * Exit from an interrupt handler, which might possibly result in entering
375 * idle mode, in other words, leaving the mode in which read-side critical
376 * sections can occur. The caller must have disabled interrupts.
377 *
378 * This code assumes that the idle loop never does anything that might
379 * result in unbalanced calls to irq_enter() and irq_exit(). If your
380 * architecture's idle loop violates this assumption, RCU will give you what
381 * you deserve, good and hard. But very infrequently and irreproducibly.
382 *
383 * Use things like work queues to work around this limitation.
384 *
385 * You have been warned.
386 *
387 * If you add or remove a call to ct_irq_exit(), be sure to test with
388 * CONFIG_RCU_EQS_DEBUG=y.
389 */
390noinstr 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 are enabled.
398 *
399 * If you add or remove a call to ct_irq_enter_irqson(), be sure to test
400 * with CONFIG_RCU_EQS_DEBUG=y.
401 */
402void 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 are enabled.
413 *
414 * If you add or remove a call to ct_irq_exit_irqson(), be sure to test
415 * with CONFIG_RCU_EQS_DEBUG=y.
416 */
417void 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
426static __always_inline void ct_kernel_exit(bool user, int offset) { }
427static __always_inline void ct_kernel_enter(bool user, int offset) { }
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
435DEFINE_STATIC_KEY_FALSE_RO(context_tracking_key);
436EXPORT_SYMBOL_GPL(context_tracking_key);
437
438static noinstr bool context_tracking_recursion_enter(void)
439{
440 int recursion;
441
442 recursion = __this_cpu_inc_return(context_tracking.recursion);
443 if (recursion == 1)
444 return true;
445
446 WARN_ONCE((recursion < 1), "Invalid context tracking recursion value %d\n", recursion);
447 __this_cpu_dec(context_tracking.recursion);
448
449 return false;
450}
451
452static __always_inline void context_tracking_recursion_exit(void)
453{
454 __this_cpu_dec(context_tracking.recursion);
455}
456
457/**
458 * __ct_user_enter - Inform the context tracking that the CPU is going
459 * to enter user or guest space mode.
460 *
461 * @state: userspace context-tracking state to enter.
462 *
463 * This function must be called right before we switch from the kernel
464 * to user or guest space, when it's guaranteed the remaining kernel
465 * instructions to execute won't use any RCU read side critical section
466 * because this function sets RCU in extended quiescent state.
467 */
468void noinstr __ct_user_enter(enum ctx_state state)
469{
470 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
471 lockdep_assert_irqs_disabled();
472
473 /* Kernel threads aren't supposed to go to userspace */
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 low level arch entry code remains and
483 * then we'll run in userspace. We can assume there won't be
484 * any RCU read-side critical section until the next call to
485 * user_exit() or ct_irq_enter(). Let's remove RCU's dependency
486 * on the tick.
487 */
488 if (state == CT_STATE_USER) {
489 instrumentation_begin();
490 trace_user_enter(0);
491 vtime_user_enter(current);
492 instrumentation_end();
493 }
494 /*
495 * Other than generic entry implementation, we may be past the last
496 * rescheduling opportunity in the entry code. Trigger a self IPI
497 * that will fire and reschedule once we resume in user/guest mode.
498 */
499 rcu_irq_work_resched();
500
501 /*
502 * Enter RCU idle mode right before resuming userspace. No use of RCU
503 * is permitted between this call and rcu_eqs_exit(). This way the
504 * CPU doesn't need to maintain the tick for RCU maintenance purposes
505 * when the CPU runs in userspace.
506 */
507 ct_kernel_exit(true, CT_RCU_WATCHING + state);
508
509 /*
510 * Special case if we only track user <-> kernel transitions for tickless
511 * cputime accounting but we don't support RCU extended quiescent state.
512 * In this we case we don't care about any concurrency/ordering.
513 */
514 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE))
515 raw_atomic_set(&ct->state, state);
516 } else {
517 /*
518 * Even if context tracking is disabled on this CPU, because it's outside
519 * the full dynticks mask for example, we still have to keep track of the
520 * context transitions and states to prevent inconsistency on those of
521 * other CPUs.
522 * If a task triggers an exception in userspace, sleep on the exception
523 * handler and then migrate to another CPU, that new CPU must know where
524 * the exception returns by the time we call exception_exit().
525 * This information can only be provided by the previous CPU when it called
526 * exception_enter().
527 * OTOH we can spare the calls to vtime and RCU when context_tracking.active
528 * is false because we know that CPU is not tickless.
529 */
530 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE)) {
531 /* Tracking for vtime only, no concurrent RCU EQS accounting */
532 raw_atomic_set(&ct->state, state);
533 } else {
534 /*
535 * Tracking for vtime and RCU EQS. Make sure we don't race
536 * with NMIs. OTOH we don't care about ordering here since
537 * RCU only requires CT_RCU_WATCHING increments to be fully
538 * ordered.
539 */
540 raw_atomic_add(state, &ct->state);
541 }
542 }
543 }
544 context_tracking_recursion_exit();
545}
546EXPORT_SYMBOL_GPL(__ct_user_enter);
547
548/*
549 * OBSOLETE:
550 * This function should be noinstr but the below local_irq_restore() is
551 * unsafe because it involves illegal RCU uses through tracing and lockdep.
552 * This is unlikely to be fixed as this function is obsolete. The preferred
553 * way is to call __context_tracking_enter() through user_enter_irqoff()
554 * or context_tracking_guest_enter(). It should be the arch entry code
555 * responsibility to call into context tracking with IRQs disabled.
556 */
557void ct_user_enter(enum ctx_state state)
558{
559 unsigned long flags;
560
561 /*
562 * Some contexts may involve an exception occuring in an irq,
563 * leading to that nesting:
564 * ct_irq_enter() rcu_eqs_exit(true) rcu_eqs_enter(true) ct_irq_exit()
565 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
566 * helpers are enough to protect RCU uses inside the exception. So
567 * just return immediately if we detect we are in an IRQ.
568 */
569 if (in_interrupt())
570 return;
571
572 local_irq_save(flags);
573 __ct_user_enter(state);
574 local_irq_restore(flags);
575}
576NOKPROBE_SYMBOL(ct_user_enter);
577EXPORT_SYMBOL_GPL(ct_user_enter);
578
579/**
580 * user_enter_callable() - Unfortunate ASM callable version of user_enter() for
581 * archs that didn't manage to check the context tracking
582 * static key from low level code.
583 *
584 * This OBSOLETE function should be noinstr but it unsafely calls
585 * local_irq_restore(), involving illegal RCU uses through tracing and lockdep.
586 * This is unlikely to be fixed as this function is obsolete. The preferred
587 * way is to call user_enter_irqoff(). It should be the arch entry code
588 * responsibility to call into context tracking with IRQs disabled.
589 */
590void user_enter_callable(void)
591{
592 user_enter();
593}
594NOKPROBE_SYMBOL(user_enter_callable);
595
596/**
597 * __ct_user_exit - Inform the context tracking that the CPU is
598 * exiting user or guest mode and entering the kernel.
599 *
600 * @state: userspace context-tracking state being exited from.
601 *
602 * This function must be called after we entered the kernel from user or
603 * guest space before any use of RCU read side critical section. This
604 * potentially include any high level kernel code like syscalls, exceptions,
605 * signal handling, etc...
606 *
607 * This call supports re-entrancy. This way it can be called from any exception
608 * handler without needing to know if we came from userspace or not.
609 */
610void noinstr __ct_user_exit(enum ctx_state state)
611{
612 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
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 while entering the kernel because it can
621 * run a RCU read side critical section anytime.
622 */
623 ct_kernel_enter(true, CT_RCU_WATCHING - state);
624 if (state == CT_STATE_USER) {
625 instrumentation_begin();
626 vtime_user_exit(current);
627 trace_user_exit(0);
628 instrumentation_end();
629 }
630
631 /*
632 * Special case if we only track user <-> kernel transitions for tickless
633 * cputime accounting but we don't support RCU extended quiescent state.
634 * In this we case we don't care about any concurrency/ordering.
635 */
636 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE))
637 raw_atomic_set(&ct->state, CT_STATE_KERNEL);
638
639 } else {
640 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE)) {
641 /* Tracking for vtime only, no concurrent RCU EQS accounting */
642 raw_atomic_set(&ct->state, CT_STATE_KERNEL);
643 } else {
644 /*
645 * Tracking for vtime and RCU EQS. Make sure we don't race
646 * with NMIs. OTOH we don't care about ordering here since
647 * RCU only requires CT_RCU_WATCHING increments to be fully
648 * ordered.
649 */
650 raw_atomic_sub(state, &ct->state);
651 }
652 }
653 }
654 context_tracking_recursion_exit();
655}
656EXPORT_SYMBOL_GPL(__ct_user_exit);
657
658/*
659 * OBSOLETE:
660 * This function should be noinstr but the below local_irq_save() is
661 * unsafe because it involves illegal RCU uses through tracing and lockdep.
662 * This is unlikely to be fixed as this function is obsolete. The preferred
663 * way is to call __context_tracking_exit() through user_exit_irqoff()
664 * or context_tracking_guest_exit(). It should be the arch entry code
665 * responsibility to call into context tracking with IRQs disabled.
666 */
667void 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}
678NOKPROBE_SYMBOL(ct_user_exit);
679EXPORT_SYMBOL_GPL(ct_user_exit);
680
681/**
682 * user_exit_callable() - Unfortunate ASM callable version of user_exit() for
683 * archs that didn't manage to check the context tracking
684 * static key from low level code.
685 *
686 * This OBSOLETE function should be noinstr but it unsafely calls local_irq_save(),
687 * involving illegal RCU uses through tracing and lockdep. This is unlikely
688 * to be fixed as this function is obsolete. The preferred way is to call
689 * user_exit_irqoff(). It should be the arch entry code responsibility to
690 * call into context tracking with IRQs disabled.
691 */
692void user_exit_callable(void)
693{
694 user_exit();
695}
696NOKPROBE_SYMBOL(user_exit_callable);
697
698void __init ct_cpu_track_user(int cpu)
699{
700 static __initdata bool initialized = false;
701
702 if (!per_cpu(context_tracking.active, cpu)) {
703 per_cpu(context_tracking.active, cpu) = true;
704 static_branch_inc(&context_tracking_key);
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 propagate to all tasks through fork
713 * This assumes that init is the only task at this early boot stage.
714 */
715 set_tsk_thread_flag(&init_task, TIF_NOHZ);
716#endif
717 WARN_ON_ONCE(!tasklist_empty());
718
719 initialized = true;
720}
721
722#ifdef CONFIG_CONTEXT_TRACKING_USER_FORCE
723void __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 */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Context tracking: Probe on high level context boundaries such as kernel,
4 * userspace, guest or idle.
5 *
6 * This is used by RCU to remove its dependency on the timer tick while a CPU
7 * runs in idle, userspace or guest mode.
8 *
9 * User/guest tracking started by Frederic Weisbecker:
10 *
11 * Copyright (C) 2012 Red Hat, Inc., Frederic Weisbecker
12 *
13 * Many thanks to Gilad Ben-Yossef, Paul McKenney, Ingo Molnar, Andrew Morton,
14 * Steven Rostedt, Peter Zijlstra for suggestions and improvements.
15 *
16 * RCU extended quiescent state bits imported from kernel/rcu/tree.c
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
29DEFINE_PER_CPU(struct context_tracking, context_tracking) = {
30#ifdef CONFIG_CONTEXT_TRACKING_IDLE
31 .dynticks_nesting = 1,
32 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
33#endif
34 .state = ATOMIC_INIT(RCU_DYNTICKS_IDX),
35};
36EXPORT_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 entry. */
42static __always_inline void rcu_dynticks_task_enter(void)
43{
44#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
45 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
46#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
47}
48
49/* Record no current task on dyntick-idle exit. */
50static __always_inline void rcu_dynticks_task_exit(void)
51{
52#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
53 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
54#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
55}
56
57/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
58static __always_inline void rcu_dynticks_task_trace_enter(void)
59{
60#ifdef CONFIG_TASKS_TRACE_RCU
61 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
62 current->trc_reader_special.b.need_mb = true;
63#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
64}
65
66/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
67static __always_inline void rcu_dynticks_task_trace_exit(void)
68{
69#ifdef CONFIG_TASKS_TRACE_RCU
70 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
71 current->trc_reader_special.b.need_mb = false;
72#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
73}
74
75/*
76 * Record entry into an extended quiescent state. This is only to be
77 * called when not already in an extended quiescent state, that is,
78 * RCU is watching prior to the call to this function and is no longer
79 * watching upon return.
80 */
81static noinstr void ct_kernel_exit_state(int offset)
82{
83 int seq;
84
85 /*
86 * CPUs seeing atomic_add_return() must see prior RCU read-side
87 * critical sections, and we also must force ordering with the
88 * next idle sojourn.
89 */
90 rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
91 seq = ct_state_inc(offset);
92 // RCU is no longer watching. Better be in extended quiescent state!
93 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && (seq & RCU_DYNTICKS_IDX));
94}
95
96/*
97 * Record exit from an extended quiescent state. This is only to be
98 * called from an extended quiescent state, that is, RCU is not watching
99 * prior to the call to this function and is watching upon return.
100 */
101static noinstr void ct_kernel_enter_state(int offset)
102{
103 int seq;
104
105 /*
106 * CPUs seeing atomic_add_return() must see prior idle sojourns,
107 * and we also must force ordering with the next RCU read-side
108 * critical section.
109 */
110 seq = ct_state_inc(offset);
111 // RCU is now watching. Better not be in an extended quiescent state!
112 rcu_dynticks_task_trace_exit(); // After ->dynticks update!
113 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !(seq & RCU_DYNTICKS_IDX));
114}
115
116/*
117 * Enter an RCU extended quiescent state, which can be either the
118 * idle loop or adaptive-tickless usermode execution.
119 *
120 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
121 * the possibility of usermode upcalls having messed up our count
122 * of interrupt nesting level during the prior busy period.
123 */
124static void noinstr ct_kernel_exit(bool user, int offset)
125{
126 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
127
128 WARN_ON_ONCE(ct_dynticks_nmi_nesting() != DYNTICK_IRQ_NONIDLE);
129 WRITE_ONCE(ct->dynticks_nmi_nesting, 0);
130 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
131 ct_dynticks_nesting() == 0);
132 if (ct_dynticks_nesting() != 1) {
133 // RCU will still be watching, so just do accounting and leave.
134 ct->dynticks_nesting--;
135 return;
136 }
137
138 instrumentation_begin();
139 lockdep_assert_irqs_disabled();
140 trace_rcu_dyntick(TPS("Start"), ct_dynticks_nesting(), 0, ct_dynticks());
141 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
142 rcu_preempt_deferred_qs(current);
143
144 // instrumentation for the noinstr ct_kernel_exit_state()
145 instrument_atomic_write(&ct->state, sizeof(ct->state));
146
147 instrumentation_end();
148 WRITE_ONCE(ct->dynticks_nesting, 0); /* Avoid irq-access tearing. */
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 can be either the
157 * idle loop or adaptive-tickless usermode execution.
158 *
159 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
160 * allow for the possibility of usermode upcalls messing up our count of
161 * interrupt nesting level during the busy period that is just now starting.
162 */
163static void noinstr ct_kernel_enter(bool user, int offset)
164{
165 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
166 long oldval;
167
168 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
169 oldval = ct_dynticks_nesting();
170 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
171 if (oldval) {
172 // RCU was already watching, so just do accounting and leave.
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_kernel_enter_state()
183 instrument_atomic_write(&ct->state, sizeof(ct->state));
184
185 trace_rcu_dyntick(TPS("End"), ct_dynticks_nesting(), 1, ct_dynticks());
186 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
187 WRITE_ONCE(ct->dynticks_nesting, 1);
188 WARN_ON_ONCE(ct_dynticks_nmi_nesting());
189 WRITE_ONCE(ct->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
190 instrumentation_end();
191}
192
193/**
194 * ct_nmi_exit - inform RCU of exit from NMI context
195 *
196 * If we are returning from the outermost NMI handler that interrupted an
197 * RCU-idle period, update ct->state and ct->dynticks_nmi_nesting
198 * to let the RCU grace-period handling know that the CPU is back to
199 * being RCU-idle.
200 *
201 * If you add or remove a call to ct_nmi_exit(), be sure to test
202 * with CONFIG_RCU_EQS_DEBUG=y.
203 */
204void noinstr ct_nmi_exit(void)
205{
206 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
207
208 instrumentation_begin();
209 /*
210 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
211 * (We are exiting an NMI handler, so RCU better be paying attention
212 * to us!)
213 */
214 WARN_ON_ONCE(ct_dynticks_nmi_nesting() <= 0);
215 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
216
217 /*
218 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
219 * leave it in non-RCU-idle state.
220 */
221 if (ct_dynticks_nmi_nesting() != 1) {
222 trace_rcu_dyntick(TPS("--="), ct_dynticks_nmi_nesting(), ct_dynticks_nmi_nesting() - 2,
223 ct_dynticks());
224 WRITE_ONCE(ct->dynticks_nmi_nesting, /* No store tearing. */
225 ct_dynticks_nmi_nesting() - 2);
226 instrumentation_end();
227 return;
228 }
229
230 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
231 trace_rcu_dyntick(TPS("Startirq"), ct_dynticks_nmi_nesting(), 0, ct_dynticks());
232 WRITE_ONCE(ct->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
233
234 // instrumentation for the noinstr ct_kernel_exit_state()
235 instrument_atomic_write(&ct->state, sizeof(ct->state));
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 context
248 *
249 * If the CPU was idle from RCU's viewpoint, update ct->state and
250 * ct->dynticks_nmi_nesting to let the RCU grace-period handling know
251 * that the CPU is active. This implementation permits nested NMIs, as
252 * long as the nesting level does not overflow an int. (You will probably
253 * run out of stack space first.)
254 *
255 * If you add or remove a call to ct_nmi_enter(), be sure to test
256 * with CONFIG_RCU_EQS_DEBUG=y.
257 */
258void noinstr ct_nmi_enter(void)
259{
260 long incby = 2;
261 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
262
263 /* Complain about underflow. */
264 WARN_ON_ONCE(ct_dynticks_nmi_nesting() < 0);
265
266 /*
267 * If idle from RCU viewpoint, atomically increment ->dynticks
268 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
269 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
270 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
271 * to be in the outermost NMI handler that interrupted an RCU-idle
272 * period (observation due to Andy Lutomirski).
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_DYNTICKS_IDX);
281 // ... but is watching here.
282
283 instrumentation_begin();
284 // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
285 instrument_atomic_read(&ct->state, sizeof(ct->state));
286 // instrumentation for the noinstr ct_kernel_enter_state()
287 instrument_atomic_write(&ct->state, sizeof(ct->state));
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("Endirq") : TPS("++="),
298 ct_dynticks_nmi_nesting(),
299 ct_dynticks_nmi_nesting() + incby, ct_dynticks());
300 instrumentation_end();
301 WRITE_ONCE(ct->dynticks_nmi_nesting, /* Prevent store tearing. */
302 ct_dynticks_nmi_nesting() + incby);
303 barrier();
304}
305
306/**
307 * ct_idle_enter - inform RCU that current CPU is entering idle
308 *
309 * Enter idle mode, in other words, -leave- the mode in which RCU
310 * read-side critical sections can occur. (Though RCU read-side
311 * critical sections can occur in irq handlers in idle, a possibility
312 * handled by irq_enter() and irq_exit().)
313 *
314 * If you add or remove a call to ct_idle_enter(), be sure to test with
315 * CONFIG_RCU_EQS_DEBUG=y.
316 */
317void noinstr ct_idle_enter(void)
318{
319 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
320 ct_kernel_exit(false, RCU_DYNTICKS_IDX + CONTEXT_IDLE);
321}
322EXPORT_SYMBOL_GPL(ct_idle_enter);
323
324/**
325 * ct_idle_exit - inform RCU that current CPU is leaving idle
326 *
327 * Exit idle mode, in other words, -enter- the mode in which RCU
328 * read-side critical sections can occur.
329 *
330 * If you add or remove a call to ct_idle_exit(), be sure to test with
331 * CONFIG_RCU_EQS_DEBUG=y.
332 */
333void noinstr ct_idle_exit(void)
334{
335 unsigned long flags;
336
337 raw_local_irq_save(flags);
338 ct_kernel_enter(false, RCU_DYNTICKS_IDX - CONTEXT_IDLE);
339 raw_local_irq_restore(flags);
340}
341EXPORT_SYMBOL_GPL(ct_idle_exit);
342
343/**
344 * ct_irq_enter - inform RCU that current CPU is entering irq away from idle
345 *
346 * Enter an interrupt handler, which might possibly result in exiting
347 * idle mode, in other words, entering the mode in which read-side critical
348 * sections can occur. The caller must have disabled interrupts.
349 *
350 * Note that the Linux kernel is fully capable of entering an interrupt
351 * handler that it never exits, for example when doing upcalls to user mode!
352 * This code assumes that the idle loop never does upcalls to user mode.
353 * If your architecture's idle loop does do upcalls to user mode (or does
354 * anything else that results in unbalanced calls to the irq_enter() and
355 * irq_exit() functions), RCU will give you what you deserve, good and hard.
356 * But very infrequently and irreproducibly.
357 *
358 * Use things like work queues to work around this limitation.
359 *
360 * You have been warned.
361 *
362 * If you add or remove a call to ct_irq_enter(), be sure to test with
363 * CONFIG_RCU_EQS_DEBUG=y.
364 */
365noinstr 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 is exiting irq towards idle
373 *
374 * Exit from an interrupt handler, which might possibly result in entering
375 * idle mode, in other words, leaving the mode in which read-side critical
376 * sections can occur. The caller must have disabled interrupts.
377 *
378 * This code assumes that the idle loop never does anything that might
379 * result in unbalanced calls to irq_enter() and irq_exit(). If your
380 * architecture's idle loop violates this assumption, RCU will give you what
381 * you deserve, good and hard. But very infrequently and irreproducibly.
382 *
383 * Use things like work queues to work around this limitation.
384 *
385 * You have been warned.
386 *
387 * If you add or remove a call to ct_irq_exit(), be sure to test with
388 * CONFIG_RCU_EQS_DEBUG=y.
389 */
390noinstr 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 are enabled.
398 *
399 * If you add or remove a call to ct_irq_enter_irqson(), be sure to test
400 * with CONFIG_RCU_EQS_DEBUG=y.
401 */
402void 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 are enabled.
413 *
414 * If you add or remove a call to ct_irq_exit_irqson(), be sure to test
415 * with CONFIG_RCU_EQS_DEBUG=y.
416 */
417void 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
426static __always_inline void ct_kernel_exit(bool user, int offset) { }
427static __always_inline void ct_kernel_enter(bool user, int offset) { }
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
435DEFINE_STATIC_KEY_FALSE(context_tracking_key);
436EXPORT_SYMBOL_GPL(context_tracking_key);
437
438static noinstr bool context_tracking_recursion_enter(void)
439{
440 int recursion;
441
442 recursion = __this_cpu_inc_return(context_tracking.recursion);
443 if (recursion == 1)
444 return true;
445
446 WARN_ONCE((recursion < 1), "Invalid context tracking recursion value %d\n", recursion);
447 __this_cpu_dec(context_tracking.recursion);
448
449 return false;
450}
451
452static __always_inline void context_tracking_recursion_exit(void)
453{
454 __this_cpu_dec(context_tracking.recursion);
455}
456
457/**
458 * __ct_user_enter - Inform the context tracking that the CPU is going
459 * to enter user or guest space mode.
460 *
461 * This function must be called right before we switch from the kernel
462 * to user or guest space, when it's guaranteed the remaining kernel
463 * instructions to execute won't use any RCU read side critical section
464 * because this function sets RCU in extended quiescent state.
465 */
466void noinstr __ct_user_enter(enum ctx_state state)
467{
468 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
469 lockdep_assert_irqs_disabled();
470
471 /* Kernel threads aren't supposed to go to userspace */
472 WARN_ON_ONCE(!current->mm);
473
474 if (!context_tracking_recursion_enter())
475 return;
476
477 if (__ct_state() != state) {
478 if (ct->active) {
479 /*
480 * At this stage, only low level arch entry code remains and
481 * then we'll run in userspace. We can assume there won't be
482 * any RCU read-side critical section until the next call to
483 * user_exit() or ct_irq_enter(). Let's remove RCU's dependency
484 * on the tick.
485 */
486 if (state == CONTEXT_USER) {
487 instrumentation_begin();
488 trace_user_enter(0);
489 vtime_user_enter(current);
490 instrumentation_end();
491 }
492 /*
493 * Other than generic entry implementation, we may be past the last
494 * rescheduling opportunity in the entry code. Trigger a self IPI
495 * that will fire and reschedule once we resume in user/guest mode.
496 */
497 rcu_irq_work_resched();
498
499 /*
500 * Enter RCU idle mode right before resuming userspace. No use of RCU
501 * is permitted between this call and rcu_eqs_exit(). This way the
502 * CPU doesn't need to maintain the tick for RCU maintenance purposes
503 * when the CPU runs in userspace.
504 */
505 ct_kernel_exit(true, RCU_DYNTICKS_IDX + state);
506
507 /*
508 * Special case if we only track user <-> kernel transitions for tickless
509 * cputime accounting but we don't support RCU extended quiescent state.
510 * In this we case we don't care about any concurrency/ordering.
511 */
512 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE))
513 raw_atomic_set(&ct->state, state);
514 } else {
515 /*
516 * Even if context tracking is disabled on this CPU, because it's outside
517 * the full dynticks mask for example, we still have to keep track of the
518 * context transitions and states to prevent inconsistency on those of
519 * other CPUs.
520 * If a task triggers an exception in userspace, sleep on the exception
521 * handler and then migrate to another CPU, that new CPU must know where
522 * the exception returns by the time we call exception_exit().
523 * This information can only be provided by the previous CPU when it called
524 * exception_enter().
525 * OTOH we can spare the calls to vtime and RCU when context_tracking.active
526 * is false because we know that CPU is not tickless.
527 */
528 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE)) {
529 /* Tracking for vtime only, no concurrent RCU EQS accounting */
530 raw_atomic_set(&ct->state, state);
531 } else {
532 /*
533 * Tracking for vtime and RCU EQS. Make sure we don't race
534 * with NMIs. OTOH we don't care about ordering here since
535 * RCU only requires RCU_DYNTICKS_IDX increments to be fully
536 * ordered.
537 */
538 raw_atomic_add(state, &ct->state);
539 }
540 }
541 }
542 context_tracking_recursion_exit();
543}
544EXPORT_SYMBOL_GPL(__ct_user_enter);
545
546/*
547 * OBSOLETE:
548 * This function should be noinstr but the below local_irq_restore() is
549 * unsafe because it involves illegal RCU uses through tracing and lockdep.
550 * This is unlikely to be fixed as this function is obsolete. The preferred
551 * way is to call __context_tracking_enter() through user_enter_irqoff()
552 * or context_tracking_guest_enter(). It should be the arch entry code
553 * responsibility to call into context tracking with IRQs disabled.
554 */
555void ct_user_enter(enum ctx_state state)
556{
557 unsigned long flags;
558
559 /*
560 * Some contexts may involve an exception occuring in an irq,
561 * leading to that nesting:
562 * ct_irq_enter() rcu_eqs_exit(true) rcu_eqs_enter(true) ct_irq_exit()
563 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
564 * helpers are enough to protect RCU uses inside the exception. So
565 * just return immediately if we detect we are in an IRQ.
566 */
567 if (in_interrupt())
568 return;
569
570 local_irq_save(flags);
571 __ct_user_enter(state);
572 local_irq_restore(flags);
573}
574NOKPROBE_SYMBOL(ct_user_enter);
575EXPORT_SYMBOL_GPL(ct_user_enter);
576
577/**
578 * user_enter_callable() - Unfortunate ASM callable version of user_enter() for
579 * archs that didn't manage to check the context tracking
580 * static key from low level code.
581 *
582 * This OBSOLETE function should be noinstr but it unsafely calls
583 * local_irq_restore(), involving illegal RCU uses through tracing and lockdep.
584 * This is unlikely to be fixed as this function is obsolete. The preferred
585 * way is to call user_enter_irqoff(). It should be the arch entry code
586 * responsibility to call into context tracking with IRQs disabled.
587 */
588void user_enter_callable(void)
589{
590 user_enter();
591}
592NOKPROBE_SYMBOL(user_enter_callable);
593
594/**
595 * __ct_user_exit - Inform the context tracking that the CPU is
596 * exiting user or guest mode and entering the kernel.
597 *
598 * This function must be called after we entered the kernel from user or
599 * guest space before any use of RCU read side critical section. This
600 * potentially include any high level kernel code like syscalls, exceptions,
601 * signal handling, etc...
602 *
603 * This call supports re-entrancy. This way it can be called from any exception
604 * handler without needing to know if we came from userspace or not.
605 */
606void noinstr __ct_user_exit(enum ctx_state state)
607{
608 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
609
610 if (!context_tracking_recursion_enter())
611 return;
612
613 if (__ct_state() == state) {
614 if (ct->active) {
615 /*
616 * Exit RCU idle mode while entering the kernel because it can
617 * run a RCU read side critical section anytime.
618 */
619 ct_kernel_enter(true, RCU_DYNTICKS_IDX - state);
620 if (state == CONTEXT_USER) {
621 instrumentation_begin();
622 vtime_user_exit(current);
623 trace_user_exit(0);
624 instrumentation_end();
625 }
626
627 /*
628 * Special case if we only track user <-> kernel transitions for tickless
629 * cputime accounting but we don't support RCU extended quiescent state.
630 * In this we case we don't care about any concurrency/ordering.
631 */
632 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE))
633 raw_atomic_set(&ct->state, CONTEXT_KERNEL);
634
635 } else {
636 if (!IS_ENABLED(CONFIG_CONTEXT_TRACKING_IDLE)) {
637 /* Tracking for vtime only, no concurrent RCU EQS accounting */
638 raw_atomic_set(&ct->state, CONTEXT_KERNEL);
639 } else {
640 /*
641 * Tracking for vtime and RCU EQS. Make sure we don't race
642 * with NMIs. OTOH we don't care about ordering here since
643 * RCU only requires RCU_DYNTICKS_IDX increments to be fully
644 * ordered.
645 */
646 raw_atomic_sub(state, &ct->state);
647 }
648 }
649 }
650 context_tracking_recursion_exit();
651}
652EXPORT_SYMBOL_GPL(__ct_user_exit);
653
654/*
655 * OBSOLETE:
656 * This function should be noinstr but the below local_irq_save() is
657 * unsafe because it involves illegal RCU uses through tracing and lockdep.
658 * This is unlikely to be fixed as this function is obsolete. The preferred
659 * way is to call __context_tracking_exit() through user_exit_irqoff()
660 * or context_tracking_guest_exit(). It should be the arch entry code
661 * responsibility to call into context tracking with IRQs disabled.
662 */
663void ct_user_exit(enum ctx_state state)
664{
665 unsigned long flags;
666
667 if (in_interrupt())
668 return;
669
670 local_irq_save(flags);
671 __ct_user_exit(state);
672 local_irq_restore(flags);
673}
674NOKPROBE_SYMBOL(ct_user_exit);
675EXPORT_SYMBOL_GPL(ct_user_exit);
676
677/**
678 * user_exit_callable() - Unfortunate ASM callable version of user_exit() for
679 * archs that didn't manage to check the context tracking
680 * static key from low level code.
681 *
682 * This OBSOLETE function should be noinstr but it unsafely calls local_irq_save(),
683 * involving illegal RCU uses through tracing and lockdep. This is unlikely
684 * to be fixed as this function is obsolete. The preferred way is to call
685 * user_exit_irqoff(). It should be the arch entry code responsibility to
686 * call into context tracking with IRQs disabled.
687 */
688void user_exit_callable(void)
689{
690 user_exit();
691}
692NOKPROBE_SYMBOL(user_exit_callable);
693
694void __init ct_cpu_track_user(int cpu)
695{
696 static __initdata bool initialized = false;
697
698 if (!per_cpu(context_tracking.active, cpu)) {
699 per_cpu(context_tracking.active, cpu) = true;
700 static_branch_inc(&context_tracking_key);
701 }
702
703 if (initialized)
704 return;
705
706#ifdef CONFIG_HAVE_TIF_NOHZ
707 /*
708 * Set TIF_NOHZ to init/0 and let it propagate to all tasks through fork
709 * This assumes that init is the only task at this early boot stage.
710 */
711 set_tsk_thread_flag(&init_task, TIF_NOHZ);
712#endif
713 WARN_ON_ONCE(!tasklist_empty());
714
715 initialized = true;
716}
717
718#ifdef CONFIG_CONTEXT_TRACKING_USER_FORCE
719void __init context_tracking_init(void)
720{
721 int cpu;
722
723 for_each_possible_cpu(cpu)
724 ct_cpu_track_user(cpu);
725}
726#endif
727
728#endif /* #ifdef CONFIG_CONTEXT_TRACKING_USER */