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, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
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 -
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>
61 #include <trace/events/rcu.h>
65 MODULE_ALIAS("rcutree");
66 #ifdef MODULE_PARAM_PREFIX
67 #undef MODULE_PARAM_PREFIX
69 #define MODULE_PARAM_PREFIX "rcutree."
71 /* Data structures. */
73 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
74 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
77 * In order to export the rcu_state name to the tracing tools, it
78 * needs to be added in the __tracepoint_string section.
79 * This requires defining a separate variable tp_<sname>_varname
80 * that points to the string being used, and this will allow
81 * the tracing userspace tools to be able to decipher the string
82 * address to the matching string.
84 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
87 struct rcu_state sname##_state = { \
88 .level = { &sname##_state.node[0] }, \
90 .fqs_state = RCU_GP_IDLE, \
91 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
93 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
94 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
96 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
97 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
98 .name = sname##_varname, \
101 DEFINE_PER_CPU(struct rcu_data, sname##_data)
103 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
106 static struct rcu_state *rcu_state;
107 LIST_HEAD(rcu_struct_flavors);
109 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
111 module_param(rcu_fanout_leaf, int, 0444);
112 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
120 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
123 * The rcu_scheduler_active variable transitions from zero to one just
124 * before the first task is spawned. So when this variable is zero, RCU
125 * can assume that there is but one task, allowing RCU to (for example)
126 * optimize synchronize_sched() to a simple barrier(). When this variable
127 * is one, RCU must actually do all the hard work required to detect real
128 * grace periods. This variable is also used to suppress boot-time false
129 * positives from lockdep-RCU error checking.
131 int rcu_scheduler_active __read_mostly;
132 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
135 * The rcu_scheduler_fully_active variable transitions from zero to one
136 * during the early_initcall() processing, which is after the scheduler
137 * is capable of creating new tasks. So RCU processing (for example,
138 * creating tasks for RCU priority boosting) must be delayed until after
139 * rcu_scheduler_fully_active transitions from zero to one. We also
140 * currently delay invocation of any RCU callbacks until after this point.
142 * It might later prove better for people registering RCU callbacks during
143 * early boot to take responsibility for these callbacks, but one step at
146 static int rcu_scheduler_fully_active __read_mostly;
148 #ifdef CONFIG_RCU_BOOST
151 * Control variables for per-CPU and per-rcu_node kthreads. These
152 * handle all flavors of RCU.
154 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
156 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
157 DEFINE_PER_CPU(char, rcu_cpu_has_work);
159 #endif /* #ifdef CONFIG_RCU_BOOST */
161 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162 static void invoke_rcu_core(void);
163 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 * Track the rcutorture test sequence number and the update version
167 * number within a given test. The rcutorture_testseq is incremented
168 * on every rcutorture module load and unload, so has an odd value
169 * when a test is running. The rcutorture_vernum is set to zero
170 * when rcutorture starts and is incremented on each rcutorture update.
171 * These variables enable correlating rcutorture output with the
172 * RCU tracing information.
174 unsigned long rcutorture_testseq;
175 unsigned long rcutorture_vernum;
178 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
179 * permit this function to be invoked without holding the root rcu_node
180 * structure's ->lock, but of course results can be subject to change.
182 static int rcu_gp_in_progress(struct rcu_state *rsp)
184 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
188 * Note a quiescent state. Because we do not need to know
189 * how many quiescent states passed, just if there was at least
190 * one since the start of the grace period, this just sets a flag.
191 * The caller must have disabled preemption.
193 void rcu_sched_qs(int cpu)
195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
199 rdp->passed_quiesce = 1;
202 void rcu_bh_qs(int cpu)
204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
206 if (rdp->passed_quiesce == 0)
207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
208 rdp->passed_quiesce = 1;
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
214 * The caller must have disabled preemption.
216 void rcu_note_context_switch(int cpu)
218 trace_rcu_utilization(TPS("Start context switch"));
220 rcu_preempt_note_context_switch(cpu);
221 trace_rcu_utilization(TPS("End context switch"));
223 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
225 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
227 .dynticks = ATOMIC_INIT(1),
228 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
229 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
230 .dynticks_idle = ATOMIC_INIT(1),
231 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
234 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
235 static long qhimark = 10000; /* If this many pending, ignore blimit. */
236 static long qlowmark = 100; /* Once only this many pending, use blimit. */
238 module_param(blimit, long, 0444);
239 module_param(qhimark, long, 0444);
240 module_param(qlowmark, long, 0444);
242 static ulong jiffies_till_first_fqs = ULONG_MAX;
243 static ulong jiffies_till_next_fqs = ULONG_MAX;
245 module_param(jiffies_till_first_fqs, ulong, 0644);
246 module_param(jiffies_till_next_fqs, ulong, 0644);
248 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
249 struct rcu_data *rdp);
250 static void force_qs_rnp(struct rcu_state *rsp,
251 int (*f)(struct rcu_data *rsp, bool *isidle,
252 unsigned long *maxj),
253 bool *isidle, unsigned long *maxj);
254 static void force_quiescent_state(struct rcu_state *rsp);
255 static int rcu_pending(int cpu);
258 * Return the number of RCU-sched batches processed thus far for debug & stats.
260 long rcu_batches_completed_sched(void)
262 return rcu_sched_state.completed;
264 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
267 * Return the number of RCU BH batches processed thus far for debug & stats.
269 long rcu_batches_completed_bh(void)
271 return rcu_bh_state.completed;
273 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
276 * Force a quiescent state for RCU BH.
278 void rcu_bh_force_quiescent_state(void)
280 force_quiescent_state(&rcu_bh_state);
282 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
285 * Record the number of times rcutorture tests have been initiated and
286 * terminated. This information allows the debugfs tracing stats to be
287 * correlated to the rcutorture messages, even when the rcutorture module
288 * is being repeatedly loaded and unloaded. In other words, we cannot
289 * store this state in rcutorture itself.
291 void rcutorture_record_test_transition(void)
293 rcutorture_testseq++;
294 rcutorture_vernum = 0;
296 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
299 * Record the number of writer passes through the current rcutorture test.
300 * This is also used to correlate debugfs tracing stats with the rcutorture
303 void rcutorture_record_progress(unsigned long vernum)
307 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
310 * Force a quiescent state for RCU-sched.
312 void rcu_sched_force_quiescent_state(void)
314 force_quiescent_state(&rcu_sched_state);
316 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
319 * Does the CPU have callbacks ready to be invoked?
322 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
324 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
325 rdp->nxttail[RCU_DONE_TAIL] != NULL;
329 * Does the current CPU require a not-yet-started grace period?
330 * The caller must have disabled interrupts to prevent races with
331 * normal callback registry.
334 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
338 if (rcu_gp_in_progress(rsp))
339 return 0; /* No, a grace period is already in progress. */
340 if (rcu_nocb_needs_gp(rsp))
341 return 1; /* Yes, a no-CBs CPU needs one. */
342 if (!rdp->nxttail[RCU_NEXT_TAIL])
343 return 0; /* No, this is a no-CBs (or offline) CPU. */
344 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
345 return 1; /* Yes, this CPU has newly registered callbacks. */
346 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
347 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
348 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
349 rdp->nxtcompleted[i]))
350 return 1; /* Yes, CBs for future grace period. */
351 return 0; /* No grace period needed. */
355 * Return the root node of the specified rcu_state structure.
357 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
359 return &rsp->node[0];
363 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
365 * If the new value of the ->dynticks_nesting counter now is zero,
366 * we really have entered idle, and must do the appropriate accounting.
367 * The caller must have disabled interrupts.
369 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
372 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
373 if (!user && !is_idle_task(current)) {
374 struct task_struct *idle __maybe_unused =
375 idle_task(smp_processor_id());
377 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
378 ftrace_dump(DUMP_ORIG);
379 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
380 current->pid, current->comm,
381 idle->pid, idle->comm); /* must be idle task! */
383 rcu_prepare_for_idle(smp_processor_id());
384 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
385 smp_mb__before_atomic_inc(); /* See above. */
386 atomic_inc(&rdtp->dynticks);
387 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
388 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
391 * It is illegal to enter an extended quiescent state while
392 * in an RCU read-side critical section.
394 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
395 "Illegal idle entry in RCU read-side critical section.");
396 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
397 "Illegal idle entry in RCU-bh read-side critical section.");
398 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
399 "Illegal idle entry in RCU-sched read-side critical section.");
403 * Enter an RCU extended quiescent state, which can be either the
404 * idle loop or adaptive-tickless usermode execution.
406 static void rcu_eqs_enter(bool user)
409 struct rcu_dynticks *rdtp;
411 rdtp = this_cpu_ptr(&rcu_dynticks);
412 oldval = rdtp->dynticks_nesting;
413 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
414 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
415 rdtp->dynticks_nesting = 0;
417 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
418 rcu_eqs_enter_common(rdtp, oldval, user);
422 * rcu_idle_enter - inform RCU that current CPU is entering idle
424 * Enter idle mode, in other words, -leave- the mode in which RCU
425 * read-side critical sections can occur. (Though RCU read-side
426 * critical sections can occur in irq handlers in idle, a possibility
427 * handled by irq_enter() and irq_exit().)
429 * We crowbar the ->dynticks_nesting field to zero to allow for
430 * the possibility of usermode upcalls having messed up our count
431 * of interrupt nesting level during the prior busy period.
433 void rcu_idle_enter(void)
437 local_irq_save(flags);
438 rcu_eqs_enter(false);
439 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
440 local_irq_restore(flags);
442 EXPORT_SYMBOL_GPL(rcu_idle_enter);
444 #ifdef CONFIG_RCU_USER_QS
446 * rcu_user_enter - inform RCU that we are resuming userspace.
448 * Enter RCU idle mode right before resuming userspace. No use of RCU
449 * is permitted between this call and rcu_user_exit(). This way the
450 * CPU doesn't need to maintain the tick for RCU maintenance purposes
451 * when the CPU runs in userspace.
453 void rcu_user_enter(void)
457 #endif /* CONFIG_RCU_USER_QS */
460 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
462 * Exit from an interrupt handler, which might possibly result in entering
463 * idle mode, in other words, leaving the mode in which read-side critical
464 * sections can occur.
466 * This code assumes that the idle loop never does anything that might
467 * result in unbalanced calls to irq_enter() and irq_exit(). If your
468 * architecture violates this assumption, RCU will give you what you
469 * deserve, good and hard. But very infrequently and irreproducibly.
471 * Use things like work queues to work around this limitation.
473 * You have been warned.
475 void rcu_irq_exit(void)
479 struct rcu_dynticks *rdtp;
481 local_irq_save(flags);
482 rdtp = this_cpu_ptr(&rcu_dynticks);
483 oldval = rdtp->dynticks_nesting;
484 rdtp->dynticks_nesting--;
485 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
486 if (rdtp->dynticks_nesting)
487 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
489 rcu_eqs_enter_common(rdtp, oldval, true);
490 rcu_sysidle_enter(rdtp, 1);
491 local_irq_restore(flags);
495 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
497 * If the new value of the ->dynticks_nesting counter was previously zero,
498 * we really have exited idle, and must do the appropriate accounting.
499 * The caller must have disabled interrupts.
501 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
504 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
505 atomic_inc(&rdtp->dynticks);
506 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
507 smp_mb__after_atomic_inc(); /* See above. */
508 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
509 rcu_cleanup_after_idle(smp_processor_id());
510 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
511 if (!user && !is_idle_task(current)) {
512 struct task_struct *idle __maybe_unused =
513 idle_task(smp_processor_id());
515 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
516 oldval, rdtp->dynticks_nesting);
517 ftrace_dump(DUMP_ORIG);
518 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
519 current->pid, current->comm,
520 idle->pid, idle->comm); /* must be idle task! */
525 * Exit an RCU extended quiescent state, which can be either the
526 * idle loop or adaptive-tickless usermode execution.
528 static void rcu_eqs_exit(bool user)
530 struct rcu_dynticks *rdtp;
533 rdtp = this_cpu_ptr(&rcu_dynticks);
534 oldval = rdtp->dynticks_nesting;
535 WARN_ON_ONCE(oldval < 0);
536 if (oldval & DYNTICK_TASK_NEST_MASK)
537 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
539 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
540 rcu_eqs_exit_common(rdtp, oldval, user);
544 * rcu_idle_exit - inform RCU that current CPU is leaving idle
546 * Exit idle mode, in other words, -enter- the mode in which RCU
547 * read-side critical sections can occur.
549 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
550 * allow for the possibility of usermode upcalls messing up our count
551 * of interrupt nesting level during the busy period that is just
554 void rcu_idle_exit(void)
558 local_irq_save(flags);
560 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
561 local_irq_restore(flags);
563 EXPORT_SYMBOL_GPL(rcu_idle_exit);
565 #ifdef CONFIG_RCU_USER_QS
567 * rcu_user_exit - inform RCU that we are exiting userspace.
569 * Exit RCU idle mode while entering the kernel because it can
570 * run a RCU read side critical section anytime.
572 void rcu_user_exit(void)
576 #endif /* CONFIG_RCU_USER_QS */
579 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
581 * Enter an interrupt handler, which might possibly result in exiting
582 * idle mode, in other words, entering the mode in which read-side critical
583 * sections can occur.
585 * Note that the Linux kernel is fully capable of entering an interrupt
586 * handler that it never exits, for example when doing upcalls to
587 * user mode! This code assumes that the idle loop never does upcalls to
588 * user mode. If your architecture does do upcalls from the idle loop (or
589 * does anything else that results in unbalanced calls to the irq_enter()
590 * and irq_exit() functions), RCU will give you what you deserve, good
591 * and hard. But very infrequently and irreproducibly.
593 * Use things like work queues to work around this limitation.
595 * You have been warned.
597 void rcu_irq_enter(void)
600 struct rcu_dynticks *rdtp;
603 local_irq_save(flags);
604 rdtp = this_cpu_ptr(&rcu_dynticks);
605 oldval = rdtp->dynticks_nesting;
606 rdtp->dynticks_nesting++;
607 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
609 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
611 rcu_eqs_exit_common(rdtp, oldval, true);
612 rcu_sysidle_exit(rdtp, 1);
613 local_irq_restore(flags);
617 * rcu_nmi_enter - inform RCU of entry to NMI context
619 * If the CPU was idle with dynamic ticks active, and there is no
620 * irq handler running, this updates rdtp->dynticks_nmi to let the
621 * RCU grace-period handling know that the CPU is active.
623 void rcu_nmi_enter(void)
625 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
627 if (rdtp->dynticks_nmi_nesting == 0 &&
628 (atomic_read(&rdtp->dynticks) & 0x1))
630 rdtp->dynticks_nmi_nesting++;
631 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
632 atomic_inc(&rdtp->dynticks);
633 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
634 smp_mb__after_atomic_inc(); /* See above. */
635 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
639 * rcu_nmi_exit - inform RCU of exit from NMI context
641 * If the CPU was idle with dynamic ticks active, and there is no
642 * irq handler running, this updates rdtp->dynticks_nmi to let the
643 * RCU grace-period handling know that the CPU is no longer active.
645 void rcu_nmi_exit(void)
647 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
649 if (rdtp->dynticks_nmi_nesting == 0 ||
650 --rdtp->dynticks_nmi_nesting != 0)
652 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
653 smp_mb__before_atomic_inc(); /* See above. */
654 atomic_inc(&rdtp->dynticks);
655 smp_mb__after_atomic_inc(); /* Force delay to next write. */
656 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
660 * __rcu_is_watching - are RCU read-side critical sections safe?
662 * Return true if RCU is watching the running CPU, which means that
663 * this CPU can safely enter RCU read-side critical sections. Unlike
664 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
665 * least disabled preemption.
667 bool notrace __rcu_is_watching(void)
669 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
673 * rcu_is_watching - see if RCU thinks that the current CPU is idle
675 * If the current CPU is in its idle loop and is neither in an interrupt
676 * or NMI handler, return true.
678 bool notrace rcu_is_watching(void)
683 ret = __rcu_is_watching();
687 EXPORT_SYMBOL_GPL(rcu_is_watching);
689 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
692 * Is the current CPU online? Disable preemption to avoid false positives
693 * that could otherwise happen due to the current CPU number being sampled,
694 * this task being preempted, its old CPU being taken offline, resuming
695 * on some other CPU, then determining that its old CPU is now offline.
696 * It is OK to use RCU on an offline processor during initial boot, hence
697 * the check for rcu_scheduler_fully_active. Note also that it is OK
698 * for a CPU coming online to use RCU for one jiffy prior to marking itself
699 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
700 * offline to continue to use RCU for one jiffy after marking itself
701 * offline in the cpu_online_mask. This leniency is necessary given the
702 * non-atomic nature of the online and offline processing, for example,
703 * the fact that a CPU enters the scheduler after completing the CPU_DYING
706 * This is also why RCU internally marks CPUs online during the
707 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
709 * Disable checking if in an NMI handler because we cannot safely report
710 * errors from NMI handlers anyway.
712 bool rcu_lockdep_current_cpu_online(void)
714 struct rcu_data *rdp;
715 struct rcu_node *rnp;
721 rdp = this_cpu_ptr(&rcu_sched_data);
723 ret = (rdp->grpmask & rnp->qsmaskinit) ||
724 !rcu_scheduler_fully_active;
728 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
730 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
733 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
735 * If the current CPU is idle or running at a first-level (not nested)
736 * interrupt from idle, return true. The caller must have at least
737 * disabled preemption.
739 static int rcu_is_cpu_rrupt_from_idle(void)
741 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
745 * Snapshot the specified CPU's dynticks counter so that we can later
746 * credit them with an implicit quiescent state. Return 1 if this CPU
747 * is in dynticks idle mode, which is an extended quiescent state.
749 static int dyntick_save_progress_counter(struct rcu_data *rdp,
750 bool *isidle, unsigned long *maxj)
752 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
753 rcu_sysidle_check_cpu(rdp, isidle, maxj);
754 return (rdp->dynticks_snap & 0x1) == 0;
758 * This function really isn't for public consumption, but RCU is special in
759 * that context switches can allow the state machine to make progress.
761 extern void resched_cpu(int cpu);
764 * Return true if the specified CPU has passed through a quiescent
765 * state by virtue of being in or having passed through an dynticks
766 * idle state since the last call to dyntick_save_progress_counter()
767 * for this same CPU, or by virtue of having been offline.
769 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
770 bool *isidle, unsigned long *maxj)
775 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
776 snap = (unsigned int)rdp->dynticks_snap;
779 * If the CPU passed through or entered a dynticks idle phase with
780 * no active irq/NMI handlers, then we can safely pretend that the CPU
781 * already acknowledged the request to pass through a quiescent
782 * state. Either way, that CPU cannot possibly be in an RCU
783 * read-side critical section that started before the beginning
784 * of the current RCU grace period.
786 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
787 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
793 * Check for the CPU being offline, but only if the grace period
794 * is old enough. We don't need to worry about the CPU changing
795 * state: If we see it offline even once, it has been through a
798 * The reason for insisting that the grace period be at least
799 * one jiffy old is that CPUs that are not quite online and that
800 * have just gone offline can still execute RCU read-side critical
803 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
804 return 0; /* Grace period is not old enough. */
806 if (cpu_is_offline(rdp->cpu)) {
807 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
813 * There is a possibility that a CPU in adaptive-ticks state
814 * might run in the kernel with the scheduling-clock tick disabled
815 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
816 * force the CPU to restart the scheduling-clock tick in this
817 * CPU is in this state.
819 rcu_kick_nohz_cpu(rdp->cpu);
822 * Alternatively, the CPU might be running in the kernel
823 * for an extended period of time without a quiescent state.
824 * Attempt to force the CPU through the scheduler to gain the
825 * needed quiescent state, but only if the grace period has gone
826 * on for an uncommonly long time. If there are many stuck CPUs,
827 * we will beat on the first one until it gets unstuck, then move
828 * to the next. Only do this for the primary flavor of RCU.
830 if (rdp->rsp == rcu_state &&
831 ULONG_CMP_GE(ACCESS_ONCE(jiffies), rdp->rsp->jiffies_resched)) {
832 rdp->rsp->jiffies_resched += 5;
833 resched_cpu(rdp->cpu);
839 static void record_gp_stall_check_time(struct rcu_state *rsp)
841 unsigned long j = ACCESS_ONCE(jiffies);
845 smp_wmb(); /* Record start time before stall time. */
846 j1 = rcu_jiffies_till_stall_check();
847 rsp->jiffies_stall = j + j1;
848 rsp->jiffies_resched = j + j1 / 2;
852 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
853 * for architectures that do not implement trigger_all_cpu_backtrace().
854 * The NMI-triggered stack traces are more accurate because they are
855 * printed by the target CPU.
857 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
861 struct rcu_node *rnp;
863 rcu_for_each_leaf_node(rsp, rnp) {
864 raw_spin_lock_irqsave(&rnp->lock, flags);
865 if (rnp->qsmask != 0) {
866 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
867 if (rnp->qsmask & (1UL << cpu))
868 dump_cpu_task(rnp->grplo + cpu);
870 raw_spin_unlock_irqrestore(&rnp->lock, flags);
874 static void print_other_cpu_stall(struct rcu_state *rsp)
880 struct rcu_node *rnp = rcu_get_root(rsp);
883 /* Only let one CPU complain about others per time interval. */
885 raw_spin_lock_irqsave(&rnp->lock, flags);
886 delta = jiffies - rsp->jiffies_stall;
887 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
888 raw_spin_unlock_irqrestore(&rnp->lock, flags);
891 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
892 raw_spin_unlock_irqrestore(&rnp->lock, flags);
895 * OK, time to rat on our buddy...
896 * See Documentation/RCU/stallwarn.txt for info on how to debug
897 * RCU CPU stall warnings.
899 pr_err("INFO: %s detected stalls on CPUs/tasks:",
901 print_cpu_stall_info_begin();
902 rcu_for_each_leaf_node(rsp, rnp) {
903 raw_spin_lock_irqsave(&rnp->lock, flags);
904 ndetected += rcu_print_task_stall(rnp);
905 if (rnp->qsmask != 0) {
906 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
907 if (rnp->qsmask & (1UL << cpu)) {
908 print_cpu_stall_info(rsp,
913 raw_spin_unlock_irqrestore(&rnp->lock, flags);
917 * Now rat on any tasks that got kicked up to the root rcu_node
918 * due to CPU offlining.
920 rnp = rcu_get_root(rsp);
921 raw_spin_lock_irqsave(&rnp->lock, flags);
922 ndetected += rcu_print_task_stall(rnp);
923 raw_spin_unlock_irqrestore(&rnp->lock, flags);
925 print_cpu_stall_info_end();
926 for_each_possible_cpu(cpu)
927 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
928 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
929 smp_processor_id(), (long)(jiffies - rsp->gp_start),
930 rsp->gpnum, rsp->completed, totqlen);
932 pr_err("INFO: Stall ended before state dump start\n");
933 else if (!trigger_all_cpu_backtrace())
934 rcu_dump_cpu_stacks(rsp);
936 /* Complain about tasks blocking the grace period. */
938 rcu_print_detail_task_stall(rsp);
940 force_quiescent_state(rsp); /* Kick them all. */
944 * This function really isn't for public consumption, but RCU is special in
945 * that context switches can allow the state machine to make progress.
947 extern void resched_cpu(int cpu);
949 static void print_cpu_stall(struct rcu_state *rsp)
953 struct rcu_node *rnp = rcu_get_root(rsp);
957 * OK, time to rat on ourselves...
958 * See Documentation/RCU/stallwarn.txt for info on how to debug
959 * RCU CPU stall warnings.
961 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
962 print_cpu_stall_info_begin();
963 print_cpu_stall_info(rsp, smp_processor_id());
964 print_cpu_stall_info_end();
965 for_each_possible_cpu(cpu)
966 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
967 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
968 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
969 if (!trigger_all_cpu_backtrace())
972 raw_spin_lock_irqsave(&rnp->lock, flags);
973 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
974 rsp->jiffies_stall = jiffies +
975 3 * rcu_jiffies_till_stall_check() + 3;
976 raw_spin_unlock_irqrestore(&rnp->lock, flags);
979 * Attempt to revive the RCU machinery by forcing a context switch.
981 * A context switch would normally allow the RCU state machine to make
982 * progress and it could be we're stuck in kernel space without context
983 * switches for an entirely unreasonable amount of time.
985 resched_cpu(smp_processor_id());
988 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
990 unsigned long completed;
995 struct rcu_node *rnp;
997 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
999 j = ACCESS_ONCE(jiffies);
1002 * Lots of memory barriers to reject false positives.
1004 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1005 * then rsp->gp_start, and finally rsp->completed. These values
1006 * are updated in the opposite order with memory barriers (or
1007 * equivalent) during grace-period initialization and cleanup.
1008 * Now, a false positive can occur if we get an new value of
1009 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1010 * the memory barriers, the only way that this can happen is if one
1011 * grace period ends and another starts between these two fetches.
1012 * Detect this by comparing rsp->completed with the previous fetch
1015 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1016 * and rsp->gp_start suffice to forestall false positives.
1018 gpnum = ACCESS_ONCE(rsp->gpnum);
1019 smp_rmb(); /* Pick up ->gpnum first... */
1020 js = ACCESS_ONCE(rsp->jiffies_stall);
1021 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1022 gps = ACCESS_ONCE(rsp->gp_start);
1023 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1024 completed = ACCESS_ONCE(rsp->completed);
1025 if (ULONG_CMP_GE(completed, gpnum) ||
1026 ULONG_CMP_LT(j, js) ||
1027 ULONG_CMP_GE(gps, js))
1028 return; /* No stall or GP completed since entering function. */
1030 if (rcu_gp_in_progress(rsp) &&
1031 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1033 /* We haven't checked in, so go dump stack. */
1034 print_cpu_stall(rsp);
1036 } else if (rcu_gp_in_progress(rsp) &&
1037 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1039 /* They had a few time units to dump stack, so complain. */
1040 print_other_cpu_stall(rsp);
1045 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1047 * Set the stall-warning timeout way off into the future, thus preventing
1048 * any RCU CPU stall-warning messages from appearing in the current set of
1049 * RCU grace periods.
1051 * The caller must disable hard irqs.
1053 void rcu_cpu_stall_reset(void)
1055 struct rcu_state *rsp;
1057 for_each_rcu_flavor(rsp)
1058 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1062 * Initialize the specified rcu_data structure's callback list to empty.
1064 static void init_callback_list(struct rcu_data *rdp)
1068 if (init_nocb_callback_list(rdp))
1070 rdp->nxtlist = NULL;
1071 for (i = 0; i < RCU_NEXT_SIZE; i++)
1072 rdp->nxttail[i] = &rdp->nxtlist;
1076 * Determine the value that ->completed will have at the end of the
1077 * next subsequent grace period. This is used to tag callbacks so that
1078 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1079 * been dyntick-idle for an extended period with callbacks under the
1080 * influence of RCU_FAST_NO_HZ.
1082 * The caller must hold rnp->lock with interrupts disabled.
1084 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1085 struct rcu_node *rnp)
1088 * If RCU is idle, we just wait for the next grace period.
1089 * But we can only be sure that RCU is idle if we are looking
1090 * at the root rcu_node structure -- otherwise, a new grace
1091 * period might have started, but just not yet gotten around
1092 * to initializing the current non-root rcu_node structure.
1094 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1095 return rnp->completed + 1;
1098 * Otherwise, wait for a possible partial grace period and
1099 * then the subsequent full grace period.
1101 return rnp->completed + 2;
1105 * Trace-event helper function for rcu_start_future_gp() and
1106 * rcu_nocb_wait_gp().
1108 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1109 unsigned long c, const char *s)
1111 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1112 rnp->completed, c, rnp->level,
1113 rnp->grplo, rnp->grphi, s);
1117 * Start some future grace period, as needed to handle newly arrived
1118 * callbacks. The required future grace periods are recorded in each
1119 * rcu_node structure's ->need_future_gp field.
1121 * The caller must hold the specified rcu_node structure's ->lock.
1123 static unsigned long __maybe_unused
1124 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1128 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1131 * Pick up grace-period number for new callbacks. If this
1132 * grace period is already marked as needed, return to the caller.
1134 c = rcu_cbs_completed(rdp->rsp, rnp);
1135 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1136 if (rnp->need_future_gp[c & 0x1]) {
1137 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1142 * If either this rcu_node structure or the root rcu_node structure
1143 * believe that a grace period is in progress, then we must wait
1144 * for the one following, which is in "c". Because our request
1145 * will be noticed at the end of the current grace period, we don't
1146 * need to explicitly start one.
1148 if (rnp->gpnum != rnp->completed ||
1149 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1150 rnp->need_future_gp[c & 0x1]++;
1151 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1156 * There might be no grace period in progress. If we don't already
1157 * hold it, acquire the root rcu_node structure's lock in order to
1158 * start one (if needed).
1160 if (rnp != rnp_root)
1161 raw_spin_lock(&rnp_root->lock);
1164 * Get a new grace-period number. If there really is no grace
1165 * period in progress, it will be smaller than the one we obtained
1166 * earlier. Adjust callbacks as needed. Note that even no-CBs
1167 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1169 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1170 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1171 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1172 rdp->nxtcompleted[i] = c;
1175 * If the needed for the required grace period is already
1176 * recorded, trace and leave.
1178 if (rnp_root->need_future_gp[c & 0x1]) {
1179 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1183 /* Record the need for the future grace period. */
1184 rnp_root->need_future_gp[c & 0x1]++;
1186 /* If a grace period is not already in progress, start one. */
1187 if (rnp_root->gpnum != rnp_root->completed) {
1188 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1190 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1191 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1194 if (rnp != rnp_root)
1195 raw_spin_unlock(&rnp_root->lock);
1200 * Clean up any old requests for the just-ended grace period. Also return
1201 * whether any additional grace periods have been requested. Also invoke
1202 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1203 * waiting for this grace period to complete.
1205 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1207 int c = rnp->completed;
1209 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1211 rcu_nocb_gp_cleanup(rsp, rnp);
1212 rnp->need_future_gp[c & 0x1] = 0;
1213 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1214 trace_rcu_future_gp(rnp, rdp, c,
1215 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1220 * If there is room, assign a ->completed number to any callbacks on
1221 * this CPU that have not already been assigned. Also accelerate any
1222 * callbacks that were previously assigned a ->completed number that has
1223 * since proven to be too conservative, which can happen if callbacks get
1224 * assigned a ->completed number while RCU is idle, but with reference to
1225 * a non-root rcu_node structure. This function is idempotent, so it does
1226 * not hurt to call it repeatedly.
1228 * The caller must hold rnp->lock with interrupts disabled.
1230 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1231 struct rcu_data *rdp)
1236 /* If the CPU has no callbacks, nothing to do. */
1237 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1241 * Starting from the sublist containing the callbacks most
1242 * recently assigned a ->completed number and working down, find the
1243 * first sublist that is not assignable to an upcoming grace period.
1244 * Such a sublist has something in it (first two tests) and has
1245 * a ->completed number assigned that will complete sooner than
1246 * the ->completed number for newly arrived callbacks (last test).
1248 * The key point is that any later sublist can be assigned the
1249 * same ->completed number as the newly arrived callbacks, which
1250 * means that the callbacks in any of these later sublist can be
1251 * grouped into a single sublist, whether or not they have already
1252 * been assigned a ->completed number.
1254 c = rcu_cbs_completed(rsp, rnp);
1255 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1256 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1257 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1261 * If there are no sublist for unassigned callbacks, leave.
1262 * At the same time, advance "i" one sublist, so that "i" will
1263 * index into the sublist where all the remaining callbacks should
1266 if (++i >= RCU_NEXT_TAIL)
1270 * Assign all subsequent callbacks' ->completed number to the next
1271 * full grace period and group them all in the sublist initially
1274 for (; i <= RCU_NEXT_TAIL; i++) {
1275 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1276 rdp->nxtcompleted[i] = c;
1278 /* Record any needed additional grace periods. */
1279 rcu_start_future_gp(rnp, rdp);
1281 /* Trace depending on how much we were able to accelerate. */
1282 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1283 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1285 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1289 * Move any callbacks whose grace period has completed to the
1290 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1291 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1292 * sublist. This function is idempotent, so it does not hurt to
1293 * invoke it repeatedly. As long as it is not invoked -too- often...
1295 * The caller must hold rnp->lock with interrupts disabled.
1297 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1298 struct rcu_data *rdp)
1302 /* If the CPU has no callbacks, nothing to do. */
1303 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1307 * Find all callbacks whose ->completed numbers indicate that they
1308 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1310 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1311 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1313 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1315 /* Clean up any sublist tail pointers that were misordered above. */
1316 for (j = RCU_WAIT_TAIL; j < i; j++)
1317 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1319 /* Copy down callbacks to fill in empty sublists. */
1320 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1321 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1323 rdp->nxttail[j] = rdp->nxttail[i];
1324 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1327 /* Classify any remaining callbacks. */
1328 rcu_accelerate_cbs(rsp, rnp, rdp);
1332 * Update CPU-local rcu_data state to record the beginnings and ends of
1333 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1334 * structure corresponding to the current CPU, and must have irqs disabled.
1336 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1338 /* Handle the ends of any preceding grace periods first. */
1339 if (rdp->completed == rnp->completed) {
1341 /* No grace period end, so just accelerate recent callbacks. */
1342 rcu_accelerate_cbs(rsp, rnp, rdp);
1346 /* Advance callbacks. */
1347 rcu_advance_cbs(rsp, rnp, rdp);
1349 /* Remember that we saw this grace-period completion. */
1350 rdp->completed = rnp->completed;
1351 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1354 if (rdp->gpnum != rnp->gpnum) {
1356 * If the current grace period is waiting for this CPU,
1357 * set up to detect a quiescent state, otherwise don't
1358 * go looking for one.
1360 rdp->gpnum = rnp->gpnum;
1361 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1362 rdp->passed_quiesce = 0;
1363 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1364 zero_cpu_stall_ticks(rdp);
1368 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1370 unsigned long flags;
1371 struct rcu_node *rnp;
1373 local_irq_save(flags);
1375 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1376 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1377 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1378 local_irq_restore(flags);
1381 __note_gp_changes(rsp, rnp, rdp);
1382 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1386 * Initialize a new grace period. Return 0 if no grace period required.
1388 static int rcu_gp_init(struct rcu_state *rsp)
1390 struct rcu_data *rdp;
1391 struct rcu_node *rnp = rcu_get_root(rsp);
1393 rcu_bind_gp_kthread();
1394 raw_spin_lock_irq(&rnp->lock);
1395 if (rsp->gp_flags == 0) {
1396 /* Spurious wakeup, tell caller to go back to sleep. */
1397 raw_spin_unlock_irq(&rnp->lock);
1400 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1402 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1404 * Grace period already in progress, don't start another.
1405 * Not supposed to be able to happen.
1407 raw_spin_unlock_irq(&rnp->lock);
1411 /* Advance to a new grace period and initialize state. */
1412 record_gp_stall_check_time(rsp);
1413 smp_wmb(); /* Record GP times before starting GP. */
1415 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1416 raw_spin_unlock_irq(&rnp->lock);
1418 /* Exclude any concurrent CPU-hotplug operations. */
1419 mutex_lock(&rsp->onoff_mutex);
1422 * Set the quiescent-state-needed bits in all the rcu_node
1423 * structures for all currently online CPUs in breadth-first order,
1424 * starting from the root rcu_node structure, relying on the layout
1425 * of the tree within the rsp->node[] array. Note that other CPUs
1426 * will access only the leaves of the hierarchy, thus seeing that no
1427 * grace period is in progress, at least until the corresponding
1428 * leaf node has been initialized. In addition, we have excluded
1429 * CPU-hotplug operations.
1431 * The grace period cannot complete until the initialization
1432 * process finishes, because this kthread handles both.
1434 rcu_for_each_node_breadth_first(rsp, rnp) {
1435 raw_spin_lock_irq(&rnp->lock);
1436 rdp = this_cpu_ptr(rsp->rda);
1437 rcu_preempt_check_blocked_tasks(rnp);
1438 rnp->qsmask = rnp->qsmaskinit;
1439 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1440 WARN_ON_ONCE(rnp->completed != rsp->completed);
1441 ACCESS_ONCE(rnp->completed) = rsp->completed;
1442 if (rnp == rdp->mynode)
1443 __note_gp_changes(rsp, rnp, rdp);
1444 rcu_preempt_boost_start_gp(rnp);
1445 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1446 rnp->level, rnp->grplo,
1447 rnp->grphi, rnp->qsmask);
1448 raw_spin_unlock_irq(&rnp->lock);
1449 #ifdef CONFIG_PROVE_RCU_DELAY
1450 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1451 system_state == SYSTEM_RUNNING)
1453 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1457 mutex_unlock(&rsp->onoff_mutex);
1462 * Do one round of quiescent-state forcing.
1464 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1466 int fqs_state = fqs_state_in;
1467 bool isidle = false;
1469 struct rcu_node *rnp = rcu_get_root(rsp);
1472 if (fqs_state == RCU_SAVE_DYNTICK) {
1473 /* Collect dyntick-idle snapshots. */
1474 if (is_sysidle_rcu_state(rsp)) {
1476 maxj = jiffies - ULONG_MAX / 4;
1478 force_qs_rnp(rsp, dyntick_save_progress_counter,
1480 rcu_sysidle_report_gp(rsp, isidle, maxj);
1481 fqs_state = RCU_FORCE_QS;
1483 /* Handle dyntick-idle and offline CPUs. */
1485 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1487 /* Clear flag to prevent immediate re-entry. */
1488 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1489 raw_spin_lock_irq(&rnp->lock);
1490 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1491 raw_spin_unlock_irq(&rnp->lock);
1497 * Clean up after the old grace period.
1499 static void rcu_gp_cleanup(struct rcu_state *rsp)
1501 unsigned long gp_duration;
1503 struct rcu_data *rdp;
1504 struct rcu_node *rnp = rcu_get_root(rsp);
1506 raw_spin_lock_irq(&rnp->lock);
1507 gp_duration = jiffies - rsp->gp_start;
1508 if (gp_duration > rsp->gp_max)
1509 rsp->gp_max = gp_duration;
1512 * We know the grace period is complete, but to everyone else
1513 * it appears to still be ongoing. But it is also the case
1514 * that to everyone else it looks like there is nothing that
1515 * they can do to advance the grace period. It is therefore
1516 * safe for us to drop the lock in order to mark the grace
1517 * period as completed in all of the rcu_node structures.
1519 raw_spin_unlock_irq(&rnp->lock);
1522 * Propagate new ->completed value to rcu_node structures so
1523 * that other CPUs don't have to wait until the start of the next
1524 * grace period to process their callbacks. This also avoids
1525 * some nasty RCU grace-period initialization races by forcing
1526 * the end of the current grace period to be completely recorded in
1527 * all of the rcu_node structures before the beginning of the next
1528 * grace period is recorded in any of the rcu_node structures.
1530 rcu_for_each_node_breadth_first(rsp, rnp) {
1531 raw_spin_lock_irq(&rnp->lock);
1532 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1533 rdp = this_cpu_ptr(rsp->rda);
1534 if (rnp == rdp->mynode)
1535 __note_gp_changes(rsp, rnp, rdp);
1536 /* smp_mb() provided by prior unlock-lock pair. */
1537 nocb += rcu_future_gp_cleanup(rsp, rnp);
1538 raw_spin_unlock_irq(&rnp->lock);
1541 rnp = rcu_get_root(rsp);
1542 raw_spin_lock_irq(&rnp->lock);
1543 rcu_nocb_gp_set(rnp, nocb);
1545 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1546 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1547 rsp->fqs_state = RCU_GP_IDLE;
1548 rdp = this_cpu_ptr(rsp->rda);
1549 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1550 if (cpu_needs_another_gp(rsp, rdp)) {
1551 rsp->gp_flags = RCU_GP_FLAG_INIT;
1552 trace_rcu_grace_period(rsp->name,
1553 ACCESS_ONCE(rsp->gpnum),
1556 raw_spin_unlock_irq(&rnp->lock);
1560 * Body of kthread that handles grace periods.
1562 static int __noreturn rcu_gp_kthread(void *arg)
1568 struct rcu_state *rsp = arg;
1569 struct rcu_node *rnp = rcu_get_root(rsp);
1573 /* Handle grace-period start. */
1575 trace_rcu_grace_period(rsp->name,
1576 ACCESS_ONCE(rsp->gpnum),
1578 wait_event_interruptible(rsp->gp_wq,
1579 ACCESS_ONCE(rsp->gp_flags) &
1581 /* Locking provides needed memory barrier. */
1582 if (rcu_gp_init(rsp))
1585 flush_signals(current);
1586 trace_rcu_grace_period(rsp->name,
1587 ACCESS_ONCE(rsp->gpnum),
1591 /* Handle quiescent-state forcing. */
1592 fqs_state = RCU_SAVE_DYNTICK;
1593 j = jiffies_till_first_fqs;
1596 jiffies_till_first_fqs = HZ;
1601 rsp->jiffies_force_qs = jiffies + j;
1602 trace_rcu_grace_period(rsp->name,
1603 ACCESS_ONCE(rsp->gpnum),
1605 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1606 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1608 (!ACCESS_ONCE(rnp->qsmask) &&
1609 !rcu_preempt_blocked_readers_cgp(rnp)),
1611 /* Locking provides needed memory barriers. */
1612 /* If grace period done, leave loop. */
1613 if (!ACCESS_ONCE(rnp->qsmask) &&
1614 !rcu_preempt_blocked_readers_cgp(rnp))
1616 /* If time for quiescent-state forcing, do it. */
1617 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1618 (gf & RCU_GP_FLAG_FQS)) {
1619 trace_rcu_grace_period(rsp->name,
1620 ACCESS_ONCE(rsp->gpnum),
1622 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1623 trace_rcu_grace_period(rsp->name,
1624 ACCESS_ONCE(rsp->gpnum),
1628 /* Deal with stray signal. */
1630 flush_signals(current);
1631 trace_rcu_grace_period(rsp->name,
1632 ACCESS_ONCE(rsp->gpnum),
1635 j = jiffies_till_next_fqs;
1638 jiffies_till_next_fqs = HZ;
1641 jiffies_till_next_fqs = 1;
1645 /* Handle grace-period end. */
1646 rcu_gp_cleanup(rsp);
1650 static void rsp_wakeup(struct irq_work *work)
1652 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1654 /* Wake up rcu_gp_kthread() to start the grace period. */
1655 wake_up(&rsp->gp_wq);
1659 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1660 * in preparation for detecting the next grace period. The caller must hold
1661 * the root node's ->lock and hard irqs must be disabled.
1663 * Note that it is legal for a dying CPU (which is marked as offline) to
1664 * invoke this function. This can happen when the dying CPU reports its
1668 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1669 struct rcu_data *rdp)
1671 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1673 * Either we have not yet spawned the grace-period
1674 * task, this CPU does not need another grace period,
1675 * or a grace period is already in progress.
1676 * Either way, don't start a new grace period.
1680 rsp->gp_flags = RCU_GP_FLAG_INIT;
1681 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1685 * We can't do wakeups while holding the rnp->lock, as that
1686 * could cause possible deadlocks with the rq->lock. Defer
1687 * the wakeup to interrupt context. And don't bother waking
1688 * up the running kthread.
1690 if (current != rsp->gp_kthread)
1691 irq_work_queue(&rsp->wakeup_work);
1695 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1696 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1697 * is invoked indirectly from rcu_advance_cbs(), which would result in
1698 * endless recursion -- or would do so if it wasn't for the self-deadlock
1699 * that is encountered beforehand.
1702 rcu_start_gp(struct rcu_state *rsp)
1704 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1705 struct rcu_node *rnp = rcu_get_root(rsp);
1708 * If there is no grace period in progress right now, any
1709 * callbacks we have up to this point will be satisfied by the
1710 * next grace period. Also, advancing the callbacks reduces the
1711 * probability of false positives from cpu_needs_another_gp()
1712 * resulting in pointless grace periods. So, advance callbacks
1713 * then start the grace period!
1715 rcu_advance_cbs(rsp, rnp, rdp);
1716 rcu_start_gp_advanced(rsp, rnp, rdp);
1720 * Report a full set of quiescent states to the specified rcu_state
1721 * data structure. This involves cleaning up after the prior grace
1722 * period and letting rcu_start_gp() start up the next grace period
1723 * if one is needed. Note that the caller must hold rnp->lock, which
1724 * is released before return.
1726 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1727 __releases(rcu_get_root(rsp)->lock)
1729 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1730 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1731 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1735 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1736 * Allows quiescent states for a group of CPUs to be reported at one go
1737 * to the specified rcu_node structure, though all the CPUs in the group
1738 * must be represented by the same rcu_node structure (which need not be
1739 * a leaf rcu_node structure, though it often will be). That structure's
1740 * lock must be held upon entry, and it is released before return.
1743 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1744 struct rcu_node *rnp, unsigned long flags)
1745 __releases(rnp->lock)
1747 struct rcu_node *rnp_c;
1749 /* Walk up the rcu_node hierarchy. */
1751 if (!(rnp->qsmask & mask)) {
1753 /* Our bit has already been cleared, so done. */
1754 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1757 rnp->qsmask &= ~mask;
1758 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1759 mask, rnp->qsmask, rnp->level,
1760 rnp->grplo, rnp->grphi,
1762 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1764 /* Other bits still set at this level, so done. */
1765 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1768 mask = rnp->grpmask;
1769 if (rnp->parent == NULL) {
1771 /* No more levels. Exit loop holding root lock. */
1775 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1778 raw_spin_lock_irqsave(&rnp->lock, flags);
1779 WARN_ON_ONCE(rnp_c->qsmask);
1783 * Get here if we are the last CPU to pass through a quiescent
1784 * state for this grace period. Invoke rcu_report_qs_rsp()
1785 * to clean up and start the next grace period if one is needed.
1787 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1791 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1792 * structure. This must be either called from the specified CPU, or
1793 * called when the specified CPU is known to be offline (and when it is
1794 * also known that no other CPU is concurrently trying to help the offline
1795 * CPU). The lastcomp argument is used to make sure we are still in the
1796 * grace period of interest. We don't want to end the current grace period
1797 * based on quiescent states detected in an earlier grace period!
1800 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1802 unsigned long flags;
1804 struct rcu_node *rnp;
1807 raw_spin_lock_irqsave(&rnp->lock, flags);
1808 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1809 rnp->completed == rnp->gpnum) {
1812 * The grace period in which this quiescent state was
1813 * recorded has ended, so don't report it upwards.
1814 * We will instead need a new quiescent state that lies
1815 * within the current grace period.
1817 rdp->passed_quiesce = 0; /* need qs for new gp. */
1818 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1821 mask = rdp->grpmask;
1822 if ((rnp->qsmask & mask) == 0) {
1823 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1825 rdp->qs_pending = 0;
1828 * This GP can't end until cpu checks in, so all of our
1829 * callbacks can be processed during the next GP.
1831 rcu_accelerate_cbs(rsp, rnp, rdp);
1833 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1838 * Check to see if there is a new grace period of which this CPU
1839 * is not yet aware, and if so, set up local rcu_data state for it.
1840 * Otherwise, see if this CPU has just passed through its first
1841 * quiescent state for this grace period, and record that fact if so.
1844 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1846 /* Check for grace-period ends and beginnings. */
1847 note_gp_changes(rsp, rdp);
1850 * Does this CPU still need to do its part for current grace period?
1851 * If no, return and let the other CPUs do their part as well.
1853 if (!rdp->qs_pending)
1857 * Was there a quiescent state since the beginning of the grace
1858 * period? If no, then exit and wait for the next call.
1860 if (!rdp->passed_quiesce)
1864 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1867 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1870 #ifdef CONFIG_HOTPLUG_CPU
1873 * Send the specified CPU's RCU callbacks to the orphanage. The
1874 * specified CPU must be offline, and the caller must hold the
1878 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1879 struct rcu_node *rnp, struct rcu_data *rdp)
1881 /* No-CBs CPUs do not have orphanable callbacks. */
1882 if (rcu_is_nocb_cpu(rdp->cpu))
1886 * Orphan the callbacks. First adjust the counts. This is safe
1887 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1888 * cannot be running now. Thus no memory barrier is required.
1890 if (rdp->nxtlist != NULL) {
1891 rsp->qlen_lazy += rdp->qlen_lazy;
1892 rsp->qlen += rdp->qlen;
1893 rdp->n_cbs_orphaned += rdp->qlen;
1895 ACCESS_ONCE(rdp->qlen) = 0;
1899 * Next, move those callbacks still needing a grace period to
1900 * the orphanage, where some other CPU will pick them up.
1901 * Some of the callbacks might have gone partway through a grace
1902 * period, but that is too bad. They get to start over because we
1903 * cannot assume that grace periods are synchronized across CPUs.
1904 * We don't bother updating the ->nxttail[] array yet, instead
1905 * we just reset the whole thing later on.
1907 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1908 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1909 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1910 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1914 * Then move the ready-to-invoke callbacks to the orphanage,
1915 * where some other CPU will pick them up. These will not be
1916 * required to pass though another grace period: They are done.
1918 if (rdp->nxtlist != NULL) {
1919 *rsp->orphan_donetail = rdp->nxtlist;
1920 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1923 /* Finally, initialize the rcu_data structure's list to empty. */
1924 init_callback_list(rdp);
1928 * Adopt the RCU callbacks from the specified rcu_state structure's
1929 * orphanage. The caller must hold the ->orphan_lock.
1931 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1934 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1936 /* No-CBs CPUs are handled specially. */
1937 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1940 /* Do the accounting first. */
1941 rdp->qlen_lazy += rsp->qlen_lazy;
1942 rdp->qlen += rsp->qlen;
1943 rdp->n_cbs_adopted += rsp->qlen;
1944 if (rsp->qlen_lazy != rsp->qlen)
1945 rcu_idle_count_callbacks_posted();
1950 * We do not need a memory barrier here because the only way we
1951 * can get here if there is an rcu_barrier() in flight is if
1952 * we are the task doing the rcu_barrier().
1955 /* First adopt the ready-to-invoke callbacks. */
1956 if (rsp->orphan_donelist != NULL) {
1957 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1958 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1959 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1960 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1961 rdp->nxttail[i] = rsp->orphan_donetail;
1962 rsp->orphan_donelist = NULL;
1963 rsp->orphan_donetail = &rsp->orphan_donelist;
1966 /* And then adopt the callbacks that still need a grace period. */
1967 if (rsp->orphan_nxtlist != NULL) {
1968 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1969 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1970 rsp->orphan_nxtlist = NULL;
1971 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1976 * Trace the fact that this CPU is going offline.
1978 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1980 RCU_TRACE(unsigned long mask);
1981 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1982 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1984 RCU_TRACE(mask = rdp->grpmask);
1985 trace_rcu_grace_period(rsp->name,
1986 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1991 * The CPU has been completely removed, and some other CPU is reporting
1992 * this fact from process context. Do the remainder of the cleanup,
1993 * including orphaning the outgoing CPU's RCU callbacks, and also
1994 * adopting them. There can only be one CPU hotplug operation at a time,
1995 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1997 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1999 unsigned long flags;
2001 int need_report = 0;
2002 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2003 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2005 /* Adjust any no-longer-needed kthreads. */
2006 rcu_boost_kthread_setaffinity(rnp, -1);
2008 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2010 /* Exclude any attempts to start a new grace period. */
2011 mutex_lock(&rsp->onoff_mutex);
2012 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2014 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2015 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2016 rcu_adopt_orphan_cbs(rsp);
2018 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2019 mask = rdp->grpmask; /* rnp->grplo is constant. */
2021 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2022 rnp->qsmaskinit &= ~mask;
2023 if (rnp->qsmaskinit != 0) {
2024 if (rnp != rdp->mynode)
2025 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2028 if (rnp == rdp->mynode)
2029 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2031 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2032 mask = rnp->grpmask;
2034 } while (rnp != NULL);
2037 * We still hold the leaf rcu_node structure lock here, and
2038 * irqs are still disabled. The reason for this subterfuge is
2039 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2040 * held leads to deadlock.
2042 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2044 if (need_report & RCU_OFL_TASKS_NORM_GP)
2045 rcu_report_unblock_qs_rnp(rnp, flags);
2047 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2048 if (need_report & RCU_OFL_TASKS_EXP_GP)
2049 rcu_report_exp_rnp(rsp, rnp, true);
2050 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2051 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2052 cpu, rdp->qlen, rdp->nxtlist);
2053 init_callback_list(rdp);
2054 /* Disallow further callbacks on this CPU. */
2055 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2056 mutex_unlock(&rsp->onoff_mutex);
2059 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2061 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2065 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2069 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2072 * Invoke any RCU callbacks that have made it to the end of their grace
2073 * period. Thottle as specified by rdp->blimit.
2075 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2077 unsigned long flags;
2078 struct rcu_head *next, *list, **tail;
2079 long bl, count, count_lazy;
2082 /* If no callbacks are ready, just return. */
2083 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2084 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2085 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2086 need_resched(), is_idle_task(current),
2087 rcu_is_callbacks_kthread());
2092 * Extract the list of ready callbacks, disabling to prevent
2093 * races with call_rcu() from interrupt handlers.
2095 local_irq_save(flags);
2096 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2098 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2099 list = rdp->nxtlist;
2100 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2101 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2102 tail = rdp->nxttail[RCU_DONE_TAIL];
2103 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2104 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2105 rdp->nxttail[i] = &rdp->nxtlist;
2106 local_irq_restore(flags);
2108 /* Invoke callbacks. */
2109 count = count_lazy = 0;
2113 debug_rcu_head_unqueue(list);
2114 if (__rcu_reclaim(rsp->name, list))
2117 /* Stop only if limit reached and CPU has something to do. */
2118 if (++count >= bl &&
2120 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2124 local_irq_save(flags);
2125 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2126 is_idle_task(current),
2127 rcu_is_callbacks_kthread());
2129 /* Update count, and requeue any remaining callbacks. */
2131 *tail = rdp->nxtlist;
2132 rdp->nxtlist = list;
2133 for (i = 0; i < RCU_NEXT_SIZE; i++)
2134 if (&rdp->nxtlist == rdp->nxttail[i])
2135 rdp->nxttail[i] = tail;
2139 smp_mb(); /* List handling before counting for rcu_barrier(). */
2140 rdp->qlen_lazy -= count_lazy;
2141 ACCESS_ONCE(rdp->qlen) -= count;
2142 rdp->n_cbs_invoked += count;
2144 /* Reinstate batch limit if we have worked down the excess. */
2145 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2146 rdp->blimit = blimit;
2148 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2149 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2150 rdp->qlen_last_fqs_check = 0;
2151 rdp->n_force_qs_snap = rsp->n_force_qs;
2152 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2153 rdp->qlen_last_fqs_check = rdp->qlen;
2154 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2156 local_irq_restore(flags);
2158 /* Re-invoke RCU core processing if there are callbacks remaining. */
2159 if (cpu_has_callbacks_ready_to_invoke(rdp))
2164 * Check to see if this CPU is in a non-context-switch quiescent state
2165 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2166 * Also schedule RCU core processing.
2168 * This function must be called from hardirq context. It is normally
2169 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2170 * false, there is no point in invoking rcu_check_callbacks().
2172 void rcu_check_callbacks(int cpu, int user)
2174 trace_rcu_utilization(TPS("Start scheduler-tick"));
2175 increment_cpu_stall_ticks();
2176 if (user || rcu_is_cpu_rrupt_from_idle()) {
2179 * Get here if this CPU took its interrupt from user
2180 * mode or from the idle loop, and if this is not a
2181 * nested interrupt. In this case, the CPU is in
2182 * a quiescent state, so note it.
2184 * No memory barrier is required here because both
2185 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2186 * variables that other CPUs neither access nor modify,
2187 * at least not while the corresponding CPU is online.
2193 } else if (!in_softirq()) {
2196 * Get here if this CPU did not take its interrupt from
2197 * softirq, in other words, if it is not interrupting
2198 * a rcu_bh read-side critical section. This is an _bh
2199 * critical section, so note it.
2204 rcu_preempt_check_callbacks(cpu);
2205 if (rcu_pending(cpu))
2207 trace_rcu_utilization(TPS("End scheduler-tick"));
2211 * Scan the leaf rcu_node structures, processing dyntick state for any that
2212 * have not yet encountered a quiescent state, using the function specified.
2213 * Also initiate boosting for any threads blocked on the root rcu_node.
2215 * The caller must have suppressed start of new grace periods.
2217 static void force_qs_rnp(struct rcu_state *rsp,
2218 int (*f)(struct rcu_data *rsp, bool *isidle,
2219 unsigned long *maxj),
2220 bool *isidle, unsigned long *maxj)
2224 unsigned long flags;
2226 struct rcu_node *rnp;
2228 rcu_for_each_leaf_node(rsp, rnp) {
2231 raw_spin_lock_irqsave(&rnp->lock, flags);
2232 if (!rcu_gp_in_progress(rsp)) {
2233 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2236 if (rnp->qsmask == 0) {
2237 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2242 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2243 if ((rnp->qsmask & bit) != 0) {
2244 if ((rnp->qsmaskinit & bit) != 0)
2246 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2252 /* rcu_report_qs_rnp() releases rnp->lock. */
2253 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2256 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2258 rnp = rcu_get_root(rsp);
2259 if (rnp->qsmask == 0) {
2260 raw_spin_lock_irqsave(&rnp->lock, flags);
2261 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2266 * Force quiescent states on reluctant CPUs, and also detect which
2267 * CPUs are in dyntick-idle mode.
2269 static void force_quiescent_state(struct rcu_state *rsp)
2271 unsigned long flags;
2273 struct rcu_node *rnp;
2274 struct rcu_node *rnp_old = NULL;
2276 /* Funnel through hierarchy to reduce memory contention. */
2277 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2278 for (; rnp != NULL; rnp = rnp->parent) {
2279 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2280 !raw_spin_trylock(&rnp->fqslock);
2281 if (rnp_old != NULL)
2282 raw_spin_unlock(&rnp_old->fqslock);
2284 rsp->n_force_qs_lh++;
2289 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2291 /* Reached the root of the rcu_node tree, acquire lock. */
2292 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2293 raw_spin_unlock(&rnp_old->fqslock);
2294 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2295 rsp->n_force_qs_lh++;
2296 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2297 return; /* Someone beat us to it. */
2299 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2300 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2301 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2305 * This does the RCU core processing work for the specified rcu_state
2306 * and rcu_data structures. This may be called only from the CPU to
2307 * whom the rdp belongs.
2310 __rcu_process_callbacks(struct rcu_state *rsp)
2312 unsigned long flags;
2313 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2315 WARN_ON_ONCE(rdp->beenonline == 0);
2317 /* Update RCU state based on any recent quiescent states. */
2318 rcu_check_quiescent_state(rsp, rdp);
2320 /* Does this CPU require a not-yet-started grace period? */
2321 local_irq_save(flags);
2322 if (cpu_needs_another_gp(rsp, rdp)) {
2323 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2325 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2327 local_irq_restore(flags);
2330 /* If there are callbacks ready, invoke them. */
2331 if (cpu_has_callbacks_ready_to_invoke(rdp))
2332 invoke_rcu_callbacks(rsp, rdp);
2336 * Do RCU core processing for the current CPU.
2338 static void rcu_process_callbacks(struct softirq_action *unused)
2340 struct rcu_state *rsp;
2342 if (cpu_is_offline(smp_processor_id()))
2344 trace_rcu_utilization(TPS("Start RCU core"));
2345 for_each_rcu_flavor(rsp)
2346 __rcu_process_callbacks(rsp);
2347 trace_rcu_utilization(TPS("End RCU core"));
2351 * Schedule RCU callback invocation. If the specified type of RCU
2352 * does not support RCU priority boosting, just do a direct call,
2353 * otherwise wake up the per-CPU kernel kthread. Note that because we
2354 * are running on the current CPU with interrupts disabled, the
2355 * rcu_cpu_kthread_task cannot disappear out from under us.
2357 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2359 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2361 if (likely(!rsp->boost)) {
2362 rcu_do_batch(rsp, rdp);
2365 invoke_rcu_callbacks_kthread();
2368 static void invoke_rcu_core(void)
2370 if (cpu_online(smp_processor_id()))
2371 raise_softirq(RCU_SOFTIRQ);
2375 * Handle any core-RCU processing required by a call_rcu() invocation.
2377 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2378 struct rcu_head *head, unsigned long flags)
2381 * If called from an extended quiescent state, invoke the RCU
2382 * core in order to force a re-evaluation of RCU's idleness.
2384 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2387 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2388 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2392 * Force the grace period if too many callbacks or too long waiting.
2393 * Enforce hysteresis, and don't invoke force_quiescent_state()
2394 * if some other CPU has recently done so. Also, don't bother
2395 * invoking force_quiescent_state() if the newly enqueued callback
2396 * is the only one waiting for a grace period to complete.
2398 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2400 /* Are we ignoring a completed grace period? */
2401 note_gp_changes(rsp, rdp);
2403 /* Start a new grace period if one not already started. */
2404 if (!rcu_gp_in_progress(rsp)) {
2405 struct rcu_node *rnp_root = rcu_get_root(rsp);
2407 raw_spin_lock(&rnp_root->lock);
2409 raw_spin_unlock(&rnp_root->lock);
2411 /* Give the grace period a kick. */
2412 rdp->blimit = LONG_MAX;
2413 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2414 *rdp->nxttail[RCU_DONE_TAIL] != head)
2415 force_quiescent_state(rsp);
2416 rdp->n_force_qs_snap = rsp->n_force_qs;
2417 rdp->qlen_last_fqs_check = rdp->qlen;
2423 * RCU callback function to leak a callback.
2425 static void rcu_leak_callback(struct rcu_head *rhp)
2430 * Helper function for call_rcu() and friends. The cpu argument will
2431 * normally be -1, indicating "currently running CPU". It may specify
2432 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2433 * is expected to specify a CPU.
2436 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2437 struct rcu_state *rsp, int cpu, bool lazy)
2439 unsigned long flags;
2440 struct rcu_data *rdp;
2442 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2443 if (debug_rcu_head_queue(head)) {
2444 /* Probable double call_rcu(), so leak the callback. */
2445 ACCESS_ONCE(head->func) = rcu_leak_callback;
2446 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2453 * Opportunistically note grace-period endings and beginnings.
2454 * Note that we might see a beginning right after we see an
2455 * end, but never vice versa, since this CPU has to pass through
2456 * a quiescent state betweentimes.
2458 local_irq_save(flags);
2459 rdp = this_cpu_ptr(rsp->rda);
2461 /* Add the callback to our list. */
2462 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2466 rdp = per_cpu_ptr(rsp->rda, cpu);
2467 offline = !__call_rcu_nocb(rdp, head, lazy);
2468 WARN_ON_ONCE(offline);
2469 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2470 local_irq_restore(flags);
2473 ACCESS_ONCE(rdp->qlen)++;
2477 rcu_idle_count_callbacks_posted();
2478 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2479 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2480 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2482 if (__is_kfree_rcu_offset((unsigned long)func))
2483 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2484 rdp->qlen_lazy, rdp->qlen);
2486 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2488 /* Go handle any RCU core processing required. */
2489 __call_rcu_core(rsp, rdp, head, flags);
2490 local_irq_restore(flags);
2494 * Queue an RCU-sched callback for invocation after a grace period.
2496 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2498 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2500 EXPORT_SYMBOL_GPL(call_rcu_sched);
2503 * Queue an RCU callback for invocation after a quicker grace period.
2505 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2507 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2509 EXPORT_SYMBOL_GPL(call_rcu_bh);
2512 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2513 * any blocking grace-period wait automatically implies a grace period
2514 * if there is only one CPU online at any point time during execution
2515 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2516 * occasionally incorrectly indicate that there are multiple CPUs online
2517 * when there was in fact only one the whole time, as this just adds
2518 * some overhead: RCU still operates correctly.
2520 static inline int rcu_blocking_is_gp(void)
2524 might_sleep(); /* Check for RCU read-side critical section. */
2526 ret = num_online_cpus() <= 1;
2532 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2534 * Control will return to the caller some time after a full rcu-sched
2535 * grace period has elapsed, in other words after all currently executing
2536 * rcu-sched read-side critical sections have completed. These read-side
2537 * critical sections are delimited by rcu_read_lock_sched() and
2538 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2539 * local_irq_disable(), and so on may be used in place of
2540 * rcu_read_lock_sched().
2542 * This means that all preempt_disable code sequences, including NMI and
2543 * non-threaded hardware-interrupt handlers, in progress on entry will
2544 * have completed before this primitive returns. However, this does not
2545 * guarantee that softirq handlers will have completed, since in some
2546 * kernels, these handlers can run in process context, and can block.
2548 * Note that this guarantee implies further memory-ordering guarantees.
2549 * On systems with more than one CPU, when synchronize_sched() returns,
2550 * each CPU is guaranteed to have executed a full memory barrier since the
2551 * end of its last RCU-sched read-side critical section whose beginning
2552 * preceded the call to synchronize_sched(). In addition, each CPU having
2553 * an RCU read-side critical section that extends beyond the return from
2554 * synchronize_sched() is guaranteed to have executed a full memory barrier
2555 * after the beginning of synchronize_sched() and before the beginning of
2556 * that RCU read-side critical section. Note that these guarantees include
2557 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2558 * that are executing in the kernel.
2560 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2561 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2562 * to have executed a full memory barrier during the execution of
2563 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2564 * again only if the system has more than one CPU).
2566 * This primitive provides the guarantees made by the (now removed)
2567 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2568 * guarantees that rcu_read_lock() sections will have completed.
2569 * In "classic RCU", these two guarantees happen to be one and
2570 * the same, but can differ in realtime RCU implementations.
2572 void synchronize_sched(void)
2574 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2575 !lock_is_held(&rcu_lock_map) &&
2576 !lock_is_held(&rcu_sched_lock_map),
2577 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2578 if (rcu_blocking_is_gp())
2581 synchronize_sched_expedited();
2583 wait_rcu_gp(call_rcu_sched);
2585 EXPORT_SYMBOL_GPL(synchronize_sched);
2588 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2590 * Control will return to the caller some time after a full rcu_bh grace
2591 * period has elapsed, in other words after all currently executing rcu_bh
2592 * read-side critical sections have completed. RCU read-side critical
2593 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2594 * and may be nested.
2596 * See the description of synchronize_sched() for more detailed information
2597 * on memory ordering guarantees.
2599 void synchronize_rcu_bh(void)
2601 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2602 !lock_is_held(&rcu_lock_map) &&
2603 !lock_is_held(&rcu_sched_lock_map),
2604 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2605 if (rcu_blocking_is_gp())
2608 synchronize_rcu_bh_expedited();
2610 wait_rcu_gp(call_rcu_bh);
2612 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2614 static int synchronize_sched_expedited_cpu_stop(void *data)
2617 * There must be a full memory barrier on each affected CPU
2618 * between the time that try_stop_cpus() is called and the
2619 * time that it returns.
2621 * In the current initial implementation of cpu_stop, the
2622 * above condition is already met when the control reaches
2623 * this point and the following smp_mb() is not strictly
2624 * necessary. Do smp_mb() anyway for documentation and
2625 * robustness against future implementation changes.
2627 smp_mb(); /* See above comment block. */
2632 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2634 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2635 * approach to force the grace period to end quickly. This consumes
2636 * significant time on all CPUs and is unfriendly to real-time workloads,
2637 * so is thus not recommended for any sort of common-case code. In fact,
2638 * if you are using synchronize_sched_expedited() in a loop, please
2639 * restructure your code to batch your updates, and then use a single
2640 * synchronize_sched() instead.
2642 * Note that it is illegal to call this function while holding any lock
2643 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2644 * to call this function from a CPU-hotplug notifier. Failing to observe
2645 * these restriction will result in deadlock.
2647 * This implementation can be thought of as an application of ticket
2648 * locking to RCU, with sync_sched_expedited_started and
2649 * sync_sched_expedited_done taking on the roles of the halves
2650 * of the ticket-lock word. Each task atomically increments
2651 * sync_sched_expedited_started upon entry, snapshotting the old value,
2652 * then attempts to stop all the CPUs. If this succeeds, then each
2653 * CPU will have executed a context switch, resulting in an RCU-sched
2654 * grace period. We are then done, so we use atomic_cmpxchg() to
2655 * update sync_sched_expedited_done to match our snapshot -- but
2656 * only if someone else has not already advanced past our snapshot.
2658 * On the other hand, if try_stop_cpus() fails, we check the value
2659 * of sync_sched_expedited_done. If it has advanced past our
2660 * initial snapshot, then someone else must have forced a grace period
2661 * some time after we took our snapshot. In this case, our work is
2662 * done for us, and we can simply return. Otherwise, we try again,
2663 * but keep our initial snapshot for purposes of checking for someone
2664 * doing our work for us.
2666 * If we fail too many times in a row, we fall back to synchronize_sched().
2668 void synchronize_sched_expedited(void)
2670 long firstsnap, s, snap;
2672 struct rcu_state *rsp = &rcu_sched_state;
2675 * If we are in danger of counter wrap, just do synchronize_sched().
2676 * By allowing sync_sched_expedited_started to advance no more than
2677 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2678 * that more than 3.5 billion CPUs would be required to force a
2679 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2680 * course be required on a 64-bit system.
2682 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2683 (ulong)atomic_long_read(&rsp->expedited_done) +
2685 synchronize_sched();
2686 atomic_long_inc(&rsp->expedited_wrap);
2691 * Take a ticket. Note that atomic_inc_return() implies a
2692 * full memory barrier.
2694 snap = atomic_long_inc_return(&rsp->expedited_start);
2697 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2700 * Each pass through the following loop attempts to force a
2701 * context switch on each CPU.
2703 while (try_stop_cpus(cpu_online_mask,
2704 synchronize_sched_expedited_cpu_stop,
2707 atomic_long_inc(&rsp->expedited_tryfail);
2709 /* Check to see if someone else did our work for us. */
2710 s = atomic_long_read(&rsp->expedited_done);
2711 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2712 /* ensure test happens before caller kfree */
2713 smp_mb__before_atomic_inc(); /* ^^^ */
2714 atomic_long_inc(&rsp->expedited_workdone1);
2718 /* No joy, try again later. Or just synchronize_sched(). */
2719 if (trycount++ < 10) {
2720 udelay(trycount * num_online_cpus());
2722 wait_rcu_gp(call_rcu_sched);
2723 atomic_long_inc(&rsp->expedited_normal);
2727 /* Recheck to see if someone else did our work for us. */
2728 s = atomic_long_read(&rsp->expedited_done);
2729 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2730 /* ensure test happens before caller kfree */
2731 smp_mb__before_atomic_inc(); /* ^^^ */
2732 atomic_long_inc(&rsp->expedited_workdone2);
2737 * Refetching sync_sched_expedited_started allows later
2738 * callers to piggyback on our grace period. We retry
2739 * after they started, so our grace period works for them,
2740 * and they started after our first try, so their grace
2741 * period works for us.
2744 snap = atomic_long_read(&rsp->expedited_start);
2745 smp_mb(); /* ensure read is before try_stop_cpus(). */
2747 atomic_long_inc(&rsp->expedited_stoppedcpus);
2750 * Everyone up to our most recent fetch is covered by our grace
2751 * period. Update the counter, but only if our work is still
2752 * relevant -- which it won't be if someone who started later
2753 * than we did already did their update.
2756 atomic_long_inc(&rsp->expedited_done_tries);
2757 s = atomic_long_read(&rsp->expedited_done);
2758 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2759 /* ensure test happens before caller kfree */
2760 smp_mb__before_atomic_inc(); /* ^^^ */
2761 atomic_long_inc(&rsp->expedited_done_lost);
2764 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2765 atomic_long_inc(&rsp->expedited_done_exit);
2769 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2772 * Check to see if there is any immediate RCU-related work to be done
2773 * by the current CPU, for the specified type of RCU, returning 1 if so.
2774 * The checks are in order of increasing expense: checks that can be
2775 * carried out against CPU-local state are performed first. However,
2776 * we must check for CPU stalls first, else we might not get a chance.
2778 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2780 struct rcu_node *rnp = rdp->mynode;
2782 rdp->n_rcu_pending++;
2784 /* Check for CPU stalls, if enabled. */
2785 check_cpu_stall(rsp, rdp);
2787 /* Is the RCU core waiting for a quiescent state from this CPU? */
2788 if (rcu_scheduler_fully_active &&
2789 rdp->qs_pending && !rdp->passed_quiesce) {
2790 rdp->n_rp_qs_pending++;
2791 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2792 rdp->n_rp_report_qs++;
2796 /* Does this CPU have callbacks ready to invoke? */
2797 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2798 rdp->n_rp_cb_ready++;
2802 /* Has RCU gone idle with this CPU needing another grace period? */
2803 if (cpu_needs_another_gp(rsp, rdp)) {
2804 rdp->n_rp_cpu_needs_gp++;
2808 /* Has another RCU grace period completed? */
2809 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2810 rdp->n_rp_gp_completed++;
2814 /* Has a new RCU grace period started? */
2815 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2816 rdp->n_rp_gp_started++;
2821 rdp->n_rp_need_nothing++;
2826 * Check to see if there is any immediate RCU-related work to be done
2827 * by the current CPU, returning 1 if so. This function is part of the
2828 * RCU implementation; it is -not- an exported member of the RCU API.
2830 static int rcu_pending(int cpu)
2832 struct rcu_state *rsp;
2834 for_each_rcu_flavor(rsp)
2835 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2841 * Return true if the specified CPU has any callback. If all_lazy is
2842 * non-NULL, store an indication of whether all callbacks are lazy.
2843 * (If there are no callbacks, all of them are deemed to be lazy.)
2845 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2849 struct rcu_data *rdp;
2850 struct rcu_state *rsp;
2852 for_each_rcu_flavor(rsp) {
2853 rdp = per_cpu_ptr(rsp->rda, cpu);
2857 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2868 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2869 * the compiler is expected to optimize this away.
2871 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2872 int cpu, unsigned long done)
2874 trace_rcu_barrier(rsp->name, s, cpu,
2875 atomic_read(&rsp->barrier_cpu_count), done);
2879 * RCU callback function for _rcu_barrier(). If we are last, wake
2880 * up the task executing _rcu_barrier().
2882 static void rcu_barrier_callback(struct rcu_head *rhp)
2884 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2885 struct rcu_state *rsp = rdp->rsp;
2887 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2888 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2889 complete(&rsp->barrier_completion);
2891 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2896 * Called with preemption disabled, and from cross-cpu IRQ context.
2898 static void rcu_barrier_func(void *type)
2900 struct rcu_state *rsp = type;
2901 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2903 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2904 atomic_inc(&rsp->barrier_cpu_count);
2905 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2909 * Orchestrate the specified type of RCU barrier, waiting for all
2910 * RCU callbacks of the specified type to complete.
2912 static void _rcu_barrier(struct rcu_state *rsp)
2915 struct rcu_data *rdp;
2916 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2917 unsigned long snap_done;
2919 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2921 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2922 mutex_lock(&rsp->barrier_mutex);
2925 * Ensure that all prior references, including to ->n_barrier_done,
2926 * are ordered before the _rcu_barrier() machinery.
2928 smp_mb(); /* See above block comment. */
2931 * Recheck ->n_barrier_done to see if others did our work for us.
2932 * This means checking ->n_barrier_done for an even-to-odd-to-even
2933 * transition. The "if" expression below therefore rounds the old
2934 * value up to the next even number and adds two before comparing.
2936 snap_done = rsp->n_barrier_done;
2937 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2940 * If the value in snap is odd, we needed to wait for the current
2941 * rcu_barrier() to complete, then wait for the next one, in other
2942 * words, we need the value of snap_done to be three larger than
2943 * the value of snap. On the other hand, if the value in snap is
2944 * even, we only had to wait for the next rcu_barrier() to complete,
2945 * in other words, we need the value of snap_done to be only two
2946 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2947 * this for us (thank you, Linus!).
2949 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
2950 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2951 smp_mb(); /* caller's subsequent code after above check. */
2952 mutex_unlock(&rsp->barrier_mutex);
2957 * Increment ->n_barrier_done to avoid duplicate work. Use
2958 * ACCESS_ONCE() to prevent the compiler from speculating
2959 * the increment to precede the early-exit check.
2961 ACCESS_ONCE(rsp->n_barrier_done)++;
2962 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2963 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2964 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2967 * Initialize the count to one rather than to zero in order to
2968 * avoid a too-soon return to zero in case of a short grace period
2969 * (or preemption of this task). Exclude CPU-hotplug operations
2970 * to ensure that no offline CPU has callbacks queued.
2972 init_completion(&rsp->barrier_completion);
2973 atomic_set(&rsp->barrier_cpu_count, 1);
2977 * Force each CPU with callbacks to register a new callback.
2978 * When that callback is invoked, we will know that all of the
2979 * corresponding CPU's preceding callbacks have been invoked.
2981 for_each_possible_cpu(cpu) {
2982 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2984 rdp = per_cpu_ptr(rsp->rda, cpu);
2985 if (rcu_is_nocb_cpu(cpu)) {
2986 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2987 rsp->n_barrier_done);
2988 atomic_inc(&rsp->barrier_cpu_count);
2989 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2991 } else if (ACCESS_ONCE(rdp->qlen)) {
2992 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2993 rsp->n_barrier_done);
2994 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2996 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2997 rsp->n_barrier_done);
3003 * Now that we have an rcu_barrier_callback() callback on each
3004 * CPU, and thus each counted, remove the initial count.
3006 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3007 complete(&rsp->barrier_completion);
3009 /* Increment ->n_barrier_done to prevent duplicate work. */
3010 smp_mb(); /* Keep increment after above mechanism. */
3011 ACCESS_ONCE(rsp->n_barrier_done)++;
3012 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3013 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3014 smp_mb(); /* Keep increment before caller's subsequent code. */
3016 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3017 wait_for_completion(&rsp->barrier_completion);
3019 /* Other rcu_barrier() invocations can now safely proceed. */
3020 mutex_unlock(&rsp->barrier_mutex);
3024 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3026 void rcu_barrier_bh(void)
3028 _rcu_barrier(&rcu_bh_state);
3030 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3033 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3035 void rcu_barrier_sched(void)
3037 _rcu_barrier(&rcu_sched_state);
3039 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3042 * Do boot-time initialization of a CPU's per-CPU RCU data.
3045 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3047 unsigned long flags;
3048 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3049 struct rcu_node *rnp = rcu_get_root(rsp);
3051 /* Set up local state, ensuring consistent view of global state. */
3052 raw_spin_lock_irqsave(&rnp->lock, flags);
3053 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3054 init_callback_list(rdp);
3056 ACCESS_ONCE(rdp->qlen) = 0;
3057 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3058 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3059 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3062 rcu_boot_init_nocb_percpu_data(rdp);
3063 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3067 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3068 * offline event can be happening at a given time. Note also that we
3069 * can accept some slop in the rsp->completed access due to the fact
3070 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3073 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3075 unsigned long flags;
3077 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3078 struct rcu_node *rnp = rcu_get_root(rsp);
3080 /* Exclude new grace periods. */
3081 mutex_lock(&rsp->onoff_mutex);
3083 /* Set up local state, ensuring consistent view of global state. */
3084 raw_spin_lock_irqsave(&rnp->lock, flags);
3085 rdp->beenonline = 1; /* We have now been online. */
3086 rdp->preemptible = preemptible;
3087 rdp->qlen_last_fqs_check = 0;
3088 rdp->n_force_qs_snap = rsp->n_force_qs;
3089 rdp->blimit = blimit;
3090 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3091 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3092 rcu_sysidle_init_percpu_data(rdp->dynticks);
3093 atomic_set(&rdp->dynticks->dynticks,
3094 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3095 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3097 /* Add CPU to rcu_node bitmasks. */
3099 mask = rdp->grpmask;
3101 /* Exclude any attempts to start a new GP on small systems. */
3102 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3103 rnp->qsmaskinit |= mask;
3104 mask = rnp->grpmask;
3105 if (rnp == rdp->mynode) {
3107 * If there is a grace period in progress, we will
3108 * set up to wait for it next time we run the
3111 rdp->gpnum = rnp->completed;
3112 rdp->completed = rnp->completed;
3113 rdp->passed_quiesce = 0;
3114 rdp->qs_pending = 0;
3115 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3117 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3119 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3120 local_irq_restore(flags);
3122 mutex_unlock(&rsp->onoff_mutex);
3125 static void rcu_prepare_cpu(int cpu)
3127 struct rcu_state *rsp;
3129 for_each_rcu_flavor(rsp)
3130 rcu_init_percpu_data(cpu, rsp,
3131 strcmp(rsp->name, "rcu_preempt") == 0);
3135 * Handle CPU online/offline notification events.
3137 static int rcu_cpu_notify(struct notifier_block *self,
3138 unsigned long action, void *hcpu)
3140 long cpu = (long)hcpu;
3141 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3142 struct rcu_node *rnp = rdp->mynode;
3143 struct rcu_state *rsp;
3145 trace_rcu_utilization(TPS("Start CPU hotplug"));
3147 case CPU_UP_PREPARE:
3148 case CPU_UP_PREPARE_FROZEN:
3149 rcu_prepare_cpu(cpu);
3150 rcu_prepare_kthreads(cpu);
3153 case CPU_DOWN_FAILED:
3154 rcu_boost_kthread_setaffinity(rnp, -1);
3156 case CPU_DOWN_PREPARE:
3157 rcu_boost_kthread_setaffinity(rnp, cpu);
3160 case CPU_DYING_FROZEN:
3161 for_each_rcu_flavor(rsp)
3162 rcu_cleanup_dying_cpu(rsp);
3165 case CPU_DEAD_FROZEN:
3166 case CPU_UP_CANCELED:
3167 case CPU_UP_CANCELED_FROZEN:
3168 for_each_rcu_flavor(rsp)
3169 rcu_cleanup_dead_cpu(cpu, rsp);
3174 trace_rcu_utilization(TPS("End CPU hotplug"));
3178 static int rcu_pm_notify(struct notifier_block *self,
3179 unsigned long action, void *hcpu)
3182 case PM_HIBERNATION_PREPARE:
3183 case PM_SUSPEND_PREPARE:
3184 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3187 case PM_POST_HIBERNATION:
3188 case PM_POST_SUSPEND:
3198 * Spawn the kthread that handles this RCU flavor's grace periods.
3200 static int __init rcu_spawn_gp_kthread(void)
3202 unsigned long flags;
3203 struct rcu_node *rnp;
3204 struct rcu_state *rsp;
3205 struct task_struct *t;
3207 for_each_rcu_flavor(rsp) {
3208 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3210 rnp = rcu_get_root(rsp);
3211 raw_spin_lock_irqsave(&rnp->lock, flags);
3212 rsp->gp_kthread = t;
3213 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3214 rcu_spawn_nocb_kthreads(rsp);
3218 early_initcall(rcu_spawn_gp_kthread);
3221 * This function is invoked towards the end of the scheduler's initialization
3222 * process. Before this is called, the idle task might contain
3223 * RCU read-side critical sections (during which time, this idle
3224 * task is booting the system). After this function is called, the
3225 * idle tasks are prohibited from containing RCU read-side critical
3226 * sections. This function also enables RCU lockdep checking.
3228 void rcu_scheduler_starting(void)
3230 WARN_ON(num_online_cpus() != 1);
3231 WARN_ON(nr_context_switches() > 0);
3232 rcu_scheduler_active = 1;
3236 * Compute the per-level fanout, either using the exact fanout specified
3237 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3239 #ifdef CONFIG_RCU_FANOUT_EXACT
3240 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3244 for (i = rcu_num_lvls - 1; i > 0; i--)
3245 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3246 rsp->levelspread[0] = rcu_fanout_leaf;
3248 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3249 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3256 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3257 ccur = rsp->levelcnt[i];
3258 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3262 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3265 * Helper function for rcu_init() that initializes one rcu_state structure.
3267 static void __init rcu_init_one(struct rcu_state *rsp,
3268 struct rcu_data __percpu *rda)
3270 static char *buf[] = { "rcu_node_0",
3273 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3274 static char *fqs[] = { "rcu_node_fqs_0",
3277 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3281 struct rcu_node *rnp;
3283 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3285 /* Silence gcc 4.8 warning about array index out of range. */
3286 if (rcu_num_lvls > RCU_NUM_LVLS)
3287 panic("rcu_init_one: rcu_num_lvls overflow");
3289 /* Initialize the level-tracking arrays. */
3291 for (i = 0; i < rcu_num_lvls; i++)
3292 rsp->levelcnt[i] = num_rcu_lvl[i];
3293 for (i = 1; i < rcu_num_lvls; i++)
3294 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3295 rcu_init_levelspread(rsp);
3297 /* Initialize the elements themselves, starting from the leaves. */
3299 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3300 cpustride *= rsp->levelspread[i];
3301 rnp = rsp->level[i];
3302 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3303 raw_spin_lock_init(&rnp->lock);
3304 lockdep_set_class_and_name(&rnp->lock,
3305 &rcu_node_class[i], buf[i]);
3306 raw_spin_lock_init(&rnp->fqslock);
3307 lockdep_set_class_and_name(&rnp->fqslock,
3308 &rcu_fqs_class[i], fqs[i]);
3309 rnp->gpnum = rsp->gpnum;
3310 rnp->completed = rsp->completed;
3312 rnp->qsmaskinit = 0;
3313 rnp->grplo = j * cpustride;
3314 rnp->grphi = (j + 1) * cpustride - 1;
3315 if (rnp->grphi >= NR_CPUS)
3316 rnp->grphi = NR_CPUS - 1;
3322 rnp->grpnum = j % rsp->levelspread[i - 1];
3323 rnp->grpmask = 1UL << rnp->grpnum;
3324 rnp->parent = rsp->level[i - 1] +
3325 j / rsp->levelspread[i - 1];
3328 INIT_LIST_HEAD(&rnp->blkd_tasks);
3329 rcu_init_one_nocb(rnp);
3334 init_waitqueue_head(&rsp->gp_wq);
3335 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3336 rnp = rsp->level[rcu_num_lvls - 1];
3337 for_each_possible_cpu(i) {
3338 while (i > rnp->grphi)
3340 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3341 rcu_boot_init_percpu_data(i, rsp);
3343 list_add(&rsp->flavors, &rcu_struct_flavors);
3347 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3348 * replace the definitions in tree.h because those are needed to size
3349 * the ->node array in the rcu_state structure.
3351 static void __init rcu_init_geometry(void)
3357 int rcu_capacity[MAX_RCU_LVLS + 1];
3360 * Initialize any unspecified boot parameters.
3361 * The default values of jiffies_till_first_fqs and
3362 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3363 * value, which is a function of HZ, then adding one for each
3364 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3366 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3367 if (jiffies_till_first_fqs == ULONG_MAX)
3368 jiffies_till_first_fqs = d;
3369 if (jiffies_till_next_fqs == ULONG_MAX)
3370 jiffies_till_next_fqs = d;
3372 /* If the compile-time values are accurate, just leave. */
3373 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3374 nr_cpu_ids == NR_CPUS)
3378 * Compute number of nodes that can be handled an rcu_node tree
3379 * with the given number of levels. Setting rcu_capacity[0] makes
3380 * some of the arithmetic easier.
3382 rcu_capacity[0] = 1;
3383 rcu_capacity[1] = rcu_fanout_leaf;
3384 for (i = 2; i <= MAX_RCU_LVLS; i++)
3385 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3388 * The boot-time rcu_fanout_leaf parameter is only permitted
3389 * to increase the leaf-level fanout, not decrease it. Of course,
3390 * the leaf-level fanout cannot exceed the number of bits in
3391 * the rcu_node masks. Finally, the tree must be able to accommodate
3392 * the configured number of CPUs. Complain and fall back to the
3393 * compile-time values if these limits are exceeded.
3395 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3396 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3397 n > rcu_capacity[MAX_RCU_LVLS]) {
3402 /* Calculate the number of rcu_nodes at each level of the tree. */
3403 for (i = 1; i <= MAX_RCU_LVLS; i++)
3404 if (n <= rcu_capacity[i]) {
3405 for (j = 0; j <= i; j++)
3407 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3409 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3414 /* Calculate the total number of rcu_node structures. */
3416 for (i = 0; i <= MAX_RCU_LVLS; i++)
3417 rcu_num_nodes += num_rcu_lvl[i];
3421 void __init rcu_init(void)
3425 rcu_bootup_announce();
3426 rcu_init_geometry();
3427 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3428 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3429 __rcu_init_preempt();
3430 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3433 * We don't need protection against CPU-hotplug here because
3434 * this is called early in boot, before either interrupts
3435 * or the scheduler are operational.
3437 cpu_notifier(rcu_cpu_notify, 0);
3438 pm_notifier(rcu_pm_notify, 0);
3439 for_each_online_cpu(cpu)
3440 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3443 #include "tree_plugin.h"