Merge tag 'locks-v3.16-2' of git://git.samba.org/jlayton/linux
[linux/fpc-iii.git] / kernel / rcu / tree.c
blobf1ba77363fbb937e41fcdd50564bb514a41cbbdc
1 /*
2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
60 #include "tree.h"
61 #include "rcu.h"
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
66 #endif
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
82 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83 static char sname##_varname[] = #sname; \
84 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85 struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
87 .call = cr, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
97 .abbr = sabbr, \
98 }; \
99 DEFINE_PER_CPU(struct rcu_data, sname##_data)
101 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
104 static struct rcu_state *rcu_state_p;
105 LIST_HEAD(rcu_struct_flavors);
107 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109 module_param(rcu_fanout_leaf, int, 0444);
110 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
112 NUM_RCU_LVL_0,
113 NUM_RCU_LVL_1,
114 NUM_RCU_LVL_2,
115 NUM_RCU_LVL_3,
116 NUM_RCU_LVL_4,
118 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
129 int rcu_scheduler_active __read_mostly;
130 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
142 * a time.
144 static int rcu_scheduler_fully_active __read_mostly;
146 #ifdef CONFIG_RCU_BOOST
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
152 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155 DEFINE_PER_CPU(char, rcu_cpu_has_work);
157 #endif /* #ifdef CONFIG_RCU_BOOST */
159 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq;
173 unsigned long rcutorture_vernum;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state *rsp)
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(int cpu)
193 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
195 if (rdp->passed_quiesce == 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
197 rdp->passed_quiesce = 1;
200 void rcu_bh_qs(int cpu)
202 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
204 if (rdp->passed_quiesce == 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
206 rdp->passed_quiesce = 1;
210 * Note a context switch. This is a quiescent state for RCU-sched,
211 * and requires special handling for preemptible RCU.
212 * The caller must have disabled preemption.
214 void rcu_note_context_switch(int cpu)
216 trace_rcu_utilization(TPS("Start context switch"));
217 rcu_sched_qs(cpu);
218 rcu_preempt_note_context_switch(cpu);
219 trace_rcu_utilization(TPS("End context switch"));
221 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
223 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
224 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
225 .dynticks = ATOMIC_INIT(1),
226 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
227 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
228 .dynticks_idle = ATOMIC_INIT(1),
229 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
232 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
233 static long qhimark = 10000; /* If this many pending, ignore blimit. */
234 static long qlowmark = 100; /* Once only this many pending, use blimit. */
236 module_param(blimit, long, 0444);
237 module_param(qhimark, long, 0444);
238 module_param(qlowmark, long, 0444);
240 static ulong jiffies_till_first_fqs = ULONG_MAX;
241 static ulong jiffies_till_next_fqs = ULONG_MAX;
243 module_param(jiffies_till_first_fqs, ulong, 0644);
244 module_param(jiffies_till_next_fqs, ulong, 0644);
246 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
247 struct rcu_data *rdp);
248 static void force_qs_rnp(struct rcu_state *rsp,
249 int (*f)(struct rcu_data *rsp, bool *isidle,
250 unsigned long *maxj),
251 bool *isidle, unsigned long *maxj);
252 static void force_quiescent_state(struct rcu_state *rsp);
253 static int rcu_pending(int cpu);
256 * Return the number of RCU-sched batches processed thus far for debug & stats.
258 long rcu_batches_completed_sched(void)
260 return rcu_sched_state.completed;
262 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
265 * Return the number of RCU BH batches processed thus far for debug & stats.
267 long rcu_batches_completed_bh(void)
269 return rcu_bh_state.completed;
271 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
274 * Force a quiescent state.
276 void rcu_force_quiescent_state(void)
278 force_quiescent_state(rcu_state_p);
280 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
283 * Force a quiescent state for RCU BH.
285 void rcu_bh_force_quiescent_state(void)
287 force_quiescent_state(&rcu_bh_state);
289 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
292 * Show the state of the grace-period kthreads.
294 void show_rcu_gp_kthreads(void)
296 struct rcu_state *rsp;
298 for_each_rcu_flavor(rsp) {
299 pr_info("%s: wait state: %d ->state: %#lx\n",
300 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
301 /* sched_show_task(rsp->gp_kthread); */
304 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
307 * Record the number of times rcutorture tests have been initiated and
308 * terminated. This information allows the debugfs tracing stats to be
309 * correlated to the rcutorture messages, even when the rcutorture module
310 * is being repeatedly loaded and unloaded. In other words, we cannot
311 * store this state in rcutorture itself.
313 void rcutorture_record_test_transition(void)
315 rcutorture_testseq++;
316 rcutorture_vernum = 0;
318 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
321 * Send along grace-period-related data for rcutorture diagnostics.
323 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
324 unsigned long *gpnum, unsigned long *completed)
326 struct rcu_state *rsp = NULL;
328 switch (test_type) {
329 case RCU_FLAVOR:
330 rsp = rcu_state_p;
331 break;
332 case RCU_BH_FLAVOR:
333 rsp = &rcu_bh_state;
334 break;
335 case RCU_SCHED_FLAVOR:
336 rsp = &rcu_sched_state;
337 break;
338 default:
339 break;
341 if (rsp != NULL) {
342 *flags = ACCESS_ONCE(rsp->gp_flags);
343 *gpnum = ACCESS_ONCE(rsp->gpnum);
344 *completed = ACCESS_ONCE(rsp->completed);
345 return;
347 *flags = 0;
348 *gpnum = 0;
349 *completed = 0;
351 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
354 * Record the number of writer passes through the current rcutorture test.
355 * This is also used to correlate debugfs tracing stats with the rcutorture
356 * messages.
358 void rcutorture_record_progress(unsigned long vernum)
360 rcutorture_vernum++;
362 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
365 * Force a quiescent state for RCU-sched.
367 void rcu_sched_force_quiescent_state(void)
369 force_quiescent_state(&rcu_sched_state);
371 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
374 * Does the CPU have callbacks ready to be invoked?
376 static int
377 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
379 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
380 rdp->nxttail[RCU_DONE_TAIL] != NULL;
384 * Return the root node of the specified rcu_state structure.
386 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
388 return &rsp->node[0];
392 * Is there any need for future grace periods?
393 * Interrupts must be disabled. If the caller does not hold the root
394 * rnp_node structure's ->lock, the results are advisory only.
396 static int rcu_future_needs_gp(struct rcu_state *rsp)
398 struct rcu_node *rnp = rcu_get_root(rsp);
399 int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1;
400 int *fp = &rnp->need_future_gp[idx];
402 return ACCESS_ONCE(*fp);
406 * Does the current CPU require a not-yet-started grace period?
407 * The caller must have disabled interrupts to prevent races with
408 * normal callback registry.
410 static int
411 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
413 int i;
415 if (rcu_gp_in_progress(rsp))
416 return 0; /* No, a grace period is already in progress. */
417 if (rcu_future_needs_gp(rsp))
418 return 1; /* Yes, a no-CBs CPU needs one. */
419 if (!rdp->nxttail[RCU_NEXT_TAIL])
420 return 0; /* No, this is a no-CBs (or offline) CPU. */
421 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
422 return 1; /* Yes, this CPU has newly registered callbacks. */
423 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
424 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
425 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
426 rdp->nxtcompleted[i]))
427 return 1; /* Yes, CBs for future grace period. */
428 return 0; /* No grace period needed. */
432 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
434 * If the new value of the ->dynticks_nesting counter now is zero,
435 * we really have entered idle, and must do the appropriate accounting.
436 * The caller must have disabled interrupts.
438 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
439 bool user)
441 struct rcu_state *rsp;
442 struct rcu_data *rdp;
444 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
445 if (!user && !is_idle_task(current)) {
446 struct task_struct *idle __maybe_unused =
447 idle_task(smp_processor_id());
449 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
450 ftrace_dump(DUMP_ORIG);
451 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
452 current->pid, current->comm,
453 idle->pid, idle->comm); /* must be idle task! */
455 for_each_rcu_flavor(rsp) {
456 rdp = this_cpu_ptr(rsp->rda);
457 do_nocb_deferred_wakeup(rdp);
459 rcu_prepare_for_idle(smp_processor_id());
460 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
461 smp_mb__before_atomic(); /* See above. */
462 atomic_inc(&rdtp->dynticks);
463 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
464 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
467 * It is illegal to enter an extended quiescent state while
468 * in an RCU read-side critical section.
470 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
471 "Illegal idle entry in RCU read-side critical section.");
472 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
473 "Illegal idle entry in RCU-bh read-side critical section.");
474 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
475 "Illegal idle entry in RCU-sched read-side critical section.");
479 * Enter an RCU extended quiescent state, which can be either the
480 * idle loop or adaptive-tickless usermode execution.
482 static void rcu_eqs_enter(bool user)
484 long long oldval;
485 struct rcu_dynticks *rdtp;
487 rdtp = this_cpu_ptr(&rcu_dynticks);
488 oldval = rdtp->dynticks_nesting;
489 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
490 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
491 rdtp->dynticks_nesting = 0;
492 rcu_eqs_enter_common(rdtp, oldval, user);
493 } else {
494 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
499 * rcu_idle_enter - inform RCU that current CPU is entering idle
501 * Enter idle mode, in other words, -leave- the mode in which RCU
502 * read-side critical sections can occur. (Though RCU read-side
503 * critical sections can occur in irq handlers in idle, a possibility
504 * handled by irq_enter() and irq_exit().)
506 * We crowbar the ->dynticks_nesting field to zero to allow for
507 * the possibility of usermode upcalls having messed up our count
508 * of interrupt nesting level during the prior busy period.
510 void rcu_idle_enter(void)
512 unsigned long flags;
514 local_irq_save(flags);
515 rcu_eqs_enter(false);
516 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
517 local_irq_restore(flags);
519 EXPORT_SYMBOL_GPL(rcu_idle_enter);
521 #ifdef CONFIG_RCU_USER_QS
523 * rcu_user_enter - inform RCU that we are resuming userspace.
525 * Enter RCU idle mode right before resuming userspace. No use of RCU
526 * is permitted between this call and rcu_user_exit(). This way the
527 * CPU doesn't need to maintain the tick for RCU maintenance purposes
528 * when the CPU runs in userspace.
530 void rcu_user_enter(void)
532 rcu_eqs_enter(1);
534 #endif /* CONFIG_RCU_USER_QS */
537 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
539 * Exit from an interrupt handler, which might possibly result in entering
540 * idle mode, in other words, leaving the mode in which read-side critical
541 * sections can occur.
543 * This code assumes that the idle loop never does anything that might
544 * result in unbalanced calls to irq_enter() and irq_exit(). If your
545 * architecture violates this assumption, RCU will give you what you
546 * deserve, good and hard. But very infrequently and irreproducibly.
548 * Use things like work queues to work around this limitation.
550 * You have been warned.
552 void rcu_irq_exit(void)
554 unsigned long flags;
555 long long oldval;
556 struct rcu_dynticks *rdtp;
558 local_irq_save(flags);
559 rdtp = this_cpu_ptr(&rcu_dynticks);
560 oldval = rdtp->dynticks_nesting;
561 rdtp->dynticks_nesting--;
562 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
563 if (rdtp->dynticks_nesting)
564 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
565 else
566 rcu_eqs_enter_common(rdtp, oldval, true);
567 rcu_sysidle_enter(rdtp, 1);
568 local_irq_restore(flags);
572 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
574 * If the new value of the ->dynticks_nesting counter was previously zero,
575 * we really have exited idle, and must do the appropriate accounting.
576 * The caller must have disabled interrupts.
578 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
579 int user)
581 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
582 atomic_inc(&rdtp->dynticks);
583 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
584 smp_mb__after_atomic(); /* See above. */
585 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
586 rcu_cleanup_after_idle(smp_processor_id());
587 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
588 if (!user && !is_idle_task(current)) {
589 struct task_struct *idle __maybe_unused =
590 idle_task(smp_processor_id());
592 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
593 oldval, rdtp->dynticks_nesting);
594 ftrace_dump(DUMP_ORIG);
595 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
596 current->pid, current->comm,
597 idle->pid, idle->comm); /* must be idle task! */
602 * Exit an RCU extended quiescent state, which can be either the
603 * idle loop or adaptive-tickless usermode execution.
605 static void rcu_eqs_exit(bool user)
607 struct rcu_dynticks *rdtp;
608 long long oldval;
610 rdtp = this_cpu_ptr(&rcu_dynticks);
611 oldval = rdtp->dynticks_nesting;
612 WARN_ON_ONCE(oldval < 0);
613 if (oldval & DYNTICK_TASK_NEST_MASK) {
614 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
615 } else {
616 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
617 rcu_eqs_exit_common(rdtp, oldval, user);
622 * rcu_idle_exit - inform RCU that current CPU is leaving idle
624 * Exit idle mode, in other words, -enter- the mode in which RCU
625 * read-side critical sections can occur.
627 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
628 * allow for the possibility of usermode upcalls messing up our count
629 * of interrupt nesting level during the busy period that is just
630 * now starting.
632 void rcu_idle_exit(void)
634 unsigned long flags;
636 local_irq_save(flags);
637 rcu_eqs_exit(false);
638 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
639 local_irq_restore(flags);
641 EXPORT_SYMBOL_GPL(rcu_idle_exit);
643 #ifdef CONFIG_RCU_USER_QS
645 * rcu_user_exit - inform RCU that we are exiting userspace.
647 * Exit RCU idle mode while entering the kernel because it can
648 * run a RCU read side critical section anytime.
650 void rcu_user_exit(void)
652 rcu_eqs_exit(1);
654 #endif /* CONFIG_RCU_USER_QS */
657 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
659 * Enter an interrupt handler, which might possibly result in exiting
660 * idle mode, in other words, entering the mode in which read-side critical
661 * sections can occur.
663 * Note that the Linux kernel is fully capable of entering an interrupt
664 * handler that it never exits, for example when doing upcalls to
665 * user mode! This code assumes that the idle loop never does upcalls to
666 * user mode. If your architecture does do upcalls from the idle loop (or
667 * does anything else that results in unbalanced calls to the irq_enter()
668 * and irq_exit() functions), RCU will give you what you deserve, good
669 * and hard. But very infrequently and irreproducibly.
671 * Use things like work queues to work around this limitation.
673 * You have been warned.
675 void rcu_irq_enter(void)
677 unsigned long flags;
678 struct rcu_dynticks *rdtp;
679 long long oldval;
681 local_irq_save(flags);
682 rdtp = this_cpu_ptr(&rcu_dynticks);
683 oldval = rdtp->dynticks_nesting;
684 rdtp->dynticks_nesting++;
685 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
686 if (oldval)
687 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
688 else
689 rcu_eqs_exit_common(rdtp, oldval, true);
690 rcu_sysidle_exit(rdtp, 1);
691 local_irq_restore(flags);
695 * rcu_nmi_enter - inform RCU of entry to NMI context
697 * If the CPU was idle with dynamic ticks active, and there is no
698 * irq handler running, this updates rdtp->dynticks_nmi to let the
699 * RCU grace-period handling know that the CPU is active.
701 void rcu_nmi_enter(void)
703 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
705 if (rdtp->dynticks_nmi_nesting == 0 &&
706 (atomic_read(&rdtp->dynticks) & 0x1))
707 return;
708 rdtp->dynticks_nmi_nesting++;
709 smp_mb__before_atomic(); /* Force delay from prior write. */
710 atomic_inc(&rdtp->dynticks);
711 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
712 smp_mb__after_atomic(); /* See above. */
713 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
717 * rcu_nmi_exit - inform RCU of exit from NMI context
719 * If the CPU was idle with dynamic ticks active, and there is no
720 * irq handler running, this updates rdtp->dynticks_nmi to let the
721 * RCU grace-period handling know that the CPU is no longer active.
723 void rcu_nmi_exit(void)
725 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
727 if (rdtp->dynticks_nmi_nesting == 0 ||
728 --rdtp->dynticks_nmi_nesting != 0)
729 return;
730 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
731 smp_mb__before_atomic(); /* See above. */
732 atomic_inc(&rdtp->dynticks);
733 smp_mb__after_atomic(); /* Force delay to next write. */
734 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
738 * __rcu_is_watching - are RCU read-side critical sections safe?
740 * Return true if RCU is watching the running CPU, which means that
741 * this CPU can safely enter RCU read-side critical sections. Unlike
742 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
743 * least disabled preemption.
745 bool notrace __rcu_is_watching(void)
747 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
751 * rcu_is_watching - see if RCU thinks that the current CPU is idle
753 * If the current CPU is in its idle loop and is neither in an interrupt
754 * or NMI handler, return true.
756 bool notrace rcu_is_watching(void)
758 int ret;
760 preempt_disable();
761 ret = __rcu_is_watching();
762 preempt_enable();
763 return ret;
765 EXPORT_SYMBOL_GPL(rcu_is_watching);
767 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
770 * Is the current CPU online? Disable preemption to avoid false positives
771 * that could otherwise happen due to the current CPU number being sampled,
772 * this task being preempted, its old CPU being taken offline, resuming
773 * on some other CPU, then determining that its old CPU is now offline.
774 * It is OK to use RCU on an offline processor during initial boot, hence
775 * the check for rcu_scheduler_fully_active. Note also that it is OK
776 * for a CPU coming online to use RCU for one jiffy prior to marking itself
777 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
778 * offline to continue to use RCU for one jiffy after marking itself
779 * offline in the cpu_online_mask. This leniency is necessary given the
780 * non-atomic nature of the online and offline processing, for example,
781 * the fact that a CPU enters the scheduler after completing the CPU_DYING
782 * notifiers.
784 * This is also why RCU internally marks CPUs online during the
785 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
787 * Disable checking if in an NMI handler because we cannot safely report
788 * errors from NMI handlers anyway.
790 bool rcu_lockdep_current_cpu_online(void)
792 struct rcu_data *rdp;
793 struct rcu_node *rnp;
794 bool ret;
796 if (in_nmi())
797 return true;
798 preempt_disable();
799 rdp = this_cpu_ptr(&rcu_sched_data);
800 rnp = rdp->mynode;
801 ret = (rdp->grpmask & rnp->qsmaskinit) ||
802 !rcu_scheduler_fully_active;
803 preempt_enable();
804 return ret;
806 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
808 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
811 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
813 * If the current CPU is idle or running at a first-level (not nested)
814 * interrupt from idle, return true. The caller must have at least
815 * disabled preemption.
817 static int rcu_is_cpu_rrupt_from_idle(void)
819 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
823 * Snapshot the specified CPU's dynticks counter so that we can later
824 * credit them with an implicit quiescent state. Return 1 if this CPU
825 * is in dynticks idle mode, which is an extended quiescent state.
827 static int dyntick_save_progress_counter(struct rcu_data *rdp,
828 bool *isidle, unsigned long *maxj)
830 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
831 rcu_sysidle_check_cpu(rdp, isidle, maxj);
832 if ((rdp->dynticks_snap & 0x1) == 0) {
833 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
834 return 1;
835 } else {
836 return 0;
841 * This function really isn't for public consumption, but RCU is special in
842 * that context switches can allow the state machine to make progress.
844 extern void resched_cpu(int cpu);
847 * Return true if the specified CPU has passed through a quiescent
848 * state by virtue of being in or having passed through an dynticks
849 * idle state since the last call to dyntick_save_progress_counter()
850 * for this same CPU, or by virtue of having been offline.
852 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
853 bool *isidle, unsigned long *maxj)
855 unsigned int curr;
856 unsigned int snap;
858 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
859 snap = (unsigned int)rdp->dynticks_snap;
862 * If the CPU passed through or entered a dynticks idle phase with
863 * no active irq/NMI handlers, then we can safely pretend that the CPU
864 * already acknowledged the request to pass through a quiescent
865 * state. Either way, that CPU cannot possibly be in an RCU
866 * read-side critical section that started before the beginning
867 * of the current RCU grace period.
869 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
870 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
871 rdp->dynticks_fqs++;
872 return 1;
876 * Check for the CPU being offline, but only if the grace period
877 * is old enough. We don't need to worry about the CPU changing
878 * state: If we see it offline even once, it has been through a
879 * quiescent state.
881 * The reason for insisting that the grace period be at least
882 * one jiffy old is that CPUs that are not quite online and that
883 * have just gone offline can still execute RCU read-side critical
884 * sections.
886 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
887 return 0; /* Grace period is not old enough. */
888 barrier();
889 if (cpu_is_offline(rdp->cpu)) {
890 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
891 rdp->offline_fqs++;
892 return 1;
896 * There is a possibility that a CPU in adaptive-ticks state
897 * might run in the kernel with the scheduling-clock tick disabled
898 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
899 * force the CPU to restart the scheduling-clock tick in this
900 * CPU is in this state.
902 rcu_kick_nohz_cpu(rdp->cpu);
905 * Alternatively, the CPU might be running in the kernel
906 * for an extended period of time without a quiescent state.
907 * Attempt to force the CPU through the scheduler to gain the
908 * needed quiescent state, but only if the grace period has gone
909 * on for an uncommonly long time. If there are many stuck CPUs,
910 * we will beat on the first one until it gets unstuck, then move
911 * to the next. Only do this for the primary flavor of RCU.
913 if (rdp->rsp == rcu_state_p &&
914 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
915 rdp->rsp->jiffies_resched += 5;
916 resched_cpu(rdp->cpu);
919 return 0;
922 static void record_gp_stall_check_time(struct rcu_state *rsp)
924 unsigned long j = jiffies;
925 unsigned long j1;
927 rsp->gp_start = j;
928 smp_wmb(); /* Record start time before stall time. */
929 j1 = rcu_jiffies_till_stall_check();
930 ACCESS_ONCE(rsp->jiffies_stall) = j + j1;
931 rsp->jiffies_resched = j + j1 / 2;
935 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
936 * for architectures that do not implement trigger_all_cpu_backtrace().
937 * The NMI-triggered stack traces are more accurate because they are
938 * printed by the target CPU.
940 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
942 int cpu;
943 unsigned long flags;
944 struct rcu_node *rnp;
946 rcu_for_each_leaf_node(rsp, rnp) {
947 raw_spin_lock_irqsave(&rnp->lock, flags);
948 if (rnp->qsmask != 0) {
949 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
950 if (rnp->qsmask & (1UL << cpu))
951 dump_cpu_task(rnp->grplo + cpu);
953 raw_spin_unlock_irqrestore(&rnp->lock, flags);
957 static void print_other_cpu_stall(struct rcu_state *rsp)
959 int cpu;
960 long delta;
961 unsigned long flags;
962 int ndetected = 0;
963 struct rcu_node *rnp = rcu_get_root(rsp);
964 long totqlen = 0;
966 /* Only let one CPU complain about others per time interval. */
968 raw_spin_lock_irqsave(&rnp->lock, flags);
969 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
970 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
971 raw_spin_unlock_irqrestore(&rnp->lock, flags);
972 return;
974 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
975 raw_spin_unlock_irqrestore(&rnp->lock, flags);
978 * OK, time to rat on our buddy...
979 * See Documentation/RCU/stallwarn.txt for info on how to debug
980 * RCU CPU stall warnings.
982 pr_err("INFO: %s detected stalls on CPUs/tasks:",
983 rsp->name);
984 print_cpu_stall_info_begin();
985 rcu_for_each_leaf_node(rsp, rnp) {
986 raw_spin_lock_irqsave(&rnp->lock, flags);
987 ndetected += rcu_print_task_stall(rnp);
988 if (rnp->qsmask != 0) {
989 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
990 if (rnp->qsmask & (1UL << cpu)) {
991 print_cpu_stall_info(rsp,
992 rnp->grplo + cpu);
993 ndetected++;
996 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1000 * Now rat on any tasks that got kicked up to the root rcu_node
1001 * due to CPU offlining.
1003 rnp = rcu_get_root(rsp);
1004 raw_spin_lock_irqsave(&rnp->lock, flags);
1005 ndetected += rcu_print_task_stall(rnp);
1006 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1008 print_cpu_stall_info_end();
1009 for_each_possible_cpu(cpu)
1010 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1011 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1012 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1013 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1014 if (ndetected == 0)
1015 pr_err("INFO: Stall ended before state dump start\n");
1016 else if (!trigger_all_cpu_backtrace())
1017 rcu_dump_cpu_stacks(rsp);
1019 /* Complain about tasks blocking the grace period. */
1021 rcu_print_detail_task_stall(rsp);
1023 force_quiescent_state(rsp); /* Kick them all. */
1026 static void print_cpu_stall(struct rcu_state *rsp)
1028 int cpu;
1029 unsigned long flags;
1030 struct rcu_node *rnp = rcu_get_root(rsp);
1031 long totqlen = 0;
1034 * OK, time to rat on ourselves...
1035 * See Documentation/RCU/stallwarn.txt for info on how to debug
1036 * RCU CPU stall warnings.
1038 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1039 print_cpu_stall_info_begin();
1040 print_cpu_stall_info(rsp, smp_processor_id());
1041 print_cpu_stall_info_end();
1042 for_each_possible_cpu(cpu)
1043 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1044 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1045 jiffies - rsp->gp_start,
1046 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1047 if (!trigger_all_cpu_backtrace())
1048 dump_stack();
1050 raw_spin_lock_irqsave(&rnp->lock, flags);
1051 if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall)))
1052 ACCESS_ONCE(rsp->jiffies_stall) = jiffies +
1053 3 * rcu_jiffies_till_stall_check() + 3;
1054 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1057 * Attempt to revive the RCU machinery by forcing a context switch.
1059 * A context switch would normally allow the RCU state machine to make
1060 * progress and it could be we're stuck in kernel space without context
1061 * switches for an entirely unreasonable amount of time.
1063 resched_cpu(smp_processor_id());
1066 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1068 unsigned long completed;
1069 unsigned long gpnum;
1070 unsigned long gps;
1071 unsigned long j;
1072 unsigned long js;
1073 struct rcu_node *rnp;
1075 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1076 return;
1077 j = jiffies;
1080 * Lots of memory barriers to reject false positives.
1082 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1083 * then rsp->gp_start, and finally rsp->completed. These values
1084 * are updated in the opposite order with memory barriers (or
1085 * equivalent) during grace-period initialization and cleanup.
1086 * Now, a false positive can occur if we get an new value of
1087 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1088 * the memory barriers, the only way that this can happen is if one
1089 * grace period ends and another starts between these two fetches.
1090 * Detect this by comparing rsp->completed with the previous fetch
1091 * from rsp->gpnum.
1093 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1094 * and rsp->gp_start suffice to forestall false positives.
1096 gpnum = ACCESS_ONCE(rsp->gpnum);
1097 smp_rmb(); /* Pick up ->gpnum first... */
1098 js = ACCESS_ONCE(rsp->jiffies_stall);
1099 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1100 gps = ACCESS_ONCE(rsp->gp_start);
1101 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1102 completed = ACCESS_ONCE(rsp->completed);
1103 if (ULONG_CMP_GE(completed, gpnum) ||
1104 ULONG_CMP_LT(j, js) ||
1105 ULONG_CMP_GE(gps, js))
1106 return; /* No stall or GP completed since entering function. */
1107 rnp = rdp->mynode;
1108 if (rcu_gp_in_progress(rsp) &&
1109 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1111 /* We haven't checked in, so go dump stack. */
1112 print_cpu_stall(rsp);
1114 } else if (rcu_gp_in_progress(rsp) &&
1115 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1117 /* They had a few time units to dump stack, so complain. */
1118 print_other_cpu_stall(rsp);
1123 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1125 * Set the stall-warning timeout way off into the future, thus preventing
1126 * any RCU CPU stall-warning messages from appearing in the current set of
1127 * RCU grace periods.
1129 * The caller must disable hard irqs.
1131 void rcu_cpu_stall_reset(void)
1133 struct rcu_state *rsp;
1135 for_each_rcu_flavor(rsp)
1136 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2;
1140 * Initialize the specified rcu_data structure's callback list to empty.
1142 static void init_callback_list(struct rcu_data *rdp)
1144 int i;
1146 if (init_nocb_callback_list(rdp))
1147 return;
1148 rdp->nxtlist = NULL;
1149 for (i = 0; i < RCU_NEXT_SIZE; i++)
1150 rdp->nxttail[i] = &rdp->nxtlist;
1154 * Determine the value that ->completed will have at the end of the
1155 * next subsequent grace period. This is used to tag callbacks so that
1156 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1157 * been dyntick-idle for an extended period with callbacks under the
1158 * influence of RCU_FAST_NO_HZ.
1160 * The caller must hold rnp->lock with interrupts disabled.
1162 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1163 struct rcu_node *rnp)
1166 * If RCU is idle, we just wait for the next grace period.
1167 * But we can only be sure that RCU is idle if we are looking
1168 * at the root rcu_node structure -- otherwise, a new grace
1169 * period might have started, but just not yet gotten around
1170 * to initializing the current non-root rcu_node structure.
1172 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1173 return rnp->completed + 1;
1176 * Otherwise, wait for a possible partial grace period and
1177 * then the subsequent full grace period.
1179 return rnp->completed + 2;
1183 * Trace-event helper function for rcu_start_future_gp() and
1184 * rcu_nocb_wait_gp().
1186 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1187 unsigned long c, const char *s)
1189 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1190 rnp->completed, c, rnp->level,
1191 rnp->grplo, rnp->grphi, s);
1195 * Start some future grace period, as needed to handle newly arrived
1196 * callbacks. The required future grace periods are recorded in each
1197 * rcu_node structure's ->need_future_gp field. Returns true if there
1198 * is reason to awaken the grace-period kthread.
1200 * The caller must hold the specified rcu_node structure's ->lock.
1202 static bool __maybe_unused
1203 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1204 unsigned long *c_out)
1206 unsigned long c;
1207 int i;
1208 bool ret = false;
1209 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1212 * Pick up grace-period number for new callbacks. If this
1213 * grace period is already marked as needed, return to the caller.
1215 c = rcu_cbs_completed(rdp->rsp, rnp);
1216 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1217 if (rnp->need_future_gp[c & 0x1]) {
1218 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1219 goto out;
1223 * If either this rcu_node structure or the root rcu_node structure
1224 * believe that a grace period is in progress, then we must wait
1225 * for the one following, which is in "c". Because our request
1226 * will be noticed at the end of the current grace period, we don't
1227 * need to explicitly start one.
1229 if (rnp->gpnum != rnp->completed ||
1230 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1231 rnp->need_future_gp[c & 0x1]++;
1232 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1233 goto out;
1237 * There might be no grace period in progress. If we don't already
1238 * hold it, acquire the root rcu_node structure's lock in order to
1239 * start one (if needed).
1241 if (rnp != rnp_root) {
1242 raw_spin_lock(&rnp_root->lock);
1243 smp_mb__after_unlock_lock();
1247 * Get a new grace-period number. If there really is no grace
1248 * period in progress, it will be smaller than the one we obtained
1249 * earlier. Adjust callbacks as needed. Note that even no-CBs
1250 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1252 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1253 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1254 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1255 rdp->nxtcompleted[i] = c;
1258 * If the needed for the required grace period is already
1259 * recorded, trace and leave.
1261 if (rnp_root->need_future_gp[c & 0x1]) {
1262 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1263 goto unlock_out;
1266 /* Record the need for the future grace period. */
1267 rnp_root->need_future_gp[c & 0x1]++;
1269 /* If a grace period is not already in progress, start one. */
1270 if (rnp_root->gpnum != rnp_root->completed) {
1271 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1272 } else {
1273 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1274 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1276 unlock_out:
1277 if (rnp != rnp_root)
1278 raw_spin_unlock(&rnp_root->lock);
1279 out:
1280 if (c_out != NULL)
1281 *c_out = c;
1282 return ret;
1286 * Clean up any old requests for the just-ended grace period. Also return
1287 * whether any additional grace periods have been requested. Also invoke
1288 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1289 * waiting for this grace period to complete.
1291 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1293 int c = rnp->completed;
1294 int needmore;
1295 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1297 rcu_nocb_gp_cleanup(rsp, rnp);
1298 rnp->need_future_gp[c & 0x1] = 0;
1299 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1300 trace_rcu_future_gp(rnp, rdp, c,
1301 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1302 return needmore;
1306 * Awaken the grace-period kthread for the specified flavor of RCU.
1307 * Don't do a self-awaken, and don't bother awakening when there is
1308 * nothing for the grace-period kthread to do (as in several CPUs
1309 * raced to awaken, and we lost), and finally don't try to awaken
1310 * a kthread that has not yet been created.
1312 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1314 if (current == rsp->gp_kthread ||
1315 !ACCESS_ONCE(rsp->gp_flags) ||
1316 !rsp->gp_kthread)
1317 return;
1318 wake_up(&rsp->gp_wq);
1322 * If there is room, assign a ->completed number to any callbacks on
1323 * this CPU that have not already been assigned. Also accelerate any
1324 * callbacks that were previously assigned a ->completed number that has
1325 * since proven to be too conservative, which can happen if callbacks get
1326 * assigned a ->completed number while RCU is idle, but with reference to
1327 * a non-root rcu_node structure. This function is idempotent, so it does
1328 * not hurt to call it repeatedly. Returns an flag saying that we should
1329 * awaken the RCU grace-period kthread.
1331 * The caller must hold rnp->lock with interrupts disabled.
1333 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1334 struct rcu_data *rdp)
1336 unsigned long c;
1337 int i;
1338 bool ret;
1340 /* If the CPU has no callbacks, nothing to do. */
1341 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1342 return false;
1345 * Starting from the sublist containing the callbacks most
1346 * recently assigned a ->completed number and working down, find the
1347 * first sublist that is not assignable to an upcoming grace period.
1348 * Such a sublist has something in it (first two tests) and has
1349 * a ->completed number assigned that will complete sooner than
1350 * the ->completed number for newly arrived callbacks (last test).
1352 * The key point is that any later sublist can be assigned the
1353 * same ->completed number as the newly arrived callbacks, which
1354 * means that the callbacks in any of these later sublist can be
1355 * grouped into a single sublist, whether or not they have already
1356 * been assigned a ->completed number.
1358 c = rcu_cbs_completed(rsp, rnp);
1359 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1360 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1361 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1362 break;
1365 * If there are no sublist for unassigned callbacks, leave.
1366 * At the same time, advance "i" one sublist, so that "i" will
1367 * index into the sublist where all the remaining callbacks should
1368 * be grouped into.
1370 if (++i >= RCU_NEXT_TAIL)
1371 return false;
1374 * Assign all subsequent callbacks' ->completed number to the next
1375 * full grace period and group them all in the sublist initially
1376 * indexed by "i".
1378 for (; i <= RCU_NEXT_TAIL; i++) {
1379 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1380 rdp->nxtcompleted[i] = c;
1382 /* Record any needed additional grace periods. */
1383 ret = rcu_start_future_gp(rnp, rdp, NULL);
1385 /* Trace depending on how much we were able to accelerate. */
1386 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1387 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1388 else
1389 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1390 return ret;
1394 * Move any callbacks whose grace period has completed to the
1395 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1396 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1397 * sublist. This function is idempotent, so it does not hurt to
1398 * invoke it repeatedly. As long as it is not invoked -too- often...
1399 * Returns true if the RCU grace-period kthread needs to be awakened.
1401 * The caller must hold rnp->lock with interrupts disabled.
1403 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1404 struct rcu_data *rdp)
1406 int i, j;
1408 /* If the CPU has no callbacks, nothing to do. */
1409 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1410 return false;
1413 * Find all callbacks whose ->completed numbers indicate that they
1414 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1416 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1417 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1418 break;
1419 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1421 /* Clean up any sublist tail pointers that were misordered above. */
1422 for (j = RCU_WAIT_TAIL; j < i; j++)
1423 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1425 /* Copy down callbacks to fill in empty sublists. */
1426 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1427 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1428 break;
1429 rdp->nxttail[j] = rdp->nxttail[i];
1430 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1433 /* Classify any remaining callbacks. */
1434 return rcu_accelerate_cbs(rsp, rnp, rdp);
1438 * Update CPU-local rcu_data state to record the beginnings and ends of
1439 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1440 * structure corresponding to the current CPU, and must have irqs disabled.
1441 * Returns true if the grace-period kthread needs to be awakened.
1443 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1444 struct rcu_data *rdp)
1446 bool ret;
1448 /* Handle the ends of any preceding grace periods first. */
1449 if (rdp->completed == rnp->completed) {
1451 /* No grace period end, so just accelerate recent callbacks. */
1452 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1454 } else {
1456 /* Advance callbacks. */
1457 ret = rcu_advance_cbs(rsp, rnp, rdp);
1459 /* Remember that we saw this grace-period completion. */
1460 rdp->completed = rnp->completed;
1461 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1464 if (rdp->gpnum != rnp->gpnum) {
1466 * If the current grace period is waiting for this CPU,
1467 * set up to detect a quiescent state, otherwise don't
1468 * go looking for one.
1470 rdp->gpnum = rnp->gpnum;
1471 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1472 rdp->passed_quiesce = 0;
1473 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1474 zero_cpu_stall_ticks(rdp);
1476 return ret;
1479 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1481 unsigned long flags;
1482 bool needwake;
1483 struct rcu_node *rnp;
1485 local_irq_save(flags);
1486 rnp = rdp->mynode;
1487 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1488 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1489 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1490 local_irq_restore(flags);
1491 return;
1493 smp_mb__after_unlock_lock();
1494 needwake = __note_gp_changes(rsp, rnp, rdp);
1495 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1496 if (needwake)
1497 rcu_gp_kthread_wake(rsp);
1501 * Initialize a new grace period. Return 0 if no grace period required.
1503 static int rcu_gp_init(struct rcu_state *rsp)
1505 struct rcu_data *rdp;
1506 struct rcu_node *rnp = rcu_get_root(rsp);
1508 rcu_bind_gp_kthread();
1509 raw_spin_lock_irq(&rnp->lock);
1510 smp_mb__after_unlock_lock();
1511 if (!ACCESS_ONCE(rsp->gp_flags)) {
1512 /* Spurious wakeup, tell caller to go back to sleep. */
1513 raw_spin_unlock_irq(&rnp->lock);
1514 return 0;
1516 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1518 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1520 * Grace period already in progress, don't start another.
1521 * Not supposed to be able to happen.
1523 raw_spin_unlock_irq(&rnp->lock);
1524 return 0;
1527 /* Advance to a new grace period and initialize state. */
1528 record_gp_stall_check_time(rsp);
1529 /* Record GP times before starting GP, hence smp_store_release(). */
1530 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1531 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1532 raw_spin_unlock_irq(&rnp->lock);
1534 /* Exclude any concurrent CPU-hotplug operations. */
1535 mutex_lock(&rsp->onoff_mutex);
1536 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1539 * Set the quiescent-state-needed bits in all the rcu_node
1540 * structures for all currently online CPUs in breadth-first order,
1541 * starting from the root rcu_node structure, relying on the layout
1542 * of the tree within the rsp->node[] array. Note that other CPUs
1543 * will access only the leaves of the hierarchy, thus seeing that no
1544 * grace period is in progress, at least until the corresponding
1545 * leaf node has been initialized. In addition, we have excluded
1546 * CPU-hotplug operations.
1548 * The grace period cannot complete until the initialization
1549 * process finishes, because this kthread handles both.
1551 rcu_for_each_node_breadth_first(rsp, rnp) {
1552 raw_spin_lock_irq(&rnp->lock);
1553 smp_mb__after_unlock_lock();
1554 rdp = this_cpu_ptr(rsp->rda);
1555 rcu_preempt_check_blocked_tasks(rnp);
1556 rnp->qsmask = rnp->qsmaskinit;
1557 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1558 WARN_ON_ONCE(rnp->completed != rsp->completed);
1559 ACCESS_ONCE(rnp->completed) = rsp->completed;
1560 if (rnp == rdp->mynode)
1561 (void)__note_gp_changes(rsp, rnp, rdp);
1562 rcu_preempt_boost_start_gp(rnp);
1563 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1564 rnp->level, rnp->grplo,
1565 rnp->grphi, rnp->qsmask);
1566 raw_spin_unlock_irq(&rnp->lock);
1567 #ifdef CONFIG_PROVE_RCU_DELAY
1568 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1569 system_state == SYSTEM_RUNNING)
1570 udelay(200);
1571 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1572 cond_resched();
1575 mutex_unlock(&rsp->onoff_mutex);
1576 return 1;
1580 * Do one round of quiescent-state forcing.
1582 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1584 int fqs_state = fqs_state_in;
1585 bool isidle = false;
1586 unsigned long maxj;
1587 struct rcu_node *rnp = rcu_get_root(rsp);
1589 rsp->n_force_qs++;
1590 if (fqs_state == RCU_SAVE_DYNTICK) {
1591 /* Collect dyntick-idle snapshots. */
1592 if (is_sysidle_rcu_state(rsp)) {
1593 isidle = 1;
1594 maxj = jiffies - ULONG_MAX / 4;
1596 force_qs_rnp(rsp, dyntick_save_progress_counter,
1597 &isidle, &maxj);
1598 rcu_sysidle_report_gp(rsp, isidle, maxj);
1599 fqs_state = RCU_FORCE_QS;
1600 } else {
1601 /* Handle dyntick-idle and offline CPUs. */
1602 isidle = 0;
1603 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1605 /* Clear flag to prevent immediate re-entry. */
1606 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1607 raw_spin_lock_irq(&rnp->lock);
1608 smp_mb__after_unlock_lock();
1609 ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS;
1610 raw_spin_unlock_irq(&rnp->lock);
1612 return fqs_state;
1616 * Clean up after the old grace period.
1618 static void rcu_gp_cleanup(struct rcu_state *rsp)
1620 unsigned long gp_duration;
1621 bool needgp = false;
1622 int nocb = 0;
1623 struct rcu_data *rdp;
1624 struct rcu_node *rnp = rcu_get_root(rsp);
1626 raw_spin_lock_irq(&rnp->lock);
1627 smp_mb__after_unlock_lock();
1628 gp_duration = jiffies - rsp->gp_start;
1629 if (gp_duration > rsp->gp_max)
1630 rsp->gp_max = gp_duration;
1633 * We know the grace period is complete, but to everyone else
1634 * it appears to still be ongoing. But it is also the case
1635 * that to everyone else it looks like there is nothing that
1636 * they can do to advance the grace period. It is therefore
1637 * safe for us to drop the lock in order to mark the grace
1638 * period as completed in all of the rcu_node structures.
1640 raw_spin_unlock_irq(&rnp->lock);
1643 * Propagate new ->completed value to rcu_node structures so
1644 * that other CPUs don't have to wait until the start of the next
1645 * grace period to process their callbacks. This also avoids
1646 * some nasty RCU grace-period initialization races by forcing
1647 * the end of the current grace period to be completely recorded in
1648 * all of the rcu_node structures before the beginning of the next
1649 * grace period is recorded in any of the rcu_node structures.
1651 rcu_for_each_node_breadth_first(rsp, rnp) {
1652 raw_spin_lock_irq(&rnp->lock);
1653 smp_mb__after_unlock_lock();
1654 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1655 rdp = this_cpu_ptr(rsp->rda);
1656 if (rnp == rdp->mynode)
1657 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
1658 /* smp_mb() provided by prior unlock-lock pair. */
1659 nocb += rcu_future_gp_cleanup(rsp, rnp);
1660 raw_spin_unlock_irq(&rnp->lock);
1661 cond_resched();
1663 rnp = rcu_get_root(rsp);
1664 raw_spin_lock_irq(&rnp->lock);
1665 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1666 rcu_nocb_gp_set(rnp, nocb);
1668 /* Declare grace period done. */
1669 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1670 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1671 rsp->fqs_state = RCU_GP_IDLE;
1672 rdp = this_cpu_ptr(rsp->rda);
1673 /* Advance CBs to reduce false positives below. */
1674 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
1675 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
1676 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1677 trace_rcu_grace_period(rsp->name,
1678 ACCESS_ONCE(rsp->gpnum),
1679 TPS("newreq"));
1681 raw_spin_unlock_irq(&rnp->lock);
1685 * Body of kthread that handles grace periods.
1687 static int __noreturn rcu_gp_kthread(void *arg)
1689 int fqs_state;
1690 int gf;
1691 unsigned long j;
1692 int ret;
1693 struct rcu_state *rsp = arg;
1694 struct rcu_node *rnp = rcu_get_root(rsp);
1696 for (;;) {
1698 /* Handle grace-period start. */
1699 for (;;) {
1700 trace_rcu_grace_period(rsp->name,
1701 ACCESS_ONCE(rsp->gpnum),
1702 TPS("reqwait"));
1703 rsp->gp_state = RCU_GP_WAIT_GPS;
1704 wait_event_interruptible(rsp->gp_wq,
1705 ACCESS_ONCE(rsp->gp_flags) &
1706 RCU_GP_FLAG_INIT);
1707 /* Locking provides needed memory barrier. */
1708 if (rcu_gp_init(rsp))
1709 break;
1710 cond_resched();
1711 flush_signals(current);
1712 trace_rcu_grace_period(rsp->name,
1713 ACCESS_ONCE(rsp->gpnum),
1714 TPS("reqwaitsig"));
1717 /* Handle quiescent-state forcing. */
1718 fqs_state = RCU_SAVE_DYNTICK;
1719 j = jiffies_till_first_fqs;
1720 if (j > HZ) {
1721 j = HZ;
1722 jiffies_till_first_fqs = HZ;
1724 ret = 0;
1725 for (;;) {
1726 if (!ret)
1727 rsp->jiffies_force_qs = jiffies + j;
1728 trace_rcu_grace_period(rsp->name,
1729 ACCESS_ONCE(rsp->gpnum),
1730 TPS("fqswait"));
1731 rsp->gp_state = RCU_GP_WAIT_FQS;
1732 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1733 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1734 RCU_GP_FLAG_FQS) ||
1735 (!ACCESS_ONCE(rnp->qsmask) &&
1736 !rcu_preempt_blocked_readers_cgp(rnp)),
1738 /* Locking provides needed memory barriers. */
1739 /* If grace period done, leave loop. */
1740 if (!ACCESS_ONCE(rnp->qsmask) &&
1741 !rcu_preempt_blocked_readers_cgp(rnp))
1742 break;
1743 /* If time for quiescent-state forcing, do it. */
1744 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1745 (gf & RCU_GP_FLAG_FQS)) {
1746 trace_rcu_grace_period(rsp->name,
1747 ACCESS_ONCE(rsp->gpnum),
1748 TPS("fqsstart"));
1749 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1750 trace_rcu_grace_period(rsp->name,
1751 ACCESS_ONCE(rsp->gpnum),
1752 TPS("fqsend"));
1753 cond_resched();
1754 } else {
1755 /* Deal with stray signal. */
1756 cond_resched();
1757 flush_signals(current);
1758 trace_rcu_grace_period(rsp->name,
1759 ACCESS_ONCE(rsp->gpnum),
1760 TPS("fqswaitsig"));
1762 j = jiffies_till_next_fqs;
1763 if (j > HZ) {
1764 j = HZ;
1765 jiffies_till_next_fqs = HZ;
1766 } else if (j < 1) {
1767 j = 1;
1768 jiffies_till_next_fqs = 1;
1772 /* Handle grace-period end. */
1773 rcu_gp_cleanup(rsp);
1778 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1779 * in preparation for detecting the next grace period. The caller must hold
1780 * the root node's ->lock and hard irqs must be disabled.
1782 * Note that it is legal for a dying CPU (which is marked as offline) to
1783 * invoke this function. This can happen when the dying CPU reports its
1784 * quiescent state.
1786 * Returns true if the grace-period kthread must be awakened.
1788 static bool
1789 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1790 struct rcu_data *rdp)
1792 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1794 * Either we have not yet spawned the grace-period
1795 * task, this CPU does not need another grace period,
1796 * or a grace period is already in progress.
1797 * Either way, don't start a new grace period.
1799 return false;
1801 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1802 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1803 TPS("newreq"));
1806 * We can't do wakeups while holding the rnp->lock, as that
1807 * could cause possible deadlocks with the rq->lock. Defer
1808 * the wakeup to our caller.
1810 return true;
1814 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1815 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1816 * is invoked indirectly from rcu_advance_cbs(), which would result in
1817 * endless recursion -- or would do so if it wasn't for the self-deadlock
1818 * that is encountered beforehand.
1820 * Returns true if the grace-period kthread needs to be awakened.
1822 static bool rcu_start_gp(struct rcu_state *rsp)
1824 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1825 struct rcu_node *rnp = rcu_get_root(rsp);
1826 bool ret = false;
1829 * If there is no grace period in progress right now, any
1830 * callbacks we have up to this point will be satisfied by the
1831 * next grace period. Also, advancing the callbacks reduces the
1832 * probability of false positives from cpu_needs_another_gp()
1833 * resulting in pointless grace periods. So, advance callbacks
1834 * then start the grace period!
1836 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
1837 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
1838 return ret;
1842 * Report a full set of quiescent states to the specified rcu_state
1843 * data structure. This involves cleaning up after the prior grace
1844 * period and letting rcu_start_gp() start up the next grace period
1845 * if one is needed. Note that the caller must hold rnp->lock, which
1846 * is released before return.
1848 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1849 __releases(rcu_get_root(rsp)->lock)
1851 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1852 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1853 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1857 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1858 * Allows quiescent states for a group of CPUs to be reported at one go
1859 * to the specified rcu_node structure, though all the CPUs in the group
1860 * must be represented by the same rcu_node structure (which need not be
1861 * a leaf rcu_node structure, though it often will be). That structure's
1862 * lock must be held upon entry, and it is released before return.
1864 static void
1865 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1866 struct rcu_node *rnp, unsigned long flags)
1867 __releases(rnp->lock)
1869 struct rcu_node *rnp_c;
1871 /* Walk up the rcu_node hierarchy. */
1872 for (;;) {
1873 if (!(rnp->qsmask & mask)) {
1875 /* Our bit has already been cleared, so done. */
1876 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1877 return;
1879 rnp->qsmask &= ~mask;
1880 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1881 mask, rnp->qsmask, rnp->level,
1882 rnp->grplo, rnp->grphi,
1883 !!rnp->gp_tasks);
1884 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1886 /* Other bits still set at this level, so done. */
1887 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1888 return;
1890 mask = rnp->grpmask;
1891 if (rnp->parent == NULL) {
1893 /* No more levels. Exit loop holding root lock. */
1895 break;
1897 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1898 rnp_c = rnp;
1899 rnp = rnp->parent;
1900 raw_spin_lock_irqsave(&rnp->lock, flags);
1901 smp_mb__after_unlock_lock();
1902 WARN_ON_ONCE(rnp_c->qsmask);
1906 * Get here if we are the last CPU to pass through a quiescent
1907 * state for this grace period. Invoke rcu_report_qs_rsp()
1908 * to clean up and start the next grace period if one is needed.
1910 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1914 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1915 * structure. This must be either called from the specified CPU, or
1916 * called when the specified CPU is known to be offline (and when it is
1917 * also known that no other CPU is concurrently trying to help the offline
1918 * CPU). The lastcomp argument is used to make sure we are still in the
1919 * grace period of interest. We don't want to end the current grace period
1920 * based on quiescent states detected in an earlier grace period!
1922 static void
1923 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1925 unsigned long flags;
1926 unsigned long mask;
1927 bool needwake;
1928 struct rcu_node *rnp;
1930 rnp = rdp->mynode;
1931 raw_spin_lock_irqsave(&rnp->lock, flags);
1932 smp_mb__after_unlock_lock();
1933 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1934 rnp->completed == rnp->gpnum) {
1937 * The grace period in which this quiescent state was
1938 * recorded has ended, so don't report it upwards.
1939 * We will instead need a new quiescent state that lies
1940 * within the current grace period.
1942 rdp->passed_quiesce = 0; /* need qs for new gp. */
1943 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1944 return;
1946 mask = rdp->grpmask;
1947 if ((rnp->qsmask & mask) == 0) {
1948 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1949 } else {
1950 rdp->qs_pending = 0;
1953 * This GP can't end until cpu checks in, so all of our
1954 * callbacks can be processed during the next GP.
1956 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1958 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1959 if (needwake)
1960 rcu_gp_kthread_wake(rsp);
1965 * Check to see if there is a new grace period of which this CPU
1966 * is not yet aware, and if so, set up local rcu_data state for it.
1967 * Otherwise, see if this CPU has just passed through its first
1968 * quiescent state for this grace period, and record that fact if so.
1970 static void
1971 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1973 /* Check for grace-period ends and beginnings. */
1974 note_gp_changes(rsp, rdp);
1977 * Does this CPU still need to do its part for current grace period?
1978 * If no, return and let the other CPUs do their part as well.
1980 if (!rdp->qs_pending)
1981 return;
1984 * Was there a quiescent state since the beginning of the grace
1985 * period? If no, then exit and wait for the next call.
1987 if (!rdp->passed_quiesce)
1988 return;
1991 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1992 * judge of that).
1994 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1997 #ifdef CONFIG_HOTPLUG_CPU
2000 * Send the specified CPU's RCU callbacks to the orphanage. The
2001 * specified CPU must be offline, and the caller must hold the
2002 * ->orphan_lock.
2004 static void
2005 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2006 struct rcu_node *rnp, struct rcu_data *rdp)
2008 /* No-CBs CPUs do not have orphanable callbacks. */
2009 if (rcu_is_nocb_cpu(rdp->cpu))
2010 return;
2013 * Orphan the callbacks. First adjust the counts. This is safe
2014 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2015 * cannot be running now. Thus no memory barrier is required.
2017 if (rdp->nxtlist != NULL) {
2018 rsp->qlen_lazy += rdp->qlen_lazy;
2019 rsp->qlen += rdp->qlen;
2020 rdp->n_cbs_orphaned += rdp->qlen;
2021 rdp->qlen_lazy = 0;
2022 ACCESS_ONCE(rdp->qlen) = 0;
2026 * Next, move those callbacks still needing a grace period to
2027 * the orphanage, where some other CPU will pick them up.
2028 * Some of the callbacks might have gone partway through a grace
2029 * period, but that is too bad. They get to start over because we
2030 * cannot assume that grace periods are synchronized across CPUs.
2031 * We don't bother updating the ->nxttail[] array yet, instead
2032 * we just reset the whole thing later on.
2034 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2035 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2036 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2037 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2041 * Then move the ready-to-invoke callbacks to the orphanage,
2042 * where some other CPU will pick them up. These will not be
2043 * required to pass though another grace period: They are done.
2045 if (rdp->nxtlist != NULL) {
2046 *rsp->orphan_donetail = rdp->nxtlist;
2047 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2050 /* Finally, initialize the rcu_data structure's list to empty. */
2051 init_callback_list(rdp);
2055 * Adopt the RCU callbacks from the specified rcu_state structure's
2056 * orphanage. The caller must hold the ->orphan_lock.
2058 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2060 int i;
2061 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2063 /* No-CBs CPUs are handled specially. */
2064 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2065 return;
2067 /* Do the accounting first. */
2068 rdp->qlen_lazy += rsp->qlen_lazy;
2069 rdp->qlen += rsp->qlen;
2070 rdp->n_cbs_adopted += rsp->qlen;
2071 if (rsp->qlen_lazy != rsp->qlen)
2072 rcu_idle_count_callbacks_posted();
2073 rsp->qlen_lazy = 0;
2074 rsp->qlen = 0;
2077 * We do not need a memory barrier here because the only way we
2078 * can get here if there is an rcu_barrier() in flight is if
2079 * we are the task doing the rcu_barrier().
2082 /* First adopt the ready-to-invoke callbacks. */
2083 if (rsp->orphan_donelist != NULL) {
2084 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2085 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2086 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2087 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2088 rdp->nxttail[i] = rsp->orphan_donetail;
2089 rsp->orphan_donelist = NULL;
2090 rsp->orphan_donetail = &rsp->orphan_donelist;
2093 /* And then adopt the callbacks that still need a grace period. */
2094 if (rsp->orphan_nxtlist != NULL) {
2095 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2096 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2097 rsp->orphan_nxtlist = NULL;
2098 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2103 * Trace the fact that this CPU is going offline.
2105 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2107 RCU_TRACE(unsigned long mask);
2108 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2109 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2111 RCU_TRACE(mask = rdp->grpmask);
2112 trace_rcu_grace_period(rsp->name,
2113 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2114 TPS("cpuofl"));
2118 * The CPU has been completely removed, and some other CPU is reporting
2119 * this fact from process context. Do the remainder of the cleanup,
2120 * including orphaning the outgoing CPU's RCU callbacks, and also
2121 * adopting them. There can only be one CPU hotplug operation at a time,
2122 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2124 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2126 unsigned long flags;
2127 unsigned long mask;
2128 int need_report = 0;
2129 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2130 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2132 /* Adjust any no-longer-needed kthreads. */
2133 rcu_boost_kthread_setaffinity(rnp, -1);
2135 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2137 /* Exclude any attempts to start a new grace period. */
2138 mutex_lock(&rsp->onoff_mutex);
2139 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2141 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2142 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2143 rcu_adopt_orphan_cbs(rsp, flags);
2145 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2146 mask = rdp->grpmask; /* rnp->grplo is constant. */
2147 do {
2148 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2149 smp_mb__after_unlock_lock();
2150 rnp->qsmaskinit &= ~mask;
2151 if (rnp->qsmaskinit != 0) {
2152 if (rnp != rdp->mynode)
2153 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2154 break;
2156 if (rnp == rdp->mynode)
2157 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2158 else
2159 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2160 mask = rnp->grpmask;
2161 rnp = rnp->parent;
2162 } while (rnp != NULL);
2165 * We still hold the leaf rcu_node structure lock here, and
2166 * irqs are still disabled. The reason for this subterfuge is
2167 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2168 * held leads to deadlock.
2170 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2171 rnp = rdp->mynode;
2172 if (need_report & RCU_OFL_TASKS_NORM_GP)
2173 rcu_report_unblock_qs_rnp(rnp, flags);
2174 else
2175 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2176 if (need_report & RCU_OFL_TASKS_EXP_GP)
2177 rcu_report_exp_rnp(rsp, rnp, true);
2178 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2179 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2180 cpu, rdp->qlen, rdp->nxtlist);
2181 init_callback_list(rdp);
2182 /* Disallow further callbacks on this CPU. */
2183 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2184 mutex_unlock(&rsp->onoff_mutex);
2187 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2189 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2193 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2197 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2200 * Invoke any RCU callbacks that have made it to the end of their grace
2201 * period. Thottle as specified by rdp->blimit.
2203 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2205 unsigned long flags;
2206 struct rcu_head *next, *list, **tail;
2207 long bl, count, count_lazy;
2208 int i;
2210 /* If no callbacks are ready, just return. */
2211 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2212 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2213 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2214 need_resched(), is_idle_task(current),
2215 rcu_is_callbacks_kthread());
2216 return;
2220 * Extract the list of ready callbacks, disabling to prevent
2221 * races with call_rcu() from interrupt handlers.
2223 local_irq_save(flags);
2224 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2225 bl = rdp->blimit;
2226 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2227 list = rdp->nxtlist;
2228 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2229 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2230 tail = rdp->nxttail[RCU_DONE_TAIL];
2231 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2232 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2233 rdp->nxttail[i] = &rdp->nxtlist;
2234 local_irq_restore(flags);
2236 /* Invoke callbacks. */
2237 count = count_lazy = 0;
2238 while (list) {
2239 next = list->next;
2240 prefetch(next);
2241 debug_rcu_head_unqueue(list);
2242 if (__rcu_reclaim(rsp->name, list))
2243 count_lazy++;
2244 list = next;
2245 /* Stop only if limit reached and CPU has something to do. */
2246 if (++count >= bl &&
2247 (need_resched() ||
2248 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2249 break;
2252 local_irq_save(flags);
2253 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2254 is_idle_task(current),
2255 rcu_is_callbacks_kthread());
2257 /* Update count, and requeue any remaining callbacks. */
2258 if (list != NULL) {
2259 *tail = rdp->nxtlist;
2260 rdp->nxtlist = list;
2261 for (i = 0; i < RCU_NEXT_SIZE; i++)
2262 if (&rdp->nxtlist == rdp->nxttail[i])
2263 rdp->nxttail[i] = tail;
2264 else
2265 break;
2267 smp_mb(); /* List handling before counting for rcu_barrier(). */
2268 rdp->qlen_lazy -= count_lazy;
2269 ACCESS_ONCE(rdp->qlen) -= count;
2270 rdp->n_cbs_invoked += count;
2272 /* Reinstate batch limit if we have worked down the excess. */
2273 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2274 rdp->blimit = blimit;
2276 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2277 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2278 rdp->qlen_last_fqs_check = 0;
2279 rdp->n_force_qs_snap = rsp->n_force_qs;
2280 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2281 rdp->qlen_last_fqs_check = rdp->qlen;
2282 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2284 local_irq_restore(flags);
2286 /* Re-invoke RCU core processing if there are callbacks remaining. */
2287 if (cpu_has_callbacks_ready_to_invoke(rdp))
2288 invoke_rcu_core();
2292 * Check to see if this CPU is in a non-context-switch quiescent state
2293 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2294 * Also schedule RCU core processing.
2296 * This function must be called from hardirq context. It is normally
2297 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2298 * false, there is no point in invoking rcu_check_callbacks().
2300 void rcu_check_callbacks(int cpu, int user)
2302 trace_rcu_utilization(TPS("Start scheduler-tick"));
2303 increment_cpu_stall_ticks();
2304 if (user || rcu_is_cpu_rrupt_from_idle()) {
2307 * Get here if this CPU took its interrupt from user
2308 * mode or from the idle loop, and if this is not a
2309 * nested interrupt. In this case, the CPU is in
2310 * a quiescent state, so note it.
2312 * No memory barrier is required here because both
2313 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2314 * variables that other CPUs neither access nor modify,
2315 * at least not while the corresponding CPU is online.
2318 rcu_sched_qs(cpu);
2319 rcu_bh_qs(cpu);
2321 } else if (!in_softirq()) {
2324 * Get here if this CPU did not take its interrupt from
2325 * softirq, in other words, if it is not interrupting
2326 * a rcu_bh read-side critical section. This is an _bh
2327 * critical section, so note it.
2330 rcu_bh_qs(cpu);
2332 rcu_preempt_check_callbacks(cpu);
2333 if (rcu_pending(cpu))
2334 invoke_rcu_core();
2335 trace_rcu_utilization(TPS("End scheduler-tick"));
2339 * Scan the leaf rcu_node structures, processing dyntick state for any that
2340 * have not yet encountered a quiescent state, using the function specified.
2341 * Also initiate boosting for any threads blocked on the root rcu_node.
2343 * The caller must have suppressed start of new grace periods.
2345 static void force_qs_rnp(struct rcu_state *rsp,
2346 int (*f)(struct rcu_data *rsp, bool *isidle,
2347 unsigned long *maxj),
2348 bool *isidle, unsigned long *maxj)
2350 unsigned long bit;
2351 int cpu;
2352 unsigned long flags;
2353 unsigned long mask;
2354 struct rcu_node *rnp;
2356 rcu_for_each_leaf_node(rsp, rnp) {
2357 cond_resched();
2358 mask = 0;
2359 raw_spin_lock_irqsave(&rnp->lock, flags);
2360 smp_mb__after_unlock_lock();
2361 if (!rcu_gp_in_progress(rsp)) {
2362 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2363 return;
2365 if (rnp->qsmask == 0) {
2366 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2367 continue;
2369 cpu = rnp->grplo;
2370 bit = 1;
2371 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2372 if ((rnp->qsmask & bit) != 0) {
2373 if ((rnp->qsmaskinit & bit) != 0)
2374 *isidle = 0;
2375 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2376 mask |= bit;
2379 if (mask != 0) {
2381 /* rcu_report_qs_rnp() releases rnp->lock. */
2382 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2383 continue;
2385 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2387 rnp = rcu_get_root(rsp);
2388 if (rnp->qsmask == 0) {
2389 raw_spin_lock_irqsave(&rnp->lock, flags);
2390 smp_mb__after_unlock_lock();
2391 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2396 * Force quiescent states on reluctant CPUs, and also detect which
2397 * CPUs are in dyntick-idle mode.
2399 static void force_quiescent_state(struct rcu_state *rsp)
2401 unsigned long flags;
2402 bool ret;
2403 struct rcu_node *rnp;
2404 struct rcu_node *rnp_old = NULL;
2406 /* Funnel through hierarchy to reduce memory contention. */
2407 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2408 for (; rnp != NULL; rnp = rnp->parent) {
2409 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2410 !raw_spin_trylock(&rnp->fqslock);
2411 if (rnp_old != NULL)
2412 raw_spin_unlock(&rnp_old->fqslock);
2413 if (ret) {
2414 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2415 return;
2417 rnp_old = rnp;
2419 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2421 /* Reached the root of the rcu_node tree, acquire lock. */
2422 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2423 smp_mb__after_unlock_lock();
2424 raw_spin_unlock(&rnp_old->fqslock);
2425 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2426 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2427 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2428 return; /* Someone beat us to it. */
2430 ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS;
2431 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2432 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2436 * This does the RCU core processing work for the specified rcu_state
2437 * and rcu_data structures. This may be called only from the CPU to
2438 * whom the rdp belongs.
2440 static void
2441 __rcu_process_callbacks(struct rcu_state *rsp)
2443 unsigned long flags;
2444 bool needwake;
2445 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2447 WARN_ON_ONCE(rdp->beenonline == 0);
2449 /* Update RCU state based on any recent quiescent states. */
2450 rcu_check_quiescent_state(rsp, rdp);
2452 /* Does this CPU require a not-yet-started grace period? */
2453 local_irq_save(flags);
2454 if (cpu_needs_another_gp(rsp, rdp)) {
2455 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2456 needwake = rcu_start_gp(rsp);
2457 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2458 if (needwake)
2459 rcu_gp_kthread_wake(rsp);
2460 } else {
2461 local_irq_restore(flags);
2464 /* If there are callbacks ready, invoke them. */
2465 if (cpu_has_callbacks_ready_to_invoke(rdp))
2466 invoke_rcu_callbacks(rsp, rdp);
2468 /* Do any needed deferred wakeups of rcuo kthreads. */
2469 do_nocb_deferred_wakeup(rdp);
2473 * Do RCU core processing for the current CPU.
2475 static void rcu_process_callbacks(struct softirq_action *unused)
2477 struct rcu_state *rsp;
2479 if (cpu_is_offline(smp_processor_id()))
2480 return;
2481 trace_rcu_utilization(TPS("Start RCU core"));
2482 for_each_rcu_flavor(rsp)
2483 __rcu_process_callbacks(rsp);
2484 trace_rcu_utilization(TPS("End RCU core"));
2488 * Schedule RCU callback invocation. If the specified type of RCU
2489 * does not support RCU priority boosting, just do a direct call,
2490 * otherwise wake up the per-CPU kernel kthread. Note that because we
2491 * are running on the current CPU with interrupts disabled, the
2492 * rcu_cpu_kthread_task cannot disappear out from under us.
2494 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2496 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2497 return;
2498 if (likely(!rsp->boost)) {
2499 rcu_do_batch(rsp, rdp);
2500 return;
2502 invoke_rcu_callbacks_kthread();
2505 static void invoke_rcu_core(void)
2507 if (cpu_online(smp_processor_id()))
2508 raise_softirq(RCU_SOFTIRQ);
2512 * Handle any core-RCU processing required by a call_rcu() invocation.
2514 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2515 struct rcu_head *head, unsigned long flags)
2517 bool needwake;
2520 * If called from an extended quiescent state, invoke the RCU
2521 * core in order to force a re-evaluation of RCU's idleness.
2523 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2524 invoke_rcu_core();
2526 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2527 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2528 return;
2531 * Force the grace period if too many callbacks or too long waiting.
2532 * Enforce hysteresis, and don't invoke force_quiescent_state()
2533 * if some other CPU has recently done so. Also, don't bother
2534 * invoking force_quiescent_state() if the newly enqueued callback
2535 * is the only one waiting for a grace period to complete.
2537 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2539 /* Are we ignoring a completed grace period? */
2540 note_gp_changes(rsp, rdp);
2542 /* Start a new grace period if one not already started. */
2543 if (!rcu_gp_in_progress(rsp)) {
2544 struct rcu_node *rnp_root = rcu_get_root(rsp);
2546 raw_spin_lock(&rnp_root->lock);
2547 smp_mb__after_unlock_lock();
2548 needwake = rcu_start_gp(rsp);
2549 raw_spin_unlock(&rnp_root->lock);
2550 if (needwake)
2551 rcu_gp_kthread_wake(rsp);
2552 } else {
2553 /* Give the grace period a kick. */
2554 rdp->blimit = LONG_MAX;
2555 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2556 *rdp->nxttail[RCU_DONE_TAIL] != head)
2557 force_quiescent_state(rsp);
2558 rdp->n_force_qs_snap = rsp->n_force_qs;
2559 rdp->qlen_last_fqs_check = rdp->qlen;
2565 * RCU callback function to leak a callback.
2567 static void rcu_leak_callback(struct rcu_head *rhp)
2572 * Helper function for call_rcu() and friends. The cpu argument will
2573 * normally be -1, indicating "currently running CPU". It may specify
2574 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2575 * is expected to specify a CPU.
2577 static void
2578 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2579 struct rcu_state *rsp, int cpu, bool lazy)
2581 unsigned long flags;
2582 struct rcu_data *rdp;
2584 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2585 if (debug_rcu_head_queue(head)) {
2586 /* Probable double call_rcu(), so leak the callback. */
2587 ACCESS_ONCE(head->func) = rcu_leak_callback;
2588 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2589 return;
2591 head->func = func;
2592 head->next = NULL;
2595 * Opportunistically note grace-period endings and beginnings.
2596 * Note that we might see a beginning right after we see an
2597 * end, but never vice versa, since this CPU has to pass through
2598 * a quiescent state betweentimes.
2600 local_irq_save(flags);
2601 rdp = this_cpu_ptr(rsp->rda);
2603 /* Add the callback to our list. */
2604 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2605 int offline;
2607 if (cpu != -1)
2608 rdp = per_cpu_ptr(rsp->rda, cpu);
2609 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2610 WARN_ON_ONCE(offline);
2611 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2612 local_irq_restore(flags);
2613 return;
2615 ACCESS_ONCE(rdp->qlen)++;
2616 if (lazy)
2617 rdp->qlen_lazy++;
2618 else
2619 rcu_idle_count_callbacks_posted();
2620 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2621 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2622 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2624 if (__is_kfree_rcu_offset((unsigned long)func))
2625 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2626 rdp->qlen_lazy, rdp->qlen);
2627 else
2628 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2630 /* Go handle any RCU core processing required. */
2631 __call_rcu_core(rsp, rdp, head, flags);
2632 local_irq_restore(flags);
2636 * Queue an RCU-sched callback for invocation after a grace period.
2638 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2640 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2642 EXPORT_SYMBOL_GPL(call_rcu_sched);
2645 * Queue an RCU callback for invocation after a quicker grace period.
2647 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2649 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2651 EXPORT_SYMBOL_GPL(call_rcu_bh);
2654 * Queue an RCU callback for lazy invocation after a grace period.
2655 * This will likely be later named something like "call_rcu_lazy()",
2656 * but this change will require some way of tagging the lazy RCU
2657 * callbacks in the list of pending callbacks. Until then, this
2658 * function may only be called from __kfree_rcu().
2660 void kfree_call_rcu(struct rcu_head *head,
2661 void (*func)(struct rcu_head *rcu))
2663 __call_rcu(head, func, rcu_state_p, -1, 1);
2665 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2668 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2669 * any blocking grace-period wait automatically implies a grace period
2670 * if there is only one CPU online at any point time during execution
2671 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2672 * occasionally incorrectly indicate that there are multiple CPUs online
2673 * when there was in fact only one the whole time, as this just adds
2674 * some overhead: RCU still operates correctly.
2676 static inline int rcu_blocking_is_gp(void)
2678 int ret;
2680 might_sleep(); /* Check for RCU read-side critical section. */
2681 preempt_disable();
2682 ret = num_online_cpus() <= 1;
2683 preempt_enable();
2684 return ret;
2688 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2690 * Control will return to the caller some time after a full rcu-sched
2691 * grace period has elapsed, in other words after all currently executing
2692 * rcu-sched read-side critical sections have completed. These read-side
2693 * critical sections are delimited by rcu_read_lock_sched() and
2694 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2695 * local_irq_disable(), and so on may be used in place of
2696 * rcu_read_lock_sched().
2698 * This means that all preempt_disable code sequences, including NMI and
2699 * non-threaded hardware-interrupt handlers, in progress on entry will
2700 * have completed before this primitive returns. However, this does not
2701 * guarantee that softirq handlers will have completed, since in some
2702 * kernels, these handlers can run in process context, and can block.
2704 * Note that this guarantee implies further memory-ordering guarantees.
2705 * On systems with more than one CPU, when synchronize_sched() returns,
2706 * each CPU is guaranteed to have executed a full memory barrier since the
2707 * end of its last RCU-sched read-side critical section whose beginning
2708 * preceded the call to synchronize_sched(). In addition, each CPU having
2709 * an RCU read-side critical section that extends beyond the return from
2710 * synchronize_sched() is guaranteed to have executed a full memory barrier
2711 * after the beginning of synchronize_sched() and before the beginning of
2712 * that RCU read-side critical section. Note that these guarantees include
2713 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2714 * that are executing in the kernel.
2716 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2717 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2718 * to have executed a full memory barrier during the execution of
2719 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2720 * again only if the system has more than one CPU).
2722 * This primitive provides the guarantees made by the (now removed)
2723 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2724 * guarantees that rcu_read_lock() sections will have completed.
2725 * In "classic RCU", these two guarantees happen to be one and
2726 * the same, but can differ in realtime RCU implementations.
2728 void synchronize_sched(void)
2730 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2731 !lock_is_held(&rcu_lock_map) &&
2732 !lock_is_held(&rcu_sched_lock_map),
2733 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2734 if (rcu_blocking_is_gp())
2735 return;
2736 if (rcu_expedited)
2737 synchronize_sched_expedited();
2738 else
2739 wait_rcu_gp(call_rcu_sched);
2741 EXPORT_SYMBOL_GPL(synchronize_sched);
2744 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2746 * Control will return to the caller some time after a full rcu_bh grace
2747 * period has elapsed, in other words after all currently executing rcu_bh
2748 * read-side critical sections have completed. RCU read-side critical
2749 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2750 * and may be nested.
2752 * See the description of synchronize_sched() for more detailed information
2753 * on memory ordering guarantees.
2755 void synchronize_rcu_bh(void)
2757 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2758 !lock_is_held(&rcu_lock_map) &&
2759 !lock_is_held(&rcu_sched_lock_map),
2760 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2761 if (rcu_blocking_is_gp())
2762 return;
2763 if (rcu_expedited)
2764 synchronize_rcu_bh_expedited();
2765 else
2766 wait_rcu_gp(call_rcu_bh);
2768 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2771 * get_state_synchronize_rcu - Snapshot current RCU state
2773 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2774 * to determine whether or not a full grace period has elapsed in the
2775 * meantime.
2777 unsigned long get_state_synchronize_rcu(void)
2780 * Any prior manipulation of RCU-protected data must happen
2781 * before the load from ->gpnum.
2783 smp_mb(); /* ^^^ */
2786 * Make sure this load happens before the purportedly
2787 * time-consuming work between get_state_synchronize_rcu()
2788 * and cond_synchronize_rcu().
2790 return smp_load_acquire(&rcu_state_p->gpnum);
2792 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2795 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2797 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2799 * If a full RCU grace period has elapsed since the earlier call to
2800 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2801 * synchronize_rcu() to wait for a full grace period.
2803 * Yes, this function does not take counter wrap into account. But
2804 * counter wrap is harmless. If the counter wraps, we have waited for
2805 * more than 2 billion grace periods (and way more on a 64-bit system!),
2806 * so waiting for one additional grace period should be just fine.
2808 void cond_synchronize_rcu(unsigned long oldstate)
2810 unsigned long newstate;
2813 * Ensure that this load happens before any RCU-destructive
2814 * actions the caller might carry out after we return.
2816 newstate = smp_load_acquire(&rcu_state_p->completed);
2817 if (ULONG_CMP_GE(oldstate, newstate))
2818 synchronize_rcu();
2820 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2822 static int synchronize_sched_expedited_cpu_stop(void *data)
2825 * There must be a full memory barrier on each affected CPU
2826 * between the time that try_stop_cpus() is called and the
2827 * time that it returns.
2829 * In the current initial implementation of cpu_stop, the
2830 * above condition is already met when the control reaches
2831 * this point and the following smp_mb() is not strictly
2832 * necessary. Do smp_mb() anyway for documentation and
2833 * robustness against future implementation changes.
2835 smp_mb(); /* See above comment block. */
2836 return 0;
2840 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2842 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2843 * approach to force the grace period to end quickly. This consumes
2844 * significant time on all CPUs and is unfriendly to real-time workloads,
2845 * so is thus not recommended for any sort of common-case code. In fact,
2846 * if you are using synchronize_sched_expedited() in a loop, please
2847 * restructure your code to batch your updates, and then use a single
2848 * synchronize_sched() instead.
2850 * Note that it is illegal to call this function while holding any lock
2851 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2852 * to call this function from a CPU-hotplug notifier. Failing to observe
2853 * these restriction will result in deadlock.
2855 * This implementation can be thought of as an application of ticket
2856 * locking to RCU, with sync_sched_expedited_started and
2857 * sync_sched_expedited_done taking on the roles of the halves
2858 * of the ticket-lock word. Each task atomically increments
2859 * sync_sched_expedited_started upon entry, snapshotting the old value,
2860 * then attempts to stop all the CPUs. If this succeeds, then each
2861 * CPU will have executed a context switch, resulting in an RCU-sched
2862 * grace period. We are then done, so we use atomic_cmpxchg() to
2863 * update sync_sched_expedited_done to match our snapshot -- but
2864 * only if someone else has not already advanced past our snapshot.
2866 * On the other hand, if try_stop_cpus() fails, we check the value
2867 * of sync_sched_expedited_done. If it has advanced past our
2868 * initial snapshot, then someone else must have forced a grace period
2869 * some time after we took our snapshot. In this case, our work is
2870 * done for us, and we can simply return. Otherwise, we try again,
2871 * but keep our initial snapshot for purposes of checking for someone
2872 * doing our work for us.
2874 * If we fail too many times in a row, we fall back to synchronize_sched().
2876 void synchronize_sched_expedited(void)
2878 long firstsnap, s, snap;
2879 int trycount = 0;
2880 struct rcu_state *rsp = &rcu_sched_state;
2883 * If we are in danger of counter wrap, just do synchronize_sched().
2884 * By allowing sync_sched_expedited_started to advance no more than
2885 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2886 * that more than 3.5 billion CPUs would be required to force a
2887 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2888 * course be required on a 64-bit system.
2890 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2891 (ulong)atomic_long_read(&rsp->expedited_done) +
2892 ULONG_MAX / 8)) {
2893 synchronize_sched();
2894 atomic_long_inc(&rsp->expedited_wrap);
2895 return;
2899 * Take a ticket. Note that atomic_inc_return() implies a
2900 * full memory barrier.
2902 snap = atomic_long_inc_return(&rsp->expedited_start);
2903 firstsnap = snap;
2904 get_online_cpus();
2905 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2908 * Each pass through the following loop attempts to force a
2909 * context switch on each CPU.
2911 while (try_stop_cpus(cpu_online_mask,
2912 synchronize_sched_expedited_cpu_stop,
2913 NULL) == -EAGAIN) {
2914 put_online_cpus();
2915 atomic_long_inc(&rsp->expedited_tryfail);
2917 /* Check to see if someone else did our work for us. */
2918 s = atomic_long_read(&rsp->expedited_done);
2919 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2920 /* ensure test happens before caller kfree */
2921 smp_mb__before_atomic(); /* ^^^ */
2922 atomic_long_inc(&rsp->expedited_workdone1);
2923 return;
2926 /* No joy, try again later. Or just synchronize_sched(). */
2927 if (trycount++ < 10) {
2928 udelay(trycount * num_online_cpus());
2929 } else {
2930 wait_rcu_gp(call_rcu_sched);
2931 atomic_long_inc(&rsp->expedited_normal);
2932 return;
2935 /* Recheck to see if someone else did our work for us. */
2936 s = atomic_long_read(&rsp->expedited_done);
2937 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2938 /* ensure test happens before caller kfree */
2939 smp_mb__before_atomic(); /* ^^^ */
2940 atomic_long_inc(&rsp->expedited_workdone2);
2941 return;
2945 * Refetching sync_sched_expedited_started allows later
2946 * callers to piggyback on our grace period. We retry
2947 * after they started, so our grace period works for them,
2948 * and they started after our first try, so their grace
2949 * period works for us.
2951 get_online_cpus();
2952 snap = atomic_long_read(&rsp->expedited_start);
2953 smp_mb(); /* ensure read is before try_stop_cpus(). */
2955 atomic_long_inc(&rsp->expedited_stoppedcpus);
2958 * Everyone up to our most recent fetch is covered by our grace
2959 * period. Update the counter, but only if our work is still
2960 * relevant -- which it won't be if someone who started later
2961 * than we did already did their update.
2963 do {
2964 atomic_long_inc(&rsp->expedited_done_tries);
2965 s = atomic_long_read(&rsp->expedited_done);
2966 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2967 /* ensure test happens before caller kfree */
2968 smp_mb__before_atomic(); /* ^^^ */
2969 atomic_long_inc(&rsp->expedited_done_lost);
2970 break;
2972 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2973 atomic_long_inc(&rsp->expedited_done_exit);
2975 put_online_cpus();
2977 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2980 * Check to see if there is any immediate RCU-related work to be done
2981 * by the current CPU, for the specified type of RCU, returning 1 if so.
2982 * The checks are in order of increasing expense: checks that can be
2983 * carried out against CPU-local state are performed first. However,
2984 * we must check for CPU stalls first, else we might not get a chance.
2986 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2988 struct rcu_node *rnp = rdp->mynode;
2990 rdp->n_rcu_pending++;
2992 /* Check for CPU stalls, if enabled. */
2993 check_cpu_stall(rsp, rdp);
2995 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2996 if (rcu_nohz_full_cpu(rsp))
2997 return 0;
2999 /* Is the RCU core waiting for a quiescent state from this CPU? */
3000 if (rcu_scheduler_fully_active &&
3001 rdp->qs_pending && !rdp->passed_quiesce) {
3002 rdp->n_rp_qs_pending++;
3003 } else if (rdp->qs_pending && rdp->passed_quiesce) {
3004 rdp->n_rp_report_qs++;
3005 return 1;
3008 /* Does this CPU have callbacks ready to invoke? */
3009 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3010 rdp->n_rp_cb_ready++;
3011 return 1;
3014 /* Has RCU gone idle with this CPU needing another grace period? */
3015 if (cpu_needs_another_gp(rsp, rdp)) {
3016 rdp->n_rp_cpu_needs_gp++;
3017 return 1;
3020 /* Has another RCU grace period completed? */
3021 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3022 rdp->n_rp_gp_completed++;
3023 return 1;
3026 /* Has a new RCU grace period started? */
3027 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
3028 rdp->n_rp_gp_started++;
3029 return 1;
3032 /* Does this CPU need a deferred NOCB wakeup? */
3033 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3034 rdp->n_rp_nocb_defer_wakeup++;
3035 return 1;
3038 /* nothing to do */
3039 rdp->n_rp_need_nothing++;
3040 return 0;
3044 * Check to see if there is any immediate RCU-related work to be done
3045 * by the current CPU, returning 1 if so. This function is part of the
3046 * RCU implementation; it is -not- an exported member of the RCU API.
3048 static int rcu_pending(int cpu)
3050 struct rcu_state *rsp;
3052 for_each_rcu_flavor(rsp)
3053 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
3054 return 1;
3055 return 0;
3059 * Return true if the specified CPU has any callback. If all_lazy is
3060 * non-NULL, store an indication of whether all callbacks are lazy.
3061 * (If there are no callbacks, all of them are deemed to be lazy.)
3063 static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
3065 bool al = true;
3066 bool hc = false;
3067 struct rcu_data *rdp;
3068 struct rcu_state *rsp;
3070 for_each_rcu_flavor(rsp) {
3071 rdp = per_cpu_ptr(rsp->rda, cpu);
3072 if (!rdp->nxtlist)
3073 continue;
3074 hc = true;
3075 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3076 al = false;
3077 break;
3080 if (all_lazy)
3081 *all_lazy = al;
3082 return hc;
3086 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3087 * the compiler is expected to optimize this away.
3089 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3090 int cpu, unsigned long done)
3092 trace_rcu_barrier(rsp->name, s, cpu,
3093 atomic_read(&rsp->barrier_cpu_count), done);
3097 * RCU callback function for _rcu_barrier(). If we are last, wake
3098 * up the task executing _rcu_barrier().
3100 static void rcu_barrier_callback(struct rcu_head *rhp)
3102 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3103 struct rcu_state *rsp = rdp->rsp;
3105 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3106 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
3107 complete(&rsp->barrier_completion);
3108 } else {
3109 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3114 * Called with preemption disabled, and from cross-cpu IRQ context.
3116 static void rcu_barrier_func(void *type)
3118 struct rcu_state *rsp = type;
3119 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3121 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3122 atomic_inc(&rsp->barrier_cpu_count);
3123 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3127 * Orchestrate the specified type of RCU barrier, waiting for all
3128 * RCU callbacks of the specified type to complete.
3130 static void _rcu_barrier(struct rcu_state *rsp)
3132 int cpu;
3133 struct rcu_data *rdp;
3134 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3135 unsigned long snap_done;
3137 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3139 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3140 mutex_lock(&rsp->barrier_mutex);
3143 * Ensure that all prior references, including to ->n_barrier_done,
3144 * are ordered before the _rcu_barrier() machinery.
3146 smp_mb(); /* See above block comment. */
3149 * Recheck ->n_barrier_done to see if others did our work for us.
3150 * This means checking ->n_barrier_done for an even-to-odd-to-even
3151 * transition. The "if" expression below therefore rounds the old
3152 * value up to the next even number and adds two before comparing.
3154 snap_done = rsp->n_barrier_done;
3155 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3158 * If the value in snap is odd, we needed to wait for the current
3159 * rcu_barrier() to complete, then wait for the next one, in other
3160 * words, we need the value of snap_done to be three larger than
3161 * the value of snap. On the other hand, if the value in snap is
3162 * even, we only had to wait for the next rcu_barrier() to complete,
3163 * in other words, we need the value of snap_done to be only two
3164 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3165 * this for us (thank you, Linus!).
3167 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3168 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3169 smp_mb(); /* caller's subsequent code after above check. */
3170 mutex_unlock(&rsp->barrier_mutex);
3171 return;
3175 * Increment ->n_barrier_done to avoid duplicate work. Use
3176 * ACCESS_ONCE() to prevent the compiler from speculating
3177 * the increment to precede the early-exit check.
3179 ACCESS_ONCE(rsp->n_barrier_done)++;
3180 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3181 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3182 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3185 * Initialize the count to one rather than to zero in order to
3186 * avoid a too-soon return to zero in case of a short grace period
3187 * (or preemption of this task). Exclude CPU-hotplug operations
3188 * to ensure that no offline CPU has callbacks queued.
3190 init_completion(&rsp->barrier_completion);
3191 atomic_set(&rsp->barrier_cpu_count, 1);
3192 get_online_cpus();
3195 * Force each CPU with callbacks to register a new callback.
3196 * When that callback is invoked, we will know that all of the
3197 * corresponding CPU's preceding callbacks have been invoked.
3199 for_each_possible_cpu(cpu) {
3200 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3201 continue;
3202 rdp = per_cpu_ptr(rsp->rda, cpu);
3203 if (rcu_is_nocb_cpu(cpu)) {
3204 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3205 rsp->n_barrier_done);
3206 atomic_inc(&rsp->barrier_cpu_count);
3207 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3208 rsp, cpu, 0);
3209 } else if (ACCESS_ONCE(rdp->qlen)) {
3210 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3211 rsp->n_barrier_done);
3212 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3213 } else {
3214 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3215 rsp->n_barrier_done);
3218 put_online_cpus();
3221 * Now that we have an rcu_barrier_callback() callback on each
3222 * CPU, and thus each counted, remove the initial count.
3224 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3225 complete(&rsp->barrier_completion);
3227 /* Increment ->n_barrier_done to prevent duplicate work. */
3228 smp_mb(); /* Keep increment after above mechanism. */
3229 ACCESS_ONCE(rsp->n_barrier_done)++;
3230 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3231 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3232 smp_mb(); /* Keep increment before caller's subsequent code. */
3234 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3235 wait_for_completion(&rsp->barrier_completion);
3237 /* Other rcu_barrier() invocations can now safely proceed. */
3238 mutex_unlock(&rsp->barrier_mutex);
3242 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3244 void rcu_barrier_bh(void)
3246 _rcu_barrier(&rcu_bh_state);
3248 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3251 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3253 void rcu_barrier_sched(void)
3255 _rcu_barrier(&rcu_sched_state);
3257 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3260 * Do boot-time initialization of a CPU's per-CPU RCU data.
3262 static void __init
3263 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3265 unsigned long flags;
3266 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3267 struct rcu_node *rnp = rcu_get_root(rsp);
3269 /* Set up local state, ensuring consistent view of global state. */
3270 raw_spin_lock_irqsave(&rnp->lock, flags);
3271 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3272 init_callback_list(rdp);
3273 rdp->qlen_lazy = 0;
3274 ACCESS_ONCE(rdp->qlen) = 0;
3275 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3276 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3277 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3278 rdp->cpu = cpu;
3279 rdp->rsp = rsp;
3280 rcu_boot_init_nocb_percpu_data(rdp);
3281 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3285 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3286 * offline event can be happening at a given time. Note also that we
3287 * can accept some slop in the rsp->completed access due to the fact
3288 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3290 static void
3291 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3293 unsigned long flags;
3294 unsigned long mask;
3295 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3296 struct rcu_node *rnp = rcu_get_root(rsp);
3298 /* Exclude new grace periods. */
3299 mutex_lock(&rsp->onoff_mutex);
3301 /* Set up local state, ensuring consistent view of global state. */
3302 raw_spin_lock_irqsave(&rnp->lock, flags);
3303 rdp->beenonline = 1; /* We have now been online. */
3304 rdp->qlen_last_fqs_check = 0;
3305 rdp->n_force_qs_snap = rsp->n_force_qs;
3306 rdp->blimit = blimit;
3307 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3308 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3309 rcu_sysidle_init_percpu_data(rdp->dynticks);
3310 atomic_set(&rdp->dynticks->dynticks,
3311 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3312 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3314 /* Add CPU to rcu_node bitmasks. */
3315 rnp = rdp->mynode;
3316 mask = rdp->grpmask;
3317 do {
3318 /* Exclude any attempts to start a new GP on small systems. */
3319 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3320 rnp->qsmaskinit |= mask;
3321 mask = rnp->grpmask;
3322 if (rnp == rdp->mynode) {
3324 * If there is a grace period in progress, we will
3325 * set up to wait for it next time we run the
3326 * RCU core code.
3328 rdp->gpnum = rnp->completed;
3329 rdp->completed = rnp->completed;
3330 rdp->passed_quiesce = 0;
3331 rdp->qs_pending = 0;
3332 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3334 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3335 rnp = rnp->parent;
3336 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3337 local_irq_restore(flags);
3339 mutex_unlock(&rsp->onoff_mutex);
3342 static void rcu_prepare_cpu(int cpu)
3344 struct rcu_state *rsp;
3346 for_each_rcu_flavor(rsp)
3347 rcu_init_percpu_data(cpu, rsp);
3351 * Handle CPU online/offline notification events.
3353 static int rcu_cpu_notify(struct notifier_block *self,
3354 unsigned long action, void *hcpu)
3356 long cpu = (long)hcpu;
3357 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3358 struct rcu_node *rnp = rdp->mynode;
3359 struct rcu_state *rsp;
3361 trace_rcu_utilization(TPS("Start CPU hotplug"));
3362 switch (action) {
3363 case CPU_UP_PREPARE:
3364 case CPU_UP_PREPARE_FROZEN:
3365 rcu_prepare_cpu(cpu);
3366 rcu_prepare_kthreads(cpu);
3367 break;
3368 case CPU_ONLINE:
3369 case CPU_DOWN_FAILED:
3370 rcu_boost_kthread_setaffinity(rnp, -1);
3371 break;
3372 case CPU_DOWN_PREPARE:
3373 rcu_boost_kthread_setaffinity(rnp, cpu);
3374 break;
3375 case CPU_DYING:
3376 case CPU_DYING_FROZEN:
3377 for_each_rcu_flavor(rsp)
3378 rcu_cleanup_dying_cpu(rsp);
3379 break;
3380 case CPU_DEAD:
3381 case CPU_DEAD_FROZEN:
3382 case CPU_UP_CANCELED:
3383 case CPU_UP_CANCELED_FROZEN:
3384 for_each_rcu_flavor(rsp)
3385 rcu_cleanup_dead_cpu(cpu, rsp);
3386 break;
3387 default:
3388 break;
3390 trace_rcu_utilization(TPS("End CPU hotplug"));
3391 return NOTIFY_OK;
3394 static int rcu_pm_notify(struct notifier_block *self,
3395 unsigned long action, void *hcpu)
3397 switch (action) {
3398 case PM_HIBERNATION_PREPARE:
3399 case PM_SUSPEND_PREPARE:
3400 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3401 rcu_expedited = 1;
3402 break;
3403 case PM_POST_HIBERNATION:
3404 case PM_POST_SUSPEND:
3405 rcu_expedited = 0;
3406 break;
3407 default:
3408 break;
3410 return NOTIFY_OK;
3414 * Spawn the kthread that handles this RCU flavor's grace periods.
3416 static int __init rcu_spawn_gp_kthread(void)
3418 unsigned long flags;
3419 struct rcu_node *rnp;
3420 struct rcu_state *rsp;
3421 struct task_struct *t;
3423 for_each_rcu_flavor(rsp) {
3424 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3425 BUG_ON(IS_ERR(t));
3426 rnp = rcu_get_root(rsp);
3427 raw_spin_lock_irqsave(&rnp->lock, flags);
3428 rsp->gp_kthread = t;
3429 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3430 rcu_spawn_nocb_kthreads(rsp);
3432 return 0;
3434 early_initcall(rcu_spawn_gp_kthread);
3437 * This function is invoked towards the end of the scheduler's initialization
3438 * process. Before this is called, the idle task might contain
3439 * RCU read-side critical sections (during which time, this idle
3440 * task is booting the system). After this function is called, the
3441 * idle tasks are prohibited from containing RCU read-side critical
3442 * sections. This function also enables RCU lockdep checking.
3444 void rcu_scheduler_starting(void)
3446 WARN_ON(num_online_cpus() != 1);
3447 WARN_ON(nr_context_switches() > 0);
3448 rcu_scheduler_active = 1;
3452 * Compute the per-level fanout, either using the exact fanout specified
3453 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3455 #ifdef CONFIG_RCU_FANOUT_EXACT
3456 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3458 int i;
3460 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3461 for (i = rcu_num_lvls - 2; i >= 0; i--)
3462 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3464 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3465 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3467 int ccur;
3468 int cprv;
3469 int i;
3471 cprv = nr_cpu_ids;
3472 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3473 ccur = rsp->levelcnt[i];
3474 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3475 cprv = ccur;
3478 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3481 * Helper function for rcu_init() that initializes one rcu_state structure.
3483 static void __init rcu_init_one(struct rcu_state *rsp,
3484 struct rcu_data __percpu *rda)
3486 static char *buf[] = { "rcu_node_0",
3487 "rcu_node_1",
3488 "rcu_node_2",
3489 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3490 static char *fqs[] = { "rcu_node_fqs_0",
3491 "rcu_node_fqs_1",
3492 "rcu_node_fqs_2",
3493 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3494 int cpustride = 1;
3495 int i;
3496 int j;
3497 struct rcu_node *rnp;
3499 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3501 /* Silence gcc 4.8 warning about array index out of range. */
3502 if (rcu_num_lvls > RCU_NUM_LVLS)
3503 panic("rcu_init_one: rcu_num_lvls overflow");
3505 /* Initialize the level-tracking arrays. */
3507 for (i = 0; i < rcu_num_lvls; i++)
3508 rsp->levelcnt[i] = num_rcu_lvl[i];
3509 for (i = 1; i < rcu_num_lvls; i++)
3510 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3511 rcu_init_levelspread(rsp);
3513 /* Initialize the elements themselves, starting from the leaves. */
3515 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3516 cpustride *= rsp->levelspread[i];
3517 rnp = rsp->level[i];
3518 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3519 raw_spin_lock_init(&rnp->lock);
3520 lockdep_set_class_and_name(&rnp->lock,
3521 &rcu_node_class[i], buf[i]);
3522 raw_spin_lock_init(&rnp->fqslock);
3523 lockdep_set_class_and_name(&rnp->fqslock,
3524 &rcu_fqs_class[i], fqs[i]);
3525 rnp->gpnum = rsp->gpnum;
3526 rnp->completed = rsp->completed;
3527 rnp->qsmask = 0;
3528 rnp->qsmaskinit = 0;
3529 rnp->grplo = j * cpustride;
3530 rnp->grphi = (j + 1) * cpustride - 1;
3531 if (rnp->grphi >= nr_cpu_ids)
3532 rnp->grphi = nr_cpu_ids - 1;
3533 if (i == 0) {
3534 rnp->grpnum = 0;
3535 rnp->grpmask = 0;
3536 rnp->parent = NULL;
3537 } else {
3538 rnp->grpnum = j % rsp->levelspread[i - 1];
3539 rnp->grpmask = 1UL << rnp->grpnum;
3540 rnp->parent = rsp->level[i - 1] +
3541 j / rsp->levelspread[i - 1];
3543 rnp->level = i;
3544 INIT_LIST_HEAD(&rnp->blkd_tasks);
3545 rcu_init_one_nocb(rnp);
3549 rsp->rda = rda;
3550 init_waitqueue_head(&rsp->gp_wq);
3551 rnp = rsp->level[rcu_num_lvls - 1];
3552 for_each_possible_cpu(i) {
3553 while (i > rnp->grphi)
3554 rnp++;
3555 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3556 rcu_boot_init_percpu_data(i, rsp);
3558 list_add(&rsp->flavors, &rcu_struct_flavors);
3562 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3563 * replace the definitions in tree.h because those are needed to size
3564 * the ->node array in the rcu_state structure.
3566 static void __init rcu_init_geometry(void)
3568 ulong d;
3569 int i;
3570 int j;
3571 int n = nr_cpu_ids;
3572 int rcu_capacity[MAX_RCU_LVLS + 1];
3575 * Initialize any unspecified boot parameters.
3576 * The default values of jiffies_till_first_fqs and
3577 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3578 * value, which is a function of HZ, then adding one for each
3579 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3581 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3582 if (jiffies_till_first_fqs == ULONG_MAX)
3583 jiffies_till_first_fqs = d;
3584 if (jiffies_till_next_fqs == ULONG_MAX)
3585 jiffies_till_next_fqs = d;
3587 /* If the compile-time values are accurate, just leave. */
3588 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3589 nr_cpu_ids == NR_CPUS)
3590 return;
3591 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3592 rcu_fanout_leaf, nr_cpu_ids);
3595 * Compute number of nodes that can be handled an rcu_node tree
3596 * with the given number of levels. Setting rcu_capacity[0] makes
3597 * some of the arithmetic easier.
3599 rcu_capacity[0] = 1;
3600 rcu_capacity[1] = rcu_fanout_leaf;
3601 for (i = 2; i <= MAX_RCU_LVLS; i++)
3602 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3605 * The boot-time rcu_fanout_leaf parameter is only permitted
3606 * to increase the leaf-level fanout, not decrease it. Of course,
3607 * the leaf-level fanout cannot exceed the number of bits in
3608 * the rcu_node masks. Finally, the tree must be able to accommodate
3609 * the configured number of CPUs. Complain and fall back to the
3610 * compile-time values if these limits are exceeded.
3612 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3613 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3614 n > rcu_capacity[MAX_RCU_LVLS]) {
3615 WARN_ON(1);
3616 return;
3619 /* Calculate the number of rcu_nodes at each level of the tree. */
3620 for (i = 1; i <= MAX_RCU_LVLS; i++)
3621 if (n <= rcu_capacity[i]) {
3622 for (j = 0; j <= i; j++)
3623 num_rcu_lvl[j] =
3624 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3625 rcu_num_lvls = i;
3626 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3627 num_rcu_lvl[j] = 0;
3628 break;
3631 /* Calculate the total number of rcu_node structures. */
3632 rcu_num_nodes = 0;
3633 for (i = 0; i <= MAX_RCU_LVLS; i++)
3634 rcu_num_nodes += num_rcu_lvl[i];
3635 rcu_num_nodes -= n;
3638 void __init rcu_init(void)
3640 int cpu;
3642 rcu_bootup_announce();
3643 rcu_init_geometry();
3644 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3645 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3646 __rcu_init_preempt();
3647 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3650 * We don't need protection against CPU-hotplug here because
3651 * this is called early in boot, before either interrupts
3652 * or the scheduler are operational.
3654 cpu_notifier(rcu_cpu_notify, 0);
3655 pm_notifier(rcu_pm_notify, 0);
3656 for_each_online_cpu(cpu)
3657 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3660 #include "tree_plugin.h"