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/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/suspend.h>
60 #include <trace/events/rcu.h>
65 * Strings used in tracepoints need to be exported via the
66 * tracing system such that tools like perf and trace-cmd can
67 * translate the string address pointers to actual text.
69 #define TPS(x) tracepoint_string(x)
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 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 = idle_task(smp_processor_id());
376 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
377 ftrace_dump(DUMP_ORIG);
378 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
379 current->pid, current->comm,
380 idle->pid, idle->comm); /* must be idle task! */
382 rcu_prepare_for_idle(smp_processor_id());
383 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
384 smp_mb__before_atomic_inc(); /* See above. */
385 atomic_inc(&rdtp->dynticks);
386 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
387 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
390 * It is illegal to enter an extended quiescent state while
391 * in an RCU read-side critical section.
393 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
394 "Illegal idle entry in RCU read-side critical section.");
395 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
396 "Illegal idle entry in RCU-bh read-side critical section.");
397 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
398 "Illegal idle entry in RCU-sched read-side critical section.");
402 * Enter an RCU extended quiescent state, which can be either the
403 * idle loop or adaptive-tickless usermode execution.
405 static void rcu_eqs_enter(bool user)
408 struct rcu_dynticks *rdtp;
410 rdtp = &__get_cpu_var(rcu_dynticks);
411 oldval = rdtp->dynticks_nesting;
412 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
413 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
414 rdtp->dynticks_nesting = 0;
416 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
417 rcu_eqs_enter_common(rdtp, oldval, user);
421 * rcu_idle_enter - inform RCU that current CPU is entering idle
423 * Enter idle mode, in other words, -leave- the mode in which RCU
424 * read-side critical sections can occur. (Though RCU read-side
425 * critical sections can occur in irq handlers in idle, a possibility
426 * handled by irq_enter() and irq_exit().)
428 * We crowbar the ->dynticks_nesting field to zero to allow for
429 * the possibility of usermode upcalls having messed up our count
430 * of interrupt nesting level during the prior busy period.
432 void rcu_idle_enter(void)
436 local_irq_save(flags);
437 rcu_eqs_enter(false);
438 rcu_sysidle_enter(&__get_cpu_var(rcu_dynticks), 0);
439 local_irq_restore(flags);
441 EXPORT_SYMBOL_GPL(rcu_idle_enter);
443 #ifdef CONFIG_RCU_USER_QS
445 * rcu_user_enter - inform RCU that we are resuming userspace.
447 * Enter RCU idle mode right before resuming userspace. No use of RCU
448 * is permitted between this call and rcu_user_exit(). This way the
449 * CPU doesn't need to maintain the tick for RCU maintenance purposes
450 * when the CPU runs in userspace.
452 void rcu_user_enter(void)
456 #endif /* CONFIG_RCU_USER_QS */
459 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
461 * Exit from an interrupt handler, which might possibly result in entering
462 * idle mode, in other words, leaving the mode in which read-side critical
463 * sections can occur.
465 * This code assumes that the idle loop never does anything that might
466 * result in unbalanced calls to irq_enter() and irq_exit(). If your
467 * architecture violates this assumption, RCU will give you what you
468 * deserve, good and hard. But very infrequently and irreproducibly.
470 * Use things like work queues to work around this limitation.
472 * You have been warned.
474 void rcu_irq_exit(void)
478 struct rcu_dynticks *rdtp;
480 local_irq_save(flags);
481 rdtp = &__get_cpu_var(rcu_dynticks);
482 oldval = rdtp->dynticks_nesting;
483 rdtp->dynticks_nesting--;
484 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
485 if (rdtp->dynticks_nesting)
486 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
488 rcu_eqs_enter_common(rdtp, oldval, true);
489 rcu_sysidle_enter(rdtp, 1);
490 local_irq_restore(flags);
494 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
496 * If the new value of the ->dynticks_nesting counter was previously zero,
497 * we really have exited idle, and must do the appropriate accounting.
498 * The caller must have disabled interrupts.
500 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
503 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
504 atomic_inc(&rdtp->dynticks);
505 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
506 smp_mb__after_atomic_inc(); /* See above. */
507 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
508 rcu_cleanup_after_idle(smp_processor_id());
509 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
510 if (!user && !is_idle_task(current)) {
511 struct task_struct *idle = idle_task(smp_processor_id());
513 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
514 oldval, rdtp->dynticks_nesting);
515 ftrace_dump(DUMP_ORIG);
516 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
517 current->pid, current->comm,
518 idle->pid, idle->comm); /* must be idle task! */
523 * Exit an RCU extended quiescent state, which can be either the
524 * idle loop or adaptive-tickless usermode execution.
526 static void rcu_eqs_exit(bool user)
528 struct rcu_dynticks *rdtp;
531 rdtp = &__get_cpu_var(rcu_dynticks);
532 oldval = rdtp->dynticks_nesting;
533 WARN_ON_ONCE(oldval < 0);
534 if (oldval & DYNTICK_TASK_NEST_MASK)
535 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
537 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
538 rcu_eqs_exit_common(rdtp, oldval, user);
542 * rcu_idle_exit - inform RCU that current CPU is leaving idle
544 * Exit idle mode, in other words, -enter- the mode in which RCU
545 * read-side critical sections can occur.
547 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
548 * allow for the possibility of usermode upcalls messing up our count
549 * of interrupt nesting level during the busy period that is just
552 void rcu_idle_exit(void)
556 local_irq_save(flags);
558 rcu_sysidle_exit(&__get_cpu_var(rcu_dynticks), 0);
559 local_irq_restore(flags);
561 EXPORT_SYMBOL_GPL(rcu_idle_exit);
563 #ifdef CONFIG_RCU_USER_QS
565 * rcu_user_exit - inform RCU that we are exiting userspace.
567 * Exit RCU idle mode while entering the kernel because it can
568 * run a RCU read side critical section anytime.
570 void rcu_user_exit(void)
574 #endif /* CONFIG_RCU_USER_QS */
577 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
579 * Enter an interrupt handler, which might possibly result in exiting
580 * idle mode, in other words, entering the mode in which read-side critical
581 * sections can occur.
583 * Note that the Linux kernel is fully capable of entering an interrupt
584 * handler that it never exits, for example when doing upcalls to
585 * user mode! This code assumes that the idle loop never does upcalls to
586 * user mode. If your architecture does do upcalls from the idle loop (or
587 * does anything else that results in unbalanced calls to the irq_enter()
588 * and irq_exit() functions), RCU will give you what you deserve, good
589 * and hard. But very infrequently and irreproducibly.
591 * Use things like work queues to work around this limitation.
593 * You have been warned.
595 void rcu_irq_enter(void)
598 struct rcu_dynticks *rdtp;
601 local_irq_save(flags);
602 rdtp = &__get_cpu_var(rcu_dynticks);
603 oldval = rdtp->dynticks_nesting;
604 rdtp->dynticks_nesting++;
605 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
607 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
609 rcu_eqs_exit_common(rdtp, oldval, true);
610 rcu_sysidle_exit(rdtp, 1);
611 local_irq_restore(flags);
615 * rcu_nmi_enter - inform RCU of entry to NMI context
617 * If the CPU was idle with dynamic ticks active, and there is no
618 * irq handler running, this updates rdtp->dynticks_nmi to let the
619 * RCU grace-period handling know that the CPU is active.
621 void rcu_nmi_enter(void)
623 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
625 if (rdtp->dynticks_nmi_nesting == 0 &&
626 (atomic_read(&rdtp->dynticks) & 0x1))
628 rdtp->dynticks_nmi_nesting++;
629 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
630 atomic_inc(&rdtp->dynticks);
631 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
632 smp_mb__after_atomic_inc(); /* See above. */
633 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
637 * rcu_nmi_exit - inform RCU of exit from NMI context
639 * If the CPU was idle with dynamic ticks active, and there is no
640 * irq handler running, this updates rdtp->dynticks_nmi to let the
641 * RCU grace-period handling know that the CPU is no longer active.
643 void rcu_nmi_exit(void)
645 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
647 if (rdtp->dynticks_nmi_nesting == 0 ||
648 --rdtp->dynticks_nmi_nesting != 0)
650 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
651 smp_mb__before_atomic_inc(); /* See above. */
652 atomic_inc(&rdtp->dynticks);
653 smp_mb__after_atomic_inc(); /* Force delay to next write. */
654 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
658 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
660 * If the current CPU is in its idle loop and is neither in an interrupt
661 * or NMI handler, return true.
663 int rcu_is_cpu_idle(void)
668 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
672 EXPORT_SYMBOL(rcu_is_cpu_idle);
674 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
677 * Is the current CPU online? Disable preemption to avoid false positives
678 * that could otherwise happen due to the current CPU number being sampled,
679 * this task being preempted, its old CPU being taken offline, resuming
680 * on some other CPU, then determining that its old CPU is now offline.
681 * It is OK to use RCU on an offline processor during initial boot, hence
682 * the check for rcu_scheduler_fully_active. Note also that it is OK
683 * for a CPU coming online to use RCU for one jiffy prior to marking itself
684 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
685 * offline to continue to use RCU for one jiffy after marking itself
686 * offline in the cpu_online_mask. This leniency is necessary given the
687 * non-atomic nature of the online and offline processing, for example,
688 * the fact that a CPU enters the scheduler after completing the CPU_DYING
691 * This is also why RCU internally marks CPUs online during the
692 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
694 * Disable checking if in an NMI handler because we cannot safely report
695 * errors from NMI handlers anyway.
697 bool rcu_lockdep_current_cpu_online(void)
699 struct rcu_data *rdp;
700 struct rcu_node *rnp;
706 rdp = &__get_cpu_var(rcu_sched_data);
708 ret = (rdp->grpmask & rnp->qsmaskinit) ||
709 !rcu_scheduler_fully_active;
713 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
715 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
718 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
720 * If the current CPU is idle or running at a first-level (not nested)
721 * interrupt from idle, return true. The caller must have at least
722 * disabled preemption.
724 static int rcu_is_cpu_rrupt_from_idle(void)
726 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
730 * Snapshot the specified CPU's dynticks counter so that we can later
731 * credit them with an implicit quiescent state. Return 1 if this CPU
732 * is in dynticks idle mode, which is an extended quiescent state.
734 static int dyntick_save_progress_counter(struct rcu_data *rdp,
735 bool *isidle, unsigned long *maxj)
737 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
738 rcu_sysidle_check_cpu(rdp, isidle, maxj);
739 return (rdp->dynticks_snap & 0x1) == 0;
743 * Return true if the specified CPU has passed through a quiescent
744 * state by virtue of being in or having passed through an dynticks
745 * idle state since the last call to dyntick_save_progress_counter()
746 * for this same CPU, or by virtue of having been offline.
748 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
749 bool *isidle, unsigned long *maxj)
754 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
755 snap = (unsigned int)rdp->dynticks_snap;
758 * If the CPU passed through or entered a dynticks idle phase with
759 * no active irq/NMI handlers, then we can safely pretend that the CPU
760 * already acknowledged the request to pass through a quiescent
761 * state. Either way, that CPU cannot possibly be in an RCU
762 * read-side critical section that started before the beginning
763 * of the current RCU grace period.
765 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
766 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
772 * Check for the CPU being offline, but only if the grace period
773 * is old enough. We don't need to worry about the CPU changing
774 * state: If we see it offline even once, it has been through a
777 * The reason for insisting that the grace period be at least
778 * one jiffy old is that CPUs that are not quite online and that
779 * have just gone offline can still execute RCU read-side critical
782 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
783 return 0; /* Grace period is not old enough. */
785 if (cpu_is_offline(rdp->cpu)) {
786 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
792 * There is a possibility that a CPU in adaptive-ticks state
793 * might run in the kernel with the scheduling-clock tick disabled
794 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
795 * force the CPU to restart the scheduling-clock tick in this
796 * CPU is in this state.
798 rcu_kick_nohz_cpu(rdp->cpu);
803 static void record_gp_stall_check_time(struct rcu_state *rsp)
805 rsp->gp_start = jiffies;
806 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
810 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
811 * for architectures that do not implement trigger_all_cpu_backtrace().
812 * The NMI-triggered stack traces are more accurate because they are
813 * printed by the target CPU.
815 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
819 struct rcu_node *rnp;
821 rcu_for_each_leaf_node(rsp, rnp) {
822 raw_spin_lock_irqsave(&rnp->lock, flags);
823 if (rnp->qsmask != 0) {
824 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
825 if (rnp->qsmask & (1UL << cpu))
826 dump_cpu_task(rnp->grplo + cpu);
828 raw_spin_unlock_irqrestore(&rnp->lock, flags);
832 static void print_other_cpu_stall(struct rcu_state *rsp)
838 struct rcu_node *rnp = rcu_get_root(rsp);
841 /* Only let one CPU complain about others per time interval. */
843 raw_spin_lock_irqsave(&rnp->lock, flags);
844 delta = jiffies - rsp->jiffies_stall;
845 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
846 raw_spin_unlock_irqrestore(&rnp->lock, flags);
849 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
850 raw_spin_unlock_irqrestore(&rnp->lock, flags);
853 * OK, time to rat on our buddy...
854 * See Documentation/RCU/stallwarn.txt for info on how to debug
855 * RCU CPU stall warnings.
857 pr_err("INFO: %s detected stalls on CPUs/tasks:",
859 print_cpu_stall_info_begin();
860 rcu_for_each_leaf_node(rsp, rnp) {
861 raw_spin_lock_irqsave(&rnp->lock, flags);
862 ndetected += rcu_print_task_stall(rnp);
863 if (rnp->qsmask != 0) {
864 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
865 if (rnp->qsmask & (1UL << cpu)) {
866 print_cpu_stall_info(rsp,
871 raw_spin_unlock_irqrestore(&rnp->lock, flags);
875 * Now rat on any tasks that got kicked up to the root rcu_node
876 * due to CPU offlining.
878 rnp = rcu_get_root(rsp);
879 raw_spin_lock_irqsave(&rnp->lock, flags);
880 ndetected += rcu_print_task_stall(rnp);
881 raw_spin_unlock_irqrestore(&rnp->lock, flags);
883 print_cpu_stall_info_end();
884 for_each_possible_cpu(cpu)
885 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
886 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
887 smp_processor_id(), (long)(jiffies - rsp->gp_start),
888 rsp->gpnum, rsp->completed, totqlen);
890 pr_err("INFO: Stall ended before state dump start\n");
891 else if (!trigger_all_cpu_backtrace())
892 rcu_dump_cpu_stacks(rsp);
894 /* Complain about tasks blocking the grace period. */
896 rcu_print_detail_task_stall(rsp);
898 force_quiescent_state(rsp); /* Kick them all. */
902 * This function really isn't for public consumption, but RCU is special in
903 * that context switches can allow the state machine to make progress.
905 extern void resched_cpu(int cpu);
907 static void print_cpu_stall(struct rcu_state *rsp)
911 struct rcu_node *rnp = rcu_get_root(rsp);
915 * OK, time to rat on ourselves...
916 * See Documentation/RCU/stallwarn.txt for info on how to debug
917 * RCU CPU stall warnings.
919 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
920 print_cpu_stall_info_begin();
921 print_cpu_stall_info(rsp, smp_processor_id());
922 print_cpu_stall_info_end();
923 for_each_possible_cpu(cpu)
924 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
925 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
926 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
927 if (!trigger_all_cpu_backtrace())
930 raw_spin_lock_irqsave(&rnp->lock, flags);
931 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
932 rsp->jiffies_stall = jiffies +
933 3 * rcu_jiffies_till_stall_check() + 3;
934 raw_spin_unlock_irqrestore(&rnp->lock, flags);
937 * Attempt to revive the RCU machinery by forcing a context switch.
939 * A context switch would normally allow the RCU state machine to make
940 * progress and it could be we're stuck in kernel space without context
941 * switches for an entirely unreasonable amount of time.
943 resched_cpu(smp_processor_id());
946 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
950 struct rcu_node *rnp;
952 if (rcu_cpu_stall_suppress)
954 j = ACCESS_ONCE(jiffies);
955 js = ACCESS_ONCE(rsp->jiffies_stall);
957 if (rcu_gp_in_progress(rsp) &&
958 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
960 /* We haven't checked in, so go dump stack. */
961 print_cpu_stall(rsp);
963 } else if (rcu_gp_in_progress(rsp) &&
964 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
966 /* They had a few time units to dump stack, so complain. */
967 print_other_cpu_stall(rsp);
972 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
974 * Set the stall-warning timeout way off into the future, thus preventing
975 * any RCU CPU stall-warning messages from appearing in the current set of
978 * The caller must disable hard irqs.
980 void rcu_cpu_stall_reset(void)
982 struct rcu_state *rsp;
984 for_each_rcu_flavor(rsp)
985 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
989 * Initialize the specified rcu_data structure's callback list to empty.
991 static void init_callback_list(struct rcu_data *rdp)
995 if (init_nocb_callback_list(rdp))
998 for (i = 0; i < RCU_NEXT_SIZE; i++)
999 rdp->nxttail[i] = &rdp->nxtlist;
1003 * Determine the value that ->completed will have at the end of the
1004 * next subsequent grace period. This is used to tag callbacks so that
1005 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1006 * been dyntick-idle for an extended period with callbacks under the
1007 * influence of RCU_FAST_NO_HZ.
1009 * The caller must hold rnp->lock with interrupts disabled.
1011 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1012 struct rcu_node *rnp)
1015 * If RCU is idle, we just wait for the next grace period.
1016 * But we can only be sure that RCU is idle if we are looking
1017 * at the root rcu_node structure -- otherwise, a new grace
1018 * period might have started, but just not yet gotten around
1019 * to initializing the current non-root rcu_node structure.
1021 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1022 return rnp->completed + 1;
1025 * Otherwise, wait for a possible partial grace period and
1026 * then the subsequent full grace period.
1028 return rnp->completed + 2;
1032 * Trace-event helper function for rcu_start_future_gp() and
1033 * rcu_nocb_wait_gp().
1035 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1036 unsigned long c, const char *s)
1038 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1039 rnp->completed, c, rnp->level,
1040 rnp->grplo, rnp->grphi, s);
1044 * Start some future grace period, as needed to handle newly arrived
1045 * callbacks. The required future grace periods are recorded in each
1046 * rcu_node structure's ->need_future_gp field.
1048 * The caller must hold the specified rcu_node structure's ->lock.
1050 static unsigned long __maybe_unused
1051 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1055 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1058 * Pick up grace-period number for new callbacks. If this
1059 * grace period is already marked as needed, return to the caller.
1061 c = rcu_cbs_completed(rdp->rsp, rnp);
1062 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1063 if (rnp->need_future_gp[c & 0x1]) {
1064 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1069 * If either this rcu_node structure or the root rcu_node structure
1070 * believe that a grace period is in progress, then we must wait
1071 * for the one following, which is in "c". Because our request
1072 * will be noticed at the end of the current grace period, we don't
1073 * need to explicitly start one.
1075 if (rnp->gpnum != rnp->completed ||
1076 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1077 rnp->need_future_gp[c & 0x1]++;
1078 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1083 * There might be no grace period in progress. If we don't already
1084 * hold it, acquire the root rcu_node structure's lock in order to
1085 * start one (if needed).
1087 if (rnp != rnp_root)
1088 raw_spin_lock(&rnp_root->lock);
1091 * Get a new grace-period number. If there really is no grace
1092 * period in progress, it will be smaller than the one we obtained
1093 * earlier. Adjust callbacks as needed. Note that even no-CBs
1094 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1096 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1097 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1098 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1099 rdp->nxtcompleted[i] = c;
1102 * If the needed for the required grace period is already
1103 * recorded, trace and leave.
1105 if (rnp_root->need_future_gp[c & 0x1]) {
1106 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1110 /* Record the need for the future grace period. */
1111 rnp_root->need_future_gp[c & 0x1]++;
1113 /* If a grace period is not already in progress, start one. */
1114 if (rnp_root->gpnum != rnp_root->completed) {
1115 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1117 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1118 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1121 if (rnp != rnp_root)
1122 raw_spin_unlock(&rnp_root->lock);
1127 * Clean up any old requests for the just-ended grace period. Also return
1128 * whether any additional grace periods have been requested. Also invoke
1129 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1130 * waiting for this grace period to complete.
1132 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1134 int c = rnp->completed;
1136 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1138 rcu_nocb_gp_cleanup(rsp, rnp);
1139 rnp->need_future_gp[c & 0x1] = 0;
1140 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1141 trace_rcu_future_gp(rnp, rdp, c,
1142 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1147 * If there is room, assign a ->completed number to any callbacks on
1148 * this CPU that have not already been assigned. Also accelerate any
1149 * callbacks that were previously assigned a ->completed number that has
1150 * since proven to be too conservative, which can happen if callbacks get
1151 * assigned a ->completed number while RCU is idle, but with reference to
1152 * a non-root rcu_node structure. This function is idempotent, so it does
1153 * not hurt to call it repeatedly.
1155 * The caller must hold rnp->lock with interrupts disabled.
1157 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1158 struct rcu_data *rdp)
1163 /* If the CPU has no callbacks, nothing to do. */
1164 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1168 * Starting from the sublist containing the callbacks most
1169 * recently assigned a ->completed number and working down, find the
1170 * first sublist that is not assignable to an upcoming grace period.
1171 * Such a sublist has something in it (first two tests) and has
1172 * a ->completed number assigned that will complete sooner than
1173 * the ->completed number for newly arrived callbacks (last test).
1175 * The key point is that any later sublist can be assigned the
1176 * same ->completed number as the newly arrived callbacks, which
1177 * means that the callbacks in any of these later sublist can be
1178 * grouped into a single sublist, whether or not they have already
1179 * been assigned a ->completed number.
1181 c = rcu_cbs_completed(rsp, rnp);
1182 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1183 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1184 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1188 * If there are no sublist for unassigned callbacks, leave.
1189 * At the same time, advance "i" one sublist, so that "i" will
1190 * index into the sublist where all the remaining callbacks should
1193 if (++i >= RCU_NEXT_TAIL)
1197 * Assign all subsequent callbacks' ->completed number to the next
1198 * full grace period and group them all in the sublist initially
1201 for (; i <= RCU_NEXT_TAIL; i++) {
1202 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1203 rdp->nxtcompleted[i] = c;
1205 /* Record any needed additional grace periods. */
1206 rcu_start_future_gp(rnp, rdp);
1208 /* Trace depending on how much we were able to accelerate. */
1209 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1210 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1212 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1216 * Move any callbacks whose grace period has completed to the
1217 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1218 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1219 * sublist. This function is idempotent, so it does not hurt to
1220 * invoke it repeatedly. As long as it is not invoked -too- often...
1222 * The caller must hold rnp->lock with interrupts disabled.
1224 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1225 struct rcu_data *rdp)
1229 /* If the CPU has no callbacks, nothing to do. */
1230 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1234 * Find all callbacks whose ->completed numbers indicate that they
1235 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1237 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1238 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1240 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1242 /* Clean up any sublist tail pointers that were misordered above. */
1243 for (j = RCU_WAIT_TAIL; j < i; j++)
1244 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1246 /* Copy down callbacks to fill in empty sublists. */
1247 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1248 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1250 rdp->nxttail[j] = rdp->nxttail[i];
1251 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1254 /* Classify any remaining callbacks. */
1255 rcu_accelerate_cbs(rsp, rnp, rdp);
1259 * Update CPU-local rcu_data state to record the beginnings and ends of
1260 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1261 * structure corresponding to the current CPU, and must have irqs disabled.
1263 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1265 /* Handle the ends of any preceding grace periods first. */
1266 if (rdp->completed == rnp->completed) {
1268 /* No grace period end, so just accelerate recent callbacks. */
1269 rcu_accelerate_cbs(rsp, rnp, rdp);
1273 /* Advance callbacks. */
1274 rcu_advance_cbs(rsp, rnp, rdp);
1276 /* Remember that we saw this grace-period completion. */
1277 rdp->completed = rnp->completed;
1278 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1281 if (rdp->gpnum != rnp->gpnum) {
1283 * If the current grace period is waiting for this CPU,
1284 * set up to detect a quiescent state, otherwise don't
1285 * go looking for one.
1287 rdp->gpnum = rnp->gpnum;
1288 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1289 rdp->passed_quiesce = 0;
1290 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1291 zero_cpu_stall_ticks(rdp);
1295 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1297 unsigned long flags;
1298 struct rcu_node *rnp;
1300 local_irq_save(flags);
1302 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1303 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1304 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1305 local_irq_restore(flags);
1308 __note_gp_changes(rsp, rnp, rdp);
1309 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1313 * Initialize a new grace period.
1315 static int rcu_gp_init(struct rcu_state *rsp)
1317 struct rcu_data *rdp;
1318 struct rcu_node *rnp = rcu_get_root(rsp);
1320 rcu_bind_gp_kthread();
1321 raw_spin_lock_irq(&rnp->lock);
1322 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1324 if (rcu_gp_in_progress(rsp)) {
1325 /* Grace period already in progress, don't start another. */
1326 raw_spin_unlock_irq(&rnp->lock);
1330 /* Advance to a new grace period and initialize state. */
1332 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1333 record_gp_stall_check_time(rsp);
1334 raw_spin_unlock_irq(&rnp->lock);
1336 /* Exclude any concurrent CPU-hotplug operations. */
1337 mutex_lock(&rsp->onoff_mutex);
1340 * Set the quiescent-state-needed bits in all the rcu_node
1341 * structures for all currently online CPUs in breadth-first order,
1342 * starting from the root rcu_node structure, relying on the layout
1343 * of the tree within the rsp->node[] array. Note that other CPUs
1344 * will access only the leaves of the hierarchy, thus seeing that no
1345 * grace period is in progress, at least until the corresponding
1346 * leaf node has been initialized. In addition, we have excluded
1347 * CPU-hotplug operations.
1349 * The grace period cannot complete until the initialization
1350 * process finishes, because this kthread handles both.
1352 rcu_for_each_node_breadth_first(rsp, rnp) {
1353 raw_spin_lock_irq(&rnp->lock);
1354 rdp = this_cpu_ptr(rsp->rda);
1355 rcu_preempt_check_blocked_tasks(rnp);
1356 rnp->qsmask = rnp->qsmaskinit;
1357 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1358 WARN_ON_ONCE(rnp->completed != rsp->completed);
1359 ACCESS_ONCE(rnp->completed) = rsp->completed;
1360 if (rnp == rdp->mynode)
1361 __note_gp_changes(rsp, rnp, rdp);
1362 rcu_preempt_boost_start_gp(rnp);
1363 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1364 rnp->level, rnp->grplo,
1365 rnp->grphi, rnp->qsmask);
1366 raw_spin_unlock_irq(&rnp->lock);
1367 #ifdef CONFIG_PROVE_RCU_DELAY
1368 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1369 system_state == SYSTEM_RUNNING)
1371 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1375 mutex_unlock(&rsp->onoff_mutex);
1380 * Do one round of quiescent-state forcing.
1382 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1384 int fqs_state = fqs_state_in;
1385 bool isidle = false;
1387 struct rcu_node *rnp = rcu_get_root(rsp);
1390 if (fqs_state == RCU_SAVE_DYNTICK) {
1391 /* Collect dyntick-idle snapshots. */
1392 if (is_sysidle_rcu_state(rsp)) {
1394 maxj = jiffies - ULONG_MAX / 4;
1396 force_qs_rnp(rsp, dyntick_save_progress_counter,
1398 rcu_sysidle_report_gp(rsp, isidle, maxj);
1399 fqs_state = RCU_FORCE_QS;
1401 /* Handle dyntick-idle and offline CPUs. */
1403 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1405 /* Clear flag to prevent immediate re-entry. */
1406 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1407 raw_spin_lock_irq(&rnp->lock);
1408 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1409 raw_spin_unlock_irq(&rnp->lock);
1415 * Clean up after the old grace period.
1417 static void rcu_gp_cleanup(struct rcu_state *rsp)
1419 unsigned long gp_duration;
1421 struct rcu_data *rdp;
1422 struct rcu_node *rnp = rcu_get_root(rsp);
1424 raw_spin_lock_irq(&rnp->lock);
1425 gp_duration = jiffies - rsp->gp_start;
1426 if (gp_duration > rsp->gp_max)
1427 rsp->gp_max = gp_duration;
1430 * We know the grace period is complete, but to everyone else
1431 * it appears to still be ongoing. But it is also the case
1432 * that to everyone else it looks like there is nothing that
1433 * they can do to advance the grace period. It is therefore
1434 * safe for us to drop the lock in order to mark the grace
1435 * period as completed in all of the rcu_node structures.
1437 raw_spin_unlock_irq(&rnp->lock);
1440 * Propagate new ->completed value to rcu_node structures so
1441 * that other CPUs don't have to wait until the start of the next
1442 * grace period to process their callbacks. This also avoids
1443 * some nasty RCU grace-period initialization races by forcing
1444 * the end of the current grace period to be completely recorded in
1445 * all of the rcu_node structures before the beginning of the next
1446 * grace period is recorded in any of the rcu_node structures.
1448 rcu_for_each_node_breadth_first(rsp, rnp) {
1449 raw_spin_lock_irq(&rnp->lock);
1450 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1451 rdp = this_cpu_ptr(rsp->rda);
1452 if (rnp == rdp->mynode)
1453 __note_gp_changes(rsp, rnp, rdp);
1454 nocb += rcu_future_gp_cleanup(rsp, rnp);
1455 raw_spin_unlock_irq(&rnp->lock);
1458 rnp = rcu_get_root(rsp);
1459 raw_spin_lock_irq(&rnp->lock);
1460 rcu_nocb_gp_set(rnp, nocb);
1462 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1463 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1464 rsp->fqs_state = RCU_GP_IDLE;
1465 rdp = this_cpu_ptr(rsp->rda);
1466 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1467 if (cpu_needs_another_gp(rsp, rdp))
1469 raw_spin_unlock_irq(&rnp->lock);
1473 * Body of kthread that handles grace periods.
1475 static int __noreturn rcu_gp_kthread(void *arg)
1480 struct rcu_state *rsp = arg;
1481 struct rcu_node *rnp = rcu_get_root(rsp);
1485 /* Handle grace-period start. */
1487 wait_event_interruptible(rsp->gp_wq,
1490 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1494 flush_signals(current);
1497 /* Handle quiescent-state forcing. */
1498 fqs_state = RCU_SAVE_DYNTICK;
1499 j = jiffies_till_first_fqs;
1502 jiffies_till_first_fqs = HZ;
1505 rsp->jiffies_force_qs = jiffies + j;
1506 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1507 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1508 (!ACCESS_ONCE(rnp->qsmask) &&
1509 !rcu_preempt_blocked_readers_cgp(rnp)),
1511 /* If grace period done, leave loop. */
1512 if (!ACCESS_ONCE(rnp->qsmask) &&
1513 !rcu_preempt_blocked_readers_cgp(rnp))
1515 /* If time for quiescent-state forcing, do it. */
1516 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1517 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1520 /* Deal with stray signal. */
1522 flush_signals(current);
1524 j = jiffies_till_next_fqs;
1527 jiffies_till_next_fqs = HZ;
1530 jiffies_till_next_fqs = 1;
1534 /* Handle grace-period end. */
1535 rcu_gp_cleanup(rsp);
1539 static void rsp_wakeup(struct irq_work *work)
1541 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1543 /* Wake up rcu_gp_kthread() to start the grace period. */
1544 wake_up(&rsp->gp_wq);
1548 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1549 * in preparation for detecting the next grace period. The caller must hold
1550 * the root node's ->lock and hard irqs must be disabled.
1552 * Note that it is legal for a dying CPU (which is marked as offline) to
1553 * invoke this function. This can happen when the dying CPU reports its
1557 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1558 struct rcu_data *rdp)
1560 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1562 * Either we have not yet spawned the grace-period
1563 * task, this CPU does not need another grace period,
1564 * or a grace period is already in progress.
1565 * Either way, don't start a new grace period.
1569 rsp->gp_flags = RCU_GP_FLAG_INIT;
1572 * We can't do wakeups while holding the rnp->lock, as that
1573 * could cause possible deadlocks with the rq->lock. Defer
1574 * the wakeup to interrupt context. And don't bother waking
1575 * up the running kthread.
1577 if (current != rsp->gp_kthread)
1578 irq_work_queue(&rsp->wakeup_work);
1582 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1583 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1584 * is invoked indirectly from rcu_advance_cbs(), which would result in
1585 * endless recursion -- or would do so if it wasn't for the self-deadlock
1586 * that is encountered beforehand.
1589 rcu_start_gp(struct rcu_state *rsp)
1591 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1592 struct rcu_node *rnp = rcu_get_root(rsp);
1595 * If there is no grace period in progress right now, any
1596 * callbacks we have up to this point will be satisfied by the
1597 * next grace period. Also, advancing the callbacks reduces the
1598 * probability of false positives from cpu_needs_another_gp()
1599 * resulting in pointless grace periods. So, advance callbacks
1600 * then start the grace period!
1602 rcu_advance_cbs(rsp, rnp, rdp);
1603 rcu_start_gp_advanced(rsp, rnp, rdp);
1607 * Report a full set of quiescent states to the specified rcu_state
1608 * data structure. This involves cleaning up after the prior grace
1609 * period and letting rcu_start_gp() start up the next grace period
1610 * if one is needed. Note that the caller must hold rnp->lock, which
1611 * is released before return.
1613 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1614 __releases(rcu_get_root(rsp)->lock)
1616 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1617 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1618 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1622 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1623 * Allows quiescent states for a group of CPUs to be reported at one go
1624 * to the specified rcu_node structure, though all the CPUs in the group
1625 * must be represented by the same rcu_node structure (which need not be
1626 * a leaf rcu_node structure, though it often will be). That structure's
1627 * lock must be held upon entry, and it is released before return.
1630 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1631 struct rcu_node *rnp, unsigned long flags)
1632 __releases(rnp->lock)
1634 struct rcu_node *rnp_c;
1636 /* Walk up the rcu_node hierarchy. */
1638 if (!(rnp->qsmask & mask)) {
1640 /* Our bit has already been cleared, so done. */
1641 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1644 rnp->qsmask &= ~mask;
1645 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1646 mask, rnp->qsmask, rnp->level,
1647 rnp->grplo, rnp->grphi,
1649 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1651 /* Other bits still set at this level, so done. */
1652 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1655 mask = rnp->grpmask;
1656 if (rnp->parent == NULL) {
1658 /* No more levels. Exit loop holding root lock. */
1662 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1665 raw_spin_lock_irqsave(&rnp->lock, flags);
1666 WARN_ON_ONCE(rnp_c->qsmask);
1670 * Get here if we are the last CPU to pass through a quiescent
1671 * state for this grace period. Invoke rcu_report_qs_rsp()
1672 * to clean up and start the next grace period if one is needed.
1674 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1678 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1679 * structure. This must be either called from the specified CPU, or
1680 * called when the specified CPU is known to be offline (and when it is
1681 * also known that no other CPU is concurrently trying to help the offline
1682 * CPU). The lastcomp argument is used to make sure we are still in the
1683 * grace period of interest. We don't want to end the current grace period
1684 * based on quiescent states detected in an earlier grace period!
1687 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1689 unsigned long flags;
1691 struct rcu_node *rnp;
1694 raw_spin_lock_irqsave(&rnp->lock, flags);
1695 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1696 rnp->completed == rnp->gpnum) {
1699 * The grace period in which this quiescent state was
1700 * recorded has ended, so don't report it upwards.
1701 * We will instead need a new quiescent state that lies
1702 * within the current grace period.
1704 rdp->passed_quiesce = 0; /* need qs for new gp. */
1705 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1708 mask = rdp->grpmask;
1709 if ((rnp->qsmask & mask) == 0) {
1710 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1712 rdp->qs_pending = 0;
1715 * This GP can't end until cpu checks in, so all of our
1716 * callbacks can be processed during the next GP.
1718 rcu_accelerate_cbs(rsp, rnp, rdp);
1720 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1725 * Check to see if there is a new grace period of which this CPU
1726 * is not yet aware, and if so, set up local rcu_data state for it.
1727 * Otherwise, see if this CPU has just passed through its first
1728 * quiescent state for this grace period, and record that fact if so.
1731 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1733 /* Check for grace-period ends and beginnings. */
1734 note_gp_changes(rsp, rdp);
1737 * Does this CPU still need to do its part for current grace period?
1738 * If no, return and let the other CPUs do their part as well.
1740 if (!rdp->qs_pending)
1744 * Was there a quiescent state since the beginning of the grace
1745 * period? If no, then exit and wait for the next call.
1747 if (!rdp->passed_quiesce)
1751 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1754 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1757 #ifdef CONFIG_HOTPLUG_CPU
1760 * Send the specified CPU's RCU callbacks to the orphanage. The
1761 * specified CPU must be offline, and the caller must hold the
1765 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1766 struct rcu_node *rnp, struct rcu_data *rdp)
1768 /* No-CBs CPUs do not have orphanable callbacks. */
1769 if (rcu_is_nocb_cpu(rdp->cpu))
1773 * Orphan the callbacks. First adjust the counts. This is safe
1774 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1775 * cannot be running now. Thus no memory barrier is required.
1777 if (rdp->nxtlist != NULL) {
1778 rsp->qlen_lazy += rdp->qlen_lazy;
1779 rsp->qlen += rdp->qlen;
1780 rdp->n_cbs_orphaned += rdp->qlen;
1782 ACCESS_ONCE(rdp->qlen) = 0;
1786 * Next, move those callbacks still needing a grace period to
1787 * the orphanage, where some other CPU will pick them up.
1788 * Some of the callbacks might have gone partway through a grace
1789 * period, but that is too bad. They get to start over because we
1790 * cannot assume that grace periods are synchronized across CPUs.
1791 * We don't bother updating the ->nxttail[] array yet, instead
1792 * we just reset the whole thing later on.
1794 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1795 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1796 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1797 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1801 * Then move the ready-to-invoke callbacks to the orphanage,
1802 * where some other CPU will pick them up. These will not be
1803 * required to pass though another grace period: They are done.
1805 if (rdp->nxtlist != NULL) {
1806 *rsp->orphan_donetail = rdp->nxtlist;
1807 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1810 /* Finally, initialize the rcu_data structure's list to empty. */
1811 init_callback_list(rdp);
1815 * Adopt the RCU callbacks from the specified rcu_state structure's
1816 * orphanage. The caller must hold the ->orphan_lock.
1818 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1821 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1823 /* No-CBs CPUs are handled specially. */
1824 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1827 /* Do the accounting first. */
1828 rdp->qlen_lazy += rsp->qlen_lazy;
1829 rdp->qlen += rsp->qlen;
1830 rdp->n_cbs_adopted += rsp->qlen;
1831 if (rsp->qlen_lazy != rsp->qlen)
1832 rcu_idle_count_callbacks_posted();
1837 * We do not need a memory barrier here because the only way we
1838 * can get here if there is an rcu_barrier() in flight is if
1839 * we are the task doing the rcu_barrier().
1842 /* First adopt the ready-to-invoke callbacks. */
1843 if (rsp->orphan_donelist != NULL) {
1844 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1845 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1846 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1847 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1848 rdp->nxttail[i] = rsp->orphan_donetail;
1849 rsp->orphan_donelist = NULL;
1850 rsp->orphan_donetail = &rsp->orphan_donelist;
1853 /* And then adopt the callbacks that still need a grace period. */
1854 if (rsp->orphan_nxtlist != NULL) {
1855 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1856 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1857 rsp->orphan_nxtlist = NULL;
1858 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1863 * Trace the fact that this CPU is going offline.
1865 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1867 RCU_TRACE(unsigned long mask);
1868 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1869 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1871 RCU_TRACE(mask = rdp->grpmask);
1872 trace_rcu_grace_period(rsp->name,
1873 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1878 * The CPU has been completely removed, and some other CPU is reporting
1879 * this fact from process context. Do the remainder of the cleanup,
1880 * including orphaning the outgoing CPU's RCU callbacks, and also
1881 * adopting them. There can only be one CPU hotplug operation at a time,
1882 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1884 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1886 unsigned long flags;
1888 int need_report = 0;
1889 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1890 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1892 /* Adjust any no-longer-needed kthreads. */
1893 rcu_boost_kthread_setaffinity(rnp, -1);
1895 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1897 /* Exclude any attempts to start a new grace period. */
1898 mutex_lock(&rsp->onoff_mutex);
1899 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1901 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1902 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1903 rcu_adopt_orphan_cbs(rsp);
1905 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1906 mask = rdp->grpmask; /* rnp->grplo is constant. */
1908 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1909 rnp->qsmaskinit &= ~mask;
1910 if (rnp->qsmaskinit != 0) {
1911 if (rnp != rdp->mynode)
1912 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1915 if (rnp == rdp->mynode)
1916 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1918 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1919 mask = rnp->grpmask;
1921 } while (rnp != NULL);
1924 * We still hold the leaf rcu_node structure lock here, and
1925 * irqs are still disabled. The reason for this subterfuge is
1926 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1927 * held leads to deadlock.
1929 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1931 if (need_report & RCU_OFL_TASKS_NORM_GP)
1932 rcu_report_unblock_qs_rnp(rnp, flags);
1934 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1935 if (need_report & RCU_OFL_TASKS_EXP_GP)
1936 rcu_report_exp_rnp(rsp, rnp, true);
1937 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1938 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1939 cpu, rdp->qlen, rdp->nxtlist);
1940 init_callback_list(rdp);
1941 /* Disallow further callbacks on this CPU. */
1942 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1943 mutex_unlock(&rsp->onoff_mutex);
1946 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1948 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1952 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1956 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1959 * Invoke any RCU callbacks that have made it to the end of their grace
1960 * period. Thottle as specified by rdp->blimit.
1962 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1964 unsigned long flags;
1965 struct rcu_head *next, *list, **tail;
1966 long bl, count, count_lazy;
1969 /* If no callbacks are ready, just return. */
1970 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1971 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1972 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1973 need_resched(), is_idle_task(current),
1974 rcu_is_callbacks_kthread());
1979 * Extract the list of ready callbacks, disabling to prevent
1980 * races with call_rcu() from interrupt handlers.
1982 local_irq_save(flags);
1983 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1985 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1986 list = rdp->nxtlist;
1987 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1988 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1989 tail = rdp->nxttail[RCU_DONE_TAIL];
1990 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1991 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1992 rdp->nxttail[i] = &rdp->nxtlist;
1993 local_irq_restore(flags);
1995 /* Invoke callbacks. */
1996 count = count_lazy = 0;
2000 debug_rcu_head_unqueue(list);
2001 if (__rcu_reclaim(rsp->name, list))
2004 /* Stop only if limit reached and CPU has something to do. */
2005 if (++count >= bl &&
2007 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2011 local_irq_save(flags);
2012 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2013 is_idle_task(current),
2014 rcu_is_callbacks_kthread());
2016 /* Update count, and requeue any remaining callbacks. */
2018 *tail = rdp->nxtlist;
2019 rdp->nxtlist = list;
2020 for (i = 0; i < RCU_NEXT_SIZE; i++)
2021 if (&rdp->nxtlist == rdp->nxttail[i])
2022 rdp->nxttail[i] = tail;
2026 smp_mb(); /* List handling before counting for rcu_barrier(). */
2027 rdp->qlen_lazy -= count_lazy;
2028 ACCESS_ONCE(rdp->qlen) -= count;
2029 rdp->n_cbs_invoked += count;
2031 /* Reinstate batch limit if we have worked down the excess. */
2032 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2033 rdp->blimit = blimit;
2035 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2036 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2037 rdp->qlen_last_fqs_check = 0;
2038 rdp->n_force_qs_snap = rsp->n_force_qs;
2039 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2040 rdp->qlen_last_fqs_check = rdp->qlen;
2041 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2043 local_irq_restore(flags);
2045 /* Re-invoke RCU core processing if there are callbacks remaining. */
2046 if (cpu_has_callbacks_ready_to_invoke(rdp))
2051 * Check to see if this CPU is in a non-context-switch quiescent state
2052 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2053 * Also schedule RCU core processing.
2055 * This function must be called from hardirq context. It is normally
2056 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2057 * false, there is no point in invoking rcu_check_callbacks().
2059 void rcu_check_callbacks(int cpu, int user)
2061 trace_rcu_utilization(TPS("Start scheduler-tick"));
2062 increment_cpu_stall_ticks();
2063 if (user || rcu_is_cpu_rrupt_from_idle()) {
2066 * Get here if this CPU took its interrupt from user
2067 * mode or from the idle loop, and if this is not a
2068 * nested interrupt. In this case, the CPU is in
2069 * a quiescent state, so note it.
2071 * No memory barrier is required here because both
2072 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2073 * variables that other CPUs neither access nor modify,
2074 * at least not while the corresponding CPU is online.
2080 } else if (!in_softirq()) {
2083 * Get here if this CPU did not take its interrupt from
2084 * softirq, in other words, if it is not interrupting
2085 * a rcu_bh read-side critical section. This is an _bh
2086 * critical section, so note it.
2091 rcu_preempt_check_callbacks(cpu);
2092 if (rcu_pending(cpu))
2094 trace_rcu_utilization(TPS("End scheduler-tick"));
2098 * Scan the leaf rcu_node structures, processing dyntick state for any that
2099 * have not yet encountered a quiescent state, using the function specified.
2100 * Also initiate boosting for any threads blocked on the root rcu_node.
2102 * The caller must have suppressed start of new grace periods.
2104 static void force_qs_rnp(struct rcu_state *rsp,
2105 int (*f)(struct rcu_data *rsp, bool *isidle,
2106 unsigned long *maxj),
2107 bool *isidle, unsigned long *maxj)
2111 unsigned long flags;
2113 struct rcu_node *rnp;
2115 rcu_for_each_leaf_node(rsp, rnp) {
2118 raw_spin_lock_irqsave(&rnp->lock, flags);
2119 if (!rcu_gp_in_progress(rsp)) {
2120 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2123 if (rnp->qsmask == 0) {
2124 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2129 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2130 if ((rnp->qsmask & bit) != 0) {
2131 if ((rnp->qsmaskinit & bit) != 0)
2133 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2139 /* rcu_report_qs_rnp() releases rnp->lock. */
2140 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2143 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2145 rnp = rcu_get_root(rsp);
2146 if (rnp->qsmask == 0) {
2147 raw_spin_lock_irqsave(&rnp->lock, flags);
2148 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2153 * Force quiescent states on reluctant CPUs, and also detect which
2154 * CPUs are in dyntick-idle mode.
2156 static void force_quiescent_state(struct rcu_state *rsp)
2158 unsigned long flags;
2160 struct rcu_node *rnp;
2161 struct rcu_node *rnp_old = NULL;
2163 /* Funnel through hierarchy to reduce memory contention. */
2164 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2165 for (; rnp != NULL; rnp = rnp->parent) {
2166 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2167 !raw_spin_trylock(&rnp->fqslock);
2168 if (rnp_old != NULL)
2169 raw_spin_unlock(&rnp_old->fqslock);
2171 rsp->n_force_qs_lh++;
2176 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2178 /* Reached the root of the rcu_node tree, acquire lock. */
2179 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2180 raw_spin_unlock(&rnp_old->fqslock);
2181 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2182 rsp->n_force_qs_lh++;
2183 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2184 return; /* Someone beat us to it. */
2186 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2187 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2188 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2192 * This does the RCU core processing work for the specified rcu_state
2193 * and rcu_data structures. This may be called only from the CPU to
2194 * whom the rdp belongs.
2197 __rcu_process_callbacks(struct rcu_state *rsp)
2199 unsigned long flags;
2200 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2202 WARN_ON_ONCE(rdp->beenonline == 0);
2204 /* Update RCU state based on any recent quiescent states. */
2205 rcu_check_quiescent_state(rsp, rdp);
2207 /* Does this CPU require a not-yet-started grace period? */
2208 local_irq_save(flags);
2209 if (cpu_needs_another_gp(rsp, rdp)) {
2210 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2212 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2214 local_irq_restore(flags);
2217 /* If there are callbacks ready, invoke them. */
2218 if (cpu_has_callbacks_ready_to_invoke(rdp))
2219 invoke_rcu_callbacks(rsp, rdp);
2223 * Do RCU core processing for the current CPU.
2225 static void rcu_process_callbacks(struct softirq_action *unused)
2227 struct rcu_state *rsp;
2229 if (cpu_is_offline(smp_processor_id()))
2231 trace_rcu_utilization(TPS("Start RCU core"));
2232 for_each_rcu_flavor(rsp)
2233 __rcu_process_callbacks(rsp);
2234 trace_rcu_utilization(TPS("End RCU core"));
2238 * Schedule RCU callback invocation. If the specified type of RCU
2239 * does not support RCU priority boosting, just do a direct call,
2240 * otherwise wake up the per-CPU kernel kthread. Note that because we
2241 * are running on the current CPU with interrupts disabled, the
2242 * rcu_cpu_kthread_task cannot disappear out from under us.
2244 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2246 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2248 if (likely(!rsp->boost)) {
2249 rcu_do_batch(rsp, rdp);
2252 invoke_rcu_callbacks_kthread();
2255 static void invoke_rcu_core(void)
2257 if (cpu_online(smp_processor_id()))
2258 raise_softirq(RCU_SOFTIRQ);
2262 * Handle any core-RCU processing required by a call_rcu() invocation.
2264 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2265 struct rcu_head *head, unsigned long flags)
2268 * If called from an extended quiescent state, invoke the RCU
2269 * core in order to force a re-evaluation of RCU's idleness.
2271 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2274 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2275 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2279 * Force the grace period if too many callbacks or too long waiting.
2280 * Enforce hysteresis, and don't invoke force_quiescent_state()
2281 * if some other CPU has recently done so. Also, don't bother
2282 * invoking force_quiescent_state() if the newly enqueued callback
2283 * is the only one waiting for a grace period to complete.
2285 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2287 /* Are we ignoring a completed grace period? */
2288 note_gp_changes(rsp, rdp);
2290 /* Start a new grace period if one not already started. */
2291 if (!rcu_gp_in_progress(rsp)) {
2292 struct rcu_node *rnp_root = rcu_get_root(rsp);
2294 raw_spin_lock(&rnp_root->lock);
2296 raw_spin_unlock(&rnp_root->lock);
2298 /* Give the grace period a kick. */
2299 rdp->blimit = LONG_MAX;
2300 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2301 *rdp->nxttail[RCU_DONE_TAIL] != head)
2302 force_quiescent_state(rsp);
2303 rdp->n_force_qs_snap = rsp->n_force_qs;
2304 rdp->qlen_last_fqs_check = rdp->qlen;
2310 * RCU callback function to leak a callback.
2312 static void rcu_leak_callback(struct rcu_head *rhp)
2317 * Helper function for call_rcu() and friends. The cpu argument will
2318 * normally be -1, indicating "currently running CPU". It may specify
2319 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2320 * is expected to specify a CPU.
2323 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2324 struct rcu_state *rsp, int cpu, bool lazy)
2326 unsigned long flags;
2327 struct rcu_data *rdp;
2329 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2330 if (debug_rcu_head_queue(head)) {
2331 /* Probable double call_rcu(), so leak the callback. */
2332 ACCESS_ONCE(head->func) = rcu_leak_callback;
2333 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2340 * Opportunistically note grace-period endings and beginnings.
2341 * Note that we might see a beginning right after we see an
2342 * end, but never vice versa, since this CPU has to pass through
2343 * a quiescent state betweentimes.
2345 local_irq_save(flags);
2346 rdp = this_cpu_ptr(rsp->rda);
2348 /* Add the callback to our list. */
2349 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2353 rdp = per_cpu_ptr(rsp->rda, cpu);
2354 offline = !__call_rcu_nocb(rdp, head, lazy);
2355 WARN_ON_ONCE(offline);
2356 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2357 local_irq_restore(flags);
2360 ACCESS_ONCE(rdp->qlen)++;
2364 rcu_idle_count_callbacks_posted();
2365 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2366 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2367 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2369 if (__is_kfree_rcu_offset((unsigned long)func))
2370 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2371 rdp->qlen_lazy, rdp->qlen);
2373 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2375 /* Go handle any RCU core processing required. */
2376 __call_rcu_core(rsp, rdp, head, flags);
2377 local_irq_restore(flags);
2381 * Queue an RCU-sched callback for invocation after a grace period.
2383 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2385 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2387 EXPORT_SYMBOL_GPL(call_rcu_sched);
2390 * Queue an RCU callback for invocation after a quicker grace period.
2392 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2394 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2396 EXPORT_SYMBOL_GPL(call_rcu_bh);
2399 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2400 * any blocking grace-period wait automatically implies a grace period
2401 * if there is only one CPU online at any point time during execution
2402 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2403 * occasionally incorrectly indicate that there are multiple CPUs online
2404 * when there was in fact only one the whole time, as this just adds
2405 * some overhead: RCU still operates correctly.
2407 static inline int rcu_blocking_is_gp(void)
2411 might_sleep(); /* Check for RCU read-side critical section. */
2413 ret = num_online_cpus() <= 1;
2419 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2421 * Control will return to the caller some time after a full rcu-sched
2422 * grace period has elapsed, in other words after all currently executing
2423 * rcu-sched read-side critical sections have completed. These read-side
2424 * critical sections are delimited by rcu_read_lock_sched() and
2425 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2426 * local_irq_disable(), and so on may be used in place of
2427 * rcu_read_lock_sched().
2429 * This means that all preempt_disable code sequences, including NMI and
2430 * non-threaded hardware-interrupt handlers, in progress on entry will
2431 * have completed before this primitive returns. However, this does not
2432 * guarantee that softirq handlers will have completed, since in some
2433 * kernels, these handlers can run in process context, and can block.
2435 * Note that this guarantee implies further memory-ordering guarantees.
2436 * On systems with more than one CPU, when synchronize_sched() returns,
2437 * each CPU is guaranteed to have executed a full memory barrier since the
2438 * end of its last RCU-sched read-side critical section whose beginning
2439 * preceded the call to synchronize_sched(). In addition, each CPU having
2440 * an RCU read-side critical section that extends beyond the return from
2441 * synchronize_sched() is guaranteed to have executed a full memory barrier
2442 * after the beginning of synchronize_sched() and before the beginning of
2443 * that RCU read-side critical section. Note that these guarantees include
2444 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2445 * that are executing in the kernel.
2447 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2448 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2449 * to have executed a full memory barrier during the execution of
2450 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2451 * again only if the system has more than one CPU).
2453 * This primitive provides the guarantees made by the (now removed)
2454 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2455 * guarantees that rcu_read_lock() sections will have completed.
2456 * In "classic RCU", these two guarantees happen to be one and
2457 * the same, but can differ in realtime RCU implementations.
2459 void synchronize_sched(void)
2461 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2462 !lock_is_held(&rcu_lock_map) &&
2463 !lock_is_held(&rcu_sched_lock_map),
2464 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2465 if (rcu_blocking_is_gp())
2468 synchronize_sched_expedited();
2470 wait_rcu_gp(call_rcu_sched);
2472 EXPORT_SYMBOL_GPL(synchronize_sched);
2475 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2477 * Control will return to the caller some time after a full rcu_bh grace
2478 * period has elapsed, in other words after all currently executing rcu_bh
2479 * read-side critical sections have completed. RCU read-side critical
2480 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2481 * and may be nested.
2483 * See the description of synchronize_sched() for more detailed information
2484 * on memory ordering guarantees.
2486 void synchronize_rcu_bh(void)
2488 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2489 !lock_is_held(&rcu_lock_map) &&
2490 !lock_is_held(&rcu_sched_lock_map),
2491 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2492 if (rcu_blocking_is_gp())
2495 synchronize_rcu_bh_expedited();
2497 wait_rcu_gp(call_rcu_bh);
2499 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2501 static int synchronize_sched_expedited_cpu_stop(void *data)
2504 * There must be a full memory barrier on each affected CPU
2505 * between the time that try_stop_cpus() is called and the
2506 * time that it returns.
2508 * In the current initial implementation of cpu_stop, the
2509 * above condition is already met when the control reaches
2510 * this point and the following smp_mb() is not strictly
2511 * necessary. Do smp_mb() anyway for documentation and
2512 * robustness against future implementation changes.
2514 smp_mb(); /* See above comment block. */
2519 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2521 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2522 * approach to force the grace period to end quickly. This consumes
2523 * significant time on all CPUs and is unfriendly to real-time workloads,
2524 * so is thus not recommended for any sort of common-case code. In fact,
2525 * if you are using synchronize_sched_expedited() in a loop, please
2526 * restructure your code to batch your updates, and then use a single
2527 * synchronize_sched() instead.
2529 * Note that it is illegal to call this function while holding any lock
2530 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2531 * to call this function from a CPU-hotplug notifier. Failing to observe
2532 * these restriction will result in deadlock.
2534 * This implementation can be thought of as an application of ticket
2535 * locking to RCU, with sync_sched_expedited_started and
2536 * sync_sched_expedited_done taking on the roles of the halves
2537 * of the ticket-lock word. Each task atomically increments
2538 * sync_sched_expedited_started upon entry, snapshotting the old value,
2539 * then attempts to stop all the CPUs. If this succeeds, then each
2540 * CPU will have executed a context switch, resulting in an RCU-sched
2541 * grace period. We are then done, so we use atomic_cmpxchg() to
2542 * update sync_sched_expedited_done to match our snapshot -- but
2543 * only if someone else has not already advanced past our snapshot.
2545 * On the other hand, if try_stop_cpus() fails, we check the value
2546 * of sync_sched_expedited_done. If it has advanced past our
2547 * initial snapshot, then someone else must have forced a grace period
2548 * some time after we took our snapshot. In this case, our work is
2549 * done for us, and we can simply return. Otherwise, we try again,
2550 * but keep our initial snapshot for purposes of checking for someone
2551 * doing our work for us.
2553 * If we fail too many times in a row, we fall back to synchronize_sched().
2555 void synchronize_sched_expedited(void)
2557 long firstsnap, s, snap;
2559 struct rcu_state *rsp = &rcu_sched_state;
2562 * If we are in danger of counter wrap, just do synchronize_sched().
2563 * By allowing sync_sched_expedited_started to advance no more than
2564 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2565 * that more than 3.5 billion CPUs would be required to force a
2566 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2567 * course be required on a 64-bit system.
2569 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2570 (ulong)atomic_long_read(&rsp->expedited_done) +
2572 synchronize_sched();
2573 atomic_long_inc(&rsp->expedited_wrap);
2578 * Take a ticket. Note that atomic_inc_return() implies a
2579 * full memory barrier.
2581 snap = atomic_long_inc_return(&rsp->expedited_start);
2584 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2587 * Each pass through the following loop attempts to force a
2588 * context switch on each CPU.
2590 while (try_stop_cpus(cpu_online_mask,
2591 synchronize_sched_expedited_cpu_stop,
2594 atomic_long_inc(&rsp->expedited_tryfail);
2596 /* Check to see if someone else did our work for us. */
2597 s = atomic_long_read(&rsp->expedited_done);
2598 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2599 /* ensure test happens before caller kfree */
2600 smp_mb__before_atomic_inc(); /* ^^^ */
2601 atomic_long_inc(&rsp->expedited_workdone1);
2605 /* No joy, try again later. Or just synchronize_sched(). */
2606 if (trycount++ < 10) {
2607 udelay(trycount * num_online_cpus());
2609 wait_rcu_gp(call_rcu_sched);
2610 atomic_long_inc(&rsp->expedited_normal);
2614 /* Recheck to see if someone else did our work for us. */
2615 s = atomic_long_read(&rsp->expedited_done);
2616 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2617 /* ensure test happens before caller kfree */
2618 smp_mb__before_atomic_inc(); /* ^^^ */
2619 atomic_long_inc(&rsp->expedited_workdone2);
2624 * Refetching sync_sched_expedited_started allows later
2625 * callers to piggyback on our grace period. We retry
2626 * after they started, so our grace period works for them,
2627 * and they started after our first try, so their grace
2628 * period works for us.
2631 snap = atomic_long_read(&rsp->expedited_start);
2632 smp_mb(); /* ensure read is before try_stop_cpus(). */
2634 atomic_long_inc(&rsp->expedited_stoppedcpus);
2637 * Everyone up to our most recent fetch is covered by our grace
2638 * period. Update the counter, but only if our work is still
2639 * relevant -- which it won't be if someone who started later
2640 * than we did already did their update.
2643 atomic_long_inc(&rsp->expedited_done_tries);
2644 s = atomic_long_read(&rsp->expedited_done);
2645 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2646 /* ensure test happens before caller kfree */
2647 smp_mb__before_atomic_inc(); /* ^^^ */
2648 atomic_long_inc(&rsp->expedited_done_lost);
2651 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2652 atomic_long_inc(&rsp->expedited_done_exit);
2656 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2659 * Check to see if there is any immediate RCU-related work to be done
2660 * by the current CPU, for the specified type of RCU, returning 1 if so.
2661 * The checks are in order of increasing expense: checks that can be
2662 * carried out against CPU-local state are performed first. However,
2663 * we must check for CPU stalls first, else we might not get a chance.
2665 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2667 struct rcu_node *rnp = rdp->mynode;
2669 rdp->n_rcu_pending++;
2671 /* Check for CPU stalls, if enabled. */
2672 check_cpu_stall(rsp, rdp);
2674 /* Is the RCU core waiting for a quiescent state from this CPU? */
2675 if (rcu_scheduler_fully_active &&
2676 rdp->qs_pending && !rdp->passed_quiesce) {
2677 rdp->n_rp_qs_pending++;
2678 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2679 rdp->n_rp_report_qs++;
2683 /* Does this CPU have callbacks ready to invoke? */
2684 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2685 rdp->n_rp_cb_ready++;
2689 /* Has RCU gone idle with this CPU needing another grace period? */
2690 if (cpu_needs_another_gp(rsp, rdp)) {
2691 rdp->n_rp_cpu_needs_gp++;
2695 /* Has another RCU grace period completed? */
2696 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2697 rdp->n_rp_gp_completed++;
2701 /* Has a new RCU grace period started? */
2702 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2703 rdp->n_rp_gp_started++;
2708 rdp->n_rp_need_nothing++;
2713 * Check to see if there is any immediate RCU-related work to be done
2714 * by the current CPU, returning 1 if so. This function is part of the
2715 * RCU implementation; it is -not- an exported member of the RCU API.
2717 static int rcu_pending(int cpu)
2719 struct rcu_state *rsp;
2721 for_each_rcu_flavor(rsp)
2722 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2728 * Return true if the specified CPU has any callback. If all_lazy is
2729 * non-NULL, store an indication of whether all callbacks are lazy.
2730 * (If there are no callbacks, all of them are deemed to be lazy.)
2732 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2736 struct rcu_data *rdp;
2737 struct rcu_state *rsp;
2739 for_each_rcu_flavor(rsp) {
2740 rdp = per_cpu_ptr(rsp->rda, cpu);
2741 if (rdp->qlen != rdp->qlen_lazy)
2752 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2753 * the compiler is expected to optimize this away.
2755 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2756 int cpu, unsigned long done)
2758 trace_rcu_barrier(rsp->name, s, cpu,
2759 atomic_read(&rsp->barrier_cpu_count), done);
2763 * RCU callback function for _rcu_barrier(). If we are last, wake
2764 * up the task executing _rcu_barrier().
2766 static void rcu_barrier_callback(struct rcu_head *rhp)
2768 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2769 struct rcu_state *rsp = rdp->rsp;
2771 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2772 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2773 complete(&rsp->barrier_completion);
2775 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2780 * Called with preemption disabled, and from cross-cpu IRQ context.
2782 static void rcu_barrier_func(void *type)
2784 struct rcu_state *rsp = type;
2785 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2787 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2788 atomic_inc(&rsp->barrier_cpu_count);
2789 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2793 * Orchestrate the specified type of RCU barrier, waiting for all
2794 * RCU callbacks of the specified type to complete.
2796 static void _rcu_barrier(struct rcu_state *rsp)
2799 struct rcu_data *rdp;
2800 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2801 unsigned long snap_done;
2803 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2805 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2806 mutex_lock(&rsp->barrier_mutex);
2809 * Ensure that all prior references, including to ->n_barrier_done,
2810 * are ordered before the _rcu_barrier() machinery.
2812 smp_mb(); /* See above block comment. */
2815 * Recheck ->n_barrier_done to see if others did our work for us.
2816 * This means checking ->n_barrier_done for an even-to-odd-to-even
2817 * transition. The "if" expression below therefore rounds the old
2818 * value up to the next even number and adds two before comparing.
2820 snap_done = rsp->n_barrier_done;
2821 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2824 * If the value in snap is odd, we needed to wait for the current
2825 * rcu_barrier() to complete, then wait for the next one, in other
2826 * words, we need the value of snap_done to be three larger than
2827 * the value of snap. On the other hand, if the value in snap is
2828 * even, we only had to wait for the next rcu_barrier() to complete,
2829 * in other words, we need the value of snap_done to be only two
2830 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2831 * this for us (thank you, Linus!).
2833 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
2834 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2835 smp_mb(); /* caller's subsequent code after above check. */
2836 mutex_unlock(&rsp->barrier_mutex);
2841 * Increment ->n_barrier_done to avoid duplicate work. Use
2842 * ACCESS_ONCE() to prevent the compiler from speculating
2843 * the increment to precede the early-exit check.
2845 ACCESS_ONCE(rsp->n_barrier_done)++;
2846 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2847 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2848 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2851 * Initialize the count to one rather than to zero in order to
2852 * avoid a too-soon return to zero in case of a short grace period
2853 * (or preemption of this task). Exclude CPU-hotplug operations
2854 * to ensure that no offline CPU has callbacks queued.
2856 init_completion(&rsp->barrier_completion);
2857 atomic_set(&rsp->barrier_cpu_count, 1);
2861 * Force each CPU with callbacks to register a new callback.
2862 * When that callback is invoked, we will know that all of the
2863 * corresponding CPU's preceding callbacks have been invoked.
2865 for_each_possible_cpu(cpu) {
2866 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2868 rdp = per_cpu_ptr(rsp->rda, cpu);
2869 if (rcu_is_nocb_cpu(cpu)) {
2870 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2871 rsp->n_barrier_done);
2872 atomic_inc(&rsp->barrier_cpu_count);
2873 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2875 } else if (ACCESS_ONCE(rdp->qlen)) {
2876 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2877 rsp->n_barrier_done);
2878 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2880 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2881 rsp->n_barrier_done);
2887 * Now that we have an rcu_barrier_callback() callback on each
2888 * CPU, and thus each counted, remove the initial count.
2890 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2891 complete(&rsp->barrier_completion);
2893 /* Increment ->n_barrier_done to prevent duplicate work. */
2894 smp_mb(); /* Keep increment after above mechanism. */
2895 ACCESS_ONCE(rsp->n_barrier_done)++;
2896 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2897 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2898 smp_mb(); /* Keep increment before caller's subsequent code. */
2900 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2901 wait_for_completion(&rsp->barrier_completion);
2903 /* Other rcu_barrier() invocations can now safely proceed. */
2904 mutex_unlock(&rsp->barrier_mutex);
2908 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2910 void rcu_barrier_bh(void)
2912 _rcu_barrier(&rcu_bh_state);
2914 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2917 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2919 void rcu_barrier_sched(void)
2921 _rcu_barrier(&rcu_sched_state);
2923 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2926 * Do boot-time initialization of a CPU's per-CPU RCU data.
2929 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2931 unsigned long flags;
2932 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2933 struct rcu_node *rnp = rcu_get_root(rsp);
2935 /* Set up local state, ensuring consistent view of global state. */
2936 raw_spin_lock_irqsave(&rnp->lock, flags);
2937 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2938 init_callback_list(rdp);
2940 ACCESS_ONCE(rdp->qlen) = 0;
2941 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2942 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2943 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2946 rcu_boot_init_nocb_percpu_data(rdp);
2947 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2951 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2952 * offline event can be happening at a given time. Note also that we
2953 * can accept some slop in the rsp->completed access due to the fact
2954 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2957 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2959 unsigned long flags;
2961 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2962 struct rcu_node *rnp = rcu_get_root(rsp);
2964 /* Exclude new grace periods. */
2965 mutex_lock(&rsp->onoff_mutex);
2967 /* Set up local state, ensuring consistent view of global state. */
2968 raw_spin_lock_irqsave(&rnp->lock, flags);
2969 rdp->beenonline = 1; /* We have now been online. */
2970 rdp->preemptible = preemptible;
2971 rdp->qlen_last_fqs_check = 0;
2972 rdp->n_force_qs_snap = rsp->n_force_qs;
2973 rdp->blimit = blimit;
2974 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2975 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2976 rcu_sysidle_init_percpu_data(rdp->dynticks);
2977 atomic_set(&rdp->dynticks->dynticks,
2978 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2979 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2981 /* Add CPU to rcu_node bitmasks. */
2983 mask = rdp->grpmask;
2985 /* Exclude any attempts to start a new GP on small systems. */
2986 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2987 rnp->qsmaskinit |= mask;
2988 mask = rnp->grpmask;
2989 if (rnp == rdp->mynode) {
2991 * If there is a grace period in progress, we will
2992 * set up to wait for it next time we run the
2995 rdp->gpnum = rnp->completed;
2996 rdp->completed = rnp->completed;
2997 rdp->passed_quiesce = 0;
2998 rdp->qs_pending = 0;
2999 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3001 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3003 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3004 local_irq_restore(flags);
3006 mutex_unlock(&rsp->onoff_mutex);
3009 static void rcu_prepare_cpu(int cpu)
3011 struct rcu_state *rsp;
3013 for_each_rcu_flavor(rsp)
3014 rcu_init_percpu_data(cpu, rsp,
3015 strcmp(rsp->name, "rcu_preempt") == 0);
3019 * Handle CPU online/offline notification events.
3021 static int rcu_cpu_notify(struct notifier_block *self,
3022 unsigned long action, void *hcpu)
3024 long cpu = (long)hcpu;
3025 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3026 struct rcu_node *rnp = rdp->mynode;
3027 struct rcu_state *rsp;
3029 trace_rcu_utilization(TPS("Start CPU hotplug"));
3031 case CPU_UP_PREPARE:
3032 case CPU_UP_PREPARE_FROZEN:
3033 rcu_prepare_cpu(cpu);
3034 rcu_prepare_kthreads(cpu);
3037 case CPU_DOWN_FAILED:
3038 rcu_boost_kthread_setaffinity(rnp, -1);
3040 case CPU_DOWN_PREPARE:
3041 rcu_boost_kthread_setaffinity(rnp, cpu);
3044 case CPU_DYING_FROZEN:
3045 for_each_rcu_flavor(rsp)
3046 rcu_cleanup_dying_cpu(rsp);
3049 case CPU_DEAD_FROZEN:
3050 case CPU_UP_CANCELED:
3051 case CPU_UP_CANCELED_FROZEN:
3052 for_each_rcu_flavor(rsp)
3053 rcu_cleanup_dead_cpu(cpu, rsp);
3058 trace_rcu_utilization(TPS("End CPU hotplug"));
3062 static int rcu_pm_notify(struct notifier_block *self,
3063 unsigned long action, void *hcpu)
3066 case PM_HIBERNATION_PREPARE:
3067 case PM_SUSPEND_PREPARE:
3068 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3071 case PM_POST_HIBERNATION:
3072 case PM_POST_SUSPEND:
3082 * Spawn the kthread that handles this RCU flavor's grace periods.
3084 static int __init rcu_spawn_gp_kthread(void)
3086 unsigned long flags;
3087 struct rcu_node *rnp;
3088 struct rcu_state *rsp;
3089 struct task_struct *t;
3091 for_each_rcu_flavor(rsp) {
3092 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3094 rnp = rcu_get_root(rsp);
3095 raw_spin_lock_irqsave(&rnp->lock, flags);
3096 rsp->gp_kthread = t;
3097 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3098 rcu_spawn_nocb_kthreads(rsp);
3102 early_initcall(rcu_spawn_gp_kthread);
3105 * This function is invoked towards the end of the scheduler's initialization
3106 * process. Before this is called, the idle task might contain
3107 * RCU read-side critical sections (during which time, this idle
3108 * task is booting the system). After this function is called, the
3109 * idle tasks are prohibited from containing RCU read-side critical
3110 * sections. This function also enables RCU lockdep checking.
3112 void rcu_scheduler_starting(void)
3114 WARN_ON(num_online_cpus() != 1);
3115 WARN_ON(nr_context_switches() > 0);
3116 rcu_scheduler_active = 1;
3120 * Compute the per-level fanout, either using the exact fanout specified
3121 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3123 #ifdef CONFIG_RCU_FANOUT_EXACT
3124 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3128 for (i = rcu_num_lvls - 1; i > 0; i--)
3129 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3130 rsp->levelspread[0] = rcu_fanout_leaf;
3132 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3133 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3140 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3141 ccur = rsp->levelcnt[i];
3142 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3146 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3149 * Helper function for rcu_init() that initializes one rcu_state structure.
3151 static void __init rcu_init_one(struct rcu_state *rsp,
3152 struct rcu_data __percpu *rda)
3154 static char *buf[] = { "rcu_node_0",
3157 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3158 static char *fqs[] = { "rcu_node_fqs_0",
3161 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3165 struct rcu_node *rnp;
3167 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3169 /* Silence gcc 4.8 warning about array index out of range. */
3170 if (rcu_num_lvls > RCU_NUM_LVLS)
3171 panic("rcu_init_one: rcu_num_lvls overflow");
3173 /* Initialize the level-tracking arrays. */
3175 for (i = 0; i < rcu_num_lvls; i++)
3176 rsp->levelcnt[i] = num_rcu_lvl[i];
3177 for (i = 1; i < rcu_num_lvls; i++)
3178 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3179 rcu_init_levelspread(rsp);
3181 /* Initialize the elements themselves, starting from the leaves. */
3183 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3184 cpustride *= rsp->levelspread[i];
3185 rnp = rsp->level[i];
3186 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3187 raw_spin_lock_init(&rnp->lock);
3188 lockdep_set_class_and_name(&rnp->lock,
3189 &rcu_node_class[i], buf[i]);
3190 raw_spin_lock_init(&rnp->fqslock);
3191 lockdep_set_class_and_name(&rnp->fqslock,
3192 &rcu_fqs_class[i], fqs[i]);
3193 rnp->gpnum = rsp->gpnum;
3194 rnp->completed = rsp->completed;
3196 rnp->qsmaskinit = 0;
3197 rnp->grplo = j * cpustride;
3198 rnp->grphi = (j + 1) * cpustride - 1;
3199 if (rnp->grphi >= NR_CPUS)
3200 rnp->grphi = NR_CPUS - 1;
3206 rnp->grpnum = j % rsp->levelspread[i - 1];
3207 rnp->grpmask = 1UL << rnp->grpnum;
3208 rnp->parent = rsp->level[i - 1] +
3209 j / rsp->levelspread[i - 1];
3212 INIT_LIST_HEAD(&rnp->blkd_tasks);
3213 rcu_init_one_nocb(rnp);
3218 init_waitqueue_head(&rsp->gp_wq);
3219 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3220 rnp = rsp->level[rcu_num_lvls - 1];
3221 for_each_possible_cpu(i) {
3222 while (i > rnp->grphi)
3224 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3225 rcu_boot_init_percpu_data(i, rsp);
3227 list_add(&rsp->flavors, &rcu_struct_flavors);
3231 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3232 * replace the definitions in rcutree.h because those are needed to size
3233 * the ->node array in the rcu_state structure.
3235 static void __init rcu_init_geometry(void)
3241 int rcu_capacity[MAX_RCU_LVLS + 1];
3244 * Initialize any unspecified boot parameters.
3245 * The default values of jiffies_till_first_fqs and
3246 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3247 * value, which is a function of HZ, then adding one for each
3248 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3250 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3251 if (jiffies_till_first_fqs == ULONG_MAX)
3252 jiffies_till_first_fqs = d;
3253 if (jiffies_till_next_fqs == ULONG_MAX)
3254 jiffies_till_next_fqs = d;
3256 /* If the compile-time values are accurate, just leave. */
3257 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3258 nr_cpu_ids == NR_CPUS)
3262 * Compute number of nodes that can be handled an rcu_node tree
3263 * with the given number of levels. Setting rcu_capacity[0] makes
3264 * some of the arithmetic easier.
3266 rcu_capacity[0] = 1;
3267 rcu_capacity[1] = rcu_fanout_leaf;
3268 for (i = 2; i <= MAX_RCU_LVLS; i++)
3269 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3272 * The boot-time rcu_fanout_leaf parameter is only permitted
3273 * to increase the leaf-level fanout, not decrease it. Of course,
3274 * the leaf-level fanout cannot exceed the number of bits in
3275 * the rcu_node masks. Finally, the tree must be able to accommodate
3276 * the configured number of CPUs. Complain and fall back to the
3277 * compile-time values if these limits are exceeded.
3279 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3280 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3281 n > rcu_capacity[MAX_RCU_LVLS]) {
3286 /* Calculate the number of rcu_nodes at each level of the tree. */
3287 for (i = 1; i <= MAX_RCU_LVLS; i++)
3288 if (n <= rcu_capacity[i]) {
3289 for (j = 0; j <= i; j++)
3291 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3293 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3298 /* Calculate the total number of rcu_node structures. */
3300 for (i = 0; i <= MAX_RCU_LVLS; i++)
3301 rcu_num_nodes += num_rcu_lvl[i];
3305 void __init rcu_init(void)
3309 rcu_bootup_announce();
3310 rcu_init_geometry();
3311 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3312 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3313 __rcu_init_preempt();
3314 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3317 * We don't need protection against CPU-hotplug here because
3318 * this is called early in boot, before either interrupts
3319 * or the scheduler are operational.
3321 cpu_notifier(rcu_cpu_notify, 0);
3322 pm_notifier(rcu_pm_notify, 0);
3323 for_each_online_cpu(cpu)
3324 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3327 #include "rcutree_plugin.h"