2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
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_wait.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/sched/debug.h>
39 #include <linux/nmi.h>
40 #include <linux/atomic.h>
41 #include <linux/bitops.h>
42 #include <linux/export.h>
43 #include <linux/completion.h>
44 #include <linux/moduleparam.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <uapi/linux/sched/types.h>
54 #include <linux/prefetch.h>
55 #include <linux/delay.h>
56 #include <linux/stop_machine.h>
57 #include <linux/random.h>
58 #include <linux/trace_events.h>
59 #include <linux/suspend.h>
60 #include <linux/ftrace.h>
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
68 #define MODULE_PARAM_PREFIX "rcutree."
70 /* Data structures. */
73 * In order to export the rcu_state name to the tracing tools, it
74 * needs to be added in the __tracepoint_string section.
75 * This requires defining a separate variable tp_<sname>_varname
76 * that points to the string being used, and this will allow
77 * the tracing userspace tools to be able to decipher the string
78 * address to the matching string.
81 # define DEFINE_RCU_TPS(sname) \
82 static char sname##_varname[] = #sname; \
83 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
84 # define RCU_STATE_NAME(sname) sname##_varname
86 # define DEFINE_RCU_TPS(sname)
87 # define RCU_STATE_NAME(sname) __stringify(sname)
90 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
91 DEFINE_RCU_TPS(sname) \
92 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
93 struct rcu_state sname##_state = { \
94 .level = { &sname##_state.node[0] }, \
95 .rda = &sname##_data, \
97 .gp_state = RCU_GP_IDLE, \
98 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, \
99 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
100 .name = RCU_STATE_NAME(sname), \
102 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
103 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
104 .ofl_lock = __SPIN_LOCK_UNLOCKED(sname##_state.ofl_lock), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126 /* panic() on RCU Stall sysctl. */
127 int sysctl_panic_on_rcu_stall __read_mostly;
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
156 static int rcu_scheduler_fully_active __read_mostly;
159 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
160 struct rcu_node *rnp, unsigned long gps, unsigned long flags);
161 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
162 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
163 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
164 static void invoke_rcu_core(void);
165 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 static void rcu_report_exp_rdp(struct rcu_state *rsp,
167 struct rcu_data *rdp, bool wake);
168 static void sync_sched_exp_online_cleanup(int cpu);
170 /* rcuc/rcub kthread realtime priority */
171 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
172 module_param(kthread_prio, int, 0644);
174 /* Delay in jiffies for grace-period initialization delays, debug only. */
176 static int gp_preinit_delay;
177 module_param(gp_preinit_delay, int, 0444);
178 static int gp_init_delay;
179 module_param(gp_init_delay, int, 0444);
180 static int gp_cleanup_delay;
181 module_param(gp_cleanup_delay, int, 0444);
184 * Number of grace periods between delays, normalized by the duration of
185 * the delay. The longer the delay, the more the grace periods between
186 * each delay. The reason for this normalization is that it means that,
187 * for non-zero delays, the overall slowdown of grace periods is constant
188 * regardless of the duration of the delay. This arrangement balances
189 * the need for long delays to increase some race probabilities with the
190 * need for fast grace periods to increase other race probabilities.
192 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
195 * Track the rcutorture test sequence number and the update version
196 * number within a given test. The rcutorture_testseq is incremented
197 * on every rcutorture module load and unload, so has an odd value
198 * when a test is running. The rcutorture_vernum is set to zero
199 * when rcutorture starts and is incremented on each rcutorture update.
200 * These variables enable correlating rcutorture output with the
201 * RCU tracing information.
203 unsigned long rcutorture_testseq;
204 unsigned long rcutorture_vernum;
207 * Compute the mask of online CPUs for the specified rcu_node structure.
208 * This will not be stable unless the rcu_node structure's ->lock is
209 * held, but the bit corresponding to the current CPU will be stable
212 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
214 return READ_ONCE(rnp->qsmaskinitnext);
218 * Return true if an RCU grace period is in progress. The READ_ONCE()s
219 * permit this function to be invoked without holding the root rcu_node
220 * structure's ->lock, but of course results can be subject to change.
222 static int rcu_gp_in_progress(struct rcu_state *rsp)
224 return rcu_seq_state(rcu_seq_current(&rsp->gp_seq));
228 * Note a quiescent state. Because we do not need to know
229 * how many quiescent states passed, just if there was at least
230 * one since the start of the grace period, this just sets a flag.
231 * The caller must have disabled preemption.
233 void rcu_sched_qs(void)
235 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
236 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
238 trace_rcu_grace_period(TPS("rcu_sched"),
239 __this_cpu_read(rcu_sched_data.gp_seq),
241 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
242 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
244 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
245 rcu_report_exp_rdp(&rcu_sched_state,
246 this_cpu_ptr(&rcu_sched_data), true);
251 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
252 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
253 trace_rcu_grace_period(TPS("rcu_bh"),
254 __this_cpu_read(rcu_bh_data.gp_seq),
256 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
261 * Steal a bit from the bottom of ->dynticks for idle entry/exit
262 * control. Initially this is for TLB flushing.
264 #define RCU_DYNTICK_CTRL_MASK 0x1
265 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
266 #ifndef rcu_eqs_special_exit
267 #define rcu_eqs_special_exit() do { } while (0)
270 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
271 .dynticks_nesting = 1,
272 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
273 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
277 * Record entry into an extended quiescent state. This is only to be
278 * called when not already in an extended quiescent state.
280 static void rcu_dynticks_eqs_enter(void)
282 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
286 * CPUs seeing atomic_add_return() must see prior RCU read-side
287 * critical sections, and we also must force ordering with the
290 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
291 /* Better be in an extended quiescent state! */
292 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
293 (seq & RCU_DYNTICK_CTRL_CTR));
294 /* Better not have special action (TLB flush) pending! */
295 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
296 (seq & RCU_DYNTICK_CTRL_MASK));
300 * Record exit from an extended quiescent state. This is only to be
301 * called from an extended quiescent state.
303 static void rcu_dynticks_eqs_exit(void)
305 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
309 * CPUs seeing atomic_add_return() must see prior idle sojourns,
310 * and we also must force ordering with the next RCU read-side
313 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
314 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
315 !(seq & RCU_DYNTICK_CTRL_CTR));
316 if (seq & RCU_DYNTICK_CTRL_MASK) {
317 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
318 smp_mb__after_atomic(); /* _exit after clearing mask. */
319 /* Prefer duplicate flushes to losing a flush. */
320 rcu_eqs_special_exit();
325 * Reset the current CPU's ->dynticks counter to indicate that the
326 * newly onlined CPU is no longer in an extended quiescent state.
327 * This will either leave the counter unchanged, or increment it
328 * to the next non-quiescent value.
330 * The non-atomic test/increment sequence works because the upper bits
331 * of the ->dynticks counter are manipulated only by the corresponding CPU,
332 * or when the corresponding CPU is offline.
334 static void rcu_dynticks_eqs_online(void)
336 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
338 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
340 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
344 * Is the current CPU in an extended quiescent state?
346 * No ordering, as we are sampling CPU-local information.
348 bool rcu_dynticks_curr_cpu_in_eqs(void)
350 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
352 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
356 * Snapshot the ->dynticks counter with full ordering so as to allow
357 * stable comparison of this counter with past and future snapshots.
359 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
361 int snap = atomic_add_return(0, &rdtp->dynticks);
363 return snap & ~RCU_DYNTICK_CTRL_MASK;
367 * Return true if the snapshot returned from rcu_dynticks_snap()
368 * indicates that RCU is in an extended quiescent state.
370 static bool rcu_dynticks_in_eqs(int snap)
372 return !(snap & RCU_DYNTICK_CTRL_CTR);
376 * Return true if the CPU corresponding to the specified rcu_dynticks
377 * structure has spent some time in an extended quiescent state since
378 * rcu_dynticks_snap() returned the specified snapshot.
380 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
382 return snap != rcu_dynticks_snap(rdtp);
386 * Do a double-increment of the ->dynticks counter to emulate a
387 * momentary idle-CPU quiescent state.
389 static void rcu_dynticks_momentary_idle(void)
391 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
392 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
395 /* It is illegal to call this from idle state. */
396 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
400 * Set the special (bottom) bit of the specified CPU so that it
401 * will take special action (such as flushing its TLB) on the
402 * next exit from an extended quiescent state. Returns true if
403 * the bit was successfully set, or false if the CPU was not in
404 * an extended quiescent state.
406 bool rcu_eqs_special_set(int cpu)
410 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
413 old = atomic_read(&rdtp->dynticks);
414 if (old & RCU_DYNTICK_CTRL_CTR)
416 new = old | RCU_DYNTICK_CTRL_MASK;
417 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
422 * Let the RCU core know that this CPU has gone through the scheduler,
423 * which is a quiescent state. This is called when the need for a
424 * quiescent state is urgent, so we burn an atomic operation and full
425 * memory barriers to let the RCU core know about it, regardless of what
426 * this CPU might (or might not) do in the near future.
428 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
430 * The caller must have disabled interrupts.
432 static void rcu_momentary_dyntick_idle(void)
434 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
435 rcu_dynticks_momentary_idle();
439 * Note a context switch. This is a quiescent state for RCU-sched,
440 * and requires special handling for preemptible RCU.
441 * The caller must have disabled interrupts.
443 void rcu_note_context_switch(bool preempt)
445 barrier(); /* Avoid RCU read-side critical sections leaking down. */
446 trace_rcu_utilization(TPS("Start context switch"));
448 rcu_preempt_note_context_switch(preempt);
449 /* Load rcu_urgent_qs before other flags. */
450 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
452 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
453 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
454 rcu_momentary_dyntick_idle();
455 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
457 rcu_note_voluntary_context_switch_lite(current);
459 trace_rcu_utilization(TPS("End context switch"));
460 barrier(); /* Avoid RCU read-side critical sections leaking up. */
462 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
465 * Register a quiescent state for all RCU flavors. If there is an
466 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
467 * dyntick-idle quiescent state visible to other CPUs (but only for those
468 * RCU flavors in desperate need of a quiescent state, which will normally
469 * be none of them). Either way, do a lightweight quiescent state for
472 * The barrier() calls are redundant in the common case when this is
473 * called externally, but just in case this is called from within this
477 void rcu_all_qs(void)
481 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
484 /* Load rcu_urgent_qs before other flags. */
485 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
489 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
490 barrier(); /* Avoid RCU read-side critical sections leaking down. */
491 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
492 local_irq_save(flags);
493 rcu_momentary_dyntick_idle();
494 local_irq_restore(flags);
496 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
498 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
499 barrier(); /* Avoid RCU read-side critical sections leaking up. */
502 EXPORT_SYMBOL_GPL(rcu_all_qs);
504 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
505 static long blimit = DEFAULT_RCU_BLIMIT;
506 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
507 static long qhimark = DEFAULT_RCU_QHIMARK;
508 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
509 static long qlowmark = DEFAULT_RCU_QLOMARK;
511 module_param(blimit, long, 0444);
512 module_param(qhimark, long, 0444);
513 module_param(qlowmark, long, 0444);
515 static ulong jiffies_till_first_fqs = ULONG_MAX;
516 static ulong jiffies_till_next_fqs = ULONG_MAX;
517 static bool rcu_kick_kthreads;
519 module_param(jiffies_till_first_fqs, ulong, 0644);
520 module_param(jiffies_till_next_fqs, ulong, 0644);
521 module_param(rcu_kick_kthreads, bool, 0644);
524 * How long the grace period must be before we start recruiting
525 * quiescent-state help from rcu_note_context_switch().
527 static ulong jiffies_till_sched_qs = HZ / 10;
528 module_param(jiffies_till_sched_qs, ulong, 0444);
530 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
531 static void force_quiescent_state(struct rcu_state *rsp);
532 static int rcu_pending(void);
535 * Return the number of RCU GPs completed thus far for debug & stats.
537 unsigned long rcu_get_gp_seq(void)
539 return READ_ONCE(rcu_state_p->gp_seq);
541 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
544 * Return the number of RCU-sched GPs completed thus far for debug & stats.
546 unsigned long rcu_sched_get_gp_seq(void)
548 return READ_ONCE(rcu_sched_state.gp_seq);
550 EXPORT_SYMBOL_GPL(rcu_sched_get_gp_seq);
553 * Return the number of RCU-bh GPs completed thus far for debug & stats.
555 unsigned long rcu_bh_get_gp_seq(void)
557 return READ_ONCE(rcu_bh_state.gp_seq);
559 EXPORT_SYMBOL_GPL(rcu_bh_get_gp_seq);
562 * Return the number of RCU expedited batches completed thus far for
563 * debug & stats. Odd numbers mean that a batch is in progress, even
564 * numbers mean idle. The value returned will thus be roughly double
565 * the cumulative batches since boot.
567 unsigned long rcu_exp_batches_completed(void)
569 return rcu_state_p->expedited_sequence;
571 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
574 * Return the number of RCU-sched expedited batches completed thus far
575 * for debug & stats. Similar to rcu_exp_batches_completed().
577 unsigned long rcu_exp_batches_completed_sched(void)
579 return rcu_sched_state.expedited_sequence;
581 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
584 * Force a quiescent state.
586 void rcu_force_quiescent_state(void)
588 force_quiescent_state(rcu_state_p);
590 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
593 * Force a quiescent state for RCU BH.
595 void rcu_bh_force_quiescent_state(void)
597 force_quiescent_state(&rcu_bh_state);
599 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
602 * Force a quiescent state for RCU-sched.
604 void rcu_sched_force_quiescent_state(void)
606 force_quiescent_state(&rcu_sched_state);
608 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
611 * Show the state of the grace-period kthreads.
613 void show_rcu_gp_kthreads(void)
615 struct rcu_state *rsp;
617 for_each_rcu_flavor(rsp) {
618 pr_info("%s: wait state: %d ->state: %#lx\n",
619 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
620 /* sched_show_task(rsp->gp_kthread); */
623 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
626 * Record the number of times rcutorture tests have been initiated and
627 * terminated. This information allows the debugfs tracing stats to be
628 * correlated to the rcutorture messages, even when the rcutorture module
629 * is being repeatedly loaded and unloaded. In other words, we cannot
630 * store this state in rcutorture itself.
632 void rcutorture_record_test_transition(void)
634 rcutorture_testseq++;
635 rcutorture_vernum = 0;
637 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
640 * Send along grace-period-related data for rcutorture diagnostics.
642 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
643 unsigned long *gp_seq)
645 struct rcu_state *rsp = NULL;
654 case RCU_SCHED_FLAVOR:
655 rsp = &rcu_sched_state;
662 *flags = READ_ONCE(rsp->gp_flags);
663 *gp_seq = rcu_seq_current(&rsp->gp_seq);
665 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
668 * Record the number of writer passes through the current rcutorture test.
669 * This is also used to correlate debugfs tracing stats with the rcutorture
672 void rcutorture_record_progress(unsigned long vernum)
676 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
679 * Return the root node of the specified rcu_state structure.
681 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
683 return &rsp->node[0];
687 * Enter an RCU extended quiescent state, which can be either the
688 * idle loop or adaptive-tickless usermode execution.
690 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
691 * the possibility of usermode upcalls having messed up our count
692 * of interrupt nesting level during the prior busy period.
694 static void rcu_eqs_enter(bool user)
696 struct rcu_state *rsp;
697 struct rcu_data *rdp;
698 struct rcu_dynticks *rdtp;
700 rdtp = this_cpu_ptr(&rcu_dynticks);
701 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
702 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
703 rdtp->dynticks_nesting == 0);
704 if (rdtp->dynticks_nesting != 1) {
705 rdtp->dynticks_nesting--;
709 lockdep_assert_irqs_disabled();
710 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
711 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
712 for_each_rcu_flavor(rsp) {
713 rdp = this_cpu_ptr(rsp->rda);
714 do_nocb_deferred_wakeup(rdp);
716 rcu_prepare_for_idle();
717 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
718 rcu_dynticks_eqs_enter();
719 rcu_dynticks_task_enter();
723 * rcu_idle_enter - inform RCU that current CPU is entering idle
725 * Enter idle mode, in other words, -leave- the mode in which RCU
726 * read-side critical sections can occur. (Though RCU read-side
727 * critical sections can occur in irq handlers in idle, a possibility
728 * handled by irq_enter() and irq_exit().)
730 * If you add or remove a call to rcu_idle_enter(), be sure to test with
731 * CONFIG_RCU_EQS_DEBUG=y.
733 void rcu_idle_enter(void)
735 lockdep_assert_irqs_disabled();
736 rcu_eqs_enter(false);
739 #ifdef CONFIG_NO_HZ_FULL
741 * rcu_user_enter - inform RCU that we are resuming userspace.
743 * Enter RCU idle mode right before resuming userspace. No use of RCU
744 * is permitted between this call and rcu_user_exit(). This way the
745 * CPU doesn't need to maintain the tick for RCU maintenance purposes
746 * when the CPU runs in userspace.
748 * If you add or remove a call to rcu_user_enter(), be sure to test with
749 * CONFIG_RCU_EQS_DEBUG=y.
751 void rcu_user_enter(void)
753 lockdep_assert_irqs_disabled();
756 #endif /* CONFIG_NO_HZ_FULL */
759 * rcu_nmi_exit - inform RCU of exit from NMI context
761 * If we are returning from the outermost NMI handler that interrupted an
762 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
763 * to let the RCU grace-period handling know that the CPU is back to
766 * If you add or remove a call to rcu_nmi_exit(), be sure to test
767 * with CONFIG_RCU_EQS_DEBUG=y.
769 void rcu_nmi_exit(void)
771 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
774 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
775 * (We are exiting an NMI handler, so RCU better be paying attention
778 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
779 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
782 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
783 * leave it in non-RCU-idle state.
785 if (rdtp->dynticks_nmi_nesting != 1) {
786 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
787 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
788 rdtp->dynticks_nmi_nesting - 2);
792 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
793 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
794 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
795 rcu_dynticks_eqs_enter();
799 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
801 * Exit from an interrupt handler, which might possibly result in entering
802 * idle mode, in other words, leaving the mode in which read-side critical
803 * sections can occur. The caller must have disabled interrupts.
805 * This code assumes that the idle loop never does anything that might
806 * result in unbalanced calls to irq_enter() and irq_exit(). If your
807 * architecture's idle loop violates this assumption, RCU will give you what
808 * you deserve, good and hard. But very infrequently and irreproducibly.
810 * Use things like work queues to work around this limitation.
812 * You have been warned.
814 * If you add or remove a call to rcu_irq_exit(), be sure to test with
815 * CONFIG_RCU_EQS_DEBUG=y.
817 void rcu_irq_exit(void)
819 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
821 lockdep_assert_irqs_disabled();
822 if (rdtp->dynticks_nmi_nesting == 1)
823 rcu_prepare_for_idle();
825 if (rdtp->dynticks_nmi_nesting == 0)
826 rcu_dynticks_task_enter();
830 * Wrapper for rcu_irq_exit() where interrupts are enabled.
832 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
833 * with CONFIG_RCU_EQS_DEBUG=y.
835 void rcu_irq_exit_irqson(void)
839 local_irq_save(flags);
841 local_irq_restore(flags);
845 * Exit an RCU extended quiescent state, which can be either the
846 * idle loop or adaptive-tickless usermode execution.
848 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
849 * allow for the possibility of usermode upcalls messing up our count of
850 * interrupt nesting level during the busy period that is just now starting.
852 static void rcu_eqs_exit(bool user)
854 struct rcu_dynticks *rdtp;
857 lockdep_assert_irqs_disabled();
858 rdtp = this_cpu_ptr(&rcu_dynticks);
859 oldval = rdtp->dynticks_nesting;
860 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
862 rdtp->dynticks_nesting++;
865 rcu_dynticks_task_exit();
866 rcu_dynticks_eqs_exit();
867 rcu_cleanup_after_idle();
868 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
869 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
870 WRITE_ONCE(rdtp->dynticks_nesting, 1);
871 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
875 * rcu_idle_exit - inform RCU that current CPU is leaving idle
877 * Exit idle mode, in other words, -enter- the mode in which RCU
878 * read-side critical sections can occur.
880 * If you add or remove a call to rcu_idle_exit(), be sure to test with
881 * CONFIG_RCU_EQS_DEBUG=y.
883 void rcu_idle_exit(void)
887 local_irq_save(flags);
889 local_irq_restore(flags);
892 #ifdef CONFIG_NO_HZ_FULL
894 * rcu_user_exit - inform RCU that we are exiting userspace.
896 * Exit RCU idle mode while entering the kernel because it can
897 * run a RCU read side critical section anytime.
899 * If you add or remove a call to rcu_user_exit(), be sure to test with
900 * CONFIG_RCU_EQS_DEBUG=y.
902 void rcu_user_exit(void)
906 #endif /* CONFIG_NO_HZ_FULL */
909 * rcu_nmi_enter - inform RCU of entry to NMI context
911 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
912 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
913 * that the CPU is active. This implementation permits nested NMIs, as
914 * long as the nesting level does not overflow an int. (You will probably
915 * run out of stack space first.)
917 * If you add or remove a call to rcu_nmi_enter(), be sure to test
918 * with CONFIG_RCU_EQS_DEBUG=y.
920 void rcu_nmi_enter(void)
922 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
925 /* Complain about underflow. */
926 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
929 * If idle from RCU viewpoint, atomically increment ->dynticks
930 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
931 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
932 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
933 * to be in the outermost NMI handler that interrupted an RCU-idle
934 * period (observation due to Andy Lutomirski).
936 if (rcu_dynticks_curr_cpu_in_eqs()) {
937 rcu_dynticks_eqs_exit();
940 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
941 rdtp->dynticks_nmi_nesting,
942 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
943 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
944 rdtp->dynticks_nmi_nesting + incby);
949 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
951 * Enter an interrupt handler, which might possibly result in exiting
952 * idle mode, in other words, entering the mode in which read-side critical
953 * sections can occur. The caller must have disabled interrupts.
955 * Note that the Linux kernel is fully capable of entering an interrupt
956 * handler that it never exits, for example when doing upcalls to user mode!
957 * This code assumes that the idle loop never does upcalls to user mode.
958 * If your architecture's idle loop does do upcalls to user mode (or does
959 * anything else that results in unbalanced calls to the irq_enter() and
960 * irq_exit() functions), RCU will give you what you deserve, good and hard.
961 * But very infrequently and irreproducibly.
963 * Use things like work queues to work around this limitation.
965 * You have been warned.
967 * If you add or remove a call to rcu_irq_enter(), be sure to test with
968 * CONFIG_RCU_EQS_DEBUG=y.
970 void rcu_irq_enter(void)
972 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
974 lockdep_assert_irqs_disabled();
975 if (rdtp->dynticks_nmi_nesting == 0)
976 rcu_dynticks_task_exit();
978 if (rdtp->dynticks_nmi_nesting == 1)
979 rcu_cleanup_after_idle();
983 * Wrapper for rcu_irq_enter() where interrupts are enabled.
985 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
986 * with CONFIG_RCU_EQS_DEBUG=y.
988 void rcu_irq_enter_irqson(void)
992 local_irq_save(flags);
994 local_irq_restore(flags);
998 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1000 * Return true if RCU is watching the running CPU, which means that this
1001 * CPU can safely enter RCU read-side critical sections. In other words,
1002 * if the current CPU is in its idle loop and is neither in an interrupt
1003 * or NMI handler, return true.
1005 bool notrace rcu_is_watching(void)
1009 preempt_disable_notrace();
1010 ret = !rcu_dynticks_curr_cpu_in_eqs();
1011 preempt_enable_notrace();
1014 EXPORT_SYMBOL_GPL(rcu_is_watching);
1017 * If a holdout task is actually running, request an urgent quiescent
1018 * state from its CPU. This is unsynchronized, so migrations can cause
1019 * the request to go to the wrong CPU. Which is OK, all that will happen
1020 * is that the CPU's next context switch will be a bit slower and next
1021 * time around this task will generate another request.
1023 void rcu_request_urgent_qs_task(struct task_struct *t)
1030 return; /* This task is not running on that CPU. */
1031 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1034 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1037 * Is the current CPU online as far as RCU is concerned?
1039 * Disable preemption to avoid false positives that could otherwise
1040 * happen due to the current CPU number being sampled, this task being
1041 * preempted, its old CPU being taken offline, resuming on some other CPU,
1042 * then determining that its old CPU is now offline. Because there are
1043 * multiple flavors of RCU, and because this function can be called in the
1044 * midst of updating the flavors while a given CPU coming online or going
1045 * offline, it is necessary to check all flavors. If any of the flavors
1046 * believe that given CPU is online, it is considered to be online.
1048 * Disable checking if in an NMI handler because we cannot safely
1049 * report errors from NMI handlers anyway. In addition, it is OK to use
1050 * RCU on an offline processor during initial boot, hence the check for
1051 * rcu_scheduler_fully_active.
1053 bool rcu_lockdep_current_cpu_online(void)
1055 struct rcu_data *rdp;
1056 struct rcu_node *rnp;
1057 struct rcu_state *rsp;
1059 if (in_nmi() || !rcu_scheduler_fully_active)
1062 for_each_rcu_flavor(rsp) {
1063 rdp = this_cpu_ptr(rsp->rda);
1065 if (rdp->grpmask & rcu_rnp_online_cpus(rnp)) {
1073 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1075 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1078 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1080 * If the current CPU is idle or running at a first-level (not nested)
1081 * interrupt from idle, return true. The caller must have at least
1082 * disabled preemption.
1084 static int rcu_is_cpu_rrupt_from_idle(void)
1086 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1087 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1091 * We are reporting a quiescent state on behalf of some other CPU, so
1092 * it is our responsibility to check for and handle potential overflow
1093 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1094 * After all, the CPU might be in deep idle state, and thus executing no
1097 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1099 raw_lockdep_assert_held_rcu_node(rnp);
1100 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1102 WRITE_ONCE(rdp->gpwrap, true);
1103 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1104 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1108 * Snapshot the specified CPU's dynticks counter so that we can later
1109 * credit them with an implicit quiescent state. Return 1 if this CPU
1110 * is in dynticks idle mode, which is an extended quiescent state.
1112 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1114 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1115 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1116 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1117 rcu_gpnum_ovf(rdp->mynode, rdp);
1124 * Handler for the irq_work request posted when a grace period has
1125 * gone on for too long, but not yet long enough for an RCU CPU
1126 * stall warning. Set state appropriately, but just complain if
1127 * there is unexpected state on entry.
1129 static void rcu_iw_handler(struct irq_work *iwp)
1131 struct rcu_data *rdp;
1132 struct rcu_node *rnp;
1134 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1136 raw_spin_lock_rcu_node(rnp);
1137 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1138 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1139 rdp->rcu_iw_pending = false;
1141 raw_spin_unlock_rcu_node(rnp);
1145 * Return true if the specified CPU has passed through a quiescent
1146 * state by virtue of being in or having passed through an dynticks
1147 * idle state since the last call to dyntick_save_progress_counter()
1148 * for this same CPU, or by virtue of having been offline.
1150 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1155 struct rcu_node *rnp = rdp->mynode;
1158 * If the CPU passed through or entered a dynticks idle phase with
1159 * no active irq/NMI handlers, then we can safely pretend that the CPU
1160 * already acknowledged the request to pass through a quiescent
1161 * state. Either way, that CPU cannot possibly be in an RCU
1162 * read-side critical section that started before the beginning
1163 * of the current RCU grace period.
1165 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1166 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1167 rdp->dynticks_fqs++;
1168 rcu_gpnum_ovf(rnp, rdp);
1173 * Has this CPU encountered a cond_resched() since the beginning
1174 * of the grace period? For this to be the case, the CPU has to
1175 * have noticed the current grace period. This might not be the
1176 * case for nohz_full CPUs looping in the kernel.
1178 jtsq = jiffies_till_sched_qs;
1179 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1180 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1181 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1182 rcu_seq_current(&rdp->gp_seq) == rnp->gp_seq && !rdp->gpwrap) {
1183 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("rqc"));
1184 rcu_gpnum_ovf(rnp, rdp);
1186 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1187 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1188 smp_store_release(ruqp, true);
1192 * A CPU running for an extended time within the kernel can
1193 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1194 * even context-switching back and forth between a pair of
1195 * in-kernel CPU-bound tasks cannot advance grace periods.
1196 * So if the grace period is old enough, make the CPU pay attention.
1197 * Note that the unsynchronized assignments to the per-CPU
1198 * rcu_need_heavy_qs variable are safe. Yes, setting of
1199 * bits can be lost, but they will be set again on the next
1200 * force-quiescent-state pass. So lost bit sets do not result
1201 * in incorrect behavior, merely in a grace period lasting
1202 * a few jiffies longer than it might otherwise. Because
1203 * there are at most four threads involved, and because the
1204 * updates are only once every few jiffies, the probability of
1205 * lossage (and thus of slight grace-period extension) is
1208 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1209 if (!READ_ONCE(*rnhqp) &&
1210 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1211 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1212 WRITE_ONCE(*rnhqp, true);
1213 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1214 smp_store_release(ruqp, true);
1215 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1219 * If more than halfway to RCU CPU stall-warning time, do a
1220 * resched_cpu() to try to loosen things up a bit. Also check to
1221 * see if the CPU is getting hammered with interrupts, but only
1222 * once per grace period, just to keep the IPIs down to a dull roar.
1224 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1225 resched_cpu(rdp->cpu);
1226 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1227 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1228 (rnp->ffmask & rdp->grpmask)) {
1229 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1230 rdp->rcu_iw_pending = true;
1231 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1232 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1239 static void record_gp_stall_check_time(struct rcu_state *rsp)
1241 unsigned long j = jiffies;
1245 j1 = rcu_jiffies_till_stall_check();
1246 /* Record ->gp_start before ->jiffies_stall. */
1247 smp_store_release(&rsp->jiffies_stall, j + j1); /* ^^^ */
1248 rsp->jiffies_resched = j + j1 / 2;
1249 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1253 * Convert a ->gp_state value to a character string.
1255 static const char *gp_state_getname(short gs)
1257 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1259 return gp_state_names[gs];
1263 * Complain about starvation of grace-period kthread.
1265 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1271 gpa = READ_ONCE(rsp->gp_activity);
1272 if (j - gpa > 2 * HZ) {
1273 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1275 (long)rcu_seq_current(&rsp->gp_seq),
1277 gp_state_getname(rsp->gp_state), rsp->gp_state,
1278 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1279 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1280 if (rsp->gp_kthread) {
1281 pr_err("RCU grace-period kthread stack dump:\n");
1282 sched_show_task(rsp->gp_kthread);
1283 wake_up_process(rsp->gp_kthread);
1289 * Dump stacks of all tasks running on stalled CPUs. First try using
1290 * NMIs, but fall back to manual remote stack tracing on architectures
1291 * that don't support NMI-based stack dumps. The NMI-triggered stack
1292 * traces are more accurate because they are printed by the target CPU.
1294 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1297 unsigned long flags;
1298 struct rcu_node *rnp;
1300 rcu_for_each_leaf_node(rsp, rnp) {
1301 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1302 for_each_leaf_node_possible_cpu(rnp, cpu)
1303 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1304 if (!trigger_single_cpu_backtrace(cpu))
1306 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1311 * If too much time has passed in the current grace period, and if
1312 * so configured, go kick the relevant kthreads.
1314 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1318 if (!rcu_kick_kthreads)
1320 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1321 if (time_after(jiffies, j) && rsp->gp_kthread &&
1322 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1323 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1324 rcu_ftrace_dump(DUMP_ALL);
1325 wake_up_process(rsp->gp_kthread);
1326 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1330 static inline void panic_on_rcu_stall(void)
1332 if (sysctl_panic_on_rcu_stall)
1333 panic("RCU Stall\n");
1336 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gp_seq)
1339 unsigned long flags;
1343 struct rcu_node *rnp = rcu_get_root(rsp);
1346 /* Kick and suppress, if so configured. */
1347 rcu_stall_kick_kthreads(rsp);
1348 if (rcu_cpu_stall_suppress)
1352 * OK, time to rat on our buddy...
1353 * See Documentation/RCU/stallwarn.txt for info on how to debug
1354 * RCU CPU stall warnings.
1356 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1358 print_cpu_stall_info_begin();
1359 rcu_for_each_leaf_node(rsp, rnp) {
1360 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1361 ndetected += rcu_print_task_stall(rnp);
1362 if (rnp->qsmask != 0) {
1363 for_each_leaf_node_possible_cpu(rnp, cpu)
1364 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1365 print_cpu_stall_info(rsp, cpu);
1369 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1372 print_cpu_stall_info_end();
1373 for_each_possible_cpu(cpu)
1374 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1376 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
1377 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1378 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1380 rcu_dump_cpu_stacks(rsp);
1382 /* Complain about tasks blocking the grace period. */
1383 rcu_print_detail_task_stall(rsp);
1385 if (rcu_seq_current(&rsp->gp_seq) != gp_seq) {
1386 pr_err("INFO: Stall ended before state dump start\n");
1389 gpa = READ_ONCE(rsp->gp_activity);
1390 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1391 rsp->name, j - gpa, j, gpa,
1392 jiffies_till_next_fqs,
1393 rcu_get_root(rsp)->qsmask);
1394 /* In this case, the current CPU might be at fault. */
1395 sched_show_task(current);
1398 /* Rewrite if needed in case of slow consoles. */
1399 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1400 WRITE_ONCE(rsp->jiffies_stall,
1401 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1403 rcu_check_gp_kthread_starvation(rsp);
1405 panic_on_rcu_stall();
1407 force_quiescent_state(rsp); /* Kick them all. */
1410 static void print_cpu_stall(struct rcu_state *rsp)
1413 unsigned long flags;
1414 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1415 struct rcu_node *rnp = rcu_get_root(rsp);
1418 /* Kick and suppress, if so configured. */
1419 rcu_stall_kick_kthreads(rsp);
1420 if (rcu_cpu_stall_suppress)
1424 * OK, time to rat on ourselves...
1425 * See Documentation/RCU/stallwarn.txt for info on how to debug
1426 * RCU CPU stall warnings.
1428 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1429 print_cpu_stall_info_begin();
1430 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1431 print_cpu_stall_info(rsp, smp_processor_id());
1432 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1433 print_cpu_stall_info_end();
1434 for_each_possible_cpu(cpu)
1435 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1437 pr_cont(" (t=%lu jiffies g=%ld q=%lu)\n",
1438 jiffies - rsp->gp_start,
1439 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1441 rcu_check_gp_kthread_starvation(rsp);
1443 rcu_dump_cpu_stacks(rsp);
1445 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1446 /* Rewrite if needed in case of slow consoles. */
1447 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1448 WRITE_ONCE(rsp->jiffies_stall,
1449 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1450 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1452 panic_on_rcu_stall();
1455 * Attempt to revive the RCU machinery by forcing a context switch.
1457 * A context switch would normally allow the RCU state machine to make
1458 * progress and it could be we're stuck in kernel space without context
1459 * switches for an entirely unreasonable amount of time.
1461 resched_cpu(smp_processor_id());
1464 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1472 struct rcu_node *rnp;
1474 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1475 !rcu_gp_in_progress(rsp))
1477 rcu_stall_kick_kthreads(rsp);
1481 * Lots of memory barriers to reject false positives.
1483 * The idea is to pick up rsp->gp_seq, then rsp->jiffies_stall,
1484 * then rsp->gp_start, and finally another copy of rsp->gp_seq.
1485 * These values are updated in the opposite order with memory
1486 * barriers (or equivalent) during grace-period initialization
1487 * and cleanup. Now, a false positive can occur if we get an new
1488 * value of rsp->gp_start and a old value of rsp->jiffies_stall.
1489 * But given the memory barriers, the only way that this can happen
1490 * is if one grace period ends and another starts between these
1491 * two fetches. This is detected by comparing the second fetch
1492 * of rsp->gp_seq with the previous fetch from rsp->gp_seq.
1494 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1495 * and rsp->gp_start suffice to forestall false positives.
1497 gs1 = READ_ONCE(rsp->gp_seq);
1498 smp_rmb(); /* Pick up ->gp_seq first... */
1499 js = READ_ONCE(rsp->jiffies_stall);
1500 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1501 gps = READ_ONCE(rsp->gp_start);
1502 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
1503 gs2 = READ_ONCE(rsp->gp_seq);
1505 ULONG_CMP_LT(j, js) ||
1506 ULONG_CMP_GE(gps, js))
1507 return; /* No stall or GP completed since entering function. */
1509 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1510 if (rcu_gp_in_progress(rsp) &&
1511 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
1512 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1514 /* We haven't checked in, so go dump stack. */
1515 print_cpu_stall(rsp);
1517 } else if (rcu_gp_in_progress(rsp) &&
1518 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
1519 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1521 /* They had a few time units to dump stack, so complain. */
1522 print_other_cpu_stall(rsp, gs2);
1527 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1529 * Set the stall-warning timeout way off into the future, thus preventing
1530 * any RCU CPU stall-warning messages from appearing in the current set of
1531 * RCU grace periods.
1533 * The caller must disable hard irqs.
1535 void rcu_cpu_stall_reset(void)
1537 struct rcu_state *rsp;
1539 for_each_rcu_flavor(rsp)
1540 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1543 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1544 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1545 unsigned long gp_seq_req, const char *s)
1547 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gp_seq, gp_seq_req,
1548 rnp->level, rnp->grplo, rnp->grphi, s);
1552 * rcu_start_this_gp - Request the start of a particular grace period
1553 * @rnp_start: The leaf node of the CPU from which to start.
1554 * @rdp: The rcu_data corresponding to the CPU from which to start.
1555 * @gp_seq_req: The gp_seq of the grace period to start.
1557 * Start the specified grace period, as needed to handle newly arrived
1558 * callbacks. The required future grace periods are recorded in each
1559 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1560 * is reason to awaken the grace-period kthread.
1562 * The caller must hold the specified rcu_node structure's ->lock, which
1563 * is why the caller is responsible for waking the grace-period kthread.
1565 * Returns true if the GP thread needs to be awakened else false.
1567 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1568 unsigned long gp_seq_req)
1571 struct rcu_state *rsp = rdp->rsp;
1572 struct rcu_node *rnp;
1575 * Use funnel locking to either acquire the root rcu_node
1576 * structure's lock or bail out if the need for this grace period
1577 * has already been recorded -- or if that grace period has in
1578 * fact already started. If there is already a grace period in
1579 * progress in a non-leaf node, no recording is needed because the
1580 * end of the grace period will scan the leaf rcu_node structures.
1581 * Note that rnp_start->lock must not be released.
1583 raw_lockdep_assert_held_rcu_node(rnp_start);
1584 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1585 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1586 if (rnp != rnp_start)
1587 raw_spin_lock_rcu_node(rnp);
1588 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1589 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1590 (rnp != rnp_start &&
1591 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1592 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1596 rnp->gp_seq_needed = gp_seq_req;
1597 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1599 * We just marked the leaf or internal node, and a
1600 * grace period is in progress, which means that
1601 * rcu_gp_cleanup() will see the marking. Bail to
1602 * reduce contention.
1604 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1605 TPS("Startedleaf"));
1608 if (rnp != rnp_start && rnp->parent != NULL)
1609 raw_spin_unlock_rcu_node(rnp);
1611 break; /* At root, and perhaps also leaf. */
1614 /* If GP already in progress, just leave, otherwise start one. */
1615 if (rcu_gp_in_progress(rsp)) {
1616 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1619 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1620 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags | RCU_GP_FLAG_INIT);
1621 rsp->gp_req_activity = jiffies;
1622 if (!rsp->gp_kthread) {
1623 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1626 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq), TPS("newreq"));
1627 ret = true; /* Caller must wake GP kthread. */
1629 /* Push furthest requested GP to leaf node and rcu_data structure. */
1630 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1631 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1632 rdp->gp_seq_needed = rnp->gp_seq_needed;
1634 if (rnp != rnp_start)
1635 raw_spin_unlock_rcu_node(rnp);
1640 * Clean up any old requests for the just-ended grace period. Also return
1641 * whether any additional grace periods have been requested.
1643 static bool rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1646 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1648 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1650 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1651 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1652 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1657 * Awaken the grace-period kthread for the specified flavor of RCU.
1658 * Don't do a self-awaken, and don't bother awakening when there is
1659 * nothing for the grace-period kthread to do (as in several CPUs
1660 * raced to awaken, and we lost), and finally don't try to awaken
1661 * a kthread that has not yet been created.
1663 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1665 if (current == rsp->gp_kthread ||
1666 !READ_ONCE(rsp->gp_flags) ||
1669 swake_up(&rsp->gp_wq);
1673 * If there is room, assign a ->gp_seq number to any callbacks on this
1674 * CPU that have not already been assigned. Also accelerate any callbacks
1675 * that were previously assigned a ->gp_seq number that has since proven
1676 * to be too conservative, which can happen if callbacks get assigned a
1677 * ->gp_seq number while RCU is idle, but with reference to a non-root
1678 * rcu_node structure. This function is idempotent, so it does not hurt
1679 * to call it repeatedly. Returns an flag saying that we should awaken
1680 * the RCU grace-period kthread.
1682 * The caller must hold rnp->lock with interrupts disabled.
1684 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1685 struct rcu_data *rdp)
1687 unsigned long gp_seq_req;
1690 raw_lockdep_assert_held_rcu_node(rnp);
1692 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1693 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1697 * Callbacks are often registered with incomplete grace-period
1698 * information. Something about the fact that getting exact
1699 * information requires acquiring a global lock... RCU therefore
1700 * makes a conservative estimate of the grace period number at which
1701 * a given callback will become ready to invoke. The following
1702 * code checks this estimate and improves it when possible, thus
1703 * accelerating callback invocation to an earlier grace-period
1706 gp_seq_req = rcu_seq_snap(&rsp->gp_seq);
1707 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1708 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1710 /* Trace depending on how much we were able to accelerate. */
1711 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1712 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccWaitCB"));
1714 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccReadyCB"));
1719 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1720 * rcu_node structure's ->lock be held. It consults the cached value
1721 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1722 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1723 * while holding the leaf rcu_node structure's ->lock.
1725 static void rcu_accelerate_cbs_unlocked(struct rcu_state *rsp,
1726 struct rcu_node *rnp,
1727 struct rcu_data *rdp)
1732 lockdep_assert_irqs_disabled();
1733 c = rcu_seq_snap(&rsp->gp_seq);
1734 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1735 /* Old request still live, so mark recent callbacks. */
1736 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1739 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1740 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1741 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1743 rcu_gp_kthread_wake(rsp);
1747 * Move any callbacks whose grace period has completed to the
1748 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1749 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1750 * sublist. This function is idempotent, so it does not hurt to
1751 * invoke it repeatedly. As long as it is not invoked -too- often...
1752 * Returns true if the RCU grace-period kthread needs to be awakened.
1754 * The caller must hold rnp->lock with interrupts disabled.
1756 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1757 struct rcu_data *rdp)
1759 raw_lockdep_assert_held_rcu_node(rnp);
1761 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1762 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1766 * Find all callbacks whose ->gp_seq numbers indicate that they
1767 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1769 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1771 /* Classify any remaining callbacks. */
1772 return rcu_accelerate_cbs(rsp, rnp, rdp);
1776 * Update CPU-local rcu_data state to record the beginnings and ends of
1777 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1778 * structure corresponding to the current CPU, and must have irqs disabled.
1779 * Returns true if the grace-period kthread needs to be awakened.
1781 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1782 struct rcu_data *rdp)
1787 raw_lockdep_assert_held_rcu_node(rnp);
1789 if (rdp->gp_seq == rnp->gp_seq)
1790 return false; /* Nothing to do. */
1792 /* Handle the ends of any preceding grace periods first. */
1793 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1794 unlikely(READ_ONCE(rdp->gpwrap))) {
1795 ret = rcu_advance_cbs(rsp, rnp, rdp); /* Advance callbacks. */
1796 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuend"));
1798 ret = rcu_accelerate_cbs(rsp, rnp, rdp); /* Recent callbacks. */
1801 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1802 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1803 unlikely(READ_ONCE(rdp->gpwrap))) {
1805 * If the current grace period is waiting for this CPU,
1806 * set up to detect a quiescent state, otherwise don't
1807 * go looking for one.
1809 trace_rcu_grace_period(rsp->name, rnp->gp_seq, TPS("cpustart"));
1810 need_gp = !!(rnp->qsmask & rdp->grpmask);
1811 rdp->cpu_no_qs.b.norm = need_gp;
1812 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1813 rdp->core_needs_qs = need_gp;
1814 zero_cpu_stall_ticks(rdp);
1816 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1817 if (ULONG_CMP_GE(rnp->gp_seq_needed, rdp->gp_seq_needed) || rdp->gpwrap)
1818 rdp->gp_seq_needed = rnp->gp_seq_needed;
1819 WRITE_ONCE(rdp->gpwrap, false);
1820 rcu_gpnum_ovf(rnp, rdp);
1824 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1826 unsigned long flags;
1828 struct rcu_node *rnp;
1830 local_irq_save(flags);
1832 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1833 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1834 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1835 local_irq_restore(flags);
1838 needwake = __note_gp_changes(rsp, rnp, rdp);
1839 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1841 rcu_gp_kthread_wake(rsp);
1844 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1847 !(rcu_seq_ctr(rsp->gp_seq) %
1848 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1849 schedule_timeout_uninterruptible(delay);
1853 * Initialize a new grace period. Return false if no grace period required.
1855 static bool rcu_gp_init(struct rcu_state *rsp)
1857 unsigned long flags;
1858 unsigned long oldmask;
1860 struct rcu_data *rdp;
1861 struct rcu_node *rnp = rcu_get_root(rsp);
1863 WRITE_ONCE(rsp->gp_activity, jiffies);
1864 raw_spin_lock_irq_rcu_node(rnp);
1865 if (!READ_ONCE(rsp->gp_flags)) {
1866 /* Spurious wakeup, tell caller to go back to sleep. */
1867 raw_spin_unlock_irq_rcu_node(rnp);
1870 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1872 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1874 * Grace period already in progress, don't start another.
1875 * Not supposed to be able to happen.
1877 raw_spin_unlock_irq_rcu_node(rnp);
1881 /* Advance to a new grace period and initialize state. */
1882 record_gp_stall_check_time(rsp);
1883 /* Record GP times before starting GP, hence rcu_seq_start(). */
1884 rcu_seq_start(&rsp->gp_seq);
1885 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("start"));
1886 raw_spin_unlock_irq_rcu_node(rnp);
1889 * Apply per-leaf buffered online and offline operations to the
1890 * rcu_node tree. Note that this new grace period need not wait
1891 * for subsequent online CPUs, and that quiescent-state forcing
1892 * will handle subsequent offline CPUs.
1894 rcu_for_each_leaf_node(rsp, rnp) {
1895 spin_lock(&rsp->ofl_lock);
1896 raw_spin_lock_irq_rcu_node(rnp);
1897 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1898 !rnp->wait_blkd_tasks) {
1899 /* Nothing to do on this leaf rcu_node structure. */
1900 raw_spin_unlock_irq_rcu_node(rnp);
1901 spin_unlock(&rsp->ofl_lock);
1905 /* Record old state, apply changes to ->qsmaskinit field. */
1906 oldmask = rnp->qsmaskinit;
1907 rnp->qsmaskinit = rnp->qsmaskinitnext;
1909 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1910 if (!oldmask != !rnp->qsmaskinit) {
1911 if (!oldmask) { /* First online CPU for rcu_node. */
1912 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1913 rcu_init_new_rnp(rnp);
1914 } else if (rcu_preempt_has_tasks(rnp)) {
1915 rnp->wait_blkd_tasks = true; /* blocked tasks */
1916 } else { /* Last offline CPU and can propagate. */
1917 rcu_cleanup_dead_rnp(rnp);
1922 * If all waited-on tasks from prior grace period are
1923 * done, and if all this rcu_node structure's CPUs are
1924 * still offline, propagate up the rcu_node tree and
1925 * clear ->wait_blkd_tasks. Otherwise, if one of this
1926 * rcu_node structure's CPUs has since come back online,
1927 * simply clear ->wait_blkd_tasks.
1929 if (rnp->wait_blkd_tasks &&
1930 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1931 rnp->wait_blkd_tasks = false;
1932 if (!rnp->qsmaskinit)
1933 rcu_cleanup_dead_rnp(rnp);
1936 raw_spin_unlock_irq_rcu_node(rnp);
1937 spin_unlock(&rsp->ofl_lock);
1939 rcu_gp_slow(rsp, gp_preinit_delay); /* Races with CPU hotplug. */
1942 * Set the quiescent-state-needed bits in all the rcu_node
1943 * structures for all currently online CPUs in breadth-first order,
1944 * starting from the root rcu_node structure, relying on the layout
1945 * of the tree within the rsp->node[] array. Note that other CPUs
1946 * will access only the leaves of the hierarchy, thus seeing that no
1947 * grace period is in progress, at least until the corresponding
1948 * leaf node has been initialized.
1950 * The grace period cannot complete until the initialization
1951 * process finishes, because this kthread handles both.
1953 rcu_for_each_node_breadth_first(rsp, rnp) {
1954 rcu_gp_slow(rsp, gp_init_delay);
1955 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1956 rdp = this_cpu_ptr(rsp->rda);
1957 rcu_preempt_check_blocked_tasks(rsp, rnp);
1958 rnp->qsmask = rnp->qsmaskinit;
1959 WRITE_ONCE(rnp->gp_seq, rsp->gp_seq);
1960 if (rnp == rdp->mynode)
1961 (void)__note_gp_changes(rsp, rnp, rdp);
1962 rcu_preempt_boost_start_gp(rnp);
1963 trace_rcu_grace_period_init(rsp->name, rnp->gp_seq,
1964 rnp->level, rnp->grplo,
1965 rnp->grphi, rnp->qsmask);
1966 /* Quiescent states for tasks on any now-offline CPUs. */
1967 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1968 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1969 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
1971 raw_spin_unlock_irq_rcu_node(rnp);
1972 cond_resched_tasks_rcu_qs();
1973 WRITE_ONCE(rsp->gp_activity, jiffies);
1980 * Helper function for swait_event_idle() wakeup at force-quiescent-state
1983 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1985 struct rcu_node *rnp = rcu_get_root(rsp);
1987 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1988 *gfp = READ_ONCE(rsp->gp_flags);
1989 if (*gfp & RCU_GP_FLAG_FQS)
1992 /* The current grace period has completed. */
1993 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2000 * Do one round of quiescent-state forcing.
2002 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2004 struct rcu_node *rnp = rcu_get_root(rsp);
2006 WRITE_ONCE(rsp->gp_activity, jiffies);
2009 /* Collect dyntick-idle snapshots. */
2010 force_qs_rnp(rsp, dyntick_save_progress_counter);
2012 /* Handle dyntick-idle and offline CPUs. */
2013 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2015 /* Clear flag to prevent immediate re-entry. */
2016 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2017 raw_spin_lock_irq_rcu_node(rnp);
2018 WRITE_ONCE(rsp->gp_flags,
2019 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2020 raw_spin_unlock_irq_rcu_node(rnp);
2025 * Clean up after the old grace period.
2027 static void rcu_gp_cleanup(struct rcu_state *rsp)
2029 unsigned long gp_duration;
2030 bool needgp = false;
2031 unsigned long new_gp_seq;
2032 struct rcu_data *rdp;
2033 struct rcu_node *rnp = rcu_get_root(rsp);
2034 struct swait_queue_head *sq;
2036 WRITE_ONCE(rsp->gp_activity, jiffies);
2037 raw_spin_lock_irq_rcu_node(rnp);
2038 gp_duration = jiffies - rsp->gp_start;
2039 if (gp_duration > rsp->gp_max)
2040 rsp->gp_max = gp_duration;
2043 * We know the grace period is complete, but to everyone else
2044 * it appears to still be ongoing. But it is also the case
2045 * that to everyone else it looks like there is nothing that
2046 * they can do to advance the grace period. It is therefore
2047 * safe for us to drop the lock in order to mark the grace
2048 * period as completed in all of the rcu_node structures.
2050 raw_spin_unlock_irq_rcu_node(rnp);
2053 * Propagate new ->gp_seq value to rcu_node structures so that
2054 * other CPUs don't have to wait until the start of the next grace
2055 * period to process their callbacks. This also avoids some nasty
2056 * RCU grace-period initialization races by forcing the end of
2057 * the current grace period to be completely recorded in all of
2058 * the rcu_node structures before the beginning of the next grace
2059 * period is recorded in any of the rcu_node structures.
2061 new_gp_seq = rsp->gp_seq;
2062 rcu_seq_end(&new_gp_seq);
2063 rcu_for_each_node_breadth_first(rsp, rnp) {
2064 raw_spin_lock_irq_rcu_node(rnp);
2065 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2066 dump_blkd_tasks(rsp, rnp, 10);
2067 WARN_ON_ONCE(rnp->qsmask);
2068 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2069 rdp = this_cpu_ptr(rsp->rda);
2070 if (rnp == rdp->mynode)
2071 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2072 /* smp_mb() provided by prior unlock-lock pair. */
2073 needgp = rcu_future_gp_cleanup(rsp, rnp) || needgp;
2074 sq = rcu_nocb_gp_get(rnp);
2075 raw_spin_unlock_irq_rcu_node(rnp);
2076 rcu_nocb_gp_cleanup(sq);
2077 cond_resched_tasks_rcu_qs();
2078 WRITE_ONCE(rsp->gp_activity, jiffies);
2079 rcu_gp_slow(rsp, gp_cleanup_delay);
2081 rnp = rcu_get_root(rsp);
2082 raw_spin_lock_irq_rcu_node(rnp); /* GP before rsp->gp_seq update. */
2084 /* Declare grace period done. */
2085 rcu_seq_end(&rsp->gp_seq);
2086 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("end"));
2087 rsp->gp_state = RCU_GP_IDLE;
2088 /* Check for GP requests since above loop. */
2089 rdp = this_cpu_ptr(rsp->rda);
2090 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2091 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2092 TPS("CleanupMore"));
2095 /* Advance CBs to reduce false positives below. */
2096 if (!rcu_accelerate_cbs(rsp, rnp, rdp) && needgp) {
2097 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2098 rsp->gp_req_activity = jiffies;
2099 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq),
2102 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags & RCU_GP_FLAG_INIT);
2104 raw_spin_unlock_irq_rcu_node(rnp);
2108 * Body of kthread that handles grace periods.
2110 static int __noreturn rcu_gp_kthread(void *arg)
2116 struct rcu_state *rsp = arg;
2117 struct rcu_node *rnp = rcu_get_root(rsp);
2119 rcu_bind_gp_kthread();
2122 /* Handle grace-period start. */
2124 trace_rcu_grace_period(rsp->name,
2125 READ_ONCE(rsp->gp_seq),
2127 rsp->gp_state = RCU_GP_WAIT_GPS;
2128 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2130 rsp->gp_state = RCU_GP_DONE_GPS;
2131 /* Locking provides needed memory barrier. */
2132 if (rcu_gp_init(rsp))
2134 cond_resched_tasks_rcu_qs();
2135 WRITE_ONCE(rsp->gp_activity, jiffies);
2136 WARN_ON(signal_pending(current));
2137 trace_rcu_grace_period(rsp->name,
2138 READ_ONCE(rsp->gp_seq),
2142 /* Handle quiescent-state forcing. */
2143 first_gp_fqs = true;
2144 j = jiffies_till_first_fqs;
2147 jiffies_till_first_fqs = HZ;
2152 rsp->jiffies_force_qs = jiffies + j;
2153 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2156 trace_rcu_grace_period(rsp->name,
2157 READ_ONCE(rsp->gp_seq),
2159 rsp->gp_state = RCU_GP_WAIT_FQS;
2160 ret = swait_event_idle_timeout(rsp->gp_wq,
2161 rcu_gp_fqs_check_wake(rsp, &gf), j);
2162 rsp->gp_state = RCU_GP_DOING_FQS;
2163 /* Locking provides needed memory barriers. */
2164 /* If grace period done, leave loop. */
2165 if (!READ_ONCE(rnp->qsmask) &&
2166 !rcu_preempt_blocked_readers_cgp(rnp))
2168 /* If time for quiescent-state forcing, do it. */
2169 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2170 (gf & RCU_GP_FLAG_FQS)) {
2171 trace_rcu_grace_period(rsp->name,
2172 READ_ONCE(rsp->gp_seq),
2174 rcu_gp_fqs(rsp, first_gp_fqs);
2175 first_gp_fqs = false;
2176 trace_rcu_grace_period(rsp->name,
2177 READ_ONCE(rsp->gp_seq),
2179 cond_resched_tasks_rcu_qs();
2180 WRITE_ONCE(rsp->gp_activity, jiffies);
2181 ret = 0; /* Force full wait till next FQS. */
2182 j = jiffies_till_next_fqs;
2185 jiffies_till_next_fqs = HZ;
2188 jiffies_till_next_fqs = 1;
2191 /* Deal with stray signal. */
2192 cond_resched_tasks_rcu_qs();
2193 WRITE_ONCE(rsp->gp_activity, jiffies);
2194 WARN_ON(signal_pending(current));
2195 trace_rcu_grace_period(rsp->name,
2196 READ_ONCE(rsp->gp_seq),
2198 ret = 1; /* Keep old FQS timing. */
2200 if (time_after(jiffies, rsp->jiffies_force_qs))
2203 j = rsp->jiffies_force_qs - j;
2207 /* Handle grace-period end. */
2208 rsp->gp_state = RCU_GP_CLEANUP;
2209 rcu_gp_cleanup(rsp);
2210 rsp->gp_state = RCU_GP_CLEANED;
2215 * Report a full set of quiescent states to the specified rcu_state data
2216 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2217 * kthread if another grace period is required. Whether we wake
2218 * the grace-period kthread or it awakens itself for the next round
2219 * of quiescent-state forcing, that kthread will clean up after the
2220 * just-completed grace period. Note that the caller must hold rnp->lock,
2221 * which is released before return.
2223 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2224 __releases(rcu_get_root(rsp)->lock)
2226 raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp));
2227 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2228 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2229 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2230 rcu_gp_kthread_wake(rsp);
2234 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2235 * Allows quiescent states for a group of CPUs to be reported at one go
2236 * to the specified rcu_node structure, though all the CPUs in the group
2237 * must be represented by the same rcu_node structure (which need not be a
2238 * leaf rcu_node structure, though it often will be). The gps parameter
2239 * is the grace-period snapshot, which means that the quiescent states
2240 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2241 * must be held upon entry, and it is released before return.
2243 * As a special case, if mask is zero, the bit-already-cleared check is
2244 * disabled. This allows propagating quiescent state due to resumed tasks
2245 * during grace-period initialization.
2248 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2249 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2250 __releases(rnp->lock)
2252 unsigned long oldmask = 0;
2253 struct rcu_node *rnp_c;
2255 raw_lockdep_assert_held_rcu_node(rnp);
2257 /* Walk up the rcu_node hierarchy. */
2259 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2262 * Our bit has already been cleared, or the
2263 * relevant grace period is already over, so done.
2265 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2268 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2269 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2270 rcu_preempt_blocked_readers_cgp(rnp));
2271 rnp->qsmask &= ~mask;
2272 trace_rcu_quiescent_state_report(rsp->name, rnp->gp_seq,
2273 mask, rnp->qsmask, rnp->level,
2274 rnp->grplo, rnp->grphi,
2276 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2278 /* Other bits still set at this level, so done. */
2279 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2282 rnp->completedqs = rnp->gp_seq;
2283 mask = rnp->grpmask;
2284 if (rnp->parent == NULL) {
2286 /* No more levels. Exit loop holding root lock. */
2290 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2293 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2294 oldmask = rnp_c->qsmask;
2298 * Get here if we are the last CPU to pass through a quiescent
2299 * state for this grace period. Invoke rcu_report_qs_rsp()
2300 * to clean up and start the next grace period if one is needed.
2302 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2306 * Record a quiescent state for all tasks that were previously queued
2307 * on the specified rcu_node structure and that were blocking the current
2308 * RCU grace period. The caller must hold the specified rnp->lock with
2309 * irqs disabled, and this lock is released upon return, but irqs remain
2312 static void __maybe_unused
2313 rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2314 struct rcu_node *rnp, unsigned long flags)
2315 __releases(rnp->lock)
2319 struct rcu_node *rnp_p;
2321 raw_lockdep_assert_held_rcu_node(rnp);
2322 if (WARN_ON_ONCE(rcu_state_p == &rcu_sched_state) ||
2323 WARN_ON_ONCE(rsp != rcu_state_p) ||
2324 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2326 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2327 return; /* Still need more quiescent states! */
2330 rnp->completedqs = rnp->gp_seq;
2331 rnp_p = rnp->parent;
2332 if (rnp_p == NULL) {
2334 * Only one rcu_node structure in the tree, so don't
2335 * try to report up to its nonexistent parent!
2337 rcu_report_qs_rsp(rsp, flags);
2341 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2343 mask = rnp->grpmask;
2344 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2345 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2346 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2350 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2351 * structure. This must be called from the specified CPU.
2354 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2356 unsigned long flags;
2359 struct rcu_node *rnp;
2362 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2363 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2367 * The grace period in which this quiescent state was
2368 * recorded has ended, so don't report it upwards.
2369 * We will instead need a new quiescent state that lies
2370 * within the current grace period.
2372 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2373 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2374 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2377 mask = rdp->grpmask;
2378 if ((rnp->qsmask & mask) == 0) {
2379 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2381 rdp->core_needs_qs = false;
2384 * This GP can't end until cpu checks in, so all of our
2385 * callbacks can be processed during the next GP.
2387 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2389 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2390 /* ^^^ Released rnp->lock */
2392 rcu_gp_kthread_wake(rsp);
2397 * Check to see if there is a new grace period of which this CPU
2398 * is not yet aware, and if so, set up local rcu_data state for it.
2399 * Otherwise, see if this CPU has just passed through its first
2400 * quiescent state for this grace period, and record that fact if so.
2403 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2405 /* Check for grace-period ends and beginnings. */
2406 note_gp_changes(rsp, rdp);
2409 * Does this CPU still need to do its part for current grace period?
2410 * If no, return and let the other CPUs do their part as well.
2412 if (!rdp->core_needs_qs)
2416 * Was there a quiescent state since the beginning of the grace
2417 * period? If no, then exit and wait for the next call.
2419 if (rdp->cpu_no_qs.b.norm)
2423 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2426 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2430 * Trace the fact that this CPU is going offline.
2432 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2434 RCU_TRACE(bool blkd;)
2435 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2436 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2438 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2441 RCU_TRACE(blkd = !!(rnp->qsmask & rdp->grpmask);)
2442 trace_rcu_grace_period(rsp->name, rnp->gp_seq,
2443 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2447 * All CPUs for the specified rcu_node structure have gone offline,
2448 * and all tasks that were preempted within an RCU read-side critical
2449 * section while running on one of those CPUs have since exited their RCU
2450 * read-side critical section. Some other CPU is reporting this fact with
2451 * the specified rcu_node structure's ->lock held and interrupts disabled.
2452 * This function therefore goes up the tree of rcu_node structures,
2453 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2454 * the leaf rcu_node structure's ->qsmaskinit field has already been
2457 * This function does check that the specified rcu_node structure has
2458 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2459 * prematurely. That said, invoking it after the fact will cost you
2460 * a needless lock acquisition. So once it has done its work, don't
2463 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2466 struct rcu_node *rnp = rnp_leaf;
2468 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2469 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2470 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2471 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2474 mask = rnp->grpmask;
2478 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2479 rnp->qsmaskinit &= ~mask;
2480 /* Between grace periods, so better already be zero! */
2481 WARN_ON_ONCE(rnp->qsmask);
2482 if (rnp->qsmaskinit) {
2483 raw_spin_unlock_rcu_node(rnp);
2484 /* irqs remain disabled. */
2487 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2492 * The CPU has been completely removed, and some other CPU is reporting
2493 * this fact from process context. Do the remainder of the cleanup.
2494 * There can only be one CPU hotplug operation at a time, so no need for
2497 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2499 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2500 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2502 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2505 /* Adjust any no-longer-needed kthreads. */
2506 rcu_boost_kthread_setaffinity(rnp, -1);
2510 * Invoke any RCU callbacks that have made it to the end of their grace
2511 * period. Thottle as specified by rdp->blimit.
2513 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2515 unsigned long flags;
2516 struct rcu_head *rhp;
2517 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2520 /* If no callbacks are ready, just return. */
2521 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2522 trace_rcu_batch_start(rsp->name,
2523 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2524 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2525 trace_rcu_batch_end(rsp->name, 0,
2526 !rcu_segcblist_empty(&rdp->cblist),
2527 need_resched(), is_idle_task(current),
2528 rcu_is_callbacks_kthread());
2533 * Extract the list of ready callbacks, disabling to prevent
2534 * races with call_rcu() from interrupt handlers. Leave the
2535 * callback counts, as rcu_barrier() needs to be conservative.
2537 local_irq_save(flags);
2538 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2540 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2541 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2542 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2543 local_irq_restore(flags);
2545 /* Invoke callbacks. */
2546 rhp = rcu_cblist_dequeue(&rcl);
2547 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2548 debug_rcu_head_unqueue(rhp);
2549 if (__rcu_reclaim(rsp->name, rhp))
2550 rcu_cblist_dequeued_lazy(&rcl);
2552 * Stop only if limit reached and CPU has something to do.
2553 * Note: The rcl structure counts down from zero.
2555 if (-rcl.len >= bl &&
2557 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2561 local_irq_save(flags);
2563 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2564 is_idle_task(current), rcu_is_callbacks_kthread());
2566 /* Update counts and requeue any remaining callbacks. */
2567 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2568 smp_mb(); /* List handling before counting for rcu_barrier(). */
2569 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2571 /* Reinstate batch limit if we have worked down the excess. */
2572 count = rcu_segcblist_n_cbs(&rdp->cblist);
2573 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2574 rdp->blimit = blimit;
2576 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2577 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2578 rdp->qlen_last_fqs_check = 0;
2579 rdp->n_force_qs_snap = rsp->n_force_qs;
2580 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2581 rdp->qlen_last_fqs_check = count;
2584 * The following usually indicates a double call_rcu(). To track
2585 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2587 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2589 local_irq_restore(flags);
2591 /* Re-invoke RCU core processing if there are callbacks remaining. */
2592 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2597 * Check to see if this CPU is in a non-context-switch quiescent state
2598 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2599 * Also schedule RCU core processing.
2601 * This function must be called from hardirq context. It is normally
2602 * invoked from the scheduling-clock interrupt.
2604 void rcu_check_callbacks(int user)
2606 trace_rcu_utilization(TPS("Start scheduler-tick"));
2607 increment_cpu_stall_ticks();
2608 if (user || rcu_is_cpu_rrupt_from_idle()) {
2611 * Get here if this CPU took its interrupt from user
2612 * mode or from the idle loop, and if this is not a
2613 * nested interrupt. In this case, the CPU is in
2614 * a quiescent state, so note it.
2616 * No memory barrier is required here because both
2617 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2618 * variables that other CPUs neither access nor modify,
2619 * at least not while the corresponding CPU is online.
2625 } else if (!in_softirq()) {
2628 * Get here if this CPU did not take its interrupt from
2629 * softirq, in other words, if it is not interrupting
2630 * a rcu_bh read-side critical section. This is an _bh
2631 * critical section, so note it.
2636 rcu_preempt_check_callbacks();
2640 rcu_note_voluntary_context_switch(current);
2641 trace_rcu_utilization(TPS("End scheduler-tick"));
2645 * Scan the leaf rcu_node structures, processing dyntick state for any that
2646 * have not yet encountered a quiescent state, using the function specified.
2647 * Also initiate boosting for any threads blocked on the root rcu_node.
2649 * The caller must have suppressed start of new grace periods.
2651 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2654 unsigned long flags;
2656 struct rcu_node *rnp;
2658 rcu_for_each_leaf_node(rsp, rnp) {
2659 cond_resched_tasks_rcu_qs();
2661 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2662 if (rnp->qsmask == 0) {
2663 if (rcu_state_p == &rcu_sched_state ||
2664 rsp != rcu_state_p ||
2665 rcu_preempt_blocked_readers_cgp(rnp)) {
2667 * No point in scanning bits because they
2668 * are all zero. But we might need to
2669 * priority-boost blocked readers.
2671 rcu_initiate_boost(rnp, flags);
2672 /* rcu_initiate_boost() releases rnp->lock */
2675 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2678 for_each_leaf_node_possible_cpu(rnp, cpu) {
2679 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2680 if ((rnp->qsmask & bit) != 0) {
2681 if (f(per_cpu_ptr(rsp->rda, cpu)))
2686 /* Idle/offline CPUs, report (releases rnp->lock). */
2687 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2689 /* Nothing to do here, so just drop the lock. */
2690 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2696 * Force quiescent states on reluctant CPUs, and also detect which
2697 * CPUs are in dyntick-idle mode.
2699 static void force_quiescent_state(struct rcu_state *rsp)
2701 unsigned long flags;
2703 struct rcu_node *rnp;
2704 struct rcu_node *rnp_old = NULL;
2706 /* Funnel through hierarchy to reduce memory contention. */
2707 rnp = __this_cpu_read(rsp->rda->mynode);
2708 for (; rnp != NULL; rnp = rnp->parent) {
2709 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2710 !raw_spin_trylock(&rnp->fqslock);
2711 if (rnp_old != NULL)
2712 raw_spin_unlock(&rnp_old->fqslock);
2717 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2719 /* Reached the root of the rcu_node tree, acquire lock. */
2720 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2721 raw_spin_unlock(&rnp_old->fqslock);
2722 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2723 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2724 return; /* Someone beat us to it. */
2726 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2727 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2728 rcu_gp_kthread_wake(rsp);
2732 * This function checks for grace-period requests that fail to motivate
2733 * RCU to come out of its idle mode.
2736 rcu_check_gp_start_stall(struct rcu_state *rsp, struct rcu_node *rnp,
2737 struct rcu_data *rdp)
2739 unsigned long flags;
2741 struct rcu_node *rnp_root = rcu_get_root(rsp);
2742 static atomic_t warned = ATOMIC_INIT(0);
2744 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress(rsp) ||
2745 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed))
2747 j = jiffies; /* Expensive access, and in common case don't get here. */
2748 if (time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2749 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2750 atomic_read(&warned))
2753 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2755 if (rcu_gp_in_progress(rsp) ||
2756 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2757 time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2758 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2759 atomic_read(&warned)) {
2760 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2763 /* Hold onto the leaf lock to make others see warned==1. */
2765 if (rnp_root != rnp)
2766 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
2768 if (rcu_gp_in_progress(rsp) ||
2769 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2770 time_before(j, rsp->gp_req_activity + HZ) ||
2771 time_before(j, rsp->gp_activity + HZ) ||
2772 atomic_xchg(&warned, 1)) {
2773 raw_spin_unlock_rcu_node(rnp_root); /* irqs remain disabled. */
2774 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2777 pr_alert("%s: g%ld->%ld gar:%lu ga:%lu f%#x %s->state:%#lx\n",
2778 __func__, (long)READ_ONCE(rsp->gp_seq),
2779 (long)READ_ONCE(rnp_root->gp_seq_needed),
2780 j - rsp->gp_req_activity, j - rsp->gp_activity,
2781 rsp->gp_flags, rsp->name,
2782 rsp->gp_kthread ? rsp->gp_kthread->state : 0x1ffffL);
2784 if (rnp_root != rnp)
2785 raw_spin_unlock_rcu_node(rnp_root);
2786 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2790 * This does the RCU core processing work for the specified rcu_state
2791 * and rcu_data structures. This may be called only from the CPU to
2792 * whom the rdp belongs.
2795 __rcu_process_callbacks(struct rcu_state *rsp)
2797 unsigned long flags;
2798 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2799 struct rcu_node *rnp = rdp->mynode;
2801 WARN_ON_ONCE(!rdp->beenonline);
2803 /* Update RCU state based on any recent quiescent states. */
2804 rcu_check_quiescent_state(rsp, rdp);
2806 /* No grace period and unregistered callbacks? */
2807 if (!rcu_gp_in_progress(rsp) &&
2808 rcu_segcblist_is_enabled(&rdp->cblist)) {
2809 local_irq_save(flags);
2810 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2811 rcu_accelerate_cbs_unlocked(rsp, rnp, rdp);
2812 local_irq_restore(flags);
2815 rcu_check_gp_start_stall(rsp, rnp, rdp);
2817 /* If there are callbacks ready, invoke them. */
2818 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2819 invoke_rcu_callbacks(rsp, rdp);
2821 /* Do any needed deferred wakeups of rcuo kthreads. */
2822 do_nocb_deferred_wakeup(rdp);
2826 * Do RCU core processing for the current CPU.
2828 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2830 struct rcu_state *rsp;
2832 if (cpu_is_offline(smp_processor_id()))
2834 trace_rcu_utilization(TPS("Start RCU core"));
2835 for_each_rcu_flavor(rsp)
2836 __rcu_process_callbacks(rsp);
2837 trace_rcu_utilization(TPS("End RCU core"));
2841 * Schedule RCU callback invocation. If the specified type of RCU
2842 * does not support RCU priority boosting, just do a direct call,
2843 * otherwise wake up the per-CPU kernel kthread. Note that because we
2844 * are running on the current CPU with softirqs disabled, the
2845 * rcu_cpu_kthread_task cannot disappear out from under us.
2847 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2849 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2851 if (likely(!rsp->boost)) {
2852 rcu_do_batch(rsp, rdp);
2855 invoke_rcu_callbacks_kthread();
2858 static void invoke_rcu_core(void)
2860 if (cpu_online(smp_processor_id()))
2861 raise_softirq(RCU_SOFTIRQ);
2865 * Handle any core-RCU processing required by a call_rcu() invocation.
2867 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2868 struct rcu_head *head, unsigned long flags)
2871 * If called from an extended quiescent state, invoke the RCU
2872 * core in order to force a re-evaluation of RCU's idleness.
2874 if (!rcu_is_watching())
2877 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2878 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2882 * Force the grace period if too many callbacks or too long waiting.
2883 * Enforce hysteresis, and don't invoke force_quiescent_state()
2884 * if some other CPU has recently done so. Also, don't bother
2885 * invoking force_quiescent_state() if the newly enqueued callback
2886 * is the only one waiting for a grace period to complete.
2888 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2889 rdp->qlen_last_fqs_check + qhimark)) {
2891 /* Are we ignoring a completed grace period? */
2892 note_gp_changes(rsp, rdp);
2894 /* Start a new grace period if one not already started. */
2895 if (!rcu_gp_in_progress(rsp)) {
2896 rcu_accelerate_cbs_unlocked(rsp, rdp->mynode, rdp);
2898 /* Give the grace period a kick. */
2899 rdp->blimit = LONG_MAX;
2900 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2901 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2902 force_quiescent_state(rsp);
2903 rdp->n_force_qs_snap = rsp->n_force_qs;
2904 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2910 * RCU callback function to leak a callback.
2912 static void rcu_leak_callback(struct rcu_head *rhp)
2917 * Helper function for call_rcu() and friends. The cpu argument will
2918 * normally be -1, indicating "currently running CPU". It may specify
2919 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2920 * is expected to specify a CPU.
2923 __call_rcu(struct rcu_head *head, rcu_callback_t func,
2924 struct rcu_state *rsp, int cpu, bool lazy)
2926 unsigned long flags;
2927 struct rcu_data *rdp;
2929 /* Misaligned rcu_head! */
2930 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2932 if (debug_rcu_head_queue(head)) {
2934 * Probable double call_rcu(), so leak the callback.
2935 * Use rcu:rcu_callback trace event to find the previous
2936 * time callback was passed to __call_rcu().
2938 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
2940 WRITE_ONCE(head->func, rcu_leak_callback);
2945 local_irq_save(flags);
2946 rdp = this_cpu_ptr(rsp->rda);
2948 /* Add the callback to our list. */
2949 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
2953 rdp = per_cpu_ptr(rsp->rda, cpu);
2954 if (likely(rdp->mynode)) {
2955 /* Post-boot, so this should be for a no-CBs CPU. */
2956 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2957 WARN_ON_ONCE(offline);
2958 /* Offline CPU, _call_rcu() illegal, leak callback. */
2959 local_irq_restore(flags);
2963 * Very early boot, before rcu_init(). Initialize if needed
2964 * and then drop through to queue the callback.
2967 WARN_ON_ONCE(!rcu_is_watching());
2968 if (rcu_segcblist_empty(&rdp->cblist))
2969 rcu_segcblist_init(&rdp->cblist);
2971 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2973 rcu_idle_count_callbacks_posted();
2975 if (__is_kfree_rcu_offset((unsigned long)func))
2976 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2977 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2978 rcu_segcblist_n_cbs(&rdp->cblist));
2980 trace_rcu_callback(rsp->name, head,
2981 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2982 rcu_segcblist_n_cbs(&rdp->cblist));
2984 /* Go handle any RCU core processing required. */
2985 __call_rcu_core(rsp, rdp, head, flags);
2986 local_irq_restore(flags);
2990 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
2991 * @head: structure to be used for queueing the RCU updates.
2992 * @func: actual callback function to be invoked after the grace period
2994 * The callback function will be invoked some time after a full grace
2995 * period elapses, in other words after all currently executing RCU
2996 * read-side critical sections have completed. call_rcu_sched() assumes
2997 * that the read-side critical sections end on enabling of preemption
2998 * or on voluntary preemption.
2999 * RCU read-side critical sections are delimited by:
3001 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3002 * - anything that disables preemption.
3004 * These may be nested.
3006 * See the description of call_rcu() for more detailed information on
3007 * memory ordering guarantees.
3009 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3011 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3013 EXPORT_SYMBOL_GPL(call_rcu_sched);
3016 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3017 * @head: structure to be used for queueing the RCU updates.
3018 * @func: actual callback function to be invoked after the grace period
3020 * The callback function will be invoked some time after a full grace
3021 * period elapses, in other words after all currently executing RCU
3022 * read-side critical sections have completed. call_rcu_bh() assumes
3023 * that the read-side critical sections end on completion of a softirq
3024 * handler. This means that read-side critical sections in process
3025 * context must not be interrupted by softirqs. This interface is to be
3026 * used when most of the read-side critical sections are in softirq context.
3027 * RCU read-side critical sections are delimited by:
3029 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3030 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3032 * These may be nested.
3034 * See the description of call_rcu() for more detailed information on
3035 * memory ordering guarantees.
3037 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3039 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3041 EXPORT_SYMBOL_GPL(call_rcu_bh);
3044 * Queue an RCU callback for lazy invocation after a grace period.
3045 * This will likely be later named something like "call_rcu_lazy()",
3046 * but this change will require some way of tagging the lazy RCU
3047 * callbacks in the list of pending callbacks. Until then, this
3048 * function may only be called from __kfree_rcu().
3050 void kfree_call_rcu(struct rcu_head *head,
3051 rcu_callback_t func)
3053 __call_rcu(head, func, rcu_state_p, -1, 1);
3055 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3058 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3059 * any blocking grace-period wait automatically implies a grace period
3060 * if there is only one CPU online at any point time during execution
3061 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3062 * occasionally incorrectly indicate that there are multiple CPUs online
3063 * when there was in fact only one the whole time, as this just adds
3064 * some overhead: RCU still operates correctly.
3066 static inline int rcu_blocking_is_gp(void)
3070 might_sleep(); /* Check for RCU read-side critical section. */
3072 ret = num_online_cpus() <= 1;
3078 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3080 * Control will return to the caller some time after a full rcu-sched
3081 * grace period has elapsed, in other words after all currently executing
3082 * rcu-sched read-side critical sections have completed. These read-side
3083 * critical sections are delimited by rcu_read_lock_sched() and
3084 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3085 * local_irq_disable(), and so on may be used in place of
3086 * rcu_read_lock_sched().
3088 * This means that all preempt_disable code sequences, including NMI and
3089 * non-threaded hardware-interrupt handlers, in progress on entry will
3090 * have completed before this primitive returns. However, this does not
3091 * guarantee that softirq handlers will have completed, since in some
3092 * kernels, these handlers can run in process context, and can block.
3094 * Note that this guarantee implies further memory-ordering guarantees.
3095 * On systems with more than one CPU, when synchronize_sched() returns,
3096 * each CPU is guaranteed to have executed a full memory barrier since the
3097 * end of its last RCU-sched read-side critical section whose beginning
3098 * preceded the call to synchronize_sched(). In addition, each CPU having
3099 * an RCU read-side critical section that extends beyond the return from
3100 * synchronize_sched() is guaranteed to have executed a full memory barrier
3101 * after the beginning of synchronize_sched() and before the beginning of
3102 * that RCU read-side critical section. Note that these guarantees include
3103 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3104 * that are executing in the kernel.
3106 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3107 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3108 * to have executed a full memory barrier during the execution of
3109 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3110 * again only if the system has more than one CPU).
3112 void synchronize_sched(void)
3114 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3115 lock_is_held(&rcu_lock_map) ||
3116 lock_is_held(&rcu_sched_lock_map),
3117 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3118 if (rcu_blocking_is_gp())
3120 if (rcu_gp_is_expedited())
3121 synchronize_sched_expedited();
3123 wait_rcu_gp(call_rcu_sched);
3125 EXPORT_SYMBOL_GPL(synchronize_sched);
3128 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3130 * Control will return to the caller some time after a full rcu_bh grace
3131 * period has elapsed, in other words after all currently executing rcu_bh
3132 * read-side critical sections have completed. RCU read-side critical
3133 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3134 * and may be nested.
3136 * See the description of synchronize_sched() for more detailed information
3137 * on memory ordering guarantees.
3139 void synchronize_rcu_bh(void)
3141 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3142 lock_is_held(&rcu_lock_map) ||
3143 lock_is_held(&rcu_sched_lock_map),
3144 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3145 if (rcu_blocking_is_gp())
3147 if (rcu_gp_is_expedited())
3148 synchronize_rcu_bh_expedited();
3150 wait_rcu_gp(call_rcu_bh);
3152 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3155 * get_state_synchronize_rcu - Snapshot current RCU state
3157 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3158 * to determine whether or not a full grace period has elapsed in the
3161 unsigned long get_state_synchronize_rcu(void)
3164 * Any prior manipulation of RCU-protected data must happen
3165 * before the load from ->gp_seq.
3168 return rcu_seq_snap(&rcu_state_p->gp_seq);
3170 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3173 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3175 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3177 * If a full RCU grace period has elapsed since the earlier call to
3178 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3179 * synchronize_rcu() to wait for a full grace period.
3181 * Yes, this function does not take counter wrap into account. But
3182 * counter wrap is harmless. If the counter wraps, we have waited for
3183 * more than 2 billion grace periods (and way more on a 64-bit system!),
3184 * so waiting for one additional grace period should be just fine.
3186 void cond_synchronize_rcu(unsigned long oldstate)
3188 if (!rcu_seq_done(&rcu_state_p->gp_seq, oldstate))
3191 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3193 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3196 * get_state_synchronize_sched - Snapshot current RCU-sched state
3198 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3199 * to determine whether or not a full grace period has elapsed in the
3202 unsigned long get_state_synchronize_sched(void)
3205 * Any prior manipulation of RCU-protected data must happen
3206 * before the load from ->gp_seq.
3209 return rcu_seq_snap(&rcu_sched_state.gp_seq);
3211 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3214 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3216 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3218 * If a full RCU-sched grace period has elapsed since the earlier call to
3219 * get_state_synchronize_sched(), just return. Otherwise, invoke
3220 * synchronize_sched() to wait for a full grace period.
3222 * Yes, this function does not take counter wrap into account. But
3223 * counter wrap is harmless. If the counter wraps, we have waited for
3224 * more than 2 billion grace periods (and way more on a 64-bit system!),
3225 * so waiting for one additional grace period should be just fine.
3227 void cond_synchronize_sched(unsigned long oldstate)
3229 if (!rcu_seq_done(&rcu_sched_state.gp_seq, oldstate))
3230 synchronize_sched();
3232 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3234 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3237 * Check to see if there is any immediate RCU-related work to be done
3238 * by the current CPU, for the specified type of RCU, returning 1 if so.
3239 * The checks are in order of increasing expense: checks that can be
3240 * carried out against CPU-local state are performed first. However,
3241 * we must check for CPU stalls first, else we might not get a chance.
3243 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3245 struct rcu_node *rnp = rdp->mynode;
3247 /* Check for CPU stalls, if enabled. */
3248 check_cpu_stall(rsp, rdp);
3250 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3251 if (rcu_nohz_full_cpu(rsp))
3254 /* Is the RCU core waiting for a quiescent state from this CPU? */
3255 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
3258 /* Does this CPU have callbacks ready to invoke? */
3259 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3262 /* Has RCU gone idle with this CPU needing another grace period? */
3263 if (!rcu_gp_in_progress(rsp) &&
3264 rcu_segcblist_is_enabled(&rdp->cblist) &&
3265 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3268 /* Have RCU grace period completed or started? */
3269 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3270 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3273 /* Does this CPU need a deferred NOCB wakeup? */
3274 if (rcu_nocb_need_deferred_wakeup(rdp))
3282 * Check to see if there is any immediate RCU-related work to be done
3283 * by the current CPU, returning 1 if so. This function is part of the
3284 * RCU implementation; it is -not- an exported member of the RCU API.
3286 static int rcu_pending(void)
3288 struct rcu_state *rsp;
3290 for_each_rcu_flavor(rsp)
3291 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3297 * Return true if the specified CPU has any callback. If all_lazy is
3298 * non-NULL, store an indication of whether all callbacks are lazy.
3299 * (If there are no callbacks, all of them are deemed to be lazy.)
3301 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3305 struct rcu_data *rdp;
3306 struct rcu_state *rsp;
3308 for_each_rcu_flavor(rsp) {
3309 rdp = this_cpu_ptr(rsp->rda);
3310 if (rcu_segcblist_empty(&rdp->cblist))
3313 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3324 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3325 * the compiler is expected to optimize this away.
3327 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3328 int cpu, unsigned long done)
3330 trace_rcu_barrier(rsp->name, s, cpu,
3331 atomic_read(&rsp->barrier_cpu_count), done);
3335 * RCU callback function for _rcu_barrier(). If we are last, wake
3336 * up the task executing _rcu_barrier().
3338 static void rcu_barrier_callback(struct rcu_head *rhp)
3340 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3341 struct rcu_state *rsp = rdp->rsp;
3343 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3344 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3345 rsp->barrier_sequence);
3346 complete(&rsp->barrier_completion);
3348 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3353 * Called with preemption disabled, and from cross-cpu IRQ context.
3355 static void rcu_barrier_func(void *type)
3357 struct rcu_state *rsp = type;
3358 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3360 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3361 rdp->barrier_head.func = rcu_barrier_callback;
3362 debug_rcu_head_queue(&rdp->barrier_head);
3363 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3364 atomic_inc(&rsp->barrier_cpu_count);
3366 debug_rcu_head_unqueue(&rdp->barrier_head);
3367 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3368 rsp->barrier_sequence);
3373 * Orchestrate the specified type of RCU barrier, waiting for all
3374 * RCU callbacks of the specified type to complete.
3376 static void _rcu_barrier(struct rcu_state *rsp)
3379 struct rcu_data *rdp;
3380 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3382 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3384 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3385 mutex_lock(&rsp->barrier_mutex);
3387 /* Did someone else do our work for us? */
3388 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3389 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3390 rsp->barrier_sequence);
3391 smp_mb(); /* caller's subsequent code after above check. */
3392 mutex_unlock(&rsp->barrier_mutex);
3396 /* Mark the start of the barrier operation. */
3397 rcu_seq_start(&rsp->barrier_sequence);
3398 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3401 * Initialize the count to one rather than to zero in order to
3402 * avoid a too-soon return to zero in case of a short grace period
3403 * (or preemption of this task). Exclude CPU-hotplug operations
3404 * to ensure that no offline CPU has callbacks queued.
3406 init_completion(&rsp->barrier_completion);
3407 atomic_set(&rsp->barrier_cpu_count, 1);
3411 * Force each CPU with callbacks to register a new callback.
3412 * When that callback is invoked, we will know that all of the
3413 * corresponding CPU's preceding callbacks have been invoked.
3415 for_each_possible_cpu(cpu) {
3416 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3418 rdp = per_cpu_ptr(rsp->rda, cpu);
3419 if (rcu_is_nocb_cpu(cpu)) {
3420 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3421 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3422 rsp->barrier_sequence);
3424 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3425 rsp->barrier_sequence);
3426 smp_mb__before_atomic();
3427 atomic_inc(&rsp->barrier_cpu_count);
3428 __call_rcu(&rdp->barrier_head,
3429 rcu_barrier_callback, rsp, cpu, 0);
3431 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3432 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3433 rsp->barrier_sequence);
3434 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3436 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3437 rsp->barrier_sequence);
3443 * Now that we have an rcu_barrier_callback() callback on each
3444 * CPU, and thus each counted, remove the initial count.
3446 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3447 complete(&rsp->barrier_completion);
3449 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3450 wait_for_completion(&rsp->barrier_completion);
3452 /* Mark the end of the barrier operation. */
3453 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3454 rcu_seq_end(&rsp->barrier_sequence);
3456 /* Other rcu_barrier() invocations can now safely proceed. */
3457 mutex_unlock(&rsp->barrier_mutex);
3461 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3463 void rcu_barrier_bh(void)
3465 _rcu_barrier(&rcu_bh_state);
3467 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3470 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3472 void rcu_barrier_sched(void)
3474 _rcu_barrier(&rcu_sched_state);
3476 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3479 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3480 * first CPU in a given leaf rcu_node structure coming online. The caller
3481 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3484 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3488 struct rcu_node *rnp = rnp_leaf;
3490 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3491 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3493 mask = rnp->grpmask;
3497 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3498 oldmask = rnp->qsmaskinit;
3499 rnp->qsmaskinit |= mask;
3500 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3507 * Do boot-time initialization of a CPU's per-CPU RCU data.
3510 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3512 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3514 /* Set up local state, ensuring consistent view of global state. */
3515 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3516 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3517 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3518 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3519 rdp->rcu_ofl_gp_seq = rsp->gp_seq;
3520 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3521 rdp->rcu_onl_gp_seq = rsp->gp_seq;
3522 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3525 rcu_boot_init_nocb_percpu_data(rdp);
3529 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3530 * offline event can be happening at a given time. Note also that we can
3531 * accept some slop in the rsp->gp_seq access due to the fact that this
3532 * CPU cannot possibly have any RCU callbacks in flight yet.
3535 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3537 unsigned long flags;
3538 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3539 struct rcu_node *rnp = rcu_get_root(rsp);
3541 /* Set up local state, ensuring consistent view of global state. */
3542 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3543 rdp->qlen_last_fqs_check = 0;
3544 rdp->n_force_qs_snap = rsp->n_force_qs;
3545 rdp->blimit = blimit;
3546 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3547 !init_nocb_callback_list(rdp))
3548 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3549 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3550 rcu_dynticks_eqs_online();
3551 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3554 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3555 * propagation up the rcu_node tree will happen at the beginning
3556 * of the next grace period.
3559 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3560 rdp->beenonline = true; /* We have now been online. */
3561 rdp->gp_seq = rnp->gp_seq;
3562 rdp->gp_seq_needed = rnp->gp_seq;
3563 rdp->cpu_no_qs.b.norm = true;
3564 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3565 rdp->core_needs_qs = false;
3566 rdp->rcu_iw_pending = false;
3567 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3568 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuonl"));
3569 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3573 * Invoked early in the CPU-online process, when pretty much all
3574 * services are available. The incoming CPU is not present.
3576 int rcutree_prepare_cpu(unsigned int cpu)
3578 struct rcu_state *rsp;
3580 for_each_rcu_flavor(rsp)
3581 rcu_init_percpu_data(cpu, rsp);
3583 rcu_prepare_kthreads(cpu);
3584 rcu_spawn_all_nocb_kthreads(cpu);
3590 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3592 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3594 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3596 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3600 * Near the end of the CPU-online process. Pretty much all services
3601 * enabled, and the CPU is now very much alive.
3603 int rcutree_online_cpu(unsigned int cpu)
3605 unsigned long flags;
3606 struct rcu_data *rdp;
3607 struct rcu_node *rnp;
3608 struct rcu_state *rsp;
3610 for_each_rcu_flavor(rsp) {
3611 rdp = per_cpu_ptr(rsp->rda, cpu);
3613 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3614 rnp->ffmask |= rdp->grpmask;
3615 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3617 if (IS_ENABLED(CONFIG_TREE_SRCU))
3618 srcu_online_cpu(cpu);
3619 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3620 return 0; /* Too early in boot for scheduler work. */
3621 sync_sched_exp_online_cleanup(cpu);
3622 rcutree_affinity_setting(cpu, -1);
3627 * Near the beginning of the process. The CPU is still very much alive
3628 * with pretty much all services enabled.
3630 int rcutree_offline_cpu(unsigned int cpu)
3632 unsigned long flags;
3633 struct rcu_data *rdp;
3634 struct rcu_node *rnp;
3635 struct rcu_state *rsp;
3637 for_each_rcu_flavor(rsp) {
3638 rdp = per_cpu_ptr(rsp->rda, cpu);
3640 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3641 rnp->ffmask &= ~rdp->grpmask;
3642 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3645 rcutree_affinity_setting(cpu, cpu);
3646 if (IS_ENABLED(CONFIG_TREE_SRCU))
3647 srcu_offline_cpu(cpu);
3652 * Near the end of the offline process. We do only tracing here.
3654 int rcutree_dying_cpu(unsigned int cpu)
3656 struct rcu_state *rsp;
3658 for_each_rcu_flavor(rsp)
3659 rcu_cleanup_dying_cpu(rsp);
3664 * The outgoing CPU is gone and we are running elsewhere.
3666 int rcutree_dead_cpu(unsigned int cpu)
3668 struct rcu_state *rsp;
3670 for_each_rcu_flavor(rsp) {
3671 rcu_cleanup_dead_cpu(cpu, rsp);
3672 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3677 static DEFINE_PER_CPU(int, rcu_cpu_started);
3680 * Mark the specified CPU as being online so that subsequent grace periods
3681 * (both expedited and normal) will wait on it. Note that this means that
3682 * incoming CPUs are not allowed to use RCU read-side critical sections
3683 * until this function is called. Failing to observe this restriction
3684 * will result in lockdep splats.
3686 * Note that this function is special in that it is invoked directly
3687 * from the incoming CPU rather than from the cpuhp_step mechanism.
3688 * This is because this function must be invoked at a precise location.
3690 void rcu_cpu_starting(unsigned int cpu)
3692 unsigned long flags;
3695 unsigned long oldmask;
3696 struct rcu_data *rdp;
3697 struct rcu_node *rnp;
3698 struct rcu_state *rsp;
3700 if (per_cpu(rcu_cpu_started, cpu))
3703 per_cpu(rcu_cpu_started, cpu) = 1;
3705 for_each_rcu_flavor(rsp) {
3706 rdp = per_cpu_ptr(rsp->rda, cpu);
3708 mask = rdp->grpmask;
3709 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3710 rnp->qsmaskinitnext |= mask;
3711 oldmask = rnp->expmaskinitnext;
3712 rnp->expmaskinitnext |= mask;
3713 oldmask ^= rnp->expmaskinitnext;
3714 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3715 /* Allow lockless access for expedited grace periods. */
3716 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3717 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3718 rdp->rcu_onl_gp_seq = READ_ONCE(rsp->gp_seq);
3719 rdp->rcu_onl_gp_flags = READ_ONCE(rsp->gp_flags);
3720 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3721 /* Report QS -after- changing ->qsmaskinitnext! */
3722 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
3724 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3727 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3730 #ifdef CONFIG_HOTPLUG_CPU
3732 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3733 * function. We now remove it from the rcu_node tree's ->qsmaskinitnext
3736 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3738 unsigned long flags;
3740 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3741 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3743 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3744 mask = rdp->grpmask;
3745 spin_lock(&rsp->ofl_lock);
3746 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3747 rdp->rcu_ofl_gp_seq = READ_ONCE(rsp->gp_seq);
3748 rdp->rcu_ofl_gp_flags = READ_ONCE(rsp->gp_flags);
3749 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3750 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3751 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
3752 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3754 rnp->qsmaskinitnext &= ~mask;
3755 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3756 spin_unlock(&rsp->ofl_lock);
3760 * The outgoing function has no further need of RCU, so remove it from
3761 * the list of CPUs that RCU must track.
3763 * Note that this function is special in that it is invoked directly
3764 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3765 * This is because this function must be invoked at a precise location.
3767 void rcu_report_dead(unsigned int cpu)
3769 struct rcu_state *rsp;
3771 /* QS for any half-done expedited RCU-sched GP. */
3773 rcu_report_exp_rdp(&rcu_sched_state,
3774 this_cpu_ptr(rcu_sched_state.rda), true);
3776 for_each_rcu_flavor(rsp)
3777 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3779 per_cpu(rcu_cpu_started, cpu) = 0;
3782 /* Migrate the dead CPU's callbacks to the current CPU. */
3783 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3785 unsigned long flags;
3786 struct rcu_data *my_rdp;
3787 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3788 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3791 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3792 return; /* No callbacks to migrate. */
3794 local_irq_save(flags);
3795 my_rdp = this_cpu_ptr(rsp->rda);
3796 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3797 local_irq_restore(flags);
3800 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3801 /* Leverage recent GPs and set GP for new callbacks. */
3802 needwake = rcu_advance_cbs(rsp, rnp_root, rdp) ||
3803 rcu_advance_cbs(rsp, rnp_root, my_rdp);
3804 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3805 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3806 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3807 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3809 rcu_gp_kthread_wake(rsp);
3810 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3811 !rcu_segcblist_empty(&rdp->cblist),
3812 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3813 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3814 rcu_segcblist_first_cb(&rdp->cblist));
3818 * The outgoing CPU has just passed through the dying-idle state,
3819 * and we are being invoked from the CPU that was IPIed to continue the
3820 * offline operation. We need to migrate the outgoing CPU's callbacks.
3822 void rcutree_migrate_callbacks(int cpu)
3824 struct rcu_state *rsp;
3826 for_each_rcu_flavor(rsp)
3827 rcu_migrate_callbacks(cpu, rsp);
3832 * On non-huge systems, use expedited RCU grace periods to make suspend
3833 * and hibernation run faster.
3835 static int rcu_pm_notify(struct notifier_block *self,
3836 unsigned long action, void *hcpu)
3839 case PM_HIBERNATION_PREPARE:
3840 case PM_SUSPEND_PREPARE:
3841 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3844 case PM_POST_HIBERNATION:
3845 case PM_POST_SUSPEND:
3846 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3847 rcu_unexpedite_gp();
3856 * Spawn the kthreads that handle each RCU flavor's grace periods.
3858 static int __init rcu_spawn_gp_kthread(void)
3860 unsigned long flags;
3861 int kthread_prio_in = kthread_prio;
3862 struct rcu_node *rnp;
3863 struct rcu_state *rsp;
3864 struct sched_param sp;
3865 struct task_struct *t;
3867 /* Force priority into range. */
3868 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3870 else if (kthread_prio < 0)
3872 else if (kthread_prio > 99)
3874 if (kthread_prio != kthread_prio_in)
3875 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3876 kthread_prio, kthread_prio_in);
3878 rcu_scheduler_fully_active = 1;
3879 for_each_rcu_flavor(rsp) {
3880 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3882 rnp = rcu_get_root(rsp);
3883 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3884 rsp->gp_kthread = t;
3886 sp.sched_priority = kthread_prio;
3887 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3889 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3892 rcu_spawn_nocb_kthreads();
3893 rcu_spawn_boost_kthreads();
3896 early_initcall(rcu_spawn_gp_kthread);
3899 * This function is invoked towards the end of the scheduler's
3900 * initialization process. Before this is called, the idle task might
3901 * contain synchronous grace-period primitives (during which time, this idle
3902 * task is booting the system, and such primitives are no-ops). After this
3903 * function is called, any synchronous grace-period primitives are run as
3904 * expedited, with the requesting task driving the grace period forward.
3905 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3906 * runtime RCU functionality.
3908 void rcu_scheduler_starting(void)
3910 WARN_ON(num_online_cpus() != 1);
3911 WARN_ON(nr_context_switches() > 0);
3912 rcu_test_sync_prims();
3913 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3914 rcu_test_sync_prims();
3918 * Helper function for rcu_init() that initializes one rcu_state structure.
3920 static void __init rcu_init_one(struct rcu_state *rsp)
3922 static const char * const buf[] = RCU_NODE_NAME_INIT;
3923 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3924 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3925 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3927 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3931 struct rcu_node *rnp;
3933 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3935 /* Silence gcc 4.8 false positive about array index out of range. */
3936 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3937 panic("rcu_init_one: rcu_num_lvls out of range");
3939 /* Initialize the level-tracking arrays. */
3941 for (i = 1; i < rcu_num_lvls; i++)
3942 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
3943 rcu_init_levelspread(levelspread, num_rcu_lvl);
3945 /* Initialize the elements themselves, starting from the leaves. */
3947 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3948 cpustride *= levelspread[i];
3949 rnp = rsp->level[i];
3950 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3951 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3952 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3953 &rcu_node_class[i], buf[i]);
3954 raw_spin_lock_init(&rnp->fqslock);
3955 lockdep_set_class_and_name(&rnp->fqslock,
3956 &rcu_fqs_class[i], fqs[i]);
3957 rnp->gp_seq = rsp->gp_seq;
3958 rnp->gp_seq_needed = rsp->gp_seq;
3959 rnp->completedqs = rsp->gp_seq;
3961 rnp->qsmaskinit = 0;
3962 rnp->grplo = j * cpustride;
3963 rnp->grphi = (j + 1) * cpustride - 1;
3964 if (rnp->grphi >= nr_cpu_ids)
3965 rnp->grphi = nr_cpu_ids - 1;
3971 rnp->grpnum = j % levelspread[i - 1];
3972 rnp->grpmask = 1UL << rnp->grpnum;
3973 rnp->parent = rsp->level[i - 1] +
3974 j / levelspread[i - 1];
3977 INIT_LIST_HEAD(&rnp->blkd_tasks);
3978 rcu_init_one_nocb(rnp);
3979 init_waitqueue_head(&rnp->exp_wq[0]);
3980 init_waitqueue_head(&rnp->exp_wq[1]);
3981 init_waitqueue_head(&rnp->exp_wq[2]);
3982 init_waitqueue_head(&rnp->exp_wq[3]);
3983 spin_lock_init(&rnp->exp_lock);
3987 init_swait_queue_head(&rsp->gp_wq);
3988 init_swait_queue_head(&rsp->expedited_wq);
3989 rnp = rcu_first_leaf_node(rsp);
3990 for_each_possible_cpu(i) {
3991 while (i > rnp->grphi)
3993 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3994 rcu_boot_init_percpu_data(i, rsp);
3996 list_add(&rsp->flavors, &rcu_struct_flavors);
4000 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4001 * replace the definitions in tree.h because those are needed to size
4002 * the ->node array in the rcu_state structure.
4004 static void __init rcu_init_geometry(void)
4008 int rcu_capacity[RCU_NUM_LVLS];
4011 * Initialize any unspecified boot parameters.
4012 * The default values of jiffies_till_first_fqs and
4013 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4014 * value, which is a function of HZ, then adding one for each
4015 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4017 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4018 if (jiffies_till_first_fqs == ULONG_MAX)
4019 jiffies_till_first_fqs = d;
4020 if (jiffies_till_next_fqs == ULONG_MAX)
4021 jiffies_till_next_fqs = d;
4023 /* If the compile-time values are accurate, just leave. */
4024 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4025 nr_cpu_ids == NR_CPUS)
4027 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4028 rcu_fanout_leaf, nr_cpu_ids);
4031 * The boot-time rcu_fanout_leaf parameter must be at least two
4032 * and cannot exceed the number of bits in the rcu_node masks.
4033 * Complain and fall back to the compile-time values if this
4034 * limit is exceeded.
4036 if (rcu_fanout_leaf < 2 ||
4037 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4038 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4044 * Compute number of nodes that can be handled an rcu_node tree
4045 * with the given number of levels.
4047 rcu_capacity[0] = rcu_fanout_leaf;
4048 for (i = 1; i < RCU_NUM_LVLS; i++)
4049 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4052 * The tree must be able to accommodate the configured number of CPUs.
4053 * If this limit is exceeded, fall back to the compile-time values.
4055 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4056 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4061 /* Calculate the number of levels in the tree. */
4062 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4064 rcu_num_lvls = i + 1;
4066 /* Calculate the number of rcu_nodes at each level of the tree. */
4067 for (i = 0; i < rcu_num_lvls; i++) {
4068 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4069 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4072 /* Calculate the total number of rcu_node structures. */
4074 for (i = 0; i < rcu_num_lvls; i++)
4075 rcu_num_nodes += num_rcu_lvl[i];
4079 * Dump out the structure of the rcu_node combining tree associated
4080 * with the rcu_state structure referenced by rsp.
4082 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4085 struct rcu_node *rnp;
4087 pr_info("rcu_node tree layout dump\n");
4089 rcu_for_each_node_breadth_first(rsp, rnp) {
4090 if (rnp->level != level) {
4095 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4100 struct workqueue_struct *rcu_gp_wq;
4101 struct workqueue_struct *rcu_par_gp_wq;
4103 void __init rcu_init(void)
4107 rcu_early_boot_tests();
4109 rcu_bootup_announce();
4110 rcu_init_geometry();
4111 rcu_init_one(&rcu_bh_state);
4112 rcu_init_one(&rcu_sched_state);
4114 rcu_dump_rcu_node_tree(&rcu_sched_state);
4115 __rcu_init_preempt();
4116 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4119 * We don't need protection against CPU-hotplug here because
4120 * this is called early in boot, before either interrupts
4121 * or the scheduler are operational.
4123 pm_notifier(rcu_pm_notify, 0);
4124 for_each_online_cpu(cpu) {
4125 rcutree_prepare_cpu(cpu);
4126 rcu_cpu_starting(cpu);
4127 rcutree_online_cpu(cpu);
4130 /* Create workqueue for expedited GPs and for Tree SRCU. */
4131 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4132 WARN_ON(!rcu_gp_wq);
4133 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4134 WARN_ON(!rcu_par_gp_wq);
4137 #include "tree_exp.h"
4138 #include "tree_plugin.h"