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 __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 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 * Return true if the specified CPU has passed through a quiescent
759 * state by virtue of being in or having passed through an dynticks
760 * idle state since the last call to dyntick_save_progress_counter()
761 * for this same CPU, or by virtue of having been offline.
763 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
764 bool *isidle, unsigned long *maxj)
769 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
770 snap = (unsigned int)rdp->dynticks_snap;
773 * If the CPU passed through or entered a dynticks idle phase with
774 * no active irq/NMI handlers, then we can safely pretend that the CPU
775 * already acknowledged the request to pass through a quiescent
776 * state. Either way, that CPU cannot possibly be in an RCU
777 * read-side critical section that started before the beginning
778 * of the current RCU grace period.
780 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
781 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
787 * Check for the CPU being offline, but only if the grace period
788 * is old enough. We don't need to worry about the CPU changing
789 * state: If we see it offline even once, it has been through a
792 * The reason for insisting that the grace period be at least
793 * one jiffy old is that CPUs that are not quite online and that
794 * have just gone offline can still execute RCU read-side critical
797 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
798 return 0; /* Grace period is not old enough. */
800 if (cpu_is_offline(rdp->cpu)) {
801 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
807 * There is a possibility that a CPU in adaptive-ticks state
808 * might run in the kernel with the scheduling-clock tick disabled
809 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
810 * force the CPU to restart the scheduling-clock tick in this
811 * CPU is in this state.
813 rcu_kick_nohz_cpu(rdp->cpu);
818 static void record_gp_stall_check_time(struct rcu_state *rsp)
820 unsigned long j = ACCESS_ONCE(jiffies);
823 smp_wmb(); /* Record start time before stall time. */
824 rsp->jiffies_stall = j + rcu_jiffies_till_stall_check();
828 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
829 * for architectures that do not implement trigger_all_cpu_backtrace().
830 * The NMI-triggered stack traces are more accurate because they are
831 * printed by the target CPU.
833 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
837 struct rcu_node *rnp;
839 rcu_for_each_leaf_node(rsp, rnp) {
840 raw_spin_lock_irqsave(&rnp->lock, flags);
841 if (rnp->qsmask != 0) {
842 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
843 if (rnp->qsmask & (1UL << cpu))
844 dump_cpu_task(rnp->grplo + cpu);
846 raw_spin_unlock_irqrestore(&rnp->lock, flags);
850 static void print_other_cpu_stall(struct rcu_state *rsp)
856 struct rcu_node *rnp = rcu_get_root(rsp);
859 /* Only let one CPU complain about others per time interval. */
861 raw_spin_lock_irqsave(&rnp->lock, flags);
862 delta = jiffies - rsp->jiffies_stall;
863 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
864 raw_spin_unlock_irqrestore(&rnp->lock, flags);
867 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
868 raw_spin_unlock_irqrestore(&rnp->lock, flags);
871 * OK, time to rat on our buddy...
872 * See Documentation/RCU/stallwarn.txt for info on how to debug
873 * RCU CPU stall warnings.
875 pr_err("INFO: %s detected stalls on CPUs/tasks:",
877 print_cpu_stall_info_begin();
878 rcu_for_each_leaf_node(rsp, rnp) {
879 raw_spin_lock_irqsave(&rnp->lock, flags);
880 ndetected += rcu_print_task_stall(rnp);
881 if (rnp->qsmask != 0) {
882 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
883 if (rnp->qsmask & (1UL << cpu)) {
884 print_cpu_stall_info(rsp,
889 raw_spin_unlock_irqrestore(&rnp->lock, flags);
893 * Now rat on any tasks that got kicked up to the root rcu_node
894 * due to CPU offlining.
896 rnp = rcu_get_root(rsp);
897 raw_spin_lock_irqsave(&rnp->lock, flags);
898 ndetected += rcu_print_task_stall(rnp);
899 raw_spin_unlock_irqrestore(&rnp->lock, flags);
901 print_cpu_stall_info_end();
902 for_each_possible_cpu(cpu)
903 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
904 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
905 smp_processor_id(), (long)(jiffies - rsp->gp_start),
906 rsp->gpnum, rsp->completed, totqlen);
908 pr_err("INFO: Stall ended before state dump start\n");
909 else if (!trigger_all_cpu_backtrace())
910 rcu_dump_cpu_stacks(rsp);
912 /* Complain about tasks blocking the grace period. */
914 rcu_print_detail_task_stall(rsp);
916 force_quiescent_state(rsp); /* Kick them all. */
919 static void print_cpu_stall(struct rcu_state *rsp)
923 struct rcu_node *rnp = rcu_get_root(rsp);
927 * OK, time to rat on ourselves...
928 * See Documentation/RCU/stallwarn.txt for info on how to debug
929 * RCU CPU stall warnings.
931 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
932 print_cpu_stall_info_begin();
933 print_cpu_stall_info(rsp, smp_processor_id());
934 print_cpu_stall_info_end();
935 for_each_possible_cpu(cpu)
936 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
937 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
938 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
939 if (!trigger_all_cpu_backtrace())
942 raw_spin_lock_irqsave(&rnp->lock, flags);
943 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
944 rsp->jiffies_stall = jiffies +
945 3 * rcu_jiffies_till_stall_check() + 3;
946 raw_spin_unlock_irqrestore(&rnp->lock, flags);
948 set_need_resched(); /* kick ourselves to get things going. */
951 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
953 unsigned long completed;
958 struct rcu_node *rnp;
960 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
962 j = ACCESS_ONCE(jiffies);
965 * Lots of memory barriers to reject false positives.
967 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
968 * then rsp->gp_start, and finally rsp->completed. These values
969 * are updated in the opposite order with memory barriers (or
970 * equivalent) during grace-period initialization and cleanup.
971 * Now, a false positive can occur if we get an new value of
972 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
973 * the memory barriers, the only way that this can happen is if one
974 * grace period ends and another starts between these two fetches.
975 * Detect this by comparing rsp->completed with the previous fetch
978 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
979 * and rsp->gp_start suffice to forestall false positives.
981 gpnum = ACCESS_ONCE(rsp->gpnum);
982 smp_rmb(); /* Pick up ->gpnum first... */
983 js = ACCESS_ONCE(rsp->jiffies_stall);
984 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
985 gps = ACCESS_ONCE(rsp->gp_start);
986 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
987 completed = ACCESS_ONCE(rsp->completed);
988 if (ULONG_CMP_GE(completed, gpnum) ||
989 ULONG_CMP_LT(j, js) ||
990 ULONG_CMP_GE(gps, js))
991 return; /* No stall or GP completed since entering function. */
993 if (rcu_gp_in_progress(rsp) &&
994 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
996 /* We haven't checked in, so go dump stack. */
997 print_cpu_stall(rsp);
999 } else if (rcu_gp_in_progress(rsp) &&
1000 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1002 /* They had a few time units to dump stack, so complain. */
1003 print_other_cpu_stall(rsp);
1008 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1010 * Set the stall-warning timeout way off into the future, thus preventing
1011 * any RCU CPU stall-warning messages from appearing in the current set of
1012 * RCU grace periods.
1014 * The caller must disable hard irqs.
1016 void rcu_cpu_stall_reset(void)
1018 struct rcu_state *rsp;
1020 for_each_rcu_flavor(rsp)
1021 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1025 * Initialize the specified rcu_data structure's callback list to empty.
1027 static void init_callback_list(struct rcu_data *rdp)
1031 if (init_nocb_callback_list(rdp))
1033 rdp->nxtlist = NULL;
1034 for (i = 0; i < RCU_NEXT_SIZE; i++)
1035 rdp->nxttail[i] = &rdp->nxtlist;
1039 * Determine the value that ->completed will have at the end of the
1040 * next subsequent grace period. This is used to tag callbacks so that
1041 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1042 * been dyntick-idle for an extended period with callbacks under the
1043 * influence of RCU_FAST_NO_HZ.
1045 * The caller must hold rnp->lock with interrupts disabled.
1047 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1048 struct rcu_node *rnp)
1051 * If RCU is idle, we just wait for the next grace period.
1052 * But we can only be sure that RCU is idle if we are looking
1053 * at the root rcu_node structure -- otherwise, a new grace
1054 * period might have started, but just not yet gotten around
1055 * to initializing the current non-root rcu_node structure.
1057 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1058 return rnp->completed + 1;
1061 * Otherwise, wait for a possible partial grace period and
1062 * then the subsequent full grace period.
1064 return rnp->completed + 2;
1068 * Trace-event helper function for rcu_start_future_gp() and
1069 * rcu_nocb_wait_gp().
1071 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1072 unsigned long c, const char *s)
1074 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1075 rnp->completed, c, rnp->level,
1076 rnp->grplo, rnp->grphi, s);
1080 * Start some future grace period, as needed to handle newly arrived
1081 * callbacks. The required future grace periods are recorded in each
1082 * rcu_node structure's ->need_future_gp field.
1084 * The caller must hold the specified rcu_node structure's ->lock.
1086 static unsigned long __maybe_unused
1087 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1091 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1094 * Pick up grace-period number for new callbacks. If this
1095 * grace period is already marked as needed, return to the caller.
1097 c = rcu_cbs_completed(rdp->rsp, rnp);
1098 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1099 if (rnp->need_future_gp[c & 0x1]) {
1100 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1105 * If either this rcu_node structure or the root rcu_node structure
1106 * believe that a grace period is in progress, then we must wait
1107 * for the one following, which is in "c". Because our request
1108 * will be noticed at the end of the current grace period, we don't
1109 * need to explicitly start one.
1111 if (rnp->gpnum != rnp->completed ||
1112 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1113 rnp->need_future_gp[c & 0x1]++;
1114 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1119 * There might be no grace period in progress. If we don't already
1120 * hold it, acquire the root rcu_node structure's lock in order to
1121 * start one (if needed).
1123 if (rnp != rnp_root)
1124 raw_spin_lock(&rnp_root->lock);
1127 * Get a new grace-period number. If there really is no grace
1128 * period in progress, it will be smaller than the one we obtained
1129 * earlier. Adjust callbacks as needed. Note that even no-CBs
1130 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1132 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1133 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1134 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1135 rdp->nxtcompleted[i] = c;
1138 * If the needed for the required grace period is already
1139 * recorded, trace and leave.
1141 if (rnp_root->need_future_gp[c & 0x1]) {
1142 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1146 /* Record the need for the future grace period. */
1147 rnp_root->need_future_gp[c & 0x1]++;
1149 /* If a grace period is not already in progress, start one. */
1150 if (rnp_root->gpnum != rnp_root->completed) {
1151 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1153 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1154 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1157 if (rnp != rnp_root)
1158 raw_spin_unlock(&rnp_root->lock);
1163 * Clean up any old requests for the just-ended grace period. Also return
1164 * whether any additional grace periods have been requested. Also invoke
1165 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1166 * waiting for this grace period to complete.
1168 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1170 int c = rnp->completed;
1172 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1174 rcu_nocb_gp_cleanup(rsp, rnp);
1175 rnp->need_future_gp[c & 0x1] = 0;
1176 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1177 trace_rcu_future_gp(rnp, rdp, c,
1178 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1183 * If there is room, assign a ->completed number to any callbacks on
1184 * this CPU that have not already been assigned. Also accelerate any
1185 * callbacks that were previously assigned a ->completed number that has
1186 * since proven to be too conservative, which can happen if callbacks get
1187 * assigned a ->completed number while RCU is idle, but with reference to
1188 * a non-root rcu_node structure. This function is idempotent, so it does
1189 * not hurt to call it repeatedly.
1191 * The caller must hold rnp->lock with interrupts disabled.
1193 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1194 struct rcu_data *rdp)
1199 /* If the CPU has no callbacks, nothing to do. */
1200 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1204 * Starting from the sublist containing the callbacks most
1205 * recently assigned a ->completed number and working down, find the
1206 * first sublist that is not assignable to an upcoming grace period.
1207 * Such a sublist has something in it (first two tests) and has
1208 * a ->completed number assigned that will complete sooner than
1209 * the ->completed number for newly arrived callbacks (last test).
1211 * The key point is that any later sublist can be assigned the
1212 * same ->completed number as the newly arrived callbacks, which
1213 * means that the callbacks in any of these later sublist can be
1214 * grouped into a single sublist, whether or not they have already
1215 * been assigned a ->completed number.
1217 c = rcu_cbs_completed(rsp, rnp);
1218 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1219 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1220 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1224 * If there are no sublist for unassigned callbacks, leave.
1225 * At the same time, advance "i" one sublist, so that "i" will
1226 * index into the sublist where all the remaining callbacks should
1229 if (++i >= RCU_NEXT_TAIL)
1233 * Assign all subsequent callbacks' ->completed number to the next
1234 * full grace period and group them all in the sublist initially
1237 for (; i <= RCU_NEXT_TAIL; i++) {
1238 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1239 rdp->nxtcompleted[i] = c;
1241 /* Record any needed additional grace periods. */
1242 rcu_start_future_gp(rnp, rdp);
1244 /* Trace depending on how much we were able to accelerate. */
1245 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1246 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1248 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1252 * Move any callbacks whose grace period has completed to the
1253 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1254 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1255 * sublist. This function is idempotent, so it does not hurt to
1256 * invoke it repeatedly. As long as it is not invoked -too- often...
1258 * The caller must hold rnp->lock with interrupts disabled.
1260 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1261 struct rcu_data *rdp)
1265 /* If the CPU has no callbacks, nothing to do. */
1266 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1270 * Find all callbacks whose ->completed numbers indicate that they
1271 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1273 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1274 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1276 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1278 /* Clean up any sublist tail pointers that were misordered above. */
1279 for (j = RCU_WAIT_TAIL; j < i; j++)
1280 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1282 /* Copy down callbacks to fill in empty sublists. */
1283 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1284 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1286 rdp->nxttail[j] = rdp->nxttail[i];
1287 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1290 /* Classify any remaining callbacks. */
1291 rcu_accelerate_cbs(rsp, rnp, rdp);
1295 * Update CPU-local rcu_data state to record the beginnings and ends of
1296 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1297 * structure corresponding to the current CPU, and must have irqs disabled.
1299 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1301 /* Handle the ends of any preceding grace periods first. */
1302 if (rdp->completed == rnp->completed) {
1304 /* No grace period end, so just accelerate recent callbacks. */
1305 rcu_accelerate_cbs(rsp, rnp, rdp);
1309 /* Advance callbacks. */
1310 rcu_advance_cbs(rsp, rnp, rdp);
1312 /* Remember that we saw this grace-period completion. */
1313 rdp->completed = rnp->completed;
1314 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1317 if (rdp->gpnum != rnp->gpnum) {
1319 * If the current grace period is waiting for this CPU,
1320 * set up to detect a quiescent state, otherwise don't
1321 * go looking for one.
1323 rdp->gpnum = rnp->gpnum;
1324 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1325 rdp->passed_quiesce = 0;
1326 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1327 zero_cpu_stall_ticks(rdp);
1331 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1333 unsigned long flags;
1334 struct rcu_node *rnp;
1336 local_irq_save(flags);
1338 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1339 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1340 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1341 local_irq_restore(flags);
1344 __note_gp_changes(rsp, rnp, rdp);
1345 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1349 * Initialize a new grace period. Return 0 if no grace period required.
1351 static int rcu_gp_init(struct rcu_state *rsp)
1353 struct rcu_data *rdp;
1354 struct rcu_node *rnp = rcu_get_root(rsp);
1356 rcu_bind_gp_kthread();
1357 raw_spin_lock_irq(&rnp->lock);
1358 if (rsp->gp_flags == 0) {
1359 /* Spurious wakeup, tell caller to go back to sleep. */
1360 raw_spin_unlock_irq(&rnp->lock);
1363 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1365 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1367 * Grace period already in progress, don't start another.
1368 * Not supposed to be able to happen.
1370 raw_spin_unlock_irq(&rnp->lock);
1374 /* Advance to a new grace period and initialize state. */
1375 record_gp_stall_check_time(rsp);
1376 smp_wmb(); /* Record GP times before starting GP. */
1378 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1379 raw_spin_unlock_irq(&rnp->lock);
1381 /* Exclude any concurrent CPU-hotplug operations. */
1382 mutex_lock(&rsp->onoff_mutex);
1385 * Set the quiescent-state-needed bits in all the rcu_node
1386 * structures for all currently online CPUs in breadth-first order,
1387 * starting from the root rcu_node structure, relying on the layout
1388 * of the tree within the rsp->node[] array. Note that other CPUs
1389 * will access only the leaves of the hierarchy, thus seeing that no
1390 * grace period is in progress, at least until the corresponding
1391 * leaf node has been initialized. In addition, we have excluded
1392 * CPU-hotplug operations.
1394 * The grace period cannot complete until the initialization
1395 * process finishes, because this kthread handles both.
1397 rcu_for_each_node_breadth_first(rsp, rnp) {
1398 raw_spin_lock_irq(&rnp->lock);
1399 rdp = this_cpu_ptr(rsp->rda);
1400 rcu_preempt_check_blocked_tasks(rnp);
1401 rnp->qsmask = rnp->qsmaskinit;
1402 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1403 WARN_ON_ONCE(rnp->completed != rsp->completed);
1404 ACCESS_ONCE(rnp->completed) = rsp->completed;
1405 if (rnp == rdp->mynode)
1406 __note_gp_changes(rsp, rnp, rdp);
1407 rcu_preempt_boost_start_gp(rnp);
1408 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1409 rnp->level, rnp->grplo,
1410 rnp->grphi, rnp->qsmask);
1411 raw_spin_unlock_irq(&rnp->lock);
1412 #ifdef CONFIG_PROVE_RCU_DELAY
1413 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1414 system_state == SYSTEM_RUNNING)
1416 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1420 mutex_unlock(&rsp->onoff_mutex);
1425 * Do one round of quiescent-state forcing.
1427 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1429 int fqs_state = fqs_state_in;
1430 bool isidle = false;
1432 struct rcu_node *rnp = rcu_get_root(rsp);
1435 if (fqs_state == RCU_SAVE_DYNTICK) {
1436 /* Collect dyntick-idle snapshots. */
1437 if (is_sysidle_rcu_state(rsp)) {
1439 maxj = jiffies - ULONG_MAX / 4;
1441 force_qs_rnp(rsp, dyntick_save_progress_counter,
1443 rcu_sysidle_report_gp(rsp, isidle, maxj);
1444 fqs_state = RCU_FORCE_QS;
1446 /* Handle dyntick-idle and offline CPUs. */
1448 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1450 /* Clear flag to prevent immediate re-entry. */
1451 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1452 raw_spin_lock_irq(&rnp->lock);
1453 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1454 raw_spin_unlock_irq(&rnp->lock);
1460 * Clean up after the old grace period.
1462 static void rcu_gp_cleanup(struct rcu_state *rsp)
1464 unsigned long gp_duration;
1466 struct rcu_data *rdp;
1467 struct rcu_node *rnp = rcu_get_root(rsp);
1469 raw_spin_lock_irq(&rnp->lock);
1470 gp_duration = jiffies - rsp->gp_start;
1471 if (gp_duration > rsp->gp_max)
1472 rsp->gp_max = gp_duration;
1475 * We know the grace period is complete, but to everyone else
1476 * it appears to still be ongoing. But it is also the case
1477 * that to everyone else it looks like there is nothing that
1478 * they can do to advance the grace period. It is therefore
1479 * safe for us to drop the lock in order to mark the grace
1480 * period as completed in all of the rcu_node structures.
1482 raw_spin_unlock_irq(&rnp->lock);
1485 * Propagate new ->completed value to rcu_node structures so
1486 * that other CPUs don't have to wait until the start of the next
1487 * grace period to process their callbacks. This also avoids
1488 * some nasty RCU grace-period initialization races by forcing
1489 * the end of the current grace period to be completely recorded in
1490 * all of the rcu_node structures before the beginning of the next
1491 * grace period is recorded in any of the rcu_node structures.
1493 rcu_for_each_node_breadth_first(rsp, rnp) {
1494 raw_spin_lock_irq(&rnp->lock);
1495 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1496 rdp = this_cpu_ptr(rsp->rda);
1497 if (rnp == rdp->mynode)
1498 __note_gp_changes(rsp, rnp, rdp);
1499 nocb += rcu_future_gp_cleanup(rsp, rnp);
1500 raw_spin_unlock_irq(&rnp->lock);
1503 rnp = rcu_get_root(rsp);
1504 raw_spin_lock_irq(&rnp->lock);
1505 rcu_nocb_gp_set(rnp, nocb);
1507 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1508 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1509 rsp->fqs_state = RCU_GP_IDLE;
1510 rdp = this_cpu_ptr(rsp->rda);
1511 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1512 if (cpu_needs_another_gp(rsp, rdp)) {
1513 rsp->gp_flags = RCU_GP_FLAG_INIT;
1514 trace_rcu_grace_period(rsp->name,
1515 ACCESS_ONCE(rsp->gpnum),
1518 raw_spin_unlock_irq(&rnp->lock);
1522 * Body of kthread that handles grace periods.
1524 static int __noreturn rcu_gp_kthread(void *arg)
1530 struct rcu_state *rsp = arg;
1531 struct rcu_node *rnp = rcu_get_root(rsp);
1535 /* Handle grace-period start. */
1537 trace_rcu_grace_period(rsp->name,
1538 ACCESS_ONCE(rsp->gpnum),
1540 wait_event_interruptible(rsp->gp_wq,
1541 ACCESS_ONCE(rsp->gp_flags) &
1543 if (rcu_gp_init(rsp))
1546 flush_signals(current);
1547 trace_rcu_grace_period(rsp->name,
1548 ACCESS_ONCE(rsp->gpnum),
1552 /* Handle quiescent-state forcing. */
1553 fqs_state = RCU_SAVE_DYNTICK;
1554 j = jiffies_till_first_fqs;
1557 jiffies_till_first_fqs = HZ;
1562 rsp->jiffies_force_qs = jiffies + j;
1563 trace_rcu_grace_period(rsp->name,
1564 ACCESS_ONCE(rsp->gpnum),
1566 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1567 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1569 (!ACCESS_ONCE(rnp->qsmask) &&
1570 !rcu_preempt_blocked_readers_cgp(rnp)),
1572 /* If grace period done, leave loop. */
1573 if (!ACCESS_ONCE(rnp->qsmask) &&
1574 !rcu_preempt_blocked_readers_cgp(rnp))
1576 /* If time for quiescent-state forcing, do it. */
1577 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1578 (gf & RCU_GP_FLAG_FQS)) {
1579 trace_rcu_grace_period(rsp->name,
1580 ACCESS_ONCE(rsp->gpnum),
1582 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1583 trace_rcu_grace_period(rsp->name,
1584 ACCESS_ONCE(rsp->gpnum),
1588 /* Deal with stray signal. */
1590 flush_signals(current);
1591 trace_rcu_grace_period(rsp->name,
1592 ACCESS_ONCE(rsp->gpnum),
1595 j = jiffies_till_next_fqs;
1598 jiffies_till_next_fqs = HZ;
1601 jiffies_till_next_fqs = 1;
1605 /* Handle grace-period end. */
1606 rcu_gp_cleanup(rsp);
1610 static void rsp_wakeup(struct irq_work *work)
1612 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1614 /* Wake up rcu_gp_kthread() to start the grace period. */
1615 wake_up(&rsp->gp_wq);
1619 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1620 * in preparation for detecting the next grace period. The caller must hold
1621 * the root node's ->lock and hard irqs must be disabled.
1623 * Note that it is legal for a dying CPU (which is marked as offline) to
1624 * invoke this function. This can happen when the dying CPU reports its
1628 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1629 struct rcu_data *rdp)
1631 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1633 * Either we have not yet spawned the grace-period
1634 * task, this CPU does not need another grace period,
1635 * or a grace period is already in progress.
1636 * Either way, don't start a new grace period.
1640 rsp->gp_flags = RCU_GP_FLAG_INIT;
1641 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1645 * We can't do wakeups while holding the rnp->lock, as that
1646 * could cause possible deadlocks with the rq->lock. Defer
1647 * the wakeup to interrupt context. And don't bother waking
1648 * up the running kthread.
1650 if (current != rsp->gp_kthread)
1651 irq_work_queue(&rsp->wakeup_work);
1655 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1656 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1657 * is invoked indirectly from rcu_advance_cbs(), which would result in
1658 * endless recursion -- or would do so if it wasn't for the self-deadlock
1659 * that is encountered beforehand.
1662 rcu_start_gp(struct rcu_state *rsp)
1664 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1665 struct rcu_node *rnp = rcu_get_root(rsp);
1668 * If there is no grace period in progress right now, any
1669 * callbacks we have up to this point will be satisfied by the
1670 * next grace period. Also, advancing the callbacks reduces the
1671 * probability of false positives from cpu_needs_another_gp()
1672 * resulting in pointless grace periods. So, advance callbacks
1673 * then start the grace period!
1675 rcu_advance_cbs(rsp, rnp, rdp);
1676 rcu_start_gp_advanced(rsp, rnp, rdp);
1680 * Report a full set of quiescent states to the specified rcu_state
1681 * data structure. This involves cleaning up after the prior grace
1682 * period and letting rcu_start_gp() start up the next grace period
1683 * if one is needed. Note that the caller must hold rnp->lock, which
1684 * is released before return.
1686 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1687 __releases(rcu_get_root(rsp)->lock)
1689 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1690 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1691 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1695 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1696 * Allows quiescent states for a group of CPUs to be reported at one go
1697 * to the specified rcu_node structure, though all the CPUs in the group
1698 * must be represented by the same rcu_node structure (which need not be
1699 * a leaf rcu_node structure, though it often will be). That structure's
1700 * lock must be held upon entry, and it is released before return.
1703 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1704 struct rcu_node *rnp, unsigned long flags)
1705 __releases(rnp->lock)
1707 struct rcu_node *rnp_c;
1709 /* Walk up the rcu_node hierarchy. */
1711 if (!(rnp->qsmask & mask)) {
1713 /* Our bit has already been cleared, so done. */
1714 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1717 rnp->qsmask &= ~mask;
1718 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1719 mask, rnp->qsmask, rnp->level,
1720 rnp->grplo, rnp->grphi,
1722 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1724 /* Other bits still set at this level, so done. */
1725 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1728 mask = rnp->grpmask;
1729 if (rnp->parent == NULL) {
1731 /* No more levels. Exit loop holding root lock. */
1735 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1738 raw_spin_lock_irqsave(&rnp->lock, flags);
1739 WARN_ON_ONCE(rnp_c->qsmask);
1743 * Get here if we are the last CPU to pass through a quiescent
1744 * state for this grace period. Invoke rcu_report_qs_rsp()
1745 * to clean up and start the next grace period if one is needed.
1747 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1751 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1752 * structure. This must be either called from the specified CPU, or
1753 * called when the specified CPU is known to be offline (and when it is
1754 * also known that no other CPU is concurrently trying to help the offline
1755 * CPU). The lastcomp argument is used to make sure we are still in the
1756 * grace period of interest. We don't want to end the current grace period
1757 * based on quiescent states detected in an earlier grace period!
1760 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1762 unsigned long flags;
1764 struct rcu_node *rnp;
1767 raw_spin_lock_irqsave(&rnp->lock, flags);
1768 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1769 rnp->completed == rnp->gpnum) {
1772 * The grace period in which this quiescent state was
1773 * recorded has ended, so don't report it upwards.
1774 * We will instead need a new quiescent state that lies
1775 * within the current grace period.
1777 rdp->passed_quiesce = 0; /* need qs for new gp. */
1778 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1781 mask = rdp->grpmask;
1782 if ((rnp->qsmask & mask) == 0) {
1783 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1785 rdp->qs_pending = 0;
1788 * This GP can't end until cpu checks in, so all of our
1789 * callbacks can be processed during the next GP.
1791 rcu_accelerate_cbs(rsp, rnp, rdp);
1793 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1798 * Check to see if there is a new grace period of which this CPU
1799 * is not yet aware, and if so, set up local rcu_data state for it.
1800 * Otherwise, see if this CPU has just passed through its first
1801 * quiescent state for this grace period, and record that fact if so.
1804 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1806 /* Check for grace-period ends and beginnings. */
1807 note_gp_changes(rsp, rdp);
1810 * Does this CPU still need to do its part for current grace period?
1811 * If no, return and let the other CPUs do their part as well.
1813 if (!rdp->qs_pending)
1817 * Was there a quiescent state since the beginning of the grace
1818 * period? If no, then exit and wait for the next call.
1820 if (!rdp->passed_quiesce)
1824 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1827 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1830 #ifdef CONFIG_HOTPLUG_CPU
1833 * Send the specified CPU's RCU callbacks to the orphanage. The
1834 * specified CPU must be offline, and the caller must hold the
1838 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1839 struct rcu_node *rnp, struct rcu_data *rdp)
1841 /* No-CBs CPUs do not have orphanable callbacks. */
1842 if (rcu_is_nocb_cpu(rdp->cpu))
1846 * Orphan the callbacks. First adjust the counts. This is safe
1847 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1848 * cannot be running now. Thus no memory barrier is required.
1850 if (rdp->nxtlist != NULL) {
1851 rsp->qlen_lazy += rdp->qlen_lazy;
1852 rsp->qlen += rdp->qlen;
1853 rdp->n_cbs_orphaned += rdp->qlen;
1855 ACCESS_ONCE(rdp->qlen) = 0;
1859 * Next, move those callbacks still needing a grace period to
1860 * the orphanage, where some other CPU will pick them up.
1861 * Some of the callbacks might have gone partway through a grace
1862 * period, but that is too bad. They get to start over because we
1863 * cannot assume that grace periods are synchronized across CPUs.
1864 * We don't bother updating the ->nxttail[] array yet, instead
1865 * we just reset the whole thing later on.
1867 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1868 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1869 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1870 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1874 * Then move the ready-to-invoke callbacks to the orphanage,
1875 * where some other CPU will pick them up. These will not be
1876 * required to pass though another grace period: They are done.
1878 if (rdp->nxtlist != NULL) {
1879 *rsp->orphan_donetail = rdp->nxtlist;
1880 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1883 /* Finally, initialize the rcu_data structure's list to empty. */
1884 init_callback_list(rdp);
1888 * Adopt the RCU callbacks from the specified rcu_state structure's
1889 * orphanage. The caller must hold the ->orphan_lock.
1891 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1894 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1896 /* No-CBs CPUs are handled specially. */
1897 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1900 /* Do the accounting first. */
1901 rdp->qlen_lazy += rsp->qlen_lazy;
1902 rdp->qlen += rsp->qlen;
1903 rdp->n_cbs_adopted += rsp->qlen;
1904 if (rsp->qlen_lazy != rsp->qlen)
1905 rcu_idle_count_callbacks_posted();
1910 * We do not need a memory barrier here because the only way we
1911 * can get here if there is an rcu_barrier() in flight is if
1912 * we are the task doing the rcu_barrier().
1915 /* First adopt the ready-to-invoke callbacks. */
1916 if (rsp->orphan_donelist != NULL) {
1917 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1918 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1919 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1920 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1921 rdp->nxttail[i] = rsp->orphan_donetail;
1922 rsp->orphan_donelist = NULL;
1923 rsp->orphan_donetail = &rsp->orphan_donelist;
1926 /* And then adopt the callbacks that still need a grace period. */
1927 if (rsp->orphan_nxtlist != NULL) {
1928 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1929 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1930 rsp->orphan_nxtlist = NULL;
1931 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1936 * Trace the fact that this CPU is going offline.
1938 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1940 RCU_TRACE(unsigned long mask);
1941 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1942 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1944 RCU_TRACE(mask = rdp->grpmask);
1945 trace_rcu_grace_period(rsp->name,
1946 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1951 * The CPU has been completely removed, and some other CPU is reporting
1952 * this fact from process context. Do the remainder of the cleanup,
1953 * including orphaning the outgoing CPU's RCU callbacks, and also
1954 * adopting them. There can only be one CPU hotplug operation at a time,
1955 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1957 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1959 unsigned long flags;
1961 int need_report = 0;
1962 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1963 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1965 /* Adjust any no-longer-needed kthreads. */
1966 rcu_boost_kthread_setaffinity(rnp, -1);
1968 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1970 /* Exclude any attempts to start a new grace period. */
1971 mutex_lock(&rsp->onoff_mutex);
1972 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1974 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1975 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1976 rcu_adopt_orphan_cbs(rsp);
1978 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1979 mask = rdp->grpmask; /* rnp->grplo is constant. */
1981 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1982 rnp->qsmaskinit &= ~mask;
1983 if (rnp->qsmaskinit != 0) {
1984 if (rnp != rdp->mynode)
1985 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1988 if (rnp == rdp->mynode)
1989 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1991 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1992 mask = rnp->grpmask;
1994 } while (rnp != NULL);
1997 * We still hold the leaf rcu_node structure lock here, and
1998 * irqs are still disabled. The reason for this subterfuge is
1999 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2000 * held leads to deadlock.
2002 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2004 if (need_report & RCU_OFL_TASKS_NORM_GP)
2005 rcu_report_unblock_qs_rnp(rnp, flags);
2007 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2008 if (need_report & RCU_OFL_TASKS_EXP_GP)
2009 rcu_report_exp_rnp(rsp, rnp, true);
2010 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2011 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2012 cpu, rdp->qlen, rdp->nxtlist);
2013 init_callback_list(rdp);
2014 /* Disallow further callbacks on this CPU. */
2015 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2016 mutex_unlock(&rsp->onoff_mutex);
2019 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2021 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2025 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2029 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2032 * Invoke any RCU callbacks that have made it to the end of their grace
2033 * period. Thottle as specified by rdp->blimit.
2035 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2037 unsigned long flags;
2038 struct rcu_head *next, *list, **tail;
2039 long bl, count, count_lazy;
2042 /* If no callbacks are ready, just return. */
2043 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2044 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2045 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2046 need_resched(), is_idle_task(current),
2047 rcu_is_callbacks_kthread());
2052 * Extract the list of ready callbacks, disabling to prevent
2053 * races with call_rcu() from interrupt handlers.
2055 local_irq_save(flags);
2056 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2058 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2059 list = rdp->nxtlist;
2060 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2061 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2062 tail = rdp->nxttail[RCU_DONE_TAIL];
2063 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2064 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2065 rdp->nxttail[i] = &rdp->nxtlist;
2066 local_irq_restore(flags);
2068 /* Invoke callbacks. */
2069 count = count_lazy = 0;
2073 debug_rcu_head_unqueue(list);
2074 if (__rcu_reclaim(rsp->name, list))
2077 /* Stop only if limit reached and CPU has something to do. */
2078 if (++count >= bl &&
2080 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2084 local_irq_save(flags);
2085 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2086 is_idle_task(current),
2087 rcu_is_callbacks_kthread());
2089 /* Update count, and requeue any remaining callbacks. */
2091 *tail = rdp->nxtlist;
2092 rdp->nxtlist = list;
2093 for (i = 0; i < RCU_NEXT_SIZE; i++)
2094 if (&rdp->nxtlist == rdp->nxttail[i])
2095 rdp->nxttail[i] = tail;
2099 smp_mb(); /* List handling before counting for rcu_barrier(). */
2100 rdp->qlen_lazy -= count_lazy;
2101 ACCESS_ONCE(rdp->qlen) -= count;
2102 rdp->n_cbs_invoked += count;
2104 /* Reinstate batch limit if we have worked down the excess. */
2105 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2106 rdp->blimit = blimit;
2108 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2109 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2110 rdp->qlen_last_fqs_check = 0;
2111 rdp->n_force_qs_snap = rsp->n_force_qs;
2112 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2113 rdp->qlen_last_fqs_check = rdp->qlen;
2114 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2116 local_irq_restore(flags);
2118 /* Re-invoke RCU core processing if there are callbacks remaining. */
2119 if (cpu_has_callbacks_ready_to_invoke(rdp))
2124 * Check to see if this CPU is in a non-context-switch quiescent state
2125 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2126 * Also schedule RCU core processing.
2128 * This function must be called from hardirq context. It is normally
2129 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2130 * false, there is no point in invoking rcu_check_callbacks().
2132 void rcu_check_callbacks(int cpu, int user)
2134 trace_rcu_utilization(TPS("Start scheduler-tick"));
2135 increment_cpu_stall_ticks();
2136 if (user || rcu_is_cpu_rrupt_from_idle()) {
2139 * Get here if this CPU took its interrupt from user
2140 * mode or from the idle loop, and if this is not a
2141 * nested interrupt. In this case, the CPU is in
2142 * a quiescent state, so note it.
2144 * No memory barrier is required here because both
2145 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2146 * variables that other CPUs neither access nor modify,
2147 * at least not while the corresponding CPU is online.
2153 } else if (!in_softirq()) {
2156 * Get here if this CPU did not take its interrupt from
2157 * softirq, in other words, if it is not interrupting
2158 * a rcu_bh read-side critical section. This is an _bh
2159 * critical section, so note it.
2164 rcu_preempt_check_callbacks(cpu);
2165 if (rcu_pending(cpu))
2167 trace_rcu_utilization(TPS("End scheduler-tick"));
2171 * Scan the leaf rcu_node structures, processing dyntick state for any that
2172 * have not yet encountered a quiescent state, using the function specified.
2173 * Also initiate boosting for any threads blocked on the root rcu_node.
2175 * The caller must have suppressed start of new grace periods.
2177 static void force_qs_rnp(struct rcu_state *rsp,
2178 int (*f)(struct rcu_data *rsp, bool *isidle,
2179 unsigned long *maxj),
2180 bool *isidle, unsigned long *maxj)
2184 unsigned long flags;
2186 struct rcu_node *rnp;
2188 rcu_for_each_leaf_node(rsp, rnp) {
2191 raw_spin_lock_irqsave(&rnp->lock, flags);
2192 if (!rcu_gp_in_progress(rsp)) {
2193 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2196 if (rnp->qsmask == 0) {
2197 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2202 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2203 if ((rnp->qsmask & bit) != 0) {
2204 if ((rnp->qsmaskinit & bit) != 0)
2206 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2212 /* rcu_report_qs_rnp() releases rnp->lock. */
2213 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2216 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2218 rnp = rcu_get_root(rsp);
2219 if (rnp->qsmask == 0) {
2220 raw_spin_lock_irqsave(&rnp->lock, flags);
2221 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2226 * Force quiescent states on reluctant CPUs, and also detect which
2227 * CPUs are in dyntick-idle mode.
2229 static void force_quiescent_state(struct rcu_state *rsp)
2231 unsigned long flags;
2233 struct rcu_node *rnp;
2234 struct rcu_node *rnp_old = NULL;
2236 /* Funnel through hierarchy to reduce memory contention. */
2237 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2238 for (; rnp != NULL; rnp = rnp->parent) {
2239 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2240 !raw_spin_trylock(&rnp->fqslock);
2241 if (rnp_old != NULL)
2242 raw_spin_unlock(&rnp_old->fqslock);
2244 rsp->n_force_qs_lh++;
2249 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2251 /* Reached the root of the rcu_node tree, acquire lock. */
2252 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2253 raw_spin_unlock(&rnp_old->fqslock);
2254 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2255 rsp->n_force_qs_lh++;
2256 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2257 return; /* Someone beat us to it. */
2259 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2260 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2261 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2265 * This does the RCU core processing work for the specified rcu_state
2266 * and rcu_data structures. This may be called only from the CPU to
2267 * whom the rdp belongs.
2270 __rcu_process_callbacks(struct rcu_state *rsp)
2272 unsigned long flags;
2273 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2275 WARN_ON_ONCE(rdp->beenonline == 0);
2277 /* Update RCU state based on any recent quiescent states. */
2278 rcu_check_quiescent_state(rsp, rdp);
2280 /* Does this CPU require a not-yet-started grace period? */
2281 local_irq_save(flags);
2282 if (cpu_needs_another_gp(rsp, rdp)) {
2283 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2285 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2287 local_irq_restore(flags);
2290 /* If there are callbacks ready, invoke them. */
2291 if (cpu_has_callbacks_ready_to_invoke(rdp))
2292 invoke_rcu_callbacks(rsp, rdp);
2296 * Do RCU core processing for the current CPU.
2298 static void rcu_process_callbacks(struct softirq_action *unused)
2300 struct rcu_state *rsp;
2302 if (cpu_is_offline(smp_processor_id()))
2304 trace_rcu_utilization(TPS("Start RCU core"));
2305 for_each_rcu_flavor(rsp)
2306 __rcu_process_callbacks(rsp);
2307 trace_rcu_utilization(TPS("End RCU core"));
2311 * Schedule RCU callback invocation. If the specified type of RCU
2312 * does not support RCU priority boosting, just do a direct call,
2313 * otherwise wake up the per-CPU kernel kthread. Note that because we
2314 * are running on the current CPU with interrupts disabled, the
2315 * rcu_cpu_kthread_task cannot disappear out from under us.
2317 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2319 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2321 if (likely(!rsp->boost)) {
2322 rcu_do_batch(rsp, rdp);
2325 invoke_rcu_callbacks_kthread();
2328 static void invoke_rcu_core(void)
2330 if (cpu_online(smp_processor_id()))
2331 raise_softirq(RCU_SOFTIRQ);
2335 * Handle any core-RCU processing required by a call_rcu() invocation.
2337 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2338 struct rcu_head *head, unsigned long flags)
2341 * If called from an extended quiescent state, invoke the RCU
2342 * core in order to force a re-evaluation of RCU's idleness.
2344 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2347 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2348 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2352 * Force the grace period if too many callbacks or too long waiting.
2353 * Enforce hysteresis, and don't invoke force_quiescent_state()
2354 * if some other CPU has recently done so. Also, don't bother
2355 * invoking force_quiescent_state() if the newly enqueued callback
2356 * is the only one waiting for a grace period to complete.
2358 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2360 /* Are we ignoring a completed grace period? */
2361 note_gp_changes(rsp, rdp);
2363 /* Start a new grace period if one not already started. */
2364 if (!rcu_gp_in_progress(rsp)) {
2365 struct rcu_node *rnp_root = rcu_get_root(rsp);
2367 raw_spin_lock(&rnp_root->lock);
2369 raw_spin_unlock(&rnp_root->lock);
2371 /* Give the grace period a kick. */
2372 rdp->blimit = LONG_MAX;
2373 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2374 *rdp->nxttail[RCU_DONE_TAIL] != head)
2375 force_quiescent_state(rsp);
2376 rdp->n_force_qs_snap = rsp->n_force_qs;
2377 rdp->qlen_last_fqs_check = rdp->qlen;
2383 * RCU callback function to leak a callback.
2385 static void rcu_leak_callback(struct rcu_head *rhp)
2390 * Helper function for call_rcu() and friends. The cpu argument will
2391 * normally be -1, indicating "currently running CPU". It may specify
2392 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2393 * is expected to specify a CPU.
2396 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2397 struct rcu_state *rsp, int cpu, bool lazy)
2399 unsigned long flags;
2400 struct rcu_data *rdp;
2402 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2403 if (debug_rcu_head_queue(head)) {
2404 /* Probable double call_rcu(), so leak the callback. */
2405 ACCESS_ONCE(head->func) = rcu_leak_callback;
2406 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2413 * Opportunistically note grace-period endings and beginnings.
2414 * Note that we might see a beginning right after we see an
2415 * end, but never vice versa, since this CPU has to pass through
2416 * a quiescent state betweentimes.
2418 local_irq_save(flags);
2419 rdp = this_cpu_ptr(rsp->rda);
2421 /* Add the callback to our list. */
2422 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2426 rdp = per_cpu_ptr(rsp->rda, cpu);
2427 offline = !__call_rcu_nocb(rdp, head, lazy);
2428 WARN_ON_ONCE(offline);
2429 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2430 local_irq_restore(flags);
2433 ACCESS_ONCE(rdp->qlen)++;
2437 rcu_idle_count_callbacks_posted();
2438 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2439 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2440 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2442 if (__is_kfree_rcu_offset((unsigned long)func))
2443 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2444 rdp->qlen_lazy, rdp->qlen);
2446 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2448 /* Go handle any RCU core processing required. */
2449 __call_rcu_core(rsp, rdp, head, flags);
2450 local_irq_restore(flags);
2454 * Queue an RCU-sched callback for invocation after a grace period.
2456 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2458 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2460 EXPORT_SYMBOL_GPL(call_rcu_sched);
2463 * Queue an RCU callback for invocation after a quicker grace period.
2465 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2467 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2469 EXPORT_SYMBOL_GPL(call_rcu_bh);
2472 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2473 * any blocking grace-period wait automatically implies a grace period
2474 * if there is only one CPU online at any point time during execution
2475 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2476 * occasionally incorrectly indicate that there are multiple CPUs online
2477 * when there was in fact only one the whole time, as this just adds
2478 * some overhead: RCU still operates correctly.
2480 static inline int rcu_blocking_is_gp(void)
2484 might_sleep(); /* Check for RCU read-side critical section. */
2486 ret = num_online_cpus() <= 1;
2492 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2494 * Control will return to the caller some time after a full rcu-sched
2495 * grace period has elapsed, in other words after all currently executing
2496 * rcu-sched read-side critical sections have completed. These read-side
2497 * critical sections are delimited by rcu_read_lock_sched() and
2498 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2499 * local_irq_disable(), and so on may be used in place of
2500 * rcu_read_lock_sched().
2502 * This means that all preempt_disable code sequences, including NMI and
2503 * non-threaded hardware-interrupt handlers, in progress on entry will
2504 * have completed before this primitive returns. However, this does not
2505 * guarantee that softirq handlers will have completed, since in some
2506 * kernels, these handlers can run in process context, and can block.
2508 * Note that this guarantee implies further memory-ordering guarantees.
2509 * On systems with more than one CPU, when synchronize_sched() returns,
2510 * each CPU is guaranteed to have executed a full memory barrier since the
2511 * end of its last RCU-sched read-side critical section whose beginning
2512 * preceded the call to synchronize_sched(). In addition, each CPU having
2513 * an RCU read-side critical section that extends beyond the return from
2514 * synchronize_sched() is guaranteed to have executed a full memory barrier
2515 * after the beginning of synchronize_sched() and before the beginning of
2516 * that RCU read-side critical section. Note that these guarantees include
2517 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2518 * that are executing in the kernel.
2520 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2521 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2522 * to have executed a full memory barrier during the execution of
2523 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2524 * again only if the system has more than one CPU).
2526 * This primitive provides the guarantees made by the (now removed)
2527 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2528 * guarantees that rcu_read_lock() sections will have completed.
2529 * In "classic RCU", these two guarantees happen to be one and
2530 * the same, but can differ in realtime RCU implementations.
2532 void synchronize_sched(void)
2534 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2535 !lock_is_held(&rcu_lock_map) &&
2536 !lock_is_held(&rcu_sched_lock_map),
2537 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2538 if (rcu_blocking_is_gp())
2541 synchronize_sched_expedited();
2543 wait_rcu_gp(call_rcu_sched);
2545 EXPORT_SYMBOL_GPL(synchronize_sched);
2548 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2550 * Control will return to the caller some time after a full rcu_bh grace
2551 * period has elapsed, in other words after all currently executing rcu_bh
2552 * read-side critical sections have completed. RCU read-side critical
2553 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2554 * and may be nested.
2556 * See the description of synchronize_sched() for more detailed information
2557 * on memory ordering guarantees.
2559 void synchronize_rcu_bh(void)
2561 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2562 !lock_is_held(&rcu_lock_map) &&
2563 !lock_is_held(&rcu_sched_lock_map),
2564 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2565 if (rcu_blocking_is_gp())
2568 synchronize_rcu_bh_expedited();
2570 wait_rcu_gp(call_rcu_bh);
2572 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2574 static int synchronize_sched_expedited_cpu_stop(void *data)
2577 * There must be a full memory barrier on each affected CPU
2578 * between the time that try_stop_cpus() is called and the
2579 * time that it returns.
2581 * In the current initial implementation of cpu_stop, the
2582 * above condition is already met when the control reaches
2583 * this point and the following smp_mb() is not strictly
2584 * necessary. Do smp_mb() anyway for documentation and
2585 * robustness against future implementation changes.
2587 smp_mb(); /* See above comment block. */
2592 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2594 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2595 * approach to force the grace period to end quickly. This consumes
2596 * significant time on all CPUs and is unfriendly to real-time workloads,
2597 * so is thus not recommended for any sort of common-case code. In fact,
2598 * if you are using synchronize_sched_expedited() in a loop, please
2599 * restructure your code to batch your updates, and then use a single
2600 * synchronize_sched() instead.
2602 * Note that it is illegal to call this function while holding any lock
2603 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2604 * to call this function from a CPU-hotplug notifier. Failing to observe
2605 * these restriction will result in deadlock.
2607 * This implementation can be thought of as an application of ticket
2608 * locking to RCU, with sync_sched_expedited_started and
2609 * sync_sched_expedited_done taking on the roles of the halves
2610 * of the ticket-lock word. Each task atomically increments
2611 * sync_sched_expedited_started upon entry, snapshotting the old value,
2612 * then attempts to stop all the CPUs. If this succeeds, then each
2613 * CPU will have executed a context switch, resulting in an RCU-sched
2614 * grace period. We are then done, so we use atomic_cmpxchg() to
2615 * update sync_sched_expedited_done to match our snapshot -- but
2616 * only if someone else has not already advanced past our snapshot.
2618 * On the other hand, if try_stop_cpus() fails, we check the value
2619 * of sync_sched_expedited_done. If it has advanced past our
2620 * initial snapshot, then someone else must have forced a grace period
2621 * some time after we took our snapshot. In this case, our work is
2622 * done for us, and we can simply return. Otherwise, we try again,
2623 * but keep our initial snapshot for purposes of checking for someone
2624 * doing our work for us.
2626 * If we fail too many times in a row, we fall back to synchronize_sched().
2628 void synchronize_sched_expedited(void)
2630 long firstsnap, s, snap;
2632 struct rcu_state *rsp = &rcu_sched_state;
2635 * If we are in danger of counter wrap, just do synchronize_sched().
2636 * By allowing sync_sched_expedited_started to advance no more than
2637 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2638 * that more than 3.5 billion CPUs would be required to force a
2639 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2640 * course be required on a 64-bit system.
2642 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2643 (ulong)atomic_long_read(&rsp->expedited_done) +
2645 synchronize_sched();
2646 atomic_long_inc(&rsp->expedited_wrap);
2651 * Take a ticket. Note that atomic_inc_return() implies a
2652 * full memory barrier.
2654 snap = atomic_long_inc_return(&rsp->expedited_start);
2657 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2660 * Each pass through the following loop attempts to force a
2661 * context switch on each CPU.
2663 while (try_stop_cpus(cpu_online_mask,
2664 synchronize_sched_expedited_cpu_stop,
2667 atomic_long_inc(&rsp->expedited_tryfail);
2669 /* Check to see if someone else did our work for us. */
2670 s = atomic_long_read(&rsp->expedited_done);
2671 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2672 /* ensure test happens before caller kfree */
2673 smp_mb__before_atomic_inc(); /* ^^^ */
2674 atomic_long_inc(&rsp->expedited_workdone1);
2678 /* No joy, try again later. Or just synchronize_sched(). */
2679 if (trycount++ < 10) {
2680 udelay(trycount * num_online_cpus());
2682 wait_rcu_gp(call_rcu_sched);
2683 atomic_long_inc(&rsp->expedited_normal);
2687 /* Recheck to see if someone else did our work for us. */
2688 s = atomic_long_read(&rsp->expedited_done);
2689 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2690 /* ensure test happens before caller kfree */
2691 smp_mb__before_atomic_inc(); /* ^^^ */
2692 atomic_long_inc(&rsp->expedited_workdone2);
2697 * Refetching sync_sched_expedited_started allows later
2698 * callers to piggyback on our grace period. We retry
2699 * after they started, so our grace period works for them,
2700 * and they started after our first try, so their grace
2701 * period works for us.
2704 snap = atomic_long_read(&rsp->expedited_start);
2705 smp_mb(); /* ensure read is before try_stop_cpus(). */
2707 atomic_long_inc(&rsp->expedited_stoppedcpus);
2710 * Everyone up to our most recent fetch is covered by our grace
2711 * period. Update the counter, but only if our work is still
2712 * relevant -- which it won't be if someone who started later
2713 * than we did already did their update.
2716 atomic_long_inc(&rsp->expedited_done_tries);
2717 s = atomic_long_read(&rsp->expedited_done);
2718 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2719 /* ensure test happens before caller kfree */
2720 smp_mb__before_atomic_inc(); /* ^^^ */
2721 atomic_long_inc(&rsp->expedited_done_lost);
2724 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2725 atomic_long_inc(&rsp->expedited_done_exit);
2729 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2732 * Check to see if there is any immediate RCU-related work to be done
2733 * by the current CPU, for the specified type of RCU, returning 1 if so.
2734 * The checks are in order of increasing expense: checks that can be
2735 * carried out against CPU-local state are performed first. However,
2736 * we must check for CPU stalls first, else we might not get a chance.
2738 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2740 struct rcu_node *rnp = rdp->mynode;
2742 rdp->n_rcu_pending++;
2744 /* Check for CPU stalls, if enabled. */
2745 check_cpu_stall(rsp, rdp);
2747 /* Is the RCU core waiting for a quiescent state from this CPU? */
2748 if (rcu_scheduler_fully_active &&
2749 rdp->qs_pending && !rdp->passed_quiesce) {
2750 rdp->n_rp_qs_pending++;
2751 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2752 rdp->n_rp_report_qs++;
2756 /* Does this CPU have callbacks ready to invoke? */
2757 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2758 rdp->n_rp_cb_ready++;
2762 /* Has RCU gone idle with this CPU needing another grace period? */
2763 if (cpu_needs_another_gp(rsp, rdp)) {
2764 rdp->n_rp_cpu_needs_gp++;
2768 /* Has another RCU grace period completed? */
2769 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2770 rdp->n_rp_gp_completed++;
2774 /* Has a new RCU grace period started? */
2775 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2776 rdp->n_rp_gp_started++;
2781 rdp->n_rp_need_nothing++;
2786 * Check to see if there is any immediate RCU-related work to be done
2787 * by the current CPU, returning 1 if so. This function is part of the
2788 * RCU implementation; it is -not- an exported member of the RCU API.
2790 static int rcu_pending(int cpu)
2792 struct rcu_state *rsp;
2794 for_each_rcu_flavor(rsp)
2795 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2801 * Return true if the specified CPU has any callback. If all_lazy is
2802 * non-NULL, store an indication of whether all callbacks are lazy.
2803 * (If there are no callbacks, all of them are deemed to be lazy.)
2805 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2809 struct rcu_data *rdp;
2810 struct rcu_state *rsp;
2812 for_each_rcu_flavor(rsp) {
2813 rdp = per_cpu_ptr(rsp->rda, cpu);
2817 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2828 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2829 * the compiler is expected to optimize this away.
2831 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2832 int cpu, unsigned long done)
2834 trace_rcu_barrier(rsp->name, s, cpu,
2835 atomic_read(&rsp->barrier_cpu_count), done);
2839 * RCU callback function for _rcu_barrier(). If we are last, wake
2840 * up the task executing _rcu_barrier().
2842 static void rcu_barrier_callback(struct rcu_head *rhp)
2844 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2845 struct rcu_state *rsp = rdp->rsp;
2847 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2848 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2849 complete(&rsp->barrier_completion);
2851 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2856 * Called with preemption disabled, and from cross-cpu IRQ context.
2858 static void rcu_barrier_func(void *type)
2860 struct rcu_state *rsp = type;
2861 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2863 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2864 atomic_inc(&rsp->barrier_cpu_count);
2865 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2869 * Orchestrate the specified type of RCU barrier, waiting for all
2870 * RCU callbacks of the specified type to complete.
2872 static void _rcu_barrier(struct rcu_state *rsp)
2875 struct rcu_data *rdp;
2876 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2877 unsigned long snap_done;
2879 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2881 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2882 mutex_lock(&rsp->barrier_mutex);
2885 * Ensure that all prior references, including to ->n_barrier_done,
2886 * are ordered before the _rcu_barrier() machinery.
2888 smp_mb(); /* See above block comment. */
2891 * Recheck ->n_barrier_done to see if others did our work for us.
2892 * This means checking ->n_barrier_done for an even-to-odd-to-even
2893 * transition. The "if" expression below therefore rounds the old
2894 * value up to the next even number and adds two before comparing.
2896 snap_done = rsp->n_barrier_done;
2897 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2900 * If the value in snap is odd, we needed to wait for the current
2901 * rcu_barrier() to complete, then wait for the next one, in other
2902 * words, we need the value of snap_done to be three larger than
2903 * the value of snap. On the other hand, if the value in snap is
2904 * even, we only had to wait for the next rcu_barrier() to complete,
2905 * in other words, we need the value of snap_done to be only two
2906 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2907 * this for us (thank you, Linus!).
2909 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
2910 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2911 smp_mb(); /* caller's subsequent code after above check. */
2912 mutex_unlock(&rsp->barrier_mutex);
2917 * Increment ->n_barrier_done to avoid duplicate work. Use
2918 * ACCESS_ONCE() to prevent the compiler from speculating
2919 * the increment to precede the early-exit check.
2921 ACCESS_ONCE(rsp->n_barrier_done)++;
2922 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2923 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2924 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2927 * Initialize the count to one rather than to zero in order to
2928 * avoid a too-soon return to zero in case of a short grace period
2929 * (or preemption of this task). Exclude CPU-hotplug operations
2930 * to ensure that no offline CPU has callbacks queued.
2932 init_completion(&rsp->barrier_completion);
2933 atomic_set(&rsp->barrier_cpu_count, 1);
2937 * Force each CPU with callbacks to register a new callback.
2938 * When that callback is invoked, we will know that all of the
2939 * corresponding CPU's preceding callbacks have been invoked.
2941 for_each_possible_cpu(cpu) {
2942 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2944 rdp = per_cpu_ptr(rsp->rda, cpu);
2945 if (rcu_is_nocb_cpu(cpu)) {
2946 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2947 rsp->n_barrier_done);
2948 atomic_inc(&rsp->barrier_cpu_count);
2949 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2951 } else if (ACCESS_ONCE(rdp->qlen)) {
2952 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2953 rsp->n_barrier_done);
2954 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2956 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2957 rsp->n_barrier_done);
2963 * Now that we have an rcu_barrier_callback() callback on each
2964 * CPU, and thus each counted, remove the initial count.
2966 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2967 complete(&rsp->barrier_completion);
2969 /* Increment ->n_barrier_done to prevent duplicate work. */
2970 smp_mb(); /* Keep increment after above mechanism. */
2971 ACCESS_ONCE(rsp->n_barrier_done)++;
2972 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2973 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2974 smp_mb(); /* Keep increment before caller's subsequent code. */
2976 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2977 wait_for_completion(&rsp->barrier_completion);
2979 /* Other rcu_barrier() invocations can now safely proceed. */
2980 mutex_unlock(&rsp->barrier_mutex);
2984 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2986 void rcu_barrier_bh(void)
2988 _rcu_barrier(&rcu_bh_state);
2990 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2993 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2995 void rcu_barrier_sched(void)
2997 _rcu_barrier(&rcu_sched_state);
2999 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3002 * Do boot-time initialization of a CPU's per-CPU RCU data.
3005 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3007 unsigned long flags;
3008 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3009 struct rcu_node *rnp = rcu_get_root(rsp);
3011 /* Set up local state, ensuring consistent view of global state. */
3012 raw_spin_lock_irqsave(&rnp->lock, flags);
3013 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3014 init_callback_list(rdp);
3016 ACCESS_ONCE(rdp->qlen) = 0;
3017 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3018 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3019 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3022 rcu_boot_init_nocb_percpu_data(rdp);
3023 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3027 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3028 * offline event can be happening at a given time. Note also that we
3029 * can accept some slop in the rsp->completed access due to the fact
3030 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3033 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3035 unsigned long flags;
3037 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3038 struct rcu_node *rnp = rcu_get_root(rsp);
3040 /* Exclude new grace periods. */
3041 mutex_lock(&rsp->onoff_mutex);
3043 /* Set up local state, ensuring consistent view of global state. */
3044 raw_spin_lock_irqsave(&rnp->lock, flags);
3045 rdp->beenonline = 1; /* We have now been online. */
3046 rdp->preemptible = preemptible;
3047 rdp->qlen_last_fqs_check = 0;
3048 rdp->n_force_qs_snap = rsp->n_force_qs;
3049 rdp->blimit = blimit;
3050 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3051 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3052 rcu_sysidle_init_percpu_data(rdp->dynticks);
3053 atomic_set(&rdp->dynticks->dynticks,
3054 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3055 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3057 /* Add CPU to rcu_node bitmasks. */
3059 mask = rdp->grpmask;
3061 /* Exclude any attempts to start a new GP on small systems. */
3062 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3063 rnp->qsmaskinit |= mask;
3064 mask = rnp->grpmask;
3065 if (rnp == rdp->mynode) {
3067 * If there is a grace period in progress, we will
3068 * set up to wait for it next time we run the
3071 rdp->gpnum = rnp->completed;
3072 rdp->completed = rnp->completed;
3073 rdp->passed_quiesce = 0;
3074 rdp->qs_pending = 0;
3075 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3077 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3079 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3080 local_irq_restore(flags);
3082 mutex_unlock(&rsp->onoff_mutex);
3085 static void rcu_prepare_cpu(int cpu)
3087 struct rcu_state *rsp;
3089 for_each_rcu_flavor(rsp)
3090 rcu_init_percpu_data(cpu, rsp,
3091 strcmp(rsp->name, "rcu_preempt") == 0);
3095 * Handle CPU online/offline notification events.
3097 static int rcu_cpu_notify(struct notifier_block *self,
3098 unsigned long action, void *hcpu)
3100 long cpu = (long)hcpu;
3101 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3102 struct rcu_node *rnp = rdp->mynode;
3103 struct rcu_state *rsp;
3105 trace_rcu_utilization(TPS("Start CPU hotplug"));
3107 case CPU_UP_PREPARE:
3108 case CPU_UP_PREPARE_FROZEN:
3109 rcu_prepare_cpu(cpu);
3110 rcu_prepare_kthreads(cpu);
3113 case CPU_DOWN_FAILED:
3114 rcu_boost_kthread_setaffinity(rnp, -1);
3116 case CPU_DOWN_PREPARE:
3117 rcu_boost_kthread_setaffinity(rnp, cpu);
3120 case CPU_DYING_FROZEN:
3121 for_each_rcu_flavor(rsp)
3122 rcu_cleanup_dying_cpu(rsp);
3125 case CPU_DEAD_FROZEN:
3126 case CPU_UP_CANCELED:
3127 case CPU_UP_CANCELED_FROZEN:
3128 for_each_rcu_flavor(rsp)
3129 rcu_cleanup_dead_cpu(cpu, rsp);
3134 trace_rcu_utilization(TPS("End CPU hotplug"));
3138 static int rcu_pm_notify(struct notifier_block *self,
3139 unsigned long action, void *hcpu)
3142 case PM_HIBERNATION_PREPARE:
3143 case PM_SUSPEND_PREPARE:
3144 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3147 case PM_POST_HIBERNATION:
3148 case PM_POST_SUSPEND:
3158 * Spawn the kthread that handles this RCU flavor's grace periods.
3160 static int __init rcu_spawn_gp_kthread(void)
3162 unsigned long flags;
3163 struct rcu_node *rnp;
3164 struct rcu_state *rsp;
3165 struct task_struct *t;
3167 for_each_rcu_flavor(rsp) {
3168 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3170 rnp = rcu_get_root(rsp);
3171 raw_spin_lock_irqsave(&rnp->lock, flags);
3172 rsp->gp_kthread = t;
3173 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3174 rcu_spawn_nocb_kthreads(rsp);
3178 early_initcall(rcu_spawn_gp_kthread);
3181 * This function is invoked towards the end of the scheduler's initialization
3182 * process. Before this is called, the idle task might contain
3183 * RCU read-side critical sections (during which time, this idle
3184 * task is booting the system). After this function is called, the
3185 * idle tasks are prohibited from containing RCU read-side critical
3186 * sections. This function also enables RCU lockdep checking.
3188 void rcu_scheduler_starting(void)
3190 WARN_ON(num_online_cpus() != 1);
3191 WARN_ON(nr_context_switches() > 0);
3192 rcu_scheduler_active = 1;
3196 * Compute the per-level fanout, either using the exact fanout specified
3197 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3199 #ifdef CONFIG_RCU_FANOUT_EXACT
3200 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3204 for (i = rcu_num_lvls - 1; i > 0; i--)
3205 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3206 rsp->levelspread[0] = rcu_fanout_leaf;
3208 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3209 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3216 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3217 ccur = rsp->levelcnt[i];
3218 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3222 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3225 * Helper function for rcu_init() that initializes one rcu_state structure.
3227 static void __init rcu_init_one(struct rcu_state *rsp,
3228 struct rcu_data __percpu *rda)
3230 static char *buf[] = { "rcu_node_0",
3233 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3234 static char *fqs[] = { "rcu_node_fqs_0",
3237 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3241 struct rcu_node *rnp;
3243 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3245 /* Silence gcc 4.8 warning about array index out of range. */
3246 if (rcu_num_lvls > RCU_NUM_LVLS)
3247 panic("rcu_init_one: rcu_num_lvls overflow");
3249 /* Initialize the level-tracking arrays. */
3251 for (i = 0; i < rcu_num_lvls; i++)
3252 rsp->levelcnt[i] = num_rcu_lvl[i];
3253 for (i = 1; i < rcu_num_lvls; i++)
3254 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3255 rcu_init_levelspread(rsp);
3257 /* Initialize the elements themselves, starting from the leaves. */
3259 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3260 cpustride *= rsp->levelspread[i];
3261 rnp = rsp->level[i];
3262 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3263 raw_spin_lock_init(&rnp->lock);
3264 lockdep_set_class_and_name(&rnp->lock,
3265 &rcu_node_class[i], buf[i]);
3266 raw_spin_lock_init(&rnp->fqslock);
3267 lockdep_set_class_and_name(&rnp->fqslock,
3268 &rcu_fqs_class[i], fqs[i]);
3269 rnp->gpnum = rsp->gpnum;
3270 rnp->completed = rsp->completed;
3272 rnp->qsmaskinit = 0;
3273 rnp->grplo = j * cpustride;
3274 rnp->grphi = (j + 1) * cpustride - 1;
3275 if (rnp->grphi >= NR_CPUS)
3276 rnp->grphi = NR_CPUS - 1;
3282 rnp->grpnum = j % rsp->levelspread[i - 1];
3283 rnp->grpmask = 1UL << rnp->grpnum;
3284 rnp->parent = rsp->level[i - 1] +
3285 j / rsp->levelspread[i - 1];
3288 INIT_LIST_HEAD(&rnp->blkd_tasks);
3289 rcu_init_one_nocb(rnp);
3294 init_waitqueue_head(&rsp->gp_wq);
3295 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3296 rnp = rsp->level[rcu_num_lvls - 1];
3297 for_each_possible_cpu(i) {
3298 while (i > rnp->grphi)
3300 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3301 rcu_boot_init_percpu_data(i, rsp);
3303 list_add(&rsp->flavors, &rcu_struct_flavors);
3307 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3308 * replace the definitions in tree.h because those are needed to size
3309 * the ->node array in the rcu_state structure.
3311 static void __init rcu_init_geometry(void)
3317 int rcu_capacity[MAX_RCU_LVLS + 1];
3320 * Initialize any unspecified boot parameters.
3321 * The default values of jiffies_till_first_fqs and
3322 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3323 * value, which is a function of HZ, then adding one for each
3324 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3326 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3327 if (jiffies_till_first_fqs == ULONG_MAX)
3328 jiffies_till_first_fqs = d;
3329 if (jiffies_till_next_fqs == ULONG_MAX)
3330 jiffies_till_next_fqs = d;
3332 /* If the compile-time values are accurate, just leave. */
3333 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3334 nr_cpu_ids == NR_CPUS)
3338 * Compute number of nodes that can be handled an rcu_node tree
3339 * with the given number of levels. Setting rcu_capacity[0] makes
3340 * some of the arithmetic easier.
3342 rcu_capacity[0] = 1;
3343 rcu_capacity[1] = rcu_fanout_leaf;
3344 for (i = 2; i <= MAX_RCU_LVLS; i++)
3345 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3348 * The boot-time rcu_fanout_leaf parameter is only permitted
3349 * to increase the leaf-level fanout, not decrease it. Of course,
3350 * the leaf-level fanout cannot exceed the number of bits in
3351 * the rcu_node masks. Finally, the tree must be able to accommodate
3352 * the configured number of CPUs. Complain and fall back to the
3353 * compile-time values if these limits are exceeded.
3355 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3356 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3357 n > rcu_capacity[MAX_RCU_LVLS]) {
3362 /* Calculate the number of rcu_nodes at each level of the tree. */
3363 for (i = 1; i <= MAX_RCU_LVLS; i++)
3364 if (n <= rcu_capacity[i]) {
3365 for (j = 0; j <= i; j++)
3367 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3369 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3374 /* Calculate the total number of rcu_node structures. */
3376 for (i = 0; i <= MAX_RCU_LVLS; i++)
3377 rcu_num_nodes += num_rcu_lvl[i];
3381 void __init rcu_init(void)
3385 rcu_bootup_announce();
3386 rcu_init_geometry();
3387 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3388 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3389 __rcu_init_preempt();
3390 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3393 * We don't need protection against CPU-hotplug here because
3394 * this is called early in boot, before either interrupts
3395 * or the scheduler are operational.
3397 cpu_notifier(rcu_cpu_notify, 0);
3398 pm_notifier(rcu_pm_notify, 0);
3399 for_each_online_cpu(cpu)
3400 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3403 #include "tree_plugin.h"